Deep learning-based adaptive intelligent assembly device for mechanical structural parts

Abstract

The application discloses a kind of based on deep learning's mechanical structural piece adaptive intelligent assembly device, belong to intelligent manufacturing field.The device is through the cooperation of each component, realize the automatic assembly of different types of mechanical structural piece, and utilize deep learning algorithm according to real-time data continuously optimization assembly strategy.Device mainly includes support platform, material feeding mechanism, electric wrench, sensor group and controller.Support platform is used to stabilize support assembly matrix;Material feeding mechanism is responsible for pushing mechanical structural piece to assembly position;Sensor group acquires assembly process data;Controller then according to the analysis result of integrated deep learning model, adaptively controls operating arm to complete accurate assembly.The application endows device self-learning ability, so that it can adapt to different assembly tasks, effectively improve assembly efficiency and quality, with wide applicability and flexibility.

Description

Technical Field

[0001] This invention relates to an adaptive intelligent assembly device for mechanical structural components based on deep learning, belonging to the field of intelligent manufacturing. Background Technology

[0002] In modern manufacturing, the assembly of mechanical structural components is a core link in the production process, and its quality and efficiency directly affect the overall performance of the product and production benefits. Traditional assembly methods mainly rely on manual operation or fixed automated equipment, which have many limitations. Manual assembly is not only inefficient but also susceptible to human factors, leading to insufficient assembly precision and consequently affecting product stability and lifespan. While fixed automated equipment improves assembly efficiency to some extent, its lack of flexibility limits its application to specific types of mechanical structural components. It struggles to achieve batch, automated assembly of components with varying shapes or non-standardized designs. Furthermore, existing equipment cannot monitor equipment status and assembly quality in real time during the assembly process. Faults or assembly errors are often difficult to detect promptly, increasing maintenance costs and time, and severely impacting production schedules.

[0003] While some auxiliary equipment for assembling mechanical structural components exists on the market, these devices mostly suffer from drawbacks such as large size, significant noise, slow installation speed, and limited applicability. With the continuous development of intelligent manufacturing technology, deep learning algorithms have demonstrated powerful capabilities in image recognition and data analysis, providing new solutions to these problems. Deep learning algorithms can automatically extract features and identify patterns and rules from large amounts of data through learning and analysis, thereby achieving intelligent control and optimization of complex systems. However, there is currently no mature solution that organically combines deep learning technology with assembly equipment to achieve adaptive optimization of assembly strategies. Most existing assembly equipment simply applies automation technology, lacking intelligent self-learning and adaptive capabilities, and cannot automatically adjust assembly strategies according to different assembly tasks to meet diverse assembly needs.

[0004] Therefore, it is particularly urgent to develop an intelligent assembly and adaptive strategy optimization device capable of adapting to various mechanical structural components of different shapes and specifications. This device should possess rapid and accurate assembly capabilities, while simultaneously monitoring equipment status and assembly quality in real time. It should also utilize deep learning algorithms to achieve adaptive optimization of assembly strategies, thereby meeting the demands of modern manufacturing for efficient, precise, and intelligent production. Summary of the Invention

[0005] This invention provides a deep learning-based adaptive intelligent assembly device for mechanical structural components. This device is compact yet powerful, capable of adapting to various shapes and specifications of mechanical structural components, achieving precise positioning and assembly of different assembly substrates in different assembly tasks. Through the collaborative work of vibration sensors, vision sensors, and other sensors, the device can monitor the equipment status and assembly quality in real time during the assembly process, ensuring that each mechanical structural component is installed correctly. Simultaneously, by analyzing and processing the large amount of data collected by the sensors using deep learning algorithms, it can automatically extract features, identify patterns, and dynamically adjust assembly strategies according to different assembly tasks, achieving intelligent assembly operations. This self-learning and adaptive capability allows the device to continuously optimize the assembly process, improve assembly efficiency and quality, and reduce equipment maintenance costs. Furthermore, the device is easy to operate and structurally flexible, making it easy to integrate into existing production lines, significantly improving the intelligence level and production efficiency of mechanical structural component assembly, and providing a highly efficient, precise, and intelligent assembly solution for modern manufacturing.

[0006] The technical solution of this invention is: an adaptive intelligent assembly device for mechanical structural parts based on deep learning, comprising a support platform 1, a material feeding mechanism 2, an electric wrench 3, a sensor group, and a controller; the support platform 1 is fixed on a fixed bracket 23 for stably supporting the assembly base to be installed; the material feeding mechanism 2 is installed on both sides of the support platform 1, and the feeding cylinder 6 and the top-feeding cylinder 7 in the material feeding mechanism 2 are respectively used to push the mechanical structural part 20 from the hopper 8 to the top-feeding position 17 and push it into the assembly position 19 of the assembly base; the electric wrench 3 is fixed on a wrench holder 21. The sensor group, including vibration sensor 4-1, vision sensor 4-2, torque sensor 4-3, and displacement sensor 4-4, is used to collect data during the assembly process. The controller is placed in the control cabinet 5 near the operating area and is used to receive and process the sensor data from the sensor group. Its built-in deep learning model evaluates the health status of the assembly process and the assembly quality of the output parts in real time based on the evaluation results. The controller then dynamically and adaptively adjusts the action parameters of the material feeding mechanism 2 and the electric wrench 3 based on the evaluation results to achieve precise assembly and strategy optimization.

[0007] Specifically, the support platform 1 is used to support the track to be installed, and grooves are made on both sides of it so that they can be embedded into the assembly base. The main purpose of this design is to facilitate the installation and removal of the assembly base.

[0008] Specifically, the material feeding mechanism 2 includes a feeding cylinder 6, a top feeding cylinder 7, and a hopper 8. The feeding cylinder 6 and the hopper 8 are both placed on a fixed support 23. The hopper 8 can accommodate multiple mechanical structural parts 20. It has a feeding channel 16 at its bottom. The piston rod 14 of the feeding cylinder 6 can reciprocate linearly in the feeding channel 16. The top feeding cylinder 7 is fixed below the fixed support 23. There is a top feeding channel 18 on one side of the feeding channel 16. There is an assembly position 19 on the side of the top feeding channel 18 away from the top feeding cylinder 7. When the assembly base comes to the working area, the piston rod 14 of the feeding cylinder 6 will push the mechanical structural parts 20 in the hopper 8 to the top feeding position 17 through the feeding channel 16. Then, the piston rod of the top feeding cylinder 7 will push them into the assembly position 19 of the assembly base. This design, by separating the feeding and ejection functions into different cylinders, achieves a more precise and controllable conveying process for the mechanical components 20. This not only improves the stability and reliability of the feeding process, but also enhances the durability and ease of maintenance of the entire mechanism by reducing reliance on a single cylinder.

[0009] Specifically, the feeding cylinder 6 includes a cylinder head 9, a buffer seal 10, an adjustable buffer screw 11, a cylinder body 12, a piston 13, a piston rod 14, vent holes 15, a feeding channel 16, and a pneumatic valve 24. The cylinder head 9 seals the cylinder chambers at both ends of the cylinder body 12. The buffer seals 10 are each installed inside the cylinder head 9. The adjustable buffer screw 11 is installed on the cylinder head 9, allowing for optimal buffering at the end of the stroke. The bottom of the cylinder body 12 is located on a fixed bracket 23. The piston 13 and piston rod 14 are connected inside the cylinder body 12. The piston rod 14 can transmit the force of compressed air. The two vent holes 15 are connected by screws. Fixed to the cylinder head 9 and connected to ports A and B of the pneumatic valve 24, port P is used to input compressed air, and ports R and S of the pneumatic valve 24 are exhaust ports; compressed air is input through port P of the pneumatic valve (24), the air is guided to port A in the pneumatic valve 24, and then delivered to the cylinder head 9 and cylinder block 12 through the vent hole 15. The gas generates high pressure at the end of the cylinder head 9, thereby pushing the piston 13, thereby causing the piston rod 14 to push the mechanical structure 20 into the top position 17 in the feeding channel 16; the exhaust gas generated during the stroke will reach port B of the pneumatic valve 24, and the valve guides this gas to the S end and allows it to be discharged. When the compressed air is guided to the cylinder through the B end, the piston 13 is pushed back, and the excess gas escapes from the R port, thus constituting the reciprocating motion of the piston 13, thereby pushing the piston rod 14 to feed. By strategically arranging and switching the ports P, A, B, R, and S of the pneumatic valve 24, the flow direction and volume of compressed air can be precisely controlled, thereby achieving the reciprocating motion of the piston 13. This control method is simple, reliable, and has a fast response speed, meeting the needs of rapid material feeding. The adjustable buffer screw 11 installed on the cylinder head effectively reduces the impact force on the piston 13 at the end of its stroke, extending the cylinder's service life while reducing noise and vibration. Using compressed air as a power source is clean, pollution-free, readily available, and easy to control.

[0010] Specifically, the electric wrench 3 is connected to the spring 22 and is used to perform assembly operations. Different tools can be installed at its end to adapt to different assembly tasks. The other end of the spring 22 is fixed to the wrench holder 21. The wrench holder 21 is a sturdy metal frame fixed on the fixed bracket 23. Its main function is to provide a stable support for the electric wrench 3.

[0011] Specifically, the sensor group includes a vibration sensor 4-1, a vision sensor 4-2, a torque sensor 4-3, and a displacement sensor 4-4; the vibration sensor 4-1 is installed on the material feeding mechanism 2 and is used to monitor the vibration of the mechanical structural component 20 during the feeding, ejection, and assembly processes. Vibration signals can reflect whether the mechanical component 20 has correctly entered the assembly position and whether there is abnormal vibration during the assembly process; vision sensor 4-2 is installed on the tool head and wrench holder 21 connected to the electric wrench 3 and facing the assembly base support platform 1, and is used to capture the installation position, assembly quality and appearance characteristics of the mechanical component 20; torque sensor 4-3 is installed on the end tool of the electric wrench 3 and is used to detect whether the torque applied during the assembly process meets the requirements. The torque data can ensure that bolts or other connecting parts are tightened correctly and avoid assembly quality problems caused by insufficient or excessive torque; displacement sensor 4-4 is installed on the piston rod of the feeding cylinder 6 and the ejecting cylinder 7 and is used to monitor the movement position of the mechanical component 20. The displacement data can ensure that the mechanical component 20 is accurately pushed to the assembly position; the sensors are connected to the data acquisition card through the CAN bus.

[0012] Specifically, the controller is located in control cabinet 5 near the operating area. It receives data from the sensor array and uses its integrated deep learning model to fuse multi-sensor data for the following decisions: First, it determines the accuracy of the positioning of the mechanical component 20 based on data from vision sensor 4-2 and a CNN model. Simultaneously, data from displacement sensor 4-4 confirms whether the loading and unloading actions are in place. Next, during the tightening process of the electric wrench 3, it integrates data from vibration sensor 4-1 and torque sensor 4-3, using an LSTM model to analyze the vibration characteristics and torque curve of the assembly process to predict the tightening quality. Finally, the model outputs a comprehensive score S, which determines whether the assembly is qualified. If it is unqualified, the controller will adaptively adjust strategies such as the operating arm posture or wrench torque to correct the deviation.

[0013] The beneficial effects of this invention are:

[0014] This invention utilizes some standard components, resulting in low overall manufacturing costs. The machine is compact, small in size, and lightweight, allowing for flexible adaptation to different spaces and environments. Employing a modular design, the entire machine includes a support platform, material feeding mechanism, operating arm, sensor array (vibration sensor, vision sensor, torque sensor, displacement sensor), controller, and actuator. Each module can be installed individually and then assembled, facilitating both manufacturing and maintenance / component replacement. The sensor array, controller, and actuator of this invention, through the following steps: sensor data acquisition → data preprocessing module → deep learning model (pre-trained) prediction → predicted value formatting → controller drive → actuator action, achieves real-time acquisition of assembly process data and real-time adjustment of assembly strategies, ensuring the accuracy and efficiency of the assembly process. The integrated design of the pneumatic collaborative feeding mechanism and intelligent control system of this invention enables automatic and precise assembly of mechanical structural components. The multi-sensor collaborative monitoring mechanism significantly improves the reliability of the assembly process, and the intelligent adaptive strategy optimizes and ensures assembly quality. The present invention has a compact structure and strong adaptability, and can be seamlessly integrated into existing production lines. It improves efficiency by more than 300% compared with traditional manual assembly, and the equipment cost is only 1 / 3 of that of automated special machines. Attached Figure Description

[0015] Figure 1 This is a flowchart of the invention;

[0016] Figure 2 This is an isometric drawing of the overall structure of the present invention;

[0017] Figure 3 This is a front view of the overall structure of the present invention;

[0018] Figure 4 These are the upper and lower isometric views of the overall structure of this invention;

[0019] Figure 5 This is a partial cross-sectional view of the material feeding mechanism of the present invention;

[0020] Figure 6 This is a side view of the support platform of the present invention;

[0021] Figure 7 This is the isometric view of the track in the assembly example provided in this invention;

[0022] Figure 8 This is an isometric view of the silo of the present invention;

[0023] Figure 9 This is the isometric view of the limiting block in the assembly example provided in this invention;

[0024] Figure 10 This is an isometric view of the feeding cylinder of the present invention;

[0025] Figure 11 This is a cross-sectional schematic diagram of the pneumatic valve of the present invention;

[0026] Figure 12 This is a front view of the wrench holder of the present invention;

[0027] Figure 13 This is an isometric view of the electric wrench of the present invention;

[0028] Figure 14 These are the left and right isometric views of the vibration sensor of this invention;

[0029] Figure 15 These are the left and right isometric views of the visual sensor of this invention;

[0030] Figure 16 These are the upper and lower isometric views of the torque sensor of this invention;

[0031] Figure 17 This is an isometric view of the displacement sensor of the present invention;

[0032] Figure 18 This is an isometric view of the controller of the present invention;

[0033] Figure 19 This is the original image simulated by the visual sensor of the present invention;

[0034] Figure 20 This invention provides a visual sensor that simulates a binarized image and a contour image.

[0035] Figure 21 This invention uses a visual sensor to simulate CNN feature maps;

[0036] The labels in the diagram are as follows: 1-Support platform, 2-Material feeding mechanism, 3-Electric wrench, 4-1 Vibration sensor, 4-2 Vision sensor, 4-3 Torque sensor, 4-4 Displacement sensor, 5-Control cabinet, 6-Feeding cylinder, 7-Top cylinder, 8-Hopper, 9-Cylinder cover, 10-Buffer seal, 11-Adjustable buffer screw, 12-Cylinder body, 13-Piston, 14-Piston rod, 15-Ventilation hole, 16-Feeding channel, 17-Top position, 18-Top channel, 19-Assembly base assembly position, 20-Mechanical structural component, 21-Wrench holder, 22-Spring, 23-Fixed bracket, 24-Pneumatic valve. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the present invention is not limited to the description.

[0038] Example 1: As Figures 1-21As shown, this invention provides an adaptive intelligent assembly device for mechanical structural components based on deep learning, including a support platform 1, a material feeding mechanism 2, an electric wrench 3, a sensor group, and a controller; the support platform 1 is fixed on a fixed bracket 23 to stably support the assembly base to be installed; the material feeding mechanism 2 is installed on both sides of the support platform 1, and the feeding cylinder 6 and the top-feeding cylinder 7 in the material feeding mechanism 2 are respectively used to push the mechanical structural component 20 from the hopper 8 to the top-feeding position 17 and push it into the assembly position 19 of the assembly base; the electric wrench 3 is fixed on a wrench holder 21. The sensor group, including vibration sensor 4-1, vision sensor 4-2, torque sensor 4-3, and displacement sensor 4-4, is used to collect data during the assembly process. The controller is placed in the control cabinet 5 near the operating area and is used to receive and process the sensor data from the sensor group. Its built-in deep learning model evaluates the health status of the assembly process and the assembly quality of the output parts in real time based on the evaluation results. The controller then dynamically and adaptively adjusts the action parameters of the material feeding mechanism 2 and the electric wrench 3 based on the evaluation results to achieve precise assembly and strategy optimization.

[0039] Furthermore, the support platform 1 can be set up and fixed on the fixed bracket 23. The support platform 1 has grooves on both sides that can be inserted into the assembly base. When working, the assembly base will be stuck and remain stationary.

[0040] Furthermore, the aforementioned material feeding mechanism 2 can be provided, which includes a feeding cylinder 6, a top-feeding cylinder 7, and a hopper 8. The feeding cylinder 6 and the hopper 8 are both placed above the fixed support 23, and the top-feeding cylinder 7 is fixed below the fixed support 23. The feeding cylinder 6 and the top-feeding cylinder 7 are perpendicular to each other. When the assembly base arrives at the working area, the piston rod 14 of the feeding cylinder 6 will push the mechanical structural component 20 in the hopper 8 to the top-feeding position 17 through the feeding channel 16. At this time, the top-feeding cylinder 7 will push it into the assembly position 19 of the assembly base through the movement of the piston rod.

[0041] Furthermore, the electric wrench 3 can be connected to the spring 22 for performing assembly operations. Different tools can be installed at its end to adapt to different assembly tasks. The other end of the spring 22 is fixed to the wrench holder 21 to prevent the electric wrench 3 from falling and also to provide shock absorption. The wrench holder 21 is a sturdy metal frame fixed to the fixed bracket 23, and its main function is to provide a stable support for the electric wrench 3.

[0042] Furthermore, the aforementioned sensor group can be configured, which includes a vibration sensor 4-1, a vision sensor 4-2, a torque sensor 4-3, and a displacement sensor 4-4; the vibration sensor 4-1 is installed on the material feeding mechanism 2 and is used to monitor the vibration of the mechanical structural component 20 during the feeding, ejection, and assembly processes. Vibration signals can reflect whether the mechanical component 20 has correctly entered the assembly position and whether there is abnormal vibration during the assembly process; vision sensor 4-2 is installed on the tool head and wrench holder 21 connected to the electric wrench 3 and facing the assembly base support platform 1, and is used to capture the installation position, assembly quality and appearance characteristics of the mechanical component 20; torque sensor 4-3 is installed on the end tool of the electric wrench 3 and is used to detect whether the torque applied during the assembly process meets the requirements. Torque data can ensure that bolts or other connecting parts are tightened correctly and avoid assembly quality problems caused by insufficient or excessive torque; displacement sensor 4-4 is installed on the piston rod of the feeding cylinder 6 and the ejecting cylinder 7 and is used to monitor the moving position of the mechanical component 20 and the action position of the operating arm. Displacement data can ensure that the mechanical component 20 is accurately pushed to the assembly position; the sensors are connected to the data acquisition card through the CAN bus.

[0043] Furthermore, the controller can be installed in a control cabinet 5 near the operating area to receive data from the sensor array and make the following decisions by fusing multi-sensor data through its integrated deep learning model: First, it determines whether the positioning of the mechanical structural component 20 is accurate based on the data from the visual sensor 4-2 and the CNN model; simultaneously, the data from the displacement sensor 4-4 is used to confirm whether the feeding and lifting actions are in place; then, during the tightening process of the electric wrench 3, it integrates the data from the vibration sensor 4-1 and the torque sensor 4-3, and uses an LSTM model to analyze the vibration characteristics and torque curve of the assembly process to predict the tightening quality; finally, the model outputs a comprehensive score S, and the assembly is judged to be qualified based on the S value. If it is unqualified, the controller will adaptively adjust the operating arm posture or wrench torque, etc., to correct the deviation.

[0044] Furthermore, the deep learning workflow described above:

[0045] Data acquisition. Vibration signal x(t) of mechanical structural component 20 during loading, unloading and assembly processes is collected by vibration sensor 4-1; image information I of installation position and assembly quality of mechanical structural component 20 is collected by vision sensor 4-2; torque data τ(t) of applied torque during assembly is collected by torque sensor 4-3; and displacement data d(t) of moving position of mechanical structural component 20 is collected by displacement sensor 4-4.

[0046] Data preprocessing. Vibration data: High-frequency noise is filtered out using a low-pass filter H(f).

[0047] xfiltered (t)=H(f)*x(t)

[0048] In the formula, H(f) is the frequency response function of the low-pass filter, used to filter out high-frequency noise. x(t) is the original vibration signal. filtered (t): The filtered vibration signal.

[0049] Fill in missing values ​​using linear interpolation:

[0050]

[0051] In the formula, x(t): the original vibration signal, x imputed (t): Vibration signal after filling in missing values.

[0052] Normalize the data to the range [-1, 1]:

[0053]

[0054] In the formula, μ: mean of vibration data, σ: standard deviation of vibration data, x normalized (t): Normalized vibration signal, x imputed (t): Vibration signal after filling in missing values.

[0055] Image data: Converting a color image to a grayscale image:

[0056] I gray =0.299R + 0.587G + 0.114B

[0057] In the formula, R, G, B: the red, green, and blue channels of the image. gray The image after grayscale conversion.

[0058] Convert a grayscale image to a binary image using a threshold T:

[0059]

[0060] In the formula, T: the threshold for binarization, I binary The binarized image, I gray The image after grayscale conversion.

[0061] Adjust the image to a fixed size (w, h):

[0062] I resized =resize(I binary ,(w,h))

[0063] In the formula, w,h: the adjusted image width and height, I resized The resized image, I binaryThe image after binarization.

[0064] Torque and displacement data: Normalize the torque and displacement data to the range [0,1].

[0065]

[0066] In the formula, τ(t): original torque data, τ min,max : Minimum and maximum values ​​of torque data, τ normalized (t): Standardized torque data. d(t): Original displacement data, d min,max : Minimum and maximum values ​​of displacement data, d normalized (t): Standardized displacement data.

[0067] Feature engineering. Vibration data: Extracting mean, variance, and peak value:

[0068]

[0069] peak x =max(|x normalized |)

[0070] In the formula, N: the number of data points, μ x : The mean of the vibration data, σ x Standard deviation of vibration data, peak x : Maximum amplitude of vibration data, x normalized (i): The normalized vibration value of the i-th sampling point, |x normalized |: Take the absolute value of the entire normalized sequence.

[0071] Use Fourier transform to convert the vibration signal from the time domain to the frequency domain:

[0072]

[0073] Extracting the main frequency component f main In the formula, X(f): frequency domain signal, x normalized (t): Normalized vibration signal.

[0074] Image data: Extracting high-level features of images using convolutional layers:

[0075] F = CNN(I) resized )

[0076] In the formula, F: the extracted feature vector, I resized Image after resizing.

[0077] Dataset partitioning. The preprocessed data is divided into a training set and a test set, with a ratio of 80% training set and 20% test set.

[0078] Train set = {(x train ,y train )}

[0079] Test set = {(x test ,y test )}

[0080] In the formula, x train ,x test y: Input data for the training and test sets. train ,y test Label data for the training and test sets.

[0081] Model building.

[0082] CNN Model: Convolutional Layers: Using convolutional kernel K to extract image features:

[0083] F conv =K*I resized

[0084] Pooling layer: using max pooling downsampling:

[0085] F pool =max(F conv )

[0086] Fully connected layer: maps feature vectors to output categories.

[0087] y CNN =softmax(W CNN F pool +b CNN )

[0088] In the formula, K: convolution kernel, F conv : Feature map after convolution, I resized The image after resizing. F pool : Feature map after pooling, W CNN ,b CNN : Weights and biases of the fully connected layer, y CNN : Output probability distribution of the CNN model.

[0089] LSTM Model: LSTM Layers: Processing Time Series Data

[0090] h t =LSTM(x normalized (t),h t-1 )

[0091] Fully connected layer: Maps the output of the LSTM to the output category.

[0092] yLSTM =softmax(W LSTM h t +b LSTM )

[0093] In the formula, h t The hidden state h of the LSTM at time step t. t-1 : The hidden state of the LSTM at time step t-1, x normalized (t): Normalized vibration signal. W LSTM ,b LSTM : Weights and biases of the fully connected layer, y LSTM : Output probability distribution of the LSTM model.

[0094] Model training. Loss function: CNN model: Cross-entropy loss:

[0095]

[0096] LSTM model: Mean squared error loss:

[0097]

[0098] In the formula, y true,i y: The true label of the i-th sampling point CNN,i ,y LSTM,i The model prediction output for the i-th sampling point. Loss function value.

[0099] Optimizer: Training using the Adam optimizer:

[0100]

[0101] In the formula, θ t : The values ​​of the model parameters at time step t, α: the learning rate, The gradient of the loss function with respect to the parameters.

[0102] Model evaluation and parameter optimization. Evaluation metrics: accuracy, recall, and F1 score.

[0103] Learning Outcomes (Deep Network Model). 1. Model Integration and Real-time Monitoring. Model Deployment: The trained deep learning model is deployed to the controller. The controller receives data collected by sensors in real time and performs real-time analysis through the model. Real-time Monitoring: The model performs real-time analysis of the collected vibration data, image data, torque data, and displacement data to predict assembly status and quality. For example: Vibration Data: By analyzing the frequency and amplitude of vibration data, it identifies whether mechanical structural component 20 is correctly installed.

[0104]

[0105] In the formula, T vibration : Abnormal threshold of vibration data, peak x : The maximum amplitude of the vibration data.

[0106] Image data: By analyzing image data, the accuracy of the installation position of mechanical structural component 20 is detected.

[0107]

[0108] In the formula, T image : Abnormal threshold of image data, y CNN : The predicted output of the CNN model.

[0109] Anomaly Detection: When the model detects an anomaly during the assembly process, the controller will trigger an alarm signal to remind the operator to handle it promptly.

[0110]

[0111] In the formula, alarm: alarm status.

[0112] 2. Intelligent Assembly and Adaptive Strategy Optimization. Dynamic Adjustment: Based on the model's output, the controller automatically adjusts the assembly strategy to adapt to various mechanical structural components of different shapes and specifications. Specific adjustment strategies include:

[0113] Adjust the manipulator's motion parameters.

[0114] Position Adjustment: Based on the shape and specifications of the mechanical structural component 20, dynamically adjust the gripping and installation positions of the operating arm.

[0115] position new =position old +Δposition

[0116] Speed ​​adjustment: The moving speed of the operating arm is dynamically adjusted according to the weight and shape of the mechanical structural component 20.

[0117] speed new =speed old adjustment factor

[0118] In the formula, position old speed old : Current position and speed of the manipulator, Δposition: Position adjustment amount, adjustment factor: Speed ​​adjustment factor.

[0119] Adjust the torque control parameters.

[0120] Torque Adjustment: Dynamically adjust the torque control parameters according to the specifications and installation requirements of mechanical structural component 20.

[0121] τ new =τ old +Δτ

[0122] In the formula, τ old : Current torque value, Δτ: Torque adjustment amount.

[0123] Adjust the feed mechanism parameters.

[0124] Feed speed adjustment: The feed speed of the material feeding mechanism 2 is dynamically adjusted according to the shape and specifications of the mechanical structural component 20.

[0125] v new =v old •feed adjustment factor

[0126] Feed position adjustment: The feed position of the material feeding mechanism 2 is dynamically adjusted according to the shape and specifications of the mechanical structural component 20.

[0127] d new =d old +Δd

[0128] In the formula, v old Current feed rate; feed adjustment factor; feed rate adjustment factor, d old : Current feed position, Δd: Feed position adjustment amount.

[0129] Adaptive learning: The device continuously learns new assembly tasks and environmental changes through deep learning models, automatically updating and optimizing assembly strategies. The specific process includes:

[0130] Data accumulation: As the assembly task progresses, the device continuously accumulates new sensor data.

[0131]

[0132] In the formula, θ new ,θ old : New and old parameters of the model, α: learning rate, Gradient operator, The loss function value of the new data.

[0133] Strategy optimization: Based on the updated model, the controller dynamically adjusts the assembly strategy to adapt to new assembly tasks and environmental changes.

[0134] 3. Model Saving and Loading. Model Saving: Save the trained deep learning model locally for easy loading and use later. Model Updates: Retrain or incrementally learn the deep learning model using newly accumulated data to optimize its parameters and structure.

[0135] save(θ CNN ,θ LSTM )

[0136] In the formula, θ CNN ,θ LSTM Parameters of CNN and LSTM models.

[0137] Model loading: Loads the saved model when the system starts up, ensuring that the system can run in real time and perform monitoring and alerts.

[0138] θ CNN ,θ LSTM =load()

[0139] In the formula, θ CNN ,θ LSTM : Parameters of the loaded CNN and LSTM models.

[0140] The installation of track limit blocks will be explained using this example:

[0141] Initially, the track equipped with the end coupler is placed on the support platform 1. The track's position is stabilized by the grooves on both sides of the support platform 1. When the track enters the working area, the piston rod 14 of the feeding cylinder 6 (20mm diameter, 80mm stroke, 0.5MPa air supply) reciprocates linearly in the feeding channel 16 under the action of compressed air, continuously pushing the limit block (0.12kg, 25mm×15mm×10mm) out of the hopper 8. The top cylinder 7 (16mm diameter, 30mm stroke) then pushes it into the track end coupler hole. Vibration sensor 4-1 (ADXL355, ±40g, 1kHz) records the vibration signal x(t) at the cylinder end cap. Displacement sensor 4-4 (magnetostrictive, 0.01mm resolution) records the displacement d(t) ∈ [0,30]mm of the piston rod 14. Loading position determination:

[0142] flag feed =1{max|x(t)|<2.5g∧d(t)∈[29.8,30.2]mm}

[0143] A 4-2 vision sensor (5MP, 25fps, 300mm mounting height) top-down endpoint captures 640×480 grayscale images in real time. raw 128-dimensional features F were extracted using a CNN. loc Calculate the pixel deviation (Δμ, Δv) between the center of the limiting block and the center of the hole:

[0144] (Δμ,Δv)=CNN loc (I raw )∈[-20,20]px

[0145] When |Δμ,Δv| < 5px, the positioning is considered "normal". An electric wrench (adjustable 800–1500 N·m) is suspended from the spring 22 damping mechanism (stiffness 50 N / m). The worker inserts an M8×20 bolt and starts the wrench. Torque sensor 4-3 (strain gauge type, ±2000 N·m, 1 kHz) records the real-time torque τ(t). peak This represents the peak torque. The vision sensor 4-2 takes a second image to check if the bolt is inserted perpendicularly into the hole. Acceptable torque range:

[0146] τ target =1200±50N\cdotpm,flag torq =1{|τ peak -1200|≤50}

[0147] The 1-second vibration segment (1000 points), positioning diagram (640×480), and torque curve (1000 points) of a complete assembly are spliced ​​together to form a sample z=[x(t),I raw ,τ(t)],W CNN ,b CNN Represents the weights and biases of the fully connected layer. The CNN output assembly quality probability is:

[0148] p ok =σ(W CNN ·CNN(I raw )+b CNN )∈[0,1]

[0149] W LSTM ,b LSTM Represents the weights and biases of the fully connected layer. LSTM outputs the probability of vibration anomalies:

[0150] p err =softmax(W LSTM ·LSTM(x(t))+b LSTM [error]

[0151] Overall Score:

[0152] S = 0.6p ok +0.4(1-p err )

[0153] A "retighten" prompt is triggered when S < 0.85. If S < 0.85, the controller immediately executes the following policy update: 1. If p okIf the value is less than 0.8 and |Δμ| > 5px, then calculate the new manipulator offset Δx. arm =k p ·Δμ(k p =0.02mm / px); 2. If |τ peak If -1200|>50N·m, then update the target torque τ of the wrench. new =τ old +γ(τ target -τ peak (γ = 0.3). After every 50 assembly iterations, the model is incrementally trained for 5 epochs using the latest 2000 samples, with a learning rate α = 10. -4 Weight decay λ = 10 -5 After training, the model parameters θ (approximately 7.8 MB) are written to the controller after CRC verification; the loading is completed within 1.2 seconds after the system restarts, ensuring "zero latency" before resuming operation.

[0154] An example of the vision processing chain for this device is as follows:

[0155] like Figure 19 As shown, draw a 240×240 grayscale image to simulate the limiting block captured by the vision sensor 4-2.

[0156] like Figure 20 As shown, the grayscale image is converted to a black and white image to highlight the limiting block and bolt holes. The outline is extracted from the binary image and outlined in green to help manual judgment of whether the positioning is correct.

[0157] like Figure 21 As shown, a 3-layer convolutional network is used to extract spatial features. The brighter the color, the more the network "pays attention" to that region, which is equivalent to the high-level feature extraction process of the CNN in the device.

[0158] This device, centered on "deep learning + multi-sensor fusion," upgrades traditional rigid special-purpose machines into self-learning intelligent assembly platforms. It features a compact structure, quiet operation, and low power consumption, allowing for plug-and-play use on existing production lines. By collecting vibration, visual, torque, and displacement data in real time and using a CNN-LSTM network for online inference, it automatically corrects the grasping posture, feed speed, and tightening torque within milliseconds, achieving high-precision, zero-defect assembly of mechanical structural parts of different shapes and specifications. Modular material bins and multi-axis cylinders support one-click switching, allowing a single unit to cover more than seven types of parts with a changeover time of less than 2 minutes. Actual efficiency is three times higher than manual labor, with an assembly pass rate of ≥99% and a downtime failure rate approaching zero. Therefore, this device is not only suitable for the installation and maintenance of track-type limit blocks but can also be extended to various small-batch scenarios such as automotive connecting rods, motor end covers, and aerospace fasteners, providing a universal, economical, and reliable "one-stop" intelligent assembly solution for flexible and intelligent manufacturing.

[0159] Matters not covered in this invention are common knowledge. The specific embodiments of this invention have been described in detail above with reference to the accompanying drawings. However, this invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of this invention.

[0160] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based adaptive intelligent assembly device for mechanical structural components, characterized in that: Includes a support platform (1), a material feeding mechanism (2), an electric wrench (3), a sensor group, and a controller; The support platform (1) is used to stably support the assembly base to be assembled; the material feeding mechanism (2) is installed on both sides of the support platform (1), and the feeding cylinder (6) and the top cylinder (7) in the material feeding mechanism (2) are used to push the mechanical structural parts (20) from the hopper (8) to the top position (17) and push them into the assembly position (19) of the assembly base; the electric wrench (3) is fixed above the wrench frame (21) and is used to perform assembly operations; the sensor group is used to collect data during the assembly process; the controller is placed in the control cabinet (5) near the operation area and is used to receive and process the sensor data in the sensor group. Its built-in deep learning model evaluates the health status of the assembly process and the assembly quality of the output parts in real time based on this; the controller then dynamically and adaptively adjusts the action parameters of the material feeding mechanism (2) and the electric wrench (3) based on the evaluation results to achieve precise assembly and strategy optimization; The material feeding mechanism (2) includes a feeding cylinder (6), a top feeding cylinder (7), and a hopper (8); the feeding cylinder (6) and the hopper (8) are both placed on a fixed support (23). The hopper (8) accommodates multiple mechanical structural parts (20), and has a feeding channel (16) at its bottom. The piston rod (14) of the feeding cylinder (6) reciprocates linearly in the feeding channel (16). The top feeding cylinder (7) is fixed below the fixed support (23). The feeding channel (6) 16) There is a top material channel (18) on one side. There is an assembly position (19) on the side of the top material channel (18) away from the top material cylinder (7). When the assembly base comes to the working area, the piston rod (14) of the feeding cylinder (6) will push the mechanical structural parts (20) in the hopper (8) to the top material position (17) through the feeding channel (16). Then, the piston rod of the top material cylinder (7) will push the mechanical structural parts (20) into the assembly position (19) of the assembly base. The controller is placed in a control cabinet (5) near the operating area to receive data from the sensor group and to make the following decisions by fusing multi-sensor data through its integrated deep learning model: First, it judges whether the positioning of the mechanical structural component (20) is accurate based on the data from the visual sensor (4-2) and the CNN model; at the same time, the data from the displacement sensor (4-4) is used to confirm whether the feeding and lifting actions are in place; then, during the tightening process of the electric wrench (3), it combines the data from the vibration sensor (4-1) and the torque sensor (4-3) and uses the LSTM model to analyze the vibration characteristics and torque curve of the assembly process to predict the tightening quality; finally, the model outputs a comprehensive score S, and the assembly is judged to be qualified by the S value.

2. The adaptive intelligent assembly device for mechanical structural components based on deep learning according to claim 1, characterized in that: The support platform (1) is fixed on the fixed bracket (23) to stably support the assembly base to be installed. The two sides of the platform have grooves that can be inserted into the assembly base.

3. The adaptive intelligent assembly device for mechanical structural components based on deep learning according to claim 1, characterized in that: The feeding cylinder (6) includes a cylinder head (9), a buffer seal (10), an adjustable buffer screw (11), a cylinder body (12), a piston (13), a piston rod (14), a vent (15), a feeding channel (16), and a pneumatic valve (24). The cylinder head (9) seals the cylinder chambers at both ends of the cylinder body (12). The buffer seals (10) are installed at the inner ends of the cylinder head (9). The adjustable buffer screw (11) is installed on the cylinder head (9) to adjust the optimal buffer at the end of the stroke. The bottom of the cylinder body (12) is located on a fixed bracket (23). The piston (13) and piston rod (14) are connected inside the cylinder body (12). The piston rod (14) can transmit the force of compressed air. The two vents (15) are fixed to the cylinder head (9) by screws and connected to the A and B ports on the pneumatic valve (24). The port P of valve (24) is used to input compressed air, and the R and S ports on the pneumatic valve (24) are exhaust ports. Compressed air is input through port P of the pneumatic valve (24), and the air is guided to port A in the pneumatic valve (24), and then transported to the cylinder head (9) and cylinder body (12) through the vent hole (15). The gas generates high pressure at the cylinder head (9) end, thereby pushing the piston (13), so that the piston rod (14) pushes the mechanical structure (20) into the top position (17) in the feeding channel (16). The exhaust gas generated during the stroke will reach port B on the pneumatic valve (24), and the valve guides this gas to the S end and allows it to be discharged. When the compressed air is guided to the cylinder through the B end, the piston (13) is pushed back, and the excess gas escapes from the R port, thus forming the reciprocating motion of the piston (13), thereby pushing the piston rod (14) to feed.

4. The adaptive intelligent assembly device for mechanical structural parts based on deep learning according to claim 1, characterized in that: The electric wrench (3) is connected to one end of the spring (22) for performing assembly operations. Different tools are installed at its end to adapt to different assembly tasks. The other end of the spring (22) is fixed to the wrench holder (21). The wrench holder (21) is a sturdy metal frame fixed to the fixed bracket (23).

5. The adaptive intelligent assembly device for mechanical structural components based on deep learning according to claim 1, characterized in that: The sensor group includes a vibration sensor (4-1), a vision sensor (4-2), a torque sensor (4-3), and a displacement sensor (4-4). The vibration sensor (4-1) is installed on the material feeding mechanism (2) to monitor the vibration of the mechanical structural component (20) during the feeding, unloading, and assembly processes. The vision sensor (4-2) is installed on the tool head and wrench holder (21) connected to the electric wrench (3) and is positioned directly opposite the assembly base support platform (1) to capture the installation position, assembly quality, and appearance characteristics of the mechanical structural component (20). The torque sensor (4-3) is installed on the end tool of the electric wrench (3) to detect whether the torque applied during the assembly process meets the requirements; the displacement sensor (4-4) is installed on the piston rod of the feeding cylinder (6) and the top cylinder (7) to monitor the movement position of the mechanical structural parts (20); the sensor is connected to the data acquisition card via the CAN bus.

Images