Engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning

Abstract

The invention discloses an engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning. Cross-model fault diagnosis is realized by using a small amount of label-free finished machine data. The method comprises the following steps: (1) constructing a digital twinborn model covering eight sub-modules on the basis of a physical entity of a testing machine in combination with dynamics and thermodynamics principles; (2) collecting testing machine data under different working conditions by using a twinborn model and a sensor technology, and establishing a twinborn and real fault database; (3) designing a step-by-step denoising module, pre-training a model on real data, and gradually calibrating and aligning to real data distribution, so as to solve the problem of data difference; (4) proposing a parallel branch feature extraction and bidirectional fusion strategy, extracting global and local features through a double-branch parallel mode, and realizing feature interaction between a source domain and a target domain; and (5) designing a multi-source domain adaptive framework, and learning different working condition characteristics through a domain discriminator and class level distribution alignment loss, thereby improving the cross-model fault diagnosis capability.

Description

Technical Field

[0001] The present invention relates to digital twin technology, and in particular to an engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning. Background Art

[0002] As a power device widely used in the fields of transportation and industry, engines are easily affected by various factors during long-term operation, leading to various faults. In order to ensure the normal operation of the engine and improve the reliability of the system, timely and accurate diagnosis of faults becomes crucial. The combination of digital twins and unsupervised domain adaptive learning technology provides a new approach to engine fault diagnosis.

[0003] Digital twin is a technology that models, simulates and monitors physical systems through digital means. It maps the real-world physical system into the digital space to achieve real-time monitoring and prediction of the system status. In the field of engines, digital twin technology can build a digital copy of the engine by simulating the operating status of the engine, capturing key parameters and behaviors. This makes fault diagnosis possible in a digital environment and provides a basis for unsupervised domain adaptive learning.

[0004] Unsupervised domain adaptive learning is a deep learning method that aims to achieve knowledge transfer between different domains without relying on domain labels. In the field of cross-model engine fault diagnosis, engines may operate under variable working conditions, loads, and environmental factors, resulting in significant differences in data distribution. Through unsupervised domain adaptive learning, this method learns and establishes the feature mapping relationship between the source domain test machine and the target domain finished machine to adapt to the changes in data distribution under different models, thereby enhancing the generalization ability and reliability of fault diagnosis.

[0005] The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning first uses digital twin technology to establish a digital model of the test engine and monitor the engine operating status in real time. Using unsupervised domain adaptive learning, this method adjusts the data distribution of the source domain and the target domain to eliminate the data differences caused by different operating conditions. After that, supervised learning is used to train the adjusted data to construct an efficient fault diagnosis model. In practical applications, the actual engine operating data is monitored in real time and input into the established model to achieve accurate diagnosis of potential engine faults.

[0006] The core advantage of this method is that it can make full use of the high simulation and real-time monitoring capabilities provided by digital twin technology. At the same time, it effectively solves the problem of data distribution differences under different models through unsupervised domain adaptive learning. This not only makes the engine fault diagnosis method more flexible and reliable, but also can adapt to various complex and changeable working conditions, providing a strong guarantee for the reliability and performance of the engine. Summary of the invention

[0007] Purpose of the invention: The present invention aims to solve the challenges of fault diagnosis under different working conditions of multi-source domain adaptive networks across models in the prior art, and to provide an innovative engine fault diagnosis method based on digital twins and unsupervised domain adaptive learning. The core of the present invention is to use digital twin technology to build a digital model of the test machine, and to reduce the difference between the twin data output by the twin model and the actual operating data through a step-by-step diffusion model method, thereby ensuring the high reliability of the twin data. In addition, the present invention uses the principle of unsupervised domain adaptive learning to learn knowledge from the source domain and adapt to the feature distribution of the target domain, thereby optimizing the performance of the model in the target domain, especially in the absence of labeled data. This method provides an effective solution for improving the performance of fault diagnosis models in environments where labeled data is scarce, enhances the generalization ability and adaptability of the model, and provides strong technical support for fault prediction and health management in industrial production.

[0008] Technical solution: The engine fault diagnosis method and device based on digital twin and unsupervised domain adaptive learning described in the present invention include:

[0009] (1) Obtain a data set with several engine parameters and corresponding fault category labels, and simulate engine faults including piston ring wear, injector clogging, and misfire. Use simulation tools to build a simulation model of the engine. Use the simulation model to simulate different types of faults and obtain a large amount of twin fault data output by the model.

[0010] (2) Using the digital twin model constructed in step (1), we simulated the normal operation of the test machine under various working conditions. At the same time, the deployed sensors collected data from the test machine in complex and variable actual working conditions to capture real-time information during its operation. Based on this data, we established a twin fault database for the test machine. And the corresponding real fault database

[0011] (3) A step-by-step diffusion model is established, and diffusion technology is used to simulate the distribution transformation process from simulation data to real data, decomposing large domain gaps into small domain gaps. Specifically, the module includes the diffusion model, which defines the same diffusion operator q and inverse operator p. Unlike the standard diffusion model that separates the diffusion process and the inverse process, the diffusion operator and the inverse operator are used simultaneously in each iterative step of the transition process to bridge the distributions of two different regions. A single step of the diffusion (or inverse) operator only slightly diffuses the distribution. The diffusion operator q is used to transform the source feature F s A total of K steps of diffusion are performed to obtain features at different diffusion levels. For each diffusion source feature map in the diffusion process, a reverse operator p with the corresponding number of steps is then applied to transform the diffusion distribution into a specific distribution.

[0012] (4) An MDA model framework was established, and unsupervised learning technology was used to learn fault features from the test machine data and transfer them to the fault diagnosis task of the finished machine. This knowledge transfer strategy helps to achieve more accurate fault prediction in the target domain and solve the problem of how to effectively perform fault diagnosis. By extracting global and local features in the data through parallel branches and learning feature knowledge from different fields (test machines and finished machines), knowledge transfer can be achieved on the category distribution.

[0013] The present invention also discloses an engine fault diagnosis device based on digital twins and unsupervised domain adaptive learning, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor is used to implement the above method when executing the computer program.

[0014] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention provides an engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning. First, the method combines digital twin technology, step-by-step diffusion model and multi-source unsupervised domain adaptation; accordingly, sensor technology and the established test machine simulation model are used to construct the test machine real fault data set and twin fault data set according to different types of faults; then, the step-by-step diffusion model is used to compensate the difference between the obtained twin data features and the real fault data; finally, by using the compensated twin data and real data, the features of the two data are obtained by using the feature extraction module, and then the knowledge on the test machine source domain is learned by the established multi-source unsupervised domain adaptive module to adapt to the feature distribution of the target domain of the finished machine, so that the model performs better in the target domain, especially on unlabeled data. This provides an effective method for improving the performance of cross-model fault diagnosis models in the absence of a large amount of labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the present invention;

[0016] Figure 2 It is a model framework diagram of the present invention;

[0017] Figure 3 It is a structural diagram of the step-by-step diffusion model of the present invention;

[0018] Figure 4 is the distribution difference map before data correction;

[0019] Figure 5 It is a distribution difference diagram after data correction. DETAILED DESCRIPTION

[0020] In order to more clearly understand the purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, this embodiment discloses an engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning, including:

[0022] (1) Three engine faults were simulated, including piston ring wear, injector nozzle blockage and misfire. According to the volumetric method, the injection pump, engine cylinder and exhaust pipe modules were established through simulation tools. The fault data generated by the simulated fault conditions was realized by modifying the parameters in the simulation model. The mathematical models of each module are briefly described.

[0023] (1-1) Fuel injection pump submodule

[0024] The fuel injection pump sub-model obtains the fuel injection amount per cylinder cycle of the fuel injection pump according to the engine speed and the fuel injection pump rack position output by the speed regulator. The relationship between the fuel injection pump single cylinder cycle fuel supply amount and the fuel injection pump rack position and engine speed is obtained according to the speed characteristic curve and load characteristic curve of the fuel pump:

[0025] g c =f(F r ,n s )

[0026] where g c It is the fuel supply of the fuel pump per cylinder cycle kg / cyl, which refers to the amount of fuel injected into each cylinder during a cycle, in kg. The speed and load of the fuel pump are made into a two-dimensional difference table, which is directly interpolated during simulation. In Simulink, a two-dimensional interpolation model Lookuptable (2-D) is used to represent this fuel pump sub-model.

[0027] (1-2) Engine cylinder submodule

[0028] The cylinder charging efficiency is only a function of the engine speed, so its change with speed can be approximately expressed by a semi-empirical formula:

[0029]

[0030] Where n s is the engine speed, in r / min; is the engine speed for determining the optimal timing, in r / min; η v is the cylinder charging efficiency; is the cylinder charging efficiency with the best timing; B is a constant, which is 0.014 for a four-stroke engine. For a four-stroke engine, the air flow through the intake valve can be divided into two parts: the intake volume and the scavenging volume, and the charge mass flow q is defined. 1 (kg / s), sweeping air mass flow rate q 2 (kg / s) and total air mass flow rate q a The concept of (kg / s):

[0031] q a =q 1 +q 2

[0032] The quality of the inflation flow can be obtained by the following formula:

[0033]

[0034] where p in Air pressure in the intake pipe, Pa; T in is the average temperature of the air in the intake pipe, K; Ri is the gas constant on the intake side, which is 286.846 J / kg·K; V is the engine cylinder displacement, m 3 ; S is the stroke coefficient, for a four-stroke engine, it is 2. The scavenging mass flow rate can be calculated by the following formula:

[0035]

[0036] When p out / p in >0.98,

[0037] When p out / p in When ≤0.98,

[0038] S q is the average value of the flow cross section during scavenging, m 3 ;k iThe constant entropy index of the intake side gas is taken as 1.4; the average exhaust temperature of the engine is calculated according to the first law of thermodynamics of the cylinder process:

[0039] q f H u -L w -N e =(q a +q f )C Pe T out -q a C Pi T in

[0040] where q f The amount of fuel injected into the cylinder per unit time, in kg / s; H u is the lower calorific value of the fuel, in J / kg; L w is the heat taken away by cooling water per unit time, in J / s; N e is the effective power of the engine, in W; C Pe is the constant pressure specific heat capacity of the intake working fluid, in J / (kg·K); C Pi is the constant pressure specific heat capacity of the exhaust medium, in J / (kg·K); T out is the average exhaust temperature in K.

[0041] Engine effective power:

[0042] N e =n e q f H u

[0043] The heat removed by cooling water is:

[0044] L w =δ w q f H u

[0045] where n e is the estimated effective efficiency of the engine; δ w is the cooling loss percentage. Arranging the above three equations, we can get:

[0046]

[0047] Where T cool is the outlet temperature of the intercooler, in K. For the engine, The value of usually varies between 0.94 and 1.0, so it can be considered to be approximately equal to 1, so the above can be simplified to:

[0048]

[0049] Engine effective thermal efficiency n e Mainly affected by a / q f The influence of cooling loss δ w Mainly affected by a / q f and engine speed n s Assume the engine exhaust temperature is K T =(H u / C Pe )(1-n e -δ w )=f(q a / q f ,n s ), and finally the average exhaust temperature of the engine is:

[0050]

[0051] (1-3) Exhaust pipe module

[0052] Consider the volume of the exhaust manifold and each exhaust branch pipe as one volume, let the volume be V e (m 3 ), the exhaust gas mass in the volume is Q e (kg), the mass flow rate through the engine exhaust valve is q Exh (kg / s), equal to the total air mass flow rate q a (kg / s), fuel mass flow rate q f The sum of (kg / s) is the mass flow rate q flowing through the supercharger turbine tur (kg / s). In the non-steady state, q Exh With q t The difference makes the mass Q in the intake pipe volume e Changes occur, namely:

[0053]

[0054] The average exhaust temperature from the cylinder into the exhaust pipe is T Exh (K), and then the air is evenly mixed in the exhaust pipe, flows out of the exhaust pipe, and enters the turbine. The temperature of the gas flowing into the turbine (i.e. the average temperature of the air in the exhaust pipe) is T tur (K), from the energy equation and the uniform mixing assumption, we can get:

[0055]

[0056] From the above three equations, we can get the temperature T of the exhaust pipe: Aout (K):

[0057]

[0058] Changes in mass and temperature in the volume cause changes in pressure. According to the thermodynamic state equation, the exhaust pressure is:

[0059]

[0060] Where T Aout is the average temperature of the exhaust pipe, in K; p Exh is the exhaust pressure of the exhaust pipe, in Pa; R e is the gas constant on the exhaust side, which is 286.354 J / kg·K; k e is the constant entropy index of the exhaust gas, which is taken as 1.33.

[0061] (2) Use convolution branches to extract local feature details, including a single one-dimensional deep convolution layer, a normalization layer, a ReLU activation function layer, and a point-by-point convolution. The Transformer branch is designed to capture the long-range feature dependencies that constitute the global representation. The branch consists of modules that use relative position encoding. A bidirectional feature fusion strategy is introduced between the convolution branch and the Transformer branch. The information in the convolution branch flows to the other branch through interaction, enhancing the global feature modeling capability. At the same time, the global features are enabled to flow from the transformer branch to the convolution branch, thereby enhancing the local representation capability.

[0062] (3) Establish a step-by-step diffusion model based on the denoising diffusion probability model (i.e., DDPM). First, let’s review the DDPM. A standard DDPM usually contains a diffusion operator q and an inverse operator p. Specifically, the diffusion operator q is used to diffuse the data distribution by adding noise, while the inverse operator p is used to concentrate the data distribution by removing noise. In DDPM, applying the diffusion operator q (or the inverse operator p) at a time can lead to one step of diffusion, which will slightly change the distribution. Given a feature map F under distribution A, the diffusion operator q is iteratively used to gradually diffuse the K-step data F, and the diffusion process is obtained, i.e., {F, F 2 , …, F K}. Mathematically, the diffusion operator q can be defined as:

[0063]

[0064] Where k∈{1,2,......,K}. k-1 is the feature map that has been diffused k-1 steps, which is the input of q, F k is the output of the diffusion operator q. 1 ,β 2 ,...,β k}\ is a fixed-variance perturbator that controls the scale of the injected perturbation. When k is large enough, F k can be approximately regarded as Gaussian-distributed data. Different from the diffusion operator q defined with fixed parameters, the reverse operator p is defined with learnable parameters θ. The goal of the standard DDPM is to train the reverse operator p to gradually convert the Gaussian noise F k into data of the desired distribution, forming a reverse process {F k , F k-1 , F k-2 ,..., F}. The inverse operator p is expressed as:

[0065] p θ (F k-1 |F k ) = N(F k-1 ; μ θ (F k , k), β k I)

[0066] where F k is the input feature map of the reverse operator p, and F k-1 is the output. μ θ (F k , k) is the mean of the output feature distribution, which can be predicted by the reverse operator p as:

[0067]

[0068] where f θ is the learnable neural network in the reverse operator p, α k = 1 - β k , In each reverse (concentration) step, the reverse operator p uses its neural network f θ to first predict the mean μ θ (F k , k) of the output feature map, and then uses the predicted mean to formulate the output distribution through equation n. Where the network f θ is shared across all reverse steps.

[0069] To train the inverse operator p to convert arbitrary random Gaussian noise into the features of distribution A, first diffuse the features F under distribution A K steps using the diffusion operator q to obtain its diffusion process. Then all the data in the diffusion process are used as supervision to guide the learning of the reverse operator. The reverse learning loss is defined as:

[0070]

[0071] where p(F K) is the initial Gaussian distribution. After reverse learning, given a random Gaussian noise feature map, the reverse operator p can gradually transform the feature map into the desired distribution A.

[0072] Inspired by the above-mentioned standard DDPM, the SBSD module is designed to utilize the diffusion technique to simulate the distribution transformation process from simulation data to real data, decomposing large domain gaps into small domain gaps. Specifically, the SBSD module contains two basic units: the diffusion operator q and the inverse operator p, which are defined the same as the diffusion model. Unlike the standard diffusion model that separates the diffusion process and the inverse process, the diffusion operator and the inverse operator are used simultaneously in each iterative step of the transition process to bridge the distributions of two different regions. Note that a single step of the diffusion (or inverse) operator only slightly diffuses (or concentrates) the distribution. As Figure 3 As shown, the diffusion operator q is used to transform the source feature F S A total of K steps of diffusion are performed to obtain features at different diffusion levels. For each diffusion source feature map in the diffusion process, a reverse operator p with the corresponding number of steps is then applied to convert (centralize) the diffusion distribution to a specific distribution. A training method is designed to train the reverse operator p in the SBSD module so that it can convert (centralize) the distribution to the real data instead of returning to the twin data. Therefore, applying a single reverse operator p at a time can slightly centralize the scattered distribution, encouraging the concentrated feature distribution to gradually approach the real data. More specifically, for those that have been diffused for k (k∈{1,2,......,K}) steps, the reverse operator p is used to centralize the diffusion features with the same number of steps (i.e. ), obtain the domain transition characteristics Therefore, when k is small, the SBSD module makes a small change to the feature distribution, thereby obtaining domain transition features with small differences compared to the twin data distribution (as shown in the figure). and ). However, as k increases from 1 to K, the twin data features become more and more diffuse, and at the same time, more and more target reverse operations are performed, making the output features gradually approach the real data, thus simulating the transition process Given twin data feature F S , feature map The relationship between (k∈{1,2,......,K}) and the SBSD module can be formulated as:

[0073]

[0074] in Represents processing characteristics Simulated SBSD module, Diffusion(·,k) means diffusion of k consecutive steps using diffusion operator q, and Reverse(·,k) means reversing k consecutive steps using reverse operator p.

[0075] The SBSD module is trained to give the initial ability to transform the twin features into the real data distribution. Specifically, the real data features F T Diffusion K steps, and then use the diffused real data features through the reverse learning loss L RL To train the reverse operator p), this loss function is named the target reverse loss, which can be formulated as:

[0076] L TR =L RL (F T )

[0077] Here, the reverse operator p trained only on the real data can transform (centralize) the diffuse twin data to the real data distribution. This is because the training goal of reverse learning is to convert the negative log-likelihood E[-logp θ (F)], which shows that L RL Learn the generation of data F from various distributions. Therefore, when L RL Applied to F T , the reverse operator p can learn to generate the real data distribution from the twin data distribution.

[0078] (4) Our convolutional branch is used to capture local feature information, including interaction layers, one-dimensional convolutional layers, normalization layers, ReLU layers, and point-by-point convolutional layers. The Transformer branch aims to capture the long-range feature dependencies that constitute the global representation, including relative position encoding and multi-head self-attention layers. At the same time, a bidirectional feature fusion strategy is designed to solve the problem of feature interaction between branches. Specifically, we transform the local feature G L and the global feature G G Establish a connection, where G L From the convolution branch, G G From the Transformer branch. G L As the gating information to fuse the global features, the hidden layer features of the convolution branch h = {h 0 ,h 1 ,…,h k} can be expressed by the formula:

[0079] h L (x)=(Conv point (x)+b)⊙σ(G L )

[0080] Among them, Conv point (·) is the point-by-point convolution operation, and σ is the sigmoid function. For the interaction part that fuses local information in the global context, it can be expressed as:

[0081] h G (x)=(Conv point (x)+b)⊙σ(G G )

[0082] Finally, the features of the two branches are fused, which can be expressed as:

[0083] X=Concat(G L ,G G )

[0084] (5) First, we use the adversarial training strategy to learn domain invariance, so that the domain classifier cannot distinguish whether the input features come from the source domain or the target domain. The loss function can be defined as:

[0085]

[0086] in, is the domain label, n is the number of samples, x i is the input data, G D (·) is the domain classifier, G f (·) is a feature extractor. To solve the negative transfer problem, the category distribution is aligned using each source and target domain pair. The category distribution alignment loss for each source and target domain pair is expressed as:

[0087]

[0088] Where L is the cross entropy loss, is the k-th domain classifier The predicted label of is x i In order to extract discriminative features, the loss of the classifier for each source domain and target domain pair can be defined as:

[0089]

[0090] in, and N T is the number of samples of the kth working condition in the source domain and the number of samples in the target domain, and D T is the i-th source domain dataset and target domain dataset, is the i-th sample in the k-th working condition in the source domain, is the i-th sample in the target domain.

[0091] A prediction result integration method based on feature similarity weighting is proposed. This method achieves effective fusion of prediction results by assigning differentiated weights to the common features of different source domains. Specifically, the weight w of the kth source domain is kIt is determined based on the similarity between the common features of the source domain and the target domain. For the common features of the kth source domain and the target domain, the similarity d k The calculation method is as follows:

[0092]

[0093] Next, the weight w k It can be expressed as:

[0094]

[0095] The joint loss of the proposed MDA can be expressed as:

[0096]

[0097] Among them, β 1 ,β 2 ,β 3 is a hyperparameter.

[0098] For each pair of source and target domains, we calculate the category distribution alignment loss, L DC The loss measures the distribution difference of each category in the source and target domains, and uses cross entropy loss to quantify the difference between the predicted results and the actual labels in each category. By minimizing the category distribution alignment loss, the model learns how to adjust the source domain features to make them closer to the category distribution of the target domain. The category distribution alignment module helps to improve the diagnostic performance of the model in the target domain, especially when the target domain data is scarce or the label space is incomplete. By adjusting the network parameters and minimizing the category distribution alignment loss, the model can learn discriminative features in different domains while reducing negative transfer caused by inconsistent category distribution. It is finally applied to multi-source domain fault diagnosis, especially in scenarios where the working condition types of the source and target domains may not be completely consistent under different working conditions.

[0099] The present invention also provides an engine fault diagnosis device based on digital twins and unsupervised domain adaptive learning, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor is used to implement the above method when executing the computer program.

[0100] By modifying the parameters of different modules in the twin model of the test machine, different types of faults are simulated and a large amount of twin source domain data is obtained as the output of the model, including normal, piston ring wear, injector blockage and misfire fault states under four working conditions of load 25% (S1), 50% (S2), 75% (S3) and 100% (S4). At the same time, the same fault state of the finished machine under two working conditions of load 25% (T1) and 50% (T2) is collected. Table 1 shows the data sets and task divisions of the source domain and target domain. The SBSD module used makes up for the difference between the twin data and the real data. Figure 2 , 3 The accuracy rate is selected as the evaluation index, and the comparison results with the previous advanced models are shown in Table 2.

[0101] Table 1 Migration task dataset division table

[0102] Task S→T A1 S1, S2, S3 → T1 A2 S1, S3, S4 → T2 A3 S2, S3, S4 → T3 A4 S1, S2, S4 → T4

[0103] Table 2 Classification performance of existing unsupervised domain adaptation methods on the dataset (%)

[0104] Method A1 A2 A3 A4 Avg DTLFD 91.12 93.46 95.98 94.66 93.80 MFSAN 92.15 94.86 96.76 95.94 94.92 Ours 94.56 95.22 97.79 96.45 96.00

[0105] In the constructed dataset, the method of the present invention has improved accuracy in cross-domain diagnosis of multiple working conditions, and is more advanced than previous models. The SBSD module used makes up for the difference between twin data and real data, so that the model can better adapt to the data distribution of the target domain, thereby improving the accuracy of domain adaptive diagnosis. Experimental results show that the highest accuracy can be achieved by providing data for different faults under different working conditions through digital twins and diagnosing the engine in combination with real data from other working conditions. In the data set recognition task including three faults and normal states, the engine fault diagnosis method based on digital twins and unsupervised domain adaptive learning proposed by the present invention achieved better recognition results than other methods.

[0106] The above disclosure is only a preferred embodiment of the present invention, which cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. An engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning, characterized in that: The following steps are involved: (1) Using simulation tools to establish a simulation model; based on the volumetric method, a digital twin model is constructed, including a compressor submodule, an intercooler submodule, an intake pipe submodule, a cylinder submodule, an exhaust pipe submodule, a turbine submodule, a supercharger rotor submodule, and a fuel injection pump submodule; (2) Using the twin model established in step (1) to simulate the health status information of the test machine during operation under various operating conditions, the deployed sensors are used to collect data from the test machine under diverse and complex operating conditions, and collect real-time data from the actual operation of the test machine; establishing a twin fault database and real fault database (3) Establish a step-by-step denoising probability module (StepByStepDiffusion, SBSD), through Real data pairs The module contains a diffusion operator q and a reverse operator p. The module can iteratively diffuse on the twin data and gradually change its distribution by converging on the real data distribution. The trained reverse operator p can focus the distribution on the real data distribution instead of returning the distribution of the twin data; (4) A parallel feature extraction module is established, including a convolution branch for extracting local features and a Transformer branch for extracting global feature information. In addition, a bidirectional feature fusion strategy is proposed to achieve the interaction between local features and global features between the convolution branch and the Transformer branch. (5) A multi-source domain adaptation framework (MDA) is established to achieve cross-machine fault diagnosis under variable working conditions. The framework can extract feature knowledge from multiple source domains and adapt to changes in working conditions. Through the domain discriminator, the framework learns the differences between source domain and target domain features. Using the proposed class-level distribution alignment loss, the framework effectively solves the problem of inconsistent label spaces between the source domain and the target domain caused by different working conditions, thereby reducing the impact of negative transfer. Finally, through the information fusion strategy, the framework integrates the prediction results of multiple source domains, realizes the knowledge transfer from the test machine to the finished machine, and improves the accuracy of fault diagnosis under multiple working conditions.

2. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 1 is characterized in that: In the step (1), a simulation model of the test engine is established using a simulation tool, which specifically includes the following steps: (1-1) By using the volumetric method, the entire test machine system is divided into several modules to achieve comprehensive simulation of the engine; by constructing parameter transfer between sub-modules, it is ensured that the modules of the system work together to form a unified simulation model; (1-2) Establishing various key submodules related to the set faults, including the engine compressor submodule, the intercooler submodule, the intake pipe submodule, the cylinder submodule, the exhaust pipe submodule, the turbine submodule, the supercharger rotor submodule and the injection pump submodule, which together constitute a complete engine simulation system; (1-3) The simulation process of the entire engine is constructed by transferring parameters between sub-modules. This ensures the synergy between the various subsystems, making the simulation results more accurate and reliable, and providing a powerful tool for the subsequent production of rich twin data.

3. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 2 is characterized in that: In step (2), a twin database of the test machine is constructed and the real database The specific steps include: (2-1) By modifying the parameters of the twin model of the test machine established in claim 2, different operating conditions and different fault states of the test machine are simulated, and the different operating conditions and fault states generated are marked; (2-2) Building a twin database for testing machines in represents the i-th working condition of the test machine twin data, i∈{1,2,…m}; (2-3) By deploying sensors on the test machine, adjusting the parts on the test machine to simulate different fault states under different working conditions, and annotating the collected data; (2-4) Constructing a real database of test machines in Represents the jth working condition of the real data of the testing machine, j∈{1,2,…m}.

4. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 3 is characterized in that: In step (3), an SBSD module is established to calibrate the twin model through big data, which specifically includes the following steps: (3-1) Diffusion technology is used to simulate the distribution conversion process from simulation data to real data, and the large domain gap is decomposed into small domain gaps; specifically, the module contains the same diffusion operator q and inverse operator p as defined in the diffusion model; the diffusion operator q is used to diffuse the data distribution by adding noise, and the inverse operator p is used to concentrate the data distribution by removing noise; (3-2) Using the diffusion operator q to calculate the source feature F s Perform a total of K steps of diffusion to obtain features at different diffusion levels; then apply the inverse operator p with the corresponding number of steps to transform (centralize) the diffusion distribution into the true data distribution; (3-3) At the same time, a target reverse loss is designed to train the reverse operator p in the module so that it can transform the distribution to the distribution of the real data instead of returning to the twin data; finally, the calibrated data D is obtained S , in Represents the calibrated data of the i-th working condition in the source domain.

5. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 4 is characterized in that: The feature extraction module in the multi-source domain adaptive framework established in step (4) specifically includes the following steps: (4-1) The source domain data D S and target domain data D T ,First, use the convolution downsampling module to downsample the data; (4-2) Use parallel branch structures to extract global and local features of the data respectively; among them, the convolution branch is used to capture local feature details, including deep convolution layer, normalization layer, activation function layer and point-by-point convolution layer; the Transformer branch captures global dependencies through multi-head self-attention and relative position encoding; (4-3) In order to realize the feature interaction between branches, a bidirectional feature fusion strategy is proposed; the global features of the Transformer branch are fused with the local features of the convolution branch; at the same time, the global features of the Transformer branch are fused in the process of extracting local features by the convolution branch to enhance the expression ability of local features.

6. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 5 is characterized in that: The domain adaptation module in the multi-source domain adaptation framework established in step (4) specifically includes the following steps: (5-1) In order to reduce or eliminate the difference in category distribution between different source domains and target domains, especially when the label space of datasets collected under different working conditions is inconsistent, we propose a multi-source domain adaptation framework (MDA), which includes a domain classifier and a category label classifier; for each source domain and target domain pair, the domain-level alignment loss and the category-level distribution alignment loss are calculated by comparing the data of the same category in the source domain and the target domain, which are used to measure the distribution difference of each category in the source domain and the target domain; the cross entropy loss is used to quantify the difference between the predicted result and the actual label of each category.

7. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 6 is characterized in that: The domain separation module in the multi-source domain adaptation framework established in step (4) specifically includes the following steps: (6-1) In order to improve the performance of the model in the multi-source domain adaptation scenario, the present invention proposes a weight allocation mechanism for integrating prediction results from different source domains; specifically, the method allocates differentiated weights to the features of each source domain to reflect its relative importance and contribution to the target domain, thereby achieving optimized integration of prediction results; (6-2) When implementing this method, we first evaluate the similarity of common features between the k-th source domain and the target domain. This similarity evaluation provides a basis for determining the source domain weights; specifically, the weight w of the k-th source domain k It is quantitatively measured by comparing the similarity between the common features of the kth source domain and the target domain; this similarity measurement not only considers the geometric proximity of the feature space.

8. The engine fault diagnosis method based on digital twin and unsupervised domain adaptive learning according to claim 7 is characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to implement the method according to any one of claims 1 to 7 when executing the computer program.

Images