A machining process precision prediction method, system, device and storage medium

Abstract

The application discloses a machining process precision prediction method, system, device and storage medium, which is applied to the technical field of intelligent manufacturing, and the method comprises the following steps: determining target process indexes and precision requirements to be predicted according to a workpiece machining task, and acquiring multi-source data such as preset parameters of a numerical control system, high-frequency time sequence vibration signals and low-frequency time sequence state signals in a machining process; time-frequency images are generated by performing time-frequency conversion on the high-frequency vibration signals, and time-frequency features are extracted; meanwhile, the multi-source data are preprocessed and time sequence statistical features are extracted; then, the two types of features and real labels are fused to construct a multi-modal feature data set to train a regression prediction model; finally, multi-modal features of a workpiece to be predicted are input into the model to output process precision prediction values, so that end-to-end accurate prediction is realized under the condition of large batch, multiple targets and multiple precisions in complex machining, and the prediction efficiency and model generalization ability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method, system, device, and storage medium for predicting the accuracy of machining processes under complex conditions. Background Technology

[0002] In the field of mechanical processing and manufacturing, the precision of workpiece machining processes is a core indicator determining the assembly performance and service life of high-end equipment products. With the rapid development of aerospace, new energy equipment, precision machine tools, and other fields, the demand for processing large batches of large-section, complex-structure workpieces continues to grow. At the same time, this has placed multi-objective and multi-level precision control requirements on workpiece machining quality. It is necessary not only to control the overall machining precision of the workpiece, but also to impose differentiated constraints on the precision of key local areas, which places higher demands on the prediction and control of machining process precision.

[0003] However, in actual production, especially when dealing with the machining of large and complex workpieces (such as aircraft structural components, large molds, and ship propellers), the process often presents characteristics such as large batch sizes, large machined surface dimensions (large cross-sections), the need to monitor multiple process parameters simultaneously (multiple objectives), and different precision requirements for different areas (multiple precision requirements). For example, for a large plane, it may be necessary to assess the overall flatness of the entire surface while also focusing on monitoring the flatness of key local areas, with significantly different tolerance ranges for the two. Simultaneously, various complex factors during machining, such as vibration, tool condition, and material properties, are coupled together, jointly affecting the final machining quality. Existing technical solutions are typically only suitable for scenarios with simple structures, small batches, and single objectives, and are insufficient to effectively handle the multi-dimensional and highly coupled process accuracy prediction problems under the aforementioned complex conditions.

[0004] To overcome these shortcomings, this application proposes a method, system, equipment, and storage medium for predicting the accuracy of machining processes. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, equipment, and storage medium for predicting the process accuracy of machining, aiming to solve the problem that existing technologies are unable to effectively predict the process accuracy of machining tasks with large batches, large cross-sections, multiple objectives, and multiple accuracy requirements.

[0006] To achieve the above objectives, this application provides the following technical solution: Firstly, this application provides a method for predicting the accuracy of machining processes, comprising the following steps: Based on the workpiece processing task requirements, determine the target process indicators to be predicted and the corresponding prediction accuracy requirements; Acquire multi-source data during the workpiece processing process, including preset processing parameters of the CNC system, high-frequency temporal vibration signals and low-frequency temporal state signals at at least one monitoring point; The high-frequency temporal vibration signal is subjected to time-frequency transformation to generate a time-frequency image, and time-frequency features are extracted from the time-frequency image using an image feature extraction model. The multi-source data is preprocessed, and time-series statistical features are extracted from the preprocessed data; The time-frequency features, the time-series statistical features, and the workpiece's true process quality labels are fused to construct a multimodal feature dataset; A pre-built regression prediction model is trained based on the aforementioned multimodal feature dataset; The multimodal features corresponding to the workpiece to be predicted are input into the trained regression prediction model, and the predicted values ​​of the machining process accuracy index are output.

[0007] Secondly, this application provides a machining process accuracy prediction system, specifically including: The data acquisition module is used to determine the target process indicators to be predicted and the corresponding prediction accuracy requirements based on the workpiece processing task requirements; and to acquire multi-source data during the workpiece processing process, including preset processing parameters of the CNC system, high-frequency time-series vibration signals and low-frequency time-series state signals at at least one monitoring point. The feature extraction module is used to perform time-frequency transformation on the high-frequency temporal vibration signal to generate a time-frequency image, and extract time-frequency features from the time-frequency image through an image feature extraction model; preprocess the multi-source data, and extract temporal statistical features from the preprocessed data; The feature fusion module is used to fuse the time-frequency features, the time-series statistical features, and the workpiece's true process quality label to construct a multimodal feature dataset. The model training and prediction module is used to train a pre-built regression prediction model based on the multimodal feature dataset; input the multimodal features corresponding to the workpiece to be predicted into the trained regression prediction model, and output the predicted value of the processing accuracy index.

[0008] Thirdly, this application provides a computer device, the computer device including a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a method for predicting the accuracy of a machining process; the processor is used to execute the program instructions stored in the memory to implement a method for predicting the accuracy of a machining process.

[0009] Fourthly, this application provides a computer-readable storage medium storing processor-executable program instructions for executing a method for predicting the accuracy of machining processes.

[0010] This application provides a method, system, equipment, and storage medium for predicting the accuracy of machining processes, which has the following beneficial effects: This application first constructs a multi-source data system covering process settings, dynamic vibration, and power changes by collecting preset machining parameters, high-frequency temporal vibration signals, and low-frequency temporal state signals from the CNC system. This overcomes the limitations of traditional methods that rely on only single or local data. Based on this, the high-frequency vibration signals undergo time-frequency transformation, and deep time-frequency features are extracted using the ConvNeXt model. Simultaneously, the multi-source data is preprocessed, and temporal statistical features are extracted. These two types of features complement each other, enabling the model to simultaneously capture both the fine-grained time-frequency patterns and macroscopic statistical regularities of the machining process, thus significantly improving its adaptability to complex working conditions. Furthermore, by fusing the above features with real process quality labels into a multimodal feature dataset and training it using a decision tree-based ensemble regression algorithm, the resulting prediction model is not only structurally simple and highly efficient in training, but also supports end-to-end output with multiple objectives and high precision, eliminating the need for manual point-by-point measurement or switching between different algorithms. Therefore, this application effectively expands the application scope of process accuracy prediction, and can cope with complex machining tasks with large batches, large cross sections, multiple targets, and high precision. At the same time, it reduces the labor intensity and operation difficulty of quality inspectors, and improves production efficiency and economic benefits. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a machining process accuracy prediction method according to Embodiment 1 of this application. Figure 2 This is a schematic diagram of the high-frequency time-series vibration signal after wavelet transform in Embodiment 1 of this application; Figure 3 This is a schematic diagram of the training process of the regression prediction model in Embodiment 1 of this application; Figure 4 This is a schematic diagram of the structure of a machining process accuracy prediction system according to Embodiment 2 of this application; Figure 5 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application; Figure 6 This is a schematic diagram of the storage medium structure of Embodiment 4 of this application. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0013] The following analysis, based on relevant technologies, examines existing solutions.

[0014] Traditional techniques for measuring and predicting machining process accuracy are typically applied to simple structures or localized areas, and are mainly suitable for tasks with relatively simple conditions, such as small batches, small surfaces, and single objectives (e.g., predicting only surface roughness). These methods have significant limitations when dealing with more complex actual production needs. (1) Insufficient scene adaptability, unable to meet the multi-target and multi-precision prediction requirements under complex working conditions. Existing technologies are only suitable for machining scenarios with small batches, small machining surfaces, simple structures, and single precision targets. They cannot be compatible with complex machining conditions with large batches, large cross sections, multiple targets, and multiple precision requirements. They cannot simultaneously output multiple sets of process precision index prediction results for the same workpiece with different dimensions and different tolerance requirements, making it difficult to adapt to the differentiated precision control requirements of high-end equipment manufacturing.

[0015] (2) Limited data source access makes it difficult to support the full-dimensional data requirements for high-precision prediction. Existing accuracy prediction schemes mostly rely on manual offline measurement inspection data. When facing the full-dimensional inspection requirements of large batches of workpieces and large cross-sections and multiple areas, there are problems such as high manual labor intensity, low inspection efficiency and insufficient data integrity. At the same time, for processing areas with complex structures and limited operating space, manual measurement is difficult to implement effectively, and it is impossible to provide comprehensive and real-time dynamic data of the processing process for accuracy prediction. This results in insufficient input feature dimensions of the prediction model and limited prediction accuracy.

[0016] (3) Insufficient feature extraction and fusion capabilities, resulting in poor prediction accuracy and generalization performance. Existing solutions mostly use single-dimensional processing features for accuracy prediction, failing to fully explore the correlation features between multimodal data such as high-frequency vibration signals, low-frequency power signals, and CNC system preset parameters during processing. This makes it impossible to comprehensively characterize the nonlinear mapping relationship between changes in processing conditions and process accuracy, leading to insufficient accuracy of prediction results and poor generalization performance for different processing conditions.

[0017] (4) The model deployment threshold is high and it is difficult to adapt to the actual application needs of industrial sites. Existing mainstream accuracy prediction models have high requirements for the size of the training dataset, and it is difficult to achieve effective convergence with small batch datasets. At the same time, they are highly dependent on hardware computing power, the model training cycle is long, and it is difficult to achieve localized and fast offline training. Some deep learning models have complex structures and poor interpretability of the prediction process, which do not meet the reliability control requirements of industrial production scenarios and are difficult to achieve stable and efficient deployment and application in production sites.

[0018] This application takes multiple states and parameters, such as working status, tool life parameters, workpiece material, vibration parameters of multiple points of workpiece fixture, vibration parameters of multiple points of machine tool spindle, and spindle power, as input to the algorithm. It also considers the interference between multiple features and various factors such as spindle feed rate and tool installation conditions during the machining process, so as to achieve end-to-end prediction of machining process accuracy values ​​under complex conditions of multiple objectives and multiple precisions.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Example 1 Please see Figure 1 This is a flowchart illustrating a machining process accuracy prediction method according to Embodiment 1 of this application; the steps include: S1: Based on the workpiece processing task requirements, determine the target process indicators to be predicted and the corresponding prediction accuracy requirements.

[0021] In this embodiment, the required process accuracy indicators and their accuracy requirements are first determined based on the specific machining task. Taking the milling of a large planar workpiece (such as a structural component of a certain type of aircraft or a large mold) as an example, the flatness quality of the machined surface needs to be evaluated after machining. Since the workpiece has a large surface area and different functional requirements in different areas, two process accuracy indicators need to be predicted simultaneously: one is local flatness, i.e., the flatness of several key areas within a large, complete plane; the other is global flatness, i.e., the overall flatness of the entire machined surface. According to the target task requirements, the allowable tolerance range for local flatness is ±0.025mm, while the allowable tolerance range for global flatness is ±0.15mm. Furthermore, depending on the machining task, the target process indicators can be extended to other geometric tolerances or dimensional accuracy indicators such as roughness, roundness, cylindricity, coaxiality, or concentricity, and the corresponding accuracy requirements are set based on the design tolerances of the specific workpiece.

[0022] S2: Acquire multi-source data during the workpiece processing process. The multi-source data includes preset processing parameters of the CNC system, high-frequency time-series vibration signals and low-frequency time-series state signals at at least one monitoring point.

[0023] In this embodiment, the CNC machining system is started after the workpiece is clamped, and three types of multi-source data are collected simultaneously during the machining process. The first type is the preset machining parameters of the CNC system. Before or during machining, key preset information is read through the CNC system's data interface, including important preset data such as spindle speed, feed rate, tool life, tool mounting height, tool material, and workpiece material information. These parameters need to be comprehensively evaluated and preset according to specific tasks and requirements. In this embodiment, the parameters of focus are: spindle feed rate, tool life parameters, and workpiece material information.

[0024] The second category is high-frequency temporal vibration signals. High-sensitivity vibration sensors are installed at key locations on the machine tool spindle and workpiece fixture to continuously collect raw vibration data throughout the machining process at a sampling frequency of 20kHz. Specifically, at the machine tool spindle, one vibration sensor is installed along each of the three orthogonal directions (X, Y, and Z, for a total of three channels) to comprehensively capture the vibration components of the spindle in three-dimensional space. At the workpiece fixture, four different monitoring points are selected (e.g., the four corners of the fixture or key support positions in contact with the workpiece), and two vibration sensors are installed at each point, orthogonally distributed at 90 degrees, to collect vibration data in the X and Y directions at that point (for a total of eight channels), thereby eliminating mutual interference between vibration components in non-orthogonal directions.

[0025] The third category is low-frequency timing signals. Spindle power is used as a representative of low-frequency signals. A power sensor mounted on the spindle drive motor collects real-time power change data of the spindle during machining at a sampling frequency of 10Hz. This signal reflects changes in cutting load, and its frequency is much lower than that of vibration signals, thus belonging to low-frequency timing data.

[0026] The above three types of data are stored in association according to the batch and number of the processed workpiece, forming the multi-source raw data basis for subsequent feature extraction and model prediction.

[0027] S3: Perform time-frequency transformation on the high-frequency time-series vibration signal to generate a time-frequency image, and extract time-frequency features from the time-frequency image using an image feature extraction model.

[0028] In this embodiment, the high-frequency temporal vibration signals from multiple channels acquired in step S2 are subjected to time-frequency transformation to generate corresponding time-frequency images. Specifically, wavelet transform is used to convert the high-frequency temporal vibration signal of each channel into a corresponding time-frequency image.

[0029] For the original vibration timing signal of each channel Choose a suitable wavelet mother function A series of wavelet basis functions are obtained by scaling and translating the parent function, and then the inner product operation is performed with the original signal; the calculation formula is as follows: ; in, As a scale factor, For displacement variables.

[0030] Please see Figure 2 This is a schematic diagram of the high-frequency time-series vibration signal after wavelet transform in Embodiment 1 of this application. Wavelet transform is a time-scale (time-frequency) analysis method for signals, characterized by multi-resolution analysis. After wavelet transform, the vibration signal of each channel is converted into a time-frequency image. The time-frequency image has strong representation capabilities in both the time and frequency domains, exhibiting high vibration frequency resolution in the low-frequency region and high time resolution in the high-frequency region, making it well-suited for extracting and mining local features of non-stationary signals.

[0031] Next, time-frequency features are extracted from the time-frequency image using an image feature extraction model. The image feature extraction model is a ConvNeXt model, which divides the time-frequency image into non-overlapping patch feature blocks using a 4×4 convolution kernel with a stride of 4. Feature processing is then performed through an inverted bottleneck structure composed of ConvNeXt blocks, and the time-frequency features are output after passing through a Linear layer. The model structure comprises four main stages, each consisting of multiple stacked ConvNeXt blocks. First, a 4×4 convolution kernel with a stride of 4 is used to divide the input time-frequency image into non-overlapping patch feature blocks. Subsequently, the four stages sequentially downsample and transform the feature map. Each ConvNeXt block employs an inverted bottleneck structure that expands the number of channels from narrow to wide and then back to narrow (i.e., first expanding the number of channels, then compressing them back to the original size). This design enhances the model's feature representation capability in high-dimensional space while reducing computational cost and accelerating convergence. After four stages of feature extraction, the feature vector corresponding to each time-frequency image is finally output through the Linear layer, which is the time-frequency feature.

[0032] S4: Preprocess the multi-source data and extract time-series statistical features from the preprocessed data.

[0033] In this embodiment, the multi-source data includes high-frequency time-series vibration signals, low-frequency time-series state signals, and preset machining parameters of the CNC system. Since the sampling frequencies and dimensions of these data differ, they cannot be directly fused and analyzed. Therefore, time-dimensional alignment and dimensionless normalization are required.

[0034] First, temporal alignment is performed. The high-frequency vibration signal is sampled at 20kHz, meaning one data point every 0.05 milliseconds; the low-frequency power signal is sampled at 10Hz, meaning one data point every 100 milliseconds. To establish a temporal correlation between the two types of signals, the high-frequency signal is segmented into segments according to the same time windows (e.g., every 100 milliseconds) based on the timestamp of the low-frequency signal. Statistical characteristics within each window are calculated, thus achieving temporal alignment between the high-frequency and low-frequency data. Simultaneously, preset processing parameters, which are constants or slowly changing quantities before or during processing, are also matched according to the time windows to ensure a complete set of preset parameter values ​​within each time window.

[0035] Secondly, dimensionless processing is performed. Since data collected by different sensors have different physical units and numerical magnitudes, directly using the raw numerical values ​​can lead to large-scale features dominating model training, masking the influence of subtle features. Therefore, standardization is performed separately for each type of data, commonly using Z-score standardization. Dimensionless processing eliminates differences in dimensions and magnitudes, preserving the distribution characteristics and relative changes of the original data.

[0036] After preprocessing, three types of statistical features are extracted from the high-frequency time-series vibration signals within each time window: standard deviation (reflecting the dispersion of vibration amplitude), skewness (reflecting the asymmetry of vibration distribution), and kurtosis (reflecting vibration impact characteristics).

[0037] The formula for calculating standard deviation is: ; The skewness formula is: ; The kurtosis formula is: ; in The number of vibration signal samples. The mean, Standard deviation This represents the value of the i-th vibration signal in the sample. The above three features describe the statistical characteristics of the vibration signal from different perspectives, and are of great significance for monitoring unstable states in the processing process.

[0038] Simultaneously, the mean values ​​are extracted from the low-frequency timing status signals and the preset machining parameters of the CNC system within each time window as statistical features. The average value of the spindle power within the window is calculated to reflect the average cutting load during that time period. For preset parameters such as spindle speed and feed rate, since they may change in segments during machining, their mean values ​​are also calculated according to the time window.

[0039] Ultimately, each time window corresponds to a complete set of temporal statistical feature vectors, including the standard deviation, skewness, and kurtosis of all high-frequency channels, as well as the mean of low-frequency power and the mean of multiple preset parameters. These temporal statistical features, together with the time-frequency features extracted in step S3 and the workpiece's true process quality label, will be used to construct the subsequent multimodal feature dataset.

[0040] S5: The time-frequency features, the time-series statistical features, and the workpiece's true process quality label are fused together to construct a multimodal feature dataset.

[0041] In this embodiment, firstly, the true label of the workpiece's process quality is obtained. After the workpiece is processed, a coordinate measuring machine (CMM) system is used to accurately measure the accuracy of the processed area. Taking flatness and geometric tolerance as an example, for workpieces that need to simultaneously predict local flatness and global flatness, labels are made as follows: For local flatness, a circular area with a diameter of 50mm is used as a sampling unit. Sampling is performed on the entire surface to be measured along a specified direction and a predetermined step size (e.g., every 30mm). The deviation of the actual measured value within the sampling area is calculated for each movement, and the maximum deviation in all sampling areas is taken as the true value of the local flatness of the workpiece. For global flatness, the CMM measurement data of the entire surface to be measured is used, and the difference between the maximum and minimum values ​​among all measurement points is taken as the true value of global flatness. The above true values ​​can be further divided into quality grades (e.g., Class I qualified products, Class II qualified products, and defective products) according to a preset tolerance range (local flatness ±0.025mm, global flatness ±0.15mm), but continuous values ​​are directly used as labels in the regression prediction model. In addition, if the target of prediction is other process indicators such as roughness or roundness, the true label shall be obtained in accordance with the measurement methods specified in the relevant national or industry standards.

[0042] Secondly, the multi-source features corresponding to each workpiece are correlated and fused. Since the processing of a workpiece may correspond to multiple time windows, and time-frequency features are usually extracted from the entire processing signal or key segments, feature alignment is required. The global mean or maximum value of the time-series statistical features for all time windows of each workpiece is taken to obtain a global time-series statistical feature vector; simultaneously, the time-frequency features of all channels are concatenated or averaged to obtain a global time-frequency feature vector. Then, the two are concatenated into a complete feature vector and associated with the true process quality label of the workpiece.

[0043] Finally, the sample data from all workpieces are aggregated to form a multimodal feature dataset. This dataset is randomly divided into a training set and a test set at a certain ratio (e.g., 4:1). The training set is used for parameter learning of the subsequent regression prediction model, and the test set is used to evaluate the model's generalization performance. During the dataset construction process, if measurement data for some workpieces is missing or abnormal, it can be removed according to preset rules or supplemented using interpolation methods to ensure data quality. This multimodal feature dataset integrates deep features in the time-frequency domain and statistical features in the time domain, and establishes a mapping relationship with the actual measured values ​​of process accuracy, providing a data foundation for subsequent training of high-precision prediction models.

[0044] S6: Train a pre-built regression prediction model based on the multimodal feature dataset.

[0045] In this embodiment, the regression prediction model is an ensemble regression model based on decision trees. It improves prediction accuracy by combining multiple weak classification and regression trees (CART trees), offering advantages such as insensitivity to feature scaling, the ability to handle mixed-type features, and strong resistance to overfitting. During the training process of the ensemble regression model, a loss function is constructed based on the second-order Taylor expansion, and the first and second derivatives are used to guide the optimization of this loss function.

[0046] Specifically, the model is built upon the gradient descent method of the previous training loss function during each training update. That is, after obtaining the fitting error for the first time, each subsequent training iteration further reduces the training error based on the previous training error. During the training optimization process, the algorithm uses a second-order Taylor expansion to establish the loss function, which combines the first and second derivatives to guide the calculation of the loss function. This method makes the training optimization process more stable and efficient.

[0047] The basic decision tree expression is:

[0048] in The weights of the leaf nodes; This is the index function that maps samples to leaf nodes.

[0049] The model's final prediction output is the sum of all decision tree predictions, expressed as: ; in For the predicted results, For the t-th regression tree, For the function space of the CART tree, This represents the multimodal feature vector of the sample.

[0050] The splitting of nodes in each decision tree is determined based on a gain formula. When attempting to split a node into left and right child nodes, the gain after the split is calculated: ; in L2 regularization for leaf weights The penalty term is the number of leaf nodes. and These are the second-order gradient sum and the first-order gradient sum of the left subtree, respectively. and Let be the sum of the second-order gradients and the sum of the first-order gradients of the right subtree, respectively.

[0051] During training, several hyperparameters need to be pre-set, including: learning rate, maximum tree depth, minimum sample weight for leaf nodes, and number of trees. When training the model using the training set data, cross-validation is used to select the optimal combination of hyperparameters. After model training is complete, the model's performance is evaluated using a test set. The samples in the test set have never been involved in training or hyperparameter tuning, thus accurately reflecting the model's generalization performance. The optimal model structure and weight parameters obtained during training are saved for subsequent online prediction of new artifacts.

[0052] Please see Figure 3 This diagram illustrates the training process of the regression prediction model in Embodiment 1 of this application. After acquiring the multimodal feature dataset and dividing it into training and test sets, an initial regression prediction model structure is first constructed. Then, the model is iteratively trained using the training set, with continuous optimization of the model structure and hyperparameters during training. After each round of training, the prediction performance of the current model is evaluated using the test set to determine if it meets the preset accuracy standard. If the evaluation meets the standard, training ends and the optimal model is saved; if not, the process returns to continue adjusting the model structure and hyperparameters, repeating the training and evaluation process until the requirements are met. Through the above closed-loop optimization process, a process accuracy regression prediction model with good generalization ability and high prediction accuracy is finally obtained.

[0053] S7: Input the multimodal features corresponding to the workpiece to be predicted into the trained regression prediction model, and output the predicted value of the machining process accuracy index.

[0054] In this embodiment, when a new workpiece enters the production line, data acquisition, feature extraction and fusion are performed according to the same process as in the training phase. Then, the obtained multimodal features are input into the optimal regression prediction model that has been trained and saved, and the predicted value of the processing accuracy of the workpiece is output.

[0055] Specifically, the workpiece to be predicted is first clamped in place and the CNC machining system is started. During machining, preset machining parameters are read from the CNC system, and high-frequency timing vibration signals of the spindle and fixture are collected at a sampling frequency of 20kHz; the spindle power signal is collected at a sampling frequency of 10Hz as a low-frequency timing status signal. After machining is completed, the collected data undergoes the same preprocessing as in the training phase.

[0056] Next, wavelet transform is performed on the high-frequency temporal vibration signal of each channel to generate the corresponding time-frequency image. Then, the trained ConvNeXt image feature extraction model is used to extract the depth time-frequency feature vector from each time-frequency image. The standard deviation, skewness, and kurtosis are calculated from the high-frequency vibration signal of each time window, and the mean is calculated from the low-frequency power signal and preset parameters. All statistical features are concatenated into a time-series statistical feature vector. Subsequently, the time-frequency features and time-series statistical features are fused to form a multimodal feature vector with the same dimension as the training samples.

[0057] Finally, the multimodal feature vector is input into the trained decision tree-based regression prediction model. The model calculates and accumulates the predicted values ​​from each of the multiple integrated regression trees, ultimately outputting predicted values ​​for one or more process accuracy indicators of the workpiece. For example, for large planar workpieces, the model simultaneously outputs local and global flatness prediction values. After receiving these predicted values, operators or the quality monitoring system can compare them with preset tolerance ranges to determine whether the workpiece's processing quality is acceptable or requires further processing. The entire prediction process is completed within seconds of processing, eliminating the need for manual point-by-point measurement, significantly improving inspection efficiency and reducing labor intensity. The prediction results can be automatically saved to a database or transmitted to the quality management system for production report statistics, process parameter traceability, and subsequent optimization analysis.

[0058] In summary, Embodiment 1 of this application first determines the target process indicators and their accuracy requirements to be predicted based on the processing task requirements. During the processing, it simultaneously acquires the preset processing parameters of the CNC system, high-frequency temporal vibration signals at at least one monitoring point, and low-frequency temporal state signals as multi-source data. Subsequently, it performs time-frequency transformation on the high-frequency temporal vibration signals to generate time-frequency images, and uses an image feature extraction model to extract deep time-frequency features from them. Simultaneously, it preprocesses the multi-source data and extracts temporal statistical features from the preprocessed data. The time-frequency features, temporal statistical features, and the true process quality labels obtained through precision measurement are fused to construct a multimodal feature dataset. Based on this dataset, a pre-constructed regression prediction model is trained. During training, a second-order Taylor expansion is used to construct the loss function, and iterative optimization is performed using first and second derivatives. Finally, for the workpiece to be predicted, its multimodal features are extracted following the same process and input into the trained regression prediction model, which can quickly output the predicted values ​​of the processing accuracy indicators. This method enables end-to-end prediction of machining process accuracy under complex conditions of large-scale, multi-target, and multi-precision machining, significantly reducing the cost and difficulty of manual inspection while ensuring prediction accuracy.

[0059] Example 2 Please see Figure 4 This is a schematic diagram of the structure of a machining process accuracy prediction system according to Embodiment 2 of this application; the specific contents include: The data acquisition module 100 is used to determine the target process indicators to be predicted and the corresponding prediction accuracy requirements according to the workpiece processing task requirements; and to acquire multi-source data during the workpiece processing process, wherein the multi-source data includes the preset processing parameters of the CNC system, high-frequency time-series vibration signals and low-frequency time-series state signals at at least one monitoring point. The feature extraction module 200 is used to perform time-frequency transformation on the high-frequency time-series vibration signal to generate a time-frequency image, and extract time-frequency features from the time-frequency image through an image feature extraction model; preprocess the multi-source data, and extract time-series statistical features from the preprocessed data; The feature fusion module 300 is used to fuse the time-frequency features, the time-series statistical features, and the workpiece's true process quality label to construct a multimodal feature dataset. The model training and prediction module 400 is used to train a pre-built regression prediction model based on the multimodal feature dataset; input the multimodal features corresponding to the workpiece to be predicted into the trained regression prediction model, and output the predicted value of the processing accuracy index.

[0060] In this embodiment, the data acquisition module 100 is used to determine the target process indicators to be predicted and their corresponding accuracy requirements based on the workpiece processing task requirements. Simultaneously, it acquires multi-source data during the processing, including reading preset processing parameters from the CNC system and collecting high-frequency temporal vibration signals and low-frequency temporal state signals at at least one monitoring point via sensors. This data is stored in association according to the workpiece number, providing raw input for subsequent processing.

[0061] The feature extraction module 200 performs time-frequency transformation on the high-frequency temporal vibration signal to generate a corresponding time-frequency image, and extracts deep time-frequency features from the time-frequency image using a pre-trained image feature extraction model. Simultaneously, it preprocesses the multi-source data, including time alignment and dimensionless processing, and then extracts temporal statistical features from the preprocessed data. These two types of features respectively characterize the time-frequency domain pattern and time-domain statistical properties of the processing.

[0062] The feature fusion module 300 fuses the time-frequency features and time-series statistical features output by the feature extraction module with the true labels of workpiece process quality obtained through precision measurement to construct a multimodal feature dataset. This dataset is divided into training and testing sets according to a certain ratio for subsequent model training and evaluation. The fused feature vectors simultaneously contain the abstract representations of deep learning and the interpretable features of traditional statistics, which can more comprehensively reflect the mapping relationship between processing status and final accuracy.

[0063] The model training and prediction module 400 trains a pre-built regression prediction model based on a constructed multimodal feature dataset. During training, a second-order Taylor expansion is used to construct the loss function, and first and second derivatives are used to guide model optimization. Hyperparameters are adjusted through cross-validation to obtain the optimal model. For new workpieces to be predicted, their multimodal features are extracted following the same process as training, and then input into the trained model. The model outputs predicted values ​​for machining process accuracy indicators, thereby achieving rapid and accurate prediction of machining quality.

[0064] For further details regarding the implementation of the technical solutions for each module in the machining process accuracy prediction system of the above embodiments, please refer to the description in the machining process accuracy prediction method of the above embodiments, which will not be repeated here.

[0065] Example 3 Please see Figure 5 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0066] The memory 52 stores program instructions for implementing the above-described method for predicting the accuracy of machining processes.

[0067] The processor 51 is used to execute program instructions stored in the memory 52 to achieve a machining process accuracy prediction.

[0068] The processor 51 can also be referred to as a CPU (Central Processing Unit).

[0069] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0070] Example 4 Please see Figure 6 This is a schematic diagram of the storage medium in Embodiment 4 of this application. The storage medium in this embodiment stores a program file 61 capable of implementing all the above methods. This program file 61 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of 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, or devices such as computers, servers, mobile phones, and tablets.

[0071] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0072] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0073] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

[0074] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.

Claims

1. A method for predicting the accuracy of machining processes, characterized in that, The method includes the following steps: Based on the workpiece processing task requirements, determine the target process indicators to be predicted and the corresponding prediction accuracy requirements; Acquire multi-source data during the workpiece processing process, including preset processing parameters of the CNC system, high-frequency temporal vibration signals and low-frequency temporal state signals at at least one monitoring point; The high-frequency temporal vibration signal is subjected to time-frequency transformation to generate a time-frequency image, and time-frequency features are extracted from the time-frequency image using an image feature extraction model. The multi-source data is preprocessed, and time-series statistical features are extracted from the preprocessed data; The time-frequency features, the time-series statistical features, and the workpiece's true process quality labels are fused to construct a multimodal feature dataset; A pre-built regression prediction model is trained based on the aforementioned multimodal feature dataset; The multimodal features corresponding to the workpiece to be predicted are input into the trained regression prediction model, and the predicted values ​​of the machining process accuracy index are output.

2. The method for predicting machining process accuracy according to claim 1, characterized in that, The steps for acquiring multi-source data during the workpiece processing specifically include: The high-frequency timing vibration signals of multiple channels at the machine tool spindle and workpiece fixture are collected at a first sampling frequency, and the low-frequency timing state signal is collected at a second sampling frequency.

3. The method for predicting machining process accuracy according to claim 2, characterized in that, The step of performing time-frequency transformation on the high-frequency time-series vibration signal to generate a time-frequency image, and extracting time-frequency features from the time-frequency image using an image feature extraction model, specifically includes: Wavelet transform is used to convert the high-frequency time-series vibration signal of each channel into a corresponding time-frequency image; The image feature extraction model is the ConvNeXt model. The time-frequency image is divided into non-overlapping Patch feature blocks through a convolution operation with a 4×4 convolution kernel and a stride of 4. Feature processing is performed through an inverted bottleneck structure composed of ConvNeXt blocks, and the time-frequency features are output after passing through a Linear layer.

4. The method for predicting machining process accuracy according to claim 3, characterized in that, The step of preprocessing the multi-source data and extracting time-series statistical features from the preprocessed data specifically includes: The preprocessing includes time dimension alignment and dimensionless normalization. The time-series statistical features include the standard deviation, skewness, and kurtosis of each channel of the high-frequency time-series vibration signal, as well as the average value of the low-frequency time-series state signal corresponding to the preset processing parameters.

5. The method for predicting machining process accuracy according to claim 1, characterized in that, The regression prediction model is an ensemble regression model based on decision trees; During the training of the ensemble regression model, a loss function is constructed based on the second-order Taylor expansion, and the first and second derivatives are used to guide the optimization of the loss function.

6. A machining process accuracy prediction system, characterized in that, For performing the machining process accuracy prediction method according to any one of claims 1 to 5, the machining process accuracy prediction system comprises: The data acquisition module is used to determine the target process indicators to be predicted and the corresponding prediction accuracy requirements based on the workpiece processing task requirements; and to acquire multi-source data during the workpiece processing process, including preset processing parameters of the CNC system, high-frequency time-series vibration signals and low-frequency time-series state signals at at least one monitoring point. The feature extraction module is used to perform time-frequency transformation on the high-frequency temporal vibration signal to generate a time-frequency image, and extract time-frequency features from the time-frequency image through an image feature extraction model; preprocess the multi-source data, and extract temporal statistical features from the preprocessed data; The feature fusion module is used to fuse the time-frequency features, the time-series statistical features, and the workpiece's true process quality label to construct a multimodal feature dataset. The model training and prediction module is used to train a pre-built regression prediction model based on the multimodal feature dataset; input the multimodal features corresponding to the workpiece to be predicted into the trained regression prediction model, and output the predicted value of the processing accuracy index.

7. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing the machining process accuracy prediction method according to any one of claims 1-5; the processor is used to execute the program instructions stored in the memory to implement machining process accuracy prediction.

8. A computer-readable storage medium, characterized in that, The system stores processor-executable program instructions for performing the machining process accuracy prediction method according to any one of claims 1-5.

Images