A deep learning-based sewing equipment working hour intelligent detection method and system

Abstract

The application discloses a kind of based on deep learning's sewing equipment work hour intelligent detection method and system, through real-time acquisition sewing equipment motor start-stop signal and cloth tension change data by sensor, and synchronous acquisition job image sequence, after pre-processing, fusion construction multimodal dataset;The multimodal dataset is input into target detection model, accurately identifies equipment operating state and operator on-site key state, forms continuous state sequence label;Further, extract the feature vector related to cutting line action from the state sequence, use deep learning algorithm to deeply mine its time series correlation law, to intelligently judge and divide effective sewing interval and invalid time interval, realize the accurate segmentation of real production time;On this basis, the duration of each effective sewing interval is accumulated to form a single-process work hour value;The application greatly improves the work hour accounting precision and management refinement level of garment manufacturing enterprises.

Description

Technical Field

[0001] This invention relates to the field of sewing industry technology, and in particular discloses a method and system for intelligent detection of sewing equipment working hours based on deep learning. Background Technology

[0002] As a crucial component of labor-intensive manufacturing, the sewing industry's production efficiency directly impacts cost control and market competitiveness. In actual workshops, time data serves as the foundation for measuring worker efficiency, assessing process rationality, and scheduling production. Accurate time statistics help managers identify bottlenecks and optimize resource allocation, thus becoming an indispensable element in promoting refined workshop management.

[0003] Currently, most sewing companies still rely on manual recording or simple timekeeping tools to calculate working hours, a method with several significant drawbacks. Manual methods are susceptible to worker fatigue, distraction, or omissions, leading to systematic data bias. Furthermore, the sewing process involves multiple continuous actions, making it difficult for humans to accurately distinguish the start and end of each step without disrupting normal operations. It's also impossible to precisely determine which time periods constitute truly effective sewing work and which are non-value-adding activities such as waiting, finishing, and thread changing. These limitations result in low overall reliability of working hour data, consequently impacting subsequent efficiency analysis and employee performance evaluation.

[0004] Looking deeper, the complexity of the sewing process presents core technical challenges. Sewing actions are highly continuous and rapid; workers need to complete a series of operations such as threading the needle, stitching, turning, and cutting the thread in a very short time. The key to whether a garment is up to standard often lies in the subtle movements a few seconds before and after the thread cutting. If this pivotal moment of thread cutting cannot be accurately captured, it's difficult to distinguish between two separate sewing processes, making it impossible to clearly define the boundary between effective sewing time and auxiliary time. For example, when sewing a shirt cuff, a worker might continuously complete multiple straight sewing sections, interspersed with several adjustments to fabric position and thread cutting. If judged solely by motor rotation, the entire process would be mistakenly considered a continuous single process, when in reality it contains multiple independent processing units, with each thread cutting representing the end of a process node. It is precisely this lack of accurate identification of critical moments like thread cutting that makes it difficult for the system to automatically break down continuous action sequences into meaningful process units in real-world production scenarios, resulting in significant discrepancies between calculated work time and actual processing conditions.

[0005] Therefore, the key issue in achieving accurate and automatic detection of sewing equipment time is how to identify key nodes such as thread cutting from rapid and continuous sewing actions in real time without interfering with the normal operation of workers, and to automatically divide the process and distinguish between effective processing time and ineffective time consumption. Summary of the Invention

[0006] This invention provides a method and system for intelligent detection of sewing equipment working hours based on deep learning, aiming to solve at least one of the defects existing in the prior art.

[0007] One aspect of the present invention relates to a deep learning-based intelligent detection method for sewing equipment working hours, comprising the following steps:

[0008] S100: Collects start / stop signals of the sewing equipment motor and fabric tension change data through sensors, synchronously acquires operation image sequences, and merges them into a multimodal dataset after data preprocessing;

[0009] S200. Input the multimodal dataset into the target detection model to identify the key states of sewing equipment operation and operator on duty, and obtain state sequence labels;

[0010] S300. Extract the thread-cutting action-related feature vectors from the state sequence labels, use deep learning algorithms to analyze the temporal correlation between the thread-cutting action-related feature vectors, determine the temporal correlation to mark the effective sewing intervals and invalid time-consuming intervals, and obtain the interval classification results.

[0011] S400: Obtain the interval classification results and calculate the duration of each valid sewing interval, accumulate them to generate the single-process working time value, dynamically correct the deviation of historical production data, and determine the total working time calculation value.

[0012] S500 generates structured report data based on the total working hours calculation value, integrates anomaly detection logic, and uses the anomaly detection logic to judge working hour fluctuations and trigger alarm signals to obtain optimized report output.

[0013] S600 extracts key indicators from the optimized report output, interfaces with the manufacturing execution system to achieve real-time data synchronization, and updates the production control database to obtain the final time detection record if the synchronization is completed.

[0014] Further, step S100 includes:

[0015] S110: Acquire the start / stop signal of the sewing equipment motor, fabric tension change data and work image sequence, and lock the valid work window according to the motor start / stop signal;

[0016] S120. Filter the fabric tension change data within the effective working window to obtain a tension trend sequence;

[0017] S130. Extract the sewing area image frame set from the work image sequence based on the abrupt change time of the tension trend sequence;

[0018] S140. Concatenate the image frame set of the sewing area with the numerical feature vectors corresponding to the tension trend sequence to generate a multimodal dataset.

[0019] Further, step S200 includes:

[0020] S210. Obtain a multimodal dataset, extract and fuse image spatial texture features and tension temporal fluctuation features to obtain a multimodal fusion feature tensor;

[0021] S220. Input the multimodal fusion feature tensor into the region proposal network to generate a set of detection boxes, which contains target category and location information;

[0022] S230. Calculate the Euclidean distance between the limb and the head of the machine based on the detection box set, and determine the single-frame state classification code based on the tension fluctuation range;

[0023] S240. Perform context constraint modeling on the original state sequence formed by single-frame state classification coding, correct the original state sequence, and obtain the state sequence labels of sewing equipment operation and operator on duty.

[0024] Further, step S300 includes:

[0025] S310. Obtain the time point of the line-cutting action in the state sequence label, and use a long short-term memory network to map the multimodal fusion feature tensor fragment into a line-cutting action-related feature vector;

[0026] S320. Input the feature vectors related to the line-cutting action into the self-attention mechanism model to construct the temporal correlation matrix;

[0027] S330. If the value of an element in the time sequence correlation matrix is ​​greater than a preset threshold, it is determined that there is continuous operation logic; otherwise, it is determined that there is an abnormal pause.

[0028] S340. Mark the valid sewing intervals according to the continuous operation logic, and mark the invalid time intervals according to the abnormal pauses. Summarize the valid sewing intervals and invalid time intervals to obtain the interval classification results.

[0029] Further, step S400 includes:

[0030] S410. Traverse the start and end timestamps of the valid sewing intervals in the interval classification results, and calculate the difference between the start and end timestamps of the valid sewing intervals to obtain the effective sewing duration.

[0031] S420: Accumulate the duration of multiple valid sewing operations within the same production batch to generate a single-process time value;

[0032] S430. Calculate the difference between the single-process time value and the historical benchmark mean to obtain the time deviation value, wherein the historical benchmark mean is calculated based on historical production data that matches the current process code;

[0033] S440. Based on the time deviation value, determine the weighted correction coefficient through a nonlinear mapping function, and use the weighted correction coefficient to correct the time value of a single process to determine the total time calculation value.

[0034] Further, step S500 includes:

[0035] S510. Obtain the total working hours calculation value, and perform serialization processing by combining the associated production batch number and process code to generate the initial structured report dataset;

[0036] S520. Parse the process codes contained in the initial structured report dataset and retrieve the time distribution records within the historical production cycle to determine the time fluctuation threshold range.

[0037] S530. Compare the total working hours with the working hours fluctuation threshold range to calculate the working hours fluctuation range.

[0038] S540. If the fluctuation range of working hours shows an overflow, an anomaly marker is generated and embedded into the initial structured report data. The report fields are then reorganized to generate an optimized report output with integrated anomaly detection logic.

[0039] Further, step S600 includes:

[0040] S610. Parse the optimized report output to extract key indicators, and construct a standard data transmission package to be synchronized based on the key indicators.

[0041] S620: Write the standard data transmission packet to be synchronized into the real-time synchronization transmission queue, and send the data stream to the manufacturing execution system through the encrypted channel;

[0042] S630: Receive the confirmation instruction returned by the manufacturing execution system for the data stream. If the verification is successful, update the production batch record in the production control database.

[0043] S640: Read the updated stored content in the production control database and append a timestamp to obtain the final working hour detection record.

[0044] Another aspect of the present invention relates to a deep learning-based intelligent sewing equipment time detection system, used to execute the above-described deep learning-based intelligent sewing equipment time detection method, comprising:

[0045] The multimodal dataset acquisition module is used to collect the start and stop signals of the sewing equipment motor and the change data of fabric tension through sensors, and simultaneously acquire the operation image sequence. After data preprocessing, the data is fused into a multimodal dataset.

[0046] The state sequence label acquisition module is used to input multimodal datasets into the target detection model, identify key states of sewing equipment operation and operator on duty, and obtain state sequence labels;

[0047] The interval classification result acquisition module is used to extract the thread-cutting action-related feature vectors from the state sequence labels, use deep learning algorithms to analyze the temporal correlation between the thread-cutting action-related feature vectors, determine the temporal correlation to mark the valid sewing intervals and invalid time-consuming intervals, and obtain the interval classification result.

[0048] The total working hours calculation module is used to obtain the interval classification results and calculate the duration of each valid sewing interval, accumulate and generate the single-process working hours value, dynamically correct the deviation of historical production data, and determine the total working hours calculation value.

[0049] The report output acquisition module is used to generate structured report data based on the total working hours calculation value. It integrates anomaly detection logic, which is used to judge working hour fluctuations and trigger alarm signals to obtain optimized report output.

[0050] The final time detection record acquisition module is used to extract key indicators from the optimized report output, connect to the manufacturing execution system interface to achieve real-time data synchronization and transmission, and update the production control database to obtain the final time detection record if the synchronization and transmission is completed.

[0051] The beneficial effects achieved by this invention are as follows:

[0052] This invention provides a deep learning-based intelligent detection method and system for sewing equipment working hours. It collects real-time data on motor start / stop signals and fabric tension changes from sensors, and simultaneously acquires a sequence of operational images. After preprocessing, these images are fused to construct a multimodal dataset. This multimodal dataset is input into a target detection model to accurately identify the equipment's operating status and the operator's key on-duty status, forming continuous state sequence labels. Furthermore, feature vectors related to thread-cutting actions are extracted from the state sequences, and deep learning algorithms are used to deeply mine their temporal correlation patterns, thereby intelligently judging and dividing effective sewing intervals and ineffective time-consuming intervals, achieving accurate segmentation of actual production time. Based on this, the duration of each effective sewing interval is accumulated to form a single-process working hour value, and dynamic deviation correction is performed using historical production data to finally obtain a reliable total working hour calculation value. Simultaneously, structured reports are generated, built-in abnormal fluctuation detection logic automatically triggers alarms, and key indicators are connected in real-time to the manufacturing execution system to synchronize data and update the production control database. The core technical effect of this invention is to solve the problems of low accuracy, poor efficiency, and inability to distinguish invalid time consumption caused by relying on manual experience to estimate working time or rough timing in traditional sewing processes. It realizes automated, accurate, and intelligent working time detection and production control based on multimodal data and time series in-depth analysis, which greatly improves the accuracy of working time calculation and the level of management refinement of garment manufacturing enterprises. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating an embodiment of the intelligent sewing equipment time detection method based on deep learning according to the present invention.

[0054] Figure 2 This is a functional block diagram of an embodiment of the intelligent sewing equipment time detection system based on deep learning of the present invention.

[0055] Explanation of icon numbers:

[0056] 10. Multimodal dataset acquisition module; 20. State sequence label acquisition module; 30. Interval classification result acquisition module; 40. Total working hours calculation value determination module; 50. Report output acquisition module; 60. Final working hours detection record acquisition module. Detailed Implementation

[0057] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0058] like Figure 1 As shown, the first embodiment of the present invention proposes a deep learning-based intelligent detection method for sewing equipment working hours, including the following steps:

[0059] Step S100: Collect the start / stop signals of the sewing equipment motor and the data on changes in fabric tension through sensors, simultaneously acquire the sequence of work images, and fuse them into a multimodal dataset after data preprocessing.

[0060] Core operational data—motor start / stop signals (reflecting whether the equipment is running) and fabric tension change data (characterizing the actual sewing process)—is collected by deploying multiple types of sensors (such as motor current sensors and tension sensors) on the sewing equipment. Simultaneously, a sequence of images of the operator's work is acquired through an industrial camera. Outlier removal and missing value filling are performed on the collected sensor time-series data, and noise reduction and frame alignment are performed on the image sequences. Finally, the heterogeneous sensor data and image data are fused into a multimodal dataset, which fully covers all dimensions of information on equipment operating status, work process, and personnel operation, providing a high-quality data foundation for subsequent status recognition and work time division.

[0061] Step S200: Input the multimodal dataset into the target detection model to identify the key states of sewing equipment operation and operator on duty, and obtain state sequence labels.

[0062] The multimodal dataset output from step S100 is input into a pre-trained target detection model (such as YOLO (YouOnly Live Once) or Faster R-CNN (Faster Region-based Convolutional Neural Network) and other deep learning models). The model performs joint recognition on the equipment operation features (motor start-stop signal timing) and image features (operator's limb movements and equipment needle bar movement status) within the dataset to accurately determine two key states: the running / stopping status of the sewing equipment and the on-duty / off-duty status of the operator. The recognition results are serialized according to the time dimension to generate a state sequence label containing timestamps and status identifiers, realizing the state visualization labeling of the entire sewing operation process, and providing a basis for subsequent determination of effective / ineffective working time intervals.

[0063] Step S300: Extract the thread-cutting action-related feature vectors from the state sequence labels, use a deep learning algorithm to analyze the temporal correlation between the thread-cutting action-related feature vectors, determine the temporal correlation to mark the effective sewing intervals and invalid time intervals, and obtain the interval classification results.

[0064] From the state sequence labels generated in step S200, extract the feature vector related to the iconic action of sewing operation—the thread-cutting action (such as the tension change feature at the moment of thread cutting, the limb feature of the operator cutting the thread in the image, and the instantaneous change feature of the motor current). Input this feature vector into a temporal deep learning algorithm (such as LSTM (Long Short-Term Memory) or TCN (Time Convolutional Network)) to analyze the temporal correlation before and after the thread-cutting action (such as effective sewing processing before thread cutting and ineffective time consumption such as material change / adjustment after thread cutting). Based on this, accurately mark the effective sewing interval (equipment operation + personnel on duty + actual fabric processing) and ineffective time consumption interval (equipment idling, personnel leaving their posts, material change and machine adjustment) in the sewing operation, and output clear interval classification results to solve the core pain point of "mixed calculation of effective / ineffective time consumption" in traditional time statistics.

[0065] Step S400: Obtain the interval classification results and calculate the duration of each valid sewing interval, accumulate them to generate the single-process working time value, dynamically correct the deviation of historical production data, and determine the total working time calculation value.

[0066] Based on the interval classification results of step S300, the duration of each valid sewing interval is counted and accumulated to obtain the initial working time value of a single sewing process. A deviation correction model is established by combining historical production data to analyze the deviation value between the current process working time and the historical average working time of similar processes. The influence of occasional abnormal factors (such as temporary equipment failures or single operation errors) is eliminated, and the initial working time value is dynamically corrected. Finally, the total working time calculation value that takes into account both accuracy and rationality is determined to ensure the objectivity and authenticity of the working time data and avoid working time deviations caused by single-dimensional statistics.

[0067] Step S500: Generate structured report data for the total working hours calculation value, integrate anomaly detection logic, the anomaly detection logic is used to judge working hour fluctuations and trigger alarm signals, and obtain optimized report output.

[0068] The total working hours calculated in step S400 are used to generate a structured report containing process name, equipment number, operator, effective working hours, invalid working hours, and total working hours according to a preset format. Anomaly detection logic is integrated into the report generation process. By setting a working hour fluctuation threshold (such as a single process working hours exceeding the historical average by ±20%), the system automatically judges whether there are abnormal fluctuations in the working hour data (such as an excessively high proportion of invalid working hours or a sudden drop in effective working hours). If an anomaly is triggered, a corresponding alarm signal is generated (such as a low equipment efficiency alarm or an abnormal personnel operation alarm). Finally, an optimized report output with integrated anomaly prompts is output, realizing the dual functions of visual presentation of working hour data and anomaly warning.

[0069] Step S600: Extract key indicators from the optimized report output, connect to the manufacturing execution system interface to achieve real-time data synchronization and transmission. If the synchronization and transmission is completed, update the production control database to obtain the final time detection record.

[0070] Extract core time indicators (such as average time per process, percentage of invalid time, and number of abnormal alarms) from the report output optimized in step S500. Connect to the factory's Manufacturing Execution System (MES) through a standardized interface to achieve real-time synchronous transmission of time data to the production control platform. After the system verifies and confirms that the data synchronization is complete, automatically update the latest time detection records to the production control database, completing the entire closed loop from "data acquisition - time calculation - report generation - system synchronization - production control database update", providing accurate time data support for factory production scheduling, efficiency analysis, and performance evaluation.

[0071] Furthermore, the deep learning-based intelligent sewing equipment time detection method provided in this embodiment includes step S100 as follows:

[0072] Step S110: Obtain the start / stop signal of the sewing equipment motor, fabric tension change data, and work image sequence, and lock the valid work window according to the start / stop signal of the motor.

[0073] In actual production, the motor start / stop signals of sewing equipment directly reflect whether the operator is performing sewing actions. For example, when the pedal is pressed, the motor starts, generating a high-level signal; when the pedal is released or the set number of stitches is reached, the motor stops, and the signal turns low. Specifically, by collecting a continuous sequence of motor start / stop signals, several effective working windows can be accurately identified. Each window corresponds to a complete sewing process, such as the complete cycle from the start of needle insertion to the lifting of the presser foot. This identification method avoids interference from non-sewing times such as idling, threading, and fabric changing, ensuring that subsequent analysis focuses on real operational data.

[0074] Step S120: Filter the fabric tension change data within the effective working window to obtain the tension trend sequence.

[0075] The tension trend value is obtained using the following formula:

[0076] (1)

[0077] In formula (1), Indicates the current time The smoothed tension trend value, at time... The final tension value after filtering represents the steady-state tension trend within the current sewing cycle. This represents the original tension sampling value within the effective working window, and the time within the effective working window. Raw data of fabric tension collected (unit: g). This represents the weighting coefficient centered on the current point, corresponding to the weight value of each sampling point within the window. Typically, the center has the highest weight and the edge has the lowest weight (such as Gaussian kernel and Hanning window). Indicates the length of the filter window used, an odd integer, defining the total number of sampling points involved in the filter calculation (e.g., ...). =5 indicates that the current time and two sampling points before and after are covered. The control logic of formula (1) is a weighted smoothing filtering algorithm based on the time window. Its core function is to filter out high-frequency noise (such as instantaneous mechanical vibration) from the original tension sampling data within the effective working window, and extract the tension trend sequence that reflects the steady-state change of the sewing process, so as to provide stable and reliable input features for intelligent detection of working time.

[0078] The fabric tension variation data within each effective work window is first processed by median filtering, and then combined with low-pass filtering to obtain a relatively smooth tension trend sequence. For example, if the original tension sampling frequency is 100Hz and the value fluctuates between 0 and 500g, the filtered tension trend sequence can clearly reflect the slow rising or falling trend caused by changes in stitches, fabric thickness, or uneven fabric feeding during the sewing process, while eliminating instantaneous jitter noise. This smoothing process makes subsequent abrupt change detection more reliable.

[0079] Step S130: Extract the sewing area image frame set from the work image sequence based on the abrupt change time of the tension trend sequence.

[0080] The set of image frames for the sewing area is obtained using the following formula:

[0081] (2)

[0082] In formula (2), This represents the set of image frames for the sewing area, which is the final set of key image frames extracted, containing all valid frames related to tension changes. Indicates time The corresponding image frame, the input original job image sequence, For image frame index. Indicates the current time The smoothed tension trend value is the smoothed value after low-pass / Gaussian filtering of the original tension signal, used to eliminate noise and accurately locate abrupt change points. This represents the frame radius extending from the point of change, or the frame threshold extending forward and backward from the moment of the tension change. For example... Then select 11 frames, from the 5 frames before to the 5 frames after the mutation point. This represents the total number of mutation points, the number of valid mutation points detected in the tension trend sequence (such as sewing start, stop, thread break, and other trigger points). The control logic of formula (2) is based on the mutation point location of the tension trend, performing neighborhood frame aggregation, and filtering out the set of key frames containing sewing actions / regions from massive operation images for subsequent region analysis, process recognition, or visual inspection.

[0083] Extracting image frames of the sewing area based on abrupt changes in the tension trend sequence is a crucial step. Preferably, a typical abrupt change is defined as a sudden increase or decrease in tension exceeding a preset threshold within a short period. For example, when sewing denim, the tension at the seam may rapidly rise from 180g to 420g and then drop back within 0.4 seconds. The corresponding image frame, along with the three frames before and after it, can form a five-frame image set, fully recording the dynamic process of the needle passing through the thick seam. Multiple consecutive abrupt changes can yield multiple sets of image frames, covering different sewing features.

[0084] Step S140: Concatenate the image frame set of the sewing area with the numerical feature vectors corresponding to the tension trend sequence to generate a multimodal dataset.

[0085] The multimodal dataset is generated using the following formula:

[0086] (3)

[0087] In formula (3), This represents a multimodal dataset, a unified set of features after concatenation, including visual features and tension temporal features, used for subsequent model training and analysis. It represents a set of image frames of the sewing area, and is a high-dimensional visual feature vector (such as features extracted by CNN, object detection features, etc.) extracted from continuous video frames of the sewing process, representing visual information such as fabric, stitches, and equipment movements. The numerical feature vector corresponding to the tension trend sequence is the time-series numerical feature vector extracted from the tension trend sequence (such as the tension data after filtering in step S120), such as tension peak value, mean, fluctuation variance, trend slope, etc., which characterizes the tension change law in the sewing process. This indicates a splicing operation, which directly splices two feature vectors along a fixed dimension (usually the feature dimension) without changing the semantics of the features themselves, only achieving feature dimension fusion. The control logic of formula (3) is a multimodal feature fusion model. The core is to splice the visual features (sewing area image) and the time-series numerical features (tension trend sequence) along the dimensions to generate a unified multimodal feature vector, providing complete feature input for subsequent tasks such as time detection and process recognition.

[0088] Numerical feature vectors corresponding to the tension trend sequence This can be derived from the following formula:

[0089] (4)

[0090] In formula (4), This represents the feature extraction function, defining the logic and rules for feature extraction, such as statistical calculations, Fourier transforms, and time window statistics. This represents the tension trend sequence, which is the tension time series data after filtering and smoothing (such as the tension value sequence sampled at 100Hz). The control logic of formula (4) is the time series feature extraction function, the core of which is to extract the one-dimensional tension trend sequence. Mapped to high-dimensional numerical eigenvectors By extracting the statistical regularity and temporal characteristics of the sequence, a quantitative basis is provided for subsequent time detection and process identification.

[0091] By concatenating a set of sewing area image frames with corresponding tension trend sequence numerical feature vectors to form a multimodal dataset, subsequent intelligent analysis capabilities can be significantly improved. For example, a dataset may contain 5 images and a 15-dimensional tension feature vector strictly time-aligned with them, including the average value before the mutation, the mutation magnitude, the rise slope, and the fall time. This image-data fusion approach allows the model to simultaneously perceive visual changes in stitch density, fabric wrinkles, and skipped stitch defects, as well as numerical changes in tension stability. In one embodiment, this multimodal dataset can be used to train a defect classification model, reducing the false negative rate by approximately 18% and the false positive rate by approximately 11% compared to an image-only approach. In another embodiment, it can be used to predict fabric feeding mechanism anomalies, extending the early warning time to 2-4 sewing cycles. Through precise segmentation of the work window using motor signals, extraction of reliable trends via tension filtering, selection of keyframes driven by mutations, and multimodal representation through image-data concatenation, the entire process forms a closed loop, resulting in higher accuracy and robustness for anomaly detection, quality traceability, and process optimization in the sewing process.

[0092] Preferably, the intelligent sewing equipment time detection method based on deep learning provided in this embodiment includes step S200 as follows:

[0093] Step S210: Obtain the multimodal dataset, extract and fuse image spatial texture features and tension temporal fluctuation features to obtain the multimodal fusion feature tensor.

[0094] The multimodal fusion feature tensor is derived using the following formula:

[0095] (5)

[0096] In formula (5), This represents the multimodal fusion feature tensor, the final feature representation after fusing texture and tension features, with unified dimensions, used for subsequent tasks. This indicates a concatenation operation along the feature dimension, which concatenates two feature vectors along the feature dimension to generate a high-dimensional fused feature vector. This represents a spatial texture fusion feature vector, which is a texture feature extracted from image / visual data, representing spatial information such as fabric appearance and stitch shape. This represents the time-series fluctuation feature vector of tension, which is a time-series feature extracted from tension sensor data, characterizing the temporal information such as tension change trend and fluctuation amplitude. and These are linear projection matrices representing image texture features and tension temporal features, respectively. Linear transformations are applied to the texture features and tension features to unify the feature dimensions and map them to the same feature space, thus eliminating modal differences. The bias vector is represented by the additive bias, which is used to adjust the offset after feature mapping and improve the model fitting ability. The control logic of formula (5) is a multimodal feature fusion and linear projection model. Through feature splicing and linear projection, the deep fusion of image spatial texture features and tension temporal fluctuation features is achieved, generating a unified multimodal fusion feature tensor, which provides a unified feature input for subsequent time detection and process analysis.

[0097] Spatial texture fusion feature vector This can be derived from the following formula:

[0098] (6)

[0099] In formula (6), This represents the number of image channels, specifically the channel dimension of the input image (e.g., grayscale). RGB image ). and These represent the height and width dimensions of the input feature map, respectively. Indicates the first The passage is in the location The gradient magnitude at a given location represents the texture intensity at that location; a larger gradient indicates a more drastic texture change. This indicates that the local texture gradient statistics are used for weighting, and are statistical features calculated based on the local window to measure the importance of texture features at that location. The local contribution weights are normalized by converting local statistics into weights between 0 and 1. After normalization, the sum of the weights is 1, highlighting the contribution of key regions. The control logic of formula (6) is a multi-channel gradient feature aggregation model, which realizes the weighted fusion of multi-channel image gradient features. The core is to characterize texture features through gradient magnitude and use local statistics and softmax normalization to highlight the feature contribution of important texture regions and generate robust spatial texture feature vectors.

[0100] Tension time-series fluctuation eigenvector This can be derived from the following formula:

[0101] (7)

[0102] In formula (7), This represents the characteristic vector of tension temporal fluctuation, which is the characteristic vector obtained after fusion and aggregation, characterizing the tension fluctuation pattern of the entire segment. The number of sampling points for a single segment of tension data indicates the length of the time series and determines the duration of the series. Indicates the first The change in the tension vector at time t, at time t and The difference in tension values ​​characterizes the magnitude of instantaneous tension change. Indicates the first The mask for each time point, a binary / weighted matrix, is used to filter valid sampling points (e.g., excluding idle and standby periods), summing only over the area covered by the mask. This indicates element-wise multiplication, bit-weighted addition, masking invalid regions and highlighting valid signals. To control the attenuation intensity as a fluctuation suppression hyperparameter: The larger the value, the stronger the suppression of large fluctuations (abnormal peaks). This term is used to nonlinearly attenuate violent fluctuations, nonlinearly suppress large fluctuations, and weaken the interference of outliers on the overall characteristics. The L2 norm squared represents the change in the tension vector, measuring the time interval. and The larger the value, the more intense the fluctuation. The control logic of formula (7) is a time series weighted aggregation model, which is used to extract fluctuation features from the tension sequence, suppress abnormal fluctuations through masking and exponential decay, and generate a robust tension time series fluctuation feature vector to provide stable features for subsequent time detection and job identification.

[0103] After acquiring the multimodal dataset, spatial texture features are first extracted from each frame of the sewing area image. The local regularity and global uniformity of the stitch arrangement on the fabric surface are captured using convolution, while the statistical fluctuation characteristics of the tension time series within the corresponding time window are calculated, such as standard deviation, peak value rate of change, and mean shift. These two types of features are then concatenated and fused along the channel dimension to form a multimodal fusion feature tensor, which simultaneously contains visual texture information and mechanical fluctuation information.

[0104] Step S220: Input the multimodal fusion feature tensor into the region proposal network to generate a set of detection boxes, which contains target category and location information.

[0105] The set of detection boxes is generated using the following formula:

[0106] (8)

[0107] In formula (8), This represents a set of detection boxes, containing structured data of several detection boxes. Each box is labeled with the target category (such as normal stitches, skipped stitches, wrinkles, and abnormal seams) and precise location coordinates (x1, y1, x2, y2), along with a confidence score. The Region Proposal Network (RPN) is a deep learning-based object detection candidate box generation model responsible for extracting potential target regions from feature maps and outputting classification and localization results. The multimodal fusion feature tensor is a high-dimensional feature vector that integrates visual features of the image (such as fabric texture, wrinkles, and stitch patterns) with temporal features of tension (such as tension fluctuation amplitude, standard deviation, and rate of change). It is the core basis for detection. The parameters represent all learnable parameters of the Region Proposal Network, including convolutional kernel weights, fully connected layer parameters, classifier weights, regressor weights, etc., which are the core parameters for the model to learn the relationship between features and detection targets. The control logic of formula (8) is driven by multimodal fusion features, and generates a set of detection boxes with class confidence and precise coordinates through the Region Proposal Network (RPN). The core logic is an end-to-end detection process of feature extraction → candidate box generation → classification and localization.

[0108] The multimodal fusion feature tensor is input into the region proposal network. The network outputs a set of detection boxes, each labeled with the target category such as normal stitches, skipped stitches, wrinkle buildup, or seam anomalies, along with precise location coordinates. For example, when the image shows obvious wrinkles in the fabric and the corresponding tension window standard deviation exceeds 1.8 times the normal value, the network tends to generate wrinkle-type detection boxes with higher confidence.

[0109] Step S230: Calculate the Euclidean distance between the limb and the head of the machine based on the detection box set, and determine the single-frame state classification code by combining the tension fluctuation range.

[0110] The Euclidean distance between the limbs and the nose is calculated using the following formula:

[0111] (9)

[0112] In formula (9), This represents the Euclidean distance between a limb and the machine head, characterizing the spatial relative position of an operator's limb (such as a hand) and the machine head, with the unit being pixels per centimeter. This represents the x-coordinate of the center of the limb detection box. This represents the ordinate of the center of the limb detection box, which is the center coordinate of the key points (hand / arm) of the limb extracted from the detection box set. This indicates the x-coordinate of the center of the head detection frame. The vertical coordinate of the center of the head detection frame is the coordinate of the center of the head reference point extracted from the detection frame set. The control logic of formula (9) is to accurately calculate the relative position distance between the operator's limbs (such as hands) and the head (equipment operation point) in the image space through the two-dimensional Euclidean distance formula. The core logic is "the square root of the sum of squares of coordinate differences to quantify the physical distance".

[0113] The state classification code for a single frame is determined using the following formula:

[0114] (10)

[0115] In formula (10), Represents single-frame state classification coding, using a binary identifier: Indicates a valid work status. This indicates an invalid / interrupted job status. It represents the Euclidean distance between a limb and the machine head, quantifying the spatial relative position of an operator's limb (such as the hand) and the operating point of the equipment (such as the machine head). This represents the distance threshold, a preset critical distance used to determine whether the operator is within the valid operating range. It indicates the amplitude of tension fluctuation, quantifies the degree of fluctuation in the equipment tension signal, and reflects the stability of the operating conditions. This represents the lower limit of the fluctuation range, the minimum threshold for tension fluctuation in effective operation. This represents the upper limit of the fluctuation range, the highest threshold of tension fluctuation for effective operation. The control logic of formula (10) is to generate the operation status code of a single frame by linking the spatial position and tension condition. The core logic is "position compliance + normal operation condition → active state (1), otherwise → inactive state (0)".

[0116] Tension fluctuation amplitude This can be derived from the following formula:

[0117] (11)

[0118] In formula (11), This indicates the amplitude of tension fluctuation, and the unit is consistent with the original tension data (such as g or N). The larger the value, the more intense the tension fluctuation. This represents the maximum value of the tension sequence, the peak data collected by the tension sensor within the effective working window. This represents the minimum value of the tension sequence, and the valley value data collected by the tension sensor within the effective working window. The control logic of formula (11) is to quantify the fluctuation range of the tension sequence by the difference between the maximum and minimum values ​​(i.e., the range). The core logic is to "use the extreme value span of the full data to characterize the severity of tension changes".

[0119] The Euclidean distance between the operator's key limb points and the machine head reference point is calculated based on the detection frame set. Preferably, a representative point is selected from the hand or arm. When the distance is less than a preset safety threshold, such as 18 cm, it is initially determined to be close to the operating area. Further confirmation is made based on the tension fluctuation range of the current frame. For example, if the tension is in a stable low value range and the distance is close, it is coded as "operator on duty and equipment operating normally". If the tension fluctuates abnormally high but the distance is still close, it is coded as "on duty but with potential abnormality". If the distance is greater than the threshold and the tension is stable, it is coded as "off duty".

[0120] Step S240: Perform context constraint modeling on the original state sequence formed by single-frame state classification coding, correct the original state sequence, and obtain the state sequence labels of sewing equipment operation and operator on duty.

[0121] The status sequence labels for sewing equipment operation and operator on-duty status are derived using the following formula:

[0122] (12)

[0123] In formula (12), The label represents the status sequence, which is the corrected sequence of sewing equipment operation and personnel on duty status, such as {0, 1, 0, 1} (0 = off duty / stopped, 1 = on duty / operating). This represents the modeling function, the core function for fusing temporal constraints and multimodal features, and includes logic such as continuity judgment and outlier correction. Indicates the operating status of sewing equipment, a sequence of equipment statuses, such as... , Indicates that the program is running. This indicates that the machine is out of service. This indicates the operator's on-duty status, a sequence of personnel statuses, such as... , Indicates being on duty. This indicates that the person is off duty. The control logic of formula (12) is to correct the original state sequence by modeling through context constraints. The core logic is to integrate the temporal correlation between equipment status and personnel status, eliminate isolated false detections, and output continuous and reliable on-duty / work status labels.

[0124] Contextual constraint modeling of the original state sequence formed by single-frame state classification encoding is a key correction step. A continuity constraint between frames is adopted: if a frame is encoded as "off-duty," but the two frames before and after it are both "on-duty and running," and the tension is not significantly interrupted, then the isolated frame is corrected to "on-duty." For example, during a continuous sewing process, a brief lifting of the hand to adjust the fabric may cause the distance to momentarily exceed the threshold, but the tension trend remains consistent. In this case, context smoothing can effectively avoid misjudging off-duty status. Through the above-mentioned multimodal feature fusion, detection box localization, single-frame encoding combining distance and tension intervals, and sequence context correction, the entire process forms a continuous and reliable label for the sewing equipment's operating status and the operator's on-duty status. This state sequence label can directly support subsequent on-duty time statistics, abnormal workstation location, and equipment idle time rejection analysis, thereby improving production data quality and personnel management accuracy. For example, in the actual scenario of sewing the back pocket of jeans, operators need to frequently adjust the position of the fabric, and the original code may have several short-term mislabelings due to absence from duty. After context constraints, the state sequence is smoother and more continuous, accurately reflecting the actual on-duty time, which helps to accurately calculate the processing cycle of a single piece and optimize the workstation layout.

[0125] Furthermore, the deep learning-based intelligent sewing equipment time detection method provided in this embodiment includes step S300 as follows:

[0126] Step S310: Obtain the time point of the line-cutting action in the state sequence label, and use a long short-term memory network to map the multimodal fusion feature tensor fragment into a line-cutting action-related feature vector.

[0127] The feature vector related to the wire-cutting action is obtained using the following formula:

[0128] (13)

[0129] In formula (13), It represents the feature vector related to the line-cutting action, a fixed-dimensional feature vector, which condenses the key temporal patterns before and after the line-cutting action. It represents a long short-term memory network, a core network for time-series modeling, and captures long dependencies and short-term fluctuations. The multimodal fusion feature tensor segment represents a feature sequence centered on the wire-cutting moment, with a fixed duration (2-4 seconds) before and after it, containing multimodal information such as tension fluctuations and visual texture. The control logic of formula (13) realizes multimodal temporal feature modeling. The core logic is to use a long short-term memory network (LSTM) to encode the temporal feature segment containing tension changes and visual texture, and output a high-dimensional feature vector focusing on the wire-cutting action, providing a representation for subsequent action recognition and time analysis.

[0130] After acquiring the state sequence labels, the time points of the thread-cutting action can be extracted. These time points typically correspond to a brief sharp drop in the tension signal accompanied by subtle changes in visual features near the machine head, such as thread residue or a brief pause in the needle bar. For a multimodal fusion feature tensor segment before and after each thread-cutting action, a fixed-length time segment is formed by extending a window of approximately 2 to 4 seconds forward and backward, centered on the thread-cutting moment.

[0131] The time sequence segment is input into a Long Short-Term Memory (LSTM) network. The network processes and fuses features frame by frame, gradually accumulating memory information on visual texture changes and tension fluctuations, and finally outputs a fixed-dimensional feature vector related to the cutting action at the end of the segment. This feature vector encapsulates key patterns such as the smoothness of fabric advance before cutting, the magnitude of the tension change at the moment of cutting, and the continuity of stitch recovery after cutting.

[0132] Step S320: Input the feature vectors related to the line-cutting action into the self-attention mechanism model to construct the temporal correlation matrix.

[0133] The temporal correlation matrix is ​​constructed using the following formula:

[0134] (14)

[0135] In formula (14), This represents a temporal correlation matrix, a two-dimensional square matrix with dimensions T×T (T being the number of time steps in the feature sequence). The matrix element values ​​reflect the similarity of feature vectors at two different time points; larger values ​​indicate stronger temporal correlation. The feature vector matrix represents all time steps of the line-cutting action. Each row / column corresponds to a feature vector of a time step, with a dimension of D×T (D is the number of feature dimensions). This represents the transpose of the eigenvector matrix, achieving the transpose of the matrix dimensions and satisfying the dimension matching rules for inner product calculation. represents the feature dimension, the number of dimensions of the feature vector at a single time step, and a normalization factor used for similarity calculation. This formula directly generates a similarity matrix between time steps through a self-attention inner product mechanism, used to capture the temporal dependencies of the line-cutting action. The control logic of formula (14) is to calculate the similarity of each time step in the feature sequence related to the line-cutting action through a self-attention inner product mechanism, generating a temporal correlation matrix. Its core function is to quantify the dependence strength of action features in the time dimension, providing a correlation structure support for subsequent temporal pattern analysis.

[0136] The feature vectors related to the wire-cutting action are input into the self-attention mechanism model. The self-attention mechanism model calculates the correlation weights between each time step within the segment, forming a temporal correlation matrix. The numerical values ​​of the elements in the temporal correlation matrix reflect the strength of the inherent logical dependence between features at different times. For example, if the tension gradually increases before wire cutting, drops sharply at the moment of wire cutting, and quickly returns to stability after wire cutting, the correlation weights between the corresponding time steps are higher, and the matrix element values ​​are larger. Preferably, the correlation threshold is set to 0.65.

[0137] Step S330: If the value of an element in the time sequence correlation matrix is ​​greater than a preset threshold, it is determined that there is continuous operation logic; otherwise, it is determined that there is an abnormal pause.

[0138] The following formula is used to define the condition for the existence of continuous operation logic:

[0139] (15)

[0140] In formula (15), Represents the temporal correlation matrix from time [time]. At the time The element values ​​are in the range [0, 1]. This represents a preset threshold, the critical value for determining continuous operation. In this scenario, it is taken as... =0.65. When Greater than the preset threshold The control logic of formula (15) determines whether there is continuous operation logic between any two time steps by comparing the element values ​​of the time sequence correlation matrix with the threshold. The core is used to judge the continuity and correlation of the process / action.

[0141] If most of the high-weight elements in the temporal correlation matrix exceed the threshold, it is determined that there is continuous operation logic before and after the current thread-cutting action, indicating that the operator has completed the normal thread-cutting-continue sewing process; conversely, if there is a significant low-value area in the temporal correlation matrix, especially if the correlation weight between the thread-cutting action and the preceding action is consistently lower than the threshold for several time steps, it is determined that there is an abnormal pause, caused by the operator leaving midway, fabric jamming, or minor equipment malfunction.

[0142] Step S340: Mark the valid sewing intervals according to the continuous operation logic, and mark the invalid time intervals according to the abnormal pauses. Summarize the valid sewing intervals and invalid time intervals to obtain the interval classification results.

[0143] The interval classification results are obtained using the following formula:

[0144] (16)

[0145] In formula (16), This represents the final set of interval classification results. The output set contains all job intervals that are classified as valid or invalid. It represents the sum of all valid sewing intervals marked by the continuous operation logic, and the set of valid sewing intervals marked by the continuous operation logic without abnormal pauses. This indicates a difference operation that removes invalid time intervals, ensuring that the set of valid sewing intervals does not contain invalid data. This represents a union operation, which is a set of invalid time intervals marked with abnormal pauses and including equipment downtime / idling / failure. This represents the set of all invalid time intervals obtained based on the abnormal pause markers. The valid sewing intervals and invalid time intervals are merged to ensure coverage of all time steps. This formula achieves a complete division of valid and invalid time intervals through difference and union operations. The control logic of formula (16) is to achieve full coverage and mutual exclusion of valid sewing intervals and invalid time intervals through set operations. The core logic is to "first remove invalid time intervals, then merge valid sewing intervals, and finally achieve binary classification of work intervals".

[0146] The effective sewing range is calculated using the following formula:

[0147] (17)

[0148] In formula (17), Indicates the first A set of valid sewing intervals, Indicates belonging to the first A set of sewing action indexes for a continuous operation, iterating through the indexes of all sewing actions. and They represent the first The start and end times of each sewing action. This represents the effective time function under continuous operation logic, the core decision function, and the input time. Output 1 indicates continuous operation, and output 0 indicates abnormal pause. The control logic of formula (17) is based on continuous operation logic, which filters out effective sewing intervals without abnormal pauses, merges continuous effective operation segments, and forms the final set of effective sewing intervals.

[0149] The invalid time interval is obtained by the following formula:

[0150] (18)

[0151] In formula (18), Indicates the first The output is a set of invalid time intervals, including the start point, end point and duration of each interval. and They represent the first The start and end times of the abnormal pause, the input sequence, and the time range representing the pause. Indicates the first Calculate the interval span based on the duration of the abnormal pause. This represents the preset minimum abnormal pause threshold, a manually set critical value (e.g., 5 seconds), which only occurs when... Only then was it judged as "invalid time consumption". express The machine status at any given time: 0 = running / job, 1 = paused / non-job. Indicates the first The duration of the abnormal pause. Indicates time Machine status, The logic and operation indicate that both "duration exceeds the threshold" and "pause state" must be met simultaneously for it to be considered invalid time consumption. The control logic of formula (18) is used to identify and extract invalid time consumption intervals / abnormal pause intervals in the work sequence. The core logic is to filter out invalid time consumption segments that meet the preset standards through the dual conditions of "duration threshold + state trigger" for subsequent time analysis.

[0152] The effective sewing interval is marked according to the continuous operation logic. This effective sewing interval starts from the stable sewing start point after the previous thread trimming and continues until the normal progress ends before the current thread trimming, representing the efficient actual processing period. Abnormal pauses are marked as invalid time intervals. For example, if the tension remains close to zero for a long time after a thread trimming and there is no visible fabric movement, and the timing relationship matrix shows a break from the previous thread trimming logic, this segment is marked as an invalid time interval, usually corresponding to idling, waiting, or abnormal interruption.

[0153] By summarizing the effective sewing intervals and the ineffective time intervals, a complete interval classification result can be obtained. This interval classification result directly reflects the actual operational composition ratio of each workstation in a complete sewing task. For example, in the continuous process of sewing the back pocket of a pair of jeans, if the original state sequence shows a total duration of 8 minutes, of which the effective sewing interval is determined to be 5 minutes and 12 seconds and the ineffective time interval is 2 minutes and 48 seconds, then the distribution of efficient sewing and non-value-adding time can be clearly distinguished. This facilitates subsequent analysis of operator cycle stability, identification of specific workstations with frequent abnormal stops, and optimization of training guidance for re-alignment after thread cutting, thereby significantly improving overall production efficiency and data analysis accuracy.

[0154] Preferably, the intelligent sewing equipment time detection method based on deep learning provided in this embodiment includes step S400 as follows:

[0155] Step S410: Traverse the start and end timestamps of the valid sewing intervals in the interval classification results, and calculate the difference between the start and end timestamps of the valid sewing intervals to obtain the effective sewing duration.

[0156] The effective sewing duration is calculated using the following formula:

[0157] (19)

[0158] In formula (19), Indicates the first The duration of an effective sewing interval, expressed in seconds per millisecond, represents the total time spent on effective sewing operations within that interval. Indicates the first The end timestamp of each valid sewing interval, the time point at which the valid sewing operation ends. Indicates the first The start timestamp of each effective sewing interval is the time point at which the effective sewing operation begins. The control logic of formula (19) is used to calculate the time length of a single effective sewing interval. By directly comparing the end timestamp with the start timestamp, the duration of the effective sewing operation is quantified, providing basic data for subsequent time utilization and efficiency analysis.

[0159] After obtaining the interval classification results, the start and end timestamps of all valid sewing intervals are iterated. For each valid sewing interval, the start timestamp is subtracted from the end timestamp to obtain the actual duration of that segment, which is the operation time for a single valid sewing. It should be noted that this duration excludes possible abnormal pauses or invalid waits after thread cutting, and only retains the continuous processing portion where the operator actually drives the sewing machine forward.

[0160] Step S420: Add up the duration of multiple valid sewing operations within the same production batch to generate a single-process time value.

[0161] The time value for a single operation is calculated using the following formula:

[0162] (20)

[0163] In formula (20), Indicates the first The single-process time value for a production batch, in seconds per minute, represents the total operation time for completing a single process in that batch. Indicates the first The duration of each effective sewing interval and the effective working time of a single sewing interval are calculated by formula (19). Indicates the first The total number of valid sewing records within a production batch. The control logic of formula (20) is used to calculate the total working time of a single process in batches. By statistically analyzing the total duration of all valid sewing intervals within the same production batch, the cumulative operation time required to complete the process of that batch is quantified.

[0164] The system accumulates the duration of multiple valid sewing operations within the same production batch. For example, during the sewing of the back pocket of a pair of jeans, the system identifies four valid sewing intervals with durations of 78 seconds, 92 seconds, 65 seconds, and 105 seconds, respectively. These are added together to obtain a total of 340 seconds, which is the single-operation time value for this process within the batch. This value reflects the actual processing time after removing non-value-added time.

[0165] Step S430: Calculate the difference between the single-process time value and the historical benchmark mean to obtain the time deviation value, wherein the historical benchmark mean is calculated based on historical production data that matches the current process code.

[0166] The time deviation value is obtained by the following formula:

[0167] (twenty one)

[0168] In formula (21), This indicates the deviation value of working hours, in time units (such as seconds or minutes); a positive value indicates that the current working hours are higher than the historical average (inefficiency is low), and a negative value indicates that the current working hours are lower than the historical average (inefficiency is high). This indicates the current working time value for a single process, the actual operation / processing time in the current period, such as the actual time consumed by a certain batch of processes. The arithmetic mean of all historical working hours that match the current process code is the historical baseline mean. The arithmetic mean of historical working hours that match the current process code, process, equipment and other conditions is derived from historical production data. The control logic of formula (21) is used to quantify the difference between the current process working hours and the historical baseline level. The difference between "current working hours - historical mean" intuitively reflects the degree of deviation in work efficiency, providing data support for efficiency analysis and abnormal early warning.

[0169] The system further compares the single-process time value with the historical baseline average. The historical baseline average is derived from a large number of past production records that perfectly match the current process code, and is obtained through statistical averaging. For example, for the back pocket process, the historical baseline average is 360 seconds. If the single-process time value calculated this time is 340 seconds, the time deviation value is -20 seconds, indicating that the efficiency of this operation is slightly higher than the historical average.

[0170] Step S440: Determine the weighted correction coefficient based on the time deviation value through a nonlinear mapping function, and use the weighted correction coefficient to correct the time value of a single process to determine the total time calculation value.

[0171] The weighted correction factor is obtained by the following formula:

[0172] (twenty two)

[0173] In formula (22), This represents the weighted correction coefficient, with a value range of [0, 1]. The larger the weighted correction coefficient, the stronger the correction to the deviation. The time deviation value is the core input and is calculated by formula (21). This represents the baseline value for time deviation, a threshold baseline used to define the "normal / abnormal" boundary of the deviation, such as setting it to 10% of the standard working hours of the process. This parameter represents the control parameter for the nonlinear steepness, which controls the "steepness" of the Sigmoid function. The larger the value, the faster the coefficient rises when the deviation exceeds the threshold. This represents the exponential decay term. The formula uses the sigmoid function to achieve a smooth, nonlinear mapping of the correction coefficient from near 0 to near 1 as the deviation increases. The control logic of formula (22) is used to map the time deviation value to a nonlinear correction coefficient between 0 and 1. The sigmoid function is used to achieve a smooth transition between the deviation and the correction intensity. The core logic is that the smaller the deviation, the closer the correction coefficient is to 0; the larger the deviation, the closer the correction coefficient is to 1, thus achieving adaptive weighted correction of the time deviation.

[0174] The total working hours are calculated using the following formula:

[0175] (twenty three)

[0176] In formula (23), This represents the total working hours calculated, which is ultimately used for cost accounting, process time statistics, or efficiency evaluation. This indicates the current working time value for a single operation, representing the actual time consumed by the operation, such as the time required to process a single operation. This represents the weighted correction coefficient, a comprehensive correction factor, with a value range of [0, +∞), used to adjust the total working hours. The formula obtains the total working hours calculation value by multiplying the working hours of a single process by a comprehensive weighted correction factor. The control logic of formula (23) is based on the working hours of a single process and the weighted correction coefficient, calculating the total working hours calculation value, and then adjusting it using the correction factor. The original working hours are adjusted non-linearly to obtain the final total working hours. .

[0177] Determine the weighted correction coefficient through a non - linear mapping function based on the working - hour deviation value. This non - linear mapping function is designed to have a mild correction amplitude when the deviation is small, and gradually amplify the adjustment strength when the deviation is large. For example, when the working - hour deviation is within ±15 seconds, the correction coefficient is close to 1.00; when the working - hour deviation reaches - 30 seconds, the correction coefficient rises to about 1.08; when the working - hour deviation exceeds - 60 seconds, the weighted correction coefficient further increases to reflect a greater recognition of super - efficient performance. Conversely, when there is a positive deviation, the correction coefficient is less than 1, which is used to appropriately reduce the overly long working hours.

[0178] Use the obtained weighted correction coefficient to finally correct the single - process working - hour value. For example, after multiplying the aforementioned 340 seconds by a correction coefficient of 1.05, the total working - hour calculation value of 357 seconds is obtained. This value not only retains the actual efficient performance but also maintains a reasonable connection with the historical benchmark system, facilitating standardized comparison across batches and personnel. Through the above - mentioned working - hour deviation analysis and non - linear correction mechanism, the system can more accurately calculate the true and effective working hours for each batch and each process, avoiding evaluation distortion caused by occasional anomalies or differences in operation proficiency. For example, when a certain operator continuously has short pauses marked as invalid time - consuming intervals, the single - process working - hour value is low, but after being adjusted upward by the correction coefficient, the total working - hour calculation value can still fairly reflect its core processing ability, thus supporting more scientific cycle - time assessment, performance appraisal, and subsequent targeted training guidance, ultimately improving the overall production efficiency and resource utilization rate of the workshop.

[0179] Furthermore, for the intelligent working - hour detection method of sewing equipment based on deep learning provided in this embodiment, step S500 includes:

[0180] Step S510: Obtain the total working - hour calculation value, perform serialization processing in combination with the associated production batch number and process code, and generate an initial structured report dataset.

[0181] Generate the initial structured report dataset through the following formula:

[0182] (24)

[0183] In formula (24), represents the initial structured report dataset, the output structured data set, which contains structured records in the dimension of [batch ID, process code, total working hours]. represents the set of production batch numbers the th batch in represents the set of process codes the th process in represents a specific batch in process the The value recorded for each work hour. This indicates an aggregation and summation operation, which accumulates all time records under the same batch and the same process. The control logic of formula (24) is used to aggregate time records of the same batch and the same process to generate a standardized, serializable structured time dataset, providing basic data input for subsequent time analysis, process optimization and cost accounting.

[0184] After obtaining the total working hours calculation value, the system associates it with the corresponding production batch number and process code to form a preliminary serialized structure. This serialization process organizes the scattered calculation results according to the logical order of batches and processes, facilitating subsequent batch analysis and traceability. For example, for the Houdou process in batch number P20260115-001, the system generates a structured record entry containing the batch number, process code, and a total working hours calculation value of 357 seconds.

[0185] Step S520: Parse the process codes contained in the initial structured report dataset and retrieve the time distribution records within the historical production cycle to determine the time fluctuation threshold range.

[0186] The threshold range for working hours fluctuation is determined using the following formula:

[0187] (25)

[0188] In formula (25), Indicates the threshold range for working hours fluctuation, tuple This indicates the upper and lower bounds of normal working hours fluctuation. It represents the historical working hours distribution mean, the arithmetic mean of historical working hours records, and characterizes the baseline working hours level. It represents the standard deviation of the historical working hours distribution, the degree of dispersion of historical working hours records, and characterizes the fluctuation range. Represents the volatility coefficient, with custom parameters (such as...) =1, 1.5, 2), controlling the sensitivity to abnormal fluctuations. The control logic of formula (25) is used to quantify the reasonable range of process time fluctuations based on statistical moments and fluctuation coefficients, and output upper and lower thresholds for judging whether the process time is abnormal, providing judgment criteria for quality monitoring and efficiency early warning.

[0189] When parsing the initial structured report data, the process code field is extracted first. For this code, the system retrieves a large number of time distribution records for similar processes from historical production cycles. These time distribution records typically cover real production data from the past three months or longer, and after statistical processing, they form relatively stable distribution characteristics, thereby determining a reasonable threshold range for time fluctuations. In one implementation, for subsequent processes, the historical time distribution shows a normal range concentrated between 320 and 400 seconds; therefore, the threshold range is set to the mean ± two standard deviations, i.e., 320 to 400 seconds.

[0190] Step S530: Compare the total working hours with the working hours fluctuation threshold range to calculate the working hours fluctuation amplitude.

[0191] The fluctuation range of working hours is calculated using the following formula:

[0192] (26)

[0193] In formula (26), This represents the fluctuation range of working hours. It is a non-negative value that indicates the magnitude of the fluctuation of total working hours relative to the baseline mean. The larger the value, the more significant the fluctuation / deviation. This represents the total working hours calculated, which is the actual total working hours for a single batch / single process (obtained by aggregation from steps S410-S420). This indicates the range of working hours fluctuation thresholds. The threshold mean is the benchmark mean of the historical working hours distribution (such as the standard working hours mean of the same period and the same process in history), which is used to compare the deviation. The control logic of formula (26) is used to quantify the deviation of the actual total working hours from the historical benchmark mean. The absolute value of the difference intuitively reflects the degree of fluctuation of the working hours, and provides a quantitative indicator for abnormal early warning and efficiency evaluation.

[0194] The total working time calculated this time, 357 seconds, is compared with the aforementioned threshold range. When calculating the working time fluctuation range, the current value is directly examined to see if it falls within the range. If it falls within the range, the fluctuation range is considered normal; if it exceeds the range, it is determined that there is an overflow. For example, when the total working time calculation value of a certain batch of follow-up processes reaches 425 seconds, exceeding the upper limit of 400 seconds, the fluctuation range is a positive overflow of 25 seconds, indicating that the operation time is significantly longer than expected.

[0195] Step S540: If the fluctuation range of working hours shows an overflow, an anomaly marker is generated and embedded into the initial structured report data. The report fields are reorganized to generate an optimized report output with integrated anomaly detection logic.

[0196] The optimized report output is obtained using the following formula:

[0197] (27)

[0198] In formula (27), This represents the optimized report output, which is a structured report dataset that integrates anomaly detection logic and includes original work hour data and anomaly labeling fields. This represents the initial structured report data, a basic work hour report, which includes basic fields such as batch, process, work hour value, and timestamp. Indicates that for the first Each work hour contains a set of embedded exception fields, which identify whether a single work hour record triggers an exception warning. Indicates the first The overflow anomaly amount per work hour, output value, with the unit consistent with work hours, is used to measure the degree of anomaly deviation. An exception is marked only when When the value is 1, abnormal information is added to the report to achieve field reorganization. The control logic of formula (27) is based on the judgment of work time fluctuation overflow, dynamically embedding abnormal fields and reorganizing the report structure to achieve a closed loop of "standardized output of normal work time data, automatic annotation of abnormal data and field enhancement", supporting the visualization and business implementation of work time analysis and abnormal analysis.

[0199] Overflow This can be derived from the following formula:

[0200] (28)

[0201] In formula (28), Indicates the first The overflow anomaly amount per work hour, output value, with the unit consistent with work hours, is used to measure the degree of anomaly deviation. This represents the actual working hours. The input value is the working hours consumed in the current record. Indicates the first The upper limit threshold for working hours is set, with a preset maximum allowed working hours. Exceeding this limit will trigger a positive overflow. This represents the lower limit threshold of the working hours. The minimum allowed working hours are preset. If the working hours are lower than this, a negative overflow is triggered (but the output is a positive deviation). When the actual working hours exceed the upper and lower limits, a positive overflow flag value is generated; otherwise, it is 0. The control logic of formula (28) is to quantify the overflow range of working hours that exceeds the normal range by using the double difference between the upper limit excess and the lower limit deficiency, and output a non-negative value as the abnormal overflow amount.

[0202] Exception marking This can be derived from the following formula:

[0203] (29)

[0204] In formula (29), The error is marked as follows: 1 = overflow error / time deviation exceeds the standard, 0 = no error. This indicates the actual working hours, the actual time spent on the current process / task (unit: seconds / minute). This indicates the baseline working hours, historical statistics, or standard working hours, serving as a benchmark reference value for the operation. It represents the permissible standard deviation of fluctuation, the statistical fluctuation range of the working time series, and the quantification of the normal fluctuation range. This represents the preset multiple threshold, the significance coefficient for anomaly detection, typically ranging from 1 to 3. A higher value indicates a more lenient judgment; when fluctuations exceed... When the deviation is more than 1 times the standard deviation, it is marked as 1, indicating an overflow anomaly. The control logic of formula (29) is used to identify abnormal work intervals where the time deviation exceeds the standard based on the baseline working hours and the fluctuation threshold. Through the idea of ​​statistical process control (SPC), it is determined whether the time deviation significantly exceeds the normal fluctuation range.

[0205] Once an overflow is detected, the system immediately generates an anomaly flag. This flag is embedded in the original report data as specific fields, such as adding "Anomaly Identifier: Time Exceeded" and "Overflow Value: +25 seconds." The report fields are then reorganized, with anomaly-related fields highlighted or prioritized, resulting in an optimized report output that integrates anomaly detection logic. This reorganized report not only retains the original accounting information but also visually highlights potential problems, allowing managers to quickly locate abnormal batches and processes. For example, another record shows that an operator completed the same back-end process with a total calculated time of 295 seconds, below the lower limit of 320 seconds, resulting in a negative overflow of 25 seconds. The system also embeds an anomaly flag, indicating "Time Too Short," suggesting a possible missed invalid time interval or rushed operation. These two types of overflow flags jointly support the comprehensiveness of fluctuation monitoring, ensuring that both excessively high and low efficiency are captured promptly. Through the above threshold comparison and anomaly embedding mechanism, the optimized report output possesses built-in early warning capabilities. This design allows production managers to prioritize batches marked as abnormal without having to manually check each batch individually. This enables them to quickly take targeted measures, such as arranging for review, adjusting staffing, or optimizing process parameters, ultimately achieving more precise control over production pace and efficient use of resources.

[0206] Preferably, the intelligent sewing equipment time detection method based on deep learning provided in this embodiment includes step S600 as follows:

[0207] Step S610: Parse the optimized report output to extract key indicators, and construct a standard data transmission package to be synchronized based on the key indicators.

[0208] The standard data transmission packet to be synchronized is constructed using the following formula:

[0209] (30)

[0210] In formula (30), This represents a standard data transmission packet to be synchronized. It is a structured and standardized dataset containing key metrics and corresponding values, used for cross-system transmission. This represents the construction function and the core mapping function, which implements the conversion logic from reports to standard transmission packets. This represents the optimized report output. Input data is used to generate a set of optimized report outputs, which includes core information such as working hours, processes, and equipment status. The input parameters are the set of key indicators to be extracted, such as [working hours, equipment status, process code]. The control logic of formula (30) is used to extract key indicators from the parsed and optimized report output, generate standardized and structured data transmission packages, realize seamless docking of downstream systems / modules, and ensure the consistency and parsability of data transmission.

[0211] Extracted key indicator set This can be derived from the following formula:

[0212] (31)

[0213] In formula (31), This represents the set of key indicators extracted, which is then used to generate a comprehensive score to quantify the overall level. This indicates the number of indicators in the report, representing the total number of indicators in the key indicator set. Indicates the first The weight of each indicator, the weight coefficient, and the value range is [0, 1]. Indicates the first The key indicators are the original values ​​of individual key indicators. The control logic of formula (31) is used to perform weighted summation on the extracted key indicators to generate a comprehensive score value. The importance of different indicators is quantified by weight, and a quantifiable comprehensive evaluation result is output.

[0214] When parsing the optimized report output, the system first focuses on extracting key indicators. These key indicators mainly include fields such as the actual total working hours, standard working hours, working hour deviation, anomaly status, and operator identification for each batch and process. Specifically, for the sewing process in batch number P20260120-003, the report extracts the actual total working hours as 412 seconds, the standard working hours as 380 seconds, the working hour deviation percentage as +8.42%, and the anomaly status as working hours exceeding the limit. It should be noted that these indicators are uniformly packaged into a structured standard data transmission package to be synchronized. The standard data transmission package to be synchronized uses a fixed field order for easy parsing by downstream systems.

[0215] Step S620: Write the standard data transmission packet to be synchronized into the real-time synchronization transmission queue, and send the data stream to the manufacturing execution system through the encrypted channel.

[0216] The real-time synchronization transmission queue is derived using the following formula:

[0217] (32)

[0218] In formula (32), Indicates time The real-time synchronization transmission queue stores the data to be synchronized in a first-in-first-out (FIFO) queue structure. Indicates time The real-time synchronous transmission queue, the queue state at the previous moment. The standard data transmission packet to be synchronized is standardized structured data (such as batch, process, working hours, equipment status, etc.) containing a unique identifier and a set of fields. This formula describes the process of writing the standard data transmission packet into the real-time synchronization transmission queue. The control logic of formula (32) is used to write the standard data transmission packet into the real-time synchronization transmission queue in an orderly manner. The data buffering, serialization and reliable transmission are realized through the enqueue operation of the queue, and finally sent to the Manufacturing Execution System (MES) through an encrypted channel.

[0219] The transmitted data stream is derived using the following formula:

[0220] (33)

[0221] In formula (33), This indicates a data stream, specifically the final transmitted data stream that has undergone encryption for secure transmission. This represents an encryption function, a mapping function that implements data encryption, supporting symmetric encryption, asymmetric encryption, and Chinese cryptographic algorithms. This represents the data packets in the queue, the input sequence of raw data packets, which contains the business data to be transmitted (such as working hours, procedures, equipment status, etc.). This represents the encryption key, a key parameter used for encryption, supporting symmetric / asymmetric / key derivation. This formula describes the process of generating an encrypted data stream through an encrypted channel. The control logic of formula (33) defines the process of encoding queued data packets based on the encryption key to generate an encrypted data stream, through an encryption function. To achieve secure transmission of queued data and ensure the confidentiality and integrity of the data stream.

[0222] Once the standard data transmission packet for synchronization is generated, it is immediately written to the real-time synchronization transmission queue. This real-time synchronization transmission queue employs a first-in, first-out (FIFO) mechanism to ensure that high-priority abnormal batch data is pushed first. When sending data streams to the manufacturing execution system via an encrypted channel, end-to-end encryption is performed using a session key specific to the current batch. For example, when the real-time synchronization transmission queue contains sewing process data that exceeds the time limit, the data stream embeds the batch number, process code, actual working hours, time deviation value, and abnormality description fields, keeping the entire packet size within 2KB to reduce network load.

[0223] Step S630: Receive the confirmation instruction returned by the manufacturing execution system for the data stream. If the verification is successful, update the production batch record in the production control database.

[0224] The verification result is obtained using the following formula:

[0225] (34)

[0226] In formula (34), This indicates the verification result, with a binary identifier: 1 = verification passed, 0 = verification failed. This indicates that the MES has received the confirmation instruction. The confirmation signal returned by the MES includes the instruction identifier, timestamp, and checksum. This represents the verification function, which maps the logic for validating the legality, completeness, and validity of the confirmation command. The formula indicates that the received confirmation command is verified; if it passes... The value is 1 if the instruction is valid and 0 otherwise. 1 = instruction is valid and triggers production control database update; 0 = instruction is invalid and triggers retry or alarm. The control logic of formula (34) is used to verify the validity of the acceptance confirmation instruction returned by the Manufacturing Execution System (MES) and output the binary verification result. If the verification passes, the record update of the production control database is triggered; if it fails, the retry logic is retained to ensure the accuracy and traceability of data synchronization.

[0227] Confirmation command This can be derived from the following formula:

[0228] (35)

[0229] In formula (35), This indicates that the MES has received the confirmation instruction. The confirmation signal returned by the MES includes the instruction ID, checksum, timestamp, etc. This represents the core logic of the manufacturing execution system, which performs verification and instruction generation on the input data stream. This represents the data stream, which is the business data to be synchronized (such as structured data like working hours, processes, and batch status). The control logic of formula (35) defines the receiving and confirmation mechanism of the manufacturing execution system (MES) for the data stream. The core logic is: after the system receives the original data stream, it generates a unique confirmation instruction feedback through integrity, validity and legality verification, forming a closed loop of "data receiving-verification-feedback".

[0230] The updated production batch record is derived using the following formula:

[0231] (36)

[0232] In formula (36), This indicates the updated production batch record, which includes the latest record with all fields such as batch number, process progress, working hours, and status. This indicates the existing production batch records, which are the original batch records to be updated in the production control database, including basic batch information and historical status. This represents the update function, the core mapping logic, and defines how to refresh the record after successful verification (e.g., incremental overwrite, field appending, state iteration). Indicates a successful verification flag, a binary / multi-valued identifier. =1 indicates that the verification passed. =0 indicates verification failure; multi-level verification results are supported (e.g., partial verification passes). This indicates an indicator function, a conditional trigger function, which is activated only if validation passes ( =1) execute the update logic, otherwise output the original record. This formula is used to update the production batch record in the production control database if the verification passes. The control logic of formula (36) defines the incremental update logic of the production batch record (PBR): only when the verification passes, the batch record in the production control database is refreshed with the update function (update) combined with the verification result (VP); if the verification fails, the original record is retained to ensure the rigor of data updates.

[0233] Upon receiving the data stream, the Manufacturing Execution System (MES) performs integrity checks and signature verification. If the verification passes, it immediately returns a confirmation instruction, which carries the batch number and confirmation timestamp. In one embodiment, upon receiving the confirmation instruction, the system immediately updates the corresponding production batch record in the production control database, changing the original pending synchronization status to a synchronized status, and simultaneously writes the latest cycle time confirmation value fed back by the MES into the record. For example, the original record for the sewing process showed pending confirmation; after receiving confirmation, it is updated to confirmed, and the actual cycle time matches the reported value.

[0234] Step S640: Read the updated stored content in the production control database and attach a timestamp to obtain the final working hour detection record.

[0235] The final time inspection record is obtained using the following formula:

[0236] (37)

[0237] In formula (37), Indicates the first The final generated time detection record contains a structured final record including time values, batch information, and timestamps. This represents the set of updated stored content in the production control database. The work hour information is the latest work hour data (such as processing time, accuracy, etc.) read from the production control database. Indicates the first The corresponding timestamp attached to each work hour item is a time identifier for data generation or synchronization, used for tracing and sorting. The record combination operator defines the logic of associating work time content with timestamps and encapsulating it into a complete record. The control logic of formula (37) is to associate and combine the updated work time data in the production control database with timestamps to generate a final work time detection record with time traceability, ensuring the traceability and time sequence accuracy of work time data.

[0238] When reading updated stored content from the production control database, the system automatically appends the current server timestamp to form a final time check record. This final time check record not only includes the original calculated value and anomaly flag, but also adds the synchronization completion time and the manufacturing execution system confirmation time. For example, for the same batch P20260120-003, a final record is generated: batch number, process code, actual working time 412 seconds, time deviation +8.42%, anomaly flag: time exceeded limit, synchronization completion time 2026-01-20 14:35:22, confirmation time 2026-01-20 14:35:45. Such a final record can be directly used for subsequent traceability analysis or performance evaluation. In one embodiment, if the time deviation of a certain process remains large but does not trigger a serious alarm, the system still pushes the time deviation information to the manufacturing execution system in a timely manner through this synchronization mechanism, so that process engineers can adjust parameters in real time and avoid small anomalies accumulating into batch problems. The entire process, from report parsing to final record generation, forms a closed-loop data link, ensuring that production data is real-time, accurate, and consistent, and improving the speed of anomaly response and the accuracy of resource scheduling.

[0239] Another aspect of the present invention relates to a deep learning-based intelligent sewing equipment time detection system, used to execute the aforementioned deep learning-based intelligent sewing equipment time detection method, including a multimodal dataset acquisition module 10, a state sequence label acquisition module 20, an interval classification result acquisition module 30, a total time calculation value determination module 40, a report output acquisition module 50, and a final time detection record acquisition module 60. The multimodal dataset acquisition module 10 is used to collect sewing equipment motor start / stop signals and fabric tension change data through sensors, simultaneously acquire operation image sequences, and fuse them into a multimodal dataset after data preprocessing. The state sequence label acquisition module 20 is used to input the multimodal dataset into a target detection model to identify the key states of sewing equipment operation and operator on-duty status, obtaining state sequence labels. The interval classification result acquisition module 30 is used to extract thread-cutting actions from the state sequence labels. The relevant feature vectors are analyzed using deep learning algorithms to determine the temporal correlation between feature vectors related to the thread-cutting action. This temporal correlation is used to label valid sewing intervals and invalid time-consuming intervals, resulting in interval classification results. The total working time calculation module 40 obtains the interval classification results and calculates the duration of each valid sewing interval, accumulating them to generate a single-process working time value. It dynamically corrects historical production data deviations to determine the total working time calculation value. The report output acquisition module 50 generates structured report data based on the total working time calculation value, integrating anomaly detection logic. This logic is used to detect working time fluctuations and trigger alarm signals, resulting in optimized report output. The final working time detection record acquisition module 60 extracts key indicators from the optimized report output, interfaces with the manufacturing execution system, and achieves real-time data synchronization. If synchronization is complete, the production control database is updated to obtain the final working time detection record.

[0240] The intelligent sewing equipment time detection method and system based on deep learning provided by this invention achieves the following beneficial effects compared with the prior art:

[0241] 1. The core is to accurately distinguish between effective and invalid working time intervals in sewing operations through multimodal data fusion and deep learning, thereby solving the problem of mixed calculation in traditional working time statistics.

[0242] 2. The key steps are the extraction of thread-cutting action features and temporal correlation analysis, which serve as the core criteria for determining the effective sewing interval and ensuring the accuracy of work time division.

[0243] 3. The entire process realizes a closed loop of "data acquisition - status identification - time calculation - anomaly warning - system synchronization", and the output time data can directly serve production control decisions.

[0244] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A deep learning-based sewing equipment working hour intelligent detection method, characterized in that, Includes the following steps: S100: Collects start / stop signals of the sewing equipment motor and fabric tension change data through sensors, synchronously acquires operation image sequences, and merges them into a multimodal dataset after data preprocessing; S200. Input the multimodal dataset into the target detection model to identify the key states of sewing equipment operation and operator on duty, and obtain state sequence labels; S300. Extract thread-cutting action-related feature vectors from the state sequence labels, use a deep learning algorithm to analyze the temporal correlation between the thread-cutting action-related feature vectors, determine the temporal correlation to mark valid sewing intervals and invalid time-consuming intervals, and obtain interval classification results; S400: Obtain the interval classification results and calculate the duration of each valid sewing interval, accumulate them to generate a single-process working time value, dynamically correct historical production data deviations, and determine the total working time calculation value. S500: Generate structured report data for the total working hours calculation value, integrate anomaly detection logic, the anomaly detection logic is used to judge working hour fluctuations and trigger alarm signals, and obtain optimized report output; S600 extracts key indicators from the optimized report output, connects to the manufacturing execution system interface, and realizes real-time data synchronization. If the synchronization is completed, it updates the production control database to obtain the final time detection record. Step S300 includes: S310. Obtain the time point of the line-cutting action in the state sequence label, and use a long short-term memory network to map the multimodal fusion feature tensor fragment into a line-cutting action-related feature vector; S320. Input the feature vector related to the wire-cutting action into the self-attention mechanism model to construct a temporal correlation matrix; S330. If the value of an element in the time sequence correlation matrix is ​​greater than a preset threshold, it is determined that there is continuous operation logic; otherwise, it is determined that there is an abnormal pause. S340. Mark the valid sewing intervals according to the continuous operation logic, and mark the invalid time intervals according to the abnormal pauses. Summarize the valid sewing intervals and the invalid time intervals to obtain the interval classification result.

2. The intelligent sewing equipment time detection method based on deep learning as described in claim 1, characterized in that, Step S100 includes: S110: Acquire the sewing equipment motor start / stop signal, fabric tension change data, and work image sequence, and lock the valid work window according to the motor start / stop signal; S120. Filter the fabric tension change data within the effective working window to obtain a tension trend sequence; S130. Extract the sewing area image frame set from the work image sequence based on the abrupt change time of the tension trend sequence; S140. The set of image frames of the sewing area is concatenated with the numerical feature vector corresponding to the tension trend sequence to generate a multimodal dataset.

3. The intelligent sewing equipment time detection method based on deep learning as described in claim 1, characterized in that, Step S200 includes: S210. Obtain a multimodal dataset, extract and fuse image spatial texture features and tension temporal fluctuation features to obtain a multimodal fusion feature tensor; S220. Input the multimodal fusion feature tensor into the region proposal network to generate a set of detection boxes, wherein the set of detection boxes contains target category and location information; S230. Calculate the Euclidean distance between the limb and the head of the machine based on the set of detection frames, and determine the single-frame state classification code in combination with the tension fluctuation range; S240. Perform context constraint modeling on the original state sequence formed by the single-frame state classification and coding, correct the original state sequence, and obtain the state sequence labels of sewing equipment operation and operator on duty.

4. The intelligent sewing equipment time detection method based on deep learning as described in claim 1, characterized in that, Step S400 includes: S410. Traverse the start timestamp and end timestamp of the effective sewing interval in the interval classification results, and calculate the difference between the start timestamp and end timestamp of the effective sewing interval to obtain the effective sewing duration. S420. Accumulate the multiple effective sewing durations within the same production batch to generate a single-process time value; S430. Calculate the difference between the single-process time value and the historical benchmark average to obtain the time deviation value, wherein the historical benchmark average is calculated based on historical production data that matches the current process code; S440. Determine the weighted correction coefficient based on the time deviation value through a nonlinear mapping function, and use the weighted correction coefficient to correct the single-process time value to determine the total time calculation value.

5. The intelligent sewing equipment time detection method based on deep learning as described in claim 1, characterized in that, Step S500 includes: S510. Obtain the total working hours calculation value, and perform serialization processing by combining the associated production batch number and process code to generate the initial structured report dataset. S520. Parse the process codes contained in the initial structured report dataset and retrieve the working hour distribution records within the historical production cycle to determine the working hour fluctuation threshold range. S530. Compare the total working hours calculation value with the working hours fluctuation threshold range to calculate the working hours fluctuation amplitude; S540. If the fluctuation range of the working hours shows an overflow, an anomaly marker is generated and embedded in the initial structured report data. The report fields are reorganized to generate an optimized report output with integrated anomaly detection logic.

6. The intelligent sewing equipment time detection method based on deep learning as described in claim 1, characterized in that, Step S600 includes: S610. Parse the optimized report output to extract key indicators, and construct a standard data transmission packet to be synchronized based on the key indicators. The standard data transmission packet to be synchronized is constructed using the following formula: ； in, This represents a standard data transmission packet to be synchronized. Indicates the constructor function. This indicates the optimized report output. This represents the set of key indicators extracted. Extracted key indicator set This can be derived from the following formula: ； in, This indicates the number of indicators in the report. Indicates the first The weight of each indicator, Indicates the first Key indicator values; S620: Write the standard data transmission packet to be synchronized into the real-time synchronization transmission queue, and send the data stream to the manufacturing execution system through the encrypted channel; S630. Receive the acceptance confirmation instruction returned by the manufacturing execution system for the data stream. If the verification is successful, update the production batch record in the production control database. S640. Read the updated stored content in the production control database and attach a timestamp to obtain the final working hour detection record.

7. The intelligent sewing equipment time detection method based on deep learning as described in claim 6, characterized in that, In step S620, the real-time synchronization transmission queue is obtained using the following formula: ； in, Indicates time Real-time synchronous transmission queue, Indicates time Real-time synchronous transmission queue, This represents a standard data transmission packet to be synchronized; The transmitted data stream is derived using the following formula: ； in, Represents a data stream. Indicates the encryption function. This represents the data packets in the queue. This represents the encryption key.

8. The intelligent sewing equipment time detection method based on deep learning as described in claim 7, characterized in that, In step S630, the verification result is obtained using the following formula: ； in, This indicates the verification result. This indicates that the confirmation instruction has been received. This represents the verification function; Receive confirmation command This can be derived from the following formula: ； in, Indicates manufacturing execution system; The updated production batch record is derived using the following formula: ； in, This indicates the updated production batch record. This indicates existing production batch records. Indicates the update function, This indicates that the verification has passed. This indicates an indicator function.

9. A deep learning-based intelligent sewing equipment time detection system, used to execute the deep learning-based intelligent sewing equipment time detection method as described in any one of claims 1 to 8, characterized in that, include: The multimodal dataset acquisition module (10) is used to collect the start and stop signals of the sewing equipment motor and the change data of fabric tension through sensors, synchronously acquire the operation image sequence, and fuse it into a multimodal dataset after data preprocessing; The state sequence label acquisition module (20) is used to input the multimodal dataset into the target detection model, identify the key states of sewing equipment operation and operator on duty, and obtain state sequence labels; The interval classification result acquisition module (30) is used to extract the thread-cutting action-related feature vectors from the state sequence labels, use a deep learning algorithm to analyze the temporal correlation between the thread-cutting action-related feature vectors, determine the temporal correlation to mark the effective sewing intervals and invalid time-consuming intervals, and obtain the interval classification result. The total working hours calculation value determination module (40) is used to obtain the interval classification results and calculate the duration of each effective sewing interval, accumulate and generate the single-process working hours value, dynamically correct the deviation of historical production data, and determine the total working hours calculation value; The report output acquisition module (50) is used to generate structured report data for the total working hours calculation value, integrate anomaly detection logic, the anomaly detection logic is used to judge working hour fluctuations and trigger alarm signals, and obtain optimized report output; The final working time detection record acquisition module (60) is used to extract key indicators from the optimized report output, connect to the manufacturing execution system interface, realize real-time synchronous transmission of data, and update the production control database if the synchronous transmission is completed to obtain the final working time detection record.

Images