A method for recognizing and tightening a shaft end nut locking position

Abstract

The application provides a shaft end nut locking position recognition and tightening control method, and belongs to the technical field of industrial intelligent manufacturing.The application comprises the following steps: installing a high-precision inductive displacement sensor to monitor the axial displacement of the nut in real time; constructing a tightening control system integrating a torque measurement and displacement monitoring unit; collecting real-time displacement data and eliminating interference through a sliding window filtering algorithm; identifying the nut locking critical point based on a displacement gradient analysis method; calculating an optimal torque function according to an axial end contact mechanics mathematical model; accurately adjusting the locking torque by adopting a temperature compensation type torque control method; and analyzing the locking process parameters and dynamically adjusting the tightening control parameter group by applying a multi-modal locking intelligent recognition model, so that the accurate recognition of the initial contact position and the intelligent optimization adjustment of the tightening parameters in the shaft end nut locking process are realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent manufacturing technology, and specifically relates to a method for identifying and controlling the tightening position of a shaft end nut. Background Technology

[0002] Shaft-end nut tightening technology is widely used in machinery, automotive manufacturing, aerospace, and other fields, and is a key process to ensure appropriate bearing preload and normal operation of rotating components. Traditional shaft-end nut tightening technology mainly relies on torque control, angle control, or a combination of torque and angle control. In industrial production sites, torque wrenches or electric / pneumatic tightening tools are typically used, and fixed torque or angle values ​​are set according to empirical formulas for tightening operations. In some cases, even skilled workers manually determine the tightening status.

[0003] However, traditional locking technologies have significant drawbacks: single torque control is susceptible to changes in the coefficient of friction, leading to large fluctuations in preload; angle control methods struggle to accurately determine the initial contact position, resulting in inaccurate locking positions; while composite control methods offer some improvement, they are still affected by various factors such as ambient temperature, surface condition, and material properties, making them unable to precisely adapt to different working conditions. More importantly, these methods typically rely on fixed parameter control rather than real-time adaptive adjustment, lacking the ability to accurately identify the initial contact position.

[0004] The core problem that existing technologies cannot solve is that they cannot achieve high-precision automatic identification of the initial contact position between the shaft end nut and the locking surface, and they lack intelligent adaptive tightening control capabilities based on real-time analysis of multimodal data, which makes it difficult to guarantee locking accuracy and consistency, especially under complex working conditions and high-precision requirements. Summary of the Invention

[0005] In view of this, the present invention provides a method for identifying and controlling the locking position of a shaft end nut, which can solve the technical problem in the prior art that the initial contact position cannot be accurately identified and adaptive tightening control can be achieved during the locking process of the shaft end nut.

[0006] This invention is implemented as follows: A method for identifying and controlling the locking position of a shaft-end nut includes: installing an inductive displacement sensor at a fixed position on the shaft-end nut locking device, aligning it with the axial displacement path of the nut, and calibrating it; constructing a tightening control system, integrating a torque measurement unit and a displacement monitoring unit, and establishing a torque-displacement correlation mathematical model; collecting real-time displacement data during the shaft-end nut locking process, and eliminating environmental interference through a sliding window filtering algorithm; identifying the nut locking critical point based on a displacement gradient analysis method, determining the initial locking position when the displacement change rate drops to within a preset threshold and multiple consecutive sampling points are stable; calculating the contact state between the nut and the shaft end based on a shaft-end contact mechanics mathematical model, and outputting the optimal torque function curve; precisely adjusting the locking torque using a temperature-compensated torque control method; and applying a multimodal locking intelligent identification model to analyze the locking process parameters and dynamically adjust the tightening control parameter set.

[0007] Among them, the displacement gradient analysis method refers to calculating the rate of change between adjacent sampling points by continuously collected displacement data, and drawing a displacement gradient curve. When the slope of the displacement gradient curve changes significantly and approaches a horizontal state, it is determined that the nut has contacted the locking surface, and the position at this time is the initial locking position.

[0008] Among them, the sliding window filtering algorithm refers to selecting a fixed length of data points as a window during the data acquisition process, calculating the average or median of the data within the window, and as new data points are added, the window slides forward, discarding the earliest data points, thereby achieving data smoothing and reducing random noise interference.

[0009] Among them, the temperature-compensated torque control method refers to arranging multiple temperature sensing elements around the torque sensor, establishing a mathematical relationship model between the ambient temperature value, the internal temperature value of the torque sensor, and the torque measurement deviation value, and correcting the measured torque value in real time under different temperature environments to eliminate the influence of temperature drift on the measurement accuracy.

[0010] Among them, the torque-displacement correlation mathematical model refers to the correlation analysis of torque measurement unit data and displacement monitoring unit data to establish the functional relationship between the two. Through the torque-displacement correlation mathematical model, it is possible to infer one parameter when measuring one parameter, or to determine the nut tightening state through the combined characteristics of the two parameters.

[0011] Among them, the shaft end contact mechanics mathematical model refers to a set of mathematical equations based on Hertzian contact theory, which is used to accurately calculate the pressure distribution and deformation between the nut and the shaft end contact surface. The shaft end contact mechanics mathematical model takes into account the elastic-plastic deformation characteristics of the material and the micro-geometry of the contact surface, and can accurately predict the stress state of the nut at different tightening stages.

[0012] The input parameters of the shaft end contact mechanics mathematical model include the material's elastic modulus, Poisson's ratio, surface roughness, thread pitch, and contact area. The output of the shaft end contact mechanics mathematical model is the optimal torque function curve.

[0013] Among them, the multimodal locking intelligent recognition model refers to an intelligent optimization model based on deep learning. The specific structure of the multimodal locking intelligent recognition model is a dual-stream attention fusion architecture, which includes torque data processing stream and displacement data processing stream. Each stream extracts features by a three-layer convolutional neural network, and then the two modal information are fused through a cross-attention mechanism. Finally, the fully connected layer outputs the optimal locking position prediction and tightening control parameter set.

[0014] The training dataset establishment step in the training process of the multimodal locking intelligent recognition model includes collecting no less than 10,000 sets of historical locking process data. Each set of data includes complete torque curve data, displacement curve data, ambient temperature record data, historical adjustment record data of tightening control parameter group, final locking position data, and locking quality evaluation result data.

[0015] Among them, the number of attention heads in the multimodal locking intelligent recognition model is dynamically adjusted according to the number of nut specification categories, the learning rate of the multimodal locking intelligent recognition model is adaptively adjusted with the change of contact area value, the optimization objective function is the weighted sum of locking position prediction error and tightening parameter optimization loss, and the network parameters are iteratively optimized through gradient descent algorithm.

[0016] This invention monitors the axial displacement of the nut in real time by installing a high-precision inductive displacement sensor, and accurately identifies the initial contact position by combining it with the displacement gradient analysis method; it establishes a multimodal locking intelligent identification model, integrates torque and displacement data for analysis and prediction; and it applies a temperature-compensated torque control method to eliminate environmental interference and achieve dynamic optimization and adjustment of tightening parameters.

[0017] Compared to traditional technologies, this invention solves the problem of inaccurate identification of the initial contact position of the nut, improving the locking position identification accuracy to the order of 0.01mm. By establishing a mathematical model of shaft end contact mechanics, combined with real-time sensor data and environmental temperature compensation, the influence of factors such as temperature fluctuations and surface roughness changes on locking quality is effectively eliminated. At the same time, the multimodal intelligent recognition model is optimized based on a large amount of historical data, and can adaptively adjust the tightening parameters for different nut specifications and working conditions, significantly improving locking consistency and reliability, and solving the core technical problems of initial contact position identification and adaptive tightening control during shaft end nut locking. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figure 1 The diagram shows a flowchart of a method for identifying and controlling the tightening position of a shaft end nut provided by the present invention. This method includes the following steps:

[0021] S01. Install the inductive displacement sensor at the fixed position of the shaft end nut locking device, align the sensing area of ​​the inductive displacement sensor with the axial displacement path of the nut, and use a standard measurement reference block to perform initial calibration of the inductive displacement sensor.

[0022] S02. Construct a tightening control system, which integrates a torque measurement unit and a displacement monitoring unit, and connects to a microprocessor controller to establish a torque-displacement correlation mathematical model.

[0023] S03. Collect real-time displacement data during the tightening process of the shaft end nut, eliminate environmental interference through a sliding window filtering algorithm, and calculate the displacement change rate to provide basic parameters for judging the nut tightening position.

[0024] S04. Identify the critical point for nut locking based on the displacement gradient analysis method. When the displacement change rate drops to within a preset threshold and multiple consecutive sampling points are stable, it is determined to be the initial locking position.

[0025] S05. Calculate the contact state between the nut and the shaft end based on the shaft end contact mechanical mathematical model. The input parameters of the shaft end contact mechanical mathematical model include the material elastic modulus, material Poisson's ratio, surface roughness value, thread pitch value, and contact area value, and output the optimal torque function curve.

[0026] S06. A temperature-compensated torque control method is adopted, which collects the ambient temperature value and the internal temperature value of the torque sensor in real time through temperature sensing elements arranged around the torque sensor, and adjusts the locking torque precisely according to the temperature-torque mapping relationship data.

[0027] S07. Apply a multimodal locking intelligent recognition model to analyze the locking process parameters. The multimodal locking intelligent recognition model predicts the optimal locking position based on historical locking data and dynamically adjusts the tightening control parameter set.

[0028] The displacement gradient analysis method refers to calculating the rate of change between adjacent sampling points by continuously collected displacement data, and plotting the displacement gradient curve. When the slope of the displacement gradient curve changes significantly and approaches a horizontal state, it is determined that the nut has contacted the locking surface, and the position at this time is the initial locking position.

[0029] The sliding window filtering algorithm refers to selecting a fixed length of data points as a window during the data acquisition process, calculating the average or median of the data within the window, and sliding the window forward as new data points are added, discarding the earliest data points, thereby achieving data smoothing and reducing random noise interference.

[0030] Among them, the temperature-compensated torque control method refers to arranging multiple temperature sensing elements around the torque sensor, establishing a mathematical relationship model between the ambient temperature value, the internal temperature value of the torque sensor, and the torque measurement deviation value, and correcting the measured torque value in real time under different temperature environments to eliminate the influence of temperature drift on measurement accuracy.

[0031] The torque-displacement correlation mathematical model refers to the correlation analysis of the data from the torque measurement unit and the data from the displacement monitoring unit to establish a functional relationship between them. Through the torque-displacement correlation mathematical model, one parameter can be inferred when measuring one parameter, or the locking state of the nut can be determined by the combined characteristics of the two parameters.

[0032] Among them, the inductive displacement sensor refers to a displacement detection device with an accuracy of not less than 0.01 mm, which is used to monitor the position change of the nut at the shaft end in the axial direction in real time, so as to ensure the accuracy and reliability of position data acquisition during the locking process.

[0033] The standard measurement reference block refers to a precision measuring tool that has been certified by metrology, with a measurement accuracy of 0.001 mm, and is used for initial calibration and periodic verification of the inductive displacement sensor.

[0034] The shaft end contact mechanics mathematical model refers to a set of mathematical equations based on Hertzian contact theory, used to accurately calculate the pressure distribution and deformation between the nut and the shaft end contact surface. The inputs of the shaft end contact mechanics mathematical model include the material's elastic modulus, Poisson's ratio, surface roughness, thread pitch, and contact area. The output of the shaft end contact mechanics mathematical model is the functional relationship between the optimal locking torque and angular displacement. The shaft end contact mechanics mathematical model considers the material's elastoplastic deformation characteristics and the micro-geometry of the contact surface, and can accurately predict the stress state of the nut at different tightening stages, providing a theoretical basis for precise control.

[0035] The elastic modulus of the material is derived from a material database and is used to calculate the deformation of the contact surface.

[0036] The Poisson's ratio of the material is derived from a material database and is used to calculate the degree of lateral deformation of the contact surface.

[0037] The surface roughness value is obtained by a surface roughness measuring instrument and is used to correct the actual contact state of the contact surface.

[0038] The thread pitch value is derived from the thread specification parameter table and is used to calculate the relationship between the rotation angle and the axial displacement.

[0039] The contact area value is calculated based on the geometric dimensions of the nut and shaft end, and is used to determine the pressure distribution range.

[0040] The ambient temperature value is obtained by an ambient temperature sensor and is used for temperature compensation calculation.

[0041] The internal temperature value of the torque sensor is obtained by measuring a temperature sensing element embedded inside the torque sensor, which is used to compensate for torque measurement errors caused by temperature changes.

[0042] The torque value is obtained by the torque measuring unit and is used to control the tightening process and determine the locking status.

[0043] The displacement change rate is calculated by dividing the displacement difference between two adjacent sampling times by the sampling time interval, and is used to determine the nut's motion state.

[0044] The initial locking position is the position where the nut first contacts the locking surface of the shaft end, which is used to initiate the precise locking control phase.

[0045] The initial position value refers to the axial position value before the nut begins to tighten, which is measured by the inductive displacement sensor and used as a reference point for displacement calculation.

[0046] The locking position value refers to the final axial position value after the nut is locked, and is used for quality assessment and recording.

[0047] The final torque value refers to the torque value measured when locking is completed, which is used to determine whether the locking meets the requirements.

[0048] The displacement curve data refers to the sequence of displacement values ​​recorded throughout the tightening process, which is used to analyze the characteristics of the tightening process and optimize the system.

[0049] The temperature parameter values ​​include the recorded ambient temperature value and the internal temperature value of the torque sensor, which are used for subsequent data analysis and system improvement.

[0050] The tightening control parameter set refers to the set of key parameters controlling the tightening process of the nut, including tightening speed value, tightening acceleration value, maximum torque limit value, torque increase rate value, and tightening angle control value. The tightening control parameter set is dynamically calculated and updated by the multimodal locking intelligent recognition model based on real-time monitoring data, and is used to accurately control the working state of the tightening tool to ensure the tightening accuracy and reliability of the nut.

[0051] The tightening speed value refers to the angular velocity of the tightening tool, measured in radians per second. It is adjusted by the servo motor drive control system and affects the time efficiency and accuracy of the tightening process.

[0052] Among them, the tightening acceleration value refers to the rate of change of the rotational angular velocity of the tightening tool, and the unit is radians per square second. It affects the stability of the tightening process and the mechanical impact on the equipment.

[0053] The maximum torque limit value refers to the maximum torque value that can be applied during the tightening process, measured in Newton-meters, to prevent damage to the threads or deformation of parts due to over-tightening.

[0054] The torque increase rate refers to the rate of change of torque value over time during the tightening process, measured in Newton-meters per second. It controls the smoothness of torque increase and avoids impact loads.

[0055] The tightening angle control value refers to the rotation angle by which the tightening continues after reaching the initial locking position, measured in degrees, and is used to precisely control the final locking position.

[0056] The multimodal locking intelligent recognition model refers to a deep learning-based intelligent optimization model. Its specific structure is a dual-stream attention fusion architecture, comprising a torque data processing stream and a displacement data processing stream. Each stream extracts features using a three-layer convolutional neural network, then fuses the two modalities through a cross-attention mechanism. Finally, a fully connected layer outputs the optimal locking position prediction and the tightening control parameter set. The training dataset establishment steps in the multimodal locking intelligent recognition model training process specifically include collecting no less than 10,000 sets of historical locking process data. Each set of data includes complete torque curve data, displacement curve data, ambient temperature records, historical adjustment records of the tightening control parameter set, final locking position data, and locking quality evaluation results. The data is categorized and labeled according to shaft diameter, nut specification, and material type attributes, and divided into training, validation, and test sets in a 7:2:1 ratio. The multimodal locking intelligent recognition model training steps specifically include first employing unsupervised training... The learning approach preprocesses and denoises the data from each modality, then constructs a tightening control parameter optimization sub-network. This sub-network is trained to learn the optimal tightening control parameter configuration under different working conditions. Simultaneously, a locking position prediction sub-network is constructed to learn the relationship between torque displacement patterns and the optimal locking position. Finally, the two sub-networks are jointly trained using a multi-task learning framework, enabling the multimodal locking intelligent recognition model to simultaneously output accurate locking position predictions and matching tightening control parameter sets. During training, the number of attention heads in the multimodal locking intelligent recognition model is dynamically adjusted according to the number of nut specification categories. The learning rate of the multimodal locking intelligent recognition model is adaptively adjusted according to changes in the contact area value. The optimization objective function is a weighted sum of the locking position prediction error and the tightening parameter optimization loss. The network parameters are iteratively optimized using a gradient descent algorithm. After training, the multimodal locking intelligent recognition model's performance is further improved through online fine-tuning in an actual production environment, achieving accurate prediction and control of the optimal locking parameters for different specification shaft-end nuts.

[0057] The specific implementation methods of the above steps are described in detail below.

[0058] The specific implementation of step S01 involves installing an inductive displacement sensor at the fixed position of the shaft-end nut locking device. This sensor has an accuracy requirement of no less than 0.01 mm. During installation, it is necessary to ensure that the sensor sensing area is precisely aligned with the axial displacement path of the nut, and the alignment error should be controlled within ±0.05 mm. Initial calibration of the sensor is performed using a metrologically certified standard measurement reference block. The calibration process employs the least squares method to fit the sensor's output characteristic curve, ensuring that the linearity error does not exceed 0.5% across the entire measurement range. At least five standard position points are selected for measurement during the calibration process, with each point measured three times and the average value taken to form a calibration curve. This step establishes a high-precision displacement measurement reference, providing fundamental data support for subsequent identification of the nut locking position.

[0059] The specific implementation of step S02 involves constructing a tightening control system integrating a torque measurement unit and a displacement monitoring unit, and connecting it to a microprocessor controller to establish a torque-displacement correlation mathematical model. The torque measurement unit uses a strain gauge torque sensor with a measurement range of 0–500 N·m and an accuracy of 0.1% of full scale. The displacement monitoring unit consists of the inductive displacement sensor installed in step S01, with a sampling frequency set to 1000 Hz to ensure the capture of transient changes. The microprocessor controller uses an industrial-grade 32-bit processor with an operating frequency of no less than 200 MHz and a built-in 16-bit AD converter with a conversion accuracy better than 0.01%. The torque-displacement correlation mathematical model is constructed using a piecewise polynomial fitting method. A first-order linear equation describes the torque-displacement relationship during the nut's free rotation phase; a second-order polynomial equation describes the relationship during the initial contact phase; and an exponential function equation describes the relationship during the locking phase. The model parameters are updated in real-time using the least squares method, and the model prediction accuracy is no less than 95%. The purpose of this step is to establish the system hardware foundation and construct the mathematical model, providing support for subsequent precise locking control.

[0060] The specific implementation of step S03 involves acquiring real-time displacement data during the nut tightening process at the shaft end and using a sliding window filtering algorithm to eliminate environmental interference. During displacement data acquisition, the sensor sampling rate is set to 1000Hz, and the acquisition resolution is 0.001 mm. The sliding window filtering algorithm has a window length of 21 data points and a window sliding step of 1 data point. For the data within each window, the maximum and minimum values ​​are first removed, and then the remaining data points are processed using a weighted average method. The center point has a weight of 1, and the weights of the data points on both sides decrease exponentially with increasing distance from the center point. The weight calculation formula is: weight value = exp(-0.2 × distance value). After processing with this algorithm, data fluctuations caused by environmental interference can be reduced by more than 90% while maintaining signal edge characteristics. The displacement change rate is obtained by calculating the displacement difference between adjacent sampling points divided by the sampling time interval, in millimeters per second. The purpose of this step is to obtain high-quality displacement data, providing a reliable parameter basis for determining the nut tightening position.

[0061] The specific implementation of step S04 is to identify the critical point of nut locking based on the displacement gradient analysis method. The displacement gradient calculation uses the central difference method; the formula for calculating the displacement gradient at time point t is the difference in displacement between two adjacent points divided by the corresponding time interval. The average displacement gradient is calculated within a continuous window, and a displacement gradient curve is plotted. A displacement change rate threshold is set to 10% of the initial free stage displacement change rate; that is, when the displacement change rate drops to within 10% of the initial value, the locking state is judged. To ensure judgment stability, the displacement change rate of at least 10 consecutive sampling points must be stable within the threshold range. Simultaneously, the second derivative of the displacement gradient curve is calculated. When the absolute value of the second derivative is less than 0.01 mm / s² and persists for more than 20 milliseconds, the initial locking position is finally determined by combining the first derivative judgment result. The purpose of this step is to accurately identify the instant the nut contacts the locking surface of the shaft end, providing a starting reference point for subsequent precise control.

[0062] The specific implementation of step S05 involves calculating the contact state between the nut and the shaft end based on a mathematical model of shaft end contact mechanics. This model is built upon Hertzian contact theory, and the input parameters include the material's elastic modulus (unit: GPa), material Poisson's ratio (dimensionless, ranging from 0.25 to 0.35), surface roughness (unit: μm, typical value 0.8 to 3.2 μm), thread pitch (unit: mm), and contact area (unit: mm). 2The model employs the finite element method (FEM), dividing the contact surface into at least 100 mesh elements and calculating the contact pressure distribution on each element. Considering the elastoplastic properties of the material, an elastoplastic correction factor is introduced when the contact stress exceeds 80% of the material's yield strength. The model outputs a curve showing the relationship between torque and angular displacement, divided into three stages: elastic deformation, transition, and stable locking. The optimal torque is selected at 110%–120% of the initial point of the stable locking stage to avoid over-tightening and thread damage. The purpose of this step is to determine the optimal locking torque value based on theoretical calculations, providing target parameters for the control system.

[0063] The specific implementation of step S06 involves employing a temperature-compensated torque control method to achieve precise adjustment of torque measurement. Four temperature sensing elements are evenly arranged around the torque sensor to measure both the ambient temperature and the internal temperature of the sensor. The temperature sensing elements are PT100 platinum resistance thermometers with a measurement accuracy of ±0.1℃. A polynomial fitting model is used to establish the temperature-torque correction coefficient relationship between temperature and torque. For every 1℃ deviation of the ambient temperature from the calibration temperature (typically 20℃), the torque measurement value needs to be corrected by approximately 0.02% to 0.05%. For changes in the internal temperature of the sensor, a second-order thermal conduction model is used to calculate the strain distribution corresponding to the temperature gradient, and then the compensation coefficient is calculated. The temperature compensation algorithm is executed in real-time in the microprocessor, with a compensation frequency of no less than 10Hz, ensuring that the torque measurement accuracy remains within ±0.5% under varying temperature conditions. The purpose of this step is to eliminate the influence of temperature changes on torque measurement and improve the locking control accuracy.

[0064] The specific implementation of step S07 involves applying a multimodal locking intelligent recognition model to analyze the locking process parameters. This intelligent model employs a dual-stream attention fusion architecture, including a torque data processing stream and a displacement data processing stream. The model input consists of real-time collected torque and displacement data, with a sampling interval of 1 millisecond and an input sequence length of 2000 data points. Data preprocessing includes normalization and time alignment, converting different physical quantities into a unified scale. After processing, the model outputs the optimal locking position prediction value and a set of tightening control parameters, including tightening speed (typically 2–5 radians / second), tightening acceleration (typically 1–3 radians / second²), maximum torque limit (determined according to nut specifications, typically 120% of the standard torque), torque increase rate (typically 50–100 N·m / second), and tightening angle control (typically 15–30 degrees). The model's prediction accuracy improves with increasing training data; when the training samples reach 10,000 sets, the locking position prediction accuracy can reach ±0.05 mm. The purpose of this step is to optimize the current locking control strategy by intelligently analyzing historical data, thereby achieving adaptive and precise control.

[0065] The detailed structure of the multimodal locking intelligent recognition model adopts a dual-stream attention fusion architecture. The torque data processing stream and the displacement data processing stream each contain three layers of convolutional neural networks for temporal feature extraction. The kernel sizes for the torque processing stream are 5, 3, and 3, with 32, 64, and 128 channels respectively; the kernel sizes for the displacement processing stream are 7, 5, and 3, with 32, 64, and 128 channels respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to improve model training stability and non-linear expressive power. After feature extraction, a cross-attention mechanism is used to fuse the two modalities. The attention mechanism employs a multi-head self-attention structure, with the number of heads dynamically adjusted according to the number of nut specifications and categories, typically 4 or 8. The fused features are processed through a three-layer fully connected network with hidden layer dimensions of 256, 128, and 64, and finally divided into two output branches: a locking position prediction branch and a tightening control parameter optimization branch. The model has approximately 500,000 parameters, and the inference speed meets the requirements for millisecond-level real-time control.

[0066] The steps for establishing the training dataset for the multimodal locking intelligent recognition model include collecting no fewer than 10,000 sets of historical locking process data. Each set of data includes complete torque curve data (sampling frequency 1000Hz, recording the entire tightening process), displacement curve data (sampling at the same frequency as torque), ambient temperature records (sampling frequency 1Hz), historical adjustment records of tightening control parameter groups, final locking position data (accuracy 0.01 mm), and locking quality evaluation results data (including indicators such as tightness and stability scores). Data collection covers different working conditions, including different shaft diameters (range 5–100 mm), different nut specifications (M6–M36), and different material types (carbon steel, stainless steel, aluminum alloy, etc.). The collected data are classified and labeled according to shaft diameter, nut specification, and material type attributes to ensure that the model learns the optimal locking strategy under different conditions. The dataset is divided into a 70% training set, a 20% validation set, and a 10% test set. The training process employs mini-batch gradient descent with a batch size of 64 and an initial learning rate of 0.001, dynamically adjusting the learning rate using cosine annealing. Training iterations are conducted for at least 100 epochs, stopping early when validation set performance no longer improves. The objective function is a weighted sum of the locking position prediction error and the tightening parameter optimization loss, with a weight ratio of 3:7, reflecting the relative importance of position accuracy and parameter optimization. After model training, online fine-tuning is performed using new data collected in a real-world production environment. The model is updated every 1000 products produced to ensure it adapts to changing production conditions. Fine-tuning uses a relatively small learning rate (typically 10% of the initial learning rate) and a limited number of iterations (10–20 epochs) to maintain model stability while improving adaptability.

[0067] The mathematical model or calculation process involved in this invention will be described in detail below.

[0068] In step S01, the calibration process of the inductive displacement sensor uses the least squares method to fit the sensor's output characteristic curve, which is specifically expressed as follows:

[0069] V o =a0+a1d+a2d 2 +ε c ;

[0070] In the formula, V o d is the sensor output voltage value, in volts (V); d is the displacement value, in millimeters (mm); a0 is the zero-point offset constant, in volts (V); a1 is the linear coefficient, in volts / mm² (V / mm); a2 is the quadratic coefficient, in volts / mm² (V / mm²). 2 );ε c This is the calibration error term, measured in volts (V), and typically ranges from ±0.005V.

[0071] The calibration equation, in quadratic polynomial form, considers the nonlinear characteristics of the sensor. The a0 term represents the zero-position offset, which can be determined by the sensor's output at the zero position; the a1 term represents the sensor's main sensitivity characteristics, typically dominating in the linear operating region; and the a2 term is used to correct for nonlinear effects at large displacements, improving measurement accuracy across the entire range. Solving for the coefficients using the least squares method minimizes the mean square error, enhancing the robustness of the calibration curve.

[0072] The parameter acquisition method is as follows: First, select 5 standard location points d. i (Including the 0 position and the maximum range position), the voltage output value is measured three times at each position to obtain the average voltage V. oi Then construct the least squares objective function J = Taking the partial derivatives with respect to a0, a1, and a2 and setting them to zero, we get:

[0073]

[0074] The system of linear equations is obtained by rearranging the equations:

[0075]

[0076] Solving this system of linear equations will yield the optimal values ​​of parameters a0, a1, and a2.

[0077] In step S02, the torque-displacement correlation mathematical model adopts a piecewise polynomial fitting method, as specifically expressed below:

[0078]

[0079] In the formula, T(d) is the torque value, in Newton-meters (N·m); d is the displacement value, in millimeters (mm); d c The critical contact displacement threshold is expressed in millimeters (mm), with typical values ​​ranging from 0.1 to 0.3 mm; d l k1 is the locking critical displacement threshold, in millimeters (mm), with a typical value of 0.5–1.0 mm; k2 is the linear coefficient of the free stage, in Newton-meter per millimeter (N·m / mm), with a typical value of 0.01–0.05 N·m / mm; k3 is the quadratic coefficient of the contact stage, in Newton-meter per square millimeter (N·m / mm). 2 Typical values ​​are 0.1–0.5 N·m / mm. 2 k3 is the primary coefficient for the contact stage, in Newton-meter per millimeter (N·m / mm), with a typical value of 0.5–2.0 N·m / mm; k4 is the amplitude coefficient for the locking stage, in Newton-meter (N·m), with a typical value of 5–20 N·m; k5 is the exponential coefficient for the locking stage, in millimeters per millimeter (mm). -1 Typical values ​​are 2–5 mm. -1 c1, c2, and c3 are constant terms for each stage, in Newton-meter (N·m), used to ensure the continuity of the function at the segmentation points.

[0080] This segmented model reflects the different physical stages of the nut tightening process: the first stage represents the nut rotating freely, with torque and displacement having an approximately linear relationship; the second stage represents the nut beginning to contact the shaft end face, with torque increasing quadratically with displacement; the third stage represents the nut entering the tightened state, with torque increasing exponentially with displacement. The exponential function can accurately describe the characteristics of material elastic deformation and rapid increase in contact surface pressure during the tightening stage.

[0081] To ensure the function is continuous at the piecewise points, the constant terms c1, c2, and c3 must satisfy the following conditions:

[0082]

[0083] Summarized as follows:

[0084]

[0085] The parameter acquisition method is as follows: torque-displacement data (d) are collected experimentally for different nut specifications. i T i The sampling frequency was 1000Hz, and parameter fitting was performed for each of the three stages. Least squares was used for the free and contact stages, while nonlinear least squares was used for parameter estimation in the locking stage. The fitting process also considered the continuity constraints of the function at the segmentation points.

[0086] In step S03, the mathematical expression of the sliding window filtering algorithm is as follows:

[0087]

[0088] In the formula, d f (t) represents the filtered displacement value in millimeters (mm); d(t+j) represents the original displacement measurement value in millimeters (mm); w j is the filter weight coefficient, dimensionless; n is the half-window length, in units of sampling points, with a value of 10 (corresponding to a 21-point window); t is the time point, in milliseconds (ms).

[0089] The formula for calculating the filter weight coefficients is:

[0090]

[0091] Among them, the weighting coefficient w j Using a normalized exponential decay function ensures The signal scale remains constant. The exponential decay function assigns higher weights to data near the center point and gradually decreases the weights of data farther from the center point, effectively suppressing noise while preserving the main characteristics of the signal. The decay factor of 0.2 for the exponential function is a value determined through experimental optimization, striking a balance between filtering effect and signal distortion.

[0092] The rate of change of displacement is calculated using the central difference method:

[0093]

[0094] In the formula, v(t) is the rate of displacement change, in millimeters per second (mm / s); Δt is the sampling time interval, in milliseconds (ms), with a value of 1 ms (corresponding to a sampling rate of 1000 Hz). The central difference method has higher accuracy than forward or backward difference methods, reducing the error caused by numerical differentiation.

[0095] In step S04, the mathematical expression of the displacement gradient analysis method is as follows:

[0096] First, calculate the displacement gradient:

[0097]

[0098] In the formula, g(t) is the displacement gradient, with units of millimeters per second (mm / s), which is actually the same as the rate of change of displacement v(t); d f (t) represents the filtered displacement value in millimeters (mm); Δt represents the sampling time interval in millimeters (ms).

[0099] Calculation of the second derivative of the displacement gradient:

[0100]

[0101] In the formula, a(t) is the second derivative of the displacement gradient, i.e., the acceleration, with units of millimeters per square second (mm / s²). 2 ).

[0102] Locking critical point judgment condition:

[0103] |g(t)|≤α·|g init |and|a(t)|≤β;

[0104] In the formula, g init α is the average displacement gradient during the initial free phase, in millimeters per second (mm / s); α is the rate of change threshold coefficient, dimensionless, with a value of 0.1; β is the acceleration threshold, in millimeters per square second (mm / s²). 2 The value is 0.01 mm / s. 2 .

[0105] To ensure the stability of the judgment, N consecutive N values ​​are required. s Each sampling point satisfies the above conditions:

[0106] And |a(t-iΔt)|≤β)=N s ;

[0107] In the formula, I(·) is an indicator function, which takes the value 1 when the condition is met, and 0 otherwise; N s The number of consecutive sampling points required for stability assessment is set to 10.

[0108] In step S05, the mathematical model of shaft end contact mechanics is constructed based on Hertzian contact theory, and its main equations are as follows:

[0109] Contact pressure distribution:

[0110]

[0111] In the formula, p(r) is the pressure at a distance r from the contact center, in megapascals (MPa); p0 is the maximum pressure at the center, in megapascals (MPa); a is the radius of the contact circle, in millimeters (mm); and r is the radial distance from the contact center, in millimeters (mm), satisfying 0≤r≤a.

[0112] Calculation of maximum pressure at the center:

[0113]

[0114] In the formula, F is the axial load, in Newtons (N); a is the radius of the contact circle, in millimeters (mm).

[0115] Calculation of contact circle radius:

[0116]

[0117] In the formula, R is the equivalent radius of curvature, in millimeters (mm); E * This is the equivalent elastic modulus, expressed in megapascals (MPa).

[0118] Calculation of equivalent elastic modulus:

[0119]

[0120] In the formula, E1 and E2 are the elastic moduli of the nut and shaft end materials, respectively, in megapascals (MPa); v1 and v2 are the Poisson's ratios of the nut and shaft end materials, respectively, dimensionless, with typical values ​​of 0.25 to 0.35.

[0121] Contact stiffness correction considering the effect of surface roughness:

[0122]

[0123] In the formula, k rough Contact stiffness taking roughness into account, in Newtons per millimeter (N / mm); k smooth γ represents the contact stiffness of an ideal smooth surface, expressed in Newtons per millimeter (N / mm); γ is the roughness influence coefficient, dimensionless, with typical values ​​ranging from 2 to 5; R a This represents the surface roughness value, expressed in micrometers (μm).

[0124] Relationship between torque and axial load:

[0125]

[0126] In the formula, T is the locking torque, in Newton-meter (N·m); F is the axial load, in Newtons (N); P is the thread pitch, in millimeters (mm); μ t ρ is the thread friction coefficient, dimensionless, with a typical value of 0.1–0.2; μ n d is the coefficient of friction of the nut end face, dimensionless, with a typical value of 0.1 to 0.15; m d represents the average diameter of the thread, in millimeters (mm); n The equivalent diameter of the frictional force on the end face of the nut is expressed in millimeters (mm).

[0127] Critical torque determination for elastoplastic modification:

[0128]

[0129] In the formula, T critThis is the critical torque value, expressed in Newton-meters (N·m); T yield The torque value corresponding to the yield strength of the material, in Newton-meter (N·m); F yield The axial load is the yield strength of the material, in Newtons (N); η is the safety factor, dimensionless, with a value of 0.8.

[0130] Optimal torque function curve:

[0131]

[0132] In the formula, T opt (θ) represents the optimal torque function curve, in Newton-meters (N·m); θ is the rotation angle from the contact point, in degrees (°); θ0 is the initial locking angle, in degrees (°), typically 0; k is the torque gain coefficient, dimensionless, ranging from 0.1 to 0.2; λ is the torque growth rate coefficient, in degrees (°). -1 Typical values ​​are 0.05–0.1°. -1 .

[0133] The function curve describes the optimal torque variation during the locking process starting from the contact point, with an initial torque of T. crit As the rotation angle increases, the torque increases slowly and eventually approaches (1+k)T. crit The exponential function form accurately simulates the gradual increase in contact surface pressure during the locking process and avoids stress concentration caused by abrupt changes. Parameter κ controls the ratio of the final torque to the critical torque, and parameter λ controls the rate of torque increase.

[0134] In step S06, the mathematical expression of the temperature-compensated torque control method is as follows:

[0135] Temperature-torque mapping relationship:

[0136] T corr =T meas ·(1+α T ·(T env -T ref )+β T ·(T int -T ref ) 2 );

[0137] In the formula, T corr This is the temperature-compensated torque value, expressed in Newton-meters (N·m); T meas The original torque value obtained from the measurement is expressed in Newton-meters (N·m); T env This refers to the ambient temperature value, expressed in degrees Celsius (°C); T int This refers to the internal temperature value of the torque sensor, in degrees Celsius (°C); Tref This is a reference temperature value, in degrees Celsius (°C), typically taken as 20°C; α T This is the linear compensation coefficient for ambient temperature, expressed in degrees Celsius (°C). -1 Typical values ​​are 0.0002–0.0005℃. -1 ;β T This is the secondary temperature compensation coefficient inside the sensor, expressed in degrees Celsius (°C). -2 Typical values ​​are 0.00001~0.00003℃. -2 .

[0138] This temperature compensation model considers both the linear effect of ambient temperature on the sensor's zero point and the nonlinear effect caused by the internal temperature gradient of the sensor. The linear term α... T ·(T env -T ref This primarily compensates for sensor zero-point drift caused by changes in ambient temperature; the quadratic term β T ·(T int -T ref ) 2 It primarily compensates for the non-uniform strain distribution effect caused by the internal temperature gradient of the sensor.

[0139] The temperature sensors are arranged in a four-point uniform distribution. The internal temperature of the sensors is calculated as follows:

[0140]

[0141] In the formula, T int This represents the average internal temperature of the torque sensor, in degrees Celsius (°C); T i This is the measured value of the i-th temperature sensing element, in degrees Celsius (°C).

[0142] Compensation coefficient α T and β T The data was obtained using an experimental calibration method: Different temperature points (10℃, 20℃, 30℃, 40℃) were set within an ambient temperature control chamber. A standard torque (20%, 40%, 60%, 80%, and 100% of the range) was applied at each temperature point, and the deviation between the measured value and the standard value was recorded. A least-squares objective function was then constructed.

[0143]

[0144] By solving for the parameter α that minimizes J T and β T The optimal compensation coefficient is obtained.

[0145] In step S07, the update formula for the tightening control parameter group in the multimodal locking intelligent recognition model is as follows:

[0146] Tightening speed value updated:

[0147]

[0148] In the formula, v new The updated tightening speed value, in radians per second (rad / s); v base The reference tightening speed value is expressed in radians per second (rad / s), with a typical value of 2–5 rad / s; δ v The speed adjustment coefficient is dimensionless and ranges from 0.5 to 0.8; T is the current measured torque value, in Newton-meters (N·m); T pred The model predicts the ideal torque value, in Newton-meters (N·m); T max This is the maximum permissible torque value, expressed in Newton-meters (N·m).

[0149] Tightening acceleration value update:

[0150]

[0151] In the formula, a new The updated tightening acceleration value is expressed in radians per square second (rad / s). 2 );a base The reference tightening acceleration value is expressed in radians per square second (rad / s). 2 Typical values ​​are 1–3 rad / s 2 ;δ a The acceleration adjustment coefficient is dimensionless and ranges from 0.6 to 0.9; d(|TT) pred |) / dt represents the rate of change of torque deviation, in Newton-meter per second (N·m / s); t ref For reference time, the unit is seconds (s), and the value is 1 second.

[0152] Maximum torque limit updated:

[0153] T max，new =T yield ·(1-σ mat )·(1+τ safety );

[0154] In the formula, T max，new The updated maximum torque limit is expressed in Newton-meters (N·m); T yield σ represents the yield torque of the material, expressed in Newton-meters (N·m); mat τ is the material property dispersion coefficient, dimensionless, with typical values ​​ranging from 0.1 to 0.15; safety The safety factor is dimensionless and typically ranges from 0.2 to 0.3.

[0155] Torque increase rate value updated:

[0156]

[0157] In the formula, The updated torque increase rate is expressed in Newton-meters per second (N·m / s). The reference torque increase rate is expressed in Newton-meters per second (N·m / s), with a typical value of 50–100 N·m / s; γ is the sensitivity coefficient, dimensionless, ranging from 0.5 to 1.0; |a(t)| is the absolute value of the current displacement acceleration, expressed in millimeters per square second (mm / s²). 2 );a ref For reference acceleration values, the unit is millimeters per square second (mm / s²). 2 The value is 0.1 mm / s. 2 .

[0158] Tightening angle control value updated:

[0159]

[0160] In the formula, θ new The updated tightening angle control value, in degrees (°); θ base The reference tightening angle value is expressed in degrees (°), with a typical value of 15–30°; Δθ is the angle compensation amount, expressed in degrees (°), with a typical value of ±5°; E ref The elastic modulus of the reference material is expressed in gigapascals (GPa), typically taken as 210 GPa (for steel); E mat R represents the actual elastic modulus of the material, expressed in gigapascals (GPa). a This represents the actual surface roughness value, in micrometers (μm); R a，ref The reference surface roughness value is in micrometers (μm) and is set to 1.6 μm.

[0161] The training objective function of the multimodal locking intelligent recognition model is:

[0162] L total =λ1·L pos +λ2·L param ;

[0163] In the formula, L total L is the total loss function; pos Loss prediction for locking position; L param The loss is optimized for tightening parameters; λ1 and λ2 are weighting coefficients, dimensionless, with values ​​of 0.3 and 0.7 respectively.

[0164] Location prediction loss calculation:

[0165]

[0166] In the formula, d pred，i The locking position value predicted by the model, in millimeters (mm); d true，i The actual optimal locking position value is expressed in millimeters (mm); N is the number of samples.

[0167] Tightening parameter optimization loss calculation:

[0168]

[0169] In the formula, p pred，i，j p is the j-th tightening parameter value predicted by the model. opt，i，j p is the actual optimal tightening parameter value for the j-th time. range，j w represents the effective range span of the j-th parameter. j The weighting coefficient for the j-th parameter is dimensionless; the five parameters correspond to the tightening speed value, tightening acceleration value, maximum torque limit value, torque increase rate value, and tightening angle control value, respectively.

[0170] Through the aforementioned mathematical model, the shaft end nut locking position identification and tightening control method can achieve high-precision locking position identification, accurate torque control, and intelligent parameter optimization, improving the reliability and consistency of nut locking. The functional relationships used in each equation have their physical or engineering significance: linear relationships are used to describe simple proportional changes, such as the relationship between torque and displacement in the free stage; quadratic polynomial relationships are used to describe nonlinear smooth changes, such as sensor calibration curves and torque changes in the initial contact stage; exponential function relationships are used to describe rapidly growing processes, such as torque changes with displacement in the locking stage, characterized by slow initial growth followed by rapid growth, consistent with material stress-strain characteristics; the power function relationship in Hertzian contact theory reflects the nonlinear relationship between contact area and load, accurately describing the mechanical behavior of the contact interface; the polynomial form in the temperature compensation model can simultaneously consider linear and nonlinear temperature effects, improving compensation accuracy. All equation parameters have been experimentally verified and optimized to ensure high predictive accuracy and stability in actual working conditions. The model comprehensively considers material properties, surface morphology, temperature influence, and dynamic process characteristics, achieving precise control of the shaft end nut locking process.

[0171] Specifically, the principle of this invention is based on the idea of ​​combining precise displacement detection with intelligent adaptive control, involving the integration of knowledge from multiple disciplines such as sensing measurement, mechanical modeling, data processing, and intelligent optimization.

[0172] First, this invention uses an inductive displacement sensor with an accuracy of no less than 0.01 mm to monitor the axial displacement of the nut in real time, and employs a standard measurement reference block for initial calibration to ensure the accuracy of the measurement reference. The acquired displacement data is processed by a sliding window filtering algorithm to eliminate environmental interference and improve signal quality. The key technical innovation lies in the application of a displacement gradient analysis method. By continuously calculating the rate of displacement change and plotting the displacement gradient curve, when the slope of the curve changes significantly and approaches a horizontal state, it can be determined that the nut has contacted the locking surface, thereby accurately identifying the initial locking position.

[0173] Secondly, this invention establishes a mathematical model of shaft end contact mechanics based on Hertzian contact theory. This model considers multiple parameters such as material elastic modulus, Poisson's ratio, surface roughness, thread pitch, and contact area, enabling precise calculation of the pressure distribution and deformation between the nut and the shaft end contact surface, providing a theoretical basis for the tightening process. Simultaneously, a temperature-compensated torque control method is employed, utilizing temperature sensing elements arranged around the torque sensor to collect ambient temperature and sensor internal temperature values ​​in real time. Based on the temperature-torque mapping data, the tightening torque is precisely adjusted, eliminating the impact of temperature drift on measurement accuracy.

[0174] The most innovative aspect of this invention is the multimodal locking intelligent recognition model. This model employs a dual-stream attention fusion architecture, including torque data processing streams and displacement data processing streams, fusing information from both modalities through a cross-attention mechanism. The model is trained based on no fewer than 10,000 sets of historical locking process data, covering complete torque curves, displacement curves, ambient temperature records, and other information. The model can dynamically calculate and update the tightening control parameter set based on real-time monitoring data, including tightening speed values, acceleration values, and maximum torque limit values, achieving adaptive optimization control for different working conditions.

[0175] This multimodal data fusion and intelligent adaptive control approach breaks through the limitations of traditional single-parameter control. It can comprehensively consider multi-dimensional information such as material properties, environmental factors, and historical data to achieve accurate modeling and intelligent control of the locking process. This solves the core technical problems of inaccurate initial contact position identification and difficulty in adaptively adjusting tightening parameters during the shaft end nut locking process.

[0176] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0177] The specific implementation of step S01 involves installing an inductive displacement sensor at the fixed position of the shaft end nut locking device. This sensor has an accuracy requirement of no less than 0.01 mm. During installation, first determine the centerline of the axial displacement path of the nut, and precisely align the sensor sensing area with this centerline, controlling the alignment error within ±0.05 mm. Then, fix the sensor bracket to ensure no displacement occurs during measurement. Next, use a metrologically certified standard measurement reference block for initial sensor calibration. The calibration process employs the least squares method to fit the sensor output characteristic curve, specifically expressed as follows: V o =a0+a1d+a2d 2 +ε c In the formula, V o d is the sensor output voltage value, in volts (V); d is the displacement value, in millimeters (mm); a0 is the zero-point offset constant, in volts (V); a1 is the linear coefficient, in volts / mm² (V / mm); a2 is the quadratic coefficient, in volts / mm² (V / mm²). 2 );ε c The calibration error term is expressed in volts (V), typically ranging from ±0.005V. During calibration, five standard locations are selected for measurement, with each location measured three times and the average value taken to form a calibration curve. To obtain the calibration equation parameters a0, a1, and a2, a least-squares objective function is constructed. The optimal parameter values ​​are determined by taking the partial derivatives of each parameter and setting them to zero, and then solving the resulting system of linear equations. This step establishes a high-precision displacement measurement benchmark, providing fundamental data support for subsequent identification of the nut tightening position. Furthermore, the nonlinear characteristics of the sensor are considered through quadratic polynomial fitting, improving measurement accuracy and reliability.

[0178] The specific implementation of step S02 involves constructing a tightening control system integrating a torque measurement unit and a displacement monitoring unit, and connecting it to a microprocessor controller to establish a torque-displacement correlation mathematical model. First, a strain gauge torque sensor is selected as the torque measurement unit, with a range of 0–500 N·m and an accuracy of 0.1% of full scale. Then, the inductive displacement sensor installed in step S01 is used as the displacement monitoring unit, with a sampling frequency set to 1000 Hz to ensure the capture of transient changes. Next, an industrial-grade 32-bit processor is selected as the microprocessor controller, with an operating frequency of no less than 200 MHz and a built-in 16-bit AD converter with a conversion accuracy better than 0.01%. Finally, a torque-displacement correlation mathematical model is established in the microprocessor controller. This model is constructed using a piecewise polynomial fitting method, specifically expressed as follows: T(d) =

[0179] In the formula, T(d) is the torque value, in Newton-meters (N·m); d is the displacement value, in millimeters (mm); d cThe critical contact displacement threshold is expressed in millimeters (mm), with typical values ​​ranging from 0.1 to 0.3 mm; d l k1 is the locking critical displacement threshold, in millimeters (mm), with a typical value of 0.5–1.0 mm; k2 is the linear coefficient of the free stage, in Newton-meter per millimeter (N·m / mm), with a typical value of 0.01–0.05 N·m / mm; k3 is the quadratic coefficient of the contact stage, in Newton-meter per square millimeter (N·m / mm). 2 Typical values ​​are 0.1–0.5 N·m / mm. 2 k3 is the primary coefficient for the contact stage, in Newton-meter per millimeter (N·m / mm), with a typical value of 0.5–2.0 N·m / mm; k4 is the amplitude coefficient for the locking stage, in Newton-meter (N·m), with a typical value of 5–20 N·m; k5 is the exponential coefficient for the locking stage, in millimeters per millimeter (mm). -1 Typical values ​​are 2–5 mm. -1 c1, c2, and c3 are constant terms for each stage, in Newton-meters (N·m), used to ensure the continuity of the function at the piecewise points. To ensure continuity at the piecewise points, the constant terms must satisfy the following condition: and The segmented model established in this step accurately reflects the different physical stages in the nut tightening process, providing a theoretical basis and algorithmic support for subsequent precise tightening control. At the same time, the model's fitting accuracy to different tightening stages is improved by combining polynomials and exponential functions.

[0180] The specific implementation of step S03 involves acquiring real-time displacement data during the tightening process of the shaft-end nut and using a sliding window filtering algorithm to eliminate environmental interference. First, the displacement sensor sampling rate is set to 1000Hz, and the acquisition resolution is set to 0.001mm to ensure the capture of minute displacement changes. Then, the acquired raw displacement data is smoothed using a sliding window filtering algorithm, as detailed below: In the formula, d f (t) represents the filtered displacement value in millimeters (mm); d(t+j) represents the original displacement measurement value in millimeters (mm); w j is the filter weight coefficient, dimensionless; n is the half-window length, in units of sampling points, with a value of 10 (corresponding to a 21-point window); t is the time point, in milliseconds (ms). The filter weight coefficients are calculated using a normalized exponential decay function: Next, the rate of change of displacement is calculated using the central difference method: In the formula, v(t) is the rate of displacement change, in millimeters per second (mm / s); Δt is the sampling time interval, in milliseconds (ms), with a value of 1 ms. This step effectively eliminates random noise interference through a sliding window filtering algorithm, improving the signal-to-noise ratio of the displacement data while preserving the main characteristics of the signal. This provides a reliable parameter basis for determining the nut tightening position, and the filtering effect can reduce environmental interference by more than 90%, while the signal edge characteristic retention rate is higher than 95%.

[0181] The specific implementation of step S04 is based on the identification of the nut tightening critical point using the displacement gradient analysis method. First, the displacement gradient and acceleration are calculated using the filtered displacement data obtained in step S03. The displacement gradient calculation employs the central difference method. In the formula, g(t) is the displacement gradient, with units of millimeters per second (mm / s), which is actually the same as the rate of change of displacement v(t); d f (t) represents the filtered displacement value in millimeters (mm); Δt represents the sampling time interval in millimeters (ms). The second derivative of the displacement gradient (acceleration) is calculated as follows: In the formula, a(t) is the second derivative of the displacement gradient, i.e., the acceleration, with units of millimeters per square second (mm / s²). 2 Then, the locking critical point judgment condition is set as follows: |g(t)|≤α·|g init |and|a(t)|≤β; where g init α is the average displacement gradient during the initial free phase, in millimeters per second (mm / s); α is the rate of change threshold coefficient, dimensionless, with a value of 0.1; β is the acceleration threshold, in millimeters per square second (mm / s²). 2 The value is 0.01 mm / s. 2 To ensure the stability of the judgment, it is required that N consecutive... s Each sampling point satisfies the above conditions: And |a(t-iΔt)|≤β)=N s In the formula, I(·) is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; N s The number of continuous sampling points required for stability assessment is set to 10. Finally, when the conditions are met, the current position is determined to be the initial locking position, the position value is recorded, and the subsequent precise locking control process is triggered. This step accurately identifies the moment the nut contacts the shaft end locking surface by analyzing displacement gradient changes, providing a starting reference point for subsequent precise control. The identification accuracy is higher than 98%, and the positioning accuracy is better than 0.02 mm.

[0182] The specific implementation of step S05 involves calculating the contact state between the nut and the shaft end based on the mathematical model of shaft end contact mechanics. First, a contact pressure distribution equation is established based on Hertzian contact theory: In the formula, p(r) is the pressure at a distance r from the contact center, in megapascals (MPa); p0 is the maximum pressure at the center, in megapascals (MPa); a is the radius of the contact circle, in millimeters (mm); and r is the radial distance from the contact center, in millimeters (mm), satisfying 0 ≤ r ≤ a. Calculation of the maximum pressure at the center: In the formula, F is the axial load, measured in Newtons (N). Contact circle radius calculation: In the formula, R is the equivalent radius of curvature, in millimeters (mm); E * This is the equivalent elastic modulus, expressed in megapascals (MPa). Calculation of the equivalent elastic modulus: In the formula, E1 and E2 are the elastic moduli of the nut and shaft end materials, respectively, in megapascals (MPa); v1 and v2 are the Poisson's ratios of the nut and shaft end materials, respectively, dimensionless, with typical values ​​of 0.25–0.35. Then, the contact stiffness is corrected by considering the effect of surface roughness: In the formula, k rough Contact stiffness taking roughness into account, in Newtons per millimeter (N / mm); k smooth γ represents the contact stiffness of an ideal smooth surface, expressed in Newtons per millimeter (N / mm); γ is the roughness influence coefficient, dimensionless, with typical values ​​ranging from 2 to 5; R a This represents the surface roughness value, in micrometers (μm). Next, the relationship between torque and axial load is established: In the formula, T is the locking torque, in Newton-meter (N·m); F is the axial load, in Newtons (N); P is the thread pitch, in millimeters (mm); μ t ρ is the thread friction coefficient, dimensionless, with a typical value of 0.1–0.2; μ n d is the coefficient of friction of the nut end face, dimensionless, with a typical value of 0.1 to 0.15; m d represents the average diameter of the thread, in millimeters (mm); n The effective diameter of the frictional force on the nut end face is given, in millimeters (mm). Considering the material's elastic-plastic properties, determine the critical torque value: In the formula, T crit This is the critical torque value, expressed in Newton-meters (N·m); T yield The torque value corresponding to the yield strength of the material, in Newton-meter (N·m); F yield The axial load corresponds to the material's yield strength, expressed in Newtons (N); η is the safety factor, dimensionless, with a value of 0.8. Finally, the optimal torque function curve is generated: In the formula, T opt(θ) represents the optimal torque function curve, in Newton-meters (N·m); θ is the rotation angle from the contact point, in degrees (°); θ0 is the initial locking angle, in degrees (°), typically 0; k is the torque gain coefficient, dimensionless, ranging from 0.1 to 0.2; λ is the torque growth rate coefficient, in degrees (°). -1 Typical values ​​are 0.05–0.1°. -1 This step, by establishing a mechanical model based on Hertzian contact theory, accurately calculated the contact state between the nut and the shaft end and the optimal torque function curve, providing target parameters for the control system. This effectively combined theoretical calculation with actual control, improving torque control accuracy to ±2%.

[0183] The specific implementation of step S06 involves using a temperature-compensated torque control method to achieve precise adjustment of torque measurement. First, four PT100 platinum resistance temperature sensing elements are evenly arranged around the torque sensor to measure the ambient temperature and the internal temperature of the sensor, with a measurement accuracy of ±0.1℃. Then, a mathematical model of the temperature-torque mapping relationship is established: T corr =T meas ·(1+α T ·(T env -T ref )+β T ·(T int -T ref ) 2 In the formula, T corr This is the temperature-compensated torque value, expressed in Newton-meters (N·m); T meas The original torque value obtained from the measurement is expressed in Newton-meters (N·m); T env This refers to the ambient temperature value, expressed in degrees Celsius (°C); T int This refers to the internal temperature value of the torque sensor, in degrees Celsius (°C); T ref This is a reference temperature value, in degrees Celsius (°C), typically taken as 20°C; α T This is the linear compensation coefficient for ambient temperature, expressed in degrees Celsius (°C). -1 Typical values ​​are 0.0002–0.0005℃. -1 ;β T This is the secondary temperature compensation coefficient inside the sensor, expressed in degrees Celsius per square degree (°C). -2 Typical values ​​are 0.00001~0.00003℃. -2 The internal temperature of the sensor is calculated by averaging the measurements from four uniformly distributed points: In the formula, T int This represents the average internal temperature of the torque sensor, in degrees Celsius (°C); T iThe value measured by the i-th temperature sensing element is expressed in degrees Celsius (°C). The temperature compensation coefficient α is then determined through experimental calibration. T and β T The experiment was conducted in an ambient temperature controlled chamber, with four temperature points set at 10℃, 20℃, 30℃, and 40℃. A standard torque (20%, 40%, 60%, 80%, and 100% of the range) was applied at each temperature point, and the deviation between the measured values ​​and the standard values ​​was recorded. A least-squares objective function was constructed: J = By solving for the parameter α that minimizes J T and β T The optimal compensation coefficient is obtained. Finally, the temperature compensation algorithm is executed in real time in the microprocessor, with the compensation frequency set to 10Hz, ensuring that the torque measurement accuracy remains within ±0.5% under varying temperature conditions. This step, by establishing a temperature-torque mapping model, effectively eliminates the influence of temperature changes on torque measurement, improves locking control accuracy, and enables the system to maintain stable operation within a temperature range of -10℃ to 50℃.

[0184] The specific implementation of step S07 involves applying a multimodal locking intelligent recognition model to analyze the locking process parameters. First, a neural network model with a dual-stream attention fusion architecture is constructed, including a torque data processing stream and a displacement data processing stream. Each stream consists of a three-layer convolutional neural network for temporal feature extraction. Then, the real-time acquired torque and displacement data are preprocessed, including normalization and time alignment, converting different physical quantities into a unified scale. Next, the tightening control parameter set is updated based on the model output, including updates to the tightening speed value. In the formula, v new The updated tightening speed value, in radians per second (rad / s); v base The reference tightening speed value is expressed in radians per second (rad / s), with a typical value of 2–5 rad / s; δ v The speed adjustment coefficient is dimensionless and ranges from 0.5 to 0.8; T is the current measured torque value, in Newton-meters (N·m); T pred The model predicts the ideal torque value, in Newton-meters (N·m); T max This is the maximum permissible torque value, in Newton-meters (N·m). Tightening acceleration value updated: In the formula, a new The updated tightening acceleration value is expressed in radians per square second (rad / s). 2 );a base The reference tightening acceleration value is expressed in radians per square second (rad / s). 2 Typical values ​​are 1–3 rad / s 2 ;δ aThe acceleration adjustment coefficient is dimensionless and ranges from 0.6 to 0.9; d(|TT) pred |) / dt represents the rate of change of torque deviation, in Newton-meter per second (N·m / s); t ref For reference time, the unit is seconds (s), and the value is 1 second. Maximum torque limit updated: T max，new =T yield ·(1-σ mat )·(1+τ safety In the formula, T max，new The updated maximum torque limit is expressed in Newton-meters (N·m); T yield σ represents the yield torque of the material, expressed in Newton-meters (N·m); mat τ is the material property dispersion coefficient, dimensionless, with typical values ​​ranging from 0.1 to 0.15; safety For safety factors, dimensionless, typical values ​​are 0.2–0.3. Torque increase rate updated: In the formula, The updated torque increase rate is expressed in Newton-meters per second (N·m / s). The reference torque increase rate is expressed in Newton-meters per second (N·m / s), with a typical value of 50–100 N·m / s; γ is the sensitivity coefficient, dimensionless, ranging from 0.5 to 1.0; |a(t)| is the absolute value of the current displacement acceleration, expressed in millimeters per square second (mm / s²). 2 );a ref For reference acceleration values, the unit is millimeters per square second (mm / s²). 2 The value is 0.1 mm / s. 2 Tightening angle control value updated: In the formula, θ new The updated tightening angle control value, in degrees (°); θ base The reference tightening angle value is expressed in degrees (°), with a typical value of 15–30°; Δθ is the angle compensation amount, expressed in degrees (°), with a typical value of ±5°; E ref The elastic modulus of the reference material is expressed in gigapascals (GPa), typically taken as 210 GPa (for steel); E mat R represents the actual elastic modulus of the material, expressed in gigapascals (GPa). a This represents the actual surface roughness value, in micrometers (μm); R a，ref The reference surface roughness value is in micrometers (μm), and is set to 1.6 μm. The model training uses the total loss function: L total =λ1·L pos +λ2·L param In the formula, L total L is the total loss function; posLoss prediction for locking position; L param The loss is optimized for tightening parameters; λ1 and λ2 are dimensionless weighting coefficients, with values ​​of 0.3 and 0.7 respectively. Location prediction loss calculation: Tightening parameter optimization loss calculation: This step utilizes a multimodal locking intelligent recognition model to achieve adaptive optimization of locking process parameters, thereby improving locking accuracy and reliability. When the model has 10,000 training data sets, the locking position prediction accuracy can reach ±0.05 mm, and the parameter optimization effect improves locking quality by more than 5%.

[0185] To better understand and implement this invention, a specific application scenario, Example 2, is provided below: A method for identifying and controlling the locking position of shaft-end nut is implemented on an automotive transmission production line. An output shaft end nut of M24×1.5 is selected as the research object; this nut is used to fix the preload of the output shaft bearing. According to research data, the locking quality of the transmission output shaft nut directly affects the bearing preload and the transmission noise level. Traditional locking methods relying on torque control have significant variability, affecting product consistency. By implementing the method of this invention, the nut locking process of a batch of 50 transmissions was optimized on the production line.

[0186] First, install the inductive displacement sensor according to step S01. Select a WD-V101 inductive displacement sensor with an accuracy of 0.005 mm. Fix the sensor on the bracket of the tightening device, with the sensing end facing the output shaft end face. Adjust the position so that the sensor axis coincides with the nut axis, controlling the deviation within ±0.03 mm. Use a standard measuring reference block to calibrate the sensor. Set the calibration points to six positions: 0, 0.2, 0.4, 0.6, 0.8, and 1.0 mm. Repeat the measurement three times at each point. The sensor output characteristic curve parameters are shown in Table 1.

[0187] Table 1 Calibration parameters of displacement sensor

[0188] [a0] 3.157 V [a1] 5.264 V / mm <![CDATA[a2]]> -0.153 <![CDATA[V / mm 2 ]]> linearity 0.22 ％ Repeatability 0.15 ％

[0189] A tightening control system was constructed, using an HBM T22 torque sensor as the torque measurement unit, with a range of 0–600 N·m and an accuracy of 0.1% of full scale. The control system uses an STM32F407 chip as the microprocessor controller, with an operating frequency of 168 MHz, and employs a 16-bit ADC for data acquisition. A torque-displacement correlation mathematical model was established. Through analysis of 50 sets of pre-acquired data, the segmented model parameters of the M24×1.5 nut were obtained, as shown in Table 2.

[0190] Table 2 Parameters of the Mathematical Model for Torque-Displacement Correlation

[0191] Freedom Phase <![CDATA[k1]]> 0.023 N·m / mm Freedom Phase <![CDATA[c1]]> 0.134 N·m Contact phase <![CDATA[k2]]> 0.328 <![CDATA[N·m / mm 2 ]]> Contact phase <![CDATA[k3]]> 1.247 N·m / mm Contact phase <![CDATA[c2]]> 0.156 N·m Locking phase <![CDATA[k4]]> 12.537 N·m Locking phase <![CDATA[k5]]> 3.285 <![CDATA[mm -1 ]]> Locking phase <![CDATA[c3]]> 2.418 N·m Contact critical point <![CDATA[d c ]]> 0.214 mm Locking critical point <![CDATA[d l ]]> 0.683 mm

[0192] Real-time displacement data during the nut tightening process was collected at a sampling rate of 1000Hz and a resolution of 0.001 mm. A sliding window filtering algorithm was applied to process the raw data, with a window length of 21 points and an exponential decay factor of 0.2. The noise levels before and after processing are compared in Table 3.

[0193] Table 3 Comparison of noise levels before and after sliding window filtering.

[0194] static state 0.00723 0.00063 91.3 Low-speed rotation (1 rad / s) 0.00842 0.00078 90.7 Medium speed rotation (3 rad / s) 0.00967 0.00091 90.6 High-speed rotation (5 rad / s) 0.01134 0.00112 90.1

[0195] The critical point for nut tightening was identified using displacement gradient analysis. The rate of change threshold coefficient α was set to 0.1, and the acceleration threshold β was set to 0.01 mm / s². 2 The number of continuous sampling points required for stability assessment was set to 10. Table 4 shows a comparison between the locking critical point identification results of 50 products and the results of manual verification, based on actual testing.

[0196] Table 4. Verification of the accuracy of locking critical point identification.

[0197] <0.01 21 42.0 42.0 0.01～0.02 16 32.0 74.0 0.02～0.03 8 16.0 90.0 0.03～0.05 4 8.0 98.0 >0.05 1 2.0 100.0

[0198] The contact state between the nut and the shaft end was calculated based on the mathematical model of shaft end contact mechanics. The M24×1.5 nut and the output shaft are both made of 40Cr steel, and the relevant parameters are shown in Table 5.

[0199] Table 5 Calculation parameters for contact mechanics

[0200] Elastic modulus of nut material <![CDATA[2.1×10 5 ]]> MPa Elastic modulus of shaft end material <![CDATA[2.1×10 5 ]]> MPa Poisson's ratio of nut material 0.28 - Poisson's ratio of shaft end material 0.28 - Nut end face surface roughness 1.6 μm Shaft end surface roughness 0.8 μm Thread pitch value 1.5 mm Average thread diameter 22.65 mm Equivalent diameter of frictional force 19.2 mm Thread friction coefficient 0.15 - end face friction coefficient 0.12 -

[0201] The calculated critical torque value is 265.3 N·m. In the optimal torque function curve parameters, the torque gain coefficient κ is 0.15, and the torque growth rate coefficient λ is 0.08°. 1 .

[0202] A temperature-compensated torque control method was adopted, with four PT100 temperature sensing elements evenly arranged around the torque sensor. The ambient temperature linear compensation coefficient α was obtained through temperature calibration experiments. T It is 0.00032℃ -1 The sensor's internal temperature secondary compensation coefficient β T It is 0.000023℃ -2 The results of torque measurement accuracy verification under different temperature conditions are shown in Table 6.

[0203] Table 6 Comparison of Torque Measurement Accuracy Before and After Temperature Compensation

[0204] 0 5.3 -1.28 -0.21 83.6 10 12.4 -0.65 -0.15 76.9 20 23.5 0.12 0.09 25.0 30 35.7 0.83 0.17 79.5 40 48.2 1.74 0.32 81.6

[0205] A multimodal locking intelligent recognition model was applied to analyze the locking process parameters. The model adopted a dual-stream attention fusion architecture and was trained using 12,000 sets of historical locking data. The prediction accuracy of the model after training is shown in Table 7.

[0206] Table 7 Performance of the Multimodal Locking Intelligent Recognition Model

[0207]

[0208]

[0209] The tightening control parameters are automatically adjusted based on the model output. The initial tightening speed is set to 3.5 rad / s, and the tightening acceleration is set to 2.0 rad / s. 2 The maximum torque limit is 318.4 N·m, the torque increase rate is 75 N·m / s, and the tightening angle control is 25°. Parameters are adjusted in real time according to the material and surface condition, and adaptive control improves the tightening quality.

[0210] The results of the comparative test between the method of this invention and the traditional locking method are shown in Table 8.

[0211] Table 8 Comparison of the Invention Method and Traditional Methods

[0212] Locking position standard deviation 0.085mm 0.026mm 69.4 Locking torque standard deviation 18.4 N·m 6.3 N·m 65.8 Standard deviation of axial preload 2.83kN 0.92kN 67.5 Average time of locking process 7.5s 5.2s 30.7 Standard deviation of bearing temperature rise 3.4℃ 1.2℃ 64.7 average vibration noise 68.5dB 65.1dB 5.0

[0213] Traditional nut tightening methods rely primarily on fixed torque values ​​or torsion angles for control, which cannot adapt to factors such as batch material variations, surface condition changes, and environmental temperature fluctuations. Traditional methods typically employ manual labor or simple mechanical devices for tightening, lacking real-time monitoring and precise control capabilities, resulting in significant fluctuations in tightening quality and impacting product performance consistency. This invention utilizes an inductive displacement sensor to monitor the nut position in real time, combined with torque measurement to construct a multi-modal monitoring system. It accurately identifies the tightening critical point through a sliding window filtering algorithm and displacement gradient analysis, establishes a mechanical model based on Hertzian contact theory to calculate the optimal torque function curve, employs temperature-compensated torque control to eliminate environmental influences, and applies an intelligent model to predict optimal tightening parameters. Compared to traditional methods, this invention achieves high-precision tightening position identification and torque control, significantly improving the consistency of tightening quality, reducing axial preload fluctuations, minimizing bearing temperature rise differences, improving the vibration and noise characteristics of the transmission, and simultaneously increasing production efficiency. Especially in mass production, the adaptive capability of this invention makes product quality more stable, providing strong support for improving the performance and ensuring the quality of automotive transmissions.

[0214] It should be noted that the variables involved in this invention are explained in detail in Tables 9 and 10 below.

[0215] Table 9. Variable Explanation Table (Part 1)

[0216]

[0217]

[0218] Table 10 Variable Explanation Table (Part Two)

[0219]

[0220] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and controlling the tightening position of a shaft end nut, characterized in that, include: Install an inductive displacement sensor at the fixed position of the shaft end nut locking device, align it with the axial displacement path of the nut, and calibrate it; construct a tightening control system, integrate a torque measurement unit and a displacement monitoring unit, and establish a torque-displacement correlation mathematical model; collect real-time displacement data during the shaft end nut tightening process, and eliminate environmental interference through a sliding window filtering algorithm; The nut locking critical point is identified based on displacement gradient analysis. When the displacement change rate drops to within a preset threshold and multiple consecutive sampling points stabilize, it is determined to be the initial locking position. The contact state between the nut and the shaft end is calculated based on the shaft end contact mechanics mathematical model, and the optimal torque function curve is output. A temperature-compensated torque control method is used to precisely adjust the locking torque. A multimodal locking intelligent recognition model is applied to analyze the locking process parameters and dynamically adjust the tightening control parameter set. The multimodal locking intelligent recognition model is a deep learning-based intelligent optimization model. The specific structure of the multimodal locking intelligent recognition model is a dual-stream attention fusion architecture, which includes a torque data processing stream and a displacement data processing stream. Each stream extracts features by a three-layer convolutional neural network, and then the two modal information are fused through a cross-attention mechanism. Finally, the optimal locking position prediction and tightening control parameter set are output by a fully connected layer.

2. The method for identifying and controlling the locking position of the shaft end nut according to claim 1, characterized in that, The displacement gradient analysis method refers to calculating the rate of change between adjacent sampling points by continuously collected displacement data, and plotting the displacement gradient curve. When the rate of change of displacement decreases to within the preset threshold range of the rate of change of displacement in the initial free stage, and the displacement gradient curve enters a stable horizontal state, it is determined that the nut has contacted the locking surface, and the position at this time is the initial locking position.

3. The method for identifying and controlling the locking position of the shaft end nut according to claim 2, characterized in that, The sliding window filtering algorithm selects a fixed length of data points as a window during data acquisition, calculates the average or median of the data within the window, and slides the window forward as new data points are added, discarding the earliest data points, thereby achieving data smoothing and reducing random noise interference.

4. The method for identifying and controlling the locking position of the shaft end nut according to claim 3, characterized in that, Temperature-compensated torque control method refers to arranging multiple temperature sensing elements around the torque sensor, establishing a mathematical relationship model between the ambient temperature value, the internal temperature value of the torque sensor, and the torque measurement deviation value, and correcting the measured torque value in real time under different temperature environments to eliminate the influence of temperature drift on measurement accuracy.

5. The method for identifying and controlling the locking position of the shaft end nut according to claim 4, characterized in that, The torque-displacement correlation mathematical model refers to the correlation analysis of torque measurement unit data and displacement monitoring unit data to establish the functional relationship between the two. Through the torque-displacement correlation mathematical model, it is possible to infer one parameter when measuring one parameter, or to determine the nut tightening state through the combined characteristics of the two parameters.

6. The method for identifying and controlling the locking position of the shaft end nut according to claim 5, characterized in that, The shaft end contact mechanics mathematical model refers to a set of mathematical equations based on Hertzian contact theory, which is used to accurately calculate the pressure distribution and deformation between the nut and the shaft end contact surface. The shaft end contact mechanics mathematical model takes into account the elastic-plastic deformation characteristics of the material and the micro-geometry of the contact surface, and can accurately predict the stress state of the nut at different tightening stages.

7. The method for identifying and controlling the locking position of the shaft end nut according to claim 6, characterized in that, The input parameters for the shaft end contact mechanics mathematical model include the material's elastic modulus, Poisson's ratio, surface roughness, thread pitch, and contact area. The output of the shaft end contact mechanics mathematical model is the optimal torque function curve.

8. The method for identifying and controlling the locking position of the shaft end nut according to claim 7, characterized in that, The steps for establishing the training dataset in the training process of the multimodal locking intelligent recognition model include collecting no less than 10,000 sets of historical locking process data. Each set of data includes complete torque curve data, displacement curve data, ambient temperature record data, historical adjustment record data of tightening control parameter group, final locking position data, and locking quality evaluation result data.

9. The method for identifying and controlling the locking position of the shaft end nut according to claim 8, characterized in that, The number of attention heads in the multimodal locking intelligent recognition model is dynamically adjusted according to the number of nut specification categories. The learning rate of the multimodal locking intelligent recognition model is adaptively adjusted with the change of contact area value. The optimization objective function is the weighted sum of the locking position prediction error and the tightening parameter optimization loss. The network parameters are iteratively optimized through the gradient descent algorithm.

Images