Training method and training apparatus for a yarn extrusion state detection model

Abstract

To provide a method and an apparatus for training a yarn-out state detection model.SOLUTION: A method comprises: obtaining a training image feature of a target yarn path according to a plurality of sample sub-image features and corresponding weight values obtained from multi-stage sample sub-images obtained from a third image obtained from a sample image; obtaining a training laser feature of the target yarn path according to a time-frequency feature determined according to a time-frequency representation result obtained from laser sample data; and obtaining a trained yarn-out state detection model by updating a parameter of a first detection model according to a loss function determined according to an actual result and a predicted result of state categories of the target yarn path obtained according to the training image feature and the training laser feature using the first detection model. The weight values are determined according to the width values of the yarn path included in the multi-stage sub-images.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present disclosure relates to the field of computers, and in particular, to a method and an apparatus for training a yarn extrusion state detection model.

Background Art

[0002] In the chemical fiber industry, the speed of extruding yarn from a melt spinning container is very fast, and the fiber yarn extruded from a spinneret is extremely thin and visually becomes a strip-shaped ghost. In the prior art, it is difficult to accurately confirm the state of the yarn visually. After abnormal situations such as yarn breakage and yarn floating occur, or blade cleaning is performed on the spinneret at a certain period, the quality of yarn extrusion is guaranteed.

Summary of the Invention

Problems to be Solved by the Invention

[0003] The present disclosure provides a method, an apparatus, a device, and a storage medium for detecting a yarn extrusion state in order to solve or alleviate one or more technical problems in the prior art.

Means for Solving the Problems

[0004] According to a first aspect, the present disclosure provides a method for detecting a yarn extrusion state, including: collecting a first image and laser reflection data of a target yarn path in a spinning container; obtaining image features of the target yarn path based on the first image; obtaining laser features of the target yarn path based on the laser reflection data; and obtaining the yarn extrusion state of the target yarn path based on the image features and the laser features by using a yarn extrusion state detection model.

[0005] According to a second aspect, the present disclosure provides a device for detecting a yarn extrusion state, including: A collection module for collecting a first image of the target yarn path within the spinning vessel and laser reflection data, Based on the aforementioned first image, an image feature determination module for obtaining the image features of the target thread path, A laser feature determination module for obtaining the laser features of the target thread path based on the aforementioned laser reflection data, The present invention provides a yarn extrusion state detection device comprising a detection module for obtaining the yarn extrusion state of the target yarn path based on the image features and laser features, using a yarn extrusion state detection model.

[0006] The third phase involves obtaining training image features of the target thread path based on sample images of the target thread path, Based on laser sample data of the target thread path, the training laser features of the target thread path are obtained, Using the first detection model, we obtain a prediction result of the thread extrusion state of the target thread path based on the training image features and training laser features. The loss function is determined based on the predicted and actual results of the yarn extrusion state along the target yarn path. The present invention provides a method for training a yarn extrusion state detection model, which includes updating the parameters of a first detection model based on a loss function to obtain a trained yarn extrusion state detection model. The fourth phase involves an image feature extraction module for obtaining training image features of the target thread path based on sample images of the target thread path, A laser feature extraction module for obtaining training laser features of the target thread path based on laser sample data of the target thread path, A prediction module for obtaining prediction results of the thread extrusion state of the target thread path based on training image features and training laser features using a first detection model, A loss function determination module for determining the loss function based on the predicted and actual results of the yarn extrusion state of the target yarn path, The present invention provides a training device for a yarn extrusion state detection model, comprising a training module for updating the parameters of a first detection model based on a loss function to obtain a trained yarn extrusion state detection model.

[0007] The fifth phase involves at least one processor and A memory that is communicably connected to at least one processor, The memory stores instructions that can be executed by the at least one processor, and the instructions provide electronic equipment that can be executed by the at least one processor, such that the at least one processor can be caused to perform a method according to any embodiment of the present disclosure.

[0008] The sixth aspect provides a non-temporary computer-readable storage medium that stores computer instructions used to cause a computer to perform a method according to any one embodiment of the present disclosure. [Effects of the Invention]

[0009] The beneficial effects of the invention disclosed herein include at least the following: By combining image processing and laser scanning technology, the accuracy and efficiency of yarn state detection can be effectively improved. Furthermore, by employing advanced machine learning methods, the system can be given better adaptability and intelligence, enabling early detection of abnormal conditions and improving the overall stability of the spinning process and the quality of the finished product. [Brief explanation of the drawing]

[0010] [Figure 1] This is a flowchart of a method for detecting the thread extrusion state according to one embodiment of the present disclosure. [Figure 2] This is a schematic diagram of the configuration of multiple thread paths according to one embodiment of the present disclosure. [Figure 3] This is a schematic diagram of the configuration of a sliding window according to one embodiment of the present disclosure. [Figure 4]It is a flowchart of a method for training a yarn extrusion state detection model according to an embodiment of the present disclosure. [Figure 5] It is a schematic configuration diagram of a detection device for a yarn extrusion state according to an embodiment of the present disclosure. [Figure 6] It is a schematic configuration diagram of a training device for a yarn extrusion state detection model according to an embodiment of the present disclosure. [Figure 7] It is a block diagram of an electronic device for realizing the method of an embodiment of the present disclosure.

Embodiments for Carrying Out the Invention

[0011] It should be understood that the content described in this part does not limit the key points or important features related to the embodiments of the present disclosure, nor does it limit the scope of the present disclosure. Other features of the present disclosure can be easily understood based on the following description.

[0012] In the drawings, unless otherwise specified, the same reference numerals in multiple drawings indicate the same or similar members or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only show some embodiments of the present disclosure and should not be regarded as limiting the scope of the present disclosure.

[0013] Hereinafter, the present invention will be described in more detail with reference to the drawings. In each figure, members having the same or similar functions are denoted by the same reference numerals. Although various aspects of the embodiments are shown in the drawings, unless otherwise specified, the drawings are not necessarily drawn to scale.

[0014] In addition, in order to better explain the present disclosure, many specific details are shown in the following specific embodiments. Those skilled in the art should understand that the present disclosure can be implemented similarly even without certain specific details. In some embodiments, methods, means, elements, circuits, etc. well known to those skilled in the art are not described in detail in order to emphasize the gist of the present disclosure.

[0015] In the related art, after the polyester melt is extruded from the spinneret, it cools and solidifies to form primary fibers. The primary fibers are oiled and bundled by an oiling nozzle, and then sent to a winder through a thread guide hook for winding and forming. After working for a certain period of time, the polyester material remaining around the yarn discharge holes of the spinneret gradually accumulates. After the remaining polyester material hardens, it blocks the yarn discharge holes, thus changing the fine cross-sectional shape of the melt, causing scratches on the fiber yarns, making the extruded yarns thinner, and potentially causing situations such as yarn floating and breakage. The conventional method is to perform blade cleaning on the spinneret at regular intervals to ensure the cleanliness of each yarn discharge hole in the spinneret. Such a regular maintenance policy lacks flexibility, may be performed frequently in some cases, wasting resources and time, or may be insufficient in some cases. Once yarn breakage occurs, the entire winding package has to be discarded.

[0016] Due to the very high speed of yarn extrusion, it is difficult to accurately capture the state of the yarn. Such blurring caused by high-speed movement not only affects the image quality but also reduces the accuracy of fault detection. The conventional technology cannot be visually detected unless obvious abnormal situations such as yarn floating and yarn breakage occur.

[0017] To at least partially solve one or more of the above-mentioned problems and other potential problems, embodiments of the present disclosure combine visual detection and laser scanning, reinforce image data and laser reflection data by an algorithm, and provide a method for detecting the yarn extrusion state that accurately detects the state of the yarn. According to the technical solution of the embodiments of the present disclosure, the accuracy and efficiency of yarn state detection can be effectively improved, abnormal states can be detected early, and the stability of the entire spinning process and the quality of the product can be improved.

[0018] FIG. 1 is a flowchart of a method for detecting the yarn extrusion state according to an embodiment of the present disclosure. As shown in FIG. 1, the method includes at least the following steps.

[0019] S110: The first image of the target yarn path inside the spinning vessel and laser reflection data are collected. In embodiments of this disclosure, the spinning vessel has a plurality of spinnerets and the same number of yarn paths, and the yarn path can be understood as the movement path of a set of fiber bundles corresponding to each spinneret. Specifically, as shown in Figure 2, the three first spinnerets 210 in the figure are located in the front row, and the two second spinnerets 220 are located in the back row, and the yarn path is a single, directly observable movement path of fiber bundles in the process in which a set of fiber yarns 230 extruded from one spinneret is coated with oil by an oiling nozzle 240 and bundled, and then enters the spinning cylinder via a yarn guide hook 250, but becomes difficult to observe once it enters the spinning cylinder. The fiber yarns 230 are in a ghost state before being bundled. The yarn guide hook may include an upper yarn guide hook and a lower yarn guide hook located above the oiling nozzle.

[0020] The first image of the target thread path is an image of at least one thread path captured by the camera device. Multiple thread paths can be captured at once, and the first image of the target thread path can be extracted from them. Each target thread path may be captured individually to obtain better image quality and a better imaging angle. The camera device may be a high-speed camera that captures thread path images using high-speed imaging technology. By capturing images at an extremely high frame rate, a high-speed imaging camera can reduce or eliminate ghosting caused by the rapid movement of the thread bundle. The visibility of the thread state can be improved by appropriate light illumination and imaging angle.

[0021] Laser reflection data is reflection signal data obtained by irradiating a bundle of threads along a target thread path using laser scanning technology. When a laser beam is irradiated onto a moving bundle of threads, scattering, reflection, or refraction phenomena occur. The signals generated by these optical interactions provide important information about the bundle of threads, and by processing the collected reflection signal data with algorithms, key parameters related to the state of the threads can be extracted.

[0022] S120: Based on the first image, image features of the target thread path are obtained. By using image processing algorithms (e.g., edge detection, contrast enhancement), image features of the threads in the target thread path, such as thread width, morphology, and distance between thread bundles, can be extracted from the first image.

[0023] S130: Based on the laser reflection data, the laser characteristics of the target thread path are obtained. By performing feature extraction on laser reflection data, laser characteristics of the target thread path, such as diameter, velocity, surface roughness, and optical properties due to laser scattering and reflection, can be obtained. By analyzing the distribution of reflection intensity, the thread diameter and surface condition can be estimated.

[0024] S140: Using a yarn extrusion state detection model, the yarn extrusion state of the target yarn path is obtained based on image features and laser features. The thread extrusion state of the thread path is output by processing image features and laser features using a pre-trained machine learning or deep learning model (e.g., convolutional neural network, support vector machine, etc.). For example, a convolutional neural network (CNN) model can be used for thread extrusion state detection because it excels at processing image data and can provide detailed predictions through probabilistic output.

[0025] The yarn extrusion state may include multiple state categories defined according to different possible conditions and risk levels for yarn formation. These state categories not only indicate the current state of the yarn but also suggest the risk level that could lead to yarn breakage. The following is a set of exemplary states.

[0026] State 1 - Normal: Description: The thread movement is normal, and no abnormalities are visible. Risk level: Low. There is almost no risk of thread breakage. State 2 - Slight wave: Description: The thread exhibits slight waves or vibrations, but these do not exceed the normal range. Risk level: Low to moderate. Indicates that fine-tuning of the machine's operating conditions may be necessary. State 3 - Obvious Wave Description: The yarn exhibits noticeable vibrations, which can affect yarn quality. Risk level: Moderate. Prolonged operation may cause quality issues or a small probability of thread breakage. Condition 4 - Twist: Description: The yarn surface is twisted or roughened, and the yarn's path is irregular. Risk level: Medium to high. In this case, the likelihood of yarn breakage or production stoppage increases even further. Condition 5 - Tension abnormality: Description: This may indicate that the tension of the thread is significantly abnormal, either too loose or too tight. Risk level: High. High probability of thread breakage. Condition 6 - Obvious damage: Description: Physical damage such as wear and tear and signs of impending breakage are clearly visible in the thread. Risk level: Very high. There is a very high probability of immediate thread breakage or serious product quality issues.

[0027] After obtaining the results of detecting the yarn extrusion state, the blade cleaning operation can be performed in a timely manner on the spinning vessels corresponding to the yarn path with a high risk level, depending on the operator's situation and production cycle time.

[0028] According to the technical proposals of the embodiments of this disclosure, by applying image processing and laser scanning technologies in combination, the accuracy and efficiency of yarn state detection can be effectively improved. Furthermore, by employing advanced machine learning methods, the system can be made more adaptable and intelligent, enabling early detection of abnormal conditions and improving the overall stability of the spinning process and the quality of the product.

[0029] In one possible embodiment, (S110) collecting a first image of the target yarn path in the spinning vessel and laser reflection data further includes the following steps:

[0030] S111: Control the patrol equipment to move in front of the spinning vessel to be detected. In the embodiments of this disclosure, the patrol equipment may be an AGV (Automated Guided Vehicle) that operates according to a pre-set route or command, or it may be a patrol equipment that moves along a fixed track, but is not limited thereto. The advantage of using patrol equipment is that one set of collection equipment can detect multiple spinning vessels one by one in multiple adjacent channels, and the mobile patrol equipment can adjust the imaging angle to collect better quality detection data.

[0031] S112: The camera and laser scanning devices of the patrol equipment are triggered, and a first image and laser reflection data of the target yarn path in the spinning vessel are collected. The target yarn path includes a movement path in which multiple yarn bundles ejected from one spinneret are gathered into one by a yarn guide hook.

[0032] In embodiments of this disclosure, multiple detection stations can be pre-set in front of the spinning vessel to avoid the front row yarn path obstructing the rear row yarn path and to facilitate imaging of yarn paths at different positions in the rear row at appropriate angles. That is, each detection station may be used to collect data from multiple yarn paths. After reaching a detection station, the lines corresponding to the detection station are designated as multiple target yarn paths, and the camera and laser scanning devices of the patrol equipment are triggered to sequentially collect a first image and laser reflection data for each target yarn path.

[0033] According to the technical proposal of the embodiments of this disclosure, by collecting data on the target yarn path at multiple detection stations using a mobile patrol device, better quality detection data can be collected.

[0034] In one possible embodiment, (S120) obtaining image features of the target thread path based on the first image further includes the following steps: S121: Morphological processing is performed on the first image to obtain the second image. S122: Based on the second image, image features of the target thread path are obtained. In the embodiments of this disclosure, the image structure of the thread is improved by utilizing morphological transformations (e.g., contraction, expansion), making the thread state features in the resulting second image more easily recognizable by the model.

[0035] In one possible embodiment, S121 further includes the following steps of performing morphological processing on the first image and obtaining a second image. Using the first sliding window, an expansion operation is performed on the first image based on the degree of sliding to obtain an expanded image. Using the second sliding window, a shrinkage operation is performed on the first image based on the degree of sliding, and a shrinkage image is obtained. A second image is obtained based on the morphological gradients of the expanded and contracted images.

[0036] In the embodiments of this disclosure, dilation is a morphological operation that has the effect of enlarging objects in an image. Specifically, one structural element is moved in the image as a sliding window, and when it overlaps with any part of an object (thread), a pixel can be added at the center of that structural element. As a result, the boundaries of the object expand outward, and larger structures are emphasized.

[0037] Shrinking is the opposite operation of dilation, having the effect of shrinking objects in an image. It uses a "sliding window," but the rule is that the center pixel of the window remains unchanged only if all pixels within the window are part of the object. As a result, the object's boundaries are shrunk inward, and details and small objects may be removed.

[0038] Morphological gradients are calculated by subtracting the condensed image from the condensed image. Object edges are enlarged by the condensation and reduced by the condensation, and are therefore emphasized by the difference between the two. This method is particularly suitable for emphasizing fine lines and details, such as fiber edges in thread images.

[0039] According to the technical proposals of the embodiments of this disclosure, the edges of the thread can be made clearer and more prominent by applying a morphological gradient to the thread image. This greatly contributes to subsequent feature extraction and model prediction. As a result, algorithms can more easily distinguish between the thread and the background, providing sharper thread contours. Such edge enhancement techniques are particularly advantageous for processing thread images generated under high-speed motion or low-contrast conditions, and enhance the model's ability to distinguish between thread states.

[0040] In one possible implementation, the first and second sliding windows employ different structural elements, and the vicinity and shape of the structural elements (or nuclei) depend on specific features to be extracted or emphasized from the image. After the fiber bundle is extruded from the spinneret, it gradually gathers into a bundle from top to bottom, i.e., the yarn path is approximately vertical. When processing morphologically, the thickness and surface roughness of the yarn should be emphasized. For this reason, the constituent elements used for shrinkage and expansion need to be designed according to the specific properties of the yarn. Considering that the yarn is approximately vertical and that horizontal winding and burrs need to be emphasized, the structural elements can be designed according to the following policy.

[0041] As a structural element for shrinkage, a vertically elongated rectangular structural element can be used to emphasize the longitudinal properties of the yarn and reduce the effects of burrs. Dimensionally, the height of the structural element needs to be greater than its width, which allows it to have a greater influence on the horizontal properties of the yarn during the shrinkage process. The resulting constituent elements contribute to reducing the effects of burrs by removing small horizontal protrusions on the yarn edges during shrinkage.

[0042] For the structural elements involved in expansion, similar rectangular structural elements can be used. To recover the longitudinal information of the yarn lost due to contraction, the structural elements involved in expansion may be slightly wider, but remain relatively long in the vertical direction. The resulting structural elements contribute to the recovery of the yarn's original thickness during the expansion process while avoiding excessive expansion of horizontal burrs.

[0043] One specific example is that the first sliding window is a rectangular structural element with an aspect ratio of 3:2 to 5:2, and the second sliding window is a rectangular structural element with an aspect ratio of 2:1 to 5:1. For example, as shown in Figure 3, the first sliding window has a length of 9*X pixels and a width of 5*X pixels. The second sliding window has a length of 9*X pixels and a width of 3*X pixels. In both cases, the vertical direction is the longitudinal direction, and X is a positive integer greater than 0.

[0044] In another example, the first sliding window is an inverted trapezoidal structural element with an aspect ratio of 3:2 to 5:2 and wider at the top and narrower at the bottom, with a top-to-bottom width ratio of 1.1:1 to 1.3:1. The second sliding window is an inverted trapezoidal structural element with an aspect ratio of 2:1 to 5:1 and a top-to-bottom width ratio of 1.1:1 to 1.3:1. The inverted trapezoidal structure is used to provide a form that is wider at the top and narrower at the bottom, approximating a thread path.

[0045] Furthermore, the selection of structural elements needs to be adjusted according to the specific application scenario and desired results. The optimal size and shape may need to be determined through experimentation.

[0046] According to the technical proposal of the embodiments of this disclosure, the structural element designed in the above manner can effectively emphasize the thickness and surface roughness of the thread while reducing the effects of horizontal burrs. This customized structural element design is a key part of morphological processing and can significantly improve the effectiveness of thread image processing.

[0047] In one possible implementation, S122 further includes the following steps to obtain image features of the target thread path based on the second image. S122-1: The second image is divided into multiple segments along the longitudinal direction of the target thread path to obtain multi-segment sub-images. S122-2: Feature extraction is performed on the multi-stage sub-images to obtain multiple first-image features. S122-3: Based on multiple first image features and their corresponding weight values, image features of the target thread path are obtained.

[0048] In the embodiments of this disclosure, the yarn paths gradually converge and narrow from top to bottom, eventually converging into a single path. As a result, the spacing between each fiber bundle becomes smaller closer to the bottom, making it more difficult to accurately identify the thickness and surface roughness of each bundle. Therefore, higher weight values ​​can be sequentially assigned to the first image features corresponding to the sub-images of each stage along the yarn path direction. That is, higher weight values ​​are assigned to the first image features extracted from multiple sub-images at the top of the yarn path, and lower weight values ​​are assigned to the first image features extracted from sub-images at the bottom of the yarn path.

[0049] According to the technical invention of the present disclosure, by assigning corresponding weight values ​​to the first image features extracted from the sub-images of each stage based on the morphological features of the thread path, the model contributes to drawing further attention to the image features of the upper region of the thread path.

[0050] In laser scanning detection, threads with different extrusion conditions will show clear differences in laser reflection or scattering data. These differences mainly stem from changes in the physical properties and motion of the threads, and specifically include the following:

[0051] (1) Yarn diameter: Yarns of different specifications may have different diameters, and the pore sizes of the micropores in the spinning plates used may differ. For example, under normal conditions, the diameter of yarn should be uniform and consistent. On the other hand, as the yarn extrusion holes gradually become blocked, the yarn diameter may change abnormally. (ii) Surface roughness: A yarn in normal condition usually has a smooth surface, but a problematic yarn may exhibit high surface roughness, which can affect the scattering pattern of the laser beam. (3) Scattering intensity: The scattering intensity changes as the surface properties of the yarn change. On a smooth surface, the scattering signal is weak and uniform, but on a rough or irregular surface, it causes amplification and pattern changes in the scattering signal. (iv) Scattering angle: During laser scanning, the scattering angle of threads in different states may differ due to differences in surface characteristics and shape. For example, the scattering angle of a normal thread is relatively stable, but the scattering angle of an abnormal thread (e.g., with altered surface roughness or cross-sectional shape) may be irregular. (5) Spectral characteristics: The spectral characteristics that occur when laser light interacts with the thread may also differ depending on the state of the thread. Depending on the physical state, the spectrum of the laser light may change slightly. (vi) Continuity of the reflected beam: A normal thread will have a continuous and uniform reflection during laser scanning, but a discontinuous thread may cause interruptions or irregularities in the reflection.

[0052] In short, by analyzing the differences in this laser scanning data, it is possible to effectively identify and distinguish abnormal characteristics of fibers, providing important information for monitoring the spinning process and controlling quality.

[0053] In one possible implementation, S122-3 further includes the following steps to obtain image features of the target thread path based on a plurality of first image features and corresponding weight values. The weight values ​​for the multi-stage subimages are determined based on the width values ​​of the thread paths included in the multi-stage subimages. Based on the weight values ​​and first image features of the multi-stage sub-images, the second image features of the multi-stage sub-images are obtained. Based on the second image features of the multi-stage sub-images, the image features of the target thread path are obtained.

[0054] In the embodiments of this disclosure, corresponding weight values ​​can be set based on the width values ​​of the thread path in each sub-image. The width values ​​may be the maximum width, minimum width, or average width. The first image feature is multiplied by the weight value to obtain the second image feature. Furthermore, the second image features of the multi-stage sub-images are integrated to finally obtain the image feature of the target thread path.

[0055] In one possible implementation, S130 further includes the following steps to obtain the laser characteristics of the target thread path based on the laser reflection data.

[0056] S131: A short-time Fourier transform is performed on the laser reflection data to obtain the time-frequency representation result. In the embodiments of this disclosure, a short-time Fourier transform (STFT) is performed on laser reflection data, a sliding window is applied to the signal, and a Fourier transform is performed on the signal within each window to obtain the time-frequency representation of the signal. Time locality is guaranteed, and the time-dependent frequency components of the signal can be observed.

[0057] S132: Determine the time-frequency features based on the time-frequency representation results. The time-frequency characteristics of laser reflection data are analyzed using the STFT result, i.e., a time-frequency graph. The time-frequency graph shows the time-dependent changes in different frequency components of the signal, providing frequency information of the signal at different points in time. The time-frequency characteristics to be identified specifically include the following:

[0058] Identifying the distribution of frequency components in a time-frequency graph involves determining which frequencies are dominant in the signal and how these frequencies change over time. For example, in laser scanning data of a thread, the prominence of certain frequencies may indicate a particular physical state or change.

[0059] The energy distribution is analyzed to evaluate the energy of frequency components at different time points. In a time-frequency graph, regions with higher energy typically appear as brighter regions. The energy distribution can indicate the dynamic characteristics of the signal, such as the stability of the signal or the occurrence of abnormal conditions.

[0060] Observe the changes over time and pay attention to the changes in frequency components over time. For example, abrupt changes in frequency may indicate a change in the thread extrusion state or some other abnormality.

[0061] S133: Based on the time-frequency characteristics, the laser characteristics of the target thread path are obtained. In the embodiments of this disclosure, important information is extracted and laser characteristics of the target thread path are obtained by performing further operations and analyses on the time-frequency features obtained from the time-frequency graph. Specifically, this may include identifying energy peaks within a specific frequency range and calculating the durations of different frequency components.

[0062] According to the technical invention of the embodiment of this disclosure, a wealth of information can be extracted from the time-frequency graph, which contributes to accurately identifying the behavior and state of the thread during laser scanning and achieving quality control.

[0063] In one possible implementation, S133 further includes the following steps to obtain the laser characteristics of the target thread path based on time-frequency features. The principal frequency features, energy peak features, and dynamic change features are extracted from the time-frequency features. Based on the dominant frequency characteristics, energy peak characteristics, and dynamic change characteristics, the laser characteristics of the target thread path are obtained.

[0064] In the embodiments of this disclosure, further feature extraction should be performed on the time-frequency features so that the extracted features capture important attributes of the signal, and the specific steps include the following: Extract the main frequency components and determine the specific values ​​of the main frequencies, as well as their duration and intensity within the overall signal. A time-frequency graph identifies energy peaks that indicate particularly high signal intensity at specific times and frequencies. Recording the location (time and frequency) and magnitude of these peaks helps identify specific characteristics and anomalies in the signal. We analyze the time-dependent changes in frequency components. We focus on how the frequency components increase, decrease, or stabilize over time. Any abrupt changes or irregular patterns in the identified frequencies may indicate an abnormal state in the thread.

[0065] Through the steps described above, multiple effective time-frequency features can be extracted from the laser signal and used to create the laser features.

[0066] In one possible implementation, S140 further includes the following steps to obtain the thread extrusion state of the target thread path based on image features and laser features using a thread extrusion state detection model.

[0067] S141: Based on image features and laser features, the thread extrusion state features of the target thread path are obtained by fusion. S142: Using a yarn extrusion state detection model, the yarn extrusion state of the target yarn path is obtained based on the yarn extrusion state characteristics.

[0068] In embodiments of this disclosure, data fusion technology can be used to integrate image features and laser features to combine datasets from two different sources and combine features from two groups into a single feature vector. The yarn extrusion state detection model is pre-trained, and by training the machine learning model, various state categories of yarn extrusion states are identified. This requires a certain amount of yarn path image data to train the model, and the trained yarn extrusion state detection model can identify yarn extrusion states quickly and accurately and maintain high accuracy even under complex or changing conditions.

[0069] The main types of yarns relating to the technical proposals of the embodiments of this disclosure may include one or more types such as partially oriented yarns (POY), fully drawn yarns (FDY), and drawn textured yarns (DTY) (or referred to as low-stretch yarns). For example, the types of yarns specifically include polyester partially oriented yarns, polyester fully drawn yarns, and polyester drawn yarns. It may also contain yarns, polyester low-stretch yarns (Polyester Draw Textured Yarns), etc.

[0070] This disclosure further provides a method for training a yarn extrusion state detection model. Figure 4 is a flowchart of the method for training a yarn extrusion state detection model. As shown in Figure 4, the method includes the following steps.

[0071] S410: Based on sample images of the target thread path, training image features of the target thread path are obtained. S420: Based on the laser sample data of the target thread path, the training laser features of the target thread path are obtained. S430: Using the first detection model, a prediction result of the thread extrusion state of the target thread path is obtained based on the training image features and training laser features. S440: Determine the loss function based on the predicted and actual results of the yarn extrusion state of the target yarn path. S450: Based on the loss function, the parameters of the first detection model are updated to obtain a trained yarn extrusion state detection model.

[0072] In one example, the structure of the first detection model comprises an input layer, a convolutional layer, a pooling layer, a fusion layer, a fully connected layer, and an output layer.

[0073] The input layer receives processed image data and laser reflection data, which may require appropriate transformation and standardization to adapt to the network input. Multiple convolutional layers can effectively extract spatial features from an image. Each convolutional layer extracts features from a different layer using a set of learnable filters. An activation function (e.g., ReLU) is used to enhance the nonlinearity of the network. Pooling layers reduce the number of parameters and computational complexity by adding a pooling layer (e.g., a max pooling layer) after a convolutional layer to reduce the spatial size of the feature map. A fusion layer is set up after convolution and pooling to combine features from the image and laser data. This can be achieved by stitching or other fusion techniques. One or more fully connected layers are used at the ends of the network to integrate previously extracted features. In the fully connected layers, the network model learns how to classify based on the extracted features. The output layer can provide multiple state-level probability distributions using the softmax activation function. The output of each neuron represents the probability that the thread is at a specific state level.

[0074] In one possible implementation, the sample image and laser sample data of the target yarn path may be pre-collected and manually labeled. Alternatively, they may be normal images and laser reflection data collected by patrol equipment during routine patrols. After yarn breakage occurs in the winding machine, or after a quality problem is detected in the subsequent visual inspection flow of the wound yarn package and its quality level is reduced, the corresponding image data and laser data collected during the patrol may be retrospectively used as sample data for abnormal problems.

[0075] In one possible implementation, S410 obtains image features of the target thread path based on the sample image, which includes the following: A third image is obtained by performing morphological processing on the sample image. Based on the third image, the training image features of the target thread path are obtained.

[0076] In one possible implementation, performing morphological processing on a sample image to obtain a third image includes the following: Using the first sliding window, an expansion operation is performed on the sample image based on the degree of sliding to obtain an expanded image. Using the second sliding window, a shrinkage operation is performed across the sample image based on the degree of sliding, and a shrinkage image is obtained. A third image is obtained based on the morphological gradients of the expanded and contracted images.

[0077] In one possible implementation, based on the third image, the training image features of the target thread path can be obtained. The third image is divided into multiple sections along the longitudinal direction of the target thread path to obtain multi-section sample sub-images. Feature extraction is performed on multi-stage sample subimages to obtain features from multiple sample subimages. Based on multiple sample sub-image features and their corresponding weight values, the first training image features of the target thread path are obtained.

[0078] In one possible implementation, the first training image feature of the target thread path is obtained based on multiple sample sub-image features and corresponding weight values. The weight values ​​for multi-stage sample subimages are determined based on the thread path width values ​​contained in multiple sample subimages. Based on the weight values ​​and training image features of the multi-stage sample subimages, a second training image feature of the multi-stage sample subimages is obtained. Based on the second training image features of the multi-stage sub-images, the training image features of the target thread path are obtained.

[0079] In one possible implementation, S420 obtains training laser features of the target thread path based on laser sample data of the target thread path, which includes the following: A short-time Fourier transform is performed on the laser sample data to obtain the time-frequency representation. The time-frequency features are determined based on the time-frequency representation results. Based on time-frequency characteristics, the training laser characteristics of the target thread path are obtained.

[0080] In one possible implementation, obtaining training laser features of a target thread path based on time-frequency features includes the following: The principal frequency features, energy peak features, and dynamic change features are extracted from the time-frequency features. Based on the dominant frequency features, energy peak features, and dynamic change features, the training laser features of the target thread path are obtained.

[0081] In one possible implementation, S430 uses a first detection model to obtain a prediction result of the thread extrusion state of the target thread path based on training image features and training laser features, which includes the following: Based on the training image features and training laser features, the training state features of the target thread path are obtained by fusion. Using the first detection model, we obtain a prediction of the thread extrusion state of the target thread path based on the training state features.

[0082] In one possible implementation, S440 determines the loss function based on the predicted and actual results of the yarn extrusion state of the target yarn path, which includes the following: In each iteration, the first detection model predicts the input training image features and training laser features and outputs the probability of each thread extrusion state. The model's prediction results are compared with the actual thread state (actual label), and the loss value is calculated based on the loss function. For multi-class classification problems, the cross-entropy loss function is typically used. The cross-entropy loss function can quantify the difference between the probability distribution of model predictions and the actual labels.

[0083] In one possible implementation, S450 updates the parameters of the first detection model based on the loss function to obtain a trained yarn extrusion state detection model, which includes the following: The entire training dataset is divided into multiple batches, and for each batch, the model parameters are updated using an optimizer based on the gradient of the loss function. This process typically repeats over multiple epochs (the number of times across the complete dataset). After each epoch, the model's performance usually improves. Through the above process, a trained yarn extrusion state detection model can be obtained, which can predict the state of the yarn based on image features and laser features.

[0084] Figure 5 is a schematic diagram of the configuration of a yarn extrusion state detection device according to one embodiment of the present disclosure. As shown in Figure 5, the device comprises at least a collection module 501, an image feature determination module 502, a laser feature determination module 503, and a detection module 504.

[0085] The acquisition module 501 collects a first image and laser reflection data of the target yarn path within the spinning vessel. The image feature determination module 502 obtains image features of the target thread path based on the first image. The laser feature determination module 503 obtains the laser features of the target thread path based on the laser reflection data. The detection module 504 uses a thread extrusion state detection model to obtain the thread extrusion state of the target thread path based on image features and laser features.

[0086] In one possible implementation, the collection module 501 is used for the following purposes: The patrol equipment is controlled to move in front of the spinning vessel that needs to be detected. The camera and laser scanning devices of the patrol equipment are triggered, and a first image and laser reflection data of the target yarn path inside the spinning vessel are collected. The target yarn path includes a travel path in which multiple yarn bundles ejected from a single spinneret are bundled together into one by a yarn guide hook.

[0087] In one possible implementation, the image feature determination module 502 is used for the following purposes: Morphological processing is performed on the first image to obtain the second image. Based on the second image, image features of the target thread path are obtained.

[0088] In one possible implementation, the image feature determination module 502 is used for the following purposes: Using the first sliding window, an expansion operation is performed on the first image based on the degree of sliding to obtain an expanded image. Using the second sliding window, a shrinkage operation is performed on the first image based on the degree of sliding, and a shrinkage image is obtained. A second image is obtained based on the morphological gradients of the expanded and contracted images.

[0089] In one possible implementation, the image feature determination module 502 is used for the following purposes: The second image is divided into multiple segments along the longitudinal direction of the target thread path to obtain multi-segment sub-images. Feature extraction is performed on the multi-stage sub-images to obtain multiple features from the first image. Based on multiple first image features and their corresponding weight values, image features of the target thread path are obtained.

[0090] In one possible implementation, the image feature determination module 502 is used for the following purposes: The weight values ​​for the multi-stage subimages are determined based on the width values ​​of the thread paths included in the multi-stage subimages. Based on the weight values ​​and first image features of the multi-stage sub-images, the second image features of the multi-stage sub-images are obtained. Based on the second image features of the multi-stage sub-images, the image features of the target thread path are obtained.

[0091] In one possible implementation, the laser feature determination module 503 is used for the following purposes: A short-time Fourier transform is performed on the laser reflection data to obtain the time-frequency representation. The time-frequency features are determined based on the time-frequency representation results. Based on the time-frequency characteristics, the laser characteristics of the target thread path are obtained.

[0092] In one possible implementation, the laser feature determination module 503 is used for the following purposes: The principal frequency features, energy peak features, and dynamic change features are extracted from the time-frequency features. Based on the dominant frequency characteristics, energy peak characteristics, and dynamic change characteristics, the laser characteristics of the target thread path are obtained.

[0093] In one possible implementation, the detection module 504 is used for the following purposes: Based on image features and laser features, the thread extrusion state characteristics of the target thread path are obtained by fusion. Using a yarn extrusion state detection model, the yarn extrusion state of the target yarn path is obtained based on the yarn extrusion state characteristics.

[0094] A description of the specific functions and embodiments of each module and submodule of the apparatus of the embodiments of this disclosure can be found by referring to the relevant descriptions of the corresponding steps in the method embodiments described above, but such descriptions will not be repeated here.

[0095] Figure 6 is a schematic diagram of the configuration of the training device for detecting the thread extrusion state. As shown in Figure 6, the device is An image feature extraction module 601 is used to obtain training image features of the target thread path based on sample images of the target thread path, A laser feature extraction module 602 is used to obtain training laser features of the target thread path based on laser sample data of the target thread path, A prediction module 603 is used to obtain a prediction result of the thread extrusion state of the target thread path based on the training image features and training laser features, using the first detection model. A loss function determination module 604 for determining the loss function based on the predicted and actual results of the yarn extrusion state of the target yarn path, The system includes a training module 605 for updating the parameters of the first detection model based on a loss function to obtain a trained yarn extrusion state detection model.

[0096] For a description of the specific functions and embodiments of each module and submodule of the apparatus of the embodiments of this disclosure, refer to the relevant descriptions of the corresponding steps in the method embodiments described above, and will not be repeated here.

[0097] Figure 7 is a block diagram of the configuration of an electronic device according to one embodiment of the present disclosure. As shown in Figure 7, the electronic device comprises a memory 710 and a processor 720, the memory 710 storing a computer program that can be executed by the processor 720. The number of memories 710 and processors 720 may be one or more. The memory 710 can store one or more computer programs, and when the one or more computer programs are executed by the electronic device, it causes the electronic device to execute the method according to the above embodiment of the method. The electronic device may further include a communication interface 730 for communicating with external devices and transmitting data interaction.

[0098] The memory 710, processor 720, and communication interface 730, if implemented independently, are connected to each other via a bus and can communicate with one another. The bus may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience, only one thick line is shown in Figure 7, but this does not mean that only one bus or one type of bus exists.

[0099] If selectively and specifically implemented, the memory 710, processor 720, and communication interface 730 can communicate with each other via an internal interface when integrated onto a single chip.

[0100] The processor may be a Central Processing Unit (CPU), or other general-purpose processors, or a Digital Signal Processor (Digital It should be understood that the general-purpose processor may be a Signal Processing (DSP), Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), or other programmable logic device, discrete gate, transistor logic device, discrete hardware component, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor may also be a processor that supports an Advanced Reduced Command Set Machine (ARM) architecture.

[0101] Furthermore, the memory may selectively include read-only memory and random access memory, and may further include non-volatile random access memory. The memory may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory may include ROM (Read-Only Memory), PROM (Programmable ROM), EPROM (Erasable PROM), EEPROM (Electrically EPROM), or flash memory. Volatile memory may include random access memory (RAM) used as an external cache. The above description is illustrative and not restrictive. Many forms of RAM are available. Examples include static random access memory (Static RAM, SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchlink dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus random access memory (Direct RAM BUS RAM, DR RAM).

[0102] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the flows or functions described in the embodiments of this disclosure are generated. The computer may be a general-purpose computer, a dedicated computer, a computer network, or another programmable device. The computer instructions may be stored on a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired connection (e.g., coaxial cable, optical fiber, digital subscriber line, DSL) or wireless connection (e.g., infrared, Bluetooth®, microwave, etc.). The computer-readable storage medium may be any available medium accessible by a computer, or a data storage device such as a server or data center that includes one or more available media. The available media may be magnetic media (e.g., floppy disk, hard disk, or magnetic tape), optical media (e.g., digital versatile disc, DVD), semiconductor media (e.g., solid state disk, SSD), etc. The computer-readable storage medium in this disclosure may also be a non-volatile storage medium, in other words, a non-temporary storage medium.

[0103] Those skilled in the art will understand that all or part of the steps for realizing the above embodiment may be performed by hardware, or completed by a program that instructs the relevant hardware, and that the program may be stored in a computer-readable storage medium, which may be a read-only memory, a magnetic disk, or an optical disk, etc.

[0104] In the description of embodiments of this disclosure, the terms “one embodiment,” “several embodiments,” “example,” “specific example,” or “several examples” mean that any specific features, structures, materials, or characteristics described in relation to such embodiment or example are included in at least one embodiment or example of this disclosure. Furthermore, any specific features, structures, materials, or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples. Also, a person skilled in the art may combine different embodiments or examples and features described herein, provided they are not inconsistent.

[0105] In the description of embodiments of this disclosure, unless otherwise specified, " / " means "or," for example, "A / B" can represent "A" or "B." In this specification, "and / or" is merely a relation that describes related objects, indicating that there may be three types of relations, for example, "A and / or B" can indicate three situations: "A" exists alone, "A" and "B" exist together, and "B" exists alone.

[0106] In the description of embodiments of this disclosure, the terms “first” and “second” are descriptive only and should not be understood as indicating or implying relative importance or the number of designated constituent elements. Thus, features limited by “first” and “second” may explicitly or implicitly include one or more such features. In the description of embodiments of this disclosure, unless otherwise specified, “multiple” means two or more.

[0107] The foregoing are merely illustrative examples of the Disclosure and do not limit the Disclosure. Any modifications, equivalent substitutions, or improvements made to the intent and principles of the Disclosure should be included within the scope of the Disclosure.

Claims

1. The third image is obtained by performing morphological processing on the sample image. The third image is divided into multiple sections along the length of the target thread path to obtain multi-section sample sub-images. Feature extraction is performed on the aforementioned multi-stage sample subimages to obtain multiple sample subimage features, Based on the aforementioned multiple sample sub-image features and corresponding weight values, the training image features of the target thread path are obtained. Performing a short-time Fourier transform on laser sample data to obtain a time-frequency representation result, Based on the aforementioned time-frequency representation results, the time-frequency characteristics are determined, Based on the aforementioned time-frequency characteristics, the training laser characteristics of the target thread path are obtained, Using the first detection model, a prediction result of the state category of the target thread path is obtained based on the training image features and the training laser features. The method involves determining a loss function based on the predicted state category of the target thread path and the actual thread state, wherein in each iteration, the first detection model makes a prediction based on the input training image features and training laser features, outputs the probability of each thread extrusion state as the model's prediction result, and calculates a loss value based on the loss function by comparing the model's prediction result with the actual thread state. This includes updating the parameters of the first detection model based on the loss function to obtain a trained yarn extrusion state detection model, The weight value is determined based on the width of the thread path included in the multi-stage sample sub-image. A training method for a yarn extrusion state detection model.

2. An image feature extraction module for obtaining training image features of the target thread path by performing morphological processing on a sample image to obtain a third image, dividing the third image into multiple stages along the length direction of the target thread path to obtain multi-stage sample sub-images, performing feature extraction on the multi-stage sample sub-images to obtain multiple sample sub-image features, and obtaining training image features of the target thread path based on the multiple sample sub-image features and corresponding weight values, A laser feature extraction module for performing a short-time Fourier transform on laser sample data to obtain a time-frequency representation result, determining time-frequency features based on the time-frequency representation result, and obtaining training laser features for the target thread path based on the time-frequency features, A prediction module for obtaining prediction results of the state category of the target thread path based on the training image features and the training laser features using a first detection model, A loss function determination module for determining a loss function based on the predicted state category of the target thread path and the actual thread state, wherein in each iteration, the first detection model makes a prediction based on the input training image features and training laser features, outputs the probability of each thread extrusion state as the model's prediction result, and the loss function determination module calculates a loss value based on the loss function by comparing the model's prediction result with the actual thread state. The system includes a training module for updating the parameters of the first detection model based on the loss function to obtain a trained yarn extrusion state detection model, The weight values ​​are determined based on the width values ​​of the thread paths included in the multi-stage sample subimages. A training device for detecting the thread extrusion state.

Images