Methods, systems, equipment and media for predicting surface integrity enhanced by water jet in thin-walled components

Abstract

This application discloses a method, system, equipment, and medium for predicting the surface integrity of thin-walled components during waterjet strengthening. This application enables real-time prediction of surface roughness, residual stress, microhardness, and other indicators of thin-walled components such as aero-engine blades and integral bladed disks during the strengthening process. The prediction results can provide a basis for online optimization of the strengthening process, effectively solving the problems of difficulty in monitoring the state of the waterjet strengthening process, complexity in characterizing the strengthening effect, and limitations in the prediction accuracy and generalization ability of traditional monitoring methods relying on feature engineering under complex working conditions.

Description

Technical Field

[0001] This application belongs to the field of surface engineering technology, and in particular relates to a method, system, equipment and medium for predicting the surface integrity of thin-walled parts by water jet enhancement. Background Technology

[0002] Waterjet strengthening technology enhances the mechanical properties of materials by comprehensively optimizing their surface integrity, providing a reliable guarantee for the long-term safe service of typical thin-walled components such as aero-engine blades and bladed disks. However, due to their weak stiffness, thin-walled components such as blades inevitably experience vibration during processing, adding nonlinearity and randomness to the strengthening process and posing a significant challenge to the uniformity and consistency of the surface quality after strengthening.

[0003] Surface integrity refers to the comprehensive characteristics of the surface quality (geometric features and physical properties) of a part after machining, which has a significant impact on the part's performance. It includes various indicators such as surface roughness, residual stress, and microhardness. Taking surface roughness as an example, this indicator is one of the most intuitive features for evaluating the surface quality of a part and is an important criterion for judging the quality of part machining. Currently, traditional characterization methods mainly include probe contact measurement and optical measurement. Although these methods can obtain accurate surface roughness values, they are essentially "post-evaluation" methods, with low detection efficiency and unable to provide real-time feedback for adjusting process parameters. Similarly, the characterization tests for indicators such as residual stress and microhardness are currently all offline, off-machine processes, unable to be coupled in real-time during machining, making it difficult to timely compensate for or repair quality deviations.

[0004] Furthermore, waterjet enhancement processes are often accompanied by severe turbulence, bubbles, and noise, creating a highly complex processing environment. Currently, there are few reports on online monitoring and surface integrity prediction during waterjet enhancement. Existing monitoring research mainly focuses on waterjet cutting, relying heavily on traditional feature engineering to extract effective information from signals and establishing a correspondence between these signals and processing states through empirical fitting or machine learning methods. However, under highly noisy and complex conditions, these methods struggle to effectively separate features strongly correlated with surface integrity, resulting in poor generalization and robustness, and failing to meet the requirements for high-precision online prediction. Summary of the Invention

[0005] This application provides a method, system, device, and medium for predicting the surface integrity of waterjet-enhanced thin-walled components, in order to solve the problems of difficulty in monitoring the state of the waterjet enhancement process, complexity in characterizing the enhancement effect, and the limited prediction accuracy and generalization ability of traditional monitoring methods that rely on feature engineering under complex working conditions.

[0006] To achieve the above objectives, this application provides the following technical solution.

[0007] According to a first aspect of this application, embodiments of this application provide a method for predicting the surface integrity of a thin-walled component enhanced by water jetting, the method comprising: Acquire vibration signals of thin-walled components during waterjet enhanced processing; The vibration signal is preprocessed for noise reduction using variational mode decomposition, including: iteratively optimizing the hyperparameters of the variational mode decomposition using particle swarm optimization algorithm, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal; The denoised time-domain vibration signal is subjected to spectrum conversion to obtain a frequency-domain vibration signal. The time-domain vibration signal and the frequency-domain vibration signal are then normalized to construct the input data for the prediction model. The prediction model input data is input into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model includes a parallel time-domain feature extraction branch and a frequency-domain feature extraction branch, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branch and the frequency-domain features extracted by the frequency-domain feature extraction branch; The time-frequency parallel deep learning model outputs the predicted surface integrity value of the thin-walled component at the current processing position.

[0008] In some possible implementations, the iterative optimization of the hyperparameters of the variational mode decomposition using a particle swarm optimization algorithm includes: The penalty factor and the number of decomposition layers of the variational mode decomposition are selected as the particle position parameters of the particle swarm. The vibration signal is subjected to variational mode decomposition based on the particle position parameters to obtain modal components, and the average value of the envelope entropy of each modal component is calculated as the fitness value. By iteratively updating the particle position and velocity until the fitness value meets the preset iteration stopping condition, the globally optimal penalty factor and decomposition layer number are determined. The step of decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters includes: decomposing the vibration signal using the globally optimal penalty factor and the number of decomposition layers, removing the high-frequency modal components, and reconstructing the remaining modal components.

[0009] In some possible implementations, the time-frequency parallel deep learning model includes: a time-domain feature extraction branch comprising multiple convolutional layers and long short-term memory network layers connected in sequence, used to extract local features and sequence dependency features of the time-domain vibration signal; a frequency-domain feature extraction branch comprising multiple convolutional layers and fully connected layers connected in sequence, used to extract energy distribution features of the frequency-domain vibration signal; and a multi-head attention mechanism module used to concatenate the feature vectors output by the time-domain feature extraction branch and the frequency-domain feature extraction branch and then perform feature weighted fusion.

[0010] In some possible implementations, the convolutional layers in both the temporal feature extraction branch and the frequency domain feature extraction branch are configured with convolutional kernel sizes decreasing sequentially; wherein, the convolutional layers in the early stages of feature extraction use larger convolutional kernels to extract long-distance dependent features, while the convolutional layers in subsequent levels use smaller convolutional kernels to extract deep and fine features.

[0011] In some possible implementations, the training process of the time-frequency parallel deep learning model includes: Vibration signal segments during the waterjet-enhanced stabilization process were acquired, and the measured surface integrity values ​​corresponding to the vibration signal segments were used as labels to construct a training dataset. The time-domain vibration signal and the corresponding frequency-domain vibration signal are mapped to a preset interval using the Min-Max normalization method, and the training dataset is normalized. The time-frequency parallel deep learning model is trained using the training dataset, the network weights are updated using the Adam optimizer, and the mean squared error is used as the loss function to calculate the prediction error between the predicted value and the label until the time-frequency parallel deep learning model converges.

[0012] In some possible implementations, the measured surface integrity value includes at least surface roughness; the label of the training dataset is the average surface roughness obtained by an optical three-dimensional surface roughness measuring instrument under the corresponding processing technology; the predicted surface integrity value is the predicted surface roughness value under the current processing technology state.

[0013] In some possible implementations, the step of iteratively optimizing the hyperparameters of the variational mode decomposition using the particle swarm optimization algorithm is performed in the offline preparation stage before online prediction; in the online prediction stage, the vibration signal acquired in real time is subjected to variational mode decomposition and reconstruction using the globally optimal penalty factor and decomposition level determined in the offline preparation stage.

[0014] According to a second aspect of this application, embodiments of this application provide a waterjet-enhanced surface integrity prediction system for thin-walled components, the system comprising: The signal acquisition module is used to acquire vibration signals on the surface of thin-walled components during waterjet enhanced processing. The signal preprocessing module is used to perform noise reduction preprocessing on the vibration signal using variational mode decomposition, including: using particle swarm optimization algorithm to iteratively optimize the hyperparameters of the variational mode decomposition, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal. The feature construction module is used to perform spectrum conversion on the denoised time-domain vibration signal to obtain the frequency-domain vibration signal, and to normalize the time-domain vibration signal and the frequency-domain vibration signal to construct the input data for the prediction model. The predictive analysis module is used to input the prediction model input data into a pre-trained time-frequency parallel deep learning model, and output the surface integrity prediction value of the thin-walled component at the current processing position through the time-frequency parallel deep learning model; wherein, the time-frequency parallel deep learning model includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches.

[0015] According to a third aspect of this application, an embodiment of this application provides an electronic device, the electronic device including a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the thin-walled component waterjet enhanced surface integrity prediction method as described in any embodiment of this application.

[0016] According to a fourth aspect of this application, embodiments of this application provide a computer-readable storage medium storing computer instructions that, when executed, implement the thin-walled component waterjet enhanced surface integrity prediction method as described in any embodiment of this application.

[0017] This application provides a method, system, device, and medium for predicting the surface integrity of thin-walled components during waterjet strengthening. It enables real-time prediction of surface roughness, residual stress, and microhardness in thin-walled components such as aero-engine blades and integral bladed disks during the strengthening process. The prediction results provide a basis for online optimization of the strengthening process, effectively solving the problems of difficulty in monitoring the state of the waterjet strengthening process, complexity in characterizing the strengthening effect, and limitations in the prediction accuracy and generalization ability of traditional monitoring methods relying on feature engineering under complex working conditions. Compared with traditional offline testing methods, the method described in this application can monitor the surface integrity of the processing location in real time during the strengthening process and considers the nonlinear factors of the processing, facilitating integration with online process control, ensuring uniformity and consistency of processing quality, and significantly improving manufacturing efficiency.

[0018] Furthermore, this application introduces a particle swarm optimization algorithm to adaptively optimize the hyperparameters (penalty factor and number of decomposition layers) of Variational Mode Decomposition (VMD), ensuring the noise reduction effect of the VMD method and effectively eliminating the influence of high-frequency noise. In addition, compared with prediction methods such as multivariate regression analysis and traditional machine learning, the time-frequency parallel deep learning model used in this application avoids the limitations of traditional feature engineering on model performance and can adaptively extract effective information related to surface integrity from vibration signals from multiple dimensions. By combining an attention mechanism, the model improves the attention to key features, achieving efficient fusion of time-domain and frequency-domain signal features, thereby establishing a nonlinear mapping relationship between vibration signals and surface integrity indices, further enhancing the model's robustness and generalization ability; and under complex conditions of strong fluid turbulence noise interference, it can achieve high-efficiency and high-precision online prediction of surface integrity. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for predicting the surface integrity of thin-walled components using water jet enhancement, as provided in an embodiment of this application.

[0021] Figure 2 for Figure 1 The flowchart of the sub-steps of step S200 is shown.

[0022] Figure 3 This is a flowchart illustrating a method for predicting the surface integrity of thin-walled components using water jet enhancement, provided in a specific embodiment of this application.

[0023] Figure 4 A schematic diagram of the network architecture of the time-frequency parallel deep learning model provided in the embodiments of this application.

[0024] Figure 5 This is a schematic diagram of the architecture of a waterjet-enhanced surface integrity prediction system for thin-walled components provided in an embodiment of this application.

[0025] Figure 6 This is a schematic diagram illustrating an application scenario of the thin-walled component waterjet enhanced surface integrity prediction system provided in an embodiment of this application.

[0026] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] This application provides a method, system, device, and medium for predicting the surface integrity of thin-walled components using waterjet strengthening. The method is primarily applied to waterjet strengthening processes for thin-walled, weak-stiffness components such as aero-engine blades and integral bladed disks. Addressing the difficulties in condition monitoring caused by strong fluid turbulence noise during the processing and the poor generalization of traditional feature-engineering-dependent monitoring methods in existing technologies, this application's method specifically achieves online prediction of surface integrity by acquiring vibration signals from the surface of the thin-walled component during waterjet strengthening, performing noise reduction preprocessing on the signals using variational mode decomposition optimized by particle swarm optimization, constructing input data for a prediction model, and analyzing the data using a pre-trained time-frequency parallel deep learning model. Based on this, this application combines variational mode decomposition with deep learning, utilizing the time-domain and frequency-domain feature extraction branches in the time-frequency parallel deep learning model to extract features in parallel, and performing weighted fusion of features through a multi-head attention mechanism module. This enables the model to adaptively extract effective information related to surface integrity from the vibration signals from multiple dimensions, establishing a nonlinear mapping relationship between the vibration signals and surface integrity indices (such as surface roughness, residual stress, and microhardness). Therefore, this application solves the problems of difficult monitoring of the state of water jet enhancement process and complex characterization process of enhancement effect in the prior art, and overcomes the shortcomings of traditional monitoring methods that rely on feature engineering, which have limited prediction accuracy and generalization ability under complex working conditions. In particular, in the complex environment of strong fluid turbulence and noise interference, this application has strong robustness and generalization ability, and can achieve high-efficiency and high-precision online prediction of surface integrity. The prediction results can provide a basis for online optimization of the enhancement process.

[0029] This application provides a method for predicting the surface integrity of thin-walled components enhanced by waterjet technology. This method can be executed by a system for predicting the surface integrity of thin-walled components enhanced by waterjet technology, which can be integrated into an electronic device. The electronic device may include at least one of a terminal and a server. Specifically, the method for predicting the surface integrity of thin-walled components enhanced by waterjet technology provided in this application can be executed by a terminal, a server, or jointly by a terminal and a server with communication capabilities. In terms of implementation architecture, the terminal may include devices with computing capabilities such as tablet computers, laptops, and desktop computers, while the server may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, and big data and artificial intelligence platforms. In one embodiment, the system for predicting the surface integrity of thin-walled components enhanced by waterjet technology can be integrated into an electronic device such as a terminal or server to implement the method for predicting the surface integrity of thin-walled components enhanced by waterjet technology provided in this application. The following will describe the system in detail from the perspective of the system for predicting the surface integrity of thin-walled components enhanced by waterjet technology, which can be integrated into an electronic device, which may include computing devices such as a terminal or server.

[0030] See Figure 1 This is a flowchart illustrating the steps of a waterjet-enhanced surface integrity prediction method for thin-walled components according to an embodiment of this application. The method may include steps S100 to S500. It should be noted that, to address the problem of insufficient high-precision online prediction in existing technologies, this application may include an offline preparation stage and an online prediction stage.

[0031] Step S100: Obtain the vibration signal of the thin-walled component surface during the water jet strengthening process.

[0032] This process includes an offline preparation phase and an online prediction phase. Both the dataset needed to build the training model and the real-time prediction required for actual processing necessitate the acquisition of vibration signals.

[0033] In the offline preparation phase, this step is used to build the training dataset. Specifically, thin-walled components such as aero-engine blades and integral bladed disks are selected as processing objects, and they are strengthened under a preset waterjet strengthening process, while vibration signals are collected simultaneously.

[0034] In some embodiments, an accelerometer is placed on the surface of the part to be strengthened to obtain vibration characteristics of the component surface. The accelerometer can be fixed by bolts, magnetic adsorption, or special adhesive to ensure close contact between the sensor and the thin-walled component and stable signal transmission. Further, to ensure data representativeness and quality, it is preferable to extract vibration signal segments during the waterjet strengthening stabilization process. Specifically, vibration signals during the waterjet strengthening stabilization process (e.g., within 120 seconds of the stabilization process) are acquired as samples, and the sampling frequency can be set to 48 kHz. Furthermore, the measured surface integrity values ​​corresponding to the acquired vibration signal segments are used as labels to construct a training dataset.

[0035] In the online prediction phase, this step is used to acquire real-time input data for the model. Specifically, using the same sensor arrangement and acquisition parameters as in the offline preparation phase, vibration signals at the current processing location are acquired in real time. These real-time acquired vibration signals will be used as the processing object in subsequent steps to predict the current surface integrity status in real time.

[0036] In some embodiments, the measured surface integrity value includes at least surface roughness. In other embodiments, the measured surface integrity value includes at least one of surface roughness, residual stress, and microhardness. Specifically, when the prediction target is surface roughness, the labels of the training dataset can be measured by an optical three-dimensional surface roughness measuring instrument, and the average surface roughness under the corresponding process can be calculated; when the prediction target is residual stress, the labels of the training dataset can be characterized by an X-ray residual stress diffractometer; when the prediction target is microhardness, the labels of the training dataset can be characterized by a Vickers hardness tester. It should be noted that, since the surface quality is uniform in the stable processing stage under a fixed process, the obtained measured value labels can represent the true state of surface integrity under that process and vibration state. In one specific embodiment, when the prediction target is surface roughness, the predicted surface integrity value is the predicted surface roughness value under the current processing state. At this time, the average value label obtained above is used as the measured surface roughness value (i.e., the training target) under that process and vibration state. Similarly, in other embodiments, for residual stress or microhardness, their corresponding measured values ​​are used as labels.

[0037] Step S200: The vibration signal is preprocessed for noise reduction using variational mode decomposition, including: iteratively optimizing the hyperparameters of the variational mode decomposition using particle swarm optimization algorithm, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal.

[0038] Due to the vibrations caused by the weak stiffness of thin-walled components and the severe fluid turbulence noise during processing, it is difficult to extract effective features directly from the original signal. Therefore, this step of the application employs Variational Mode Decomposition (VMD) technology for adaptive noise reduction. It should be noted that the VMD method involves two main hyperparameters: the penalty factor α and the number of decomposition layers k. The number of decomposition layers k determines the number of modes after signal decomposition, reflecting the signal's resolution in the frequency domain. If k is less than the number of effective components in the original signal, mode aliasing will occur; if k is greater than the number of effective components, useless spurious components will be generated. The penalty factor α determines the bandwidth of the modal components and is inversely proportional to the bandwidth. If α is too small, information redundancy will occur between modal components; conversely, it may cause the loss of some information in the decomposed signal. Therefore, the penalty factor α and the number of decomposition layers k are crucial for avoiding mode aliasing and ensuring the effectiveness of VMD noise reduction.

[0039] In some embodiments, the step of iteratively optimizing the hyperparameters of the variational mode decomposition using a particle swarm optimization algorithm is performed during the offline preparation stage before online prediction. Specifically, this application uses a particle swarm optimization algorithm (PSO) to adaptively iteratively optimize the two hyperparameters (penalty factor α and decomposition level k) during the offline preparation stage. The specific optimization process includes: See Figure 2 Step S201: Initialize particle swarm parameters. This may specifically include initializing the number, position, and velocity of particles. In some embodiments, the number of particles may be set to 20. Furthermore, the penalty factor α and the number of decomposition layers k of the variational mode decomposition are selected as the particle position parameters of the particle swarm.

[0040] Step S202: Calculate the fitness value. Based on the particle position parameters (i.e., the penalty factor α and the number of decomposition layers k for each individual particle), the vibration signal is subjected to variational mode decomposition to obtain modal components, and the average value of the envelope entropy of each modal component is calculated as the fitness value. Since the envelope entropy reflects the sparsity of the signal, the smaller the entropy value, the less noise interference in the signal and the more obvious the characteristics. Therefore, the optimization objective of the particle swarm is to minimize the fitness value.

[0041] Step S203: Update individual and global optimal positions. Based on the fitness values ​​obtained in the previous steps, update the individual optimal position and the global optimal position of the particle.

[0042] Step S204: Iterate until the stopping condition is met. The particle position and velocity are iteratively updated until the fitness value meets a preset iteration stopping condition, determining the globally optimal penalty factor and decomposition level. That is, the final globally optimal position is the optimal penalty factor α and decomposition level k. In some embodiments, the iteration stopping condition is that the fitness value has not decreased in 50 consecutive iterations.

[0043] Step S205: Signal reconstruction and noise reduction. The vibration signal is subjected to variational mode decomposition using the globally optimal penalty factor and decomposition level determined in the previous steps. High-frequency mode components in the decomposition result are removed, and the remaining mode components (i.e., effective mode components) are reconstructed to obtain a time-domain vibration signal with high-frequency noise removed.

[0044] It should be noted that although step S200 is described as a single step in the flowchart, in some embodiments of this application, this step is executed in two phases: The first phase (offline preparation phase) involves executing the iterative optimization of hyperparameters using the particle swarm optimization algorithm (i.e., steps S201 to S204), determining and storing the globally optimal parameter combination, and then executing step S205 based on the globally optimal parameter combination to perform variational mode decomposition and reconstruction of the vibration signal during the stable processing phase, achieving noise reduction. The second phase (online prediction phase) involves calling the stored globally optimal parameters and performing the signal decomposition and reconstruction described in step S205 on the real-time acquired vibration signal. By combining offline optimization and online calling, both the accuracy of parameter selection and the real-time requirements of online prediction are ensured.

[0045] Step S300: Perform spectrum conversion on the denoised time-domain vibration signal to obtain the frequency-domain vibration signal, and normalize the time-domain vibration signal and the frequency-domain vibration signal to construct the input data for the prediction model.

[0046] In this step, after obtaining the denoised time-domain vibration signal using the aforementioned steps, it is formatted to adapt to the input requirements of the deep learning model (described below). It should be noted that the data preprocessing method differs slightly depending on the current stage (offline preparation or online prediction). In the offline preparation stage, to construct a training dataset containing a large number of samples, the denoised long-term time-domain vibration signal obtained in the aforementioned steps (e.g., within 120 seconds of the stabilization phase) needs to be sliced ​​without overlap to generate multiple sample segments of fixed length. In the online prediction stage, the system directly acquires a fixed-length signal with the same length as the aforementioned sample segments. Therefore, slicing is not required in the online stage; instead, the currently acquired fixed-length signal segments are directly processed. Subsequently, the same spectral transformation and normalization operations are performed on both the sample segments obtained offline and the fixed-length signal segments acquired online. Specifically, the signal segments undergo spectral transformation (e.g., using Fast Fourier Transform, FFT) to obtain the corresponding frequency-domain vibration signal. Furthermore, the data needs to be normalized to eliminate the influence of different units on model training and prediction, and to accelerate model convergence. In some embodiments, the Min-Max normalization method is used to map the time-domain vibration signal and the corresponding frequency-domain vibration signal to a preset interval (such as the [0,1] interval).

[0047] It is worth noting that during the offline preparation phase, the training dataset is normalized; while during the online prediction phase, the normalized real-time collected time-domain vibration signals and frequency-domain vibration signals are combined to construct the input data for the prediction model.

[0048] Before describing the following online prediction process (i.e., subsequent steps S400 to S500), it should be noted that the method described in this application also relies on a pre-completed model building and training process (i.e., the offline preparation phase). In this phase, a training dataset including labels for vibration signals and measured surface integrity values ​​is constructed. This training dataset is then input into the time-frequency parallel deep learning model described below for training until the model converges, thereby obtaining the pre-trained time-frequency parallel deep learning model required for step S400.

[0049] In this embodiment, during the offline preparation phase, the time-frequency parallel deep learning model is trained using the training dataset constructed in the aforementioned steps. The Adam optimizer can be used to update the network weights, and the mean squared error (MSE) is used as the loss function to calculate the prediction error between the predicted value and the label until the time-frequency parallel deep learning model converges.

[0050] Step S400: Input the prediction model input data into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model (hereinafter referred to as the deep learning model) includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches.

[0051] In this step, a pre-trained deep learning model is used to perform adaptive feature extraction and regression analysis. It should be noted that both the vibration signal and the surface integrity after processing are strongly correlated with the waterjet strengthening process (e.g., jet pressure, abrasive flow rate), and both exhibit similar changing patterns with process variations. Therefore, this application uses the vibration signal as a reference to characterize surface integrity (e.g., surface roughness, residual stress). Taking residual stress as an example, the impact intensity reflected by the vibration signal is intrinsically related to the plastic deformation that causes residual stress in the material. Therefore, the deep learning model can predict residual stress by adaptively extracting feature information from the vibration signal. Furthermore, this application employs a time-frequency parallel architecture because the flow field environment of the waterjet strengthening process is extremely complex, accompanied by intense fluid turbulence and a large amount of abrasive scattering, resulting in a large amount of noise mixed in the vibration signal, and exhibiting time-varying characteristics. In this case, features extracted solely from the time-domain signal are insufficient and difficult to establish an effective correlation with surface integrity. Physically speaking, the time-domain signal reflects the transient impact and energy fluctuations during processing, while the frequency-domain signal reflects the energy distribution and periodic components of the processing. Therefore, this application utilizes the complementarity of the two signals to simultaneously extract multi-domain features of the vibration signal in the time and frequency domains, so as to ensure sufficient and comprehensive information for prediction modeling.

[0052] In some embodiments, the time-frequency parallel deep learning model includes: a temporal feature extraction branch, which includes multiple convolutional layers (CNN) and long short-term memory (LSTM) layers connected in sequence, utilizing the convolutional layers to extract local features of the temporal vibration signal and utilizing the LSTM layers to extract sequence dependency features; and a frequency domain feature extraction branch, which includes multiple convolutional layers and fully connected layers connected in sequence, used to extract the energy distribution features of the frequency domain vibration signal. The convolutional layers in both the temporal and frequency domain feature extraction branches employ a configuration with progressively decreasing kernel sizes to balance the breadth and depth of feature extraction. Specifically, the initial convolutional layers in feature extraction use larger kernels to extract long-distance dependency features in the signal and quickly extract spatial features; subsequent convolutional layers use smaller kernels to extract deeper, more refined features, thereby improving the model's expressive power and accuracy. Furthermore, each convolutional layer's output is connected to a pooling layer for downsampling the features output by the convolutional layer.

[0053] Furthermore, the time-frequency parallel deep learning model also includes a multi-head attention mechanism module. This module concatenates the feature vectors output from the time-domain feature extraction branch and the frequency-domain feature extraction branch, and then performs weighted feature fusion to achieve effective integration of these two feature paths. The multi-head attention mechanism represents the importance of feature information through weighting, enabling the deep learning model to automatically filter out redundant information and adaptively extract effective information from the feature vectors from multiple dimensions. This achieves efficient fusion of multi-domain signal features and enhances the robustness and generalization ability of the deep learning model.

[0054] Step S500: Output the predicted surface integrity value of the thin-walled component at the current processing position through the time-frequency parallel deep learning model.

[0055] In the actual online prediction process for enhanced water jet propagation, the system continuously acquires real-time vibration signals from the surface of thin-walled components. It should be noted that during the online prediction phase, the globally optimal penalty factor and decomposition level determined in the offline preparation phase are used to perform variational mode decomposition and reconstruction on the acquired vibration signals. That is, there is no need to perform time-consuming particle swarm optimization calculations again, thus ensuring the real-time performance of the online processing. Subsequently, in order to construct feature vectors that meet the input requirements of the deep learning model, data preprocessing (including spectrum conversion and normalization) is performed on the aforementioned denoised reconstructed signal. The processed time-domain and frequency-domain data are then input into the trained deep learning model to obtain the current surface integrity prediction value.

[0056] The method described in the above embodiments will be further illustrated below with examples.

[0057] This application will use the example of integrating a thin-walled component waterjet-enhanced surface integrity prediction method into an electronic device (such as an industrial control computer or a high-performance server) to illustrate the method of this application embodiment.

[0058] See Figure 3 As shown, the method may include steps S1000 to S5000.

[0059] It should be noted that, in order to address the problem that existing technologies cannot meet the requirements for high-precision online prediction, this application may include an offline preparation stage and an online prediction stage.

[0060] Step S1000: The electronic device acquires the vibration signal of the surface of the thin-walled component during the water jet strengthening process.

[0061] This process includes an offline preparation phase and an online prediction phase.

[0062] In the offline preparation phase, this step is used to build the training dataset. In this embodiment, a waterjet enhancement processing platform is constructed. During the offline preparation phase, an accelerometer (such as a piezoelectric accelerometer) is bolted to the surface of the thin-walled component to be enhanced, ensuring close contact between the sensor and the thin-walled component to guarantee the stability of high-frequency vibration signal transmission.

[0063] To verify the model's generalization ability under different process conditions, this embodiment selected multiple combinations of different waterjet enhancement process parameters. Specifically, it included nine sets of process parameters (WJP-1 to WJP-9), encompassing different jet pressures (50 MPa to 150 MPa) and abrasive flow rates (200 g / min to 1000 g / min). The specific combinations of process parameters for each set are shown in Table 1 below.

[0064] Based on the process parameters set in Table 1 above, the electronic equipment acquires the corresponding vibration signals.

[0065] Furthermore, the electronic device is set to a sampling frequency of 48kHz. Under preset waterjet enhancement process parameters, vibration signals during the stable processing stage of waterjet enhancement (e.g., within 120 seconds of the stable processing stage) are preferably captured as samples. The electronic device then uses the measured surface integrity values ​​corresponding to the acquired vibration signal segments as labels to construct a training dataset. In this embodiment, surface roughness is used as the prediction target. The labels of the training dataset are measured using an optical three-dimensional surface roughness measuring instrument, and the average surface roughness under the corresponding processing technology is calculated.

[0066] In the online prediction phase, this step is used to acquire real-time input data for the model. Specifically, using the same sensor arrangement and acquisition parameters as in the offline preparation phase, vibration signals at the current processing location are acquired in real time. These real-time acquired vibration signals will be used as the processing object in subsequent steps to predict the current surface integrity status in real time.

[0067] Step S2000: The electronic device uses variational mode decomposition to perform noise reduction preprocessing on the vibration signal, including: using particle swarm optimization algorithm to iteratively optimize the hyperparameters of the variational mode decomposition, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal.

[0068] For the fluid turbulence noise contained in the acquired raw signal (i.e., vibration signal), the electronic device processes it in this step using variational mode decomposition (VMD) optimized by particle swarm optimization (PSO). In this embodiment, the step of iteratively optimizing the hyperparameters of variational mode decomposition using particle swarm optimization is performed in the offline preparation stage before online prediction.

[0069] The specific optimization process may include: initializing particle swarm parameters, including particle number, position, and velocity, wherein the particle number is set to 20, and selecting the penalty factor α and the decomposition level k of variational mode decomposition as particle position parameters; performing variational mode decomposition on the vibration signal according to the particle position parameters to obtain modal components, and calculating the average value of the envelope entropy of each modal component as the fitness value; iteratively updating the particle position and velocity until the fitness value meets a preset iteration stopping condition (e.g., the fitness value has not decreased in 50 consecutive iterations), and determining the globally optimal penalty factor and decomposition level. Further, the electronic device stores the globally optimal penalty factor and decomposition level as optimization results for direct retrieval in subsequent online prediction stages.

[0070] Subsequently, the electronic device decomposes the vibration signal using the globally optimal penalty factor and the number of decomposition layers, removes the modal components containing high-frequency noise (i.e., high-frequency modal components), and reconstructs the remaining modal components (i.e., effective modal components) to obtain the noise-reduced time-domain vibration signal.

[0071] It should be noted that although step S2000 is described as a whole step in the flowchart, in some embodiments of this application, this step is executed in two phases in terms of timing: the first phase (offline preparation phase) executes the process of iteratively optimizing hyperparameters using the particle swarm optimization algorithm, determines and stores the globally optimal parameter combination, and performs variational mode decomposition and reconstruction on the vibration signal of the stable processing phase based on the globally optimal parameter combination to achieve noise reduction; the second phase (online prediction phase) calls the stored globally optimal parameters and performs decomposition and reconstruction on the real-time acquired vibration signal. By combining offline optimization and online calling, both the accuracy of parameter selection and the real-time requirements of online prediction are ensured.

[0072] Step S3000: The electronic device performs spectrum conversion on the noise-reduced time-domain vibration signal to obtain a frequency-domain vibration signal, and normalizes the time-domain vibration signal and the frequency-domain vibration signal to construct the prediction model input data.

[0073] After obtaining the denoised time-domain vibration signal using the aforementioned steps, the electronic device formats it to adapt to the input requirements of the deep learning model. It should be noted that the data preprocessing method differs slightly depending on the current stage (offline preparation or online prediction). In the offline preparation stage, to construct a training dataset containing a large number of samples, the denoised long-term time-domain vibration signal obtained in the aforementioned steps (e.g., within 120 seconds of the stabilization phase) needs to be sliced ​​without overlap to generate multiple sample segments of fixed length. In the online prediction stage, the system directly acquires a fixed-length signal with the same length as the aforementioned sample segments. Therefore, slicing is not required in the online stage; instead, the currently acquired fixed-length signal segments are directly processed. Subsequently, both the sample segments obtained from offline slicing and the fixed-length signal segments acquired online undergo the same spectral transformation and normalization operations. Specifically, the electronic device performs a Fast Fourier Transform (FFT) on the signal segments to obtain the corresponding frequency-domain vibration signal. Furthermore, in this embodiment, the electronic device uses the Min-Max normalization method to map the time-domain vibration signal and the corresponding frequency-domain vibration signal to a preset interval (such as the [0,1] interval).

[0074] It is worth noting that during the offline preparation phase, the training dataset is normalized; while during the online prediction phase, the electronic device combines the normalized real-time acquired time-domain vibration signal and frequency-domain vibration signal to construct the input data for the prediction model.

[0075] Before describing the following online prediction process (i.e., subsequent steps S4000 to S5000), it should be noted that the method described in this application also relies on a pre-completed model building and training process (i.e., the offline preparation stage). In this stage, a training dataset including labels for vibration signals and measured surface integrity values ​​is constructed. This training dataset is then input into the time-frequency parallel deep learning model described below for training until the model converges, thereby obtaining the pre-trained time-frequency parallel deep learning model required for step S4000.

[0076] In this embodiment, during the offline preparation phase, the electronic device uses the Adam optimizer to iteratively update the network weights, with the initial learning rate set to 0.001 and the mean squared error (MSE) used as the loss function, until the time-frequency parallel deep learning model converges.

[0077] Step S4000: The electronic device inputs the prediction model input data into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches.

[0078] In this embodiment, the electronic device utilizes a pre-trained deep learning model to perform adaptive feature extraction and regression analysis. In some embodiments, such as Figure 4 As shown, a time-frequency parallel deep learning model can include a time-domain feature extraction branch and a frequency-domain feature extraction branch. The time-domain feature extraction branch is used to receive and process the time-domain input (…). Figure 4 The identifier in the text is "Time Input," and the frequency domain feature extraction branch is used to receive and process the frequency domain input. Figure 4 The identifier in the code is FrequencyInput). The temporal feature extraction branch consists of four sequentially connected convolutional layers (in...). Figure 4 The middle is labeled Conv) and 1 long short-term memory network layer ( Figure 4 (Identified as LSTM). To effectively acquire temporal features at different scales, the convolutional layers are configured as follows: The first convolutional layer uses a kernel of size 64 with 8 channels to extract long-range dependency features in the signal and quickly extract spatial features. It is followed by a pooling layer (in... Figure 4 The pooling layer is labeled "pool," and its kernel size is 2, meaning the stride is set to 2 by default. The subsequent second, third, and fourth convolutional layers use kernels of 7, 5, and 3 respectively, with 16, 32, and 64 channels, used to extract deep, fine features. Each layer's output is connected to the same pooling layer (kernel size 2) for downsampling. Furthermore, a random deactivation layer is added at the output of the pooling layer after the fourth convolutional layer (in... Figure 4 The LSTM layer, marked as dropout, is used to prevent overfitting. It receives the output after random deactivation and has 128 hidden units to extract the temporal dependency features of the vibration signal, outputting a 128-dimensional temporal feature vector.

[0079] Accordingly, the frequency domain feature extraction branch contains four convolutional layers (Conv) and one fully connected layer (in) connected sequentially. Figure 4 The four convolutional layers in this branch are configured with kernel sizes of 32, 7, 5, and 3, and channel numbers of 8, 16, 32, and 64, respectively, to achieve multi-scale feature extraction of spectral energy distribution. Each convolutional layer is also followed by a pooling layer (in...). Figure 4The pooling layer is labeled "pool" with a kernel size of 2 (meaning the stride is set to 2 by default). A dropout layer is also provided at the output of the pooling layer after the fourth convolutional layer in this branch. Furthermore, a fully connected layer is placed at the end of this branch to map the randomly dropout-processed convolutional output into a 128-dimensional frequency domain feature vector.

[0080] exist Figure 4 In the illustrated embodiment, the time-frequency parallel deep learning model further includes a multi-head attention mechanism module (in... Figure 4 The above time-domain and frequency-domain feature vectors are fused using a Multi-Head Attention (MGA) mechanism. Specifically, the time-domain and frequency-domain feature vectors are concatenated to form a joint feature vector with dimension 256, which is then input into the MGA module. The MGA module adaptively calculates the attention weights for each feature dimension to enhance key features and suppress redundant information. The fused features output by the MGA module are first processed by a dropout layer before being input into a fully connected layer with one node for feature integration and regression analysis, and the predicted surface roughness value is output. Figure 4 The label is Surface RoughnessOutput.

[0081] Step S5000: The electronic device outputs the predicted surface integrity value of the thin-walled component at the current processing position through the time-frequency parallel deep learning model.

[0082] In this step, the electronic device outputs a surface integrity prediction value through a time-frequency parallel deep learning model. During the actual online prediction process for enhanced water jet propagation, the system continuously acquires real-time vibration signals from the surface of the thin-walled component. It should be noted that during the online prediction phase, the electronic device directly utilizes the globally optimal penalty factor and decomposition layer number determined in the offline preparation phase to perform variational mode decomposition and reconstruction on the real-time acquired vibration signals. That is, there is no need to perform time-consuming particle swarm optimization calculations again, thus ensuring the real-time performance of the online processing. Subsequently, in order to construct feature vectors that meet the input requirements of the deep learning model, the electronic device needs to perform data preprocessing (including spectrum conversion and normalization) on the aforementioned denoised reconstructed signal, and input the processed time-domain and frequency-domain data into the trained deep learning model to obtain the current surface integrity prediction value.

[0083] To verify the effectiveness of each module in this model, ablation experiments were also conducted in the embodiments of this application. The performance comparison results of the time-frequency parallel deep learning model constructed in this application with other comparative models will be described below. The specific comparative models and their architectures are described below: The comparison model 1 (TCNN) retains only the CNN architecture of the time-domain feature extraction branch in the aforementioned time-frequency parallel deep learning model to extract time-domain signal features, and performs regression prediction through fully connected layers.

[0084] Compared with the aforementioned time-frequency parallel deep learning model, Model 2 (MCNN) retains the CNN architecture of time-domain feature extraction branch and frequency-domain feature extraction branch, but uses fully connected layers instead of LSTM layers, and the time-domain and frequency-domain features are concatenated and directly passed through the fully connected layers for regression prediction.

[0085] Compared with the aforementioned time-frequency parallel deep learning models, Model 3 (MCNN-LSTM) removes the multi-head attention mechanism module (or attention pooling layer), retains the time-domain feature extraction branch and the frequency-domain feature extraction branch, and the time-domain and frequency-domain features are concatenated and directly passed through the fully connected layer for regression prediction.

[0086] Furthermore, the hyperparameters and training parameters of each comparative model are consistent with those of the aforementioned time-frequency parallel deep learning model. The dataset used in this experiment was constructed from vibration signals obtained under the WJP-1 to WJP-5 processes as shown in Table 1 above. In addition, to avoid the randomness of the model training results, each model was trained separately 5 times, and the mean absolute percentage error (MAPE) was used as the evaluation index for prediction performance. The MAPE statistics and average values ​​of each model in the 5 training iterations are shown in Table 2 below.

[0087]

[0088] As shown in Table 2 above, the time-frequency parallel deep learning model combined with the attention mechanism proposed in this application achieves the best prediction accuracy, with an average MAPE of only 0.99% and a standard deviation of 0.099 after 5 training iterations. Compared with other comparative models, this model not only improves prediction accuracy but also exhibits better stability, fully demonstrating the effectiveness of combining the time-frequency parallel architecture with the attention mechanism in predicting the surface integrity of waterjet-enhanced thin-walled components.

[0089] It should also be noted that the beneficial effects of the thin-walled component waterjet enhanced surface integrity prediction method provided in the above embodiments can be found in the relevant descriptions in the preceding embodiments, and will not be repeated here.

[0090] See Figure 5This application also provides a surface integrity prediction system 100 for waterjet-strengthened thin-walled components. This system integrates waterjet processing hardware and deep learning algorithm software to achieve high-precision online prediction of the surface integrity of thin-walled components such as aero-engine turbine blades during the waterjet strengthening process. The system 100 may include a signal acquisition module 110, a signal preprocessing module 120, a feature construction module 130, and a prediction analysis module 140.

[0091] In this embodiment, the signal acquisition module 110 can be based on Figure 6 The waterjet processing hardware platform 200 shown is used to acquire vibration signals on the surface of thin-walled components during waterjet strengthening processing. The signal acquisition module 110 is used not only to construct a training dataset containing labels of measured surface integrity values, but also to acquire vibration signals in real time during the online prediction phase.

[0092] like Figure 6 As shown, the waterjet processing hardware platform 200 may include: a robotic arm 210, a water tank 220, a jet process monitoring and control platform 230 (hereinafter referred to as the control platform 230), a turntable fixture 240, a blade clamp 250, an accelerometer 260, a jet tool 270, and a turbine blade 280 as the processing object. During operation, the turbine blade 280 is clamped onto the turntable fixture 240 by the blade clamp 250 and placed inside the water tank 220. The robotic arm 210 clamps the jet tool 270 and, under the command control of the control platform 230, performs waterjet reinforcement on the blade portion of the turbine blade 280 according to a preset processing path. The accelerometer 260 is installed at the root of the turbine blade 280 to sense vibrations caused by the waterjet impact. The signal acquisition module 110 transmits the signals acquired by the accelerometer 260 to the control platform 230 (i.e., the electronic device in this embodiment) via cable for processing and storage.

[0093] In some embodiments, the signal acquisition module 110 is configured to set a sampling frequency of 48 kHz under the control of the control platform 230. It should be noted that the signal acquisition module 110 employs different acquisition methods at different stages. In the offline preparation stage, the signal acquisition module 110 acquires long-term vibration signals during the stable processing phase under the set process (e.g., within 120 seconds of the stable processing phase), facilitating the subsequent construction of a large number of samples through slicing. In the online prediction stage, the signal acquisition module 110 acquires signal segments (i.e., fixed-length signals) in real time that match the input length of the prediction model to meet the time response requirements of real-time prediction.

[0094] In addition, the signal acquisition module 110 is also used to associate measured values ​​of surface integrity. During the offline preparation phase, the average surface roughness of the turbine blade 280 under the corresponding process is measured using an optical three-dimensional surface roughness measuring instrument, or residual stress is measured using an X-ray residual stress diffractometer. These measured values ​​are then used as labels and associated with the corresponding vibration signal segments to construct a training dataset. It should be noted that the labels are measured values ​​of surface integrity indices (such as average surface roughness or residual stress values). Since the waterjet enhancement process under a fixed process is basically stable, and the resulting surface quality is uniform, these measured values ​​can represent the expected surface integrity under the process and corresponding vibration state. Based on this, the deep learning model learns the mapping relationship between vibration state and surface integrity indices under different process parameters. During online prediction, it can identify the current processing state based on the real-time input vibration signal and output the predicted surface integrity value at the current time and location.

[0095] In this embodiment, the signal preprocessing module 120 is used to perform noise reduction preprocessing on the vibration signal using variational mode decomposition. Specifically, the signal preprocessing module 120 is configured to iteratively optimize the hyperparameters of the variational mode decomposition using a particle swarm optimization algorithm, and decompose and reconstruct the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal. Since the complex processing environment of water jetting involves severe fluid turbulence, resulting in a large amount of noise in the signal, the signal preprocessing module 120 integrates an adaptive optimization mechanism based on the particle swarm optimization algorithm (PSO). In specific implementation, the signal preprocessing module 120 is configured as follows: In the offline preparation stage before online prediction, it performs an iterative optimization of the hyperparameters (penalty factor and number of decomposition layers) of variational mode decomposition using the particle swarm optimization algorithm, aiming to minimize the average value of the envelope entropy of each modal component, determines the globally optimal parameter combination and stores it, and performs variational mode decomposition and reconstruction on the vibration signal of the stable processing stage according to the globally optimal parameter combination to achieve noise reduction processing; In the online prediction stage, it calls the globally optimal penalty factor and number of decomposition layers determined in the offline preparation stage to directly perform variational mode decomposition on the real-time acquired vibration signal, removes modal components containing high-frequency noise, and reconstructs the signal of the remaining effective modal components, thereby ensuring the noise reduction effect while meeting the real-time requirements of online processing.

[0096] In this embodiment, the feature construction module 130 performs spectral conversion on the denoised time-domain vibration signal to obtain a frequency-domain vibration signal, and normalizes the time-domain vibration signal and the frequency-domain vibration signal to construct the input data for the prediction model. In specific implementations, the data processing logic varies depending on the signal morphology at different stages.

[0097] During the offline preparation phase, since the acquired signal is a long-term vibration signal during a stable processing phase under a set process (e.g., within 120 seconds of the stable processing phase), the feature construction module 130 needs to perform non-overlapping time slices on the noise-reduced signal output by the signal preprocessing module 120 and use Fast Fourier Transform (FFT) to convert each time slice into a corresponding frequency domain signal. Subsequently, the feature construction module 130 uses the Min-Max normalization method to map the time-domain vibration signal and the corresponding frequency-domain vibration signal to a preset interval (e.g., the interval [0,1]). Next, the feature construction module 130 combines the normalized time-domain vibration signal and the frequency-domain vibration signal to construct a (time-frequency) joint feature vector, which serves as the input to the training dataset for the subsequent deep learning model.

[0098] During the online prediction phase, since the real-time acquired vibration signal has been configured as a signal segment matching the input length of the prediction model, the feature construction module 130 does not need to perform a slicing operation. Instead, it directly performs spectral conversion and normalization processing on the noise-reduced real-time acquired time-domain vibration signal to construct the input data for the prediction model.

[0099] In this embodiment, the prediction analysis module 140 is used to input the prediction model input data into a pre-trained time-frequency parallel deep learning model, and output the surface integrity prediction value of the thin-walled component at the current processing position through the time-frequency parallel deep learning model. The prediction analysis module 140 internally deploys a time-frequency parallel deep learning model incorporating an attention mechanism. Specifically, the time-frequency parallel deep learning model includes parallel temporal feature extraction branches (including 4 convolutional layers and 1 LSTM layer) and frequency domain feature extraction branches (including 4 convolutional layers and 1 fully connected layer), as well as a multi-head attention mechanism module for fusing the temporal features extracted by the temporal feature extraction branch and the frequency features extracted by the frequency domain feature extraction branch. In the online prediction phase, the prediction analysis module 140 receives data transmitted by the feature construction module 130, performs adaptive feature extraction and regression analysis through the aforementioned deep learning network, and outputs the surface roughness prediction value under the current processing state in real time.

[0100] It should be noted that the time-frequency parallel deep learning model called by the prediction and analysis module 140 is pre-trained during the offline preparation phase. Specifically, during the offline preparation phase, the training dataset is input into the time-frequency parallel deep learning model for training. The Adam optimizer is used to iteratively update the network weights, and the mean squared error (MSE) is used as the loss function until the time-frequency parallel deep learning model converges.

[0101] This application also provides an electronic device, such as... Figure 7 As shown, it illustrates a structural schematic diagram of the electronic device 500 involved in an embodiment of this application. Specifically: The electronic device 500 may include components such as a processor 501 with one or more processing cores, a memory 502 with one or more computer-readable storage media, a power supply 503, an input unit 504, and a communication unit 505. Those skilled in the art will understand that... Figure 7 The electronic device 500 shown does not constitute a structural limitation on the electronic device 500, and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: The processor 501 is the control center of the electronic device 500. It connects various parts of the electronic device 500 via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 502, and by calling data stored in the memory 502, it performs various functions and processes data of the electronic device 500, thereby providing overall monitoring of the electronic device 500. Optionally, the processor 501 may include one or more processing cores; preferably, the processor 501 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 501.

[0102] The memory 502 can be used to store software programs and modules. The processor 501 executes various functional applications and data processing by running the software programs and modules stored in the memory 502. The memory 502 mainly includes a program storage area and a data storage area. The program storage area can store the operating system, computer programs required for at least one function, etc.; the data storage area can store data created based on the use of the electronic device 500, etc. In addition, the memory 502 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 502 may also include a memory controller to provide the processor 501 with access to the memory 502.

[0103] The electronic device 500 also includes a power supply 503 that supplies power to the various components. Preferably, the power supply 503 can be logically connected to the processor 501 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 503 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0104] The electronic device 500 may also include an input unit 504, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0105] The electronic device 500 may also include a communication unit 505. In some embodiments, the communication unit 505 may include a wireless module, and the electronic device 500 may perform short-range wireless transmission through the wireless module of the communication unit 505.

[0106] Although not shown, the electronic device 500 may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 501 in the electronic device 500 loads the executable files corresponding to the processes of one or more computer programs into the memory 502 according to the following instructions, and the processor 501 runs the computer programs stored in the memory 502 to realize various functions.

[0107] It should be noted that the instructions executed by processor 501 mainly correspond to the online prediction process, and call the optimized parameters and trained model determined in the offline preparation phase. Specifically, this includes: The vibration signal of the thin-walled component surface during waterjet enhanced processing is acquired; the vibration signal is preprocessed for noise reduction using variational mode decomposition, including: iteratively optimizing the hyperparameters of the variational mode decomposition using particle swarm optimization algorithm, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal; the noise-reduced time-domain vibration signal is subjected to spectral transformation to obtain the frequency-domain vibration signal, and the time-domain vibration signal and the frequency-domain vibration signal are normalized to construct the prediction model input data; the prediction model input data is input into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches; the time-frequency parallel deep learning model outputs the predicted surface integrity value of the thin-walled component at the current processing position.

[0108] It should be noted that the steps described above, namely "iteratively optimizing the hyperparameters of the variational mode decomposition using the particle swarm optimization algorithm" and "pre-training the time-frequency parallel deep learning model," are completed during the offline preparation phase. Specifically, for the "pre-trained time-frequency parallel deep learning model," during the offline preparation phase, the Adam optimizer is used to iteratively update the network weights, and the mean squared error is used as the loss function until the time-frequency parallel deep learning model converges.

[0109] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0110] As can be seen from the above, the electronic device 500 in this embodiment achieves online prediction of surface integrity by acquiring the vibration signal of the thin-walled component surface during water jet strengthening processing, performing noise reduction preprocessing on the signal using variational mode decomposition optimized by particle swarm optimization algorithm, constructing prediction model input data, and analyzing and processing it using a pre-trained time-frequency parallel deep learning model. Based on this, by combining variational mode decomposition with deep learning, and utilizing the time-domain and frequency-domain feature extraction branches in the time-frequency parallel deep learning model to extract features in parallel, and performing feature weighted fusion through a multi-head attention mechanism module, the model can adaptively extract effective information related to surface integrity from the vibration signal from multiple dimensions, establishing a nonlinear mapping relationship between the vibration signal and surface integrity indicators (such as surface roughness, residual stress, microhardness, etc.). Therefore, the electronic device 500 in this embodiment solves the problems of difficult monitoring of the state of water jet enhancement process and complex process of characterizing enhancement effect in the prior art, and overcomes the defects of traditional monitoring methods that rely on feature engineering, which have limited prediction accuracy and generalization ability under complex working conditions. In particular, in the complex environment of strong fluid turbulence and noise interference, the electronic device 500 in this embodiment has strong robustness and generalization ability, and can realize high-efficiency and high-precision online prediction of surface integrity. The prediction results can provide a basis for online optimization of the enhancement process.

[0111] Those skilled in the art will understand that all or part of the steps in the methods described in the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0112] Therefore, embodiments of this application provide a computer-readable storage medium storing computer instructions that can be loaded by a processor to execute steps in any of the thin-walled component waterjet-enhanced surface integrity prediction methods provided in this application. It should be noted that, as described above, this method includes a coordinated offline preparation phase and an online prediction phase. In a preferred embodiment, the computer instructions are configured to execute the online prediction process and invoke the optimized parameters and trained model determined in the offline preparation phase. For example, the instructions can execute the following steps: The vibration signal of the thin-walled component surface during waterjet enhanced processing is acquired; the vibration signal is preprocessed for noise reduction using variational mode decomposition, including: iteratively optimizing the hyperparameters of the variational mode decomposition using particle swarm optimization algorithm, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal; the noise-reduced time-domain vibration signal is subjected to spectral transformation to obtain the frequency-domain vibration signal, and the time-domain vibration signal and the frequency-domain vibration signal are normalized to construct the prediction model input data; the prediction model input data is input into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches; the time-frequency parallel deep learning model outputs the predicted surface integrity value of the thin-walled component at the current processing position.

[0113] It should also be noted that the steps described above, namely "iteratively optimizing the hyperparameters of the variational mode decomposition using the particle swarm optimization algorithm" and "pre-training the time-frequency parallel deep learning model," are completed during the offline preparation phase. Specifically, for the "pre-trained time-frequency parallel deep learning model," during the offline preparation phase, the Adam optimizer is used to iteratively update the network weights, and the mean squared error is used as the loss function until the time-frequency parallel deep learning model converges.

[0114] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0115] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0116] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the thin-walled component waterjet-enhanced surface integrity prediction methods provided in the embodiments of this application, the beneficial effects that any of the thin-walled component waterjet-enhanced surface integrity prediction methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0117] This application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the thin-walled component waterjet enhanced surface integrity prediction method in the above embodiments of this application.

[0118] It should also be noted that the beneficial effects achievable by the electronic device, computer-readable storage medium, and computer program product provided in the above embodiment can be found in the relevant descriptions in the preceding embodiments. Further details will not be repeated here.

[0119] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods. The above descriptions are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the surface integrity of waterjet-enhanced thin-walled components, characterized in that, include: Acquire vibration signals on the surface of thin-walled components during waterjet enhanced processing; The vibration signal is preprocessed for noise reduction using variational mode decomposition, including: iteratively optimizing the hyperparameters of the variational mode decomposition using particle swarm optimization algorithm, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal; The denoised time-domain vibration signal is subjected to spectrum conversion to obtain a frequency-domain vibration signal. The time-domain vibration signal and the frequency-domain vibration signal are then normalized to construct the input data for the prediction model. The prediction model input data is input into a pre-trained time-frequency parallel deep learning model, wherein the time-frequency parallel deep learning model includes a parallel time-domain feature extraction branch and a frequency-domain feature extraction branch, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branch and the frequency-domain features extracted by the frequency-domain feature extraction branch; The time-frequency parallel deep learning model outputs the predicted surface integrity value of the thin-walled component at the current processing position.

2. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 1, characterized in that, The iterative optimization of the hyperparameters of the variational mode decomposition using the particle swarm optimization algorithm includes: The penalty factor and the number of decomposition layers of the variational mode decomposition are selected as the particle position parameters of the particle swarm. The vibration signal is subjected to variational mode decomposition based on the particle position parameters to obtain modal components, and the average value of the envelope entropy of each modal component is calculated as the fitness value. By iteratively updating the particle position and velocity until the fitness value meets the preset iteration stopping condition, the globally optimal penalty factor and decomposition layer number are determined. The step of decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters includes: decomposing the vibration signal using the globally optimal penalty factor and the number of decomposition layers, removing the high-frequency modal components, and reconstructing the remaining modal components.

3. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 1, characterized in that, The time-frequency parallel deep learning model includes: a time-domain feature extraction branch comprising multiple convolutional layers and long short-term memory network layers connected in sequence, used to extract local features and sequence dependency features of the time-domain vibration signal; a frequency-domain feature extraction branch comprising multiple convolutional layers and fully connected layers connected in sequence, used to extract energy distribution features of the frequency-domain vibration signal; and a multi-head attention mechanism module used to concatenate the feature vectors output by the time-domain feature extraction branch and the frequency-domain feature extraction branch and then perform feature weighted fusion.

4. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 3, characterized in that, The convolutional layers in both the temporal feature extraction branch and the frequency domain feature extraction branch are configured with convolutional kernel sizes decreasing sequentially. Specifically, the convolutional layers in the initial stages of feature extraction use larger kernel sizes to extract long-distance dependent features, while the convolutional layers in subsequent levels use smaller kernel sizes to extract deep and refined features.

5. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 1, characterized in that, The training process of the time-frequency parallel deep learning model includes: Vibration signal segments during the waterjet-enhanced stabilization process were acquired, and the measured surface integrity values ​​corresponding to the vibration signal segments were used as labels to construct a training dataset. The time-domain vibration signal and the corresponding frequency-domain vibration signal are mapped to a preset interval using the Min-Max normalization method, and the training dataset is normalized. The time-frequency parallel deep learning model is trained using the training dataset, the network weights are updated using the Adam optimizer, and the mean squared error is used as the loss function to calculate the prediction error between the predicted value and the label until the time-frequency parallel deep learning model converges.

6. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 5, characterized in that, The measured surface integrity value includes at least surface roughness; the labels of the training dataset are the average surface roughness values ​​under the corresponding processing technology obtained by an optical three-dimensional surface roughness measuring instrument; the predicted surface integrity value is the predicted surface roughness value under the current processing technology state.

7. The method for predicting the surface integrity of thin-walled parts enhanced by water jet as described in claim 1, characterized in that, The step of iteratively optimizing the hyperparameters of the variational mode decomposition using the particle swarm optimization algorithm is performed in the offline preparation stage before online prediction; in the online prediction stage, the globally optimal penalty factor and decomposition level determined in the offline preparation stage are used to perform variational mode decomposition and reconstruction on the vibration signal acquired in real time.

8. A system for predicting the surface integrity of thin-walled components enhanced by water jetting, characterized in that, include: The signal acquisition module is used to acquire vibration signals on the surface of thin-walled components during waterjet enhanced processing. The signal preprocessing module is used to perform noise reduction preprocessing on the vibration signal using variational mode decomposition, including: using particle swarm optimization algorithm to iteratively optimize the hyperparameters of the variational mode decomposition, and decomposing and reconstructing the vibration signal according to the determined optimal hyperparameters to obtain the noise-reduced time-domain vibration signal. The feature construction module is used to perform spectrum conversion on the denoised time-domain vibration signal to obtain the frequency-domain vibration signal, and to normalize the time-domain vibration signal and the frequency-domain vibration signal to construct the input data for the prediction model. The predictive analysis module is used to input the prediction model input data into a pre-trained time-frequency parallel deep learning model, and output the surface integrity prediction value of the thin-walled component at the current processing position through the time-frequency parallel deep learning model; wherein, the time-frequency parallel deep learning model includes parallel time-domain feature extraction branches and frequency-domain feature extraction branches, as well as a multi-head attention mechanism module for fusing the time-domain features extracted by the time-domain feature extraction branches and the frequency-domain features extracted by the frequency-domain feature extraction branches.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the thin-walled component waterjet enhanced surface integrity prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device stores computer instructions that, when executed, implement the thin-walled component waterjet enhanced surface integrity prediction method as described in any one of claims 1 to 7.

Images