DE112022000106T5 - Transmission fault diagnosis and signal acquisition method, apparatus and electronic device - Google Patents

Abstract

The present invention relates to a method for transmission fault diagnosis and signal detection, a device and an electronic device, the diagnostic method comprising obtaining a real signal of the transmission and an analog signal generated by the detection method, the real signal being a real torque signal, a real vibration signal and a includes real rotation signal; each of the analog signal and the real signal is input to a predetermined fault diagnosis network to obtain a fault diagnosis result of the transmission; wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting the fault diagnosis results, and the predetermined fault diagnosis network is generated based on the training of the analog signals and the real signals. The invention offers a technical solution that improves the quality of the sensor data and increases the accuracy of the transmission fault diagnosis.

Description

technical field

The invention relates to the field of algorithm design, in particular to a method for transmission fault diagnosis and signal acquisition, an apparatus and an electronic device.

State of the art

Wind turbines operate in harsh environmental conditions such as sandstorms, lightning and rainfall all year round, and the requirements for their operation and maintenance are becoming more stringent to ensure the reliability and safety of their main components. The gearbox is an important power transmission element for the wind turbines, which is time-consuming and costly to maintain and replace, while its health condition is closely related to the stable operation of the wind turbines. In recent years, researchers have extensively applied signal analysis methods to diagnose gear failures, and their methods mostly use time-domain and frequency-domain statistical features, wavelet transform, fast Fourier transform, empirical modal decomposition, etc. to extract features from gear failure signals extract and then create an error diagnosis. However, in practice, transmissions exhibit complex and variable operating conditions, different signal components, and a mix of operating modes, resulting in a non-stationary signal. Common methods of signal analysis can be hampered by spurious components and poor resolution of the characteristic signals, making it impossible to accurately determine the fault condition of a transmission. At the same time, the final sensor detection signal is mixed with a large amount of ambient noise and has a low signal-to-noise ratio. In addition, the existing sensors in wind farms can only detect characteristic signals such as wind speed, voltage and current and cannot detect or accurately measure physical quantities such as component voltage and acceleration, so they cannot accurately reflect the fault condition of the gearbox. Improving the quality of sensor data is therefore essential for accurate fault diagnosis in transmissions.

The development of digital twin technology offers an effective strategy for real-time condition assessment of transmissions in a variable environment by creating virtual simulation models that can simulate the evolution of faults during actual transmission operation and maintain a highly accurate simulation with the actual physical transmission. The integrated multiphysics and multiscale simulation of the transmission allows for accurate monitoring of the health of each individual system of the transmission and real-time output of the twin system acquisition data. The final fusion of model and sensor data effectively compensates for the deficiencies of the data captured by physical sensors. Currently, some developers have developed a model for diagnosing gearbox failures based on digital twin technology and deep neural networks for failure detection (see patent document CN113505655A ), and got some results. However, the current digital twin models are mostly based on a 3D modeling platform for user interface creation, the models are too simple and the signal data for analysis is simple, the simulated failure data is not deep enough to characterize the failure, and the quality of the Data needs to be improved, affecting the ultimate accuracy of transmission fault diagnosis. There is therefore an urgent need to further improve the quality of the sensor data and thus the accuracy of the fault diagnosis in transmissions.

content of the invention

In view of this, the present embodiment of the invention provides a transmission fault diagnosis and signal acquisition method, apparatus and electronic device, thereby improving the quality of sensor data and increasing the accuracy of transmission fault diagnosis.

According to a first aspect, the present invention provides a method for signal acquisition of a transmission, the method comprising: creating a digital twin model based on material and geometry parameters of gears in a transmission; obtaining a real torque signal of the transmission, the real torque signal comprising first signal data for correcting the parameters of a digital twin model and second signal data for calculating an analog signal; inputting the first signal data into the digital twin model to calibrate the parameters of the digital twin model; Entering the second signal data into the corrected digital twin mo dell to generate an analog signal, the analog signal including an analog vibration signal and an analog rotation signal.

Optionally, creating a digital twin model based on material and geometry parameters of the gears in a transmission includes: obtaining material parameters and geometric parameters of the active and driven wheels, respectively, in the transmission, the material parameters including at least one of the following parameters: mass parameters, mass inertia parameters, stiffness parameters and damping parameters; generating equations for the differential dynamics of the transmission based on the physical structure of the mesh between the active and driven wheels using the material parameters and the geometric parameters; calculating the simulation matrix for the output of the analog signal using the equations of the differential dynamics of the transmission; Creating a mathematical model to characterize the gear load relationship using the simulation matrix, the inertia matrix, the damping matrix and the stiffness matrix and using the mathematical model as a digital twin model, where the inertia matrix is formed based on the mass parameters and the inertia parameters, and where the damping matrix is based on the Based on the damping parameters and the geometry parameters and the stiffness matrix is formed on the basis of the stiffness parameters and the geometry parameters.

According to a second aspect, the invention provides a method for transmission fault diagnosis, comprising: obtaining a real signal of the transmission and an analog signal generated as in any of the optional embodiments of the first aspect, the real signal being a real torque signal, a real vibration signal and a real rotation signal; inputting the analog signal or the real signal into a predetermined fault diagnosis network to obtain a fault diagnosis result for the transmission; wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting the fault diagnosis results, and the predetermined fault diagnosis network is generated based on the training of the analog signals and the real signals.

Optionally, the step of extracting features of the target signal based on the first neural network using the analog signal and/or the real signal as the target signal comprises: assembling the target signal into a two-dimensional plane vector using the number of sampled time points and the number of the sensors of the target signal as two-dimensional plane axes; performing a one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution kernels of the same dimension to obtain a plurality of one-dimensional original features; separately aggregating each of the one-dimensional original features to obtain a plurality of one-dimensional time-series features; two-dimensionally merging a plurality of one-dimensional time-series features to obtain a new two-dimensional plane vector and adjusting the number and dimensionality of the one-dimensional convolution kernels; wherein the new two-dimensional plane vector is used as the two-dimensional plane vector to obtain a plurality of one-dimensional original features based on the adjusted one-dimensional convolution kernel, wherein to the step of one-dimensional convolution computation of the two-dimensional plane vector based on a plurality of one-dimensional convolution kernels of the same dimension returning a predetermined number of times to obtain a plurality of target timing features; wherein the plurality of target time features are assembled into a one-dimensional target vector and the one-dimensional target vector is fed into a predetermined fully connected layer to transform the one-dimensional target vector via the predetermined fully connected layer into a plurality of one-dimensional abstract features having the same number of Elements have like the number of time points sampled, where the number of elements of the one-dimensional abstract features is the same as the number of sensors.

Optionally, a one-dimensional convolution calculation is performed on the two-dimensional plane vector using the same one-dimensional convolution kernel based on multiple dimensions to obtain multiple one-dimensional original features, including the following steps: for each sensor element, a plurality of elements from the two-dimensional plane vector with a predetermined number of sampling times is extracted in a sliding manner and the plurality of elements having a predetermined number of sampling times are formed into a plurality of one-dimensional vectors to be convolved; where each one-dimensional vector to be convolved is successively convolved with the current one-dimensional convolution kernel and computed to get the current one-dimensional original ones to generate features; on this basis, each one-dimensional vector to be convolved is convolved in turn with the current one-dimensional convolution kernel to generate the current one-dimensional original feature, and all one-dimensional convolution kernels are traversed until the one-dimensional original feature corresponding to each one-dimensional convolution kernel is obtained .

Optionally, the step of outputting failure diagnosis results based on the second neural network includes: performing a hierarchical attention process on the plurality of one-dimensional abstract features based on a linear perceptron and merging the plurality of one-dimensional abstract features to obtain an output vector; and inputting the output vector to a predetermined classifier to cause the predetermined classifier to output the fault diagnosis results.

Optionally, based on a linear perceptron, a hierarchical attentional process is performed for the plurality of one-dimensional abstract features and the plurality of one-dimensional abstract features is merged to obtain an output vector, including the following steps: Inputting the plurality of one-dimensional abstract features into the linear perceptron , to obtain a variety of hidden traits; calculating a plurality of weight coefficients corresponding to the plurality of hidden features using a softmax function, the weight coefficients being used to characterize the hierarchical attention of the hidden features; weight and sum the plurality of hidden features using the plurality of weight coefficients to obtain the output vector.

According to a third aspect, the invention provides an apparatus for transmission fault diagnosis, comprising: a signal acquisition module for acquiring a real signal of a transmission and an analog signal generated as in any of the optional embodiments of the first aspect, the real signal being a real torque signal , a real vibration signal and a real rotation signal; a fault diagnosis module for inputting the analog signal or the real signal into a predetermined fault diagnosis network to obtain a fault diagnosis result for the transmission; wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting the fault diagnosis results, and the predetermined fault diagnosis network is generated based on the training of the analog signals and the real signals.

According to a fourth aspect, an embodiment of the invention provides an electronic device comprising: a memory and a processor, said memory and said processor being communicatively coupled, said memory storing computer instructions, said processor executes said computer instructions and thereby performs the method described in the first aspect or in one of the optional embodiments of the first aspect.

According to a fifth aspect, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium storing computer instructions, the computer instructions being used to cause the computer to thereby perform the method described in the first aspect or any of the optional embodiments of the first aspect.

The technical solution provided by the present application has the following advantages:

This patent application provides a technical solution to create a digital model of the digital twin model based on the material parameters and geometric dimensional parameters of the gears in the gearbox, after which the parameters of the digital twin model are corrected by simulation in combination with the real torque signal of the gearbox. The digital twin model corresponds more closely to the actual working conditions of the gearbox than traditional techniques to make the data quality of the analog signal more realistic. The real torque signal is thus used as input to the model, and the analog vibration signal and the analog rotation signal are output. This not only widens the range of transmission failure data, but also leads to higher analog signal quality, thereby improving the fault diagnosis accuracy of subsequent transmission failure diagnosis based on the transmission real signal data and transmission analog signal data.

In addition, the present invention uses a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism to output fault diagnosis results during fault diagnosis to further enhance useful sample features and weaken useless sample features, thereby further improving the accuracy of fault diagnosis. wherein the feature extraction step of the first neural network converts the error signals collected by the plurality of sensors from two-dimensional data to one-dimensional data, then performs one-dimensional convolution, and then iteratively splits the resulting one-dimensional features into two-dimensional data for multi-stage one-dimensional, two-dimensional and one-dimensional feature extraction. The result is a signal with more pronounced features and the same dimensionality as the original data. This makes it easier to classify the subsequent error messages and increases diagnostic accuracy. The linear perceptron-based hierarchical attention processing of feature vectors increases the non-linearity of the neural network to better account for the degradation characteristics of gears, making the fault diagnosis results closer to the real situation and improving the accuracy of gear fault diagnosis.

character list

• 1 eine schematische Darstellung der Schritte eines Verfahrens zur Signalerfassung eines Getriebes in einer Ausführungsform der vorliegenden Erfindung;
• 2 ein Flussblockdiagramm eines Verfahrens zur Signalerfassung eines Getriebes in einer Ausführungsform der vorliegenden Erfindung;
• 3 eine schematische Darstellung eines typischen geraden zylindrischen Zahnradpaares, das nach dem Stand der Technik ineinandergreift;
• 4 eine schematische Darstellung der Schritte eines Verfahrens zur Getriebefehlerdiagnose in einer Ausführungsform der vorliegenden Erfindung;
• 5 ein schematisches Flussdiagramm eines Verfahrens zur Getriebefehlerdiagnose in einer Ausführungsform der vorliegenden Erfindung;
• 6 eine schematische Darstellung der Struktur des ersten neuronalen Netzes in einer Ausführungsform der vorliegenden Erfindung;
• 7 ein schematisches Flussdiagramm eines anderen Verfahrens zur Getriebefehlerdiagnose in einer Ausführungsform der vorliegenden Erfindung;
• 8 ein Diagramm der Variation der Verlustfunktion eines tiefen zeitlichen Diagnosenetzes mit der Anzahl der Iterationen in einer Ausführungsform der vorliegenden Erfindung;
• 9 eine schematische Darstellung der Struktur einer Vorrichtung zur Signalerfassung eines Getriebes in einer Ausführungsform der vorliegenden Erfindung;
• 10 eine schematische Darstellung der Struktur eines elektronischen Geräts in einer Ausführungsform der vorliegenden Erfindung;

The features and advantages of the present invention will be more clearly understood by reference to the attached figures. The accompanying figures are schematic and should not be construed as limiting the invention in any way. Show it:

• 1 a schematic representation of the steps of a method for signal detection of a transmission in an embodiment of the present invention;
• 2 a flow block diagram of a method for signal detection of a transmission in an embodiment of the present invention;
• 3 Figure 12 is a schematic representation of a typical prior art right cylindrical gear pair meshing;
• 4 a schematic representation of the steps of a method for transmission fault diagnosis in an embodiment of the present invention;
• 5 a schematic flowchart of a method for transmission fault diagnosis in an embodiment of the present invention;
• 6 a schematic representation of the structure of the first neural network in an embodiment of the present invention;
• 7 a schematic flow diagram of another method for transmission fault diagnosis in an embodiment of the present invention;
• 8th a diagram of the variation of the loss function of a deep temporal diagnostic network with the number of iterations in an embodiment of the present invention;
• 9 Fig. 12 is a schematic structure diagram of a transmission signal detection apparatus in one embodiment of the present invention;
• 10 12 is a schematic representation of the structure of an electronic device in an embodiment of the present invention;

In order to make the purpose, the technical solutions and the advantages of the embodiments of the invention more clear, the technical solutions in the embodiments of the invention are described clearly and fully in the following in conjunction with the accompanying drawings in the embodiments of the invention, it being understood that the described embodiments are part and not all of the embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work also fall within the scope of the present invention.

As in the 1 and 2 shown, a method for detecting signals from transmissions in one embodiment includes the following steps in particular:

Step S101: Creation of a digital twin model based on material and geometry parameters of the gears in a gearbox.

Step S102: Obtaining the real torque signal of the transmission, the real torque signal comprising first signal data for correcting the parameters of the digital twin model and second signal data for calculating the analog signal, and feeding the first signal data into the digital twin model for correcting the parameters of the digital twin model.

Step S103: input the second signal data into the corrected digital twin model to generate an analog signal, the analog signal including an analog vibration signal and an analog rotation signal.

In particular, in this embodiment, the signal quality of the simulated error signal is further improved, thus creating a digital twin model of the transmission based on the material parameters and geometric dimensional parameters of the gears in the transmission, including but not limited to the gear size, the gear material, the material stiffness, the Damping between the gears and the gear mass, as opposed to traditional techniques that simply use a three-dimensional model to create a digital twin model. This allows the digital twin model to simulate conditions that are closer to real-life transmission operating conditions. Traditional digital twin models, which are typically simulations of a single vibration signal. In this embodiment, the geometric and physical parameters in the digital twin model are corrected by establishing an equation between the vibration and rotation signals and the torque signal, and then first using the real torque signal based on finite element simulation software. The real torque signal is then input to the corrected digital twin model as an excitation to obtain a well-formed analog vibration signal and analog rotation signal. As a result, more types of analog signals are output from the digital twin model, and the quality of the analog signals is better, which improves the accuracy of the subsequent transmission fault diagnosis.

• Schritt 1: Ermittlung von Materialparametern und geometrischen Abmessungsparametern der aktiven und angetriebenen Räder im Getriebe, wobei die Materialparameter mindestens einen der folgenden Parameter umfassen: Masseparameter, Trägheitsparameter, Steifigkeitsparameter und Dämpfungsparameter.
• Schritt 2: Erstellen von Differentialdynamikgleichungen für das Getriebe auf der Grundlage der physikalischen Struktur des Eingriffs zwischen den Haupt- und Antriebsrädern unter Verwendung von Materialparametern und geometrischen Maßparametern.
• Schritt 3: Berechnen der Simulationsmatrix für das analoge Ausgangssignal mit Hilfe der Differentialdynamikgleichungen des Getriebes.
• Schritt 4: Unter Verwendung der Simulationsmatrix, der Massenträgheitsmatrix, der Dämpfungsmatrix und der Steifigkeitsmatrix wird ein mathematisches Modell zur Charakterisierung der Zahnradlastbeziehung erstellt und das mathematische Modell als digitales Zwillingsmodell verwendet.

wobei die Massenträgheitsmatrix auf der Grundlage von Massenparametern und Trägheitsparametern konstruiert wird, die Dämpfungsmatrix auf der Grundlage von Dämpfungsparametern und geometrischen Dimensionsparametern und die Steifigkeitsmatrix auf der Grundlage von Steifigkeitsparametern und geometrischen Dimensionsparametern.Specifically, in one embodiment, step S101 above includes the following steps in particular:

• Step 1: Determination of material parameters and geometric dimensional parameters of the active and driven wheels in the transmission, the material parameters comprising at least one of the following parameters: mass parameters, inertia parameters, stiffness parameters and damping parameters.
• Step 2: Create differential dynamics equations for the gearbox based on the physical structure of the mesh between the main and drive gears using material parameters and geometric dimensional parameters.
• Step 3: Calculate the simulation matrix for the analog output signal using the differential dynamics equations of the gearbox.
• Step 4: Using the simulation matrix, inertia matrix, damping matrix and stiffness matrix, a mathematical model is created to characterize the gear load relationship and the mathematical model is used as a digital twin model.

wherein the mass inertia matrix is constructed based on mass parameters and inertial parameters, the damping matrix based on damping parameters and geometric dimension parameters, and the stiffness matrix based on stiffness parameters and geometric dimension parameters.

• 1) Es wird angenommen, dass die Bauteile starr sind, und die elastische Verformung der Zahnräder und Seile wird ignoriert.
• 2) Es wird angenommen, dass sich die ineinandergreifenden Zahnräder nicht voneinander lösen.
• 3) Das Phänomen des umgekehrten Schocks sollte ignoriert werden.

Specifically, in this embodiment, a mathematical equation model is created based on the gear load relationship, thereby enabling creation of a digital twin model. First, material parameters such as mass parameters, inertia parameters, stiffness parameters and damping parameters are determined, followed by the geometric parameters that describe the size of the gear. The inertia matrix, damping matrix and stiffness matrix are then constructed and this implementation uses the following assumptions in building the gearbox simulation model to design the ideal state gearbox:

• 1) The components are assumed to be rigid and the elastic deformation of the gears and cables is ignored.
• 2) It is assumed that the meshed gears do not disengage from each other.
• 3) The reverse shock phenomenon should be ignored.

$$\text{D}\ddot{\text{a}}\text{mpfungsmatrix}\left[c\right]=\left[\begin{array}{cccc}{c}_{py}+c_m& c_m{R}_p& -c_m& c_m{R}_g\\ c_m{R}_p& c_m{R}_p^2& -c_m{R}_p& c_m{R}_g{R}_p\\ -c_m& -c_m{R}_p& c_m+c_{gy}& -c_m{R}_g\\ -c_m{R}_g& -c_m{R}_p{R}_g& c_m{R}_g& -c_m{R}_g^2\end{array}\right]\text{'}$$

$$\text{Steifigkeitsmatrix}\left[k\right]=\left[\begin{array}{cccc}{k}_{py}+k_m& k_m{R}_p& -k_m& k_m{R}_g\\ k_m{R}_p& k_m{R}_p^2& -k_m{R}_p& k_m{R}_g{R}_p\\ -k_m& -k_m{R}_p& k_m+k_{gy}& -k_m{R}_g\\ -k_m{R}_g& -k_m{R}_p{R}_g& k_m{R}_g& -k_m{R}_g^2\end{array}\right],$$

wobei die Massenparameter m_p und m_g sind (p und g für die aktiven bzw. angetriebenen Räder), die Trägheitsparameter I_p und I_g sind, die Steifigkeitsparameter k_m und k_py, k_gy sind, die Dämpfungsparameter C_m und c_py, c_gy sind, wobei k_m, c_m die kombinierte Steifigkeit und die kombinierte Dämpfung des Eingriffs des Zahnradpaares ist, k_py und k_gy sind die Steifigkeiten der Translationsschwingungen des angetriebenen und des aktiven Rades, c_py und c_gy die Dämpfung der Translationsschwingungen des angetriebenen und des aktiven Rades, und die geometrischen Abmessungen R_p, R_g.The matrix of parameters required to build the model is as follows. $$\text{Mass Tr}\ddot{\text{a}}\text{ghty matrix}\left[m\right]=\left[\begin{array}{cccc}{m}_p& 0& 0& 0\\ 0& I_p& 0& 0\\ 0& 0& m_G& 0\\ 0& 0& 0& I_G\end{array}\right],$$

$$\text{D}\ddot{\text{a}}\text{vaccination matrix}\left[c\right]=\left[\begin{array}{cccc}{c}_{py}+c_m& c_m{R}_p& -c_m& c_m{R}_G\\ c_m{R}_p& c_m{R}_p^2& -c_m{R}_p& c_m{R}_G{R}_p\\ -c_m& -c_m{R}_p& c_m+c_{Gy}& -c_m{R}_G\\ -c_m{R}_G& -c_m{R}_p{R}_G& c_m{R}_G& -c_m{R}_G^2\end{array}\right]\text{'}$$

$$\text{stiffness matrix}\left[k\right]=\left[\begin{array}{cccc}{k}_{py}+k_m& k_m{R}_p& -k_m& k_m{R}_G\\ k_m{R}_p& k_m{R}_p^2& -k_m{R}_p& k_m{R}_G{R}_p\\ -k_m& -k_m{R}_p& k_m+k_{Gy}& -k_m{R}_G\\ -k_m{R}_G& -k_m{R}_p{R}_G& k_m{R}_G& -k_m{R}_G^2\end{array}\right],$$

where the mass parameters are m _p and m _g (p and g for the active and driven wheels, respectively), the inertia parameters are I _p and I _g , the stiffness parameters are _km and k _py , k _gy , the damping parameters are C _m and c _py , c _gy , where km , _{cm is} the combined stiffness and damping of the meshing of the gear pair, k _py and k _gy are the translational vibration stiffnesses of the driven and active wheels, c _py and c _{gy are} the translational vibration damping of the driven and the active wheel, and the geometric dimensions R _p , R _g .

wobei F_k und F_c die elastische Eingriffskraft bzw. die viskose Eingriffskraft sind, F_p und F_g die dynamischen Eingriffskräfte der Verzahnung an den aktiven und angetriebenen Zahnrädern sind, die bei der Lösung der Gleichung eliminiert werden. θ_p und θ_g sind die Drehvariable, die zur Charakterisierung des Drehsignals verwendet werden, y_p und y_g sind die translatorische Größe, die zur Charakterisierung des Schwingungssignals verwendet wird, T_p0, T_g0 wird zur Charakterisierung des Drehmomentsignals verwendet. Das obige kinetische Modell ermöglicht die Lösung der analogen Matrix, die zur Ausgabe des analogen Signals, d. h. der Translation Array Drehung {δ} = {y_p θ_p y_g θ_g}^T verwendet wird.Then the differential dynamics equations for the gearbox are constructed according to the physical structure of the gearbox, as in 3 shown with the expressions: $$\{\begin{array}{l}\left\{\delta \right\}={\left\{y_p{\theta }_p{y}_G{\theta }_G\right\}}^T\hfill \\ \overline{y_p}=y_p+R_p{\theta }_p\hfill \\ \overline{y_G}=y_G-R_G{\theta }_G\hfill \\ f_k=k_m\left(\overline{y_p}-\overline{y_G}\right)=k_m\left(y_p+R_p{\theta }_p-y_G+R_G{\theta }_G\right)\hfill \\ f_c=c_m\left(y_p-y_G\right)=c_m\left(y_p+R_p{\theta }_p-y_G+R_G{\theta }_G\right)\hfill \\ f_p=f_k+f_c\hfill \\ f_G=-f_p\hfill \\ m_p{y}_p+c_{py}{y}_p+k_{py}{y}_p=-f_p\hfill \\ I_p{\theta }_p=-f_p{R}_p-T_p\hfill \\ \left[m\right]\left\{\delta \right\}+\left[c\right]\left\{\delta \right\}+\left[k\right]\left\{\delta \right\}=\left\{p\right\}\hfill \\ I_G{\theta }_G=-f_G{R}_G-T_G\hfill \end{array}$$

where F _k and F _c are the elastic meshing force and viscous meshing force, respectively, F _p and F _g are the dynamic meshing forces of the meshing on the active and driven gears, which are eliminated in solving the equation. θ _p and θ _g are the rotational variable used to characterize the rotational signal, y _p and y _g are the translational quantity used to characterize the vibration signal, T _p0 , T _g0 is used to characterize the torque signal. The above kinetic model enables the solution of the analog matrix used to output the analog signal, i.e. the translation array rotation {δ}={y _p θ _p y _g θ _g } ^T .

Then the real torque signals T _p0 and T _g0 , which are monitored in the physical sensors, are passed to the digital twin model, which is updated in real time with the parameters T _p and T _g of the digital twin model can be updated, and the material and geometry parameters can be corrected using finite element simulation software simulations. The updated load matrix is as follows: $$\left\{p\right\}=\left\{\begin{array}{c}0\\ -{\mathrm{<figure-callout\; id="0"\; label="T\; p"\; filenames="DE112022000106T5\_0021.png"\; state="\{\{state\}\}">T</figure-callout>}}_p\\ 0\\ -T_G\end{array}\right\}$$

The simulation model can be solved in Matlab after the parametric variables are determined, and the digital twin model based on the load relation has the form [m]{δ̈̈} + [c]{δ̇} + [k]{δ} = { p}, thereby maintaining a high degree of consistency between the digital twin model and the physical model. Through the above steps, the simulated conditions of the digital twin model are brought closer to the real conditions of transmission operation, so that the digital twin model outputs more types of simulated signals and better quality of the simulated signals, thus improving the accuracy of the subsequent fault diagnosis of the transmission.

As in the 4 and 5 shown, in one embodiment, a method for diagnosing transmission faults is provided, which in particular comprises the following steps:

Step S201: Obtaining a real signal of the transmission and an analog signal generated based on the methods in the above steps S101 to S103, the real signal including a real torque signal, a real vibration signal and a real rotation signal, the analog signal comprises an analog vibration signal and an analog rotation signal;

Step S202: Inputting the analog signal or the real signal to a predetermined fault diagnosis network to obtain a fault diagnosis result for the transmission.

The predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, the predetermined fault diagnosis network being generated based on training with simulated signals and real signals.

In particular, in this embodiment, the analog signal generated in steps S101 to S103 is used together with the real signal as input data for the predetermined fault diagnosis network, which improves the signal type and signal quality, thereby improving the accuracy of the fault diagnosis network in classifying the faulty signal. Furthermore, in this embodiment, the predetermined fault diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results. The second neural network, based on a hierarchical attention mechanism, can form multi-level attention by dividing the features extracted by the first neural network into multiple levels with a logical subtotal. Attention is drawn to the relevant parts of the scene and the unimportant parts are ignored. The attentional mechanism is essentially an efficient allocation of information processing resources. By assigning different weights to the model input features, focusing attention on information considered more critical to the present, and highlighting the impact of key features on the model, the attention mechanism can solve the information overload problem and improve model efficiency and prediction performance. This leads to more accurate fault diagnosis results based on deeper and more critical feature information.

In particular, in one embodiment, the analog and/or real signal is used as the target signal and the feature extraction of the target signal is performed based on the first neural network, which specifically comprises the following steps.

Step 5: the target signal is assembled into a two-dimensional plane vector using the number of sampled time points and the number of sensors as the two-dimensional plane axis.

Step 6: a plurality of one-dimensional original features are obtained by performing a one-dimensional convolution calculation on a two-dimensional plane vector based on a plurality of one-dimensional convolution kernels of the same dimension.

Step 7: Separately aggregate the raw one-dimensional features to obtain a plurality of one-dimensional temporal features.

Step 8: a plurality of one-dimensional temporal features are two-dimensionally assembled to obtain a new two-dimensional plane vector, and the number and dimensionality of the one-dimensional convolution kernels are adjusted.

Step 9: the new two-dimensional plane vector is used as a two-dimensional plane vector based on the adjusted one-dimensional convolution kernel and returns to step six until a predetermined number of times to return to obtain a plurality of target timing features.

Step 10: merging the plurality of target time features into a one-dimensional target vector and inputting the one-dimensional target vector into a predetermined fully connected layer to transform the one-dimensional target vector into a plurality of one-dimensional abstract features that have the same number of elements as the number of sampled points in time over the predetermined fully connected layer, where the number of elements of the one-dimensional abstract features is the same as the number of sensors.

Specifically, in this embodiment, a one-dimensional convolution method, which differs from the prior art, is used to obtain feature data with the same dimension as the input signal data and with more varied features. Like in the 6 As shown, a convolutional neural network consists of one or more convolutional layers, activation functions and combined layers as a supervised learning neural network. The mesh optimizes the objective function by the gradient descent method and is iteratively trained to update the mesh parameters. The convolutional layer and the pooling layer are the core components of the convolutional neural network for feature extraction. The input for each node in the convolutional layer is only the neuron in the upper layer of the neural network within the local domain, convolutional layers take them to a deeper level of analysis to get features with a higher level of abstraction. The pooling layer often takes the maximum of these windows as output, then slides on and processes each depth slice of the input data body individually, reducing the mesh parameters.

1. 1. gleitende Extraktion einer Vielzahl von Sätzen von Elementen einer vorbestimmten Anzahl von Abtastzeitpunkten aus einem zweidimensionalen Ebenenvektor für jedes Sensorelement, und Bilden einer Vielzahl von eindimensionalen Vektoren, die aus der Vielzahl von Sätzen von Elementen der vorbestimmten Anzahl von Abtastzeitpunkten gefaltet werden sollen.
2. 2. Falten jedes eindimensionalen Vektors, der mit dem aktuellen eindimensionalen Faltungskern gefaltet werden soll, um die aktuellen eindimensionalen Rohmerkmale zu erzeugen.
3. 3. auf der Grundlage von Schritt 2 alle 1D-Faltungskerne durchlaufen, bis die 1D-Rohmerkmale, die jedem 1D-Faltungskern entsprechen, erhalten werden.

concretely, the target signal is composed into a two-dimensional plane vector using the number of sampled time points and the number of sensors as a two-dimensional plane axis. That is, the signal data originally input to the first neural network is two-dimensional planar data with the sampling time as the y-axis and the serial number of the signal sensor as the x-axis. The first extracted signal features are then obtained by a one-dimensional convolution calculation of the two-dimensional planar data based on several one-dimensional convolution kernels of the same size. Compared to the existing technology that uses a one-dimensional convolution kernel to directly convolve the one-dimensional input signals, the present embodiment combines multiple signals into two-dimensional data and then performs one-dimensional convolution, so that different signal sequences are mixed and the obtained feature information is more comprehensive and representative are. In addition, compared to the prior art, one-dimensional convolution kernels are used directly on two-dimensional, planar data for row-by-row or column-by-column calculation of convolutions. The present embodiment differs from it by the following specific operational steps:

1. 1. sliding extraction of a plurality of sets of elements of a predetermined number of sampling times from a two-dimensional plane vector for each sensor element, and forming a plurality of one-dimensional vectors to be convolved from the plurality of sets of elements of the predetermined number of sampling times.
2. 2. Convolve each one-dimensional vector to be convolved with the current one-dimensional convolution kernel to produce the current raw one-dimensional features.
3. 3. Based on step 2, iterate through all 1D convolution kernels until the raw 1D features corresponding to each 1D convolution kernel are obtained.

Concretely, for example, the x-axis of the 2D planar data represents the sensor number and the y-axis represents the number of sampling times, and a predetermined number of consecutive sampling times (e.g. two) is determined on the y-axis, and then the data for all sensors corresponding to the sampling instants are fetched via slider (e.g., starting with the first row, two data rows on the 2D planar data are taken at a time, i.e. two sampling instants until all data rows are traversed). The data acquired at the specified sampling times are then concatenated with the sensor serial number to form a 1D sequence (e.g. after the acquisition of two data rows each, the two rows are concatenated to form a 1D sequence). Then the one-dimensional sequence data obtained from each stitching is convolved once until the entire one-dimensional sequence data is convolved, and the results of multiple calculations are obtained, and then the results of each calculation are assembled into a one-dimensional feature vector, that is, the original one-dimensional features will get. Finally, based on step 2, all one-dimensional convolution kernels are traversed until the original one-dimensional features corresponding to each one-dimensional convolution kernel are obtained.

wobei i=1,2,L,n-t+1, ⊕ die Kombination der verschiedenen Zeitfenster darstellt. Der Vektor x_i:i+t-1 wird mit dem j-ten Faltungskern k_j ∈ ℝ^tm bearbeitet und das im i-ten Zeitschritt extrahierte Merkmal wird wie folgt dargestellt: $$z_j^i=ƒ\left({\mathbf{k}}_j\cdot {\mathbf{x}}_{i:i+t-1}+b_j\right)$$

wobei b_j und ƒ der Bias-Term bzw. die Aktivierungsfunktion sind. Die eindimensionalen Rohmerkmale z_j, die aus dem Eingangsvektor extrahiert werden, erhält man, wenn sich der Faltungskernel im Eingangsvektor von oben nach unten bewegt. $${\mathbf{z}}_j=\lfloor z_j^1{z}_j^2\text{L }{z}_j^{n-t-1}\rfloor$$

i.e. given two-dimensional sensor data X = [x ₁ x ₂ x _i x _n ] ^T ∈R ^n×m , where n is the length of the sensor time series, then the data corresponding to the ith time step x _i = [ x _1,i x _2,i L x _m,i ] ^T , where m is the number of sensors. In addition, let the vector X _i:i+t-1 represent the ith input sample as follows: $${\mathbf{x}}_{i:i+t-1}={\mathbf{x}}_i\oplus {\mathbf{x}}_{i+1}\oplus \text{L}\oplus {\mathbf{x}}_{i+t-1}$$

where i=1,2,L,n-t+1, ⊕ represents the combination of the different time windows. The vector x _i:i+t-1 is processed with the j-th convolution kernel k _j ∈ ℝ ^tm and the feature extracted in the i-th time step is represented as follows: $${\mathrm{e.g}}_j^i=ƒ\left({\mathbf{k}}_j\cdot {\mathbf{x}}_{i:i+t-1}+b_j\right)$$

where b _j and ƒ are the bias term and the activation function, respectively. The one-dimensional raw features z _j extracted from the input vector are obtained when the convolution kernel moves from top to bottom in the input vector. $${\mathbf{e.g}}_j=\lfloor {\mathrm{e.g}}_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="DE112022000106T5\_0018.png,DE112022000106T5\_0019.png"\; state="\{\{state\}\}">j</figure-callout>}}^1{\mathrm{e.g}}_j^2\text{L}{\mathrm{e.g}}_j^{n-t-1}\rfloor$$

Because a single convolution kernel can only extract one feature map from the input vector, multiple convolution kernels are used to operate on the input vector so that a more complete error feature is extracted.

wobei das Zeitmerkmal das Ergebnis einer Poolingoperation ist, pool () die maximale Pooling-Funktion, p die Pooling-Größe und s die Schrittweite ist.Then, in the seventh step, the multiple raw one-dimensional features output from the convolution layer are downsampled using the pooling layer to obtain multiple one-dimensional temporal features. On the one hand, the pooling layer extracts the most important parts of each feature mapping, on the other hand, this operation significantly reduces the number of feature dimensions, making it ideal for processing high-dimensional data. The maximum pooling operation is expressed as follows: $${\mathbf{y}}_j=\text{pool}\left({\mathbf{e.g}}_j,p,s\right)$$

where the time property is the result of a pooling operation, pool () is the maximum pooling function, p is the pooling size, and s is the increment.

In this embodiment, to further improve the extracted features to make them more complete and representative, the obtained multiple one-dimensional temporal features are spliced again in two dimensions to obtain a new two-dimensional plane vector, and the number and dimensionality of the original one-dimensional convolution kernel is adjusted. The new two-dimensional plane vector is then used as the two-dimensional plane vector, and steps six through seven are repeated a specified number of times using the adjusted one-dimensional convolution kernel. In this embodiment, three features are extracted, resulting in multiple target time features with even better feature performance.

Finally, the multiple target timing features are assembled first and last into a one-dimensional target vector, and then the one-dimensional target vector is fed into a predetermined fully connected layer to transform it into multiple one-dimensional abstract features that have the same number of elements as the number of fully sampled instants of the signal, and where the number of elements of each one-dimensional abstract feature is the same as the number of sensors by increasing the dimensionality and the number of one-dimensional target vectors over the predetermined fully connected layer is adjusted. This ensures that the signal input data corresponds exactly to the dimension of the feature data and avoids the problem of lower accuracy due to the smaller amount of data in feature extraction.

• Schritt 11: Separate Dimensionserweiterung der mehreren eindimensionalen abstrakten Merkmale.
• Schritt 12: die erweiterte Vielzahl eindimensionaler abstrakter Merkmale wird für hierarchische Aufmerksamkeit auf der Grundlage eines linearen Perzeptrons verarbeitet, und die Vielzahl eindimensionaler abstrakter Merkmale wird fusioniert, um einen Ausgangsvektor zu erhalten.
• Schritt 13: Der Ausgangsvektor wird in den vorbestimmten Klassifikator eingegeben, um den vorbestimmten Klassifikator zu veranlassen, ein Fehlerdiagnoseergebnis auszugeben.

In one embodiment, the output of a fault diagnosis result based on the second neural network includes the following steps in particular:

• Step 11: Separate dimensional expansion of the multiple one-dimensional abstract features.
• Step 12: the extended plurality of one-dimensional abstract features are processed for hierarchical attention based on a linear perceptron, and the plurality of one-dimensional abstract features are fused to obtain an output vector.
• Step 13: The output vector is input to the predetermined classifier to cause the predetermined classifier to output a failure diagnosis result.

Specifically, in this embodiment, the hierarchical attention mechanism is used to fuse high-level feature information (i.e., one-dimensional abstract features) from the output of the first neural network. Before that, the one-dimensional abstract features are first dimensionally expanded. Attention networks can dimensionally expand the input sequence in three ways to further improve feature accuracy. The first method repeats the first moment of the data in the time dimension a predetermined number of times. In a complex physical system, the value of the first moment can greatly affect the direction and course of the physical system, so emphasizing the importance of the first moment after dimensional expansion can lead to more accurate one-dimensional abstract features. The two remaining treatments are subtracted and multiplied from the first moment of the data at each time point to view the degradation information in multiple dimensions and help the attention network learn degradation features more comprehensively and richly.

The hidden states of the fault features are then computed for each moment based on a linear perceptron to increase the non-linearity of the network and better match the degraded features of the transmission. The hidden states are then normalized and the normalized input sequence is then multiplied by the weights to achieve feature fusion. Finally, the feature vectors output by the hierarchical attention network are fed into a predefined classifier (e.g. softmax classifier, logistic classifier) in order to achieve a fault diagnosis of the transmission.

• Schritt 14: Mehrere eindimensionale abstrakte Merkmale werden nacheinander in das lineare Perzeptron eingegeben, um mehrere versteckte Merkmale zu erhalten.
• Schritt 15: Eine Vielzahl von Gewichtskoeffizienten, die der Vielzahl von versteckten Merkmalen entsprechen, werden von der Softmax-Funktion berechnet, und die Gewichtskoeffizienten werden verwendet, um die hierarchische Aufmerksamkeit der versteckten Merkmale zu charakterisieren.
• Schritt 16: Der Ausgangsvektor wird durch Gewichtung und Summierung einer Vielzahl von versteckten Merkmalen unter Verwendung einer Vielzahl von Gewichtungskoeffizienten erhalten.

Specifically, in one embodiment, step 12 above specifically includes the following steps:

• Step 14: Multiple one-dimensional abstract features are sequentially input to the linear perceptron to obtain multiple hidden features.
• Step 15: A plurality of weight coefficients corresponding to the plurality of hidden features are computed by the softmax function and the weight coefficients are used to characterize the hidden features' hierarchical attention.
• Step 16: The output vector is obtained by weighting and summing a plurality of hidden features using a plurality of weighting coefficients.

In particular, it is assumed that the features of the second input of the neural network are represented as H=[h ₁ h ₂ L h _i L h _n ], where b _i represents the one-dimensional abstract feature at the ith time step. First, b _i is input to a linear perceptron to learn the hidden feature expression u _i , which is expressed as: $${\mathbf{and}}_i=\text{tanh}\left({\mathbf{wh}}_i+\mathbf{b}\right)$$

Here, W and b denote the weights or distortions of the linear perceptron.

wobei u_s ein zufällig initialisierter Vektor ist, der kontinuierlich mit der Anzahl der Iterationen aktualisiert wird. Die endgültige Gewichtung der Fehlerinformationen in jedem Zeitschritt mit Hilfe der Gewichtungskoeffizienten ergibt den Ausgangsvektor s für das zweite neuronale Netz: $$\text{s}={\displaystyle \sum _i{\lambda }_i{\mathbf{b}}_i}$$

The weight coefficient is then calculated for each time step, with a larger coefficient indicating that the time step contains more error information, and the weight coefficient is calculated by the softmax function, expressed as $${\lambda }_i=\frac{\text{ex}\left({\mathbf{and}}_i^{\tau }{\mathbf{and}}_s\right)}{{\displaystyle \sum _i\text{ex}\left({\mathbf{and}}_i^{\tau }{\mathbf{and}}_s\right)}}$$

where u _s is a randomly initialized vector that is continuously updated with the number of iterations. The final weighting of the error information in each time step using the weighting coefficients gives the output vector s for the second neural network: $$\text{s}={\displaystyle \sum _i{\lambda }_i{\mathbf{b}}_i}$$

The above steps result in an accurate calculation of an output vector containing more complete and accurate features for error signal classification.

In this embodiment, the parameters of each layer of the deep temporal diagnostic network model composed of the first neural network and the second neural network are shown in Table 1. Table 1 Description of the structure of the temporal depth diagnosis network model network structure shift number Description details one-dimensional convolution networks layer1 one-dimensional convolution one-dimensional pooling number of filters=32 convolution kernels=2 * m step length=1 padding=1 activation functions=ReLU layer2 one-dimensional convolution one-dimensional pooling number of filters=64 convolution kernels=2 step length=1 padding=1 activation functions=ReLU layer3 one-dimensional convolution one-dimensional pooling number of filters=128 convolution kernels=2 step length=1 padding=1 activation functions=ReLU layer4 full connection Output =m*n activation functions =ReLU hierarchical attention mechanism layer5 attention weights Output =32 activation functions =ReLU layer6 weighting Input = number of features * window size Output = eigenvector classifier layer7 full connection Output=4 Activation Functions=ReLU layer8 classification Softmax function one-dimensional convolution networks layer1 one-dimensional convolution one-dimensional pooling number of filters=32 convolution kernels=2 * m step length=1 padding=1 activation functions=ReLU layer2 one-dimensional convolution one-dimensional pooling number of filters=64 convolution kernels=2 step length=1 padding=1 activation functions=ReLU layer3 one-dimensional convolution one-dimensional pooling number of filters=128 convolution kernels=2 step length=1 padding=1 activation functions=ReLU layer4 full connection Output =m*n activation functions =ReLU hierarchical attention mechanism layer5 attention weights Output =32 activation functions =ReLU layer6 weighting Input = number of features * window size Output = eigenvector classifier layer7 full connection Output=4 Activation Functions=ReLU layer8 classification Softmax function

• Um den Betriebszustand des Getriebes in Echtzeit zu überwachen, wird das Getriebe zunächst in ein Prüfsystem eingesetzt. Das System besteht aus einer Getriebediagnoseplattform, einem Schwingungssensor, einem Rotationssensor, einem Drehmomentsensor, einem Signalkabel, einem Datenerfassungsgerät und einem Computer. Das Prüfsystem ist einfach zu bedienen, die Teile können leicht ausgetauscht werden und es können normale Getriebebedingungen, Zahnbruch, Lochfraß, Verschleiß und andere Bedingungen simuliert werden. Die Getriebe haben große und kleine Zahnräder mit einem Modul von 75 für die großen und 55 für die kleinen Zahnräder und werden mit Tauchöl geschmiert.

In one embodiment, an example application for experimental testing and simulation of a transmission is as follows:

• In order to monitor the operating status of the transmission in real time, the transmission is first placed in a test system. The system consists of a transmission diagnostic platform, a vibration sensor, a rotation sensor, a torque sensor, a signal cable, a data acquisition device and a computer. The test system is easy to operate, parts can be easily replaced, and normal gear conditions, tooth breakage, pitting, wear and tear and other conditions can be simulated. The gears have large and small gears with a module of 75 for the large and 55 for the small gears and are lubricated with immersion oil.

The initial input parameters for the dynamics model are first measured in terms of basic gear geometry and material parameters, including mass parameters m _p and m _g , inertia parameters I _p and I _g , stiffness parameters _km and k _py , k _gy , damping parameters _cm and c _py , c _gy as well as the geometry measures R _p and R _g . Subsequently, the parameter matrix and the differential dynamics model of the transmission can be constructed and the dynamics model for the rotational variables θ _p and θ _g and the translational (vibration) variables y _p and y _g in the translational arrangement rotation {δ} = {y _p θ _p y _g θ _g } ^T can be solved, allowing the monitoring and acquisition of the rotation signals θ _p0 and θ _g0 , the vibration signals y _p0 and y _g0 , and the torque signals T _p0 and T _g0 in the simulation model.

The torque signals T _p0 and T _g0 from the physical sensors are then fed to the digital twin model, which updates the digital twin model parameters T _p and T _g in real time, giving the updated load matrix: $$\left\{p\right\}=\left\{\begin{array}{c}0\\ -{\mathrm{<figure-callout\; id="0"\; label="T\; p"\; filenames="DE112022000106T5\_0021.png"\; state="\{\{state\}\}">T</figure-callout>}}_p\\ 0\\ -T_G\end{array}\right\}$$

The simulation model can be solved in Matlab after the parametric variables have been determined, and the model has the form [m]{δ̈} + [c]{δ̇} + [k]{δ} = {p}, giving a high degree of Consistency between the digital twin model and the physical model is guaranteed.

Finally, the digital twin model can output monitoring rotation signals θ _p0 , θ _g0 and vibration signals y _p0 , y _g0 , while the model output can be compared in real-time with the physical sensor measurement of rotation signals and vibration signals as an indicator variable reflecting the operating condition of the transmission to provide real-time monitoring of the to achieve gear condition.

In this experiment, the sensor signals for normal, pitted, cracked and worn gears were collected at a speed of 880 rpm with no load at a sampling rate of 5120 Hz, and the sensor signals for all gears were also output in the simulation model. The details are shown in Table 2. There are 3,300 samples in total, 1,000 in normal condition, 1,000 in worn condition, 800 in pitted condition and 500 with broken teeth. Each sample contains the physical sensor recordings of the transmission in continuous operation as well as the rotation signals θ _p0 , θ _g0 and vibration signals y _p0 , y _g0 from the digital twin model. They form the data set for the deep-time diagnostic model used to train the network parameters and test the network performance.

Table 2 Information on the number of samples collected for the experiment status Number of samples normal 1000 pitting 800 broken tooth 500 wear and tear 1000

The raw data does not directly reflect the variations in each parameter, as the data monitored by the different sensors represent different physical properties, resulting in different magnitudes for each parameter. To avoid the influence of the magnitude and range of the predicted process parameters, the data must be uniformly normalized and the values of all attributes fixed to an identical range of values. Normalization reduces the differences between values, avoids problems with data bias, facilitates subsequent training, and improves the accuracy and convergence of the predictive model. With this in mind, the min-max approach to normalization should be used. Assume the input X = {x ₁ , x ₂ , ···,x _k } ∈ ℝ ^N×K , where N is the time step and K is the number of sensors. The i-th vector is normalized according to the maximum-minimum normalization method as follows: $$x_i=\frac{2\left(x_i-\text{at least}\left(x_i\right)\right)}{\text{Max}\left(x_i\right)-\text{at least}\left(x_i\right)}$$

where min(x _i ), max(x _i ) represent the maximum and minimum values of the vector xi, respectively.

Because there are four categories of transmission faults and to facilitate direct entry of the classification results into the network, a unique thermal code was used for each fault, the results of which are shown in Table 3. Table 3 Unique Heat Code for Transmission Failure Types status expected node output node 1 node 2 Node 3 node 4 normal 1 0 0 0 pitting 0 1 0 0 broken tooth 0 0 1 0 wear and tear 0 0 0 1

In the training process, 80% of the samples from each complete training subset are randomly selected as the training set and the remaining 20% as the validation set. The hyperparameters of the deep diagnostic time series mesh were selected and adjusted based on the test results of the model on the validation set. These hyperparameters were determined by weighing prediction accuracy against computational cost, and the model's final training flowchart is in shown.

For each round of training, the samples are randomly divided into a series of small batches of 32 samples each and entered into the training system. The features are first extracted from the original data using three one-dimensional convolution units, and the local features are extracted from the data using multiple filters in each convolution layer. The size of each convolution kernel in the convolution layer is 2*1 and is shifted one step at a time, and the padding value is set to 1 for each time series to ensure no edge information is lost. A non-linear activation function is then applied to the output of the convolution layer to improve the mesh representation. The output is finally passed through a pooling layer to remove redundant information. The window size is set to 2 in the pooling layer and shifted by two steps, i. H. feature mapping is completed within a convolution unit. When the one-dimensional convolutional network extracts the signal error features, the error features are fed into the fully connected layer to obtain higher-level error features, and the output data has dimensions m*n, which is the original dimension of the input data. Hierarchical attentional mechanisms are then employed to fuse information about high-level traits. It calculates the hidden state of the erroneous feature at each point in time and normalizes it using the softmax function, then multiplies the input sequence by the weights to achieve feature fusion. Finally, the feature vector of the hierarchical attention network is fed into a softmax classifier to enable fault diagnosis of the transmission. The deep temporal diagnostic network uses a cross-entropy loss function as the objective function, backpropagation learning to update the weights in the network, and an Adam optimizer with adaptive tuning capabilities for optimization. The model is trained once and the learning rate is set to 0.0001 for stable convergence. By default, the maximum number of training sessions for the model is 700.

A deep temporal diagnostic network was constructed using the pytorch deep learning framework, setting the number of iterations of the network to 700, with 16 samples per input, and a learning rate of 0.0001. The grid parameters have been optimized using gradient descent and the training curve of the grid loss function is in shown.

Out of 8th it can be seen that after 700 iterations the network error has decreased to a value small enough to prove that the network can effectively adapt the training patterns. The temporal depth diagnosis was applied to a sample of 660 tests to assess the training effect of the network. The results of the network to diagnose the depth of time for the four faults in the gearbox are shown in the table. As can be seen from Table 4, the network had the highest accuracy of 98% in detecting normal state gears, with four samples being misdiagnosed and two each being diagnosed as pitting and tooth fracture defects. The net had the lowest detection accuracy for pitting errors in gears, but it was still over 90%. 10 of the samples were misdiagnosed as other conditions, two of them as gear breakage and eight as wear defects, probably because there were some similarities in the characteristics of pitting and wear defects that resulted in the mesh's inability to effectively discriminate , which error belonged to which error. The mesh achieved good results in detecting broken gears and wear errors with an accuracy of 96% and 94% respectively. With a total of 660 test patterns, a total of 30 patterns were misdiagnosed, and the overall detection rate of the model was 95.5%, indicating that the temporal deep diagnostic network was able to effectively identify the four faults in the gearbox.

Table 4 Results of the failure prediction of the temporal depth diagnosis network type of disorder normal pitting broken tooth wear and tear normal 98% 1% 1% 0 pitting 0% 93.75% 1.25% 5% broken tooth 1% 1% 96% 2% wear and tear 1% 3% 2% 94%

With the help of the above steps, this patent application provides a technical solution to create a digital model of the digital twin model based on the material parameters and geometric dimensional parameters of the gears in the gearbox, after which the parameters of the digital twin model are determined by simulation in combination with the real one Torque signal of the gearbox can be corrected. The digital twin model corresponds more closely to the actual working conditions of the gearbox than traditional techniques to make the data quality of the analog signal more realistic. The real torque signal is thus used as input to the model, and the analog vibration signal and the analog rotation signal are output. This not only widens the range of transmission failure data, but also leads to higher analog signal quality, thereby improving the fault diagnosis accuracy of subsequent transmission failure diagnosis based on the transmission real signal data and transmission analog signal data.

In addition, the present invention uses a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism to output fault diagnosis results during fault diagnosis to further strengthen useful sample features and weaken useless sample features, thereby further improving the accuracy of fault diagnosis. wherein the feature extraction step of the first neural network converts the error signals collected by the plurality of sensors from two-dimensional data to one-dimensional data, then performs one-dimensional convolution, and then iteratively splits the resulting one-dimensional features into two-dimensional data for multi-stage one-dimensional, two-dimensional and one-dimensional feature extraction. The result is a signal with more pronounced features and the same dimensionality as the original data. This makes it easier to classify the subsequent error messages and increases diagnostic accuracy. The linear perceptron-based hierarchical attention processing of feature vectors increases the non-linearity of the neural network to better account for the degradation characteristics of gears, making the fault diagnosis results closer to the real situation and improving the accuracy of gear fault diagnosis.

As in 9 As shown, this embodiment also offers a transmission failure diagnosis device comprising the following.

A signal acquisition module 201 for acquiring a real signal of a transmission and an analog signal generated by an embodiment of a transmission signal acquisition method, the real signal comprising a real torque signal, a real vibration signal and a real rotation signal. Details can be found in the corresponding description of step S201 in the above embodiment of the method, which is not repeated here.

Fault diagnosis module 202 for inputting the analog and real signals into a predefined fault diagnosis network to obtain a fault diagnosis result for the transmission. Details can be found in the corresponding description of step S201 in the above embodiment of the method, which is not repeated here.
wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, the predetermined fault diagnosis network being generated based on training with simulated signals and real signals.

The embodiment of the present invention provides a gear failure diagnosis apparatus for performing a gear failure diagnosis method in the above embodiment, which is implemented in the same manner and principle as described in detail in the corresponding description of the above method embodiment and will not be repeated become.

Through the synergy of the above components, the present application offers a technical solution to create a digital model of the digital twin model based on the material parameters and geometric parameters of the gears inside the gearbox, after which the parameters of the digital twin model are determined by simulation in combination with the real ones Torque signals of the transmission are corrected. The digital twin model corresponds more closely to the actual working conditions of the gearbox than traditional techniques to make the data quality of the analog signal more realistic. The real torque signal is thus used as input to the model, and the analog vibration signal and the analog rotation signal are output. This not only widens the range of transmission failure data, but also leads to higher analog signal quality, thereby improving the fault diagnosis accuracy of subsequent transmission failure diagnosis based on the transmission real signal data and transmission analog signal data.

In addition, the present invention uses a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism to output fault diagnosis results during fault diagnosis to further strengthen useful sample features and weaken useless sample features, thereby further improving the accuracy of fault diagnosis. wherein the feature extraction step of the first neural network converts the error signals collected by the plurality of sensors from two-dimensional data to one-dimensional data, then performs one-dimensional convolution, and then iteratively splits the resulting one-dimensional features into two-dimensional data for multi-stage one-dimensional, two-dimensional and one-dimensional feature extraction. The result is a signal with more pronounced features and the same dimensionality as the original data. This makes it easier to classify the subsequent error messages and increases diagnostic accuracy. The linear perceptron-based hierarchical attention processing of feature vectors increases the non-linearity of the neural network to better account for the degradation characteristics of gears, making the fault diagnosis results closer to the real situation and improving the accuracy of gear fault diagnosis.

10 Figure 1 shows an electronic device of an embodiment of the present invention, the device comprising a processor 901 and a memory 902, which may be connected via a bus or otherwise, the connection via a bus in is used as an example.

The processor 901 can be a central processing unit (CPU). Processor 901 can also be other general purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination act on these building blocks. discrete logic chips with gates or transistors, discrete hardware components, or a combination of these types of chips.

Memory 902 serves as a non-transitory, computer-readable storage medium that can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as B. the program instructions/modules corresponding to the method in the embodiment described above. Processor 901 carries out the method in the embodiment described above by executing the non-transitory software programs, instructions and modules stored in memory 902, thereby performing various functional applications of the processor and data processing.

Memory 902 may include a stored programs area and a stored data area, where the stored programs area may store the operating system and applications required for at least one function; the stored data area can store the data generated by the processor 901 and so on. In addition, memory 902 may include high-speed random access memory as well as non-transitory memory such as memory. B. comprise at least one disk storage device, a flash memory device or other non-transitory solid state storage device. In some embodiments, memory 902 optionally includes memory that is remote relative to processor 901, and these remote memories may be connected to processor 901 via a network. Examples of the above networks include the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

The one or more modules are stored in memory 902 and when executed by processor 901 perform the method in the method embodiment described above.

The specific details of the above electronic devices can be understood by reference to the corresponding descriptions and effects in the above method embodiments and will not be repeated here.

It will be appreciated by those skilled in the art that all or part of the processes in the methods of the above embodiments may be performed by a computer program to instruct the appropriate hardware to do so, and the implemented program may be stored on a computer readable storage medium, which, when executed, may include processes such as those of the method embodiments described above. The storage medium can be a floppy disk, optical disk, read only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid state drive (SSD) act etc.; the storage medium can also comprise a combination of the storage types mentioned above.

Although embodiments of the invention are described in connection with the accompanying drawings, various modifications and variations can be made by those skilled in the art without departing from the spirit and scope of the invention, and such modifications and variations fall within the scope defined by the appended claims.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• CN 113505655A [0003]

Claims (9)

A method for signal acquisition of a transmission, characterized in that the method comprises: creating a digital twin model based on material and geometry parameters of the gears in a transmission; obtaining a real torque signal of the transmission, the real torque signal comprising first signal data for correcting the parameters of a digital twin model and second signal data for calculating an analog signal; inputting the first signal data into the digital twin model to calibrate the parameters of the digital twin model; inputting the second signal data into the corrected digital twin model to generate an analog signal, the analog signal including an analog vibration signal and an analog rotation signal; wherein the creation of a digital twin model based on material parameters and geometric parameters of the gears in the transmission comprises: obtaining material parameters and geometric parameters of the active or driven wheels in the transmission, the material parameters comprising at least one of the following parameters: mass parameters, mass inertia parameters, stiffness parameters and damping parameters; generating equations for the differential dynamics of the transmission based on the physical structure of the mesh between the active and driven wheels using the material parameters and the geometric parameters; calculating the simulation matrix for the output of the analog signal using the equations of the differential dynamics of the gearbox; creating a mathematical model to characterize the gear load relationship using the simulation matrix, the inertia matrix, the damping matrix and the stiffness matrix and using the mathematical model as the digital twin model, the inertia matrix being formed based on the mass parameters and the inertia parameters, and the damping matrix on based on the damping parameters and the geometry parameters and the stiffness matrix is formed on the basis of the stiffness parameters and the geometry parameters. Method for transmission error diagnosis of a transmission, characterized in that the method comprises the following: Obtaining a real signal of the transmission and an analogue signal, which is determined by the method according to claim 1 is generated, wherein the real signal includes a real torque signal, a real vibration signal, and a real rotation signal; inputting the analog signal or the real signal into a predetermined fault diagnosis network to obtain a fault diagnosis result for the transmission; wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting the fault diagnosis results, and the predetermined fault diagnosis network is generated based on the training of the analog signals and the real signals. procedure after claim 2 , characterized in that the step of extracting features of the target signal based on the first neural network using the analog signal and/or the real signal as the target signal comprises: assembling the target signal into a two-dimensional plane vector using the number of sampled time points and the number of sensors of the target signal as two-dimensional plane axes; performing a one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution kernels of the same dimension to obtain a plurality of one-dimensional original features; separately aggregating each of the one-dimensional original features to obtain a plurality of one-dimensional time-series features; two-dimensionally merging a plurality of one-dimensional time-series features to obtain a new two-dimensional plane vector and adjusting the number and dimensionality of the one-dimensional convolution kernels; wherein the new two-dimensional plane vector is used as the two-dimensional plane vector to generate a plurality of one-dimensional original features based on the adjusted one obtaining a dimensional convolution kernel, returning to the step of one-dimensional convolution computing the two-dimensional plane vector based on a plurality of one-dimensional convolution kernels of the same dimension until a predetermined number of times of returning to obtain a plurality of target time features; wherein the plurality of target time features are assembled into a one-dimensional target vector and the one-dimensional target vector is fed into a predetermined fully connected layer to transform the one-dimensional target vector via the predetermined fully connected layer into a plurality of one-dimensional abstract features having the same number of Elements have like the number of time points sampled, where the number of elements of the one-dimensional abstract features is the same as the number of sensors. procedure after claim 3 , characterized in that a one-dimensional convolution calculation is performed on the two-dimensional plane vector using the same one-dimensional convolution kernel based on multiple dimensions to obtain multiple one-dimensional original features, including the following steps: for each sensor element, a plurality of elements from the two-dimensional plane vector with a predetermined number of sampling times is smoothly extracted and the plurality of elements with a predetermined number of sampling times are formed into a plurality of one-dimensional vectors to be convolved; wherein each one-dimensional vector to be convolved is sequentially convolved with the current one-dimensional convolution kernel and computed to produce the current one-dimensional original features; on this basis, each one-dimensional vector to be convolved is convolved in turn with the current one-dimensional convolution kernel to generate the current one-dimensional original feature, and all one-dimensional convolution kernels are traversed until the one-dimensional original feature corresponding to each one-dimensional convolution kernel is obtained . procedure after claim 3 , characterized in that the step of outputting fault diagnosis results based on the second neural network comprises: separate dimensional expansion of a plurality of one-dimensional abstract features; wherein the augmented plurality of one-dimensional abstract features are subjected to linear perceptron-based hierarchical attention processing and the plurality of one-dimensional abstract features are merged to obtain an output vector; inputting the output vector to a predetermined classifier to cause the predetermined classifier to output the failure diagnosis result. procedure after claim 5 , characterized in that , based on a linear perceptron, a hierarchical attention process is performed for the plurality of one-dimensional abstract features and the plurality of one-dimensional abstract features is merged to obtain an output vector, including the following steps: inputting the plurality of one-dimensional abstract features into the linear perceptron to obtain a variety of hidden features; calculating a plurality of weight coefficients corresponding to the plurality of hidden features using a softmax function, the weight coefficients being used to characterize the hierarchical attention of the hidden features; weight and sum the plurality of hidden features using the plurality of weight coefficients to obtain the output vector. Device for transmission fault diagnosis of a transmission, characterized in that the device comprises the following: a signal acquisition module for obtaining a real signal of the transmission and an analogue signal, which is determined by the method according to claim 1 is generated, wherein the real signal includes a real torque signal, a real vibration signal, and a real rotation signal; a fault diagnosis module for inputting the analog signal and the real signal into a predetermined fault diagnosis network to obtain a fault diagnosis result for the transmission; wherein the predetermined fault diagnosis network comprises a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting the fault diagnosis results, and the predetermined fault diagnosis network is generated based on the training of the analog signals and the real signals. Electronic device, characterized in that the device comprises a memory and a processor, wherein the memory and the processor are communicatively connected to each other, and wherein computer instructions are stored in the memory, wherein the processor executes the computer instructions, whereby the method according to any one of Claims 1 - 6 is carried out. A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, the computer instructions being used to cause the computer to perform a method according to any one of Claims 1 until 6 to perform.

Images