Hydraulic system fault prediction method for forging hydraulic press based on digital twin system

Abstract

The present application relates to a kind of hydraulic system fault prediction method of forging hydraulic press based on digital twin system, it includes: S1, the geometric parameter of the hydraulic system of forging hydraulic press is acquired, and digital twin model is constructed;S2, the digital twin model of the hydraulic system of forging hydraulic press is realized by using DNN depth neural network based on the order reduction, and the parameterized model of the hydraulic system of forging hydraulic press is established;S3, the order reduction of the digital twin model of forging hydraulic press is completed, and the fault diagnosis and prediction of hydraulic system are realized.The present application adopts multiscale physical model, realizes the data environment interaction of physical space and virtual body space, finally realizes the intelligent decision and maintenance of forging hydraulic press.The present application can realize the fault diagnosis, online monitoring and performance optimization of hydraulic system, reach predictive maintenance, and then realize the development of virtual verification to virtual-real interaction full closed-loop optimization driven by data, comply with the development needs of intelligent equipment, provide new platform for the design and development and performance upgrade of forging equipment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, and specifically relates to a method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system. Background Technology

[0002] Forging hydraulic presses are the mother machines of the manufacturing industry, widely used in aerospace, nuclear power, supercritical and combined cycle power generation, marine engineering, and other fields. They play a vital role in national economic and defense security, and their technological level reflects a country's manufacturing development level. Forging hydraulic presses are heavy and complex equipment integrating mechanical, electrical, hydraulic, control, and sensing technologies. Their design and manufacturing cycles are long, with large investments, high risks, and high costs for online testing. Furthermore, the high pressure and flow rate in the hydraulic system, coupled with high-energy-density alternating impact loads, easily lead to fatigue damage to hydraulic components and the system, causing frequent system failures and, in severe cases, fires. The processed objects are of high value, and equipment failures can easily result in the scrapping of processed parts. Therefore, to address these issues, this invention takes forging hydraulic presses as the research object and constructs a digitally driven predictive maintenance system for forging hydraulic presses. This system provides a simulation environment for the design of forging hydraulic press equipment, reducing design cycles and costs, predicting and optimizing its control performance, realizing full life-cycle status monitoring of the frequently failing hydraulic control system, and performing intelligent fault diagnosis and performance prediction. However, establishing a realistic mapping relationship is a key issue and the foundation for applying digital twin technology to system optimization and fault diagnosis.

[0003] The core issue in constructing a digital twin system is realizing the interaction between the physical world and the virtual information world. Digital twin technology is an effective way to integrate and interact with the physical world and the digital information world. As a mapping from physical space to a digital twin model, the construction of a digital twin of a forging press, and the fusion and interaction of virtual and physical data, are important foundations for realizing a twin system. The construction of the twin model is a crucial foundational step, and choosing an appropriate modeling method is critical for subsequent research. Furthermore, due to the nonlinear nature of the hydraulic press system, its traditional mathematical model is too complex and cannot meet the real-time requirements of the twin platform. Therefore, it is necessary to reduce the order of the digital virtual entity.

[0004] In recent years, the theoretical development of digital twins has been rapid. Domestic and international research on the interaction between digital twins and physical entities of forging hydraulic presses has provided a unified reference architecture. However, this architecture lacks flexibility, fails to consider the varying real-time requirements of the interaction process, and struggles to achieve efficient utilization of architectural resources. Furthermore, traditional hydraulic press modeling methods cannot feed back the forging press's operating condition information to the digital model in real time, lacking real-time interaction between the simulation model and physical data, thus constituting offline simulation. Therefore, it is necessary to study the construction of digital twins for forging hydraulic presses and explore methods for building such systems, thereby promoting industrial upgrading and intelligent development in the forging equipment industry. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system. This invention employs a multi-scale physical model and, through multidisciplinary collaborative simulation, establishes an interactive data environment between the physical and virtual spaces, ultimately achieving intelligent decision-making and maintenance of the forging hydraulic press. This invention enables intelligent fault diagnosis, online monitoring, and performance optimization upgrades for the hydraulic press, achieving predictive maintenance. Furthermore, it realizes a data-driven, closed-loop optimization development from virtual verification to virtual-physical interaction, meeting the needs of intelligent equipment development and providing a new platform for the design, development, and performance upgrade of forging equipment.

[0006] Specifically, this invention provides a method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system, the specific steps of which are as follows:

[0007] S1: Obtain the geometric parameters of the hydraulic system of the forging hydraulic press and construct a digital twin model;

[0008] By acquiring key geometric real-time deformation parameters of the hydraulic system of the forging hydraulic press through sensors, the digital twin model of the hydraulic system of the forging hydraulic press is reconstructed in real time.

[0009] S2: Utilize a DNN-based deep neural network to reduce the order of the digital twin model of the hydraulic system of the forging hydraulic press and establish a parameterized model of the hydraulic system of the forging hydraulic press;

[0010] S21: Use the pruning operation in the DNN model reduction method to remove convolutional kernels with small contributions in the digital twin model of the forging hydraulic press hydraulic system; set the i-th convolutional layer of the neural network in the digital twin model to F. i * W i F i * ∈R C , representing the C groups of signals input to the convolutional layer; W i ∈R C ' ×C The convolution kernel represents the weight matrix set in the convolutional layer, outputs C' sets of new data, and finally prunes W. i A convolution kernel with a small contribution; the expression for the convolution operation is:

[0011]

[0012] Where y is the output of the convolutional layer; x c b is the input to the convolutional layer; c is the bias parameter; C is the total number of input signals in the layer batch; c is the input signal number; W c The convolution kernel of the weight matrix set for the layer;

[0013] S22: Quantize the weight parameters in the digital twin model of the forging hydraulic press hydraulic system using the quantization processing in the DNN model order reduction method; use W... l Let {W} represent the weight set of the l-th layer of the neural network in the digital twin model, where L is the total number of layers. The trained model is represented by {W}. l The parameters of the fully connected layer and the convolutional layer are represented by 2D and 4D respectively, and the weight set W of the l-th layer is used to represent the parameters. l Quantization into low precision The quantized parameters are powers of 0 or 2, satisfying the quantization set P. l Specifically:

[0014] P l ={±2 n1 ,...,±2 n2 ,0};

[0015] Among them, P l Let n1 be the quantization set; n1 and n2 are the first quantization parameter and the second quantization parameter, respectively, and satisfy n2≤n1;

[0016] S3: Complete the reduction of the order of the digital twin model of the forging hydraulic press to realize fault diagnosis and prediction of the hydraulic system;

[0017] The digital twin model of the forging hydraulic press established in step S1 is obtained. The pruning operation in step S21 and the quantization processing in step S22 are performed respectively to obtain the reduced-order digital twin model of the forging hydraulic press. The operation of the hydraulic system of the forging hydraulic press is monitored online, and key geometric parameters are collected in real time to realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press.

[0018] Preferably, in step S1, the key geometric real-time deformation parameters of the forging hydraulic press hydraulic system are obtained through sensors, specifically as follows:

[0019] The key real-time geometric deformation parameters of the hydraulic system of the forging hydraulic press include: pressure parameters, displacement parameters, flow parameters, electrical parameters, environmental data, and force and torque data;

[0020] Pressure parameters in the hydraulic system of the forging press are collected using high-response pressure sensors; displacement parameters in the hydraulic system of the forging press are collected using displacement sensors; flow parameters are collected using flow sensors; electrical parameters, including current and voltage, are collected from the internal electrical parameters of each piece of equipment using electrical sensors; environmental data, including temperature, is collected using environmental sensors; and force and torque data at potential fault points in the forging press system are monitored using force sensors.

[0021] Preferably, the digital twin model of the hydraulic system of the forging hydraulic press in step S1 includes: a physical space, a virtual space that carries the digital twin system, and a communication interface connecting the physical space and the virtual space.

[0022] The physical space includes the physical space of the forging hydraulic press, which consists of a mechanical system, an electrical system, a hydraulic system, a control system, and sensors. The sensors are used to collect various data of the hydraulic press during operation, including pressure, flow rate, and temperature, and send the collected data to the virtual space via a communication interface.

[0023] The virtual space carrying the digital twin system includes a multi-dimensional virtual model, which includes a geometric expression module for establishing the system and a physically parameterized expression module for reflecting the relationship between the physical world and the virtual space.

[0024] Preferably, in step S21, the pruning operation in the DNN model reduction method is used to remove convolution kernels with small contributions from the digital twin model of the forging hydraulic press hydraulic system. Specifically:

[0025] First, the input variables of the neural network convolution channel in the digital twin model of the hydraulic system are substituted to make the following equation hold:

[0026]

[0027] in, The input variables for the convolution channels;

[0028] Simplifying the expression in step S21, we get the following expression:

[0029]

[0030] in, The input variables for the convolution channels;

[0031] The set of input variables for convolution channels The elements are independent of each other, representing a subset of the inputs to the digital twin model of the hydraulic system. Substituting into the above equation, we get:

[0032]

[0033] Where S is a subset of the inputs to the digital twin system;

[0034] If the above equation always holds true, then... The convolution kernels of other elements are redundant and are removed, i.e., pruning is performed.

[0035] Preferably, in step S22, the weight parameters in the digital twin model of the forging hydraulic press hydraulic system are quantized using the quantization processing in the DNN model reduction method, specifically as follows:

[0036] This makes the subset variable s of the input to the digital twin system = max(abs(W) l The formula for calculating the first quantization parameter n1 is as follows:

[0037]

[0038] Where floor represents the rounding operation; l represents the layer number of the DNN neural network model; and s represents a subset of the input variables of the digital twin system.

[0039] Meanwhile, based on the pre-set bias parameter b, the formula for calculating the second quantization parameter n2 is as follows:

[0040] n2 = n1 + 1 - 2 b-2 ;

[0041] At this point, the quantization set P can be determined. l Therefore Each item in the list can be quantified as:

[0042]

[0043] in, W is the quantization weight output for the l-th layer; l Let α be the weight set of the l-th layer; α and β be the quantization set P. l The first and second elements in the array are adjacent; sgn is a function that returns an integer variable; i and j are the first and second position numbers of the quantization weights; abs is the absolute value operation;

[0044] For the l-th layer, the grouping of weight parameters is as follows:

[0045]

[0046] in, For groups that need to be quantified; Groups that need further training;

[0047] To simplify the expression, a binary matrix T is introduced here. l :

[0048]

[0049] Among them, T l It is a binary matrix;

[0050] Therefore, the quantization process of the l-th layer is as follows:

[0051]

[0052] Among them, E(W) l ) represents the quantization result of the l-th layer; L(W l R(W) represents the model loss of the l-th layer; l ) represents the regularization term for the l-th layer; λ is the regularization coefficient; L is the total number of layers in the weight set;

[0053] The above formula can be further simplified as follows:

[0054]

[0055] Among them, the quantization set P l and binary matrix T l Given that the learning rate is set to γ, and using stochastic gradient descent, the above equation can be transformed into:

[0056]

[0057] Where γ is the learning rate;

[0058] The quantification process is now complete.

[0059] Preferably, the online diagnosis and prediction of hydraulic system faults in the forging hydraulic press in step S3 specifically includes:

[0060] S31: Determine the various fault states of the hydraulic system of the forging hydraulic press in the physical space, simulate different fault states, and collect data parameters under different fault states;

[0061] S32: Use sensors to collect data on the forging hydraulic press in the physical space, including raw data on flow rate, pressure, temperature at key locations in the system, as well as the displacement of the forging hydraulic press, and classify the raw data according to the fault type;

[0062] S33: Collect data on the types of faults that may occur during the operation of the forging hydraulic press in the physical space. Data preprocessing is required to remove noise, outliers and missing values ​​from the raw data to obtain the processed data.

[0063] S34: Import the processed data into the parametric model of the hydraulic system of the forging hydraulic press to realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] This invention provides a method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system. It integrates the geometric and physical models of the actual forging hydraulic press system, digitally mapping the system from multiple perspectives. Specifically, it utilizes multidisciplinary simulation software to establish a digital virtual entity of the forging hydraulic press and combines advanced sensor technology to extract a finite number of geometric deformation parameters of the physical entity in real time. This data-driven online real-time reconstruction of the press's high-confidence reduced-order model facilitates the fusion of multiple models in complex physical systems, achieving maximum consistency between the physical and virtual spaces. Furthermore, this invention establishes a three-dimensional display and control platform within the virtual space, creating a human-computer interaction system and improving the visualization performance of the digital twin platform. Attached Figure Description

[0066] Figure 1 This is a flowchart of the hydraulic system fault prediction method for a forging hydraulic press based on a digital twin system according to the present invention;

[0067] Figure 2 This is a schematic diagram of the fault diagnosis structure of the present invention;

[0068] Figure 3 This is a Simulink control program diagram of the forging hydraulic press of the present invention;

[0069] Figure 4 Screenshot of the three-dimensional display and control platform for the virtual space of the present invention;

[0070] Figure 5 This is a diagram showing the relationship between the hydraulic press displacement and the main pump pressure according to the present invention.

[0071] Figure 6 This diagram illustrates the processing of displacement data and pressure data at the main pump inlet of the forging hydraulic press according to the present invention. Detailed Implementation

[0072] Exemplary embodiments, features, and aspects of the present invention will now be described in detail with reference to the accompanying drawings. Although various parameter values ​​and aspects of the embodiments are shown, the simulation process need not be performed with exactly the same parameters and aspects unless specifically indicated.

[0073] The present invention provides a hydraulic system fault prediction method for a forging hydraulic press based on a digital twin system, such as... Figure 1 As shown, a digital twin model is constructed, and a model reduction method based on deep neural networks (DNN) is used to reduce the dimensionality of the digital twin model, thereby reducing the dimensionality of the digital twin model of the forging hydraulic press and enabling fault diagnosis and prediction; this includes:

[0074] S1: Obtain the geometric parameters of the hydraulic system of the forging hydraulic press and construct a digital twin model.

[0075] Real-time deformation parameters of key geometric features in the forging hydraulic press are acquired using sensors, enabling real-time reconstruction of the press's digital twin model. These key geometric parameters include pressure, displacement, flow, electrical parameters, environmental data, and force and torque data. High-response pressure sensors acquire pressure parameters within the hydraulic system; displacement sensors acquire displacement parameters; flow sensors acquire flow parameters; electrical sensors acquire internal electrical parameters (including current and voltage); environmental sensors acquire environmental data (including temperature); and force sensors monitor force and torque data at potential fault points within the forging press system.

[0076] like Figure 2 This is a schematic diagram of the fault diagnosis structure of the present invention, a digital twin model of a forging hydraulic press, including: a physical space, a virtual space carrying the digital twin system, and a communication interface connecting the physical space and the virtual space. The physical space includes the forging hydraulic press, which consists of a mechanical system, an electrical system, a hydraulic system, a control system, and sensors. The sensors are used to collect various data of the hydraulic press during operation, such as pressure, flow rate, and temperature, and send the collected data to the virtual space via the communication interface.

[0077] Simulink control program for forging hydraulic presses, such as Figure 3 As shown, data collected by sensors in the physical space, such as pressure, flow rate at key locations in the hydraulic system, and displacement of the moving crossbeam of the forging hydraulic press, are used to deploy and run the Simulink control algorithm of the forging hydraulic press through the xPC (External Mode for PC) real-time target toolkit in MATLAB. This toolkit provides a communication interface between the host computer and the slave computer, enabling real-time data transmission and parameter adjustment during the operation of the forging hydraulic press in the physical space, as well as real-time monitoring and debugging of the control algorithm's performance.

[0078] After sensor data in physical space is transmitted to MATLAB via a communication interface, it has two possible destinations:

[0079] (1) By adding Byte Packing and UDP Send blocks to the Simulink program, defining the data type as 'double', configuring the IP as local transfer '127.0.0.1', and defining the port as '23333', data acquired by sensors in the physical space is sent to Unity3D via the UDP protocol. Simultaneously, a C# script is written in Unity3D to listen for data from Simulink on port '23333'. In this patent, the Socket class in Unity3D is used to implement UDP data reception; that is, a UDP Socket object is created using the C# Socket class, and the ReceiveFrom function is used to receive data from Simulink. After receiving the data, it is parsed and processed in Unity3D.

[0080] There are two ways to handle this:

[0081] 1) The transmitted data can be directly visualized using the XCharts component in Unity3D, displaying key data such as flow rate, pressure, and oil temperature at critical points during the operation of the forging hydraulic press in real time in the virtual space. This operation requires writing corresponding C# scripts based on the visualization chart type; 2) The transmitted data can also be applied to the virtual 3D model of the forging hydraulic press in Unity3D via C# scripts, specifically to the moving parts of the forging hydraulic press body: the hydraulic rod and the moving crossbeam. This script transmits the displacement data of the moving crossbeam of the forging hydraulic press in physical space to the 3D model in virtual space, achieving synchronous virtual-real operation and interaction of the forging hydraulic press. Figure 5 The diagram shown illustrates the relationship between the hydraulic press displacement and the main pump pressure according to the present invention. Figure 6 This is a schematic diagram of data processing.

[0082] (2) Using JDBC drivers, MATLAB is connected to a MySQL database to store data obtained from the physical space for later retrieval. The communication interface connecting the physical and virtual spaces is mainly used for information transmission, including wireless and wired communication interfaces. The control and transmission of the forging hydraulic press in the physical space are connected via a wired communication interface using xPC, while other information transmissions are connected via wireless communication interfaces using protocols such as UDP and TCP / IP. The virtual space carrying the digital twin system includes a multidimensional virtual model used to establish an accurate, multidimensional digital mapping of the physical world. Because the hydraulic press system is nonlinear, its traditional mathematical model is too complex to meet the real-time requirements of the digital twin platform, requiring a reduction in the order of the multidimensional virtual model. This multidimensional virtual model includes: a geometric expression module for establishing the system's geometric model; and a physical parameterization expression module for reflecting the relationship between the physical world and the virtual space.

[0083] The geometric representation module will establish a press dynamics model based on component motion relationships and three-dimensional geometric constraints; considering the characteristics of nonlinear flow and time-varying load in the press hydraulic system, a high-precision model of the press hydraulic system will be established; considering the characteristics of parameter perturbation, dead zone, and saturation, a high-precision mathematical model of the electrical drive system and sensing elements will be established.

[0084] The virtual space also includes the accuracy of health assessment, diagnosis, and prediction algorithms for forging hydraulic presses in a cloud-edge collaborative environment, as well as the deployment requirements for resource optimization, in order to achieve cloud-edge resource collaboration and real-time, low-energy fault diagnosis.

[0085] like Figure 4 The image shown is a screenshot of the three-dimensional display and control platform for the virtual space of the present invention. This three-dimensional display and control platform uses virtual reality technology to perform immersive rendering of multi-dimensional virtual models, multi-scale visualization technology based on deep learning, and provides an interactive interface between humans and the system. It also develops a basic support platform, as well as health management, fault tracing, and life prediction apps to realize cloud-based equipment status assessment and operation and maintenance strategy formulation, edge-based fault diagnosis and early warning and operation and maintenance plan execution, and dynamic optimization of the operation and maintenance process through cloud-edge collaboration.

[0086] S2: Utilize a DNN-based deep neural network to reduce the order of the digital twin model of the hydraulic system of the forging hydraulic press, and establish a parameterized model of the hydraulic system of the forging hydraulic press.

[0087] S21: Use the pruning operation in the DNN model reduction method to remove convolution kernels with small contributions from the digital twin model of the hydraulic system of the forging hydraulic press. Specifically:

[0088] First, the input variables of the neural network convolution channel in the digital twin model of the hydraulic system are substituted to make the following equation hold:

[0089]

[0090] in, is the input variable for the convolution channel.

[0091] Simplifying the expression in step S21, we get the following expression:

[0092]

[0093] in, is the input variable for the convolution channel.

[0094] The set of input variables for convolution channels The elements are independent of each other, representing a subset of the inputs to the digital twin model of the hydraulic system. Substituting into the above equation, we get:

[0095]

[0096] Where S is a subset of the inputs to the digital twin system.

[0097] If the above equation always holds true, then... The convolution kernels of other elements are redundant and are removed, i.e., pruning is performed.

[0098] In a digital twin model, the i-th convolutional layer of a neural network is defined as F. i * W i F i * ∈R C , representing the C groups of signals input to the convolutional layer; W i ∈R C'×C , representing the convolution kernel with weights set in the convolutional layer, outputs C' sets of new data, and finally prunes W. i A convolution kernel with a small contribution; the expression for the convolution operation is:

[0099]

[0100] Where y is the output of the convolutional layer; x c b is the input to the convolutional layer; c is the bias parameter; C is the total number of input signals in the layer batch; c is the input signal number; W c The convolution kernel of the weight matrix set for the layer.

[0101] S22: Using the quantization process in the DNN model reduction method, the weight parameters in the digital twin model of the forging hydraulic press hydraulic system are quantized, specifically as follows:

[0102] This makes the subset variable s of the input to the digital twin system = max(abs(W)l The formula for calculating the first quantization parameter n1 is as follows:

[0103]

[0104] Where floor is the rounding operation; l is the layer number of the DNN neural network model; and s is a subset of the input variables of the digital twin system.

[0105] Meanwhile, based on the pre-set bias parameter b, the formula for calculating the second quantization parameter n2 is as follows:

[0106] n2 = n1 + 1 - 2 b-2 .

[0107] At this point, the quantization set P can be determined. l Therefore Each item in the list can be quantified as:

[0108]

[0109] in, W is the quantization weight output for the l-th layer; l Let α be the weight set of the l-th layer; α and β be the quantization set P. l The first and second elements in the array are adjacent; sgn is a function that returns an integer variable; i and j are the first and second position numbers of the quantization weights; abs is the absolute value operation.

[0110] For the l-th layer, the grouping of weight parameters is as follows:

[0111]

[0112] in, For groups that need to be quantified; The group that needs further training.

[0113] To simplify the expression, a binary matrix T is introduced here. l :

[0114]

[0115] Among them, T l It is a binary matrix.

[0116] Therefore, the quantization process of the l-th layer is as follows:

[0117]

[0118] Among them, E(W) l ) represents the quantization result of the l-th layer; L(W l R(W) represents the model loss of the l-th layer;l ) represents the regularization term for the l-th layer; λ is the regularization coefficient; and L is the total number of layers in the weight set.

[0119] The above formula can be further simplified as follows:

[0120]

[0121] Among them, the quantization set P l and binary matrix T l Given that the learning rate is set to γ, and using stochastic gradient descent, the above equation can be transformed into:

[0122]

[0123] Where γ is the learning rate;

[0124] The quantification process is now complete.

[0125] Use W l Let {W} represent the set of weights for the l-th layer of the neural network in the digital twin model, where L is the total number of layers. The trained model is represented by {W}. l The parameters of the fully connected layer and the convolutional layer are represented by 2D and 4D respectively, and the weight set W of the l-th layer is used to represent the parameters. l Quantization into low precision The quantized parameters are powers of 0 or 2, satisfying the quantization set P. l Specifically:

[0126] P l ={±2 n1 ,...,±2 n2 ,0};

[0127] Among them, P l Let n be the quantization set; n1 and n2 are the first quantization parameter and the second quantization parameter, respectively, and n2 ≤ n1.

[0128] S3: Complete the reduction of the order of the digital twin model of the forging hydraulic press, and realize the fault diagnosis and prediction of the hydraulic system.

[0129] Obtain the digital twin model of the forging hydraulic press established in step S1, perform the pruning operation in step S21 and the quantization processing in step S22 respectively, and obtain the reduced-order digital twin model of the forging hydraulic press. Monitor the operation of the hydraulic system of the forging hydraulic press online, collect key geometric parameters in real time, and realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press. Specifically:

[0130] S31: Determine the various fault states of the hydraulic system of the forging hydraulic press in the physical space, simulate different fault states, and collect data parameters under different fault states.

[0131] S32: Data is collected from the forging hydraulic press in physical space using sensors, including raw data on flow rate, pressure, temperature, and displacement at key locations in the system. This raw data is then categorized according to fault type. Since faults in the forging hydraulic press may result from the combined action of multiple components in the hydraulic system, and reproducing actual production faults could pose safety hazards, a controlled variable method is employed. This involves simulating the fault of a single component in the hydraulic system, collecting physical parameters at key locations, and identifying the mapping relationship between the fault state of this component and the fault state of the entire hydraulic system. This allows the system to be reflected from a single component, reducing potential safety hazards during fault simulation. Table 1 below simulates the fault of "forgings failing to form properly" that may occur during the operation of the forging hydraulic press.

[0132] Table 1 Simulation process of forging failure to form normally

[0133]

[0134]

[0135] S33: Collect data on the types of faults that may occur during the operation of the forging hydraulic press in the physical space. Data preprocessing is required to remove noise, outliers, and missing values ​​from the raw data to obtain the processed data.

[0136] S34: Import the processed data into the parametric model of the hydraulic system of the forging hydraulic press to realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press.

[0137] This invention addresses a fault where adjusting the overflow valve prevents the forging from forming properly. A decision tree algorithm is chosen, and Table 2 shows the accuracy results of 10 rounds of 10-fold cross-validation. The initial average accuracy of the decision tree algorithm is 0.666296. The accuracy of the 10 rounds of 10-fold cross-validation is slightly improved from the previous 65%-66%, reaching approximately 68%-69%. After the improvements made in this invention, the accuracy of the decision tree is significantly improved, stabilizing at around 89%-90%.

[0138] The accuracy results of 210 10-fold cross-validation are as follows:

[0139]

[0140] The beneficial effects of this invention are as follows: This invention provides a method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system. First, by comprehensively utilizing multiple software programs, all elements in the physical space, including the forging hydraulic press composed of mechanical, electrical, hydraulic, control, and sensor systems, are modeled at multiple levels, including geometric and physical dimensions. This allows the virtual space to reproduce the real physical space to the greatest extent possible, resulting in a digital twin model of the forging hydraulic press. Second, by combining advanced sensor technology, API interface technology, and model reduction and reconstruction technology, data collected by sensors in the physical space is transmitted in real time to the digital twin in the virtual space via a communication interface. The digital twin is then reduced in size using a DNN-based digital virtual model reduction method for the forging hydraulic press, resulting in a three-dimensional display and control platform for the forging hydraulic press. Finally, by utilizing virtual reality technology to immerse the multi-dimensional virtual model in rendering, employing deep learning-based multi-scale visualization technology, and providing an interactive interface between humans and the system, a basic support platform was developed. Furthermore, apps for health management, fault tracing, and lifespan prediction were developed to achieve functions such as cloud-based equipment status assessment and maintenance strategy formulation, edge-based fault diagnosis, early warning, and maintenance plan execution, and dynamic optimization of the cloud-edge collaborative maintenance process. After the improvements made in this invention, the accuracy was significantly improved, stabilizing at approximately 89%-90%.

[0141] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system, characterized in that, It includes: S1: Obtain the geometric parameters of the hydraulic system of the forging hydraulic press and construct a digital twin model; By acquiring key geometric real-time deformation parameters of the hydraulic system of the forging hydraulic press through sensors, the digital twin model of the hydraulic system of the forging hydraulic press is reconstructed in real time. S2: Utilize a DNN-based deep neural network to reduce the order of the digital twin model of the hydraulic system of the forging hydraulic press and establish a parameterized model of the hydraulic system of the forging hydraulic press; S21: Use the pruning operation in the DNN model reduction method to remove convolutional kernels with small contributions from the digital twin model of the forging hydraulic press hydraulic system; set the i-th convolutional layer in the neural network of the digital twin model to be... ,in , representing the batch input of the convolutional layer Group signals; The convolution kernel represents the weight matrix set in the convolutional layer, and the output is... Set up new data and finally prune. A convolution kernel with a small contribution; the expression for the convolution operation is: ； in, This is the output of the convolutional layer; Input for the convolutional layer; These are bias parameters; The total number of batch input signals for the convolutional layer; Number the input signal; The convolution kernel of the weight matrix set for the c-th input signal of the convolutional layer; S22: Quantize the weight parameters in the digital twin model of the forging hydraulic press's hydraulic system using the quantization process in the DNN model order reduction method; To represent the first neural network in a digital twin model The weight set of the layer, The total number of layers is used in the trained model. To represent; the parameters of fully connected layers and convolutional layers are represented in 2D and 4D respectively, and the 3rd... Layer weight set Quantization into low precision The quantized parameters are powers of 0 or 2, satisfying the quantization set. Specifically: ； in, For quantized sets; and These are the first quantization parameter and the second quantization parameter, respectively, and satisfy the following conditions: ; S3: Complete the reduction of the order of the digital twin model of the forging hydraulic press to realize fault diagnosis and prediction of the hydraulic system; The digital twin model of the forging hydraulic press established in step S1 is obtained. The pruning operation in step S21 and the quantization processing in step S22 are performed respectively to obtain the reduced-order digital twin model of the forging hydraulic press. The operation of the hydraulic system of the forging hydraulic press is monitored online, and key geometric parameters are collected in real time to realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press.

2. The method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system according to claim 1, characterized in that: The key geometric real-time deformation parameters of the hydraulic system of the forging hydraulic press obtained by the sensor in step S1 are as follows: The key real-time geometric deformation parameters of the hydraulic system of the forging hydraulic press include: pressure parameters, displacement parameters, flow parameters, electrical parameters, environmental data, and force and torque data; Pressure parameters in the hydraulic system of the forging press are collected using high-response pressure sensors; displacement parameters in the hydraulic system of the forging press are collected using displacement sensors; flow parameters are collected using flow sensors; electrical parameters, including current and voltage, are collected from the internal electrical parameters of each piece of equipment using electrical sensors; environmental data, including temperature, is collected using environmental sensors; and force and torque data at potential fault points in the forging press system are monitored using force sensors.

3. The method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system according to claim 1, characterized in that: The digital twin model of the hydraulic system of the forging hydraulic press in step S1 includes: a physical space, a virtual space that carries the digital twin system, and a communication interface that connects the physical space and the virtual space. The physical space includes the physical space of the forging hydraulic press, which consists of a mechanical system, an electrical system, a hydraulic system, a control system, and sensors. The sensors are used to collect various data of the hydraulic press during operation, including pressure, flow rate, and temperature, and send the collected data to the virtual space via a communication interface. The virtual space carrying the digital twin system includes a multi-dimensional virtual model, which includes a geometric expression module for establishing the system and a physically parameterized expression module for reflecting the relationship between the physical world and the virtual space.

4. The method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system according to claim 1, characterized in that: In step S21, the pruning operation in the DNN model reduction method is used to remove convolution kernels with small contributions from the digital twin model of the forging hydraulic press hydraulic system. Specifically: First, the input variables of the neural network convolution channel in the digital twin model of the hydraulic system are substituted to make the following equation hold: ； in, The input variables for the convolution channels; Simplifying the expression in step S21, we get the following expression: ； in, The output variable of the convolution channel; The set of input variables for convolution channels The elements are independent of each other, representing a subset of the inputs to the digital twin model of the hydraulic system. Substituting into the above equation, we get: ； in, A subset of inputs to a digital twin system; If the above equation always holds true, then... The convolution kernels of other elements are redundant and are removed, i.e., pruning is performed.

5. The method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system according to claim 1, characterized in that: In step S22, the quantization process in the DNN model reduction method is used to quantize the weight parameters in the digital twin model of the forging hydraulic press hydraulic system. Specifically: This makes the subset of variables input to the digital twin system First quantization parameter The calculation formula is as follows: ； in, This is a rounding operation; This refers to the layer number of the DNN neural network model; A subset of variables that are input to a digital twin system; At the same time, according to the preset bias parameters Second quantization parameter The calculation formula is as follows: ； At this point, the quantization set can be determined. Therefore Each item in the list can be quantified as: ； in, For the first Layer quantization weight output; For the first The weight set of the layer; and For quantized sets The first and second elements in the array are adjacent to each other; This function returns an integer variable; i and j are the first and second position numbers of the quantization weights. To perform the absolute value operation; For the The layer, the grouping of weight parameters, is as follows: ； in, For groups that need to be quantified; Groups that need further training; To simplify the expression, a binary matrix is ​​introduced here. : ； in, It is a binary matrix; Therefore The quantization process of the layer is as follows: ； in, For the first The quantization results of the layer; For the first The model loss of the layer; For the first The regularization term for the layer; The regularization coefficient is used. This represents the total number of layers in the weight set. The above formula can be further simplified as follows: ； Among them, quantization set and binary matrix Given that the learning rate is set to... Using stochastic gradient descent, the above equation can be transformed into: ； in, The learning rate; The quantification process is now complete.

6. The method for predicting hydraulic system faults in a forging hydraulic press based on a digital twin system according to claim 1, characterized in that: Step S3, which involves implementing online fault diagnosis and prediction of the hydraulic system of the forging hydraulic press, specifically includes: S31: Determine the various fault states of the hydraulic system of the forging hydraulic press in the physical space, simulate different fault states, and collect data parameters under different fault states; S32: Use sensors to collect data on the forging hydraulic press in the physical space, including raw data on flow rate, pressure, temperature at key locations in the system, as well as the displacement of the forging hydraulic press, and classify the raw data according to the fault type; S33: Collect data on the types of faults that may occur during the operation of the forging hydraulic press in the physical space. Data preprocessing is required to remove noise, outliers and missing values ​​from the raw data to obtain the processed data. S34: Import the processed data into the parametric model of the hydraulic system of the forging hydraulic press to realize online diagnosis and prediction of hydraulic system faults of the forging hydraulic press.

Images