Aero-engine titanium alloy centrifugal impeller milling vibration and roughness prediction system
By constructing a deep learning system for data acquisition, preprocessing, and SE-ResNet18 model, the nonlinear coupling problem between vibration and roughness during the milling of titanium alloy centrifugal impellers for aero-engines was solved, achieving real-time high-precision surface roughness prediction and improving the machining quality control capability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-05-26
- Publication Date
- 2026-06-23
Smart Images
Figure CN122262931A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aero-engine manufacturing and intelligent manufacturing technology, specifically relating to a milling vibration and roughness prediction system for aero-engine titanium alloy centrifugal impellers. Background Technology
[0002] Compressor blades for aero-engines are core components determining the overall aerodynamic performance and service reliability of the engine. The surface roughness of their milled surfaces directly affects the initiation and propagation rate of fatigue cracks and their corrosion resistance, making them a key indicator for aero-engine airworthiness compliance verification and processing quality control. TC11 titanium alloy, due to its high specific strength and excellent high-temperature resistance, has become the mainstream manufacturing material for centrifugal compressor impeller blades in aero-engines. However, this material has poor machinability, and dynamic vibrations in the tool-workpiece system during milling can easily cause surface quality fluctuations. Achieving accurate correlation prediction between vibration and surface roughness has become a core technical challenge in the high-performance manufacturing of titanium alloy compressor blades, directly impacting the processing quality of life-limited components and the overall service safety of aero-engines.
[0003] Surface roughness prediction has always been a research hotspot in the field of machining. Traditional methods are mostly based on linear or low-order nonlinear fitting algorithms such as response surface models and polynomial regression, which achieve prediction by establishing a mapping relationship between machining parameters and roughness. In existing research, scholars have used titanium alloys as the research object and constructed a regression prediction model of cutting parameters and surface roughness based on the response surface method, realizing roughness prediction under conventional milling conditions and analyzing the influence of cutting parameters. However, such methods are difficult to characterize the strong nonlinear coupling relationship between milling vibration and roughness, and cannot capture the deep machining information such as tool wear and machining stability contained in the vibration signal. In the complex curved surface milling scenario of aero-engine blades, there are problems such as insufficient prediction accuracy and poor adaptability to changing working conditions, which cannot meet the quality control requirements of high-precision machining of aero-engines.
[0004] With the development of deep learning technology, residual networks, thanks to their residual connection structure, effectively solve the gradient vanishing problem in deep networks and are widely used for feature extraction of machining signals. The introduction of channel attention mechanisms can improve the network's ability to express key signal components through adaptive feature recalibration, providing a new technical approach for high-precision mapping of machining vibration signals and surface roughness. However, current domestic and international research mostly applies deep learning models to roughness prediction of general mechanical parts, with limited specific research on milling scenarios for aero-engine compressor blades. A vibration-roughness correlation prediction method suitable for titanium alloy centrifugal impeller blades has not yet been developed, and existing research does not fully consider the machining quality control requirements of aero-engine life-limited components, making it difficult to directly apply to online monitoring and early warning of actual machining quality of aero-engine impellers. Summary of the Invention
[0005] To address the aforementioned issues, this invention provides a prediction system for milling vibration and roughness of titanium alloy centrifugal impellers for aero-engines, comprising a data acquisition module, a data preprocessing and feature conversion module, an SE-ResNet deep learning prediction model module, and an online monitoring and early warning module. The data acquisition module is used to synchronously acquire the vibration acceleration signal at the blade root and the surface roughness data at the corresponding machining position during the milling process. The data preprocessing and feature transformation module is used to remove outliers, perform spatiotemporal transformation, extract frequency domain features, and encode two-dimensional images of Gram angle field in the original data to construct a standardized dataset with a one-to-one mapping between vibration and roughness. The SE-ResNet deep learning prediction model module is used to train the SE-ResNet18 model with a fusion channel attention mechanism based on a standardized dataset to obtain the optimal prediction model weights. The online monitoring and early warning module is used to deploy the trained model to realize real-time prediction and early warning of surface roughness during the processing. The modules are connected through standardized data interfaces.
[0006] Preferably, the data acquisition module uses the centrifugal compressor impeller of an aero-engine made of TC11 titanium alloy as the acquisition object, and uses an IEPE accelerometer attached to the root of the impeller to collect vibration signals. The sampling frequency is set to 20000Hz. A laser three-dimensional confocal microscope is used to collect surface roughness data. The spindle speed, feed rate, tool space coordinates, and machining status signals are collected synchronously through the OPCUA interface of the machine tool CNC system.
[0007] Preferably, the data acquisition module conducts a pre-milling trial cut before formal data acquisition. A gradient level is set with the spindle speed as the single variable to conduct a control trial cut. Based on the results of the pre-experiment, a training set and a validation set with non-overlapping working conditions are selected. The working conditions of the training set are a spindle speed of 900 rpm, a feed rate of 180 mm / min, an axial cutting depth of 2 mm, and a radial cutting depth of 0.1 mm. The working conditions of the validation set are a spindle speed of 1800 rpm, and the other process parameters are consistent with the training set. During the acquisition process, a point coordinate mapping index table of vibration signal and surface roughness data is established.
[0008] Preferably, the data preprocessing and feature conversion module standardizes the surface roughness data, divides the contour height data of the entire processing area into equal segments according to a sampling length of 0.28 mm, calculates the arithmetic mean deviation Ra value of the contour of each unit, and divides the measured roughness value into 17 continuous graded intervals with an interval of 0.11 μm, corresponding to category codes 0 to 16.
[0009] Preferably, the data preprocessing and feature conversion module preprocesses the vibration signal: converts the original spatial domain discrete data into a continuous time domain sampling sequence through the feed speed, uses the 3σ criterion to remove abnormal sampling points and performs linear interpolation, takes 224 continuous time domain data points as a sample unit, uses fast Fourier transform to extract frequency domain feature frequencies, removes the 50Hz power frequency and its harmonics, and retains feature frequencies that are strongly correlated with milling process parameters.
[0010] Preferably, the data preprocessing and feature conversion module uses Gram angle field to convert the one-dimensional time-domain vibration signal into a two-dimensional image: the single vibration sequence is normalized to the interval [-1,1], the normalized sequence is converted into angle values through polar coordinate mapping, the cosine of the sum of any two angle values is calculated to construct a 224×224 two-dimensional matrix, and the matrix values are linearly mapped to the gray value range of 0 to 255 to generate a single-channel grayscale image.
[0011] Preferably, the data preprocessing and feature conversion module performs data augmentation on the generated Gram angle field image, including random horizontal flipping, random vertical flipping, random rotation within ±15°, and adjustment of brightness and contrast within ±0.2°. The module matches the corresponding roughness average value as a label according to 224 data points as an interval, and constructs a training set and an independent validation set.
[0012] Preferably, the SE-ResNet deep learning prediction model module adopts the SE-ResNet18 model, which integrates the SE channel attention module into each residual basic unit of the traditional ResNet18. The model input is a single-channel grayscale image of 224×224×1. After feature extraction by multiple SE-ResNet basic units, the predicted probabilities of 17 roughness categories are output through global average pooling layers and fully connected layers.
[0013] Preferably, the SE channel attention module compresses the feature map of each channel into a scalar representation of the global importance of the channel through global average pooling, and recalibrates the original feature map by generating channel weight vectors through two fully connected layers, adaptively adjusting the weights of the feature channels to strengthen important features that are strongly correlated with the processing state.
[0014] Preferably, the online monitoring and early warning module deploys a pre-trained model on the edge computing terminal and uses FastAPI to build an HTTP interface. When the machine tool outputs a cutting start signal and the tool enters the effective range of finishing, it triggers continuous sampling of vibration signals. When 224 effective data points are collected continuously, it triggers model inference. The single-sample inference delay is controlled within 10ms. When the vibration amplitude or roughness prediction value exceeds the threshold, it triggers an emergency warning.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention solves the problem that traditional prediction methods are difficult to characterize the strong nonlinear coupling relationship between milling vibration and roughness. By using GAF encoding, one-dimensional non-stationary vibration signals are losslessly converted into two-dimensional image features, which fully preserves the temporal characteristics and local abrupt change information of the vibration signals.
[0016] The SE-ResNet18 model used in this invention combines residual connections and channel attention mechanisms, achieving high prediction accuracy after 50 training rounds and exhibiting good generalization ability across speed ranges.
[0017] This invention enables precise point-by-point prediction of surface roughness based on real-time vibration signals during machining. It overcomes the limitations of traditional prediction methods that rely solely on fixed machining parameters and are difficult to characterize the dynamic characteristics of the milling process. This invention provides a feasible technical path for real-time monitoring of the milling quality and optimization of process parameters for aero-engine compressor blades, and has significant engineering application value for ensuring the machining quality and airworthiness compliance of aero-engine life-limited components. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall framework of the system of the present invention; Figure 2 This is a spatiotemporal mapping diagram of vibration signals during the entire milling process of the titanium alloy centrifugal impeller for aero-engines, as presented in this invention. Figure 3 This is an example diagram of the two-dimensional feature transformation of the Gram angle field of the milling vibration signal according to the present invention; Figure 4 This is the roughness prediction confusion matrix diagram for the optimal training round of the SE-ResNet prediction model of the present invention; Figure 5 The graph shows the evolution of accuracy and loss function during the training process of the SE-ResNet prediction model of this invention. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0020] This invention discloses a prediction system for milling vibration and roughness of titanium alloy centrifugal impellers for aero-engines. By constructing a closed loop encompassing data acquisition, preprocessing, model training, and online monitoring, this system solves the problems of traditional prediction methods, such as difficulty in characterizing the strong nonlinear coupling relationship between milling vibration and roughness, low prediction accuracy, and poor generalization ability. It achieves real-time, high-precision prediction of impeller surface roughness, providing technical support for quality control and airworthiness compliance verification of life-limited components for aero-engines.
[0021] I. System Overall Architecture and Workflow refer to Figure 1 This prediction system includes a data acquisition module, a data preprocessing and feature conversion module, an SE-ResNet deep learning prediction model module, and an online monitoring and early warning module. The modules are seamlessly connected through standardized data interfaces to form a complete process for predicting and controlling processing quality.
[0022] The system first uses a data acquisition module to synchronously collect vibration acceleration signals at the blade root and surface roughness data at the corresponding machining location during the milling process. After acquisition, the data is automatically transmitted to the preprocessing and feature transformation module, which sequentially performs outlier removal, spatiotemporal transformation, frequency domain feature extraction, and GAF two-dimensional image encoding to construct a standardized dataset with a one-to-one mapping between vibration and roughness. Subsequently, the dataset is input into the SE-ResNet deep learning model module for training and validation to obtain the optimal prediction model weights. Finally, the trained model is deployed to the online monitoring and early warning module to achieve real-time prediction and anomaly warning of surface roughness during the machining process.
[0023] II. Data Acquisition Module The data acquisition module is the foundation of the entire system. It is responsible for acquiring the raw vibration signals and surface roughness data during the milling process, providing a high-quality data source for subsequent model training.
[0024] 2.1 Object and Material Properties This embodiment focuses on the impeller of a centrifugal compressor for a certain type of aero-engine. The impeller material is TC11 titanium alloy, which has become the mainstream manufacturing material for centrifugal compressor impeller blades due to its high specific strength and excellent high-temperature resistance. The initial microstructure and actual chemical composition of the specimen were measured using a scanning electron microscope (ZEISSEVO18, 20kV) equipped with EDS. The measured chemical composition is shown in Table 1. The mechanical property data of the specimen obtained by room temperature tensile testing are shown in Table 2.
[0025] Table 1. Measured chemical composition (wt.%) of TC11 titanium alloy Element Al Mo Zr Si Ti content 5.20 3.80 1.70 0.20 margin Table 2 Tensile mechanical properties of TC11 titanium alloy at room temperature Yield strength (MPa) Tensile strength (MPa) Elongation (%) Young's modulus (GPa) Poisson's ratio 917 987 20 114 0.33 2.2 Equipment and Tooling Setup 1. Machining equipment: A five-axis linkage machining center is used for impeller milling. Carbide ball-end taper cutters are used in the blade finishing stage. The tool material and structural parameters are consistent with the actual machining conditions of the titanium alloy centrifugal impeller to improve tool rigidity and avoid tool interference risks.
[0026] 2. Vibration signal acquisition equipment: The American DYTRAN3225F IEPE accelerometer is used. Based on the conclusions of previous research, the blade root structure has greater stiffness and the vibration signal stability is significantly better than that of the blade tip. Therefore, the sensor is attached to the root of the impeller. The sampling frequency of the matching data acquisition device is set to 20000Hz.
[0027] 3. Surface Roughness Testing Equipment: A laser three-dimensional confocal microscope is used to quantitatively test the surface roughness of the milled blades. During the test, contour height morphology data are first collected from multiple test points on the blade surface. After selecting a specified measurement length as the evaluation interval according to the surface roughness evaluation specification, the roughness characteristic value Ra is calculated based on the standard calculation formula of the contour arithmetic mean deviation.
[0028] 4. Auxiliary equipment: The machine tool CNC system is equipped with an OPCUA interface for real-time acquisition of spindle speed, feed rate, tool space coordinates, and machining status signals, providing support for data synchronization and online monitoring.
[0029] 2.3 Implementation of Data Acquisition Plan 2.3.1 Preliminary experiments and determination of process parameters Before formal data collection, a pre-milling test was conducted. The feed rate, axial depth of cut, and radial depth of cut were set to fixed values (180 mm / min, 2 mm, and 0.1 mm, respectively). The spindle speed was used as the only variable, and four sets of gradient levels (900 rpm, 1800 rpm, 3600 rpm, and 4500 rpm) were set up for comparative test cutting.
[0030] Waveform characteristic analysis of vibration signals at different spindle speeds revealed that the spindle speed directly determines the periodic entry and exit frequency of the milling cutter teeth and the workpiece. Changes in spindle speed cause an overall shift in the milling characteristic frequency and its harmonic components within the frequency domain. Low spindle speeds combined with large cutting depths easily induce strong self-excited chatter, causing the vibration to change from stable, periodic, small-amplitude fluctuations to low-frequency, large-amplitude, violent oscillations. The time-domain waveform exhibits typical beat vibration characteristics or quasi-periodic large-amplitude fluctuations, and the dominant vibration frequency is often close to the natural frequency of the milling process system. Specifically, at spindle speeds of 900 rpm and 1800 rpm, the milling process exhibits stable cutting dynamics.
[0031] 2.3.2 Formal Experimental Data Acquisition and Synchronization Based on the preliminary experimental results, two sets of process parameters were selected to construct a dataset to ensure that the operating conditions of the training set and the validation set did not overlap, in order to verify the generalization ability of the model under variable speed conditions: Training set operating conditions: spindle speed 900 rpm, feed rate 180 mm / min, axial cutting depth 2 mm, radial cutting depth 0.1 mm, 1200 sets of valid samples were collected.
[0032] Validation set conditions: spindle speed 1800 rpm, other process parameters are completely consistent with the training set, 400 valid samples were collected.
[0033] During data acquisition, the spatial coordinates and machining time of each sampling point were synchronously recorded through the OPCUA interface of the machine tool CNC system. A point coordinate mapping index table between vibration signals and surface roughness data was established to provide a basis for subsequent data matching. The global spatiotemporal mapping results of the vibration acceleration signals obtained during the entire stable milling process of the blade are shown below. Figure 2 As shown, the signal time domain range covers the entire interval from the start of milling to the end of machining, and completely preserves the vibration amplitude fluctuation characteristics corresponding to different machining positions, providing a standardized raw data basis for subsequent signal preprocessing and feature conversion.
[0034] III. Data Preprocessing and Feature Transformation Module This module is responsible for cleaning, standardizing, and transforming the raw data output by the data acquisition module, converting one-dimensional non-stationary vibration signals into two-dimensional image features that can be recognized by deep learning models. It is a key link connecting data acquisition and model training.
[0035] 3.1 Surface Roughness Data Processing First, the surface roughness data collected by the laser three-dimensional confocal microscope was standardized: taking a blade specimen with a total processing area of 70 mm as an example, the contour height data of the entire processing area was divided into equal segments according to a sampling length of 0.28 mm, resulting in 250 sets of effective contour data units. The arithmetic mean deviation Ra value of the contour corresponding to each unit was then calculated.
[0036] To achieve precise matching between roughness grading and vibration signal characteristics, the frequency of the tool vibration acceleration signal during milling is used as the core basis. The smallest resolvable unit of vibration signal amplitude variation, 0.11 μm, is selected as the uniform interval. The measured roughness values are divided into 17 continuous grading intervals, corresponding to category codes 0 to 16, as shown in Table 3. This step size setting ensures that each roughness grading interval's corresponding machining process state (including tool chatter, chip breakage frequency, etc.) has its own unique vibration characteristic mode, effectively avoiding frequency domain aliasing and random noise interference of the vibration signal.
[0037] Table 3 Correspondence between surface roughness grading intervals and category codes Category coding Roughness Ra classification interval (μm) Category coding Roughness Ra classification interval (μm) 0 [1.00,1.11] 9 [1.99,2.10] 1 [1.11,1.22] 10 [2.10,2.21] 2 [1.22,1.33] 11 [2.21,2.32] 3 [1.33,1.44] 12 [2.32,2.43] 4 [1.44,1.55] 13 [2.43,2.54] 5 [1.55,1.66] 14 [2.54,2.65] 6 [1.66,1.77] 15 [2.65,2.76] 7 [1.77,1.88] 16 [2.76,2.87] 8 [1.88,1.99] - - For example, the measured roughness value of the measuring point at the -20mm position in the middle of the machining area under the stable milling condition of the blade is 1.783μm, which belongs to the interval [1.77,1.88] and corresponds to category code 7.
[0038] 3.2 Vibration signal preprocessing The original vibration data is a discrete dataset of spatial coordinate positions and vibration acceleration amplitudes, which needs to be converted into standardized time-domain samples through the following preprocessing steps: 1. Spatial Domain to Time Domain Conversion: To adapt to the Fast Fourier Transform (FFT)'s requirement for uniformly sampled time-domain signals as input, the original discrete spatial domain data is converted into a continuous time-domain sampling sequence. The conversion formula is as follows: in, For the first The processing time corresponding to each data point For the first The spatial coordinates of each data point The constant feed rate of the milling cutter (180 mm / min in this embodiment) is used. The origin of the spatial coordinate system is the starting milling position of the blade's machining area to be measured. ), corresponding to the starting point of the processing time .
[0039] 2. Outlier Removal and Data Cleaning: Using the 3σ criterion, the mean amplitude μ and standard deviation σ of a single-segment time-domain vibration signal are calculated. Outlier sampling points whose amplitudes exceed the interval [μ−3σ,μ+3σ] are removed and replaced with linear interpolation results from the five consecutive valid sampling points. NaN and null values are also linearly interpolated using the five consecutive data points. The criteria for invalid samples are: the presence of more than five consecutive null values or zero values, or outliers exceeding the 3σ criterion with a proportion exceeding 10%.
[0040] 3. Sample segmentation: To adapt to the input requirements of subsequent deep learning models, 224 consecutive time-domain data points are used as an independent single-segment sample unit to form a full sample set.
[0041] 4. Frequency Domain Feature Extraction: The one-dimensional time-domain vibration signal is converted into a frequency-domain signal using Fast Fourier Transform (FFT), and the peak characteristic frequencies in the spectral curve are extracted. The FFT implementation parameters are as follows: Sampling frequency ; Single-segment FFT analysis sample length point; Spectrum analysis range: 0 to 10000 Hz; Frequency resolution ; Window function: Hanning window, used to suppress spectral leakage. Preprocessing: Remove the values corresponding to the 50Hz power frequency and its harmonics, retaining only the characteristic frequencies strongly correlated with milling process parameters. In this embodiment, the measured fundamental frequency of a typical feature is approximately 650 Hz, corresponding to a feature period of 1.53 ms, providing a benchmark for subsequent periodic matching processing of the sequence.
[0042] 3.3 Two-dimensional feature transformation of Gram's corner field (GAF) To address the time-domain characteristics of vibration signals—globally stationary but with strong local abrupt changes—a Gram angle field (GAF) is introduced to achieve lossless conversion from one-dimensional time-series signals to two-dimensional images. This method can characterize the local abrupt changes of vibration signals at high resolution, providing fine-grained transient processing information input for deep learning models. The specific conversion steps are as follows: 1. Sequence linear normalization: This involves normalizing a single one-dimensional time-domain vibration acceleration sequence. ( These are the first and second parts of the original sequence. The original values of vibration acceleration at each location are normalized to the interval [-1, 1] to eliminate the influence of amplitude dimensions. For the normalized first One data value, For the first in the original sequence The original values of vibration acceleration at each location.
[0043] 2. Polar coordinate mapping: Converting the normalized sequence into polar coordinate encoding, using angle values to represent the signal amplitude: in, The value range is [0, π].
[0044] 3. Cosine Similarity Matrix Construction: By calculating the cosine of the sum of any two angle values, a 224×224 two-dimensional GAF matrix is constructed. They are the first The, the The angle values of each data point after polar coordinate mapping The first in the Gram angle field matrix Line number The elements of the column.
[0045] The matrix elements take values in the range [-1, 1].
[0046] 4. Gray-scale mapping: The values of the GAF matrix are linearly mapped to a gray-scale value range of 0 to 255, generating a single-channel gray-scale image, which serves as the input feature of the SE-ResNet model. A typical two-dimensional GAF feature image of a vibration signal generated through the above conversion process is shown below. Figure 3As shown, different texture densities and brightness distributions in the image correspond to different temporal fluctuation patterns of the vibration signal, which can intuitively distinguish the differences in vibration characteristics between steady cutting and micro-flutter states, providing standardized input features with clear physical meaning for subsequent deep learning models.
[0047] 3.4 Data Augmentation and Final Dataset Construction To improve the model's generalization ability, the following data augmentation operations were performed on the GAF images to enrich the sample dimensions while preserving the core vibrational features: Random horizontal flip: Perform a horizontal mirror transformation. Random vertical flip: Perform a vertical mirror transformation Random rotation: Set the random rotation angle range to ±15°, and fill blank areas with zero values. Brightness and contrast adjustment: Brightness adjustment range ±0.2, contrast adjustment range ±0.2 After data augmentation, sample matching was performed: 224 data points from the vibration signal sampling were divided into intervals, with a corresponding interval length of 0.28 mm. The average surface roughness within each interval was taken as the label for that sample, and the coordinates of the midpoint of the interval were recorded as the location label, ensuring a strict one-to-one correspondence between each vibration signal sample and a roughness category label. Finally, 1200 samples from the 900 rpm condition were divided into a training set, and 400 samples from the 1800 rpm condition were divided into an independent validation set, completing the construction of the vibration-roughness mapping dataset.
[0048] IV. SE-ResNet18 Deep Learning Prediction Model Module This module is the core of the entire system, responsible for learning the nonlinear mapping relationship between milling vibration signals and surface roughness, thereby achieving high-precision roughness prediction. This invention employs the SE-ResNet18 model with a fusion channel attention mechanism. This model combines the advantages of residual connection structures in solving the gradient vanishing problem in deep networks with the adaptive enhancement of key features by the channel attention mechanism.
[0049] 4.1 Overall Model Structure The SE-ResNet18 model, based on the traditional ResNet18, incorporates an SE channel attention module into each residual basic unit. This module adaptively adjusts the weights of feature channels, strengthening important features strongly correlated with the processing state and suppressing redundant secondary features. The complete structural parameters of each layer of the model are shown in Table 4.
[0050] Table 4. Structural parameters of each layer of the SE-ResNet18 network level name Core Operations Number of input channels Number of output channels kernel size Step length filling Input layer Image input 1 64 7×7 2 3 Conv1 Convolution + BN + ReLU 1 64 7×7 2 3 MaxPool Max pooling 64 64 3×3 2 1 Layer1 SEBasicBlock×2 64 64 3×3 1 1 Layer2 SEBasicBlock×2 64 128 3×3 2 1 Layer 3 SEBasicBlock×2 128 256 3×3 2 1 Layer 4 SEBasicBlock×2 256 512 3×3 2 1 AvgPool Adaptive average pooling 512 512 1×1 1 0 Dropout Random inactivation 512 512 - - - FC Fully Connected Layer Linear transformation 512 17 - - - The model input layer receives a single-channel GAF grayscale image with a size of 224×224×1 as input features; the feature extraction module is composed of multiple layers of SE-ResNet basic units stacked together, which gradually enhances and abstracts the vibration feature signals; after feature extraction, the network is sequentially connected to a global average pooling layer and a fully connected layer, and the training is optimized through end-to-end supervised learning, finally outputting the predicted probabilities of 17 roughness categories.
[0051] 4.2 Working principle of SE channel attention module The SE module adaptively adjusts feature weights by explicitly modeling channel dependencies, specifically through two steps: squeezing and excitation. Compression steps: The feature map of each channel is compressed into a scalar using a global average pooling operation. This characterizes the global importance of the channel: in, and These represent the height and width of the feature map, respectively. Indicates the first Each channel is located in The value at that location.
[0052] Activation steps: Generate channel weight vectors through two fully connected layers. Recalibrate the original feature map: in, For the output of the compression step, and This is the weight matrix. Represents the ReLU activation function. This represents the Sigmoid activation function. Represents the set of real numbers. For dimensional values, This represents the channel compression ratio. The final recalibrated feature map is as follows: in, This is the original feature map. This is the recalibrated feature map.
[0053] 4.3 Model Training and Performance Validation 4.3.1 Training Hyperparameter Settings The complete hyperparameters for model training are shown in Table 5.
[0054] Table 5 Model Training Hyperparameter Settings Hyperparameter categories Specific parameter values Batch size (BATCH_SIZE) 32 Maximum number of training epochs (EPOCHS) 110 (Triggered early stop, premature termination) Initial learning rate <![CDATA[1×10 -4 ]]> Optimizer <![CDATA[Adam optimizer, β1 = 0.9, β2 = 0.999]]> Weight decay coefficient (L2 regularization) 1.00E-04 loss function Cross-entropy loss function Learning rate decay strategy It decays to 90% of its current value every 10 epochs. Early stop mechanism With a patience value of 8, training stops if the loss does not decrease after 8 consecutive epochs on the validation set. Best model saves benchmark Minimum validation set loss Software environment Python 3.8+, PyTorch 1.10+, torchvision 0.11+, CUDA 11.3+, with dependencies including pandas, scikit-learn, PIL, and tqdm. Hardware environment CPU: Intel(R) Core(TM) i5-14400F; GPU: NVIDIA RTX 5060 (≥24GB VRAM); RAM: ≥32GB 4.3.2 Training Process and Result Analysis The constructed vibration-roughness dataset was input into the SE-ResNet18 model for training. During the training process, the accuracy and loss function changes of the training and validation sets were monitored simultaneously.
[0055] 1. Convergence Analysis: The model reached engineering convergence in the 50th training epoch. At this point, the validation set classification accuracy reached a peak of 90.33%, with fluctuations of less than 1% over five consecutive training epochs. The training set loss remained low with no significant fluctuations, and the test set loss no longer showed a significant decrease with training iterations. The confusion matrix between the true roughness and predicted roughness of the test set in the 50th training epoch is shown below. Figure 4 As shown, the diagonal elements of the matrix account for a high proportion and are concentrated in distribution. There are only a few misjudgments in adjacent roughness grade intervals, indicating that the model can accurately predict the roughness grade under most processing conditions, and the overall performance meets the accuracy requirements of the processing site.
[0056] Local Polyak-Lojasiewicz (PL) conditional analysis shows that during the 39th to 48th training epochs, the squared norm of the model gradient remained stable in the range of 0.8 to 1.2 and was always greater than the PL condition threshold, satisfying the local PL convergence condition and exhibiting stable linear convergence characteristics.
[0057] 2. Overfitting Analysis: Reference Figure 5 When the number of training epochs increased to 80, the overall classification accuracy of the model dropped to 78.62%, and the validation set loss increased significantly, while the training set accuracy and loss remained at the optimal level, which is consistent with the typical characteristics of overfitting in deep learning models. The core reason is that the random noise and redundant features in the dataset accumulated in the later stages of training, and the model overlearned the non-essential noise features in the training set, rather than the inherent mapping between milling vibration signals and surface roughness. At the same time, the model had converged to the global optimum after 50 training epochs, and the subsequent unconstrained continuous iterations caused the model to overfit to the features specific to the training set data, losing its ability to generalize to unknown test set data.
[0058] 3. Confusion Matrix Analysis: The confusion matrix of the 50-round training model has concentrated diagonal elements, a low misclassification rate, stable inter-class discrimination for each roughness category, and stronger robustness to small fluctuations in features; while the confusion matrix of the 80-round overfitting model has significantly more off-diagonal elements, a large increase in misclassified samples, and is highly sensitive to small perturbations in the input data.
[0059] 4.3.3 Model Performance Quantification Indicators The SE-ResNet18 model trained for 50 rounds achieved a classification accuracy of 90.33% on the independent validation set, accurately predicting the impeller surface roughness level based on processing vibration signals. Since this model is a classification model rather than a regression model, and the roughness intervals are manually defined with small inter-class differences, it primarily meets the engineering requirement of online judgment of roughness level compliance based on vibration signals. Therefore, regression indices such as MAE, RMSE, and R² were not calculated.
[0060] V. Online Monitoring and Early Warning Module This module is responsible for deploying the trained SE-ResNet18 model to industrial sites to achieve real-time prediction and early warning of surface roughness during processing, which is a key link in the system's engineering application.
[0061] 5.1 System Deployment Plan 1. Hardware deployment: The IEPE accelerometer is attached to the impeller root, and the sampling frequency of the matching data acquisition device is set to 20000Hz; the spindle speed, feed rate, tool space coordinates, and machining status signals are collected in real time through the OPCUA interface of the machine tool CNC system.
[0062] 2. Software Deployment: Deploy the pre-trained SE-ResNet model on the edge computing terminal, use FastAPI to build an HTTP interface, and send the dataset received and cleaned by the front end to the back end model for inference.
[0063] 5.2 Synchronous Triggering Logic for Signal Acquisition and Model Inference When the machine tool outputs a cutting start signal and the tool enters the preset finishing effective range, it triggers continuous sampling of vibration signals and labels each sampling point with spatial coordinates.
[0064] When the number of consecutive collection points reaches 224 and the sample coverage area is completely within the effective processing area, a model inference is triggered, and the next sample collection is carried out simultaneously.
[0065] The data acquisition stops when the machine tool outputs a pause or end signal, or when the tool leaves the effective range.
[0066] An emergency warning is triggered when the vibration amplitude or surface roughness prediction value exceeds the safety threshold or when a machine tool malfunction alarm occurs.
[0067] 5.3 Real-time Prediction Performance and Applicable Operating Conditions The latency of single-sample inference at the edge is controlled within 10ms, which can meet the requirements of real-time monitoring of the machining process. The system maps the prediction results to the corresponding Ra interval, associates them with machining coordinates, and generates a position-roughness mapping relationship, providing data support for machining quality control.
[0068] This prediction system is applicable to five-axis simultaneous milling of centrifugal compressor impeller blades (including precision-machined areas such as blade profiles and blade roots) for aero-engines. The applicable material is TC11 titanium alloy (nominal chemical composition Ti-6.5Al-3.5Mo-1.5Zr-0.3Si). Under the current experimental verification conditions, the applicable process parameters range is: spindle speed 900 rpm to 4500 rpm, feed rate 180 mm / min, axial depth of cut 2 mm, radial depth of cut 0.1 mm, and surface roughness Ra value 1.00 μm to 2.80 μm. The applicable parameter range can be further expanded by conducting experiments under more operating conditions and increasing the dataset.
[0069] VI. Overall System Effects and Beneficial Effects Through the coordinated operation of the above four modules, this invention constructs a complete prediction system for the correlation between milling vibration and surface roughness of titanium alloy centrifugal impellers for aero-engines, realizing full-process automation from raw data acquisition to real-time early warning of machining quality.
[0070] This invention addresses the challenge of traditional prediction methods in characterizing the strong nonlinear coupling between milling vibration and surface roughness. By employing GAF encoding, it losslessly converts one-dimensional non-stationary vibration signals into two-dimensional image features, fully preserving the temporal characteristics and local abrupt changes of the vibration signal. The SE-ResNet18 model, combining residual connections and channel attention mechanisms, achieves a prediction accuracy of 90.33% after 50 training rounds and demonstrates good generalization ability across various speed ranges. It enables precise point-by-point prediction of surface roughness based on real-time vibration signals during machining, overcoming the limitations of traditional prediction methods that rely solely on fixed machining parameters and struggle to characterize the dynamic characteristics of the milling process. This provides a feasible technical path for real-time monitoring of the milling quality and optimization of process parameters for aero-engine compressor blades, and has significant engineering application value for ensuring the machining quality and airworthiness compliance of aero-engine life-limited components.
[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A system for predicting chatter and roughness in the milling of a titanium alloy centrifugal impeller for an aeroengine, the system comprising: It includes a data acquisition module, a data preprocessing and feature transformation module, an SE-ResNet deep learning prediction model module, and an online monitoring and early warning module; The data acquisition module is used to synchronously acquire the vibration acceleration signal at the blade root and the surface roughness data at the corresponding machining position during the milling process. The data preprocessing and feature transformation module is used to remove outliers, perform spatiotemporal transformation, extract frequency domain features, and encode two-dimensional images of Gram angle field in the original data to construct a standardized dataset with a one-to-one mapping between vibration and roughness. The SE-ResNet deep learning prediction model module is used to train the SE-ResNet18 model with a fusion channel attention mechanism based on a standardized dataset to obtain the optimal prediction model weights. The online monitoring and early warning module is used to deploy the trained model to realize real-time prediction and early warning of surface roughness during the processing. The modules are connected through standardized data interfaces.
2. The prediction system of claim 1, wherein, The data acquisition module uses the centrifugal compressor impeller of an aero-engine made of TC11 titanium alloy as the data acquisition object. It uses an IEPE accelerometer attached to the root of the impeller to collect vibration signals, with the sampling frequency set to 20000Hz. It uses a laser three-dimensional confocal microscope to collect surface roughness data, and synchronously collects spindle speed, feed rate, tool space coordinates, and machining status signals through the OPCUA interface of the machine tool CNC system.
3. The prediction system of claim 2, wherein, Before formal data acquisition, the data acquisition module conducts a pre-milling trial cut test, setting a gradient level with the spindle speed as the single variable to conduct a comparative trial cut. Based on the pre-experiment results, a training set and a validation set with non-overlapping working conditions are selected. The working conditions of the training set are a spindle speed of 900 rpm, a feed rate of 180 mm / min, an axial cutting depth of 2 mm, and a radial cutting depth of 0.1 mm. The working conditions of the validation set are a spindle speed of 1800 rpm, with other process parameters consistent with the training set. During the acquisition process, a point coordinate mapping index table of vibration signal and surface roughness data is established.
4. The prediction system of claim 1, wherein, The data preprocessing and feature conversion module standardizes the surface roughness data, divides the contour height data of the entire processing area into equal segments according to a sampling length of 0.28 mm, calculates the arithmetic mean deviation Ra value of the contour of each unit, and divides the measured roughness value into 17 continuous graded intervals with an interval of 0.11 μm, corresponding to category codes 0 to 16.
5. The prediction system of claim 1, wherein, The data preprocessing and feature conversion module preprocesses the vibration signal: the original spatial domain discrete data is converted into a continuous time domain sampling sequence through the feed speed, abnormal sampling points are eliminated using the 3σ criterion and linear interpolation is performed, and 224 continuous time domain data points are used as a sample unit. The frequency domain feature frequencies are extracted using fast Fourier transform, the 50Hz power frequency and its harmonics are eliminated, and the feature frequencies that are strongly correlated with the milling process parameters are retained.
6. The prediction system of claim 1, wherein, The data preprocessing and feature conversion module uses Gram angle field to convert one-dimensional time-domain vibration signals into two-dimensional images: the single vibration sequence is normalized to the interval [-1,1], the normalized sequence is converted into angle values through polar coordinate mapping, the cosine of the sum of any two angle values is calculated to construct a 224×224 two-dimensional matrix, and the matrix values are linearly mapped to the gray value range of 0 to 255 to generate a single-channel grayscale image.
7. The prediction system of claim 6, wherein, The data preprocessing and feature conversion module performs data augmentation on the generated Gram angle field image, including random horizontal flipping, random vertical flipping, random rotation within ±15°, and adjustment of brightness and contrast within ±0.
2. The module matches the corresponding roughness average value as a label according to an interval of 224 data points to construct a training set and an independent validation set.
8. The prediction system of claim 1, wherein, The SE-ResNet deep learning prediction model module adopts the SE-ResNet18 model, which integrates the SE channel attention module into each residual basic unit of the traditional ResNet18. The model input is a single-channel grayscale image of 224×224×1. After feature extraction by multiple SE-ResNet basic units, the predicted probabilities of 17 roughness categories are output through global average pooling layers and fully connected layers.
9. The prediction system according to claim 8, characterized in that, The SE channel attention module compresses the feature map of each channel into a scalar representation of the global importance of the channel through global average pooling. It generates channel weight vectors through two fully connected layers to recalibrate the original feature map, adaptively adjusting the weights of the feature channels and strengthening important features that are strongly correlated with the processing state.
10. The prediction system according to claim 1, characterized in that, The online monitoring and early warning module deploys a pre-trained model on the edge computing terminal and uses FastAPI to build an HTTP interface. When the machine tool outputs a cutting start signal and the tool enters the effective range of finishing, it triggers continuous sampling of vibration signals. When 224 effective data points are collected continuously, it triggers model inference. The single-sample inference delay is controlled within 10ms. When the vibration amplitude or roughness prediction value exceeds the threshold, it triggers an emergency warning.