A physical information neural network-based automatic drilling and riveting equipment riveting quality prediction method

Abstract

The application discloses a kind of automatic drilling and riveting equipment pressure riveting quality prediction methods based on physical information neural network, comprising: obtaining production data, based on processing quality to production data is labeled, and label and production data are formed data set;Based on the error equation is constructed, and based on the error equation, initial model is constructed;Utilize data set and based on segmented constraint modeling strategy, initial model is trained to obtain quality evaluation model for evaluating processing quality.The method provided by the application can solve the problems of insufficient physical consistency of pure data-driven model, poor adaptability to equipment degradation state, and difficulty in balancing prediction accuracy and generalization ability in existing automatic drilling and riveting equipment pressure riveting quality prediction methods.

Description

Technical Field

[0001] This invention belongs to the field of intelligent aircraft assembly, and particularly relates to a method for predicting the riveting quality of automatic drilling and riveting equipment based on physical information neural networks. Background Technology

[0002] As a type of aerospace process equipment, automated drilling and riveting equipment, despite significant progress in data-driven machining quality prediction models in recent years, still faces major challenges in its practical application within the aerospace manufacturing field. Traditional data-driven algorithms relying on statistical correlation and black-box learning struggle to meet the stringent requirements of reliability, robustness, and applicability across all operating conditions for aerospace process equipment. In the aerospace manufacturing sector, existing data-driven machining quality prediction models for process equipment still face a series of pressing problems, primarily in the following two aspects: First, the core issue lies in the opaque "black box" nature of the models, resulting in severely insufficient generalization of predictions for machining quality across process clusters, making it difficult to meet the stringent safety and reliability requirements of the aerospace industry. Due to the lack of interpretable evidence at the process mechanism level, these methods often rely on empirical experiments rather than theoretical understanding to continuously adjust and optimize the model, leading to high trial-and-error costs and poor repeatability. Second, black-box models generally rely on a large number of labeled samples to maintain stable performance, but the scale of available labeled data in aerospace manufacturing scenarios is limited and difficult to cover complex operating conditions. First, aerospace manufacturing processes are highly controlled, prohibiting arbitrary changes to processing parameters or manufacturing conditions for data collection, which severely limits the scope of experimental design. Second, obtaining key quality indicators often relies on destructive testing, offline measurement, or high-cost testing processes, resulting in long and costly label generation cycles, further reducing the amount of usable data. More importantly, because high-consistency manufacturing is the industry's baseline, normal samples far outnumber faulty or abnormal samples, leading to a severely unbalanced data distribution. Black-box models are highly susceptible to overfitting and degradation of generalization ability, severely restricting their practical application.

[0003] Patent document CN114330030A discloses a method for controlling the quality of drilling and riveting of aircraft structural parts based on digital twins. The method involves installing robotic automatic drilling and riveting equipment in the aircraft assembly physical workshop. This equipment can complete the entire process of hole positioning, hole drilling, rivet feeding, and riveting. At the same time, a digital twin system corresponding to the physical workshop is established on a computer. The automatic drilling and riveting equipment transmits the center coordinates of the completed riveting holes and riveting quality information to the digital twin system in real time via a local area network and TCP / IP data interface. Through real-time integration and interaction between the physical workshop and the digital twin system, problematic riveting points can be detected and remedied in a timely manner, effectively avoiding problems such as lagging riveting status monitoring and cumulative riveting errors.

[0004] Patent document CN117909924A discloses a method, system, medium, and product for predicting the roughness of machined parts based on physical information machine learning. The method includes preprocessing multi-source time-series data of machined workpieces to obtain preprocessed multi-source time-series data; inputting machining parameters into a surface roughness mechanism model with tolerable accuracy to obtain theoretical values ​​of surface roughness of the machined workpiece; inputting the preprocessed multi-source time-series data and theoretical values ​​of surface roughness into a spatial feature prediction model of surface roughness to obtain spatial predicted values; inputting the preprocessed multi-source time-series data and theoretical values ​​of surface roughness into a temporal feature prediction model of surface roughness to obtain temporal predicted values; and aggregating the spatial and temporal predicted values ​​based on an attention mechanism to obtain the predicted surface roughness of the machined workpiece. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network. This method can solve the problems of insufficient physical consistency of pure data-driven models, poor adaptability to equipment degradation, and difficulty in balancing prediction accuracy and generalization ability in existing methods for predicting the riveting quality of automatic drilling and riveting equipment.

[0006] To achieve the objectives of this invention, the following technical solution is provided: a method for predicting the riveting quality of an automatic drilling and riveting equipment based on a physical information neural network, comprising the following steps: Acquire production data, label the production data based on processing quality, and combine the labels and production data into a dataset; An error equation is constructed based on the processing error, and an initial model is constructed based on the error equation. The initial model is trained using the dataset and a segmented constraint modeling strategy to obtain a quality assessment model for evaluating processing quality.

[0007] Based on real-time production data from automated drilling and riveting equipment, this invention introduces the error mechanism and mechanical constraints during the operation of the process equipment into the processing quality prediction model. Through a modeling approach that integrates data-driven and physical mechanisms, it achieves high-precision prediction of riveting quality evaluation indicators.

[0008] Specifically, the production data includes the power, current, load, speed, and length gauge reading of the tool spindle during the hole-making process, and the pressure of the upper and lower augers and the pressure of the presser foot during the riveting process.

[0009] Specifically, the processing quality refers to the amount of riveting interference after press riveting.

[0010] Specifically, the processing error includes micro-errors and macro-errors; The micro-errors include the evolution mechanism of mechanical behavior caused by changes in material geometry and microstructure; The macroscopic error includes the error mechanism caused by the decline in system processing capability due to performance degradation during the operation of process equipment.

[0011] Specifically, the expression for the error equation is as follows: ; in, This represents the overall error. This represents the set of error terms in the microscopic model. Indicates the error term; This represents the set of error terms in the macroscopic model. Indicates the error term; , Indicates the weight.

[0012] Specifically, the segmented constraint modeling strategy includes the early stage of equipment operation and the equipment usage stage; The processing error in the early stage of equipment operation lies in the operating parameters and / or material properties; The processing error during the equipment's use phase lies in the amplification of the transmission chain gap and / or the accumulation of measurement chain errors.

[0013] Specifically, the initial model includes: The algorithm input module is used to receive input production data; The feature extraction module is used to extract potential features from production data; The core prediction module is used to capture the dependencies between feature sequences and output the final results.

[0014] Specifically, during the training process, a loss function is used to train the model, and the loss function includes data consistency constraints and physical consistency constraints.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: In the process of riveting quality prediction, error modeling theory and mechanical constraints oriented towards process equipment are introduced, enabling the prediction model to simultaneously meet the requirements of data regularity and physical consistency. The prediction accuracy and generalization ability are superior to traditional pure data-driven methods. After the model is trained, the riveting quality prediction only involves the forward inference of the neural network, which has high computational efficiency and the ability to perform real-time prediction and online evaluation in the production site of automatic drilling and riveting equipment. Through a phased constraint modeling strategy, the model can adapt to the entire process of equipment operation from healthy to deteriorating, improving the stability of prediction results under long-term operating conditions. At the same time, the error model retains micro and macro error weight information, enhancing the interpretability of the model and providing effective support for riveting process optimization and equipment condition analysis. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for predicting the riveting quality of an automatic drilling and riveting equipment based on a physical information neural network, provided in this embodiment. Figure 2 This is a schematic diagram of the model provided in this embodiment; Figure 3 This is a schematic diagram of the loss function provided in this embodiment; Figure 4 This is a schematic diagram of the loss curve for model testing provided in this embodiment; Figure 5 This is a schematic diagram of the mean absolute error of the model test provided in this embodiment; Figure 6 This is a schematic diagram of the mean absolute percentage error of the model test provided in this embodiment; Figure 7 This is a schematic diagram of the mean square error of the model test provided in this embodiment; Figure 8 This is a schematic diagram of the root mean square error of the model test provided in this embodiment; Figure 9 This is a schematic diagram illustrating the physical consistency of the model test provided in this embodiment; Figure 10 This is a schematic diagram of the loss weights for model testing provided in this embodiment; Figure 11 This is a schematic diagram of the physical weights for the model test provided in this embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown in this embodiment, a method for predicting the riveting quality of an automatic drilling and riveting equipment based on a physical information neural network is provided, including the following steps: Acquire production data, label the production data based on processing quality, and combine the labels and production data into a dataset; An error equation is constructed based on the processing error, and an initial model is constructed based on the error equation. The initial model is trained using the dataset and a segmented constraint modeling strategy to obtain a quality assessment model for evaluating processing quality.

[0019] More specifically, the processing error includes micro-errors and macro-errors, and the micro-errors include the mechanical behavior evolution mechanism caused by changes in material geometry and microstructure; The macroscopic error includes the error mechanism caused by the decline in system processing capability due to performance degradation during the operation of process equipment.

[0020] Each process in aircraft assembly corresponds to a complex nonlinear mathematical-mechanical mechanism, essentially reflecting a multi-field coupling mechanism between materials, forces, deformation, and energy transfer. Under different operating conditions, operations such as clamping, drilling, and riveting lead to the continuous evolution of material geometry, stress distribution, and microstructure, thereby inducing a series of nonlinear responses such as elastoplastic deformation, residual stress accumulation, and local damage. These evolutionary processes are not only highly dependent on process parameters (such as clamping / riveting / drilling forces), but also exhibit significant temporal, path-dependent, and structural heterogeneity, making the mathematical expression of the process highly nonlinear, strongly coupled, and highly complex. Therefore, the microscopic model encompasses the entire mathematical-mechanical mechanism of a single process, with the core objective of revealing and quantifying the influence of these complex mechanisms on assembly quality. That is, by describing key processes such as material deformation, interface contact, and friction and wear, an interpretable mapping from process parameters to the final formed quality can be achieved. Specifically, microscopic models can not only characterize the local mechanical behavior in the process, but also provide physical consistency constraints and traceable causal links for subsequent processing prediction models, thereby constructing a causal relationship and mechanism framework between "materials-processes-equipment".

[0021] Unlike the microscopic mechanical behavior of materials, the macroscopic machining quality is mainly affected by the performance degradation of process equipment and changes in functional coupling relationships. During long-term operation, aerospace process equipment inevitably experiences wear, loosening, and thermal drift in its transmission, actuation, and other functional components, leading to motion instability and other degradation phenomena, introducing a systematic source of machining errors. Therefore, the key to constructing a macroscopic machining error mechanism model lies in starting from the overall equipment structure and characterizing the coupling mechanism and error transmission path of each functional part within the equipment. First, the spatial dependence and dynamic correlation between functional parts are confirmed through a hierarchical structural model of the equipment design coupling relationship. For example, the hierarchical relationship consisting of the spindle-feed-clamping-measuring system not only clarifies the level of action of different modal errors but also reveals the transmission path and amplification law of errors in the structural topology. This hierarchical structural model provides a traceable theoretical model for identifying the equipment error chain. Second, based on this, it is necessary to further analyze the functional degradation mechanism of key machining parts. Different functional components exhibit different types and forms of degradation; for example, the power actuator may experience displacement lag due to the gap between the lead screw and the guide rail. Degradation behavior, exhibiting nonlinearity, cumulativeity, and unobservable characteristics under dynamic operating conditions, is a major driving force behind the formation and evolution of macroscopic processing errors. Based on this, by integrating a hierarchical structure model with functional degradation mechanism analysis, the parameter chain influencing processing errors can be ultimately determined—a causal chain composed of functional degradation, process operating parameters, process dynamic response, and final quality indicators. This parameter chain essentially reflects the causal process of "equipment state evolution—error generation—error propagation—processing quality deviation," enabling an interpretable mapping of complex equipment from performance degradation to processing errors and then to the final quality response. This provides a systematic and physically consistent input dimension constraint for quality prediction models.

[0022] like Figure 2 The diagram shown is a schematic representation of the model provided in this embodiment, and its specific details are as follows: For the algorithm input module, the original sample contains 4 feature parameters, which are reconstructed into a sample of size after standardization. The tensor, in which This indicates the batch size, where 1 represents the number of input channels and 4 represents the feature dimension, to meet the input requirements of a one-dimensional convolutional network.

[0023] For the feature extraction module, the convolutional feature extraction stage employs a single-layer one-dimensional convolutional network with 19 kernels, a kernel size of 3, and a stride of 1. A padding strategy is used at the boundaries to maintain the feature length. After the convolution operation, a ReLU activation function is introduced to enhance non-linear expressiveness. Subsequently, a max-pooling layer (pooling window size of 2) is used to downsample the feature sequence, thereby reducing model complexity and suppressing local noise interference. After processing by this module, the feature tensor size is reduced from... Convert to To meet the input format requirements of recurrent neural networks, the convolutional output features are adjusted through a dimension rearrangement operation. , where 2 represents the time step length and 19 is the feature dimension of each time step.

[0024] For the core prediction module, the sequence features are input into a Bidirectional Long Short-Term Memory (BiLSTM) network for temporal modeling. The BiLSTM network consists of two stacked layers, each with 35 hidden units, and employs a bidirectional structure to simultaneously capture the positive and negative dependencies of the feature sequences. After BiLSTM encoding, the network outputs features with a dimension of [missing information]. 70% of the values ​​are derived from the concatenation of the forward and backward hidden states. The output of the BiLSTM at the last time step is selected as the global feature representation, and it is projected onto the one-dimensional output space through a fully connected linear mapping layer to obtain the prediction result of the interference quantity. During model training, a multi-objective joint loss function is used, including data error terms, relative error terms, physical consistency terms, and analytical supervision terms. A dynamic annealing strategy is employed to gradually increase the weight of physical constraints in the total loss, thereby effectively improving the model's stability and engineering applicability while ensuring fitting accuracy.

[0025] like Figure 3 The diagram shown illustrates the loss function provided in this embodiment. The overall loss in this embodiment is defined as follows: ; ; in, These are the weights of each loss term. It is the data fitting loss, which physically requires the network to predict the true interferometric amount at each measurement point in the training set. Its function is to reflect the degradation state of the actual equipment, and it is defined as: ; This is the relative error loss. Since the evaluation index for the interference amount corresponds to a range, the data fitting loss is more sensitive to larger values, while small interference amounts contribute almost nothing to training. This leads to training biased towards optimizing riveting conditions with large rivets (i.e., larger interference values), while predictions for riveting conditions with small rivets (i.e., smaller interference values) may exhibit varying degrees of distortion. Therefore, the function is to normalize the error for each condition according to its amplitude, avoiding model bias towards certain conditions, defined as: ; in, This is a numerical stabilization parameter used to avoid numerical singularities in the relative error term when the interferometric quantity approaches zero. This parameter does not introduce additional physical assumptions, its value is much smaller than the typical scale of the interferometric quantity, and it has no significant impact on the direction of model optimization or the final prediction results. This is a loss due to mechanical constraints. Axial riveting force. The analytical expression originates from the elastoplastic mechanical equilibrium relationship during the riveting process. Based on this, a global mechanical consistency loss is constructed, and the axial force is derived from the displacement field output by the network through numerical integration and compared with the theoretical calculation value. This term forces the network to satisfy the overall mechanical conservation, preventing the learned function from having solutions that violate physical laws, thereby significantly improving the stability and physical reliability of the prediction results, and is defined as: ; in, It is the derivative of the riveting force with respect to the radial coordinate. This refers to displacement field constraints. Since unconstrained neural networks may exhibit smooth but incorrectly shaped displacement curves or physically uninterpretable field patterns, the primary function of these constraints is to constrain the implicit displacement distribution shape of the neural network, reducing the feasible solution space. This constraint is defined as: ; like Figures 4 to 11 As shown, this embodiment presents various prediction performance test results for the model. The data points are roughly symmetrically distributed along both sides of the diagonal. First, as... Figure 4 As shown in the figure, the overall loss function evolution curve (LOSS) reveals that the total model loss experiences a brief fluctuation in the early stages of training before rapidly decreasing and gradually stabilizing in the later stages. This indicates that the constructed multi-objective joint loss function can be effectively optimized under the control of the dynamic annealing strategy, without gradient oscillations or divergence. Furthermore, as... Figure 7 As shown, the data-driven mean squared error (MSE) loss exhibits a monotonically rapid decreasing trend and converges to a low level after about 100 epochs.

[0026] In terms of test performance, such as Figure 5 , Figure 6 as well as Figure 8 As shown, the model's MAE, RMSE, and MAPE curves show a certain degree of increase and fluctuation in the early stages of training as the physical constraint weights gradually increase, but they stabilize within a relatively narrow range in the middle and later stages and do not continuously deteriorate as training progresses. Figure 9As shown, this phenomenon indicates that as physical consistency constraints are gradually introduced, the model proactively sacrifices some pure data fitting accuracy in exchange for the physical rationality and generalization stability of the prediction results, demonstrating an effective balance between data-driven objectives and physical prior constraints. It is noteworthy that no significant oscillations or divergences were observed in the later stages of the tests, indicating that the adopted continuous annealing weight scheduling strategy has good stability in multi-loss coupled optimization scenarios.

[0027] Furthermore, such as Figure 10 As shown, the dynamic evolution curve of the loss weights clearly demonstrates the mechanism of the proposed annealing mechanism. Among them, the data-driven MSE loss weights maintain a dominant position throughout the training process. The relative error loss and the weight of the physical consistency loss gradually increase with the training progress and eventually tend to saturate. This allows the model to prioritize data fitting in the early stage of training and gradually strengthen the influence of physical constraints in the middle and later stages, thereby avoiding excessive interference from physical terms to network parameter updates in the early stage.

[0028] From the perspective of physical consistency, such as Figure 11 The physical residuals shown continuously decreased during training and converged to a low level in the later stages, indicating that the model's predictions gradually satisfied the introduced global mechanical constraints without any physical residual bounce or instability. This result verifies that the proposed method still possesses good trainability and convergence after the introduction of physical constraints. Meanwhile, the learned physical parameters all exhibited smooth, monotonic convergence behavior, and no drastic oscillations or numerical divergence occurred throughout the training process, demonstrating that the model can stably identify the implicit physical parameters under the combined effect of data and physical constraints, exhibiting good physical interpretability.

[0029] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0030] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0031] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the riveting quality of an automatic drilling and riveting equipment based on a physical information neural network, characterized in that, Includes the following steps: Acquire production data, label the production data based on processing quality, and combine the labels and production data into a dataset; An error equation is constructed based on the processing error, and an initial model is constructed based on the error equation. The initial model is trained using the dataset and a segmented constraint modeling strategy to obtain a quality assessment model for evaluating processing quality.

2. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The production data includes the power, current, load, speed, and length gauge readings of the tool spindle during the hole-making process, and the pressure of the upper and lower augers and the pressure foot during the riveting process.

3. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The processing quality refers to the amount of riveting interference after press riveting.

4. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The processing error includes micro-error and macro-error; The micro-errors include the evolution mechanism of mechanical behavior caused by changes in material geometry and microstructure; The macroscopic error includes the error mechanism caused by the decline in system processing capability due to performance degradation during the operation of process equipment.

5. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The expression for the error equation is as follows: ； in, This represents the overall error. This represents the set of error terms in the microscopic model. Indicates the error term; This represents the set of error terms in the macroscopic model. Indicates the error term; , Indicates the weight.

6. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The segmented constraint modeling strategy includes the early stage of equipment operation and the equipment usage stage; The processing error in the early stage of equipment operation lies in the operating parameters and / or material properties; The processing error during the equipment's use phase lies in the amplification of the transmission chain gap and / or the accumulation of measurement chain errors.

7. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, The initial model includes: The algorithm input module is used to receive input production data; The feature extraction module is used to extract potential features from production data; The core prediction module is used to capture the dependencies between feature sequences and output the final results.

8. The method for predicting the riveting quality of automatic drilling and riveting equipment based on a physical information neural network according to claim 1, characterized in that, During training, a loss function is used to train the model, which includes data consistency constraints and physical consistency constraints.

Images