Mechatronics assembly method and system based on force feedback control

Abstract

The invention relates to the technical field of force feedback control, in particular to a mechanical-electrical integration assembly method and system based on force feedback control. The method comprises the following steps: acquiring electromechanical integrated assembly sensing data; carrying out data preprocessing on the obtained electromechanical integrated assembly sensing data; constructing a multivariable nonlinear prediction model, and training the model by using the preprocessed data set; generating a control signal based on adaptive genetic and fuzzy control algorithms; and electromechanical integrated assembly is realized based on a control system. And force feedback control is realized through cooperation of various algorithms such as PID, adaptive heredity and fuzzy, the assembly action can be adjusted in real time according to the assembly force, and the quality problem caused by improper assembly force, part position deviation and the like is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of force feedback control, and in particular to a mechatronic assembly method and system based on force feedback control. Background Art

[0002] As the manufacturing industry develops towards intelligent and high-precision manufacturing, mechatronic assembly, as a key link, has a direct impact on product quality and production efficiency. In actual production, traditional assembly technology faces many challenges, while existing automated assembly technology also has significant shortcomings. These issues together constitute the technical background of this invention.

[0003] In traditional assembly, parts of the process rely heavily on manual experience. Assembly workers rely on their accumulated tactile experience and visual judgment to complete component assembly, a method significantly influenced by human factors. Differences in skill level and work status among different workers lead to uneven assembly quality, making it difficult to ensure product consistency. Furthermore, manual assembly is inefficient and cannot meet the demands of large-scale, batch production. Furthermore, while some simple mechanical positioning assembly technologies can reduce manual reliance to a certain extent, they lack the precise perception and control of key parameters such as force and position during the assembly process, enabling only relatively crude assembly operations and failing to meet the assembly requirements of high-precision electromechanical products.

[0004] In terms of automated assembly technology, although there are already many automated assembly equipment on the market that can perform basic assembly functions, improve production efficiency, and reduce labor costs, they generally lack precise force feedback control mechanisms. In the mechatronic assembly process, the magnitude and direction of the assembly force have a significant impact on the installation accuracy and connection reliability of components. Equipment that lacks precise force feedback control cannot sense changes in assembly force in real time. When encountering component size deviations, differences in material properties, or changes in the assembly environment, the assembly action cannot be adjusted in time, which can easily lead to component damage, excessive or insufficient assembly clearances, and other problems, resulting in unstable assembly quality. In the assembly of precision electronic components, excessive assembly force may damage the components, while too little assembly force cannot guarantee the firmness of the connection. In the shaft-hole assembly of mechanical components, improper assembly force control will affect the matching accuracy of the shaft and hole, and thus affect the performance of the entire mechanical system.

[0005] From the perspective of industry development trends, with the rapid development of aerospace, high-end equipment manufacturing, precision instruments and other fields, the requirements for the precision, reliability, and stability of mechatronic products are becoming increasingly stringent. The shortcomings of traditional assembly techniques and existing automated assembly technologies have severely constrained the development of these industries. To meet market demand for high-quality mechatronic products and enhance the global competitiveness of my country's manufacturing industry, there is an urgent need to research and develop a mechatronic assembly method and system that can achieve precise force feedback control, thereby improving assembly accuracy and quality and realizing an automated and intelligent assembly process. Summary of the Invention

[0006] In order to solve the above-mentioned problems, the present invention provides a mechatronic assembly method and system based on force feedback control.

[0007] In a first aspect, the present invention provides a mechatronic assembly method based on force feedback control, which adopts the following technical solution:

[0008] A mechatronic assembly method based on force feedback control, comprising:

[0009] Acquire mechatronic assembly sensor data;

[0010] Perform data preprocessing on the acquired mechatronic assembly sensor data;

[0011] Build a multivariate nonlinear prediction model and train the model using the preprocessed dataset;

[0012] Generate control signals based on adaptive genetic and fuzzy control algorithms;

[0013] Realize mechatronic assembly based on control system.

[0014] Furthermore, the data preprocessing of the acquired mechatronic assembly sensor data includes adopting a mean filtering algorithm. For a sequence x1, x2, ..., xn containing n data, the filtered data yi is expressed as:

[0015] Among them, m is the filter window size, and this algorithm is used to remove random noise in the data.

[0016] Furthermore, the multivariate nonlinear prediction model is constructed and the model is trained using the preprocessed data set, including determining the key variables in the assembly process based on the preprocessed mechatronic assembly sensor data, describing the dynamic coupling relationship between the variables through a state space model, constructing the relationship between the key variables through prior knowledge and mechanisms, and introducing a multilayer perceptron (MLP). The key variables of the key variables are constructed using prior knowledge and mechanisms as input, and the predicted values of the output layer are obtained through nonlinear transformation of several hidden layers, wherein the number of neurons in the output layer is determined according to the dimension of the predicted value.

[0017] Furthermore, the construction of a multivariate nonlinear prediction model and training the model using a preprocessed data set also includes feature engineering on the input layer, calculating derived features in addition to directly using the original sensor data, including calculating the rate of change of assembly force and acceleration of position; in the output layer, in addition to predicting the position posture P and assembly force F of the component, additional evaluation indicators of assembly quality are output, including assembly accuracy and assembly reliability. Through multi-task learning, the model can learn multiple related tasks at the same time, thereby improving the generalization ability and overall performance of the model; in the hidden layer, a hierarchical architecture design is adopted, and hidden layers at different levels are responsible for extracting features at different levels of abstraction, among which the bottom hidden layer extracts local and detailed features, and the high-level hidden layer extracts global and abstract features.

[0018] Furthermore, the multivariate nonlinear prediction model is constructed and the model is trained using the preprocessed data set, including using stratified sampling partitioning and time series data partitioning methods to divide the acquired data into training sets, validation sets and test sets, and training the multivariate nonlinear prediction model by using the backpropagation BP algorithm, wherein the error of the output is calculated and the error is backpropagated to the hidden layer and the input layer, and then the bias of the connection weights between neurons is adjusted according to the error through the chain rule.

[0019] Furthermore, the method of constructing a multivariable nonlinear prediction model and training the model using the preprocessed data set also includes combining particle swarm optimization with a genetic algorithm to quickly search for the approximate area of the global optimal solution through information sharing and iterative updates between particles; and using the genetic algorithm to perform a fine search in the local area through selection, crossover, and mutation operations. The update formula of the particle swarm optimization algorithm is:

[0020] υ i (t+1)=wυ i (t)+c1r1(p best，i -x i (t))+c2r2(g best -x i (t))

[0021] x i (t+1)=x i (t)+υ i (t+1)

[0022] where vi(t) is the velocity of the ith particle at time t, xi(t) is the position of the ith particle at time t, \omega is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p best，i is the historical optimal position of the i-th particle, g best is the global optimal position.

[0023] Furthermore, the construction of a multivariate nonlinear prediction model and training the model using the preprocessed data set also includes introducing a simulated annealing method to accept inferior solutions with a certain probability, thereby preventing the algorithm from falling into a local optimum. Assuming the current parameter is θ, the updated parameter is θ', the loss function is L(θ), and the temperature is T, the probability of accepting the inferior solution is:

[0024]

[0025] As the training progresses, the temperature T gradually decreases and the algorithm gradually converges to the global optimal solution.

[0026] Furthermore, the construction of a multivariate nonlinear prediction model and training the model using the preprocessed dataset also includes constructing a reinforcement learning environment, treating the assembly process as a Markov decision process, and training the model by adopting a deep deterministic policy gradient method, wherein the goal of the actor network is to maximize the cumulative reward, and its update formula is:

[0027]

[0028] where θ μ are the parameters of the actor network, θ Q are the parameters of the critic network, Q(s, a|θ Q ) is the action value output by the critic network, μ(s|θ μ ) is the action output by the actor network.

[0029] Furthermore, the control signal is generated based on the adaptive genetic and fuzzy control algorithm, including optimizing the parameters Kp, Ki, and Kd of the PID controller using the adaptive genetic algorithm and finding the optimal parameter combination by simulating the biological evolution process, converting the input error e and the error change rate ec into fuzzy linguistic variables using the fuzzy control algorithm, formulating fuzzy control rules based on assembly experience and expert knowledge, and performing reasoning using the Mamdani reasoning method based on the input fuzzification results and the fuzzy rule base, and finally converting the total output fuzzy set into a precise control quantity.

[0030] In a second aspect, a mechatronic assembly system based on force feedback control comprises:

[0031] A data acquisition module is configured to acquire mechatronic assembly sensor data;

[0032] A preprocessing module is configured to perform data preprocessing on the acquired mechatronic assembly sensor data;

[0033] The training module is configured to construct a multivariate nonlinear prediction model and train the model using the preprocessed data set;

[0034] The control module is configured to generate control signals based on adaptive genetic and fuzzy control algorithms; and implement mechatronic assembly based on the control system.

[0035] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device to implement the mechatronic assembly method based on force feedback control.

[0036] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement the mechatronic assembly method based on force feedback control.

[0037] In summary, the present invention has the following beneficial technical effects:

[0038] The invention is based on the electromechanical integration assembly technology effect of force feedback control

[0039] High-precision sensors collect multi-source data, including force, position, and vision, during the assembly process in real time. After data preprocessing and fusion, a comprehensive model is constructed to accurately describe the relationships between various factors in the assembly process. Force feedback control is collaboratively implemented using multiple algorithms, including PID, adaptive genetics, and fuzzy algorithms. This allows for real-time adjustment of assembly movements based on assembly force, effectively avoiding quality issues caused by improper assembly force and component position deviations. In the assembly of precision electronic components or mechanical parts through shafts and holes, assembly errors can be controlled within extremely small limits, ensuring component installation accuracy and connection reliability, significantly improving product quality consistency, and reducing scrap and rework rates.

[0040] This technical solution enables an automated and intelligent mechatronic assembly process, eliminating reliance on manual experience, reducing labor costs, and addressing the inherent inefficiency of manual assembly. The collaborative operation of multiple algorithms and the dynamic model update mechanism enable the system to rapidly adapt to varying assembly conditions and product requirements, eliminating the need for frequent manual debugging and intervention. Even with complex assembly tasks or product changeovers, assembly operations can be completed efficiently, significantly shortening assembly cycles, meeting the demands of large-scale, batch production, and significantly improving production efficiency.

[0041] The comprehensive model incorporates prior knowledge and mechanistic models, combining multiple nonlinear models with innovative structural designs. Through dynamic updating mechanisms and continuous learning capabilities, the system can adapt to different types of mechatronic product assembly, including aerospace, high-end equipment manufacturing, precision instruments, and other fields that require extremely high assembly precision. Furthermore, it maintains excellent assembly performance despite complex situations such as individual component differences and environmental changes. This effectively addresses the limited application scenarios of traditional and existing automated assembly technologies, expands the application boundaries of mechatronic assembly technology, and enhances the technology's versatility and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a mechatronic assembly method based on force feedback control according to Example 1 of the present invention;

[0043] Figure 2 This is a diagram verifying the effect of Example 1 of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below with reference to the accompanying drawings.

[0045] Example 1

[0046] Reference Figure 1 , a mechatronic assembly method based on force feedback control of this embodiment includes:

[0047] Acquire mechatronic assembly sensor data;

[0048] Perform data preprocessing on the acquired mechatronic assembly sensor data;

[0049] Build a multivariate nonlinear prediction model and train the model using the preprocessed dataset;

[0050] Generate control signals based on adaptive genetic and fuzzy control algorithms;

[0051] Realize mechatronic assembly based on control system.

[0052] Specifically:

[0053] S1. Acquisition of mechatronic assembly sensor data

[0054] Force sensors, position sensors, and vision sensors are used to collect real-time force data (magnitude, direction, etc.), component position and posture information, and assembly environment and component image information during the assembly process. Force sensors can be strain gauge force sensors, piezoelectric force sensors, etc.; position sensors can use encoders, grating scales, etc.; and vision sensors can use industrial cameras.

[0055] S2. Preprocess the acquired data and perform data filtering: Using the mean filtering algorithm, for a sequence of n data x1, x2, ..., xn, the calculation formula for the filtered data yi is:

[0056]

[0057] Where m is the filter window size. This algorithm can effectively remove random noise from the data. Data normalization: Use linear normalization to map the data to the [0,1] interval. For data x, the normalized value x' is calculated as follows:

[0058]

[0059] Where xmax and xmin are the maximum and minimum values in the data sequence respectively.

[0060] Data Fusion:

[0061] Using the weighted average fusion algorithm, let the measurement value of sensor i be xi, the weight be wi, and The fused data y is:

[0062]

[0063] The weight wi is determined according to the accuracy and reliability of the sensor.

[0064] S3. Build a comprehensive model

[0065] 3.1 Determining model variables and relationships

[0066] (1) Variable determination

[0067] Based on the original assembly force vector F (including forces Fx, Fy, Fz along the x, y, and z axes and moments Mx, My, Mz around each axis), the component position and posture vector P (composed of the component's coordinate position in three-dimensional space (x, y, z) and the rotation angles (α, β, γ) around the x, y, and z axes), the component characteristic parameter vector C (covering the component's material properties such as elastic modulus and Poisson's ratio, and geometric dimensions such as length, diameter, thickness, etc.), the environmental parameter vector E (including ambient temperature, humidity, etc.), and the time variable t, the variables are further subdivided. For the component characteristic parameter C, the surface roughness of the component is considered because the surface roughness affects the friction between the components, thereby affecting the assembly force and assembly accuracy. In addition, the characteristic parameters A of the assembly tool are introduced, such as the stiffness and damping of the assembly tool. These parameters will affect the force transmission and motion control during the assembly process.

[0068] For environmental parameters E, in addition to temperature and humidity, the influence of the ambient magnetic field should also be considered. In some high-precision mechatronic assemblies, the ambient magnetic field may interfere with the measurement accuracy of the sensor or generate additional forces on components containing magnetic materials.

[0069] (2) Relationship Analysis

[0070] In addition to analyzing the static relationships between variables, their dynamic coupling relationships must also be studied in depth. The relationship between the assembly force F and the component position and posture P depends not only on the force and position at the current moment, but also on the rate of change of the force and the historical trajectory of the position. This dynamic coupling relationship can be described by establishing a dynamic system model, including a state-space model. Assuming the system's state vector is X = [P^T,\dotP^T]^T (where \dotP is the rate of change of position and posture), the state equation can be expressed as \dotX = f(X,F,C,E,A,t), and the output equation is Y = g(X,F,C,E,A,t), where Y can be the output variable to be predicted, including the final assembly quality indicator.

[0071] Consider interactions between variables. Ambient temperature and component material properties can jointly influence component thermal expansion, which in turn affects assembly clearance and assembly force. Quantify the strength and influence of these interactions through experimental design and data analysis.

[0072] 3.2 Introduction of prior knowledge and mechanism model

[0073] (1) Deepening of physical mechanism modeling

[0074] For different types of assembly operations, more detailed physical mechanism models are established. In shaft-hole assembly, in addition to considering the torque preload relationship, a contact mechanics model between the shaft and the hole is also established. According to Hertz contact theory, the contact stress and deformation of the shaft and hole under different assembly forces are calculated, thereby more accurately predicting the force and position changes during the assembly process. Assuming the radius of the shaft is r1, the radius of the hole is r2, the elastic modulus of the materials are E1 and E2 respectively, the Poisson's ratio is \nu1 and \nu2 respectively, and the contact force is Fc, the maximum contact stress \sigmamax in the contact area can be calculated using the following formula:

[0075]

[0076] Where a is the semi-axis length of the contact ellipse, which can be calculated according to the relevant formula of Hertz contact theory.

[0077] For complex assembly systems, we build a multi-physics coupling model. In the assembly of electrical components, we consider the coupling of electric and magnetic fields with mechanical forces. Using finite element analysis software, we perform multi-physics simulations of the assembly process, extracting key physical parameters and laws and integrating them into a comprehensive model.

[0078] (2) Construction of knowledge graph of domain expert experience

[0079] The experience and knowledge of domain experts are collected to construct a knowledge graph. A knowledge graph represents expert knowledge in the form of a graph, where nodes represent entities (such as assembly operations, component types, and failure modes), and edges represent relationships between entities (such as causal relationships and associations). An expert might know that under a specific assembly process, when the assembly force exceeds a certain threshold, a failure mode prone to component damage will occur. This knowledge is represented as an edge in the knowledge graph.

[0080] Leverage knowledge graphs for reasoning and decision-making. During model building and assembly, reasoning is performed based on the current assembly state and the knowledge graph to predict potential problems and provide corresponding solutions. When the assembly force is detected approaching a certain threshold, the system can adjust the assembly strategy based on the knowledge graph to avoid component damage.

[0081] 3.3 Selecting the model function form

[0082] (1) Diversification and expansion of nonlinear models

[0083] In addition to multilayer perceptrons (MLPs), other neural network types such as recurrent neural networks (RNNs) and their variants (such as long short-term memory (LSTM) and gated recurrent units (GRUs)) are introduced. Because the assembly process has a time series nature, RNNs and their variants are better able to process sequential data and capture temporal dependencies within the assembly process. LSTM networks are used to model time series data of assembly forces and positions, predicting the assembly state at future moments.

[0084] Consider using a convolutional neural network (CNN) to process image data collected by vision sensors. CNNs have powerful feature extraction capabilities and can extract assembly-related features from images, such as component shape, position, and surface defects. Fusing the CNN-extracted features with other sensor data and inputting them into a comprehensive model improves model accuracy and robustness.

[0085] (2) Strengthening the attention mechanism

[0086] Compared to traditional attention mechanisms, the multi-head attention mechanism can simultaneously focus on different aspects of the input data and extract richer feature information. In assembly models, the multi-head attention mechanism can focus on different components of assembly force and different dimensions of component position, thereby more accurately capturing the relationship between variables.

[0087] The attention mechanism is combined with time series analysis to achieve dynamic attention allocation. At different assembly stages, the attention weight is dynamically adjusted according to the changes in time series data, so that the model pays more attention to key time points and variables.

[0088] 3.4 Model structure design

[0089] (1) Optimization of input and output layers

[0090] Feature engineering is performed on the input layer. In addition to directly using raw sensor data, some derived features are calculated. These include the rate of change of assembly force and position acceleration. These derived features provide more information and help the model better learn the relationships between variables.

[0091] The output layer not only predicts the component's position and posture P and assembly force F, but also outputs assembly quality evaluation metrics such as assembly accuracy and reliability. Through multi-task learning, the model simultaneously learns multiple related tasks, improving its generalization and overall performance.

[0092] (2) Hierarchical structure design of hidden layers

[0093] The hidden layers are designed using a hierarchical architecture, with different hidden layers responsible for extracting features at different levels of abstraction. The bottom hidden layers extract local, detailed features, while the higher hidden layers extract global, abstract features. This hierarchical architecture enables the model to better handle complex assembly data.

[0094] The introduction of skip connections (such as residual blocks in ResNet) alleviates the vanishing gradient problem in deep neural networks and improves the training efficiency and stability of the model. Skip connections allow information to be directly transferred between hidden layers at different levels, making it easier for the model to learn the identity mapping, thereby enabling the construction of deeper network structures.

[0095] 3.5 Dynamic update mechanism of the model

[0096] (1) Data-driven adaptive updates

[0097] Online learning algorithms, such as online support vector machines (OnlineSVM), are used to update the model in real time. These algorithms can quickly adjust model parameters as new data arrives, adapting to dynamic changes in the assembly process. When new failure modes or anomalies are detected during the assembly process, the online learning algorithm can promptly update the model, improving its ability to handle these new situations.

[0098] Design a data selection strategy to prioritize data that has a significant impact on model updates for training. Focus on learning data that has a large error with the current model predictions or has new feature patterns to improve the efficiency and effectiveness of model updates.

[0099] (2) Model fusion and transfer learning

[0100] When encountering new assembly tasks or significantly changing the assembly environment, we employ model fusion and transfer learning. We fuse existing models with sub-models trained for the new task or environment, leveraging existing knowledge and experience. If some assembly operations in the new task are similar to those in the previous task, we can transfer relevant parameters and features from the previous model to the new model, reducing training time and data requirements.

[0101] A model library is established to store different types of assembly models and sub-models. In actual applications, appropriate models are selected from the model library for combination and adjustment based on specific assembly tasks and environmental conditions, achieving rapid model updates and optimization.

[0102] S4. Comprehensive model training process

[0103] 4.1 Data Preparation

[0104] (1) Synchronous multimodal data acquisition: In addition to collecting data from force sensors, position sensors, and visual sensors, as well as recording component characteristic parameters, environmental parameters, and time variables, acoustic sensor data is further introduced. During the assembly process, different assembly states may produce different sound characteristics, such as the collision and friction of components. By synchronously collecting acoustic signals, they are incorporated into model training as an additional information source. Let the sound signal collected by the acoustic sensor be S(t), which is a function of time t. It is combined with other data to form a sample (Fi, Ci, Ei, Pi, ti, Si).

[0105] (2) Simulated data generation: To expand the diversity of training data, physical simulation software is used to generate simulated assembly data. Based on known physical mechanism models and assembly process parameters, the assembly process under different working conditions is simulated to obtain simulated force, position, posture and other data. Simulated data can help the model learn some situations that may rarely occur in actual collected data, thereby improving the model's generalization ability.

[0106] (3) Stratified sampling: Considering that actual assembly data may have class imbalance, the number of samples of different types of assembly failures varies greatly. Stratified sampling is used to divide the data into training, validation, and test sets, ensuring that each subset contains samples of various types, and the proportions are consistent with the original data. Assuming that there are K categories in the data set and the number of samples in the kth category is nk, when dividing, samples from each category are extracted into the training, validation, and test sets according to the same proportion.

[0107] (3) Time series data partitioning: Since assembly data has time series characteristics, the time sequence must be considered when partitioning the data. A sliding window approach can be used to treat a continuous period of time series data as a sample. A time window of length T is selected, and the data from time t to time + T is used as a sample. The training set, validation set, and test set are then partitioned in chronological order to ensure that the model can learn the temporal dependencies of the data during training.

[0108] 4.2 Selecting a training algorithm

[0109] (1) Hybrid Optimization Algorithm

[0110] Combining Particle Swarm Optimization with Genetic Algorithms: Traditional optimization algorithms can become stuck in local optimal solutions when dealing with complex nonlinear models. Combining the Particle Swarm Optimization (PSO) algorithm with the Genetic Algorithm (GA) leverages the strengths of both. PSO uses information sharing and iterative updates between particles to quickly locate the approximate region of the global optimal solution. Genetic algorithms, through operations such as selection, crossover, and mutation, enable more refined searches in local areas.

[0111] The update formula of the particle swarm optimization algorithm is:

[0112] υ i (t+1)=wυ i (t)+c1r1(p best，i -x i (t))+c2r2(g best -x i (t))

[0113] x i (t+1)=x i (t)+υ i (t+1)

[0114] where vi(t) is the velocity of the ith particle at time t, xi(t) is the position of the ith particle at time t, \omega is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p best，i is the historical optimal position of the i-th particle, g best is the global optimal position.

[0115] Genetic algorithm operations include selection, crossover, and mutation. Selection can use a roulette wheel selection method, where the probability of each individual being selected is proportional to its fitness value. Crossover can use single-point crossover or multi-point crossover methods. Mutation can randomly change the genes of an individual.

[0116] Gradient descent algorithm based on simulated annealing:

[0117] When using the gradient descent algorithm to update model parameters, the idea of simulated annealing is introduced. The simulated annealing algorithm can accept inferior solutions with a certain probability, thus preventing the algorithm from falling into a local optimum. Let the current parameter be θ, the updated parameter be θ', the loss function be L(θ), and the temperature be T. The probability of accepting an inferior solution is:

[0118]

[0119] As the training progresses, the temperature T gradually decreases and the algorithm gradually converges to the global optimal solution.

[0120] (2) Reinforcement learning training

[0121] Constructing a reinforcement learning environment: The assembly process is considered a Markov decision process (MDP), defining a state space S, an action space A, a reward function R, and a state transition probability P. The state space S can be composed of the assembly force F, the component position and posture P, component characteristic parameters C, environmental parameters E, the time variable t, and the acoustic signal S. The action space A can be the control actions of the assembly tool, such as movement speed and rotation angle. The reward function R is used to evaluate the quality of each action, giving positive rewards when the assembly force is close to the target value and the position and posture are accurate, and negative rewards when an assembly failure occurs.

[0122] Deep reinforcement learning algorithm: The Deep Deterministic Policy Gradient (DDPG) algorithm is used for training. DDPG is a policy gradient-based deep reinforcement learning algorithm suitable for problems in continuous action spaces. It consists of an actor network (ActorNetwork) and a critic network (CriticNetwork). The actor network generates actions, and the critic network evaluates the value of actions.

[0123] The goal of the actor network is to maximize the cumulative reward, and its update formula is:

[0124]

[0125] where θ μ are the parameters of the actor network, θ Q are the parameters of the critic network, Q(s, a|θ Q ) is the action value output by the critic network, μ(s|θ μ ) is the action output by the actor network.

[0126] The goal of the critic network is to minimize the temporal difference error, and its update formula is:

[0127]

[0128] in

[0129] y i =r i +γQ′(s i+1 , μ′(s i+1 |θ μ′ )|θ Q′ ), ri is the immediate reward, γ is the discount factor, θ μ′ and θ Q′ are the parameters of the target network.

[0130] 4.3 Model Evaluation and Optimization

[0131] (1) Comprehensive evaluation of multiple indicators

[0132] In addition to using indicators such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R^2) to evaluate model performance, new evaluation indicators are introduced. The assembly success rate indicator is introduced, which is the proportion of assembly operations predicted by the model that can be successfully completed; the fault prediction accuracy indicator is introduced to evaluate the model's ability to predict assembly faults. Assuming the number of samples of successful assembly is nsuccess and the total number of samples is ntotal, the assembly success rate is Assume that the number of samples that correctly predict faults is n correct , the actual number of fault samples is n fault , then the fault prediction accuracy is

[0133] The weighted comprehensive evaluation method is used to assign different weights to different indicators according to their importance and calculate the comprehensive evaluation score. Let w1, w2, w3, and w4 be the weights of MSE, MAE, assembly success rate, and fault prediction accuracy, respectively, and The comprehensive evaluation score S = w1MSE+w2MAE+w3SR+w4FPA.

[0134] (2) Model fusion and integrated optimization

[0135] Bagging ensemble method: Use the Bagging (Bootstrap Aggregating) method to build multiple sub-models, and then average or vote the prediction results of these sub-models. The specific steps are as follows:

[0136] Perform sampling with replacement from the original training data to obtain M bootstrap sample sets.

[0137] A sub-model Mi (i=1,2,\cdots,M) is trained for each bootstrap sample set.

[0138] For new input data x, the prediction results of the M sub-models are averaged (for regression problems) or voted (for classification problems) to obtain the final prediction result.

[0139] Stacking ensemble method: The stacking method combines multiple base models of different types and then uses a meta-model to relearn the prediction results of the base models. The specific steps are as follows:

[0140] Divide the training data into K folds.

[0141] For each fold, the base model is trained using the data of the remaining K1 folds, and then predictions are made on the data of that fold to obtain the prediction results of the base model.

[0142] The prediction results of all folded base models are used as new features and together with the original training data to form a new training set for training the meta-model.

[0143] For new input data x, the base model is first used for prediction, and then the prediction results are input into the meta-model to obtain the final prediction results.

[0144] 4.4 Continuous Learning and Adaptive Adjustment of Model

[0145] (1) Incremental learning and concept drift processing

[0146] In the actual assembly process, data distribution may change due to process improvements, component updates, and other factors, resulting in concept drift. Incremental learning algorithms, such as online support vector machines (SVMs) or online neural networks (NNs), are used to continuously update the model. When concept drift is detected, the model parameters are adjusted based on the new data to adapt the model to the new data distribution.

[0147] Use a drift detection algorithm, such as the Page-Hinkley test, to detect changes in the data distribution. Let xt be the feature of the t-th sample, \barxt be the feature mean of the first t samples, and St be the cumulative deviation, which is defined as follows:

[0148] Where δ is a threshold. When St exceeds a preset threshold, it is considered that concept drift has occurred and the model needs to be updated.

[0149] (2) Model Adaptation Based on Meta-Learning

[0150] The idea of meta-learning is introduced to enable the model to learn how to quickly adapt to new assembly tasks or environments. The goal of meta-learning is to train on multiple different tasks so that the model can learn a universal learning strategy. Using the model-agnostic meta-learning (MAML) algorithm, during the training process, pre-training is first performed on multiple tasks, and then rapid fine-tuning is performed on the new task. Let the model parameters be \theta, and the loss function on the jth task be Lj(\theta), then the update formula of the MAML algorithm is:

[0151] Where α and β are learning rates, and N is the number of tasks. In this way, the model can quickly adjust parameters and improve its adaptability when encountering new assembly tasks.

[0152] (3) Model evaluation and optimization

[0153] Model evaluation metrics: Use metrics such as mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R^2) to evaluate model performance on the validation and test sets. MSE reflects the average squared error between the predicted and true values; smaller values indicate more accurate model predictions. MAE directly measures the average absolute difference between the predicted and true values. R^2 ranges from [0, 1]; closer to 1 indicates a better fit for the model.

[0154] Model optimization:

[0155] Adjust the model's structure and parameters based on the model's evaluation results on the validation set. If the model performs well on the training set but has large errors on the validation set, it may be overfitting. You can optimize the model by adding regularization terms (such as L1 or L2 regularization), reducing the number of hidden layer neurons, or using dropout techniques. If the model has large errors on both the training and validation sets, the model structure may be too simple to capture the data characteristics. Consider increasing the number of hidden layers or neurons, or replacing it with a more complex model structure. Repeat the training, evaluation, and optimization process until the model achieves good performance on both the validation and test sets.

[0156] S5. Using multiple algorithms to realize force feedback control

[0157] During the mechatronic assembly process, multiple algorithms work together to achieve precise force feedback control, ensuring high assembly precision and stability. The following is an in-depth explanation of each algorithm and the derivation of formulas.

[0158] 5.1PID Control Algorithm

[0159] The PID controller adjusts the error through three steps: proportional, integral, and derivative, and outputs a control signal. In the assembly scenario, taking assembly force control as an example, let the target assembly force be Ftarget and the actual measured assembly force be Factual. The error e(t) is:

[0160] e(t)=F tar get -F actual

[0161] The output u(t) of the PID controller consists of a proportional term, an integral term, and a differential term, and its formula is as follows:

[0162]

[0163] Kp is the proportional coefficient, used to quickly respond to errors. A larger Kp value results in a faster system response, but excessively large values can cause system oscillation. Ki is the integral coefficient, used to eliminate steady-state errors. By integrating and accumulating the error, the output is gradually adjusted to bring the system to the target value. Kd is the differential coefficient, used to predict error trends and adjust the output in advance based on the error rate of change, enhancing system stability. In practical applications, appropriate Kp, Ki, and Kd parameter values must be determined based on assembly conditions and system characteristics, either empirically or through tuning algorithms.

[0164] 5.2 Adaptive Genetic Algorithm

[0165] Adaptive genetic algorithm is used to optimize the parameters Kp, Ki, and Kd of the PID controller and to find the optimal parameter combination by simulating the biological evolution process (selection, crossover, and mutation).

[0166] (1) Encoding: Kp, Ki, and Kd are encoded to form a chromosome. Assuming binary encoding is used, each parameter is mapped to a certain number of binary bits. Each parameter is represented by a 16-bit binary number, so a chromosome consists of 16 times 3 = 48 binary bits.

[0167] (2) Fitness function: The fitness function is used to evaluate the quality of each chromosome (parameter combination) and is designed based on the error performance during the assembly process. Assuming the sampling period is T, the weight coefficients w1, w2, and w3, the fitness function J is:

[0168]

[0169] Where e(t) is the assembly force error at time t, w1 is used to weigh the importance of the current error, w2 focuses on the cumulative effect of the error, and w3 measures the smoothness of the error change. By adjusting the weight coefficients, different performance indicators can be emphasized according to actual needs.

[0170] 3. Genetic manipulation

[0171] Selection: Roulette wheel selection method is used. The probability Pi of each chromosome being selected is proportional to its fitness value Ji. The calculation formula is:

[0172]

[0173] Where N is the population size. Through this method, chromosomes with high fitness have a greater probability of being selected into the next generation.

[0174] Crossover: Perform a single-point or multi-point crossover operation with a crossover probability Pc. Assuming a single-point crossover is chosen, a crossover point is randomly selected and the gene segments of the two parent chromosomes after the crossover point are exchanged to generate two daughter chromosomes.

[0175] Mutation: Perform mutation operations on the genes on the chromosome with a mutation probability Pm, that is, randomly change the value of the binary bit to increase the diversity of the population and prevent the algorithm from falling into local optimality.

[0176] By repeating genetic operations and evolving for multiple generations, the chromosome that optimizes the fitness function is found. After decoding, the optimal Kp, Ki, and Kd parameter values are obtained and used in the PID controller.

[0177] 5.3 Fuzzy Control Algorithm

[0178] Fuzzy control algorithms convert human experience and knowledge into fuzzy rules, process imprecise inputs, and output precise control quantities. In assembly force feedback control, the fuzzy control process is as follows:

[0179] 1. Fuzzification: Convert the input error e and error change rate ec into fuzzy linguistic variables. Define the fuzzy sets of error e and error change rate ec as {NB, NM, NS, ZO, PS, PM, PB}, which represent negative large, negative medium, negative small, zero, positive small, positive medium, and positive large, respectively. Let the domain of error e be [E, E] and the domain of error change rate ec be [EC, EC], and calculate the degree to which the input value belongs to each fuzzy set through the membership function μ(x). Using the triangular membership function, for the fuzzy set NS (negative small), its membership function is:

[0180]

[0181] Where a and b are parameters set according to the domain.

[0182] 2. Fuzzy rule base: Fuzzy control rules are formulated based on assembly experience and expert knowledge and expressed in the form of "ifthen".

[0183] if e is NB and ec is NB then u is PB if e is ZO and ec is PS then uis NS

[0184] These rules describe the control strategies that should be adopted under different combinations of errors and error change rates.

[0185] 3. Fuzzy Inference: Based on the fuzzified input and the fuzzy rule base, the Mamdani inference method is used for inference. For each rule, the matching degree of its antecedent is calculated (i.e., the minimum value of the membership of the input value to the fuzzy set of the rule's antecedent). The fuzzy set of the rule's consequent is then truncated based on this matching degree to obtain the output fuzzy set of each rule. Finally, the output fuzzy sets of all rules are combined to obtain the total output fuzzy set.

[0186] 4. Defuzzification: Convert the total output fuzzy set into an accurate control quantity u. The centroid method is used for defuzzification, and the calculation formula is:

[0187]

[0188] Where xi is a discrete point in the output universe, μ(xi) is the membership degree corresponding to the point, and n is the number of discrete points. The u obtained by defuzzification is used as the adjustment signal for assembly force control.

[0189] 5.4 Algorithm Collaboration Mechanism

[0190] In actual assembly, the PID control algorithm provides basic control adjustments, the adaptive genetic algorithm periodically optimizes PID parameters to adapt to different assembly conditions, and the fuzzy control algorithm makes rapid control decisions based on experience in complex and uncertain assembly environments. These three algorithms work together through data interaction and decision fusion. When errors are large in the initial stages of assembly, the fuzzy control algorithm prioritizes outputting larger control variables for rapid adjustments. As errors decrease, the PID control algorithm takes effect, making precise adjustments. Simultaneously, the adaptive genetic algorithm optimizes PID parameters based on the error manifestations during assembly, ensuring efficient and accurate assembly throughout the entire process.

[0191] S6. Execute assembly operations and quality inspection: Based on the control signals output by the force feedback control algorithm, the assembly tool is driven to perform assembly actions. Sensor data is monitored in real time to adjust the assembly process. After assembly is completed, sensor data is used for quality inspection. If it fails, feedback is provided to the control system, and parameters are readjusted for corrective assembly.

[0192] Experimental verification

[0193] 1. Verify the technical effects of the mechatronic assembly solution based on force feedback control in improving assembly accuracy, increasing production efficiency, and expanding adaptability to application scenarios. Compare traditional assembly methods with existing automated assembly technologies to highlight the advantages of this solution.

[0194] Representative electromechanical product components were selected for assembly experiments, including precision electronic components (such as the assembly of chips and circuit boards), small mechanical components (such as shaft-hole assembly), and electromechanical modules of medium complexity (such as the assembly of small motors), corresponding to scenarios with different precision requirements and assembly complexity.

[0195] The experimental group of this scheme: uses a mechatronic assembly method and system based on force feedback control for assembly.

[0196] Traditional assembly control group: experienced assembly workers use manual experience and simple mechanical positioning tools to perform assembly.

[0197] Existing automated assembly control group: Assembly was performed using automated assembly equipment on the market that lacked precise force feedback control.

[0198] 3. Unify the experimental site, ensure consistency of environmental conditions such as temperature, humidity, and light, and record environmental parameters.

[0199] Each experimental group was equipped with assembly parts, tools, and consumables of the same specifications.

[0200] (1) Start the mechatronic assembly system based on force feedback control. The system collects assembly process data in real time through force sensors, position sensors and vision sensors.

[0201] After being filtered, normalized and fused by the preprocessing module, the data is input into the comprehensive model for calculation. The model combines multiple algorithms (PID control algorithm, adaptive genetic algorithm, fuzzy control algorithm, etc.) to output control signals.

[0202] The execution module drives the assembly tool to complete the assembly operation according to the control signal. During the assembly process, the system monitors the data in real time and dynamically adjusts the assembly action.

[0203] After the assembly is completed, the detection and feedback module will detect the assembly quality. If it is unqualified, it will be fed back to the control system for correction of the assembly.

[0204] (2) Traditional assembly control group operation

[0205] Assembly workers use simple mechanical positioning tools based on their own experience, and rely on vision and feel to assemble parts.

[0206] During the assembly process, workers judge whether the assembly force and position are appropriate. After the assembly is completed, manual quality inspection is carried out. If it fails, rework is required.

[0207] (3) Operation of existing automated assembly control group

[0208] When the existing automated assembly equipment is started, it performs assembly operations according to preset programs and lacks real-time force feedback adjustment functions.

[0209] After assembly is completed, quality inspection is carried out using conventional testing equipment. If it fails to meet the standards, the machine will be shut down for manual adjustment and reassembly.

[0210] 4. Data Collection

[0211] (1) Assembly accuracy related data

[0212] Use high-precision measuring instruments (such as three-coordinate measuring machines, laser rangefinders, etc.) to measure the position deviation, dimensional accuracy, fit clearance and other data of assembly parts.

[0213] Record the position error of each component in the x, y, and z axis directions as well as the rotation angle error.

[0214] (2) Production efficiency related data

[0215] Record the time it takes to complete the assembly of a single component in each group of experiments, and count the number of assemblies per unit time.

[0216] Record the rework time and downtime adjustment time caused by assembly quality issues.

[0217] (3) Other data

[0218] Record the assembly force data during the assembly process, including the force magnitude and direction change curve.

[0219] Record the loss function value, parameter adjustment of each algorithm and other data during the comprehensive model training process.

[0220] Record the assembly performance data of different experimental groups when facing individual differences in components (such as slight deviations in component dimensions) and environmental changes (such as temperature fluctuations).

[0221] V. Data Analysis and Verification

[0222] (1) Assembly accuracy analysis

[0223] Calculate the mean and standard of the assembly error of each group of experiments, draw the error distribution histogram and box plot, compare the differences in assembly accuracy among the three groups of experiments, and analyze the advantages of this scheme in controlling assembly errors.

[0224] Hypothesis testing was used to verify whether there was a significant difference in assembly accuracy between the experimental group of this scheme and other control groups.

[0225] (2) Production efficiency analysis

[0226] Calculate the average assembly time, assembly quantity per unit time and rework rate for each group of experiments.

[0227] Draw a production efficiency comparison bar chart to intuitively display the production efficiency differences among the three groups of experiments and analyze the effectiveness of this solution in improving production efficiency.

[0228] Perform analysis of variance (ANOVA) to determine whether there are significant differences in production efficiency indicators between the experimental group of this scheme and other control groups.

[0229] (3) Comprehensive performance analysis

[0230] Combined with assembly accuracy and production efficiency data, comprehensive performance indicators are calculated, including the number of qualified products per unit time, to comprehensively evaluate the performance of each solution.

[0231] Analyze the adaptability of this solution and other control groups under different experimental conditions (individual differences in components, environmental changes) to verify the technical effect of this solution in expanding application scenarios.

[0232] Analyze the comprehensive model training data and algorithm parameter adjustment data to verify the effectiveness and stability of the model and algorithm.

[0233] As shown in Table 1,

[0234]

[0235] By comparing Table 1, it can be clearly and intuitively seen that the mechatronic assembly solution based on force feedback control has significant advantages over traditional assembly methods and existing automated assembly technologies in terms of assembly accuracy, production efficiency and comprehensive performance, which strongly verifies the good technical effect of the solution.

[0236] like Figure 2 As shown in the figure, from the box plot of the qualified rate of the fit clearance, the experimental group of this scheme has a short box, a high median close to the maximum value, and a short whisker, indicating a high qualified rate and concentrated data; the traditional assembly control group has a long box and whisker, and a low median, indicating a low qualified rate and discrete data; the existing automated assembly control group has a moderate length of box and whisker, and the median and dispersion are also at an intermediate level.

[0237] In summary, through experiments and data analysis, it is expected that the experimental group of this solution will achieve significantly lower mean assembly accuracy errors than both the traditional assembly control group and the existing automated assembly control group. In terms of production efficiency, the experimental group will achieve shorter average assembly time, a higher number of assemblies per unit time, and a lower rework rate. In terms of overall performance and adaptability to application scenarios, the experimental group will also significantly outperform the other two groups. This demonstrates that the mechatronic assembly solution based on force feedback control has excellent technical performance and can meet the requirements of high-precision, high-efficiency, and highly adaptable mechatronic assembly.

[0238] Example 2

[0239] This embodiment provides a mechatronic assembly system based on force feedback control, comprising:

[0240] The data acquisition module is configured as

[0241] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for implementing a mechatronics assembly method based on force feedback control.

[0242] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a mechatronic assembly method based on force feedback control.

[0243] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A mechatronic assembly method based on force feedback control, characterized in that: include: Acquire mechatronic assembly sensor data; Perform data preprocessing on the acquired mechatronic assembly sensor data; Build a multivariate nonlinear prediction model and train the model using the preprocessed dataset; Generate control signals based on adaptive genetic and fuzzy control algorithms; Realize mechatronic assembly based on control system.

2. The mechatronic assembly method based on force feedback control according to claim 1, characterized in that: The data preprocessing of the acquired mechatronic assembly sensor data includes adopting a mean filtering algorithm. For a sequence x1, x2, ..., xn containing n data, the filtered data yi is expressed as: Among them, m is the filter window size, and this algorithm is used to remove random noise in the data.

3. The mechatronic assembly method based on force feedback control according to claim 2, characterized in that: The method comprises constructing a multivariate nonlinear prediction model and training the model using a preprocessed data set, including determining key variables in the assembly process based on preprocessed mechatronic assembly sensor data, describing the dynamic coupling relationship between the variables through a state space model, constructing the mutual relationship between the key variables through prior knowledge and mechanisms, and introducing a multilayer perceptron (MLP). The key variables of the key variables are constructed using prior knowledge and mechanisms as input, and the predicted values of the output layer are obtained through nonlinear transformation of several hidden layers, wherein the number of neurons in the output layer is determined according to the dimension of the predicted values.

4. The mechatronic assembly method based on force feedback control according to claim 3, characterized in that: The method of constructing a multivariate nonlinear prediction model and training the model using a preprocessed data set also includes feature engineering on the input layer, calculating derived features in addition to directly using the original sensor data, including calculating the rate of change of assembly force and acceleration of position; in the output layer, in addition to predicting the position posture P and assembly force F of the component, additional evaluation indicators of assembly quality are output, including assembly accuracy and assembly reliability. Through multi-task learning, the model can learn multiple related tasks at the same time, thereby improving the generalization ability and overall performance of the model; in the hidden layer, a hierarchical architecture design is adopted, and hidden layers at different levels are responsible for extracting features at different levels of abstraction, among which the bottom hidden layer extracts local and detailed features, and the high-level hidden layer extracts global and abstract features.

5. The mechatronic assembly method based on force feedback control according to claim 4, characterized in that: The method comprises constructing a multivariate nonlinear prediction model and training the model using a preprocessed data set, comprising dividing the acquired data into a training set, a validation set and a test set using a stratified sampling partitioning and a time series data partitioning method, and training the multivariate nonlinear prediction model by using a back propagation BP algorithm, wherein the error of the output is calculated and the error is back propagated to the hidden layer and the input layer, and then the bias of the connection weights between neurons is adjusted according to the error by the chain rule.

6. The mechatronic assembly method based on force feedback control according to claim 5, characterized in that: The multivariable nonlinear prediction model is constructed and trained using the preprocessed data set, and further includes using a combination of particle swarm optimization and genetic algorithm to quickly search for the approximate area of the global optimal solution through information sharing and iterative updates between particles; and using the genetic algorithm to perform a fine search in the local area through selection, crossover, and mutation operations. The update formula of the particle swarm optimization algorithm is: v i (t+1)=ωv i (t)+c1r1(p best,i -x i (t))+c2r2(g best -x i (t)) x i (t+1)=x i (t)+v i (t+1) where vi(t) is the velocity of the ith particle at time t, xi(t) is the position of the ith particle at time t, \omega is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, and p best,i is the historical optimal position of the i-th particle, g best is the global optimal position.

7. The mechatronic assembly method based on force feedback control according to claim 6, characterized in that: The multivariate nonlinear prediction model is constructed and trained using the preprocessed data set, and the simulated annealing method is introduced to accept inferior solutions with a certain probability, thereby preventing the algorithm from falling into a local optimum. Assuming the current parameter is θ, the updated parameter is θ', the loss function is L(θ), and the temperature is T, the probability of accepting the inferior solution is: As the training progresses, the temperature T gradually decreases and the algorithm gradually converges to the global optimal solution.

8. The mechatronic assembly method based on force feedback control according to claim 7, characterized in that: The multivariate nonlinear prediction model is constructed and trained using the preprocessed dataset, and further includes constructing a reinforcement learning environment, treating the assembly process as a Markov decision process, and training the model using a deep deterministic policy gradient method, wherein the goal of the actor network is to maximize the cumulative reward, and its update formula is: where θ μ are the parameters of the actor network, θ Q are the parameters of the critic network, Q(s, a|θ Q ) is the action value output by the critic network, μ(s|θ μ ) is the action output by the actor network.

9. The mechatronic assembly method based on force feedback control according to claim 8, characterized in that: The control signal generation method based on the adaptive genetic and fuzzy control algorithm includes optimizing the parameters Kp, Ki, and Kd of the PID controller by using the adaptive genetic algorithm and finding the optimal parameter combination by simulating the biological evolution process. Then, the fuzzy control algorithm is used to convert the input error e and the error change rate ec into fuzzy linguistic variables. The fuzzy control rules are formulated according to assembly experience and expert knowledge. The Mamdani reasoning method is used to perform reasoning based on the input fuzzification results and the fuzzy rule base. Finally, the total output fuzzy set is converted into a precise control variable.

10. A mechatronic assembly system based on force feedback control, characterized in that: include: A data acquisition module is configured to acquire mechatronic assembly sensor data; A preprocessing module is configured to perform data preprocessing on the acquired mechatronic assembly sensor data; The training module is configured to construct a multivariate nonlinear prediction model and train the model using the preprocessed data set; A control module configured to generate a control signal based on an adaptive genetic and fuzzy control algorithm; Realize mechatronic assembly based on control system.

Images