A thin-wall special-shaped cavity shaft full-process deformation closed-loop control method

Abstract

The present application relates to the technical field of thin-walled special-shaped part precision machining, and discloses a thin-walled special-shaped cavity shaft full-process deformation closed-loop control method, original deformation observation data at the current time is generated, the original deformation observation data is fused with the deformation state parameters at the end of the last process by Kalman filtering, the thin-walled special-shaped cavity shaft full-section deformation field distribution data under the current process is reconstructed, a mixed model of a fusion time cycle network module and a graph space network module is input, and a predicted deformation state vector at the end of the current process is output; the process parameter compensation amount that makes the deviation sum of squares minimum is solved on the current remaining machining path, and the optimal compensation increment at the current time is output; the optimal compensation increment is converted into a control instruction and applied in real time to the machining process of the current process, and the execution is repeated until all processes are completed; the present application realizes accurate prediction and closed-loop control in the thin-walled special-shaped cavity shaft full-process manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of precision machining technology for thin-walled irregular-shaped parts, specifically to a closed-loop control method for deformation of thin-walled irregular-shaped cavity shafts throughout the entire process. Background Technology

[0002] Thin-walled irregular cavity shafts are a type of thin-walled rotating parts with complex internal cavity structures and irregular cross-sections. Due to their thin walls, low stiffness, and complex structure, they are prone to deformation during the manufacturing process of forging, heat treatment, and machining, resulting in high dimensional deviations and scrap rates. Current processing technologies are unable to assess the overall evolution trend of deformation, accurately capture the collaborative evolution law of deformation in the spatiotemporal dimension, store the deformation state of each process in isolation, and cannot use the results of previous processes as the initial state input for subsequent processes, thus failing to form a closed-loop control chain for the entire process. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a closed-loop control method for deformation of thin-walled irregular cavity shafts throughout the entire process.

[0004] The first aspect of this invention provides a closed-loop control method for the deformation of a thin-walled irregular cavity shaft throughout its entire process, the method comprising the following steps: During the current process, the strain signal of the inner wall of the thin-walled irregular cavity and the displacement signal of the outer wall are collected in real time, and the strain signal and displacement signal are synchronized in time and matched in space coordinates to generate the original deformation observation data at the current moment. The original deformation observation data is fused with the deformation state parameters at the end of the previous process by Kalman filtering to reconstruct the full cross-sectional deformation field distribution data of the thin-walled irregular cavity shaft under the current process. The process parameter sequence of the current process, the full cross-section deformation field distribution data, and the deformation state sequence of the historical processes are input into a hybrid model that integrates the time cyclic network module and the graph space network module, and the predicted deformation state vector at the end of the current process is output. Using the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, the process parameter compensation amount that minimizes the sum of squared deviations is solved on the current remaining processing path, and the optimal compensation increment at the current moment is output. The optimal compensation increment is converted into control commands and applied in real time to the current processing step, and repeated until all steps are completed.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the original deformation observation data and the deformation state parameters at the end of the previous process are fused using Kalman filtering to reconstruct the deformation field distribution data of the entire cross-section of the thin-walled irregular cavity shaft under the current process, including: Obtain the deformation state parameter vector passed at the end of the previous process and use it as the initial state estimate of the Kalman filter algorithm for the current process; Obtain the original deformation observation data at the current moment to form the observation vector at the current moment; The state estimate from the previous time step is recursively applied to the current time step through the state transition matrix to obtain the prior state estimate for the current time step. At the same time, the prior estimated covariance matrix is ​​updated based on the process noise covariance matrix. Calculate the Kalman gain matrix, and then weight and fuse the observation vector at the current time with the state prior estimate to obtain the optimal state estimate at the current time. Update the covariance matrix of the optimal state estimate, use the optimal state estimate at each section as the interpolation node, and use the radial basis function interpolation method to expand the deformation data at the discrete sections into a full-section deformation field distribution data that is continuously distributed along the axis of the part, and output the reconstructed full-section deformation field at the current moment.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the deformation state parameter vector includes the radial and axial deformation components of each cross section of the thin-walled irregular cavity shaft, and the original deformation observation data includes the strain readings of each strain sensing node of the inner cavity wall and the displacement readings of the laser displacement sensor of the outer wall.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the processing procedure of the time loop network module includes: The process parameters time series of the current process and the deformation state series of the historical processes are input into the time loop network module, which includes two long short-term memory network branches. Based on the time series of process parameters of the current process, a long short-term memory network branch is used to read the process parameter vector at each sampling time step in turn. The hidden state of each time step is generated through forget gate, input gate, cell state and output gate. The hidden state output by the last time step is used as the encoding vector of the time evolution feature of the process parameters. The deformation state sequence of historical processes is input into another branch of the long short-term memory network. After forward propagation calculation, the encoded vector of historical deformation evolution features is output. The time evolution feature encoding vector of process parameters and the historical deformation evolution feature encoding vector are concatenated along the feature dimension to form the final output feature vector of the time recurrent network module.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the processing procedure of the graph space network module includes: Construct a spatial topology graph structure for the thin-walled irregular cavity shaft, and use the deformation values ​​corresponding to each key section in the full-section deformation field distribution data obtained from the current process reconstruction as the initial feature vectors of each node in the graph space network; Each node aggregates the feature vectors of its neighboring nodes through edge connections. After linear transformation of the aggregation result, it is weighted and merged with the feature vector of the node itself to obtain the output feature of the first layer of graph convolution. Using the output features of the first layer as input, the feature aggregation and linear transformation of adjacent nodes are repeated to expand the receptive field of each node, enabling each node to capture the spatial influence of its second-order neighbor nodes. The final feature vectors of all nodes are concatenated in node order to form the output feature vector of the graph space network module.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of inputting the process parameter sequence of the current process, the full-section deformation field distribution data, and the deformation state sequence of historical processes into a hybrid model that integrates a time-cycle network module and a graph-space network module, and outputting the predicted deformation state vector at the end of the current process, includes: The temporal evolution feature vector output by the temporal recurrent network module and the spatial topological feature vector output by the graph space network module are used as inputs to the attention fusion layer. The attention scores of the temporal evolution feature vector and the spatial topological feature vector are calculated separately. The two attention scores are then normalized to obtain the normalized fusion weights of the temporal and spatial features. The fused feature vector is obtained by weighted summation using the normalized fusion weights. The fused feature vector is then subjected to linear transformation and nonlinear activation to output the predicted deformation state vector at the end of the current process.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the step of using the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, solving for the process parameter compensation amount that minimizes the sum of squared deviations on the current remaining processing path, and outputting the optimal compensation increment at the current moment includes: At the current sampling time, the predicted deformation state vector is obtained. The predicted deformation state vector is compared with the preset target deformation state vector, the deviation vector is calculated, and the deviation vector is used as the tracking error term of the optimization problem. Construct a finite-time domain optimization problem and constraints for the current sampling time, and solve it to obtain the optimal process parameter compensation sequence for each time step in the control time domain; Extract the compensation amount at the first moment from the optimal compensation amount sequence obtained from the solution as the instantaneous optimal compensation increment at the current sampling moment, and output the optimal compensation increment.

[0011] Optionally, in the seventh implementation of the first aspect of the present invention, the finite-time domain optimization problem includes a tracking error term and a control increment penalty term, wherein the tracking error term is the sum of squares of the deviations between the predicted deformation and the target deformation at each time step in the prediction time domain, and the control increment penalty term is the product of the sum of squares of the changes in the control quantity at adjacent times in the control time domain and the weighting coefficient. The constraints of the finite-time optimization problem include upper and lower limits on the compensation amounts of each process parameter and upper and lower limits on the rate of change of the compensation amounts.

[0012] A second aspect of the present invention provides a closed-loop control device for the full-process deformation of a thin-walled irregular cavity shaft. The closed-loop control device includes a memory and at least one processor. The memory stores instructions. The at least one processor invokes the instructions in the memory to cause the closed-loop control device to execute the various steps of the closed-loop control method for the full-process deformation of a thin-walled irregular cavity shaft as described in any of the preceding claims.

[0013] A third aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the various steps of the full-process deformation closed-loop control method for thin-walled irregular cavity shafts as described in any of the preceding claims.

[0014] In the technical solution provided by this invention, during the execution of the current process, strain signals of the inner wall of the thin-walled irregular cavity shaft and displacement signals of the outer wall are collected in real time. The strain signals and displacement signals are synchronized in time and matched in spatial coordinates to generate the original deformation observation data at the current moment. The original deformation observation data is fused with the deformation state parameters at the end of the previous process using Kalman filtering to reconstruct the full-section deformation field distribution data of the thin-walled irregular cavity shaft under the current process. The process parameter sequence of the current process, the full-section deformation field distribution data, and the deformation state sequence of historical processes are input into a hybrid model that integrates a time-loop network module and a graph-space network module, outputting the predicted deformation state vector at the end of the current process. Using the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, the process parameter compensation amount that minimizes the sum of squared deviations is calculated on the current remaining processing path, and the optimal compensation increment at the current moment is output. The optimal compensation increment is then... The data is converted into control commands and applied in real time to the current processing step, repeating until all steps are completed. This invention achieves collaborative acquisition and fusion of internal cavity and external wall deformation data under the same spatiotemporal reference. By using Kalman filtering and radial basis function interpolation, discrete observation data is reconstructed into a continuous full-section deformation field, filling the blind spot in the overall deformation assessment. Through parallel fusion of time-loop network branches and graph space network branches via an attention mechanism, a deep joint modeling of the temporal and spatial topological features of deformation evolution is achieved, significantly improving prediction accuracy. A constrained quadratic programming problem is constructed and solved by rolling forward at each sampling time, enabling the compensation amount to be continuously and dynamically updated according to the real-time deformation state. The compensation increment is automatically converted into equipment control commands and verified for execution. At the same time, the deformation state after the completion of the process is reduced in dimension and transmitted to the next process, re-triggering the subsequent closed-loop process. This achieves accurate prediction and closed-loop control in the entire manufacturing process of thin-walled irregular cavity shafts. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart of the closed-loop control method for full-process deformation of a thin-walled irregular cavity shaft provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the closed-loop control device for the full-process deformation of a thin-walled irregular cavity shaft provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the closed-loop control method for the full-process deformation of a thin-walled irregular cavity shaft provided in this embodiment of the invention includes the following steps: Step 101: During the current process, the strain signal of the inner wall of the thin-walled irregular cavity and the displacement signal of the outer wall are collected in real time, and the strain signal and displacement signal are synchronized in time and matched in space coordinates to generate the original deformation observation data at the current moment. In this embodiment, during the current process, wireless passive strain sensing nodes arranged at multiple key cross-sections of the inner wall of the thin-walled irregular cavity and laser displacement sensor arrays arranged at corresponding cross-sections of the outer wall are synchronously triggered at a fixed sampling frequency. Strain readings at each sensing node and displacement readings at each laser measuring point are collected respectively. After obtaining the raw electrical signal data, baseline drift correction, high-frequency noise filtering and power frequency interference notch processing are performed on each channel signal in sequence. The processed strain signal and displacement signal are mapped to the unified workpiece coordinate system of the part according to the spatial layout coordinates of their respective sensing nodes. On this basis, the two types of signals are timestamped to ensure that the strain data and displacement data collected at the same time correspond to the stress state of the part under the same spatial posture. Then, the time-aligned and spatially registered strain data and displacement data are associated and integrated according to the sensor number and cross-section position to form the raw deformation observation data vector at the sampling time, indexed by each key cross-section and including strain components and displacement components.

[0019] Step 102: Perform Kalman filtering to fuse the original deformation observation data with the deformation state parameters at the end of the previous process, and reconstruct the deformation field distribution data of the entire cross section of the thin-walled irregular cavity shaft under the current process; In this embodiment, the deformation state parameter vector stored in the system database at the end of the previous process is extracted. This vector arranges the multidimensional deformation components of all key sections of the thin-walled irregular cavity shaft in the local coordinate system in a fixed order. This vector is used as the initial state estimation mean of the Kalman filter iteration of the current process. The state estimation covariance matrix stored synchronously at the end of the previous process is extracted. This matrix records the joint uncertainty between the estimated values ​​of each deformation component. This matrix is ​​used as the initial state estimation covariance of the Kalman filter iteration of the current process. At the current sampling time, the original deformation observation data is read. This dataset contains the strain readings of each sensing node of the inner cavity wall and the displacement readings of each laser measuring point of the outer wall. The readings are arranged into the original observation vector according to the physical number of the sensor. The original observation vector is processed to remove outliers. The difference between each element in the current observation vector and the corresponding element of the observation vector at the previous time is calculated. If the absolute value of the difference exceeds the preset dynamic threshold, the element is marked as a failure value and replaced with the observation value at the corresponding position at the previous time. The forward calculation process for time updates using Kalman filtering involves multiplying the mean of the optimal state estimate from the previous time step with the state transition matrix to obtain the mean of the prior state estimate for the current time step. This state transition matrix is ​​determined by the process type and part material properties and is fixed during algorithm initialization. The covariance matrix of the optimal estimate from the previous time step is then multiplied on the left and right by the state transition matrix and its transpose, respectively. The process noise covariance matrix is ​​then added to the product to obtain the prior state estimate covariance matrix for the current time step. The measurement update process using Kalman filtering first involves multiplying the prior state estimate... The covariance matrix is ​​multiplied on the left by the transpose of the observation matrix. This product is then multiplied on the left and right by the observation matrix, transposed, and summed with the observation noise covariance matrix, then inverted. Finally, the two intermediate results are multiplied together to obtain the Kalman gain matrix. The observation residual vector at the current time is calculated by subtracting the predicted observation vector obtained by mapping the state prior estimate mean through the observation matrix from the actual observation vector at the current time. The Kalman gain matrix is ​​multiplied by the observation residual vector to obtain the state correction increment. This correction increment is accumulated onto the state prior estimate mean to generate the optimal state estimate mean at the current time. A difference matrix is ​​constructed by multiplying the identity matrix, the Kalman gain matrix, and the observation matrix. This difference matrix is ​​then left-multiplied by the state prior estimate covariance matrix to obtain the optimal estimate covariance matrix for the current time step, which is used for iteration at the next sampling time step. The three-dimensional spatial coordinate offsets of the center points of each key section are extracted from the mean of the optimal state estimate at the current time step. The deformation values ​​at these discrete points and their corresponding axial coordinates are used to form an interpolation node set. A quadratic basis function form for the radial basis interpolation function is defined. All interpolation nodes are traversed, and the basis function matrix is ​​calculated using the Euclidean distance between nodes as the independent variable. The basis function matrix is ​​then regularized. The process involves adding a small regularization parameter on the diagonal to ensure matrix invertibility, then solving the linear equations to obtain the weight coefficients corresponding to each basis function. A dense sequence of interpolation positions is generated along the part's axis. For each interpolation position, the Euclidean distance to all interpolation nodes is calculated. Using this distance as the independent variable, the values ​​of each basis function are calculated, multiplied by the corresponding weight coefficients, and summed to obtain the continuous deformation value at that position. The continuous deformation values ​​at all interpolation positions are arranged in axial coordinate order and cross-sectionally matched with the 3D model of the part in this process to assemble the full-section deformation field distribution data matrix reconstructed at the current moment.

[0020] Step 103: Input the process parameter sequence of the current process, the full cross-section deformation field distribution data, and the deformation state sequence of the historical processes into the hybrid model that integrates the time cyclic network module and the graph space network module, and output the predicted deformation state vector at the end of the current process. In this embodiment, a sequence of process parameters is read in real time from the current process. This sequence records the temporal changes of cutting speed, feed rate, depth of cut, spindle speed, coolant pressure, and clamping force in the order of sampling time. A sliding window truncation process is applied to the process parameter sequence, using the current time as the window endpoint and truncating a fixed number of sampling points forward to form an input window matrix. If the actual number of sampling points is insufficient for the window length, zero values ​​are padded at the beginning. For each type of process parameter in the input window matrix, its mean and standard deviation within that window are calculated. The instantaneous value at the current time is subtracted from the mean, and then divided by the standard deviation to complete the online standardization mapping process for that parameter. The standardized process parameters are then... The number of window matrices is sequentially input into the encoder part of the time-recurrent network module at each time step. In the first time step, the initial hidden state and initial cell state of the Long Short-Term Memory (LSTM) network are all set to zero vectors. Within each time step, the LSM network performs gating operations: first, the forget gate activation value is calculated to determine the proportion of information retained in the previous cell state; second, the input gate activation value is calculated to determine the proportion of current input information written into the cell state; third, the current input and the previous hidden state are used to generate candidate cell states through the hyperbolic tangent function; then, the cell state is updated according to the ratio of the forget gate and the input gate; finally, the output gate activation value is calculated, and the hidden state output of the current time step is generated based on the new cell state. The hidden state output at each time step is temporarily stored in the sequence buffer. After all time steps within the window have been processed, the hidden state of the last time step is extracted as the encoding vector of the time evolution feature of the current process parameters. At the same time, the deformation state vectors of all historical processes before the current process are retrieved from the system database and arranged into a historical deformation state sequence according to the execution order of the processes. The historical deformation state sequence is differentially processed, and the difference sequence of deformation state vectors between adjacent processes is calculated in turn. The mean and standard deviation of the difference sequence are calculated and normalized. The normalized difference sequence is used as the input of the second long short-term memory network branch. The initial hidden state and cell state of the second long short-term memory network branch are also set to zero. The normalized historical deformation difference is read in sequentially according to the time step, and the same gating operation as in step 5 is performed to obtain the hidden state of the last time step of the branch, which is used as the encoding vector of the historical deformation evolution feature. The two feature encoding vectors obtained are concatenated end to end in terms of dimension to form the joint output feature vector of the time recurrent network module, and then subjected to discard regularization, and a portion of the dimensions of the vector are randomly set to zero according to a set ratio. The graph spatial network module is executed to extract the deformation values ​​corresponding to the preset key sections from the reconstructed full-section deformation field distribution data, and to construct the initial feature matrix of each graph node. Based on the preset cross-section adjacency table of the part's geometry, the straight-line distance between the center points of the corresponding sections of any two adjacent nodes is calculated, and the reciprocal of this distance is used as the initial value of the edge weight connecting the two nodes. The adjacent edge weights of all nodes are normalized row-wise, so that the sum of the weights of each node pointing to all its neighboring nodes equals one. Simultaneously, a self-loop edge is added to each node, with the self-loop weight being the smaller value between the sum of the original edge weights of that node and one. The first layer of graph convolution is executed, multiplying the feature vectors of all its neighboring nodes for each target node by the corresponding... The normalized edge weights are summed to obtain the neighbor aggregation feature vector. The initial feature vector of the target node itself is then combined with the neighbor aggregation feature vector element-wise with weighted merging. The merging coefficient is determined by the learnable parameters of the graph convolution layer. The merging result is then transformed by a non-linear activation function to obtain the output features of the first graph convolution layer. The output features of the first graph convolution layer are used as the input features of the second graph convolution layer. The calculation process is repeated to obtain the output features of the second graph convolution layer, thus expanding the receptive field of each node to the range of second-order neighbors. The output features of the second graph convolution layer of all nodes are concatenated into a one-dimensional long vector according to the preset node order. This vector is then transformed by a fully connected mapping layer to output a fixed-dimensional graph space topological feature encoding vector. Attention fusion computation is performed by mapping the joint feature vector output by the time loop network module and the topological feature vector output by the graph space network module to the same semantic dimension space through their respective linear mapping layers. The norm of the two mapped feature vectors is calculated separately, and each norm is used as the initial attention score for that feature vector. The two initial scores are then transformed by a normalized exponential function so that the sum of the two scores equals one, resulting in normalized attention weight coefficients. The original joint feature vector and topological feature vector are multiplied by their corresponding normalized attention weight coefficients, and the two weighted results are added element-wise to generate a fused comprehensive feature vector. The fused comprehensive feature vector is then sequentially input into a three-layer fully connected feedforward network. The first layer undergoes a linear transformation followed by nonlinear activation and dropout regularization. The second layer undergoes a linear transformation followed by nonlinear activation. The third layer undergoes only a linear transformation and sets the output dimension to be consistent with the dimension of the deformation state vector. The output value of the third fully connected network is de-standardized, that is, multiplied by the standard deviation of the deformation data of the current process and added to the mean, to obtain the final predicted deformation state vector at the current time for the end time of the current process.

[0021] Step 104: Taking the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, solve for the process parameter compensation amount that minimizes the sum of squared deviations on the current remaining processing path, and output the optimal compensation increment at the current moment. In this embodiment, during the initialization phase, the hyperparameters for rolling time-domain optimization are set, including the number of discrete time steps in the prediction time domain, the number of discrete time steps in the control time domain, and the time interval between adjacent sampling moments, ensuring that the length of the control time domain is not greater than the length of the prediction time domain. At the current sampling moment, the predicted deformation state vector at the end of the current process is obtained, and the target deformation state vector for the process is loaded from the system database. The difference between the two is calculated to obtain the tracking error vector. A reference trajectory sequence in the prediction time domain is constructed, starting from the actual deformation state at the current moment and ending at the target deformation state, and generating a reference deformation state sequence for each time step in the prediction time domain according to a preset reference trajectory attenuation coefficient. The decision variables of the optimization problem are defined as a set of process parameter compensation amounts corresponding to each moment in the control time domain, and the dimension of the compensation amount at each moment is consistent with the number of adjustable process parameters for the current process. The first component of the optimization objective function is established, and each time step in the prediction time domain is traversed to calculate the time step. The sum of squared deviations between the predicted deformation state and the reference deformation state is accumulated across all time steps and multiplied by a preset tracking error weighting factor. A second component of the optimization objective function is established by iterating through the decision variables at adjacent time points in the control time domain, calculating the difference vector between the compensation amount at the next time point and the compensation amount at the previous time point, calculating the sum of squared differences of this difference vector, accumulating the sum, and then multiplying by a preset smoothing suppression weighting factor. The first and second components of the optimization objective function are added together to form a complete quadratic objective function expression. In this expression, the coefficients of the quadratic terms of the decision variables form the Hessian matrix, and the coefficients of the linear terms form the gradient vector. Physical upper and lower limits are set for the compensation amount of each process parameter in the decision variables. These limits are jointly determined by the maximum adjustment range of the equipment and the process safety boundary. A rate of change constraint is set between the compensation amounts of the same process parameter at adjacent time points in the decision variables, meaning that the absolute value of the difference between the compensation amount at the next time point and the compensation amount at the previous time point must not exceed the maximum allowable change amount of that parameter in a single step. The magnitude constraints and rate of change constraints are uniformly organized into an inequality constraint matrix form, where each constraint condition corresponds to a row in the matrix, and the constraint matrix, lower constraint vector, and upper constraint vector are respectively organized. The constructed objective function is combined with the constructed constraints to form a standard constrained quadratic programming mathematical model. The preset quadratic programming solver is called to numerically solve the above mathematical model. The solver uses the effective set algorithm for iteration. In each iteration, the current effective constraint set is first determined, and then the equality-constrained quadratic programming subproblem is solved on the effective constraint boundary, and the Lagrange multipliers are updated to check the optimality conditions. During the solver iteration process, if the Hessian matrix is ​​detected to be non-positive definite, an incremental perturbation value is added to its diagonal for regularization correction until the matrix becomes a positive definite matrix, ensuring the numerical stability of the solution process. After the solver converges, it outputs the optimal process parameter compensation sequence corresponding to each time step in the control time domain. If the solver fails to converge within the set maximum number of iterations, the optimal solution at the previous sampling time is used as the initial guess value for a hot-start solution. The compensation value at the first time step is extracted from the optimal compensation sequence obtained from the solution and used as the instantaneous optimal compensation increment applied at the current sampling time. The instantaneous optimal compensation increment is smoothly transitioned to the actual compensation amount applied at the previous sampling time. The weighted average of the two is calculated as the final output of the current optimal compensation increment to prevent abrupt changes in the compensation instruction. The remaining part of the optimal compensation amount sequence obtained at the current time, except for the first time, is cached in memory as the initial guess value for the hot start solution at the next sampling time to accelerate the convergence speed of the solution in subsequent time steps. The finally determined current optimal compensation increment, the corresponding process parameter name, and the execution timestamp are encapsulated into an optimization result instruction package.

[0022] Step 105: Convert the optimal compensation increment into a control command and apply it in real time to the current process, repeating the process until all processes are completed.

[0023] In this embodiment, the system receives an optimization result instruction packet, from which the optimal compensation increment value and its corresponding process parameter type identifier are parsed. Based on the current machining environment configuration parameters, the system identifies the brand and model of the CNC system currently in use and the communication protocol type of the process equipment controller. If the current machining process is turning or milling, the optimal compensation increment is decomposed into offset components of each motion axis of the machine tool. The specific decomposition method maps the process parameter compensation amount to the coordinate space of each axis according to the machine tool kinematic chain structure. The offset components of each axis are converted into coordinate origin offset compensation instructions conforming to the current CNC system instruction format. The compensation instructions include an instruction type code, axis identifier, offset direction, and offset value. The coordinate origin offset compensation instructions are written into the real-time compensation register of the CNC system. In the next interpolation cycle, the CNC system automatically superimposes the offset in the register onto the motion instructions of each axis, achieving real-time correction of the machining trajectory. If the current process is heat treatment or surface treatment, the optimal compensation increment is converted into a temperature setpoint correction or a holding time extension, and then converted into an analog control signal output to the temperature control module or timing module of the process equipment. After the compensation command is issued, the feedback signal of the actuator is read. For the CNC system, the actual position feedback value of the coordinate axis is read; for the temperature control system, the thermocouple temperature feedback value is read to verify whether the compensation command has been executed correctly. If the feedback value does not reach the target value of the compensation command, the feedback deviation is calculated and a secondary correction command is generated for additional output until the feedback deviation converges to within the allowable error range. After the current compensation command is executed, a timer is started to wait for the next sampling interval. When the sampling interval ends, the algorithm thread is woken up to prepare for the next round of data acquisition. Before proceeding to the next round of data acquisition, it is determined whether the current process has reached its end point. The processing progress is determined by monitoring the percentage of the length of the processed path executed by the CNC system relative to the total path length. If the processing progress percentage is less than 100%, the sampling counter is incremented and execution is resumed to begin a new round of deformation observation, reconstruction, prediction, and optimization calculations. If the processing progress percentage has reached 100%, the current process's end-of-process procedure is executed. First, all dynamic compensation instructions are stopped, and all process parameter compensation values ​​are reset to zero. After the current process has completely stopped, the measuring equipment is used to collect the final dimensions of all key sections of the thin-walled irregular cavity shaft, obtaining the high-dimensional vector of the actual deformation state at the end of the current process. The high-dimensional vector of the actual deformation state is preprocessed for dimensionality reduction. The mean of each dimension in the vector is calculated and centered to shift the data distribution center to the origin. The centered deformation state vector is input into a pre-trained principal component analysis dimensionality reduction model. This model obtains a projection matrix through eigenvalue decomposition. The high-dimensional vector is multiplied by the projection matrix and projected onto a low-dimensional feature space to obtain low-dimensional encoded features. Simultaneously, the reconstruction error vector reconstructed from the low-dimensional encoded feature into the high-dimensional space is calculated, and the norm of the reconstruction error vector is calculated. If the norm exceeds a preset threshold, the full-dimensional state transfer mode is triggered, retaining the original high-dimensional vector without dimensionality reduction. The dimensionality-reduced low-dimensional encoded feature or the original high-dimensional vector is packaged together with the current process identifier and actual process parameter records as the current process completion status data package and stored in the inter-process state transfer database. The identifier of the next process and its corresponding initial process parameter baseline value are read from the process scheduling table, and the stored status data package is marked as the initial deformation state input of the next process. The next process is set as the new current process, the sampling counter is reset to zero, all cached historical observation data is cleared, and the state estimation covariance matrix of the Kalman filter is reset to the initial value. The new current process processing flow is started, and it automatically jumps when the first sampling time arrives, starting the full-process closed-loop control of the new process until all processes in the process route table are marked as completed, at which point the overall closed-loop control process terminates.

[0024] Figure 2 This is a schematic diagram of a closed-loop deformation control device for a thin-walled irregular cavity shaft throughout its entire process, provided in an embodiment of the present invention. The closed-loop deformation control device 300 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the closed-loop deformation control device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the closed-loop deformation control device 300 to implement the method provided in the above embodiment.

[0025] The thin-walled irregular cavity shaft full-process deformation closed-loop control device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 2The structure of the thin-walled irregular cavity shaft full-process deformation closed-loop control device shown does not constitute a limitation on the computer device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0026] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform each step of the full-process deformation closed-loop control method for thin-walled irregular cavity shafts provided in the above embodiments.

[0027] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0028] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout its entire process, characterized in that, The method includes the following steps: During the current process, the strain signal of the inner wall of the thin-walled irregular cavity and the displacement signal of the outer wall are collected in real time, and the strain signal and displacement signal are synchronized in time and matched in space coordinates to generate the original deformation observation data at the current moment. The original deformation observation data is fused with the deformation state parameters at the end of the previous process by Kalman filtering to reconstruct the full cross-sectional deformation field distribution data of the thin-walled irregular cavity shaft under the current process. The process parameter sequence of the current process, the full cross-section deformation field distribution data, and the deformation state sequence of the historical processes are input into a hybrid model that integrates the time cyclic network module and the graph space network module, and the predicted deformation state vector at the end of the current process is output. Using the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, the process parameter compensation amount that minimizes the sum of squared deviations is solved on the current remaining processing path, and the optimal compensation increment at the current moment is output. The optimal compensation increment is converted into control commands and applied in real time to the current processing step, and repeated until all steps are completed.

2. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 1, characterized in that, The original deformation observation data is fused with the deformation state parameters at the end of the previous process using Kalman filtering to reconstruct the deformation field distribution data of the entire cross-section of the thin-walled irregular cavity shaft under the current process, including: Obtain the deformation state parameter vector passed at the end of the previous process and use it as the initial state estimate of the Kalman filter algorithm for the current process; Obtain the original deformation observation data at the current moment to form the observation vector at the current moment; The state estimate from the previous time step is recursively applied to the current time step through the state transition matrix to obtain the prior state estimate for the current time step. At the same time, the prior estimated covariance matrix is ​​updated based on the process noise covariance matrix. Calculate the Kalman gain matrix, and then weight and fuse the observation vector at the current time with the state prior estimate to obtain the optimal state estimate at the current time. Update the covariance matrix of the optimal state estimate, use the optimal state estimate at each section as the interpolation node, and use the radial basis function interpolation method to expand the deformation data at the discrete sections into a full-section deformation field distribution data that is continuously distributed along the axis of the part, and output the reconstructed full-section deformation field at the current moment.

3. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 2, characterized in that, The deformation state parameter vector includes the radial and axial deformation components of each section of the thin-walled irregular cavity shaft, and the original deformation observation data includes the strain readings of each strain sensing node on the inner cavity wall and the displacement readings of the laser displacement sensor on the outer wall.

4. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 1, characterized in that, The processing procedure of the time loop network module includes: The process parameters time series of the current process and the deformation state series of the historical processes are input into the time loop network module, which includes two long short-term memory network branches. Based on the time series of process parameters of the current process, a long short-term memory network branch is used to read the process parameter vector at each sampling time step in turn. The hidden state of each time step is generated through forget gate, input gate, cell state and output gate. The hidden state output by the last time step is used as the encoding vector of the time evolution feature of the process parameters. The deformation state sequence of historical processes is input into another branch of the long short-term memory network. After forward propagation calculation, the encoded vector of historical deformation evolution features is output. The time evolution feature encoding vector of process parameters and the historical deformation evolution feature encoding vector are concatenated along the feature dimension to form the final output feature vector of the time recurrent network module.

5. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 1, characterized in that, The processing procedure of the graph space network module includes: Construct a spatial topology graph structure for the thin-walled irregular cavity shaft, and use the deformation values ​​corresponding to each key section in the full-section deformation field distribution data obtained from the current process reconstruction as the initial feature vectors of each node in the graph space network; Each node aggregates the feature vectors of its neighboring nodes through edge connections. After linear transformation of the aggregation result, it is weighted and merged with the feature vector of the node itself to obtain the output feature of the first layer of graph convolution. Using the output features of the first layer as input, the feature aggregation and linear transformation of adjacent nodes are repeated to expand the receptive field of each node, enabling each node to capture the spatial influence of its second-order neighbor nodes. The final feature vectors of all nodes are concatenated in node order to form the output feature vector of the graph space network module.

6. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout the entire process as described in claim 1, characterized in that, The process parameter sequence of the current process, the full-section deformation field distribution data, and the deformation state sequence of historical processes are input into a hybrid model that integrates a time-cycle network module and a graph-space network module. The model outputs a predicted deformation state vector at the end of the current process, including: The temporal evolution feature vector output by the temporal recurrent network module and the spatial topological feature vector output by the graph space network module are used as inputs to the attention fusion layer. The attention scores of the temporal evolution feature vector and the spatial topological feature vector are calculated separately. The two attention scores are then normalized to obtain the normalized fusion weights of the temporal and spatial features. The fused feature vector is obtained by weighted summation using the normalized fusion weights. The fused feature vector is then subjected to linear transformation and nonlinear activation to output the predicted deformation state vector at the end of the current process.

7. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 1, characterized in that, The process involves using the deviation between the predicted deformation state vector and the target deformation state vector as the control objective, solving for the process parameter compensation amount that minimizes the sum of squared deviations on the current remaining processing path, and outputting the optimal compensation increment at the current moment, including: At the current sampling time, the predicted deformation state vector is obtained. The predicted deformation state vector is compared with the preset target deformation state vector, the deviation vector is calculated, and the deviation vector is used as the tracking error term of the optimization problem. Construct a finite-time domain optimization problem and constraints for the current sampling time, and solve it to obtain the optimal process parameter compensation sequence for each time step in the control time domain; Extract the compensation amount at the first moment from the optimal compensation amount sequence obtained from the solution as the instantaneous optimal compensation increment at the current sampling moment, and output the optimal compensation increment.

8. The closed-loop control method for deformation of a thin-walled irregular cavity shaft throughout all processes as described in claim 7, characterized in that, The finite-time optimization problem includes a tracking error term and a control increment penalty term, wherein the tracking error term is the sum of squares of the deviations between the predicted deformation and the target deformation at each time step in the prediction time domain, and the control increment penalty term is the product of the sum of squares of the changes in the control quantity at adjacent time steps in the control time domain and the weighting coefficient. The constraints of the finite-time optimization problem include upper and lower limits on the compensation amounts of each process parameter and upper and lower limits on the rate of change of the compensation amounts.

9. A closed-loop control device for the full-process deformation of a thin-walled irregular cavity shaft, characterized in that, The thin-walled irregular cavity shaft full-process deformation closed-loop control device includes a memory and at least one processor. The memory stores instructions. The at least one processor calls the instructions in the memory to cause the thin-walled irregular cavity shaft full-process deformation closed-loop control device to execute each step of the thin-walled irregular cavity shaft full-process deformation closed-loop control method as described in any one of claims 1-8.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement each step of the closed-loop deformation control method for thin-walled irregular cavity shafts as described in any one of claims 1-8.

Images