A method and system for predicting the risk of a custom suit lapel body fit

Abstract

This application relates to the field of intelligent manufacturing of clothing, specifically to a method and system for predicting the fit risk of custom-made suit armholes. First, a method for predicting the fit risk of custom-made suit armholes is proposed, including: extracting key pattern geometric parameters from a pattern file, processing them to form a pattern feature vector, where the key pattern geometric parameters include at least two of the following: armhole arc length, back armhole depth, and shoulder slope angle; acquiring RGB images of the suit armhole area under simulated standard postures, locating the region of interest image, and extracting visual feature vectors; concatenating and standardizing the pattern feature vector and visual feature vector to form a fused feature vector; inputting a preset logistic regression model to calculate the risk probability value and the contribution of key pattern geometric parameters; comparing the risk probability value with a decision threshold, and generating and outputting structured warning information if it exceeds a high-risk warning threshold. Second, based on the above method, a system for predicting the fit risk of custom-made suit armholes is proposed.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing of clothing, and in particular to a method and system for predicting the fit risk of custom suit armholes. Background Technology

[0002] In the field of bespoke suit manufacturing, the armhole structure is a crucial part connecting the body and sleeves. Its design parameters must be dynamically matched with the wearer's shoulder and arm shape, the fabric's mechanical properties, and the sewing process. Armhole fit defects usually manifest as twisting, pulling, or looseness. These problems, caused by deviations between pattern design, fabric drape, and actual production conditions, severely affect wearing comfort and appearance, and are core indicators for measuring the quality of high-end bespoke suits.

[0003] Some existing technologies attempt to circumvent these shortcomings. For example, offline verification methods based on CAD design drawings assess structural rationality through statistical relationships of pattern geometric parameters, but their analysis only applies to static design data and cannot reflect actual changes during the manufacturing process. Three-dimensional human body scanning and physical simulation technologies can accurately simulate wearing conditions; however, the high equipment cost and complex data acquisition process make them difficult to scale up in fast-paced production lines. Furthermore, traditional machine vision inspection mainly targets superficial defects such as loose threads and stains, lacking the ability to identify deeper structural problems caused by mismatches between the pattern and body shape.

[0004] Therefore, how to provide a solution that can predict armhole fit risk in real time and at low cost is a technical problem that urgently needs to be solved in the current custom suit manufacturing industry. Summary of the Invention

[0005] This application first proposes a method for predicting the fit risk of custom-made suit armholes to solve the above-mentioned problems. Secondly, this application proposes a system for predicting the fit risk of custom-made suit armholes.

[0006] As the first aspect of this application, a method for predicting the fit risk of custom-made suit armholes is proposed, including... S1: Extract key pattern geometric parameters from the computer-aided design pattern file of the custom suit order. The key pattern geometric parameters include at least two of the following: armhole arc length, back armhole depth, and shoulder slope angle. Perform data cleaning and standardization on the key pattern geometric parameters to form a pattern feature vector. S2: After the sleeve-attaching process is completed in the custom suit, RGB images of the armhole area of ​​the suit are collected in a simulated wearing state; the RGB images are preprocessed to locate the region of interest (ROI) image of the armhole, and the visual feature vector is obtained based on the ROI image through the histogram of oriented gradients (HOG). S3: Compare the dimension of the pattern feature vector with the preset pattern standard dimension, and compare the dimension of the visual feature vector with the preset visual standard dimension. When the dimensions are consistent, concatenate the pattern feature vector and the visual feature vector and perform standardization processing to form a fused feature vector. S4: Input the fused feature vector into a preset logistic regression model. The logistic regression model calculates and outputs a risk probability value based on the input fused feature vector. The risk probability value is used to characterize the possibility that the armhole has a fit defect. At the same time, the contribution of the key pattern geometric parameters to the risk probability is output. S5: Compare the risk probability value with a preset decision threshold. If the risk probability value exceeds the decision threshold, generate and output a structured early warning message containing at least one of the following: order identifier, risk probability value, feature contribution analysis, and process handling suggestions.

[0007] Optionally, the data cleaning in step S1 is to check whether the values ​​of the key pattern geometric parameters are within a preset physiological and technological reasonable range; The standardization process is Z-score standardization, which calculates the mean of the geometric parameters of each key pattern based on historical order data. and standard deviation For parameter values Its standardized value The calculation formula is: .

[0008] Optionally, in step S2: S221: Convert the RGB image to the YUV color space and extract the luminance channel to obtain a grayscale image; S222: Perform histogram equalization on the grayscale image to enhance contrast, and use Gaussian filtering for smoothing and noise reduction; S223: Using a method based on template matching and / or feature point detection, the region of interest (ROI) image of the armhole is located, wherein: The template matching process involves sliding a pre-stored standard armhole region template image onto the pre-processed image to calculate the normalized cross-correlation (NCC) number, and selecting the region with the largest coefficient as the Region of Interest (ROI). The feature point detection identifies the edge feature points of the armhole using a pre-trained detection model, and the polygonal region enclosed by the edge feature points is taken as the ROI.

[0009] Optionally, in step S2, the step of obtaining the visual feature vector from the ROI image specifically includes: Calculate the gradient information of each pixel in the ROI image in the x and y directions; The complete ROI image is divided into non-overlapping cell units based on the gradient information. Adjacent cell units are grouped into blocks, and the histogram vectors of all cell units within a block are concatenated and then normalized using the L2 norm. The histogram vectors of all normalized blocks are concatenated sequentially to form the visual feature vector of fixed dimensions.

[0010] Optionally, in step S2, after forming the visual feature vector, when the dimension of the visual feature vector exceeds a preset threshold, principal component analysis (PCA) is used to perform dimensionality reduction processing to obtain a dimensionality-reduced visual feature vector. The visual feature vector or the reduced-dimensional visual feature vector is subjected to Z-score standardization so that its mean is 0 and its standard deviation is 1.

[0011] Optionally, the training method for the preset logistic regression model in S4 is as follows: S401: Collect historical custom suit order samples containing fused feature vectors and corresponding defect labels to construct a training dataset. The defect labels are determined based on production system quality inspection records or customer complaint data. S402: Define the logistic regression model: ,in, , The predicted probability of a sample in the training dataset. It is the sigmoid activation function. For the weight vector parameters, The fused feature vector of the aforementioned historical custom suit order samples. For bias terms; S403: The weight vector parameters are optimized using a binary cross-entropy loss function and gradient descent. and bias terms ; S404: Evaluate and optimize the weight vector parameters using a preset validation set. and bias terms The performance of the logistic regression model is then saved once it reaches a predetermined performance standard.

[0012] Optionally, step S404 specifically includes: (1) Initialize and set the weight vector and bias ; (2) Calculate the predicted probability of samples in the training dataset. ; (3) Calculate the loss function Regarding each parameter and gradient: ; in, The feature dimension of the fused feature vector; for The weight vector; For bias terms; (4) Update the parameters along the negative gradient direction. ; Updated to: ;in, The learning rate; (5) Repeat steps (2)-(4) until the logistic regression model converges to the predetermined performance standard.

[0013] Optionally, The decision threshold is set to 0.68, representing the risk probability value. When <0.68, output a low-risk warning; when When the value is ≥0.68, a high-risk warning is issued.

[0014] Optionally, the structured early warning information in step S5 may further include spatiotemporal context information, which includes at least one of the process, production line, workstation number, and timestamp that triggered the early warning.

[0015] As a second aspect of this application, based on the aforementioned method for predicting the fit risk of a custom-made suit armhole, a system for predicting the fit risk of a custom-made suit armhole is proposed, comprising: The pattern data processing module is used to receive key pattern geometric parameters and process them to form pattern feature vectors. The image acquisition and processing module is used to acquire and process RGB images of the armhole area after the sleeve-attaching process of a custom-made suit is completed, forming a visual feature vector. The feature fusion module is used to verify, splice, and standardize the pattern feature vector and the visual feature vector to form a fused feature vector. The risk prediction module is used to input the fused feature vector into a preset logistic regression model, calculate and output the risk probability value and the contribution of key pattern geometric parameters to the risk probability; The early warning output module is used to compare the risk probability value with a preset decision threshold and output structured early warning information based on the comparison result.

[0016] The beneficial effects of this application are as follows: 1. By combining the geometric parameters of the CAD pattern with the visual features of the armhole area after manufacturing, a fused feature vector reflecting the "design intent" and "actual manufacturing state" is constructed. A logistic regression model is then applied to this feature vector to obtain the risk probability of armhole fit defects. This method can provide interpretable risk warnings in the early stages of production, thereby reducing rework and improving work efficiency.

[0017] 2. By performing color space conversion, histogram equalization, and Gaussian filtering on the RGB image, the image contrast is effectively enhanced and noise is suppressed. Template matching or feature point detection is used to locate the ROI. The above preprocessing steps provide quality assurance for visual feature extraction and help to improve the accuracy of subsequent risk prediction.

[0018] 3. By using gradient histograms, key local shape and edge information of the armhole area can be captured. It has good robustness to changes in lighting and geometric deformation, and can stably reflect the visual representation of fit defects such as pleats and twists.

[0019] 4. When the dimension of the visual feature vector exceeds a preset threshold, PCA dimensionality reduction is used to remove feature redundancy and reduce computational cost; the dimensionality-reduced visual features are then standardized using Z-score to unify the feature scale. The visual features processed in the above way have more stable quality in the fusion stage, and the model training efficiency is also improved accordingly.

[0020] 5. Defect tags are derived from quality inspection records or customer complaint data in the production system. Their distribution more closely resembles the defect patterns on the actual production line, resulting in a more accurate model for predicting actual orders. This solution employs a logistic regression model, which inherently possesses linear interpretability. After training, the weight of each fused feature directly reflects its contribution to the key geometric parameters of the pattern, facilitating the tracing and analysis of key factors influencing the results during the prediction phase. Furthermore, a pre-defined validation set is used to evaluate model performance. Only models meeting set criteria (e.g., cross-validation accuracy of at least 85%) are saved and deployed, ensuring the reliability and stability of prediction results after deployment.

[0021] 6. Using the point where the F1 score on the precision-recall curve is maximized as the baseline threshold, and setting multi-level decision thresholds based on this, tiered alerts are output. Multi-level thresholds better balance precision and recall than a single fixed threshold, reducing false negatives and false positives.

[0022] 7. Extract the top K feature dimensions that contribute most to the risk probability, convert them into natural language descriptions, and match them with corresponding process control suggestions. In this way, the early warning results not only provide risk probability but also intuitive causal analysis and improvement directions, making it easier for process personnel to operate and use, reducing additional analysis and research time, and improving production efficiency.

[0023] 8. The structured early warning information includes the triggering process, production line / workstation number, and timestamp, which provides a spatiotemporal context for tracing the specific stage at which the problem occurred. Combined with automatically generated process handling suggestions, it can shorten the production line response and processing time. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for predicting the fit risk of a custom-made suit armhole in this specific embodiment. Detailed Implementation

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] As the first aspect of this specific embodiment, such as Figure 1 As shown, a method for predicting the fit risk of custom-made suit armholes is proposed, including the following steps: S1: Extract key pattern geometric parameters from the computer-aided design pattern file of the custom suit order. The key pattern geometric parameters include at least two of the following: armhole arc length, back armhole depth, and shoulder slope angle. Perform data cleaning and standardization on the key pattern geometric parameters to form a pattern feature vector. S2: After the sleeve-attaching process is completed in the custom suit, RGB images of the armhole area of ​​the suit are collected in a simulated wearing state; the RGB images are preprocessed to locate the region of interest (ROI) image of the armhole, and the visual feature vector is obtained based on the ROI image through the histogram of oriented gradients (HOG). S3: Compare the dimensions of the pattern feature vector with the preset pattern standard dimensions, and compare the dimensions of the visual feature vector with the preset visual standard dimensions. When the dimensions are consistent, concatenate the pattern feature vector and the visual feature vector and perform standardization processing to form a fused feature vector. S4: Input the fused feature vector into the preset logistic regression model. The logistic regression model calculates and outputs the risk probability value based on the input fused feature vector. The risk probability value is used to characterize the possibility of fit defects in the armhole. At the same time, the contribution of key pattern geometric parameters to the risk probability is output. S5: Compare the risk probability value with the preset decision threshold. If the risk probability value exceeds the decision threshold, generate and output a structured early warning message containing at least one of the following: order identifier, risk probability value, feature contribution analysis, and process handling suggestions.

[0027] The above solution constructs a fused feature vector reflecting both the "design intent" and the "actual manufacturing state" by combining the geometric parameters of the CAD pattern with the visual characteristics of the armhole area after manufacturing. A logistic regression model is then applied to this feature vector to obtain the probability of armhole fit defects. This method can provide interpretable risk warnings in the early stages of production, thereby reducing rework and improving work efficiency.

[0028] The specific steps in step S1 are as follows: S11: Analyze the computer-aided design pattern file (CAD) of the custom suit order and extract key pattern geometric parameters.

[0029] Specifically, the system obtains and parses the CAD pattern files for customized suit orders through a pre-defined data interface. The CAD pattern file formats include, but are not limited to, industry-standard formats such as DXF (Drawing Exchange Format), AAMA (American Apparel Manufacturers Association) standard formats, or proprietary formats of specific apparel CAD software (such as Richpeace, Gerber, and Lectra). Geometric metadata from the CAD pattern files can be read using built-in or invoked parsing libraries; for example, the ezdxf library can be used for DXF format.

[0030] Key pattern geometry parameters are located and extracted from the structured data layer of the parsed file. These parameters include: armhole arc length, back armhole depth, shoulder slope angle, chest width, and back width. It should be noted that at least two of the armhole arc length, back armhole depth, and shoulder slope angle are key pattern geometry parameters, while chest width and back width are optional. The extraction of key pattern geometry parameters can be achieved by parsing the coordinates of the endpoints of corresponding line segments and arcs in the CAD file, or by analyzing the dimensions already annotated. For example, the armhole arc length can be obtained by calculating the arc path through numerical integration.

[0031] The specific definitions of armhole arc length, back armhole depth, shoulder slope angle, chest width, and back width are as follows: Armhole arc length: defined as the length of the complete arc of the armhole curve from the front shoulder point through the armpit point to the back shoulder point. This parameter directly determines the matching length of the armhole arc and is one of the key indicators affecting the fit of the armhole circumference.

[0032] Back armhole depth: This is the distance from the bottom of the armhole vertically upwards along the side seam to the reference horizontal line of the back armhole. This parameter reflects the amount of space required for back movement. Insufficient depth can easily lead to tightness in the back, while excessive depth may cause excess fabric to accumulate under the armpits.

[0033] Shoulder slope angle: This is the angle of inclination of the shoulder line relative to the horizontal reference line. This parameter affects the fit of the shoulder and the visual streamline. Abnormal angles (such as excessive sloping or shrugging of the shoulders) will change the stress distribution of the armhole.

[0034] Chest width: This is the horizontal distance between the two armpit points on the front of the chest. This parameter is related to the volume of the thoracic cavity and affects the space in front of the armhole.

[0035] Back width: This is the horizontal distance between the two armpit points on the back. This parameter is related to the back muscles and activity level, and affects the space at the back of the armhole.

[0036] S12: Perform data cleaning and standardization on the key pattern geometry parameters.

[0037] The data cleaning process involves checking whether the values ​​are within a reasonable physiological and technological range (for example, the armhole arc length of an adult suit is usually between 40cm and 60cm). Abnormal values ​​that are significantly outside the range are marked. In some implementations, the historical average of the same model can be used for replacement. In other implementations, manual review is triggered for professional personnel to make a judgment.

[0038] The standardization process specifically involves converting all parameter values ​​to a unified unit (such as millimeters) and processing them using the Z-score (standard score) standardization method. The mean value of each parameter is then calculated based on historical order data. ) and standard deviation ( For any parameter value Its standardized value The calculation formula is: .

[0039] Standardization eliminates the impact of differences in the dimensions and numerical ranges of different parameters on subsequent models, bringing all features to the same scale and facilitating efficient learning by models such as logistic regression.

[0040] It should be noted that the Z-score standardization process described above is only one implementation method. Depending on the specific data distribution or model requirements, other standardization methods such as min-max standardization and decimal scaling standardization can be selected for processing.

[0041] Data cleaning and outlier removal are performed on key pattern geometric parameters, and Z-score standardization is performed based on the mean and standard deviation of historical order data. This unifies the dimensions and distribution scale of each parameter, improves the data quality and consistency of pattern feature vectors, and enhances the stability of subsequent feature fusion and model prediction accuracy.

[0042] S13: Construct the pattern feature vector.

[0043] N key pattern geometric parameters ( Arrange them in a predetermined order (e.g., [armhole arc length, back armhole depth, shoulder slope angle, etc.]) and combine them into an N-dimensional real vector. Understandably, this real vector is the pattern feature vector. In some implementations, the pattern feature vector is constructed using parameter values ​​that have undergone data cleaning and parameter standardization.

[0044] In one specific embodiment, including three key pattern geometric parameters (three dimensions): armhole arc length: 0.5; back armhole depth: -0.2; shoulder slope angle: 1.1, its pattern feature vector is: .

[0045] The specific steps in step S2 are as follows: S21: After the sleeve-attaching process is completed for the custom-made suit, fix the suit in a standard posture that simulates human wearing, and capture the RGB image of its armhole area.

[0046] Regarding the acquisition of RGB images, the simulated wearing state can be either the actual wearing state of a human body or a state formed by simulating human clothing through other humanoid simulated physical structures. A single image can be obtained by acquiring images of the armhole area through an image acquisition system. The image acquisition system includes an image acquisition station set after the sleeve-attaching process. The image acquisition station is equipped with an imaging unit, an illumination unit, and a triggering and control system. The imaging unit includes a camera with a resolution of no less than 2 million pixels and a global shutter, such as the Basler acA2000-50gm (an industrial area array camera manufactured by Basler). The illumination unit can use at least one set of ring-shaped white LED light sources.

[0047] In some implementations, the armhole area is illuminated at an angle of approximately 30 degrees. When the hanging system carrying the semi-finished suit reaches the predetermined position, a photoelectric sensor triggers the camera to automatically take a picture. In other implementations, the operator manually triggers the camera after the sleeves are attached. During the shooting, the suit is fixed in a posture simulating human wearing, and the camera's optical axis is perpendicular (or nearly perpendicular) to the plane where the armhole is located, targeting the right armhole (or uniformly the left armhole), and capturing an RGB color image. It should be noted that the RGB image is a color image containing only the three color channels: red (R), green (G), and blue (B), without any other additional channels.

[0048] S22: Preprocess the RGB image and locate the region of interest (ROI) image of the armhole.

[0049] The preprocessing of RGB images using standard libraries such as OpenCV and PIL includes the following steps: S221: Convert the RGB image to the YUV color space and extract the luminance channel to obtain a grayscale image; specifically, convert the RGB image to the YUV color space and extract its luminance (Y) channel to obtain a grayscale image. This reduces the interference of color information on subsequent shape feature extraction while preserving necessary texture and edge information.

[0050] S222: Perform histogram equalization on the grayscale image to enhance contrast, and use Gaussian filtering for smoothing and noise reduction; specifically, for the grayscale image... Histogram equalization is performed to enhance / improve contrast; and a Gaussian filter is used for mild smoothing to suppress sensor noise. The Gaussian filter kernel size can be set to 3x3. .

[0051] S223: The region of interest (ROI) image of the armhole is located using a method based on template matching and / or feature point detection.

[0052] Specifically, a method based on template matching and / or feature point detection can be used to locate the armhole region as the ROI. Template matching or feature point detection can be used independently, or both can be used simultaneously.

[0053] Specifically, the template matching method involves: pre-storing a template image of the standard armhole region, and then sliding the normalized cross-correlation (NCC) coefficient across the preprocessed image to calculate the region with the largest coefficient as the ROI.

[0054] Specifically, the feature point detection method involves using a pre-trained lightweight keypoint detection model (e.g., based on...). A convolutional neural network (CNN) identifies edge feature points of the armpit (such as shoulder points and armpit points), and the polygonal region enclosed by these edge feature points is the Region of Interest (ROI).

[0055] The extracted region of interest (ROI) image is denoted as... It serves as the direct input for visual feature extraction.

[0056] By preprocessing the RGB image, the image contrast is effectively enhanced and noise is suppressed. Template matching or feature point detection is used to locate the ROI. The above preprocessing steps provide quality assurance for visual feature extraction and help to improve the accuracy of subsequent risk prediction.

[0057] S23: Visual feature vectors are obtained from ROI images using the Histogram of Oriented Gradients (HOG) with fixed dimensions.

[0058] Understandably, the Histogram of Oriented Gradients (HOG) is robust to changes in illumination and small deformations, and it is also computationally fast. Therefore, this embodiment uses the Histogram of Oriented Gradients (HOG) descriptor pair. The representation is performed, and the calculation process outputs a feature vector of fixed dimensions.

[0059] The specific steps for obtaining the visual feature vector are as follows: (1) Calculate the gradient information of each pixel in the ROI image in the x and y directions: For the ROI image For each pixel, calculate its gradient values ​​(gradient components) in the x and y directions. and The calculation is performed using a one-dimensional discrete differential template [-1, 0, 1] and its transpose: ; ; Then the gradient magnitude of each pixel is calculated. : ; Calculate the gradient direction for each pixel. : ; To accommodate unsigned gradients, the gradient direction angle of each pixel is quantized to the range of 0-180 degrees. The gradient information includes the gradient value, gradient magnitude, and gradient direction.

[0060] (2) Divide the complete ROI image into non-overlapping cell units based on the gradient information: Divide the entire ROI image into several non-overlapping "cell units", each cell unit being a small square region (e.g., For each pixel within a cell unit, its gradient magnitude G is accumulated into the histogram interval (Bin) of that direction according to its gradient direction θ. In one implementation, the 0-180 degree range can be divided into 9 intervals of 20 degrees each.

[0061] (3) Combine adjacent cell units into blocks, concatenate the histogram vectors of all cell units within a block, and then perform L2 norm normalization: In order to improve the robustness of features to image illumination and shadows, combine adjacent cell units into larger "blocks", concatenate the histogram vectors of all cell units within a block, and then perform L2 norm normalization on the concatenated vector:

[0062] in It is a very small positive number to prevent division by zero. In addition, it is understandable that blocks can overlap. For example, a block consists of 2x2 cell units with a sliding step of 1 cell unit.

[0063] (4) Concatenate the histogram vectors of all normalized blocks in sequence to form the visual feature vector of fixed dimension: Concatenate the histogram vectors of all normalized blocks in the image in sequence to form a one-dimensional vector of fixed length, that is, the fixed-dimensional HOG image feature descriptor, which is the visual feature vector.

[0064] Understandably, the above process ensures that for any input image, as long as the ROI size and HOG parameters (cell unit size, number of orientation intervals, and block size) are fixed, the dimension of the extracted feature vector remains fixed, thus providing consistent input for subsequent comparison or classification analysis.

[0065] In some implementations, the visual feature vector generated in step S23 has a high dimension, exceeding a preset threshold (e.g., more than 1000 dimensions). In this case, dimensionality reduction techniques such as principal component analysis (PCA) can be used to reduce it to a fixed, lower dimension (e.g., 64 or 128 dimensions) to generate a dimensionality-reduced visual feature vector. This reduces computational load while improving generalization ability. Furthermore, the actual value of the preset threshold can be adjusted by those skilled in the art.

[0066] Furthermore, it is necessary to standardize the visual feature vector or the reduced-dimensional visual feature vector. Standardization refers to Z-score standardization of the visual feature vector or the reduced-dimensional visual feature vector after dimensionality reduction, so that its mean is 0 and its standard deviation is 1. The standardization parameters (mean and standard deviation) can be calculated based on the visual feature vector or the reduced-dimensional visual feature vector of the preset historical ROI image. The data of the visual feature vector or the reduced-dimensional visual feature vector of the historical ROI image can be obtained through experiments or relevant factory records, which will be understood by those skilled in the art and will not be elaborated here.

[0067] Understandably, after dimensionality reduction and standardization, a fixed dimension is obtained. The real-valued vector, i.e., the reduced-dimensional visual feature vector. For example, in an embodiment employing 64-dimensional HOG features, It can be [0.12, -0.85, 0.43, ..., 0.02], for a total of 64 values.

[0068] When the dimension of the visual feature vector exceeds a preset threshold, PCA dimensionality reduction is used to remove feature redundancy and reduce computational cost. The dimensionality-reduced visual features are then standardized using Z-scores to unify the feature scale. Visual features processed in this way exhibit more stable quality during the fusion stage, and the model training efficiency is correspondingly improved.

[0069] S3: Compare the dimensions of the pattern feature vector with the preset pattern standard dimensions, and compare the dimensions of the visual feature vector with the preset visual standard dimensions. When the dimensions are consistent, concatenate the pattern feature vector and the visual feature vector and perform standardization processing to form a fused feature vector. Specifically, it includes the following steps: S31: Feature vector dimension alignment and verification.

[0070] Before the splicing operation, the pattern feature vector is first processed. and visual feature vectors The dimensions are validated to ensure they are consistent with the feature dimensions used during model training, thereby ensuring correct model inference.

[0071] Obtain the pattern feature vectors respectively Dimensions and visual feature vectors Dimensions , The number of key pattern geometric parameters extracted in step S1. The fixed dimension is set in step S2.

[0072] Will Compared with the preset pattern standard dimensions To make a comparison, Compared with the preset visual standard dimensions A comparison is performed; if the comparison matches, the verification passes. or If the error occurs, it is determined to be an anomaly in feature extraction. The system records the error log and triggers the feature extraction process for that order to be re-executed or transferred to manual review. Preset pattern standard dimensions. The number of key geometric parameters for the pattern design is 2-5 in this embodiment; the preset visual standard dimension. This is the fixed dimension of the visual feature vector extracted by the Histogram of Oriented Gradients (HOG).

[0073] Check the pattern feature vector and visual feature vectors Check if the value contains illegal values, such as NaN (not a number) or Inf (infinity). If an illegal value is found, it is also considered an exception.

[0074] S32: Feature vector of spliced ​​pattern and visual feature vector.

[0075] After the verification is successful, and Two vectors are concatenated along a specific axis to generate a fused feature vector. .

[0076] Specifically, if , Then the concatenated vector is: ; Preferably, scientific computing libraries such as Python's NumPy can be used. The function completes the splicing operation.

[0077] Understandable, after splicing, The total dimension is For example, when =3, When =64, =67.

[0078] The advantages of using a direct splicing method are as follows: 1. Zero computational overhead: The splicing operation itself introduces almost no additional computational cost.

[0079] II. Lossless Information: Completely Preserved and Preserving the original information and avoiding information loss or distortion due to complex transformations is crucial for the interpretability of subsequent feature weights.

[0080] Third, model-friendly: Linear models such as logistic regression and their nonlinear extensions (such as those with multinomial features) are fully capable of learning in the concatenated high-dimensional space. and The interaction between them does not require manually preset fusion rules.

[0081] S33: Output fused feature vector.

[0082] although and While each feature has been standardized in its respective step, the combined features from different sources may still exhibit scale differences. To improve the stability and convergence speed of model training, it is preferable to perform global re-standardization, enabling the model to learn the weights of all features in a more balanced manner.

[0083] The re-standardization can be achieved using the same Z-score method as in steps S1 and S2, based on the fused feature vectors of all historical samples. Calculate each feature dimension mean and standard deviation ,in, , its first Dimension value Standardized as: ; Standardized parameters Saved, standardized parameters can be directly accessed. right The transformation is performed to obtain the final standardized fusion feature vector used as input in step S4. .

[0084] Understandably, the fused feature vector is a one-dimensional array of real numbers, which simultaneously contains the pattern geometry information of the customized suit and the visual appearance information of the armhole area.

[0085] In step S4, a pre-trained, highly interpretable logistic regression model is used to analyze the fused feature vector generated in step S3 and output a quantified risk probability value. Understandably, step S40 precedes step S4: pre-training the logistic regression model. The specific steps of S40 are as follows: S401: Collect historical custom-made suit order samples containing fused feature vectors and corresponding defect labels to construct a training dataset. Defect labels are determined based on production system quality inspection records or customer complaint data. Specifically: Collect M historical custom-made suit order samples, each sample Includes feature vectors and tags Where M can be 200-500, and the feature vector That is, the fused feature vector of the order. , dimension ;Label Let be a binary value, representing whether the armhole of the order ultimately has a fit defect requiring rework, defined as: This indicates the presence of a defect (positive class). This indicates no defects (negative category), which is understandable. The label is determined based on the final quality inspection records of the production system or customer complaint data.

[0086] S402: Define the logistic regression model, specifically: Logistic regression models learn a mapping from the feature space to the interval [0, 1] for an input feature vector. (dimension) The logistic regression model calculates the probability that it belongs to the positive class (defective) using the following function. : ; in, The sigmoid activation function will linearly combine... Compress to the (0, 1) interval. For feature vectors The linear weighted sum plus a bias term: ; in, For a dimension The weight vector, For scalar bias terms, and These are the parameters that the logistic regression model needs to learn through training; The fused feature vector of the historical custom suit order sample.

[0087] S403: Employs a binary cross-entropy loss function and optimizes the weight vector parameters using gradient descent. and bias terms .

[0088] In this embodiment, the binary cross-entropy loss function is used as the training objective. For the entire training set, the loss function... Defined as: ; in, The model is for the first The probability of predicting a sample as a positive class (defect) is given; the training objective is to find a set of parameters ( , ), making the loss function Minimize, that is, the probability distribution predicted by the model is closest to the true label distribution.

[0089] S404: Evaluate the performance of the logistic regression model using a preset validation set, and save it when the predetermined performance standard is met.

[0090] In this embodiment, gradient descent or its variants (such as mini-batch gradient descent or Adam optimizer) are used to iteratively optimize the parameters. .

[0091] (1) Initialization: Set the weight vector Initialize with small random numbers close to zero, such as samples obtained from a normal distribution with a mean of 0 and a standard deviation of 0.01, and bias them. Initialize to 0.

[0092] (2) Forward propagation: For the current parameter ( , ), calculate the predicted probability of samples in all training datasets. .

[0093] (3) Gradient calculation: Calculate the loss function Regarding each parameter and The gradient (partial derivative) of the logistic regression model has a relatively simple form: ; ; (4) Parameter update: Update the parameters along the negative gradient direction in standard gradient descent: ; in, The learning rate controls the step size for each parameter update. In a specific embodiment, It can be set to 0.01.

[0094] (5) Iteration and Termination: Repeat steps (2)-(4) until the predetermined number of iterations is reached, or the change in the loss function value is less than the predetermined threshold, so that the logistic regression model converges to the predetermined performance standard.

[0095] Furthermore, to prevent overfitting, an L2 regularization term can be added to the loss function when the amount of data is small. ,in It is a regularization strength hyperparameter to encourage the logistic regression model to learn smaller weights and improve its generalization ability.

[0096] Furthermore, after the logistic regression model is trained, its performance is evaluated using a pre-defined validation set or through cross-validation. Metrics such as accuracy, precision, recall, and AUC-ROC value can be used for evaluation. Once the logistic regression model reaches the predetermined performance standard (cross-validation accuracy ≥ 85%), the final model parameters are... Feature standardization parameters and decision threshold Serialized and saved for loading during the online prediction phase.

[0097] It is understandable that, for the implementation of customized orders that have completed steps S1 to S3, real-time risk prediction can be performed based on the logistic regression model trained in step S40.

[0098] Defect tags are derived from quality inspection records or customer complaint data in the production system. Their distribution more closely resembles the defect patterns on the actual production line, resulting in a more accurate prediction of actual orders for the trained model. This solution employs a logistic regression model, which inherently possesses linear interpretability. After training, the weight of each fused feature directly reflects its contribution to the key geometric parameters of the pattern design. Therefore, it can output the contribution of these key geometric parameters to the risk probability, facilitating the tracing and analysis of key factors influencing the results during the prediction phase. Furthermore, a pre-defined validation set is used to evaluate model performance. Only models that meet set criteria (e.g., cross-validation accuracy of at least 85%) are saved and deployed, ensuring the reliability and stability of prediction results after deployment.

[0099] Specifically, when it is necessary to predict the fit of a custom-made suit armhole in real time, the model parameters saved during the training phase can be loaded from storage. When a new customized order is received, the fused feature vector (Right now When performing a test, first check whether its dimension D is consistent with the dimension during training (i.e., the preset standard dimension).

[0100] Use the loaded weight and bias Calculate input features linear weighted sum :

[0101] Substitute z into the sigmoid function to obtain the risk probability value. : in, A continuous value between 0 and 1, used to directly quantify the probability that the current armhole has a fit defect.

[0102] Among them, when When it approaches 1, such as A value greater than 0.8 indicates that, based on the current design parameters and visual state, the model is highly suspected of having defects.

[0103] Among them, when When it approaches 0, such as When the value is less than 0.2, it indicates that the model's risk assessment is very low.

[0104] Among them, when When it is the middle value, such as When the value is 0.5, it indicates that the model's judgment is uncertain and the risk is moderate.

[0105] Furthermore, it can simultaneously output the contribution of each feature dimension to the final risk probability, for the th feature in the standardized fused feature vector. The weight coefficients corresponding to each feature are: The contribution of this feature dimension to the risk probability is approximately... The geometric parameters of each key pattern correspond to the dimensions. By sorting by numerical values, the relative contribution of each parameter to the final risk probability can be determined, allowing process managers to intuitively understand the source of risk. For example, "The risk warning is mainly due to the armhole arc length parameter being X units larger than the standard value (contribution + Y)" or "An abnormal gradient feature in the rear armhole region of the image (contribution + Z)".

[0106] In step S5, the risk probability value output in step S4 can be... This translates into specific, executable production decision instructions, and drives the generation and distribution of early warning information, forming a closed loop from risk perception to on-site intervention, enabling early and precise intervention in potential quality problems.

[0107] Specifically, it includes the following steps: S51: Set decision threshold.

[0108] Specifically, after model training is complete, evaluation is performed using a validation dataset that was not used in training. This can be achieved by plotting a precision-recall curve or a receiver operating characteristic (ROCCurve) curve, and selecting the probability value that maximizes the F1 score as the baseline threshold. Where the F1 score is the harmonic mean of precision and recall. Experiments have verified that when... The model achieves its highest F1 score on the validation set when the F1 score is 0.68. The F1 score is a metric used in machine learning and statistics to evaluate the performance of classification models.

[0109] The baseline threshold is set at the point where the F1 score on the precision-recall curve is maximized. Multiple decision thresholds are then set based on this, generating tiered alerts. Multi-level thresholds better balance precision and recall compared to a single fixed threshold, reducing false negatives and missed positives.

[0110] Furthermore, multiple threshold levels can be set to support more refined early warning classification management. Multiple thresholds can be set, each defining a different risk level.

[0111] In a preferred embodiment, the decision threshold is set to 0.68. A value <0.68 indicates a low-risk indicator, which should only be recorded or lightly indicated on the dashboard for process personnel's reference; when... A value ≥0.68 indicates a high-risk warning, triggering a formal alert and requiring mandatory review.

[0112] Furthermore, record the actual handling results after each warning, such as manual confirmation of accuracy, identification of defects and adjustment, and false alarms. Regularly (e.g., monthly) statistically analyze the warning accuracy rate. If the accuracy rate remains below the target value (e.g., 85%), adjust the threshold. For example, it could be increased to 0.70 to reduce the false alarm rate.

[0113] S52: Compare the risk probability value with the preset decision threshold. If the threshold is exceeded, generate and output a warning message.

[0114] Specifically, if ≥ , If the current warning threshold is met, the current order's sleeve hole is deemed to have a high risk, triggering the warning process and generating a warning message. Otherwise, the process ends, the order is marked as "prediction normal," and subsequent production continues.

[0115] In this embodiment, the generated early warning information is a structured data packet, and includes at least one key field such as order identifier, risk probability value, and feature contribution analysis.

[0116] The key fields are defined as follows: Order identifiers: such as order number, style code, customer ID.

[0117] Risk probability value: The specific value (e.g., 0.78) and the corresponding risk level (e.g., "high risk").

[0118] Feature contribution analysis: Analyze the feature contribution calculated in S4 and extract the features that contributed the most to this early warning. Features (K is a positive integer, such as...) =3), the contribution of the first K feature dimensions and the corresponding feature meanings are converted into natural language descriptions that are easy for process engineers to understand, such as: "The main reasons for the warning are: 1) The armhole arc length parameter is 1.2cm larger than the standard value (contribution +0.15); 2) Abnormal horizontal gradient texture is detected in the back armhole area in the image (contribution +0.10); 3) The shoulder slope angle is smaller (contribution +0.05)." The top K feature dimensions that contribute most to the risk probability are extracted, converted into natural language descriptions, and matched with corresponding process control suggestions. In this way, the early warning results, in addition to risk probability, also provide intuitive causal analysis and improvement directions, making them easier for process personnel to use, reducing additional analysis and research time, and improving production efficiency.

[0119] In some implementations, the structured early warning information in step S5 further includes spatiotemporal context information and process handling suggestions generated based on feature contribution analysis. The spatiotemporal context information includes at least one of the following: the process that triggered the early warning, the production line, the workstation number, and the timestamp. This facilitates rapid tracing and location of the specific stage where the problem occurred, and combined with automatically generated process handling suggestions, can significantly shorten response and processing time.

[0120] The structured early warning information includes the triggering process, production line / workstation number, and timestamp, which provides a spatiotemporal context for tracing the specific stage at which the problem occurred. Combined with automatically generated process handling suggestions, it can shorten the production line response and processing time.

[0121] In other implementations, the data packet also includes action recommendations, which may be preliminary recommendations based on historical processing data and expert rules, such as: "It is recommended to focus on checking the fit of the back armhole and try back armhole retraction." Understandably, the generated early warning information can be output through multiple channels and in multiple formats.

[0122] In some implementations, the alerts can be output via integrated push notifications from the production system. Specifically, structured alert information can be pushed in real-time to the Manufacturing Execution System (MES) or the shop floor digital dashboard system through application programming interfaces (APIs) or message brokers (such as RabbitMQ and Kafka). On the order process card in the MES, a highlighted (e.g., red) warning icon (e.g., an exclamation mark) will dynamically appear next to the "sleeve-attaching" process. Clicking the icon expands to view the detailed alert content constructed in S52.

[0123] In other implementations, the alert can also be delivered via targeted notifications. Specifically, based on preset rules, the alert information can be automatically pushed to relevant responsible persons, such as the on-duty quality inspector, team leader, or process engineer, via enterprise communication software (such as DingTalk or WeChat Work) or SMS. The notification message is a simplified version, including the order number, risk level, and a summary of the core reason.

[0124] In other implementations, local audio-visual cues can also be used. For example, at the workstation where the warning is triggered, an immediate physical cue can be given by a connected signal light (red light) or a buzzer (short beep) to guide the on-site operator's attention.

[0125] Furthermore, quality inspectors or process engineers who receive an alert can conduct an on-site review of the target sleeve based on the prompts. The MES alert interface provides action options buttons, such as "Confirmed No Error (False Alarm)," "Defect Confirmed, Adjusted," and "Defect Confirmed, Rework Required."

[0126] Understandably, regardless of the chosen approach, the final outcome of this warning, along with the person and time of action, must be recorded. If the warning is "adjusted" or "rework required," the specific measures taken can be further recorded (e.g., "retract the armhole by 2mm").

[0127] In some implementations, the collected "early warning-disposal" paired data (structured early warning information) can be added to the historical dataset as a new training dataset (feature x and label y updated according to the disposal result) to incrementally train or periodically retrain the logistic regression model, enabling the model to adapt to process changes and new body shapes.

[0128] To further elaborate on this application, 100 simulated orders generated based on historical data distribution are selected as the test set. Each order has completed CAD pattern design, and after the sleeve-attaching process, the RGB image of the right armhole area is collected according to the standard posture. The method of this application, the prediction method using CAD pattern parameters, and the prediction method using only visual images are compared and verified through the following specific experiments.

[0129] The method for predicting the fit risk of custom-made suit armholes proposed in this application (hereinafter referred to as "this method") is as follows: S1 obtains the pattern feature vector: Three key geometric parameters were extracted from the CAD pattern files of each order: armhole arc length (range 42–58 cm), back armhole depth (range 12–18 cm), and shoulder slope angle (range 10–20 degrees). Based on the historical order data of the workshop over the past 500 days, the mean (μ) and standard deviation (σ) of each parameter were calculated, and the parameters of the test set were standardized using Z-scores to obtain a 3D pattern feature vector. .

[0130] S2 obtains the visual feature vector: Each armhole image was preprocessed, and after locating the Region of Interest (ROI), a 64-dimensional Histogram of Oriented Gradients (HOG) feature was extracted. Z-score normalization was also performed based on historical image data to obtain a 64-dimensional visual feature vector. .

[0131] S3 Feature Fusion: Will and Concatenate into a 67-dimensional fused feature vector .

[0132] S4 Model Prediction and Decision Making: Will The input is fed into the pre-trained logistic regression model. This model is trained using the fused features of the aforementioned 500 historical orders and their corresponding defect labels (based on the final quality inspection records). On the independent validation set (100 orders), the model's discriminative ability AUC-ROC value is 0.89, and the cross-validation accuracy is 85%.

[0133] The model outputs risk probability values ​​for 100 orders in the test set. (Range 0 to 1).

[0134] Comparison of S5 risk probability value with decision threshold: Set the decision threshold T = 0.65 (this threshold is determined based on the principle of maximizing the F1 score on the validation set). When When the value is ≥0.65, the system generates a high-risk warning, issuing warnings for 18 out of 100 orders. Final quality inspection (manual verification) in the workshop confirms: Of the 18 orders that received warnings, 14 actually had armhole fit defects requiring rework (such as tight back armholes or wrinkles at the sleeve cap), resulting in a warning accuracy of 77.8%. Three other orders with actual defects were not flagged (missed warnings). The test set contained 17 orders with actual defects, and the method of this invention successfully detected 14 of them, achieving an overall defect detection rate of 82.4%.

[0135] Comparative Example 1: Prediction method using only CAD template parameters: We used the exact same test set of 100 orders as in this method, but only their 3D pattern feature vectors. As input, predictions were made using a similar logistic regression model trained solely on pattern features (trained on pattern data from the same 500 historical orders, with a validation set AUC-ROC of 0.72).

[0136] Using the same decision-making process as this method (threshold T=0.65), this method issued warnings for 20 orders. Final quality inspection confirmed: Of the 20 orders that received warnings, 12 had actual defects, resulting in a warning accuracy rate of 60.0%. Eight orders with actual defects went undetected. The overall defect detection rate was 70.6% (12 / 17).

[0137] Comparative Example 2: Prediction methods using only visual images: We used the exact same test set of 100 orders as in this method, but only their 64-dimensional visual feature vectors. As input, predictions are made using a similar logistic regression model trained solely on visual features (trained on the same historical image data, with a validation set AUC-ROC of 0.81).

[0138] Using the same decision-making process as this method (threshold T=0.65), this method issued warnings for 22 orders. Final quality inspection confirmed: Of the 22 orders that received warnings, 13 had actual defects, resulting in a warning accuracy rate of 59.1%.

[0139] Four genuine defective orders went undetected.

[0140] The overall defect detection rate was 76.5% (13 / 17).

[0141] Specifically, the comparison of the effects is as follows:

[0142] Comparative tests revealed that the CAD parameter method is primarily effective at detecting "inherent design defects" caused by excessive deviations between the pattern design and the standard body shape. However, its limitations are significant: firstly, it cannot detect defects introduced during manufacturing due to process execution deviations (such as inadequate shaping or uneven seam tension); secondly, it is insensitive to complex defects arising from the interaction between the customer's unique body shape and the pattern. Therefore, its defect detection rate is relatively low, and its false alarm rate is high. While the visual image method can effectively identify obvious physical defects (such as severe wrinkles or lifting) that have already appeared after manufacturing, it also has significant limitations: firstly, it is insensitive to early, minor, or potential defects (such as tension abnormalities caused by pattern mismatch that have not yet fully manifested); secondly, due to the lack of a design intent reference, it is difficult to distinguish between similar visual manifestations caused by "process execution errors" and "pattern-body shape mismatch," thus resulting in a high false alarm rate.

[0143] The comparison results between this method and Comparative Examples 1 and 2 show that single-data-source methods, due to their inherent limitations (CAD methods lack real-time manufacturing status information, and image methods lack design baseline information), are significantly insufficient in predicting sleeve fit defects, exhibiting either limited detection rates, high false alarm rates, or both. This invention, by fusing CAD pattern geometric parameters from the design phase with visual image features after manufacturing, constructs a fused feature vector that simultaneously encompasses both "design intent" and "manufacturing status," thereby enabling a more comprehensive identification of composite defects caused individually or jointly by "pattern-body mismatch" and "process execution" deviations. Therefore, this invention achieves superior and balanced performance in terms of early warning accuracy, defect detection rate, and overall model discriminative ability (AUC-ROC), verifying the effectiveness and superiority of the multi-source information fusion scheme in achieving early and accurate risk warning.

[0144] Therefore, this invention does not simply splice CAD parameters with visual features, but rather achieves a feature-level deep fusion of precise geometric parameters from the design end and real-time visual states from the manufacturing end in the specific scenario of sleeve fit prediction. This fusion produces a synergistic effect: CAD parameters provide design benchmarks and theoretical constraints for fit, while visual features reflect the actual execution state and early deformation signals during the manufacturing process. The combination of these two allows the model to effectively distinguish between defects caused by "improper design," "process deviations," and the combined effects of both—something that single-data-source methods cannot achieve. Example data shows that this invention brings non-obvious performance improvements through feature fusion, especially in improving early warning accuracy (significantly reducing false alarms) and enhancing the detection capability of complex defects, demonstrating the technological advancement and practical value of the fusion method.

[0145] The above are merely preferred embodiments of this application and do not limit the patent scope of this invention. Any equivalent structural or procedural transformations made using the content of this specification, or direct or indirect applications to other related technical fields, are included within the patent protection scope of this invention.

Claims

1. A method for predicting the fit risk of custom-made suit armholes, characterized in that, Includes the following steps: S1: Extract key pattern geometric parameters from the computer-aided design pattern file of the custom suit order. The key pattern geometric parameters include at least two of the following: armhole arc length, back armhole depth, and shoulder slope angle. Perform data cleaning and standardization on the key pattern geometric parameters to form a pattern feature vector. S2: After the sleeve-attaching process is completed in the custom suit, RGB images of the armhole area of ​​the suit are collected in a simulated wearing state; the RGB images are preprocessed to locate the region of interest (ROI) image of the armhole, and the visual feature vector is obtained based on the ROI image through the histogram of oriented gradients (HOG). S3: Compare the dimension of the pattern feature vector with the preset pattern standard dimension, and compare the dimension of the visual feature vector with the preset visual standard dimension. When the dimensions are consistent, concatenate the pattern feature vector and the visual feature vector and perform standardization processing to form a fused feature vector. S4: Input the fused feature vector into a preset logistic regression model. The logistic regression model calculates and outputs a risk probability value based on the input fused feature vector. The risk probability value is used to characterize the possibility that the armhole has a fit defect. At the same time, the contribution of the key pattern geometric parameters to the risk probability is output. S5: Compare the risk probability value with a preset decision threshold. If the risk probability value exceeds the decision threshold, generate and output a structured early warning message containing at least one of the following: order identifier, risk probability value, feature contribution analysis, and process handling suggestions.

2. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, The data cleaning in step S1 is to check whether the values ​​of the key pattern geometry parameters are within the preset physiological and technological reasonable range; The standardization process is Z-score standardization, which calculates the mean of the geometric parameters of each key pattern based on historical order data. and standard deviation For parameter values Its standardized value The calculation formula is: .

3. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, In step S2: S221: Convert the RGB image to the YUV color space and extract the luminance channel to obtain a grayscale image; S222: Perform histogram equalization on the grayscale image to enhance contrast, and use Gaussian filtering for smoothing and noise reduction; S223: Using a method based on template matching and / or feature point detection, the region of interest (ROI) image of the armhole is located, wherein: The template matching process involves sliding a pre-stored standard armhole region template image onto the pre-processed image to calculate the normalized cross-correlation (NCC) number, and selecting the region with the largest coefficient as the Region of Interest (ROI). The feature point detection identifies the edge feature points of the armhole using a pre-trained detection model, and the polygonal region enclosed by the edge feature points is taken as the ROI.

4. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, In step S2, the step of obtaining the visual feature vector from the ROI image specifically includes: Calculate the gradient information of each pixel in the ROI image in the x and y directions; The complete ROI image is divided into non-overlapping cell units based on the gradient information. Adjacent cell units are grouped into blocks, and the histogram vectors of all cell units within a block are concatenated and then normalized using the L2 norm. The histogram vectors of all normalized blocks are concatenated sequentially to form the visual feature vector of fixed dimensions.

5. The method for predicting the fit risk of custom-made suit armholes as described in claim 4, characterized in that, In step S2, after the visual feature vector is formed, when the dimension of the visual feature vector exceeds a preset threshold, principal component analysis (PCA) is used to perform dimensionality reduction to obtain a dimensionality-reduced visual feature vector. The visual feature vector or the reduced-dimensional visual feature vector is subjected to Z-score standardization so that its mean is 0 and its standard deviation is 1.

6. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, The training method for the preset logistic regression model in S4 is as follows: S401: Collect historical custom suit order samples containing fused feature vectors and corresponding defect labels to construct a training dataset. The defect labels are determined based on production system quality inspection records or customer complaint data. S402: Define the logistic regression model: ,in, , The predicted probability of a sample in the training dataset. It is the sigmoid activation function. For the weight vector parameters, The fused feature vector of the aforementioned historical custom suit order samples. For bias terms; S403: The weight vector parameters are optimized using a binary cross-entropy loss function and gradient descent. and bias terms ; S404: Evaluate and optimize the weight vector parameters using a preset validation set. and bias terms The performance of the logistic regression model is then saved once it reaches a predetermined performance standard.

7. The method for predicting the fit risk of custom-made suit armholes as described in claim 6, characterized in that, Step S404 specifically includes: (1) Initialize and set the weight vector and bias ; (2) Calculate the predicted probability of samples in the training dataset. ; (3) Calculate the loss function Regarding each parameter and gradient: ； in, The feature dimension of the fused feature vector; for The weight vector; For bias terms; (4) Update the parameters along the negative gradient direction. ; Updated to: ;in, The learning rate; (5) Repeat steps (2)-(4) until the logistic regression model converges to the predetermined performance standard.

8. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, The decision threshold is set to 0.68, representing the risk probability value. When the value is less than 0.68, a low-risk warning is displayed. when When the value is ≥0.68, a high-risk warning is issued.

9. The method for predicting the fit risk of custom-made suit armholes as described in claim 1, characterized in that, The structured early warning information in step S5 also includes spatiotemporal context information, which includes at least one of the process, production line, workstation number, and timestamp that triggered the early warning.

10. A system for predicting the fit risk of custom-made suit armholes, characterized in that, The method for predicting the fit risk of custom-made suit armholes according to any one of claims 1 to 9 includes: The pattern data processing module is used to receive key pattern geometric parameters and process them to form pattern feature vectors. The image acquisition and processing module is used to acquire and process RGB images of the armhole area after the sleeve-attaching process of a custom-made suit is completed, forming a visual feature vector. The feature fusion module is used to verify, splice, and standardize the pattern feature vector and the visual feature vector to form a fused feature vector. The risk prediction module is used to input the fused feature vector into a preset logistic regression model, calculate and output the risk probability value and the contribution of key pattern geometric parameters to the risk probability; The early warning output module is used to compare the risk probability value with a preset decision threshold and output structured early warning information based on the comparison result.

Images