A screw locking quality online evaluation and classification method based on real-time torque curve feature recognition

Abstract

The application discloses a screw locking quality online evaluation and classification method based on real-time torque curve feature recognition, which comprises the following steps: collecting torque values and angle values in the locking process in real time and performing filtering pretreatment; dividing the torque-time curve into four characteristic stages of a starting section, a screwing-in section, a fitting section and a locking section, extracting key feature parameters of each stage respectively, and constructing a multi-dimensional feature vector; performing dimension reduction on the feature vector by using principal component analysis; constructing a defect classification model based on an improved weighted support vector machine, and relieving the sample imbalance problem by a class weighting strategy; inputting the new locking curve into the classification model after online feature extraction and dimension reduction, and outputting a defect type judgment result. By means of four-stage fine decomposition of the torque curve and multi-dimensional feature extraction, the application can accurately identify various defect types such as a loose tooth, false locking, screw and hole mismatch, and foreign matter in the hole, and has the advantages of good online real-time performance, strong small sample adaptability and the like.

Description

Technical Field

[0001] This invention belongs to the field of industrial assembly quality control technology, specifically relating to an online assessment and classification method for screw fastening quality based on real-time torque curve waveform feature extraction and deep learning. It is applicable to the quality inspection and defect classification of screw fastening processes in automatic screw fastening machines, intelligent electric screwdrivers, and manual assembly production lines. Background Technology

[0002] Threaded fasteners are one of the most widely used connection methods in mechanical assembly, and their assembly quality directly affects the structural strength, operational reliability, and safety of the product. Especially in the manufacturing of small household appliances, a single product such as a coffee maker often involves dozens or even hundreds of screw fastening points. A defect in the connection of any one of these fastening points can lead to product malfunction or safety hazards. Therefore, real-time quality monitoring and defect detection of the screw fastening process has always been an important research topic in the field of industrial assembly.

[0003] Currently, common methods used in the industry to evaluate screw tightening quality include direct torque control, torque control-angle monitoring, and angle control-torque monitoring. These methods all use whether the final tightening parameters (such as the final torque value and the final angle value) meet preset standards as the criterion for judgment, and are essentially static testing strategies based on whether the "result meets the standard".

[0004] In recent years, with the development of intelligent electric screwdriver technology and the advancement of sensor technology, real-time acquisition of torque and angle data during the tightening process has become possible, and process analysis methods based on tightening curves have gradually become a research hotspot. Regarding intelligent electric screwdrivers and torque monitoring, relevant technical solutions have already been disclosed in patents. For example, Chinese invention patent CN115284207B discloses a method, device, and storage medium for identifying the holding torque of an intelligent electric screwdriver. This method, by real-time acquisition of torque data during the operation of the intelligent electric screwdriver, establishes a sliding window of preset duration for newly added torque data, establishes a sliding window comparison mechanism, identifies the contact point, and determines the holding torque, thereby improving the accuracy of contact torque identification to a certain extent. Chinese invention patent CN119115858B further discloses a device for identifying the holding torque of an intelligent electric screwdriver. By real-time acquisition of torque data during the operation of the intelligent electric screwdriver, it establishes a sliding window of preset duration as the first sliding window for newly added torque data, establishes a sliding window comparison mechanism, identifies the contact point, and determines the holding torque, thereby improving the accuracy of contact torque identification. In addition, Chinese invention patent application CN121834220A discloses a tightening curve intelligent analysis method and system, which uses a combination of deep learning feature extraction and adaptive threshold dynamic search to obtain the contact point angle position, and uses an autoencoder-based distribution adaptive recognition and multi-strategy fusion method to obtain the difference threshold, thus realizing intelligent analysis of the tightening process.

[0005] In the area of ​​tightening curve state identification and quality assessment, existing research has attempted to introduce artificial intelligence algorithms for anomaly detection and failure cause analysis. For example, Chinese invention patent CN113869502A discloses a bolt tightening failure cause analysis method based on deep neural networks, which automatically identifies the cause of failure by performing Gaussian mixture model clustering on unqualified curves and then combining it with multinomial regression. Chinese invention patent CN116226692A discloses a tightening anomaly identification method based on dynamic time planning, which identifies whether the tightening process is abnormal by calculating the dynamic time warping distance between tightening curves. Chinese invention patent CN119512246B discloses a screw feeding pressure adaptive adjustment method, which evaluates the tightening quality by online monitoring of the torque range in the tightening state sequence. The patent document "Intelligent Identification Method of Abnormal Assembly Curve of Fasteners" applied for by SAIC Volkswagen Automotive Co., Ltd., publication number CN119168971A, achieves automated acquisition and identification of tightening curves by using convolutional neural network artificial intelligence algorithms. In the field of acoustic emission signal-based detection, Chinese invention patent application CN120372171A discloses a method for detecting screw tightening anomalies, which detects anomalies by collecting the original vibration signals generated by the electric screwdriver under test during screw tightening. Furthermore, Chinese invention patent CN118404317B discloses a method, apparatus, tightening system, and storage medium for tightening workpieces. It determines multiple decision frames for the tightening curve based on bolt quality and tightening process, and identifies anomalies by determining whether the tightening curve correctly enters and exits each decision frame according to the rules.

[0006] The aforementioned technologies have played a positive role in improving the quality inspection level of screw fastening. However, those skilled in the art have discovered in their research and practice that the existing solutions still have the following technical defects and unresolved technical problems: First, the ability to identify defect types is limited. Existing methods mainly focus on whether the tightening endpoint parameters meet the preset range, or can only output a binary judgment result of "pass / fail," failing to classify specific defect types in a fine-grained manner. However, in actual production, the causes and solutions for different types of defects, such as stripped threads, false locking, mismatch between screws and holes, and foreign objects in holes, are drastically different. Traditional single torque value judgment methods are considered by the industry to be inadequate for accurately identifying latent defects such as "false tightening," "stripped threads," and "torque decay." If only an unqualified signal is output, process engineers cannot know the specific cause of the unqualified result, and process improvement lacks clear direction.

[0007] Second, there is insufficient refinement in the processing of local waveform features. While existing technologies focus on the rate of change of the torque curve or the identification of contact points, they emphasize the identification of a few key points rather than the feature analysis of the entire tightening process. Different stages of the tightening process—starting, screwing in, contacting, locking, and holding pressure—have different representational meanings for tightening quality. For example, the torque fluctuation characteristics of the screwing-in stage are highly correlated with stripped teeth defects, while the peak arrival time of the locking stage and the attenuation rate of the holding pressure stage are closely related to false locking defects. Existing technologies lack a systematic method for independently extracting features from each stage and establishing defect correspondences, resulting in the waste of a large amount of discriminative feature information.

[0008] Third, static threshold methods are insensitive to process anomalies. Existing methods mostly use fixed torque or angle thresholds for judgment. However, there are many nonlinear and uncertain factors in the fastening process, such as fluctuations in the dimensions of incoming screws (the diameter tolerance of M3 screws in different batches can reach ±0.05mm), differences in the surface condition of workpieces (fluctuations in electroplating thickness, oil residue), and changes in workshop ambient temperature. Static thresholds cannot adapt to these dynamic changes, resulting in a high failure rate. According to statistics, the failure rate of traditional static threshold methods for process anomalies can be as high as 15% or more.

[0009] Fourth, the scarcity of abnormal samples makes model training difficult. Deep learning-based methods require a large amount of labeled data for model training. However, in actual production environments, abnormal conditions such as stripped threads, seizing threads, and foreign objects in holes occur relatively infrequently, with fault samples often accounting for less than 5% of the total samples. Under these conditions, traditional deep neural network models are prone to underfitting or overfitting due to imbalanced samples, resulting in severely insufficient generalization ability and making it difficult to promote and apply them on actual production lines.

[0010] Fifth, insufficient real-time capability. In some of the aforementioned existing technologies, analysis and output can only be performed after tightening is complete. However, if certain anomalies (such as stripped threads occurring during the threading stage) can be identified and the locking process interrupted in a timely manner, subsequent non-conforming operations can be avoided. On mass production assembly lines, early detection of anomalies and prevention of defective products from flowing into subsequent processes has significant economic implications.

[0011] In summary, there is an urgent need in the existing technology for a screw fastening quality assessment method that can perform refined, multi-stage feature extraction of the torque curve during the screw fastening process, accurately identify and classify various defect types, and possess online real-time capability and resistance to sample scarcity, so as to meet the "zero defect" requirement for assembly quality in the field of intelligent manufacturing. Summary of the Invention

[0012] The purpose of this invention is to overcome the shortcomings of the prior art and provide an online assessment and classification method for screw fastening quality based on real-time torque curve feature recognition. This method achieves fine-grained identification and classification of various defect types, such as stripped threads, false fastening, screw-hole mismatch, and foreign objects in the hole, by decomposing the torque-time curve into four stages and extracting multi-dimensional features, combined with a dynamic classifier based on an improved weighted support vector machine. It also has online real-time assessment capabilities and small-sample learning capabilities, thereby improving the accuracy and operability of screw fastening quality detection.

[0013] To achieve the above objectives, this invention provides an online assessment and classification method for screw fastening quality based on real-time torque curve feature recognition. The method includes the following steps: Step S1: Real-time acquisition and preprocessing of torque curve High-precision torque sensors and angle encoders are integrated on the tightening shaft of screw-locking equipment, such as automatic screw machines or intelligent electric screwdrivers, to collect torque values ​​T(t) and angle values ​​θ(t) in real time during the screw-locking process at a sampling frequency of not less than 500Hz, where t represents time.

[0014] The collected raw torque data undergoes digital filtering, employing either moving average filtering or median filtering to eliminate high-frequency noise interference, resulting in smoothed torque-time curves T(t) and torque-angle curves T(θ). Each curve is then aligned along its time axis to ensure precise synchronization of torque and angle data at the same timestamp.

[0015] Step S2: Four-stage characteristic decomposition of the torque curve The preprocessed torque-time curve T(t) is divided into the following four characteristic stages, each corresponding to an interval with different physical meanings during the locking process: 1. Setup Phase: From the start of the electric screwdriver's operation to the first contact of the screw head with the workpiece surface. The torque characteristics of this phase include the peak starting torque T_start_peak and the duration Δt_start.

[0016] 2. Run-down Phase: From the moment the screw head contacts the workpiece surface until it is about to fully engage with it. During this phase, the torque increases steadily. Key characteristics include the average slope k_run and the run-down time Δt_run.

[0017] 3. Sealing Phase: This is the transition point from the moment the screw head contacts the workpiece surface to the completion of the sealing process and the point where the torque begins to accelerate. The key characteristics of this phase are the sealing torque value T_seating and the sealing phase duration Δt_seating. The sealing point is identified using the sliding window rate of change ratio method: a 20ms forward sliding window is established centered on the current torque data point, and a backward sliding window is established using the historical torque data from the previous 50ms. The torque change rate R_front and the torque change rate R_rear of the backward sliding window are calculated respectively. When R_front / R_rear ≥ 3.0, the point is marked as the initial sealing point, and its position is corrected based on real-time data within the following 5ms.

[0018] 4. Tightening Phase: This phase extends from the point of final contact until the electric screwdriver reaches the target torque value. Characteristics of this phase include the peak torque T_peak, the time to reach the peak torque Δt_peak, and the torque rise slope k_tight. After reaching the target torque value, the electric screwdriver enters the holding phase, and the torque decay rate δ_T_hold during this phase is extracted as a supplementary feature.

[0019] Step S3: Construction and Dimensionality Reduction of Multidimensional Feature Vectors The feature parameters extracted in step S2 are quantized and calibrated to construct the following multidimensional feature vector: F=[T_start_peak,Δt_start,k_run,Δt_run,T_seating,Δt_seating,T_peak,Δt_peak,k_tight,δ_T_hold,Δθ_total] Where Δθ_total is the total rotation angle during the entire locking process.

[0020] The multidimensional feature vector is standardized to eliminate scale differences between different units. Furthermore, principal component analysis (PCA) is used to reduce the dimensionality of the feature vector, retaining principal components with a cumulative variance contribution rate of over 95%, reducing feature redundancy and improving the computational efficiency of subsequent classifiers.

[0021] Step S4: Constructing a defect classification model based on an improved weighted support vector machine A defect classification model is constructed using an improved weighted support vector machine (W-SVM). The model input is the dimensionality-reduced feature vector, and the model output includes multiple defect types. The specific steps are as follows: S4.1: Using the dimensionality-reduced feature vectors from step S3 as model input and the pre-labeled multiple defect types as model output, construct a multi-classification training dataset.

[0022] S4.2: For each defect category k, the category weight w_k is defined as follows: w_k=1 / n_k×α Where n_k is the number of training samples for the k-th defect, and α is the global adjustment factor (α∈[1,10]). This weight setting allows defect categories with fewer samples to receive higher weights, thereby alleviating the sample imbalance problem.

[0023] S4.3: A multi-class decision framework is constructed using a one-vs-one (OVO) strategy. A binary SVM sub-classifier is trained between every two defect categories, for a total of C(N,2)=10 sub-classifiers (when N=5).

[0024] S4.4: The kernel function of each sub-classifier adopts the radial basis function (RBF), and the kernel parameter γ and the penalty factor C are optimized by grid search combined with five-fold cross-validation.

[0025] S4.5: Establish characteristic maps and mathematical judgment logic for typical defects. Based on historical locking data, establish torque curve characteristic maps corresponding to the following five types of typical defects through manual re-inspection and marking, and formulate rigorous mathematical judgment logic for each type of defect: For stripped thread defects (i.e., thread damage), the torque curve is characterized by continuous fluctuations in torque during the threading section, which cannot rise steadily, and the peak torque cannot reach the set value. The corresponding mathematical judgment logic is: the standard deviation of the torque during the threading section σ_T_run is greater than 0.3 N·m, and the average slope of the threading section k_run is less than 0.5 N·m / s.

[0026] For false locking defects (i.e., screws not fully engaged), the torque curve is characterized by the torque reaching the target torque value too early, followed by a rapid decrease in torque during the holding phase after the peak torque. The corresponding mathematical judgment logic is: the time to reach the peak torque Δt_peak is less than 60% of the historical average for this workstation, and the torque decay rate δ_T_hold during the holding phase is greater than 15%.

[0027] For screw-hole mismatch defects, the torque curve is characterized by an abnormally high peak torque during the starting phase, or a sharp increase followed by a sudden drop in torque during the screwing-in phase. The corresponding mathematical judgment logic is: the starting peak torque T_start_peak is greater than 30% of the target torque, and a negative slope range appears during the screwing-in phase.

[0028] For defects involving foreign objects inside the hole, the torque curve is characterized by one or more unexplained torque spikes in the screw-in section. The corresponding mathematical judgment logic is: the torque change rate dT / dt in the screw-in section is greater than 0.8 N·m / ms for a duration exceeding 50 ms, and occurs at least once.

[0029] For a qualified locking device, its torque curve is characterized by all characteristic parameters being within the normal range at each stage. The corresponding mathematical judgment logic is: all characteristic parameters are within the preset qualified range, and the torque attenuation rate δ_T_hold in the pressure holding stage is less than or equal to 8%.

[0030] The threshold parameters in the above judgment logic can be customized according to the specific product process requirements.

[0031] Step S5: Online Assessment and Classification The newly acquired locking torque curves undergo feature decomposition in step S2 and feature extraction and dimensionality reduction in step S3. The curves are then input into all sub-classifiers for voting classification. The defect category with the highest number of votes is output as the final classification result. If the difference between the highest and second-highest votes is less than a preset threshold (e.g., 3 votes), a "suspicious" label is output, suggesting manual review. This sample is then added to the annotation pool for incremental model updates.

[0032] Furthermore, regarding online model updates and real-time feedback, the following applies: During actual production line operation, the following real-time feedback mechanism is implemented: The quality classification results of each product are recorded and linked to the product serial number to form a traceable quality file; when the number of consecutive occurrences of the same type of defect exceeds a preset threshold (e.g., 5 times), the system automatically issues a warning signal and displays the specific defect type on the workstation display screen, while also providing process adjustment suggestions; newly labeled samples are periodically added to the training set for incremental training and updates of the W-SVM model; samples marked as "suspicious" are added to the training set after manual review and confirmation, and given priority in participating in incremental model updates. III. Technical Effects Compared with the prior art, the present invention has the following beneficial technical effects: 1. Finer granularity of defect identification and significantly improved classification accuracy. This invention performs a four-stage fine decomposition of the torque curve, extracting feature parameters with clear physical meaning from each stage. Combined with typical defect feature maps and multi-classification discrimination logic, it can accurately identify four typical locking defects: stripped threads, false locks, screw-hole mismatch, and foreign objects in the hole. As shown in the comparative experiments in the specific implementation section, the overall accuracy of this invention reaches 96.7%, which is 22.4 percentage points higher than the traditional static threshold method (74.3%) and 8.0 percentage points higher than the CNN-BiLSTM deep learning method (88.7%).

[0033] 2. Enhanced adaptability to imbalanced sample problems. This invention employs an improved weighted support vector machine model, assigning higher weights to scarce fault samples through a class weighting strategy, effectively mitigating the underfitting problem caused by the scarcity of abnormal samples in industrial settings. Compared to traditional deep neural networks where the recognition accuracy drops sharply when the proportion of abnormal samples is less than 5%, this invention maintains high classification accuracy even under small sample conditions.

[0034] 3. Excellent online real-time performance, meeting the real-time monitoring needs of the production line. The feature extraction and classification judgment of this invention have low computational complexity, and the complete evaluation of a single locking can output the result within 1 second after locking is completed, which is far lower than the time cost of traditional manual visual inspection or offline detection.

[0035] 4. High interpretability, facilitating rapid problem localization by process engineers. The four-stage feature decomposition method of this invention provides clear physical interpretability for the classification results. When the system identifies a "slippery tooth" defect, process engineers can directly correlate it to "abnormal torque fluctuation in the screw-in section," thereby quickly locating the root cause of the problem.

[0036] 5. The model possesses continuous learning capabilities. Through the labeling of suspicious samples and the incremental update mechanism of the model, the classification model of this invention can continuously learn new defect patterns during the production process, gradually improving its adaptability to newly emerging abnormal working conditions. Attached Figure Description

[0037] Figure 1 This is an overall flowchart of the online assessment and classification method for screw fastening quality based on real-time torque curve feature recognition according to the present invention; Figure 2 This diagram illustrates the four-stage feature decomposition of the torque-time curve in the screw fastening process, which is part of an online assessment and classification method for screw fastening quality based on real-time torque curve feature recognition, according to the present invention. Detailed Implementation

[0038] Please refer to the following: Figures 1 to 2 The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] Example 1: Implementation and Deployment of Automatic Screw Fastening Machine This embodiment uses an automatic screw-fastening machine in the coffee machine assembly line of an electrical appliance manufacturing company as an application scenario to describe in detail the specific implementation of the technical solution of the present invention. This production line involves 22 screw-fastening stations, and each coffee machine uses approximately 50 screws, representing a typical high-volume, high-frequency screw-fastening scenario.

[0040] Step 1: Hardware Deployment A Kistler 5074A dynamic torque sensor (range 0~5 N·m, accuracy ±0.3%, sampling frequency 1 kHz) and a magnetic angle encoder (resolution 0.05°) are integrated on the tightening shaft of the automatic screw fastening machine. Torque and angle signals are synchronously acquired via a data acquisition card at a sampling frequency of 1 kHz. The acquisition card output is connected to the industrial computer at the workstation via a USB 3.0 interface. Real-time data acquisition and processing software is installed on the industrial computer at the workstation. The data buffer adopts a circular queue structure (capacity 5000 points, corresponding to approximately 5 seconds of fastening process) to ensure continuous data recording without loss.

[0041] Step 2: Torque curve preprocessing and stage division The acquired torque-time series T(t) was preprocessed. First, a moving average filter with a window length of 5 was used to eliminate high-frequency noise. After filtering, the root mean square noise of the torque signal was reduced from 0.05 N·m to below 0.01 N·m. The filtered torque data was then processed according to... Figure 2 The diagram shows an automatic four-stage division.

[0042] The end point of the starting phase is marked as t0, the moment when the torque value first exceeds 0.02 N·m from 0. The start and end points of the screw-in phase are t0 and the identification position of the contact point t_seating, respectively. The contact point is identified using the sliding window rate of change ratio method: a 20 ms forward sliding window is established centered on the current torque data point, and a backward sliding window is established using the historical torque data of the previous 50 ms. The torque change rate R_front of the forward sliding window and the torque change rate R_rear of the backward sliding window are calculated respectively. When R_front / R_rear ≥ 3.0, the point is marked as the initial contact point. Based on the real-time data within the next 5 ms, if the torque value corresponding to the initial contact point fluctuates more than ±0.05 N·m, a linear extrapolation method is used for position correction, and the final contact point t_seating is determined.

[0043] Step 3: Feature Parameter Extraction The four-stage feature extraction method of this invention is used to extract features from the locking curves of 50 consecutive samples, and the feature parameters at each stage are calculated. In this embodiment, the specific acceptable range of feature parameters for the coffee machine body assembly station is as follows:

[0044] Step 4: Defect Sample Acquisition and Feature Map Construction This embodiment obtains various types of defect samples by artificially creating defects: Stripped thread samples: 25 samples were collected using screws with damaged threads. The characteristics included continuous torque fluctuations during the screw-in section, with a standard deviation σ_T_run > 0.35 N·m and an average slope k_run < 0.4 N·m / s during the screw-in section.

[0045] Fake lock samples: A 0.5mm thick shim was placed at the bottom of the screw hole to prevent the screw from fully engaging. A total of 25 samples were collected. The characteristics included a peak torque time Δt_peak < 250ms (the historical average for this station is 650ms), and a torque decay rate δ_T_hold > 18% during the pressure holding phase.

[0046] Foreign matter samples: Metal scrap (approximately 0.2 mg) was pre-placed in the screw holes, and a total of 20 samples were collected. The characteristic feature was a single torque spike during the screw insertion phase (torque change rate dT / dt > 1.0 N·m / ms, spike lasting approximately 80 ms).

[0047] Screw-hole mismatch samples: 20 samples were collected using M3.5 screws screwed into M3 threaded holes. The characteristic feature is that the peak torque T_start_peak reaches 1.8 N·m (36% of the target torque) during the starting phase, and a sudden drop in torque occurs during the screwing-in phase (negative slope 0.6 N·m / s).

[0048] Qualified samples: 300 samples were collected and fastened according to normal process conditions.

[0049] Step 5: Model Training and Parameter Optimization The 390 samples were randomly divided into a training set (273 samples) and a test set (117 samples) in a 7:3 ratio. Five-fold cross-validation was used to optimize the parameters of the W-SVM model on the training set. The RBF kernel was used, with the kernel parameter γ searching in the range [2^-5, 2^-3, 2^-1, 2^1, 2^3] and the penalty factor C searching in the range [0.1, 1, 10, 100, 1000]. The optimal parameter combination was obtained through grid search: γ = 2^1, C = 100.

[0050] The calculation process for category weights is as follows: For qualified samples n_normal=210, the weight w_normal=1 / 210×α (taking α=3, w_normal≈0.0143); for slipped tooth samples n_slip=17, the weight w_slip=1 / 17×3≈0.1765; for false lock samples n_false=18, the weight w_false≈0.1667; for foreign object samples n_foreign=14, the weight w_foreign≈0.2143; for mismatched samples n_mismatch=14, the weight w_mismatch≈0.2143. It can be seen that the category weight of faulty samples is approximately 12~15 times that of qualified samples, effectively focusing on scarce faulty samples.

[0051] Step 6: Online Assessment and Classification A trained W-SVM classification model is deployed on the production line to perform online evaluation of the tightening process for each screw. The evaluation process is as follows: after tightening, the system automatically extracts the four-stage feature parameters of the torque curve, compares them with the acceptable range, inputs them into the classification model, and outputs the classification result. The entire process takes about 0.8 seconds, meeting the real-time requirements of a 30-second cycle time production line.

[0052] Example 2: Comparative Experiment and Verification of Technical Effects This embodiment verifies the improved technical performance of the present invention compared to existing technologies through comparative experiments. The present invention is compared with the following two existing technical solutions on the same test set: Control Group 1: Traditional static threshold method. This method uses direct torque control, judging the fastening quality solely based on whether the final torque value is within the acceptable range of 4.5~5.5 N·m. This is the mainstream testing method currently used by the vast majority of production lines.

[0053] Control Group 2: CNN-BiLSTM deep learning method. The screw fastening detection method disclosed in CN202510863748 was used to construct a CNN-BiLSTM network model for end-to-end fastening state detection.

[0054] Comparative experimental conditions: Test set sample: 300 independent samples, including 210 qualified locks, 20 stripped threads, 25 fake locks, 20 with foreign objects in the hole, and 25 screws that do not match the hole.

[0055] Test hardware environment: Intel Core i7-12700 CPU, 32GB RAM, no GPU acceleration.

[0056] Evaluation metrics: Accuracy, Precision, Recall, F1 score, and time taken per evaluation.

[0057] All tests were conducted under the same conditions, and each method was tested three times and the average value was taken.

[0058] Comparative experimental results:

[0059] Results analysis: The comparative experimental results show that the method of this invention achieves an overall accuracy of 96.7%, which is 22.4 percentage points higher than the traditional static thresholding method (74.3%) and 8.0 percentage points higher than the CNN-BiLSTM method (88.7%). In terms of single-class defect detection performance, the traditional static thresholding method has significantly low precision and recall for all four defect classes (all below 65%), failing to effectively distinguish between different types of locking defects. The CNN-BiLSTM method shows a significant improvement in overall accuracy, but its precision and recall for rare defect categories are slightly lower than the overall level, indicating the impact of sample imbalance. The method of this invention achieves precision and recall exceeding 92% for all defect categories, demonstrating good consistency in detection performance between rare and common defects.

[0060] In terms of evaluation time, the traditional static threshold method is the fastest (0.15 seconds), but its detection capability is limited; the CNN-BiLSTM method takes the longest (2.34 seconds), which is difficult to meet the real-time requirements of high-frequency assembly scenarios with a 30-second cycle; the method of this invention takes 0.86 seconds, which takes into account both high detection accuracy and good real-time performance.

[0061] Regarding training sample requirements, the CNN-BiLSTM method requires a large amount of labeled data (≥1000 items), while the method of this invention only requires about 50 labeled samples per type of defect to obtain good classification performance, which greatly reduces the requirement for abnormal sample data.

[0062] Example 3: Verification of the effect in actual production application In this embodiment, the method of the present invention was tested on an actual coffee machine assembly line for a period of 3 months to verify its long-term stability and practical effect in a real production environment.

[0063] Trial operation conditions: Trial operation stations: the three stations with the highest screw usage in the production line (screw locking station for funnel components, screw locking station for heating pots, and screw locking station for base).

[0064] Trial operation period: 3 months (approximately 90 production days in total).

[0065] Total number of screws inspected: approximately 82,500.

[0066] Data recording method: The system automatically records the torque curve, characteristic parameters, classification results and timestamp of each screw.

[0067] Trial run results statistics:

[0068] Typical application cases: During the second month of trial operation, the system detected eight consecutive "slipped thread" defect alarms at the locking and heating cooker station. The torque curves automatically recorded by the system showed that the standard deviation of the torque during the screw-in section of these eight locking events all exceeded 0.4 N·m, while the average slope of the screw-in section was less than 0.3 N·m / s, consistent with typical characteristics of slipped threads. The system recommended checking the screw feeder after triggering the alarm.

[0069] After inspecting the screw feeder as instructed, on-site process personnel discovered that the feeder's vibratory feeding device had suffered track wear due to long-term use, resulting in thread scratches on some screws during transport. After replacing the feeder track, the frequency of stripped screw defects decreased from 8 times / day to 0 times / day. Before the application of this invention, such stripped screw problems were often misdiagnosed as "improper electric screwdriver parameter settings" or "operator technique issues," causing process personnel to waste significant time troubleshooting in the wrong direction. This invention, by providing precise defect type classification, reduces problem localization time from an average of 45 minutes to 5 minutes.

[0070] In addition, the incremental update mechanism of the model conducted five incremental training sessions during the trial operation (approximately 1-2 times per month). After each incremental training session was completed offline on the server, the updated model was pushed to the computers at each workstation. The incremental training enabled the model to effectively adapt to the torque curve drift caused by seasonal temperature fluctuations in the workshop. The overall accuracy of the model remained stable at over 95% during the three-month trial operation.

[0071] Example 4: Parameter Configuration and Model Transfer for Different Product Lines This embodiment demonstrates the parameter configuration and model transfer capabilities of the method of the present invention across different product lines.

[0072] Migration Scenario: After deploying the model for the coffee machine production line (Product A), the method is migrated to the microwave oven production line (Product B) in the same factory. The screw specifications and fastening process parameters differ between the two product lines: the coffee machine uses M3.0×6 stainless steel self-tapping screws with a target fastening torque of 4.5~5.5 N·m; the microwave oven uses M3.5×8 stainless steel self-tapping screws with a target fastening torque of 8.5~10.5 N·m.

[0073] Migration Steps: First, retain the four-stage feature decomposition framework and the mathematical definitions of each stage's features unchanged, only recalibrating the acceptable range of each stage's feature parameters based on the technological characteristics of the microwave oven product. Second, reset the judgment logic thresholds for each defect type according to the technological requirements of the microwave oven product. Third, collect approximately 200 samples from the microwave oven production line (150 qualified + 50 defective), fine-tune the classifier parameters of the weighted SVM model, retaining the basic model parameters already trained on the coffee machine product as initial values, and only using the microwave oven samples for a small number of iterative updates.

[0074] Transfer Results: After completing the above three steps (total time approximately 2 person-days), offline testing was conducted on the microwave oven production line. The test set contained 100 samples (70 qualified samples + 30 samples with various faults). The overall accuracy of the method of this invention reached 94.8%, which is higher than that of directly using the CNN-BiLSTM method on the production line (accuracy after transfer is only 83.2%) or the traditional static thresholding method (accuracy 71.5%).

[0075] Example 5: Suspicious Sample Handling and Model Incremental Update Mechanism This embodiment details the handling mechanism for suspicious samples and the incremental update process for the model.

[0076] The identification mechanism for suspicious samples: When the difference between the number of votes of the highest-voted category and the second-highest-voted category in the classification voting results of a sample by the W-SVM model is less than a preset threshold (set to 3 votes in this embodiment), the system determines that the sample is a "suspicious" sample.

[0077] During the trial run (82,500 screws), a total of 412 samples were marked as "suspicious" by the system, accounting for 0.50%.

[0078] Manual re-inspection and labeling process: Suspicious samples are automatically exported to the labeling pool after the end of production for the day. Quality inspectors manually re-inspect each suspicious sample, and the correct label is entered into the system after confirmation.

[0079] Model incremental update process: An incremental model update is performed every two weeks, adding newly labeled samples to the training set and retraining the W-SVM model. The classification accuracy of the updated model is tested on the retained validation set (approximately 200 independent samples), and deployment is only permitted when the accuracy is at least 98% of the original model. During the trial run, a total of six incremental model updates were performed, and the model's accuracy on the validation set gradually increased from 96.7% to 97.6%. The proportion of "suspicious" samples gradually decreased from 0.67% at the beginning of the trial run to 0.41% after the sixth update, indicating a gradual improvement in the model's classification confidence.

[0080] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for online assessment and classification of screw fastening quality based on real-time torque curve feature recognition, characterized in that, Includes the following steps: Step S1: Real-time acquisition and preprocessing of torque curves—Integrate a torque sensor and an angle encoder on the tightening shaft of the screw-locking device to acquire the torque value T(t) and angle value θ(t) during the tightening process in real time at a preset sampling frequency, and filter the acquired raw torque data to obtain smoothed torque-time curve T(t) and torque-angle curve T(θ). Step S2: Four-stage feature decomposition of torque curve—The preprocessed torque-time curve T(t) is divided into four feature stages with different physical meanings: starting stage, screwing in stage, fitting stage and locking stage. Key feature parameters reflecting the changes in the mechanical state of the locking are extracted from each stage to form a multi-dimensional feature vector F. Step S3: Construction and dimensionality reduction of multidimensional feature vectors—Quantize and calibrate the key feature parameters, construct multidimensional feature vectors F, and perform standardization and principal component analysis to reduce the dimensionality of the multidimensional feature vectors; Step S4: Construction of a defect classification model based on an improved weighted support vector machine—A defect classification model is constructed using an improved weighted support vector machine, with the dimensionality-reduced feature vectors as the model input and multiple defect types as the model output; where, for each defect category k, the category weight w_k = 1 / n_k × α is defined, where n_k is the number of training samples for the k-th defect category, and α is a global adjustment factor used to assign higher weights to scarce defect categories when the training samples are imbalanced; Step S5: Online evaluation and classification—After the newly acquired locking torque curve undergoes feature decomposition in step S2 and feature extraction and dimensionality reduction in step S3, it is input into the defect classification model for classification, and the defect type determination result is output.

2. The method according to claim 1, characterized in that, In step S2: The starting segment begins from the start of the electric screwdriver and ends when the screw head first contacts the workpiece surface. The peak starting torque T_start_peak and the starting segment time Δt_start are extracted as feature parameters. The screw-in section extends from when the screw head contacts the workpiece surface to when the screw head is about to adhere to the workpiece surface. The average slope k_run and the screw-in time Δt_run of the screw-in section are extracted as feature parameters. The bonding segment begins when the screw head contacts the workpiece surface and bonding is completed, marking the turning point where the torque begins to accelerate. The torque value T_seating at the bonding point and the bonding stage duration Δt_seating are extracted as feature parameters. The bonding point is identified using the sliding window change rate ratio method: a front sliding window with a first preset duration is established centered on the current torque data point, and a back sliding window is established with historical torque data for a second preset duration before this point. The torque change rate R_front of the front sliding window and the torque change rate R_rear of the back sliding window are calculated respectively. When R_front / R_rear≥3.0, the point is marked as the initial bonding point, and the bonding point position is corrected according to subsequent real-time data. The locking section stops from the point of complete bonding until the electric screwdriver reaches the target torque value. The peak torque T_peak, the time to reach the peak torque Δt_peak, and the torque rise slope k_tight of the locking section are extracted as feature parameters. After the electric screwdriver reaches the target torque value, it enters the pressure holding stage. The torque attenuation rate δ_T_hold during the pressure holding stage is further extracted as a supplementary feature. At the same time, the total rotation angle Δθ_total of the entire locking process is recorded.

3. The method according to claim 2, characterized in that, The multidimensional feature vector F is represented as: F=[T_start_peak,Δt_start,k_run,Δt_run,T_seating,Δt_seating,T_peak,Δt_peak,k_tight,δ_T_hold,Δθ_total].

4. The method according to claim 1, characterized in that, In step S4, the improved weighted support vector machine is used to construct the defect classification model, specifically including: Using the dimensionality-reduced feature vectors from step S3 as model input and pre-labeled multiple defect types as model output, a multi-classification training dataset is constructed. A one-to-one strategy is used to construct a multi-classification decision framework. A binary SVM sub-classifier is trained between every two defect categories, and a total of C(N,2) sub-classifiers are trained, where N is the total number of defect categories. The kernel function of each sub-classifier adopts a radial basis function, and the kernel parameter γ and the penalty factor C are optimized by grid search combined with cross-validation. When classifying new samples, the voting results of all subclassifiers are statistically analyzed, and the defect category with the most votes is output as the final classification result.

5. The method according to claim 1, characterized in that, In step S4, a torque curve feature map and mathematical judgment logic corresponding to at least one of the following typical defects are pre-established: The slippage defect is characterized by a torque curve that shows continuous fluctuations in torque during the screw-in segment and an inability to rise steadily, with the peak torque failing to reach the set value. The mathematical judgment logic is: the standard deviation of the torque during the screw-in segment σ_T_run>0.3N·m and the average slope of the screw-in segment k_run<0.5N·m / s. The false lock defect is characterized by a torque curve that reaches the target torque value too early, followed by a rapid decrease in torque during the holding phase after the peak torque. The mathematical judgment logic is: the time to reach the peak torque Δt_peak < 60% of the historical average of this station, and the attenuation rate during the holding phase δ_T_hold > 15%. The screw-hole mismatch defect is characterized by an abnormally high peak torque in the starting phase, or a sudden drop in torque after a sharp increase in the screw-in phase. The mathematical judgment logic is: the peak torque T_start_peak > 30% of the target torque, and the screw-in phase has a negative slope range. If there is a foreign object defect inside the hole, its torque curve will show one or more torque spikes of unknown cause in the screw-in section; the mathematical judgment logic is: the duration of the torque change rate dT / dt in the screw-in section is greater than 0.8 N·m / ms and the number of occurrences is ≥1.

6. The method according to claim 4, characterized in that, It also includes a suspicious sample handling mechanism: when the difference between the highest and second highest votes in the classification voting results of the defect classification model is less than a preset threshold, the sample is marked as a suspicious sample, and after manual review and confirmation, it is added to the training set to incrementally train and update the defect classification model.

7. The method according to claim 1, characterized in that, In step S3, the standardization process is used to eliminate the scale differences in the dimensions between different feature parameters; the dimensionality reduction process uses principal component analysis to retain principal components with a cumulative variance contribution rate of more than 95% to reduce feature redundancy.

8. The method according to claim 1, characterized in that, It also includes a real-time feedback step: the quality classification results of each product are recorded and bound to the product serial number to form a traceable quality file; when the number of consecutive occurrences of the same type of defect exceeds the preset threshold, the system automatically issues an early warning signal and displays the specific defect type on the workstation display screen, while providing process adjustment suggestions.

9. A screw fastening quality online assessment and classification system based on real-time torque curve feature recognition, characterized in that, include: The torque data acquisition module is used to integrate a torque sensor and an angle encoder on the tightening shaft of the screw-locking device to acquire torque and angle values ​​in real time during the screw-locking process. The curve processing and feature extraction module is used to filter the collected torque data and divide the torque-time curve into four characteristic stages: the starting stage, the screwing-in stage, the fitting stage, and the locking stage. Key feature parameters are extracted for each stage to construct a multi-dimensional feature vector. The feature dimensionality reduction module is used to perform standardization and dimensionality reduction processing on the multidimensional feature vectors. The defect classification model module is based on an improved weighted support vector machine. It takes the dimensionality-reduced feature vector as input and outputs the defect type determination result. as well as The results display and feedback module is used to display the classification results and issue an early warning signal when the number of consecutive occurrences of the same type of defect exceeds a preset threshold.

Images