Improved aero-engine support system vibration tracing method based on otpa and bi-lstm fusion

Abstract

The present application belongs to the technical field of aero-engine vibration tracing, and particularly relates to an improved OTPA and Bi-LSTM fusion aero-engine support system vibration tracing method, characterized in that the influence of engine speed variation and real physical transmission process on vibration signal tracing is comprehensively considered, a vibration transmission rate model conforming to the real physical law is obtained based on the improved forward OTPA network, the traditional neural network training process is improved by introducing the transmission rate model as prior physical information, and the internal vibration source prediction result conforming to the physical law and having the minimum error is obtained. Compared with the prior art, the present application has the beneficial effects that by combining the prior physical transmission law with the excellent feature extraction capability of deep learning, not only a reliable intelligent tracing model is provided for the multi-source vibration decoupling of the rotor system, but also a new digital technical approach is provided for the accurate positioning of the aero-engine complex vibration fault and the vibration reduction design.

Description

Technical Field

[0001] This invention belongs to the field of vibration tracing technology for aero-engines, and particularly relates to an improved vibration tracing method for aero-engine support systems that integrates OTPA and Bi-LSTM. Background Technology

[0002] As highly complex thermodynamic rotating machines, aero-engines operate under conditions of high temperature, high pressure, high speed, and strong vibration for extended periods. Their internal components endure extremely high mechanical loads and thermal shocks, leading to a high failure rate and a significant multi-mode composite failure characteristic. Furthermore, due to the complex operating conditions of aero-engines and the ease with which vibration signals are distorted or masked during transmission through structural and accessory systems, cross-structural system faults are among the most easily overlooked and most difficult to diagnose vibration faults. Therefore, accurately revealing the transmission patterns of vibration signals within the entire system and quantifying the vibration contribution of each transmission path is a core challenge in overcoming the bottleneck of primary fault diagnosis. Achieving efficient and accurate vibration source tracing can provide reliable theoretical support for engine fault diagnosis, and elucidating its vibration transmission mechanism has become an important research direction in the field.

[0003] Existing research on vibration signal tracing mainly focuses on gearboxes and individual mechanical components, achieving relatively effective fault source identification through physical parameter analysis or basic deep learning methods. However, the vibration signals of aero-engine support systems exhibit significant non-stationary, strongly coupled, and variable parameter characteristics. During the transmission of internal vibration signals to the casing monitoring points, the signals are not only subjected to amplitude-frequency modulation by path transmission characteristics but also experience multi-source coupling effects with other vibration sources, ultimately forming a composite vibration response with nonlinear characteristics on the casing surface. Relying solely on sensor measurement points on the casing surface cannot clearly characterize the internal excitation vibration information under different operating conditions. Traditional methods struggle to accurately identify multi-condition fault sources in aero-engines under complex vibration transmission paths, resulting in difficulties in vibration signal tracing. Summary of the Invention

[0004] The purpose of this invention is to provide an improved vibration source tracing method for aero-engine support systems that integrates OTPA and Bi-LSTM, overcoming the shortcomings of existing technologies. It comprehensively considers the influence of engine speed changes and the actual physical transmission process on vibration signal tracing. Based on the improved forward OTPA network, a vibration transmissibility model that conforms to the actual physical laws is obtained. By introducing this transmissibility model as prior physical information, the traditional neural network training process is improved, and the prediction results of internal vibration sources that conform to physical laws and have the smallest error are obtained. This enables accurate reproduction of vibration signals at each excitation point and solves the problem of difficulty in tracing vibration signals.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An improved vibration source tracing method for aero-engine support systems, integrating OTPA and Bi-LSTM, comprehensively considers the influence of engine speed variations and the actual physical transmission process on vibration signal tracing. Based on an improved forward OTPA network, a vibration transmissibility model conforming to real physical laws is obtained. By introducing this transmissibility model as prior physical information, the traditional neural network training process is improved, yielding internal vibration source prediction results that conform to physical laws and have the smallest error. The method includes the following steps: S1) Collect vibration signals of the aero-engine casing and internal support points under different operating conditions, establish an OTPA model, and verify the rationality of the model using coherence analysis; S2) The acquired vibration signal is preprocessed to extract the transient rotational speed, and then subjected to Fourier transform to obtain a complex frequency domain feature matrix containing amplitude and phase. S3) Establish a speed-adaptive dynamic OTPA neural network model, and dynamically interpolate and fuse tensors based on speed weights to generate a multi-speed transferability matrix; S4) The transfer matrix is ​​identified and solidified by backpropagation optimization through minimizing the complex field propagation error in the complex-valued neural network. S5) Establish a vibration source tracing model based on the fusion of improved OTPA and Bi-LSTM neural network; S6) The transfer rate matrix in the improved OTPA is embedded as prior physical information into the global loss function of the neural network, and the vibration source tracing model is subjected to closed-loop joint training based on prior physical information constraints to realize vibration signal source tracing. S7) Based on the predicted excitation sources obtained from the vibration source tracing model, the contribution of each excitation source is quantified using the improved OTPA method to identify and locate the fault location and the main vibration source.

[0006] Further, step S1) specifically involves: collecting vibration acceleration signals from the casing response point and the internal support excitation source of the aero-engine under different operating conditions to establish an OTPA model; calculating the coherence coefficient between the excitation source and the casing response; verifying the rationality of establishing the OTPA model using coherence analysis; and calculating the coherence coefficient according to the following formula:

[0007] In the formula: Indicates the output signal The self-power spectrum, This represents the portion of the output signal power spectrum caused by factors other than the input signal, such as noise interference, system nonlinearity, and missing excitation sources. is the partial coherence coefficient of the system.

[0008] Furthermore, step S2) specifically involves: using short-time Fourier transform to adaptively track the fundamental frequency to extract a smooth transient rotational speed sequence; then, after the vibration signal undergoes fast Fourier transform, it is adaptively normalized to obtain a complex frequency domain feature matrix containing amplitude and phase, and the dataset is divided as input to the neural network.

[0009] Further, step S3) specifically involves: converting the complex linear mapping matrix to be learned in the neural network into an equivalent transfer rate matrix; then, obtaining the corresponding floating-point index value using the transient rotational speed at the current moment, and extracting the decimal part of the floating-point index value as the rotational speed interpolation weight; subsequently, dynamically linearly interpolating the complex transfer rates corresponding to adjacent rotational speed intervals to generate a multi-rotational speed transfer rate matrix adapted to the current rotational speed in real time. The calculation formula is as follows:

[0010]

[0011]

[0012] In the formula: This is the floating-point index value for the rotational speed. The transient rotational speed at the current moment, The preset minimum boundary speed, The resolution size of the divided speed range; For speed interpolation weights, This indicates rounding down the floating-point index; This represents a dynamically generated multi-speed transfer rate matrix. and These represent the complex transfer rates corresponding to adjacent speed ranges located by rounding down and up, respectively.

[0013] Further, step S4) specifically involves: in the network layer where the activation function is removed, the generated transfer rate matrix and the pivot complex frequency domain signal are multiplied by a decoupled tensor with real and imaginary parts to predict the chassis response, and backpropagation is performed by calculating the complex domain mean square error. The relevant calculation formula is as follows;

[0014] In the formula, To propagate errors in the complex domain of the forward OTPA network, and For the first The real and imaginary parts of a sample's actual chassis response. This represents the total number of samples in the batch.

[0015] Furthermore, step S5) specifically involves: mapping the operating condition label to a condition vector, concatenating it with the historical frequency domain features of the casing, inputting it into a bidirectional LSTM to extract the temporal pattern, and establishing an improved vibration tracing model that integrates OTPA and Bi-LSTM.

[0016] Furthermore, step S6) specifically involves: introducing a dynamic peak weighting mechanism into the reconstruction loss for the excitation source to strengthen the resonance penalty; simultaneously, substituting the excitation source signal predicted by the inverse network into the fixed forward transitivity matrix to reconstruct the casing response, and calculating the error between it and the actual casing response as the physical constraint loss; finally, optimizing the inverse LSTM network by combining the two through a balancing coefficient, as shown in the following formula:

[0017]

[0018] In the formula: Physical constraint loss for reconstructing the casing response; The positive OTPA transfer rate matrix is ​​for solidification. The actual measured amplitude of the casing response. The excitation source signal is predicted by the inverse network; For the global joint loss function, To address the data reconstruction loss from the incentive source, This is the balance coefficient for physical constraint losses.

[0019] Furthermore, step S7) specifically involves: taking cruise operation as an example, after obtaining the predicted excitation source signal, the vibration contribution of each support path is calculated using the improved OTPA method to achieve accurate quantification and fault location of the main vibration source.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1) This invention improves the accuracy of multi-source vibration tracing in complex rotor systems with an intelligent tracing method. It comprehensively considers the influence of engine speed changes and the actual physical transmission process on vibration signal tracing. Based on the improved forward OTPA network, a vibration transmissibility model that conforms to the actual physical laws is obtained. By introducing this transmissibility model as prior physical information, the traditional neural network training process is improved, and the prediction results of internal vibration sources that conform to physical laws and have the smallest error are obtained. The tracing signals are accurate and reliable, and can provide a reference for the investigation and diagnosis of complex internal faults of aero engines. 2) This invention uses engine speed, operating conditions and vibration data of the casing surface as network input, uses the solidified vibration transmissibility and resonance peak weighting function as global loss function, and takes the accurate restoration of vibration sources at each internal excitation point and the calculation of vibration contribution as optimization objectives to achieve accurate source tracing and fault location of complex vibration signals inside the aero-engine. 3) By combining the prior physical transmission law with the excellent feature extraction capability of deep learning, this invention not only provides a reliable intelligent source tracing model for the decoupling of multi-source vibrations in rotor systems, but also provides a brand-new digital technology approach for the precise location and vibration reduction design of complex vibration faults in aero-engines. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the control flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the OTPA model established in an embodiment of the present invention; Figure 3 This is a schematic diagram of the recoherence coefficients from each excitation point to the casing measurement point in an embodiment of the present invention; Figure 4 This is a schematic diagram of the improved OTPA neural network architecture in an embodiment of the present invention; Figure 5 This is a schematic diagram of the transfer rate of the full-band excitation point to the casing measurement point in an embodiment of the present invention; Figure 6 This is a schematic diagram of the Bi-LSTM neural network structure in an embodiment of the present invention; Figure 7 This is a comparison diagram of the vibration source signal and the actual signal at each excitation point in the embodiment of the present invention; Figure 8 This is a schematic diagram showing the contribution of each excitation point to the casing measurement point under frequency switching in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The components of the embodiments of the present invention described and shown in the accompanying drawings can typically be arranged and designed in many different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0024] See Figure 1 This is a schematic diagram of the control flow of an embodiment of the vibration signal tracing method for aero-engine support systems based on the fusion of improved OTPA and Bi-LSTM, which includes the following steps: S1) Collect vibration signals of the aero-engine casing and internal support points under different operating conditions to establish an OTPA model, see Figures 2-3 The rationality of the model was verified using coherence analysis. Specifically, vibration acceleration signals from the casing response point and internal support excitation source of the aero-engine under different operating conditions were collected to establish an OTPA model. The re-coherence coefficient between the excitation source and the casing response was calculated, and the rationality of the established OTPA model was verified using re-coherence analysis. The re-coherence coefficient was calculated according to the following formula: Formula (1) In the formula Indicates the output signal The self-power spectrum, This represents the portion of the output signal power spectrum caused by factors other than the input signal, such as noise interference, system nonlinearity, and missing excitation sources. is the partial coherence coefficient of the system.

[0025] S2) The acquired vibration signal is preprocessed to extract transient rotational speed and obtain a complex frequency domain feature matrix containing amplitude and phase through Fourier transform. Specifically, the fundamental frequency is adaptively tracked using short-time Fourier transform to extract a smooth transient rotational speed sequence. Then, the vibration signal is subjected to fast Fourier transform and adaptive normalization is performed to obtain a complex frequency domain feature matrix containing amplitude and phase. The dataset is then divided and used as the input to the neural network.

[0026] S3) Establish a speed-adaptive dynamic OTPA neural network model, and dynamically interpolate and fuse tensors based on speed weights to generate a multi-speed transmissibility matrix; see Figure 4Specifically, the complex linear mapping matrix to be learned in the neural network is equivalent to a transfer rate matrix. Then, the corresponding floating-point index value is obtained using the transient rotational speed at the current moment, and the decimal part of the floating-point index value is extracted as the rotational speed interpolation weight. Then, the complex transfer rates corresponding to adjacent rotational speed intervals are dynamically linearly interpolated to generate a multi-rotational speed transfer rate matrix adapted to the current rotational speed in real time. The calculation formula is as follows: Formula (2) Formula (3) Formula (4) In the formula: This is the floating-point index value for the rotational speed. The transient rotational speed at the current moment, The preset minimum boundary speed, The resolution size of the divided speed range; For speed interpolation weights, This indicates rounding down the floating-point index; This represents a dynamically generated multi-speed transfer rate matrix. and These represent the complex transfer rates corresponding to adjacent speed ranges located by rounding down and up, respectively.

[0027] S4) Optimize by minimizing the complex-domain propagation error in a complex-valued neural network through backpropagation, identifying and solidifying the propagation matrix; see [link / see details] Figure 5 Specifically, in the network layer where the activation function is removed, the generated transfer rate matrix and the pivot complex frequency domain signal are multiplied by a decoupled tensor with real and imaginary parts to predict the chassis response, and backpropagation is performed by calculating the complex domain mean square error. The relevant formulas are as follows. Formula (5) In the formula, To propagate errors in the complex domain of the forward OTPA network, and For the first The real and imaginary parts of a sample's actual chassis response. This represents the total number of samples in the batch.

[0028] S5) Establish a vibration source tracing model based on the fusion of improved OTPA and Bi-LSTM neural network; specifically, map the operating condition label into a condition vector, concatenate it with the historical frequency domain features of the casing, input it into a bidirectional LSTM to extract the time series pattern, and establish an improved vibration source tracing model fused with OTPA and Bi-LSTM, see [link to relevant documentation]. Figure 6 .

[0029] S6) The improved OTPA transmissivity matrix is ​​embedded as prior physical information into the global loss function of the neural network to perform closed-loop joint training of the vibration source tracing model based on prior physical information constraints, thereby achieving vibration signal source tracing. Specifically, a dynamic peak weighting mechanism is introduced into the reconstruction loss for the excitation source to strengthen the resonance penalty. At the same time, the excitation source signal predicted by the inverse network is substituted into the fixed forward transmissivity matrix to reconstruct the casing response, and the error between it and the real casing response is calculated as the physical constraint loss. Finally, the inverse LSTM network is optimized by combining the two through a balancing coefficient. The calculation formula is as follows: Formula (6) Formula (7) In the formula, Physical constraint loss for reconstructing the casing response; The positive OTPA transfer rate matrix is ​​for solidification. The actual measured amplitude of the casing response. The excitation source signal is predicted by the inverse network; For the global joint loss function, To address the data reconstruction loss from the incentive source, This is the balance coefficient for physical constraint losses.

[0030] S7) Based on the predicted excitation sources obtained from the vibration source tracing model, the contribution of each excitation source is quantified using the improved OTPA method to identify and locate the fault location and the main vibration source.

[0031] Taking cruise operation as an example, after obtaining the predicted excitation source signal, the vibration contribution of each support path is calculated using the improved OTPA method, so as to achieve accurate quantification and fault diagnosis of the main vibration sources.

[0032] In the embodiments, see Figures 7-8 The model's final predicted signal accurately reproduced the main characteristic frequency components of the measured signal. Particularly at 110Hz (low-voltage rotor frequency) and 220Hz (high-voltage rotor frequency) and their higher harmonics, the predicted curves highly coincided with the peak positions of the measured data, with amplitude errors controlled within 3dB. The model's predicted curves maintained the same attenuation trend as the measured signals. This indicates that the model successfully learned the transmission laws at different speeds, achieving accurate vibration source tracing. Simultaneously, using the improved OTPA method, the contribution of the excitation source signal obtained from the tracing was analyzed, determining the main vibration transmission paths and vibration sources of each casing at high and low voltage frequencies and their harmonics, thereby achieving fault location.

[0033] The structural parameter constraint optimization method for the shock resistance capability of the aero-engine connecting casing in this embodiment comprehensively considers the impact of engine speed changes and the actual physical transmission process on vibration signal tracing. Based on the improved forward OTPA network, a vibration transmissibility model that conforms to the actual physical laws is obtained. By introducing this transmissibility model as prior physical information, the traditional neural network training process is improved, and the prediction results of internal vibration sources that conform to physical laws and have the smallest error are obtained. The tracing signal is accurate and reliable, providing a reference for the investigation and diagnosis of complex internal faults of aero-engines.

[0034] Traditional OTPA (Operational Time-Induced Persistent Dynamics) can only perform "forward" (excitation → response) operations, not "inverse" (response → excitation) operations. Purely data-driven methods can perform inverse operations, but lack physical constraints. This invention is the first to explicitly propose and solve the core engineering challenge of "inversely deriving the internal excitation source when only the external response is available, while ensuring that the result conforms to physical laws." By introducing the core concept of physical information neural networks—using physical laws as a loss function constraint—into the field of aero-engine vibration tracing, and replacing the traditional "analytical physical equations" with a "transitivity matrix learned from data" as the physical prior, this "data-driven physical prior" concept has not been reported in existing literature and has outstanding substantive characteristics.

[0035] This invention designs a complete closed-loop architecture of "forward learning → solidification → inverse prediction → physical constraint verification," enabling the model to simultaneously optimize prediction accuracy and physical consistency during training, rather than a simple two-stage separate method of "first calculating the transitivity in the forward direction, and then training the inverse network separately." This invention embeds the solidified OTPA transitivity matrix as prior physical information into the loss function of the Bi-LSTM inverse source tracing network, forming a closed-loop joint training architecture with physical constraints. This concept is groundbreaking in the field of aero-engine vibration source tracing and has clear technical boundaries with existing technologies (traditional OTPA, pure data-driven methods, and traditional PINN).

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An improved method for vibration source tracing of an aero-engine support system integrating OTPA and Bi-LSTM, characterized in that, Taking into account the impact of engine speed variations and the actual physical transmission process on vibration signal source tracing, a vibration transmissibility model conforming to real physical laws was obtained based on an improved forward OTPA network. By introducing this transmissibility model as prior physical information, the traditional neural network training process was improved, resulting in an internal vibration source prediction result that conforms to physical laws and has the smallest error. The specific steps include: S1) Collect vibration signals of the aero-engine casing and internal support points under different operating conditions, establish an OTPA model, and verify the rationality of the model using coherence analysis; S2) The acquired vibration signal is preprocessed to extract the transient rotational speed, and then subjected to Fourier transform to obtain a complex frequency domain feature matrix containing amplitude and phase. S3) Establish a speed-adaptive dynamic OTPA neural network model, and dynamically interpolate and fuse tensors based on speed weights to generate a multi-speed transferability matrix; S4) The transfer matrix is ​​identified and solidified by backpropagation optimization through minimizing the complex field propagation error in the complex-valued neural network. S5) Establish a vibration source tracing model based on the fusion of improved OTPA and Bi-LSTM neural network; S6) The transfer rate matrix in the improved OTPA is embedded as prior physical information into the global loss function of the neural network, and the vibration source tracing model is subjected to closed-loop joint training based on prior physical information constraints to realize vibration signal source tracing. S7) Based on the predicted excitation sources obtained from the vibration source tracing model, the contribution of each excitation source is quantified using the improved OTPA method to identify and locate the fault location and the main vibration source.

2. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S1) specifically involves: collecting vibration acceleration signals from the casing response point and the internal support excitation source of the aero-engine under different operating conditions to establish an OTPA model; calculating the coherence coefficient between the excitation source and the casing response; verifying the rationality of establishing the OTPA model using coherence analysis; and calculating the coherence coefficient according to the following formula: In the formula: Indicates the output signal The self-power spectrum, This represents the portion of the output signal power spectrum caused by factors other than the input signal, such as noise interference, system nonlinearity, and missing excitation sources. is the partial coherence coefficient of the system.

3. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S2) specifically involves: using short-time Fourier transform to adaptively track the fundamental frequency to extract a smooth transient rotational speed sequence; then, after the vibration signal undergoes fast Fourier transform, it is adaptively normalized to obtain a complex frequency domain feature matrix containing amplitude and phase, and the dataset is divided as the input to the neural network.

4. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S3) specifically involves: The complex linear mapping matrix to be learned in the neural network is equivalent to a transfer rate matrix. Then, the corresponding floating-point index value is obtained using the transient rotational speed at the current moment, and the fractional part of this floating-point index value is extracted as the rotational speed interpolation weight. This is then used to dynamically linearly interpolate the complex transfer rates corresponding to adjacent rotational speed intervals, generating a multi-rotational speed transfer rate matrix adapted to the current rotational speed in real time. The calculation formula is as follows: In the formula: This is the floating-point index value for the rotational speed. The transient rotational speed at the current moment, The preset minimum boundary speed, The resolution size of the divided speed range; For speed interpolation weights, This indicates rounding down the floating-point index; This represents a dynamically generated multi-speed transfer rate matrix. and These represent the complex transfer rates corresponding to adjacent speed ranges located by rounding down and up, respectively.

5. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S4) specifically involves: in the network layer where the activation function is removed, the generated transfer rate matrix and the pivot complex frequency domain signal are multiplied by a decoupled tensor with real and imaginary parts to predict the chassis response, and backpropagation is performed by calculating the complex domain mean square error. The relevant calculation formulas are as follows. In the formula, To propagate errors in the complex domain of the forward OTPA network, and For the first The real and imaginary parts of a sample's actual chassis response. This represents the total number of samples in the batch.

6. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S5) specifically involves mapping the operating condition label to a condition vector, concatenating it with the historical frequency domain features of the casing, inputting it into a bidirectional LSTM to extract the timing pattern, and establishing an improved vibration source tracing model that integrates OTPA and Bi-LSTM.

7. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S6) specifically involves: introducing a dynamic peak weighting mechanism into the reconstruction loss for the excitation source to strengthen the resonance penalty; simultaneously, substituting the excitation source signal predicted by the inverse network into the fixed forward transitivity matrix to reconstruct the casing response, and calculating the error between it and the actual casing response as the physical constraint loss; finally, optimizing the inverse LSTM network by combining the two through a balancing coefficient, as shown in the following formula: In the formula: Physical constraint loss for reconstructing the casing response; The positive OTPA transfer rate matrix is ​​for solidification. The actual measured amplitude of the casing response. The excitation source signal is predicted by the inverse network; For the global joint loss function, To address the data reconstruction loss from the incentive source, This is the balance coefficient for physical constraint losses.

8. The improved OTPA and Bi-LSTM fusion vibration tracing method for aero-engine support systems according to claim 1, characterized in that, Step S7) specifically involves: taking cruise operation as an example, after obtaining the predicted excitation source signal, the vibration contribution of each support path is calculated using the improved OTPA method to achieve accurate quantification and fault location of the main vibration source.

Images