Torque and angle tightening curve anomaly judgment method based on multi-modal anomaly model

Abstract

This invention relates to a method for judging anomalies in torque-angle tightening curves based on a multimodal anomaly model, belonging to the technical field of anomaly judgment methods. The judgment method includes the following steps: S1: Standardizing the torque-angle tightening curve data to be analyzed and enhancing the features of key curve segments through local magnification; S2: Constructing a multimodal knowledge base to store normal and abnormal sample images, quantitative descriptions of geometric features, and anomaly judgment criteria for each type of curve; S3: Inputting the preprocessed curve to be analyzed into a large multimodal anomaly judgment model to first determine whether the curve is abnormal. If it is abnormal, the specific anomaly category is further determined through a multimodal anomaly classification model. This application solves the problems of low accuracy and weak generalization ability of traditional single-modal detection methods through a two-level reasoning logic of pre-judgment of whether it is abnormal and fine classification of anomaly types. It is applicable to high-precision tightening processes in automobile manufacturing and mechanical assembly.

Description

Technical Field

[0001] This invention relates to the field of anomaly judgment methods, and in particular to an anomaly judgment method for torque and angle tightening curves based on a multimodal anomaly model. Background Technology

[0002] In the field of mechanical assembly, the torque-angle tightening curve directly reflects the quality of threaded connections. Traditional anomaly detection methods rely on manual experience or single sensor data, which suffers from high subjectivity and a high rate of missed detections. For example, during the tightening of cylinder head bolts in an automotive engine, if the torque-angle curve shows anomalies such as "torque climb delay" or "angle overshoot," it may lead to bolt breakage or loosening of the connection, causing safety hazards. This invention combines image visual features with text rule knowledge through multimodal fusion technology to achieve intelligent and high-precision anomaly judgment. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, the present invention provides a method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model, in order to solve the problems mentioned in the background art.

[0004] To achieve the above-mentioned objective, this invention provides a method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model. The judgment method includes the following steps: S1: Standardize the torque-angle tightening curve data to be analyzed, and enhance the features of key curve segments by local magnification; S2: Construct a multimodal knowledge base to store normal and abnormal sample images, quantitative descriptions of geometric features, and abnormal judgment criteria for each type of curve; S3: Input the preprocessed curve to be analyzed into the multimodal anomaly judgment model. First, determine whether the curve is abnormal. If it is abnormal, then further determine the specific anomaly category through the multimodal anomaly classification model.

[0005] Furthermore, in step S1, the local magnification includes: magnifying the preset interest area of ​​the curve by 5 to 10 times at the pixel level while preserving the coordinate ratio of the original curve. The preset interest area includes: torque mutation point and angle saturation segment.

[0006] Furthermore, in step S2, the multimodal knowledge base includes: For each type of curve, there are at least three magnified normal sample images and corresponding geometric feature parameters, wherein the geometric feature parameters include: slope, peak torque, and angle range; Sample images, textual descriptions of abnormal features, and judgment thresholds for each type of anomaly, wherein the textual descriptions of abnormal features include: torque not reaching the threshold, and angle sudden drop.

[0007] Furthermore, in step S3, the multimodal anomaly judgment model adopts a dual-branch structure: an image branch and a text branch. The image branch extracts local magnified features of the curve, and the text branch analyzes the quantitative description of the geometric features. After fusion, an anomaly probability value is output. When the probability value exceeds a preset threshold, it is judged as an anomaly.

[0008] Furthermore, the anomaly probability value output after the image branch and the text branch data are fused is between 0 and 1, and the preset threshold is 0.85.

[0009] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves high-precision anomaly identification and classification of tightening curves through image processing to enhance key features, construct a multimodal knowledge base, and implement a phased anomaly detection process. This method integrates image geometric features and textual descriptions, employing a two-level reasoning logic of pre-judgment of anomaly status followed by detailed anomaly type classification. This addresses the low accuracy and weak generalization ability of traditional single-modal detection methods, making it suitable for high-precision tightening processes in automotive manufacturing and mechanical assembly. Attached Figure Description

[0010] Figure 1 This is a flowchart of an embodiment of the present invention; Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the invention, but not all embodiments. The embodiments of the present invention are described below with reference to the accompanying drawings.

[0012] refer to Figure 1 A method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model includes the following steps: S1: Data Preprocessing and Image Processing Standardization process: Convert the original torque-angle curve data into a unified coordinate system, such as converting the torque unit to N·m and the angle unit to degrees, to eliminate the problem of inconsistent dimensions caused by equipment differences; Specifically, torque (N·m, accuracy ±0.5%) and angle (degrees, accuracy ±1°) data can be collected in real time using a tightening shaft sensor. The sampling frequency is set to 1kHz, and 1000-2000 data points are collected in a single tightening process to form a two-dimensional curve dataset including timestamps. Z-score standardization is used to eliminate dimensional differences.

[0013] Local Magnification Enhancement: An adaptive region magnification algorithm is used to locally crop and magnify feature points (such as torque peaks and angle plateaus) in the curve. The magnification factor is dynamically adjusted according to the curve curvature; the greater the curvature, the higher the magnification factor, highlighting geometric features such as slope changes and inflection point distribution. Adaptive Magnification Implementation: The marked area is locally cropped, and a bilinear interpolation algorithm is used for pixel-level magnification.

[0014] S2: Construction of a Multimodal Knowledge Base The knowledge base architecture consists of a three-tier structure: data layer, feature layer, and rule layer.

[0015] Data layer: Stores raw tightening curve data, normal and abnormal sample images. Each abnormal type contains at least one typical sample, and the number of normal curve samples is greater than or equal to 3. Feature layer: Extracts the geometric features of the curve through edge detection and Hough transform algorithm, such as the coordinates of extreme points, curve length, and number of inflection points, and converts them into quantitative descriptive text, such as: "Peak torque 25 N·m, corresponding angle 180°". Rule layer: Define the abnormal judgment logic, such as "if the peak torque is 20% lower than the set threshold, it is abnormal" and "if the angle does not reach the preset range, it is judged as under-tightening".

[0016] Specifically as follows: Data layer: Stores raw data based on a relational database (such as MySQL), including: Normal sample set: For each type of tightening process (such as M10 bolts, aluminum alloy materials), store the original data of ≥3 normal curves, local magnified images (resolution 1024×768 pixels, PNG format) and corresponding process parameters (such as target torque, tightening speed). Abnormal sample set: Stored according to the abnormality type (such as "insufficient torque", "angle exceeding limit", "curve fluctuation"). Each type contains ≥5 samples. Each sample is associated with: magnified image, abnormal feature parameters (such as peak torque deviation rate, angle fluctuation amplitude), and process conditions in which the abnormality occurred (such as ambient temperature, bolt batch).

[0017] Feature layer: Features are extracted using OpenCV and Python natural language processing libraries, including: Image features: Canny edge detection is used to extract curve contours, and Hough transform is used to fit the equations of straight lines / curves to calculate key parameters; Text features: Convert image feature parameters into structured descriptive text in the format of "feature type: parameter value ± error range, physical meaning", for example, "peak torque: 25.3 N·m ± 0.2 N·m, located at an angle of 180° ± 5°, corresponding to the plastic deformation stage of the tightening stage".

[0018] Rule layer: Anomaly judgment logic is represented based on production rules, with the format "IF [condition 1] AND [condition 2] THEN [anomaly type], confidence level: [score]", for example: "IF peak torque < 0.8 × target torque AND angle > 1.2 × target angle THEN Anomaly type = thread slippage, confidence level = 0.95" "IF torque curve fluctuation amplitude > 5% × Tmax AND fluctuation frequency > 10Hz THEN anomaly type = sensor failure, confidence level = 0.90".

[0019] S3: Multimodal Anomaly Detection Model Model structure: Employs a large model architecture with a dual-base base. Anomaly Prediction Model: Input a magnified local image of the curve to be analyzed and geometric feature text. Extract features through an image encoder (such as ResNet) and a text encoder (such as BERT), fuse them, and output the anomaly probability through a fully connected layer. When the probability is ≥0.85, trigger the anomaly classification process. Anomaly classification model: Based on the multimodal features (image + text) of anomaly curves, the model matches the sample to be analyzed with the anomaly type template in the knowledge base through contrastive learning, and outputs the Top-1 anomaly category, such as "overtightening", "slippage", and "sensor failure".

[0020] Specifically, a dual-branch fusion architecture is adopted, as follows: Image branches: Input: A magnified version of the curve image; Feature extraction: A pre-trained ResNet-50 model (with the last fully connected layer removed) is used to output a 2048-dimensional image feature vector; Feature enhancement: The self-attention mechanism is used to highlight the curve contour area and suppress background noise.

[0021] Text branches: Input: Structured descriptive text generated by the feature layer, with a text length of less than or equal to 512 characters; Feature extraction: The BERT-base model (pre-trained in Chinese) is used to output a 768-dimensional text feature vector; Rule integration: The anomaly judgment conditions of the rule layer are transformed into embedded vectors and then fused with the text feature vectors in a weighted manner (the weights are dynamically adjusted according to the rule confidence).

[0022] Feature fusion and output: Image feature vectors and text feature vectors are fused by element-level multiplication. The fused vectors are then input into a 3-layer fully connected network, which outputs anomaly probability P (0≤P≤1). When P≥0.85 (which can be adjusted according to process requirements), the vectors are considered anomalies.

[0023] Anomaly classification model training: Training dataset: Annotated samples from the knowledge base are used, with a normal sample to abnormal sample ratio of 1:1 (classes are balanced by oversampling), and the total number of samples is greater than or equal to 10,000; Loss function: Focal Loss is used to address the class imbalance problem; Optimizer and training strategy: Adam optimizer (learning rate 1e-4, weight decay 1e-5), early stopping (patience=5) is used to prevent overfitting. The model is trained on NVIDIA A100 GPU, with about 200 iterations. After convergence, the accuracy on the validation set is ≥97%.

[0024] Application scenarios and advantages: Scenario: High-precision assembly scenarios such as tightening cylinder heads for automobile engines and connecting bolts for wind power equipment; Advantages: Multimodal fusion improves judgment robustness, with a 23% improvement in accuracy compared to single-modal image detection; secondary reasoning logic reduces false positives, with anomaly classification precision@1 ≥ 95%; the knowledge base supports dynamic updates, and the coverage of anomaly types can be expanded by adding new samples.

[0025] The technical solution of the present invention has been described above in conjunction with specific embodiments. However, it should be noted that the above descriptions are only for explaining the solution of the present invention and should not be construed as a specific limitation on the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can conceive of other specific embodiments or equivalent substitutions of the present invention without creative effort, and all such embodiments or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model, characterized in that, The determination method includes the following steps: S1: Standardize the torque-angle tightening curve data to be analyzed, and enhance the features of key curve segments by local magnification; S2: Construct a multimodal knowledge base to store normal and abnormal sample images, quantitative descriptions of geometric features, and abnormal judgment criteria for each type of curve; S3: Input the preprocessed curve to be analyzed into the multimodal anomaly judgment model. First, determine whether the curve is abnormal. If it is abnormal, then further determine the specific anomaly category through the multimodal anomaly classification model.

2. The method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model according to claim 1, characterized in that, In step S1, the local magnification includes: magnifying the preset interest area of ​​the curve by 5 to 10 times at the pixel level while preserving the coordinate ratio of the original curve. The preset interest area includes: torque abrupt change point and angle saturation segment.

3. The method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model according to claim 1, characterized in that, In step S2, the multimodal knowledge base includes: For each type of curve, there are at least three magnified normal sample images and corresponding geometric feature parameters, wherein the geometric feature parameters include: slope, peak torque, and angle range; Sample images, textual descriptions of abnormal features, and judgment thresholds for each type of anomaly, wherein the textual descriptions of abnormal features include: torque not reaching the threshold, and angle sudden drop.

4. The method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model according to claim 1, characterized in that, In step S3, the multimodal anomaly judgment model adopts a dual-branch structure: an image branch and a text branch. The image branch extracts the local magnified features of the curve, and the text branch analyzes the quantitative description of the geometric features. After fusion, an anomaly probability value is output. When the probability value exceeds a preset threshold, it is judged as an anomaly.

5. The method for judging anomalies in torque and angle tightening curves based on a multimodal anomaly model according to claim 4, characterized in that, The anomaly probability value output after the data fusion of the image branch and the text branch is between 0 and 1, and the preset threshold is 0.85.

Images