A typical mechanical and electrical system performance prediction method based on small sample data continuation

Abstract

The present application relates to a kind of typical mechanical and electrical system performance prediction method based on small sample data extension, applied in the field of intelligent mechanical and electrical manufacturing technology, comprising: S1, based on digital twinning mechanical and electrical system real-time model building;S2, based on time series data enhancement method TimeGAN mechanical and electrical system small sample data extension;S3, mechanical and electrical system performance prediction and real-time monitoring technology.The present application is aimed at the reduction of model generalization ability and prediction accuracy caused by small sample data in mechanical and electrical system, difficult to detect and diagnose and other problems, proposes a kind of small sample data extension method TimeGAN, and on this basis, mechanical and electrical system is expanded to performance prediction;While the real-time characteristics of digital twinning are ingeniously combined with mechanical and electrical system, further strengthen the centralized management and predictive maintenance of mechanical and electrical system, provide strong support for its life prediction, fault diagnosis, performance optimization and other fields of research.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, and specifically relates to a method for predicting the performance of typical electromechanical systems based on small sample data extension. Background Technology

[0002] With the continuous development of industrial production and the widespread application of machinery and equipment, especially driven by internet-based intelligent manufacturing, electromechanical systems are widely used due to their advantages such as high production efficiency, long service life, and low maintenance costs. However, as electromechanical systems operate and are used for extended periods, the performance of their internal components gradually degrades, leading to system malfunctions. To address these issues, performance prediction for electromechanical systems is receiving increasing attention.

[0003] In recent years, a significant amount of research has been devoted to performance prediction of electromechanical systems. Machine learning and data mining techniques have gradually become the main tools in this research. For example, statistical model-based methods, including Logistic Regression, Decision Trees, and Naive Bayes, are applied to performance prediction by establishing regression or classification models. Furthermore, deep learning methods based on neural networks and support vector machines, through multi-level nonlinear mapping and information extraction, model and predict the complexity and randomness inherent in electromechanical systems, improving the accuracy and practicality of model predictions.

[0004] However, in the performance prediction of electromechanical systems, traditional prediction methods cannot effectively study the performance of electromechanical systems without sufficient monitoring information, especially when historical data records are limited and of low quality. To address these issues of small sample sizes and imbalanced data, deep learning technology, due to its powerful model fitting ability, strong adaptability, and high automation, is increasingly being applied to the field of electromechanical system performance prediction. The successful application of deep learning algorithms such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) on time series data is noteworthy. Furthermore, models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are also being used in the performance prediction of electromechanical systems. Traditional electromechanical system modeling and simulation are offline simulations, unable to feed back the real-time operating conditions of the real-world object to the model. Digital twin technology aims to create a virtual twin that closely resembles the real-world object. Unlike traditional offline simulations, digital twin technology drives the twin's updates through real-time sensor data, thereby showcasing the real-time state of the real-world object. Meanwhile, digital twin technology also enables applications such as predictive maintenance and intelligent operation. By combining sensor data and machine learning algorithms, it continuously learns and optimizes the operating status of electromechanical systems, achieving real-time monitoring, early warning, and fault diagnosis. This real-time capability not only improves the reliability and availability of the system but also meets the future development needs of industrial intelligence. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a data extension method for small sample sizes to alleviate problems such as reduced generalization ability and prediction accuracy of electromechanical system models, as well as difficulties in detection and diagnosis caused by small sample sizes. Simultaneously, leveraging the real-time characteristics of digital twins, the invention achieves real-time updates of the electromechanical system's operating conditions and related models by inputting real-time sensor data. Furthermore, it integrates other relevant information such as electromechanical system model parameters, operating status, and automatic control into the user interface, further enhancing the processing and analysis of the electromechanical system.

[0006] The present invention proposes a method for predicting the performance of typical electromechanical systems based on small sample data extension, specifically as follows:

[0007] S1: Real-time model building of electromechanical systems based on digital twins;

[0008] The digital twin electromechanical system real-time model comprises two parts: a digital twin real-time geometric model of the electromechanical system and a digital twin real-time data model.

[0009] A real-time digital twin geometric model of the electromechanical system is built. The model's external dimensions and the operating status of its movable parts must be consistent with the physical prototype. The virtual geometric model needs to provide real-time feedback on various physical information and operating status of the real physical prototype to achieve synchronization between the virtual and physical prototypes.

[0010] A real-time data model of a digital twin of an electromechanical system is built. Sensor data is collected in real time through an acquisition system and uploaded to a host computer for storage and backup for subsequent research. At the same time, the real-time data is fed into a data extension model and a machine learning prediction model to realize the real-time updating and output of the model.

[0011] S2: Small sample data extension of electromechanical systems based on the time series data augmentation method TimeGAN;

[0012] To address the problem of reduced prediction accuracy and difficulty in detection and diagnosis of machine learning-related models due to small data samples in electromechanical systems, this invention proposes a time-series data augmentation method called TimeGAN. TimeGAN extends time-series sensor data from electromechanical systems. The core idea of ​​TimeGAN is similar to GAN (Generative Adversarial Network), which trains a generator and discriminator against each other until a Nash equilibrium is reached. The generator can then extend new data with the same underlying patterns as the original data. However, TimeGAN adds the concept of an autoencoder to GAN to better handle the characteristics of time-series data.

[0013] An autoencoder consists of an encoder and a decoder, with the core components being an embedding function *e* and a recovery function *r*, respectively. The embedding function *e* maps the static and temporal features of the original data to a low-dimensional latent space, while the recovery function *r* maps the learned low-dimensional features back to the original data feature space. This use of low-dimensional representation can improve the speed and efficiency of time series generation and provide better results. The embedding function *e* and the recovery function *r* are defined as follows:

[0014]

[0015] Where S and X represent the static and temporal characteristics of the data, respectively, and H S H X ...

[0016] h S ,h 1:T =e(S,X) 1:T (3)

[0017] h S =e S (s) (4)

[0018] h t =e X (h S ,h t-1 ,x t (5)

[0019] In the formula, h S It is the low-dimensional latent vector corresponding to the static feature; h 1:T h is the low-dimensional latent vector corresponding to the temporal features from the beginning to time T. t h t-1 These represent the low-dimensional latent vectors corresponding to the temporal features at time t and time t-1, respectively; e S With e X These are the embedding functions that transform static features and temporal features into low-dimensional latent vectors, respectively.

[0020] The recovery function r restores the low-dimensional latent vector into static and temporal features. This recovery function r is implemented using a feedforward neural network.

[0021] S,X 1:T =r(h S ,h 1:T (6)

[0022] S = r S (h S(7)

[0023] X t =r X (h t (8)

[0024] In the formula, S and X 1:T These represent the latent vectors h of static features, respectively, obtained by the recovery function r. S With temporal feature latent vector h 1:T Recovering static and temporal features, r S and r X These are the recovery networks corresponding to static features and temporal features, respectively, X. t Let be the temporal characteristics that are restored by the recovery function at time t.

[0025] S3: Performance prediction and real-time monitoring of electromechanical systems;

[0026] The electromechanical system performance prediction involves using a machine learning model to predict the performance indicators of the electromechanical system. The machine learning model is combined with the small sample data extension method TimeGAN in S2, which aims to alleviate the negative impact of the small sample problem on model training and further improve the accuracy of model prediction.

[0027] Specifically, the initial data and TimeGAN extended data are synchronously input into the machine learning model of the electromechanical system, and the quality of the trained model is evaluated based on the root mean square error (RMSE) and mean square error (MSE). Then, the performance operation status of the electromechanical system is judged by the predicted values ​​of the electromechanical system performance indicators.

[0028] The real-time monitoring of the electromechanical system mainly involves the design of a digital twin user interface. The aim is to allow users to grasp the overall operating status of the electromechanical system through the real-time system-related information displayed on the interface. Specifically, the real-time geometric model of the electromechanical system in S1 is displayed on the main interface, with virtual and real synchronization. The operating status of the physical prototype can be observed in real time through the virtual geometric model. For the real-time data model, storage functions for model data, predicted data, and sensor data are added for subsequent research. The performance index values ​​predicted by the machine learning model are displayed in real-time on the data interface, and various model operating parameters are added, allowing users to make adaptive changes based on the operating status and model evaluation indicators. Sensor data is displayed in real-time on the data interface, and curves are plotted to more intuitively reflect the system status.

[0029] The beneficial effects of this invention are as follows:

[0030] (1) To address the problem of small data samples in electromechanical systems, a data extension method, TimeGAN, is proposed. This method is combined with the performance prediction of electromechanical systems to provide high-quality data support and significantly reduce the impact of data quality and quantity issues on the research.

[0031] (2) The real-time characteristics of digital twins are combined with the performance prediction of electromechanical systems. The machine learning model is updated by real-time sensor data, and the predicted values ​​of electromechanical system performance indicators are output in real time to construct a curve that reflects the trend of system operation status. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method for predicting the performance of typical electromechanical systems based on small sample data, as described in this invention.

[0033] Figure 2 This is a 3D diagram of an electro-hydraulic actuator, a typical example of an electromechanical system.

[0034] Figure 3 This is a schematic diagram of the EHA system's real-time data acquisition system.

[0035] Figure 4a This is a comparison chart of the original EHA motor speed data and the TimeGAN extended data.

[0036] Figure 4b This is a comparison chart of the raw data of the pressure signal in the left chamber of the EHA hydraulic cylinder and the TimeGAN extended data.

[0037] Figure 4c This is a comparison chart of the raw data of the pressure signal in the right chamber of the EHA hydraulic cylinder and the TimeGAN extended data.

[0038] Figure 4d This is a comparison chart of the raw EHA motor temperature data and the TimeGAN extended data.

[0039] Figure 5a This is the PCA evaluation diagram for the TimeGAN extension method.

[0040] Figure 5b This is the t-SNE evaluation plot of the TimeGAN extension method.

[0041] Figure 6 This is a flowchart of the EHA system performance prediction method.

[0042] Figure 7 This is a comparison chart of RMSE of different machine learning models based on TimeGAN extended data.

[0043] Figure 8a This is a graph showing the predicted performance of EHA metrics with only the original input data.

[0044] Figure 8b This is a graph showing the predicted performance of EHA metrics using the original data plus TimeGAN extended data.

[0045] Figure 9 This is the main interface of the EHA real-time monitoring system.

[0046] Figure 10 This is the display interface of the EHA real-time monitoring system data module. Detailed Implementation

[0047] The following is in conjunction with the appendix Figure 1-10 The present invention provides a detailed description of the performance prediction method for typical electromechanical systems based on small sample data extension, using examples of typical electro-hydraulic actuators (EHAs) in electromechanical systems.

[0048] This invention addresses the issue of reduced performance prediction accuracy in EHA systems due to the influence of small sample sizes. It proposes a data extension method, TimeGAN, to expand the small sample data in the EHA system. Based on this, a machine learning CNN-BiLSTM-Attention model is used to predict the performance indicators of the EHA system to reflect changes in performance status. At the same time, a digital twin real-time system is built to provide the model with real-time sensor data input, and the predicted performance indicators output by the model in real time reflect the trend of the system's operating status.

[0049] This invention mainly consists of three steps, such as... Figure 1 As shown, it includes:

[0050] S1: Real-time model building of EHA system based on digital twin;

[0051] The real-time model of the EHA system based on digital twins includes: the real-time geometric model of the EHA system and the real-time data model of the EHA system.

[0052] The real-time geometric model of the EHA system measures the specific dimensions of each component. The servo motor consists of a cylinder with an outer diameter of 150mm, an inner diameter of 56mm, and a length of 358mm, and a connecting plate. The valve block is a cuboid with a length of 170mm, a width of 107mm, and a height of 206mm. The hydraulic cylinder is a cuboid with a length of 95mm, a width of 310mm, and a height of 130mm. The piston rod consists of a cylinder with a length of 534mm and a diameter of 40mm, and a piston with a diameter of 56mm. The oil tank is a cuboid with a length of 145mm, a width of 133mm, and a height of 155mm. The three-dimensional geometric model of the EHA system is drawn using Solidworks. Figure 2As shown, the servo motor's connecting plate is connected to the valve block via M10 screws; one side of the valve block is connected to the hydraulic pump and oil tank via 12 M6 screws, and the other side is connected to the hydraulic cylinder via 10 M6 screws; the piston rod moves back and forth inside the hydraulic cylinder and is sealed with O-rings via front and rear end caps. The 3D model is saved as an STL file and imported into a QT interface written in Python; the piston rod is the main moving part of the EHA system, and its specific movement is defined as the difference between two adjacent data acquisitions from the displacement sensor, achieving the requirement of synchronization between the physical prototype and the virtual geometric model of the EHA system.

[0053] The real-time data model of the EHA system includes sensor data such as displacement, temperature, rotational speed, and pressure, data generated by the TimeGAN model and machine learning model, and predicted data. This data is backed up and stored for future research. Specifically, the displacement sensor outputs 1V to 4V, corresponding to a piston rod movement distance of -50mm to 50mm; the temperature sensor measures the temperature of the servo motor and hydraulic oil, with a range of 20℃ to 150℃; the rotational speed sensor measures the real-time rotational speed of the servo motor, with a range of -3300rpm to 3300rpm; and the pressure sensor measures the pressure in the two chambers of the hydraulic cylinder, with a range of 0MPa to 5MPa. Figure 3 As shown, each sensor is connected to a data acquisition card to acquire and transmit data. The data acquisition card then transmits the data to the host computer. The Nidaqmx module is called in Python to read the sensor data, perform data processing, and plot real-time data curves. The model-generated data, prediction data, and sensor data are stored in a MySQL database.

[0054] S2: Small sample data extension of the EHA system based on the time series data augmentation method TimeGAN;

[0055] Taking EHA historical sensor data and historical performance indicators as examples, this paper studies the practicality of the TimeGAN data extension method.

[0056] Step 1: Perform linear normalization on the EHA data.

[0057] X std =(XX) min ) / (X max -X min (9)

[0058] Among them, X std X, X min X maxThese represent the standard value after normalization, the current feature data, the maximum value of the feature data, and the minimum value of the feature data, respectively. A sliding window is used to preprocess the normalized data. The data in each window is used to generate new data points or sequences. By sliding a fixed-size window in the data, the model's ability to capture time series is enhanced.

[0059] Step 2: Input the data from each window into the embedding function e and the recovery function r. The embedding function converts the data into a form that is easy for the model to understand and process, while the recovery function ensures that the dimensions and format of the generated data are consistent with the original data, thus making the generated samples realistic and comparable. The data reconstruction process leads to the reconstruction loss L. R :

[0060]

[0061] in, The mathematical expectation representing tense and static characteristics; S, These represent the static features and the reconstructed static features, respectively; X t , These represent the temporal characteristics at time t and the reconstructed temporal characteristics, respectively.

[0062] Step 3: Train the TimeGAN generator and discriminator in the latent vector space containing the embedding function e and the recovery function r, thereby introducing the adversarial loss L. U :

[0063]

[0064] In the formula, y S y t These represent the static and temporal features output by the generator, respectively. This represents the static and temporal features of the discriminator output, in the context of adversarial loss L. U In this process, the generator aims to minimize this loss to generate more realistic samples; the discriminator aims to minimize the gap between the real data and the real labels, and maximize the gap between the real data and the generated labels.

[0065] Step 4: To ensure the generator accurately learns the static and temporal features of real data, the TimeGAN training process introduces a supervised loss Li. S This is used to measure the difference between the static and temporal feature distributions of real data and those of generated data.

[0066]

[0067] In the formula, g X (h S ,ht-1 ,z t ) represents the generator's output, which is located in the random vector z. t Under the incentive of the generator, it is generated and optimized during the gradient descent process.

[0068] By jointly training with three loss functions, saving the TimeGAN output and performing inverse normalization, the EHA extended data generation is completed.

[0069] The EHA historical data uses the average of all data points within 1 minute as 1 point, and the system state is reflected by observing the trend changes of the curves of 40 point values ​​over 40 minutes. After training the TimeGAN model, a sliding window of extended data and original data is randomly printed, as shown in Figure 4. The extended data has identified the potential patterns in the initial data; in Figure 4, the solid line represents the original data, and the dashed line represents the TimeGAN extended data. Figure 4a This refers to the change in motor speed over 40 minutes, ranging from 2900 rpm to 2950 rpm. Figure 4b The demonstration shows the pressure in the left chamber of the hydraulic cylinder, ranging from 3.5 MPa to 5.5 MPa; Figure 4c This refers to the pressure change in the right chamber of the hydraulic cylinder, ranging from 3.5 MPa to 5.5 MPa. Figure 4d The temperature variation of the servo motor ranges from 30°C to 90°C. Figure 5a good Figure 5b The evaluation of the trained model is shown, which is divided into principal component analysis (PCA) and dimensionality reduction visualization analysis (t-SNE); where black dots represent the original sensor data and red dots represent TimeGAN extended data. Figure 5a The PCA axes range from -1.5 to 1.5, representing the standard deviation unit range along the principal component direction. The purpose of drawing PCA plots is to observe the distribution of different datasets (original data and extended data) along the principal component direction in order to assess their similarity or difference in the feature space. Figure 5b The t-SNE coordinates range from -15 to 15. In the t-SNE algorithm, by taking coordinate values ​​from large negative to large positive values, it can better represent the complex relationships in the high-dimensional space after dimensionality reduction. Figure 5a , Figure 5b The red dots (extended data) can basically cover the black dots (original data), indicating that the TimeGAN method can effectively identify the potential patterns in the EHA system dataset.

[0070] S3: EHA System Performance Prediction and Real-time Monitoring;

[0071] The EHA system performance prediction involves using a machine learning CNN-BiLSTM-Attention model to predict EHA system performance metrics. The specific process is as follows: Figure 6 As shown in the figure; in this invention, the instruction response time of the EHA system is used as a performance prediction indicator, and the change in the operating status of the EHA system is judged by the magnitude of the predicted value of this indicator.

[0072] Historical EHA sensor data and historical EHA performance index data are split into training and testing sets at a ratio of 0.8:0.2 and imported into the machine learning CNN-BiLSTM-Attention prediction model for training. After the model recognizes the potential correlation between the two, it can obtain real-time performance index data by inputting real-time sensor data, thereby realizing real-time prediction of the EHA system's operating status.

[0073] In the CNN-BiLSTM-Attention model, the convolutional neural network (CNN) extracts local features and reduces dimensionality from the input signal data. The convolutional layer performs nonlinear activation transformation on the output, and the pooling layer is used to reduce the feature dimensionality and extract the most significant features. Then, the fully connected layer performs linear combination and nonlinear transformation on the features to increase the network's nonlinear capability.

[0074]

[0075] f(x) = max(0,x) (14)

[0076]

[0077] in, Here, f(·) is the input signal, f(·) is the activation function, down(·) is the downsampling function, and N is the number of input features to be mapped. It is the j-th output feature of the l-th layer. Represents the convolution kernel. This indicates the deviation term.

[0078] The processed feature sequence is input into a bidirectional long short-term memory (BiLSTM) network, which consists of two LSTM layers. The forward LSTM layer processes the input sequence forward, while the backward LSTM layer processes it backward. The outputs of the forward and backward layers are concatenated to obtain the final output y of the BiLSTM. t :

[0079]

[0080] Where, x t It is the input at time t. It is the forward hidden state at time t-1. It is the forward cell state at time t-1. This represents the forward hidden state at time t+1. This represents the forward unit state at time t+1. t This represents the final output of the combined two LSTM layers, where each time step corresponds to a vector representation containing the hidden state information for that time step.

[0081] Attention mechanisms distinguish important and redundant information by assigning different weights; the attention mechanism affects the output vector y of the BiLSTM. t Perform a weighted summation. Calculate y. t Score s t Thus, we obtain y t The degree of influence on the output value; the impact on the score s t The score is numerically transformed using the Softmax function to obtain the weight coefficient 'a'. t ; and according to the weighting coefficient a t and y t The final Attention output value o is calculated. t :

[0082] s t =tanh(W h y t +b h (19)

[0083] a t =Softmax(s t (20)

[0084]

[0085] Among them, W h b represents the weights of Attention. h This is the bias term for Attention.

[0086] Note the mechanism output o t Then it enters the Flatten layer to reduce the dimensionality to a one-dimensional array, and then enters the output layer to output the predicted values ​​of the performance indicators, thus completing the model training.

[0087] The test set data is input into the training model, and the quality of the training model is judged by comparing the root mean square error (RMSE) between the predicted output performance index and the historical performance index.

[0088] Using historical datasets from the EHA system as an example, the proposed EHA system performance prediction method based on TimeGAN data extension is validated. Figure 7As shown, compared with inputting only raw sensor data (Number of input = 0), the prediction performance of the TimeGAN-generated data improved by 29.26%. Furthermore, in all three different prediction models, TimeGAN-generated data significantly improved the prediction accuracy of EHA system performance indicators. Specific examples are shown below. Figure 8a , Figure 8b As shown, the implementation is the raw data, the dashed line is the predicted value of the machine learning model, and the vertical axis is the EHA performance index, i.e., instruction response time. Figure 8a It involves inputting only the raw EHA data into the machine learning model for prediction; Figure 8b The prediction is performed with the original input data plus a set of TimeGAN extended data. The prediction performance of the machine learning model with TimeGAN extended data is significantly better than that of the model without extended data.

[0089] The EHA system provides real-time monitoring, meaning it establishes a digital twin user interface that communicates with the physical prototype in real time. Figure 9 , 10 As shown, the user interface is composed of a UI written in Python, which includes the EHA system digital twin real-time geometric model, real-time data display and curve trend plotting of each sensor, EHA system control mode selection and related parameter adjustment, EHA system thermal imager real-time temperature display interface, EHA system performance index real-time predicted value display, EHA system data acquisition card device connection and data storage, etc.

[0090] The EHA digital twin geometric model uses a piston rod as the main moving part and a hydraulic pump and motor as the main temperature rise components. The geometric model is split and saved in STL format and imported into Python. The movement of the piston rod geometric model is correlated with information collected by displacement sensors, defined as the difference between two consecutive collected values ​​to achieve virtual-real synchronization between the geometric model and the physical prototype. For the main temperature rise components: the servo motor and hydraulic pump, color functions are constructed to correlate the appearance color of their geometric models with their temperature sensor data. Once the temperature exceeds a set threshold, the color changes from green to red.

[0091] The real-time data from the sensors includes pressure signals from the two chambers of the EHA system hydraulic cylinder, piston rod displacement signals, servo motor temperature signals, hydraulic pump temperature signals, and servo motor speed signals.

[0092] The EHA system control mode selection and parameter adjustment; the data acquisition card outputs voltage signals to the servo driver, which then controls the motor to drive the hydraulic pump at different speeds; among them, the manual control mode controls the system by manually inputting the values ​​output by the data acquisition card, while the automatic control mode mainly controls the EHA system based on the piston rod displacement signals of the sensor and the target command.

[0093] The real-time predicted values ​​of the EHA performance indicators are displayed; the predicted values ​​of the performance indicators are displayed on the main interface of the data interface and plotted as curves with real-time data from various sensors.

[0094] The EHA system thermal imager temperature display interface; due to its highly integrated nature, the EHA system is prone to abnormal temperature rise problems. Therefore, a thermal imaging camera is used, and the thermal imager temperature display webpage is called through the user interface using Python to monitor the temperature of each component of the EHA system in real time.

Claims

1. A method for predicting the performance of typical electromechanical systems based on small sample data extension, characterized in that, Includes the following steps: S1: Real-time model building of electromechanical systems based on digital twins; The digital twin electromechanical system real-time model comprises two parts: a digital twin real-time geometric model of the electromechanical system and a digital twin real-time data model; A real-time geometric model of a digital twin of an electromechanical system is built. The model's external dimensions and the operating status of its movable parts must be consistent with the physical prototype. The virtual geometric model needs to provide real-time feedback on various physical information and operating status of the real physical prototype to achieve the requirement of virtual-real synchronization. A real-time data model of a digital twin of an electromechanical system is built. Sensor data is collected in real time through an acquisition system and uploaded to a host computer for storage and backup for subsequent research. At the same time, the real-time data is fed into a data extension model and a machine learning prediction model to realize the real-time updating and output of the model. S2: Small sample data extension of electromechanical systems based on the time series data augmentation method TimeGAN; TimeGAN, a time series data augmentation method, extends the time-series sensor data of electromechanical systems. Through adversarial training between the generator and the discriminator, it reaches Nash equilibrium. The generator can extend new data with the same underlying patterns as the original data. TimeGAN adds an autoencoder on the basis of GAN to better handle the characteristics of time series data. S3: Performance prediction and real-time monitoring of electromechanical systems; The performance prediction of electromechanical systems uses a machine learning model to predict the performance indicators of electromechanical systems. The machine learning model is combined with the TimeGAN method for small sample data extension. The initial data and TimeGAN extended data are synchronously input into the machine learning model of the electromechanical system. The quality of the trained model is evaluated based on the root mean square error (RMSE) and mean square error (MSE). The performance operation status of the electromechanical system is judged by the predicted values ​​of the electromechanical system performance indicators. The real-time monitoring of the electromechanical system is designed as a digital twin user interface, allowing users to grasp the overall operating status of the electromechanical system through the real-time system-related information displayed on the interface. The real-time geometric model of the electromechanical system is displayed on the main interface, synchronizing virtual and real, allowing users to observe the operating status of the physical prototype in real time through the virtual geometric model. For the real-time data model, storage functions for model data, predicted data, and sensor data are added for subsequent research. The performance index values ​​predicted by the machine learning model are displayed in real time on the data interface, and the operating parameters of each model are added, allowing users to make adaptive changes based on the operating status and model evaluation indicators. Sensor data is displayed in real time on the data interface, and curves are plotted to more intuitively reflect the system status.

2. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 1, characterized in that: In step S1, the electromechanical system is an EHA system, including displacement, temperature, speed, and pressure sensor data, TimeGAN model and machine learning model generated data, and prediction data. The data is backed up and stored for subsequent research. Among them, the displacement sensor output voltage is 1V to 4V, corresponding to the piston rod movement distance of -50mm to 50mm; the temperature sensor is used to measure the temperature of the servo motor and hydraulic oil, with a value range of 20℃ to 150℃; the speed sensor is used to measure the real-time speed of the servo motor, with a value range of -3300rpm to 3300rpm; the pressure sensor is used to measure the pressure of the two chambers of the hydraulic cylinder, with a value range of 0MPa to 5MPa. Each sensor is connected to a data acquisition card to realize data acquisition and transmission. The data acquisition card then transmits the data to the host computer. The Nidaqmx module is called in Python to read the sensor data, perform data processing, and plot real-time data curves. The model generated data, prediction data, and sensor data are stored in a MySQL database.

3. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 1, characterized in that: In step S2, the autoencoder includes an encoder and a decoder, whose core components are an embedding function e and a recovery function r, respectively. The embedding function e maps the static and temporal features of the original data to a low-dimensional latent space, and the recovery function r maps the learned low-dimensional features back to the original data feature space. The embedding function e and the recovery function r are defined as follows: Where S and X represent the static and temporal characteristics of the data, respectively, and H S H X These represent the low-dimensional latent spaces corresponding to static and temporal features, respectively.

4. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 3, characterized in that: The embedding function e is implemented using a recurrent neural network, which maps the static and temporal features of the training data into a low-dimensional latent vector. The embedding function e is implemented using a recurrent neural network. h S ,h 1:T =e(S,X 1:T ) (3) h S =e S (s) (4) h t =e X (h S ,h t-1 ,x t ) (5) In the formula, h S It is the low-dimensional latent vector corresponding to the static feature; h 1:T h is the low-dimensional latent vector corresponding to the temporal features from the beginning to time T. t h t-1 These represent the low-dimensional latent vectors corresponding to the temporal features at time t and time t-1, respectively; e S With e X These are the embedding functions that transform static features and temporal features into low-dimensional latent vectors, respectively.

5. The method for predicting the performance of a typical electromechanical system based on small sample data extension according to claim 3, characterized in that: The recovery function r restores the low-dimensional latent vector into static and temporal features. This recovery function r is implemented using a feedforward neural network. S,X 1:T =r(h S ,h 1:T ) (6) S=r S (h S ) (7) X t =r X (h t ) (8) In the formula, S and X 1:T These represent the latent vectors h of static features, respectively, obtained by the recovery function r. S With temporal feature latent vector h 1:T Recovering static and temporal features, r S and r X These are the recovery networks corresponding to static features and temporal features, respectively, X. t Let be the temporal characteristics that are restored by the recovery function at time t.

6. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 1 or 3, characterized in that: Perform linear normalization: X std =(X-X min ) / (X max -X min ) (9) Among them, X std X, X min X max These represent the standard value after normalization, the current feature data, the maximum value of the feature data, and the minimum value of the feature data, respectively. A sliding window is used to preprocess the normalized data. The data in each window is used to generate new data points or sequences. By sliding a fixed-size window in the data, the model's ability to capture time series is enhanced.

7. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 6, characterized in that: The data from each window is input sequentially into the embedding function e and the recovery function r. The embedding function transforms the data into a form that is easy for the model to understand and process, while the recovery function ensures that the dimensions and format of the generated data remain consistent with the original data, thus making the generated samples realistic and comparable. The data reconstruction process leads to the reconstruction loss L. R : in, The mathematical expectation representing tense and static characteristics; S, These represent the static features and the reconstructed static features, respectively; X t , These represent the temporal characteristics at time t and the reconstructed temporal characteristics, respectively.

8. The method for predicting the performance of a typical electromechanical system based on small sample data extension according to claim 7, characterized in that: In the latent vector space containing the embedding function e and the recovery function r, the generator and discriminator of TimeGAN are trained, leading to the adversarial loss L. U : In the formula, y S y t These represent the static and temporal features output by the generator, respectively. This represents the static and temporal features of the discriminator output, in the context of adversarial loss L. U In this process, the generator aims to minimize this loss in order to generate more realistic samples. The discriminator aims to minimize the gap between the real data and the real labels, and maximize the gap between the real data and the generated labels.

9. The method for predicting the performance of a typical electromechanical system based on small sample data extension as described in claim 8, characterized in that: To enable the generator to accurately learn the static and temporal features of real data, TimeGAN introduces a supervised loss L during training. S This is used to measure the difference between the static and temporal feature distributions of real data and those of generated data. In the formula, g X (h S ,h t-1 ,z t ) represents the generator's output, which is located in the random vector z. t Under the incentive of the generator, it is generated and optimized during the gradient descent process.

10. The method for predicting the performance of a typical electromechanical system based on small sample data extension according to claim 1, characterized in that: In step S3, a convolutional neural network (CNN) is used to extract local features and reduce the dimensionality of the input signal data. The convolutional layer performs nonlinear activation transformation on the output, and the pooling layer is used to reduce the feature dimensionality and extract the most significant features. Then, a fully connected layer is used to perform linear combination and nonlinear transformation on the features to increase the nonlinearity of the network. f(x) = max(0,x) (14) in, Here, f(·) is the input signal, f(·) is the activation function, down(·) is the downsampling function, and N is the number of input features to be mapped. It is the j-th output feature of the l-th layer. Represents the convolution kernel. Indicates the deviation term; The processed feature sequence is input into a bidirectional long short-term memory (BiLSTM) network, which consists of two LSTM layers. The forward LSTM layer processes the input sequence forward, while the backward LSTM layer processes it backward. The outputs of the forward and backward layers are concatenated to obtain the final output y of the BiLSTM. t : Where, x t It is the input at time t. It is the forward hidden state at time t-1. It is the forward cell state at time t-1. This represents the forward hidden state at time t+1. The forward cell state at time t+1; y t This represents the final output of the combined two LSTM layers, where each time step corresponds to a vector representation containing the hidden state information of that time step; Attention mechanisms distinguish important and redundant information by assigning different weights; the attention mechanism affects the output vector y of the BiLSTM. t Perform a weighted summation; calculate y t Score s t Thus, we obtain y t The degree of influence on the output value; the impact on the score s t The score is numerically transformed using the Softmax function to obtain the weight coefficient 'a'. t ; and according to the weighting coefficient a t and y t The final Attention output value o is calculated. t : S t =tanh(W h y t +b h ) (19) a t =Softmax(s t ) (20) Among them, W h b represents the weights of Attention. h This represents the bias term of Attention; Note the mechanism output o t Then it enters the Flatten layer to reduce the dimensionality to a one-dimensional array, and then enters the output layer to output the predicted values ​​of the performance indicators, thus completing the model training; The test set data is input into the training model, and the quality of the training model is judged by comparing the root mean square error (RMSE) between the predicted output performance index and the historical performance index.

Images