A method, system and device for extracting a tip time of arrival

Abstract

The application discloses a blade tip arrival time extraction method, system and device, relates to the technical field of rotating blade non-contact vibration monitoring and signal processing, and comprises the following steps: collecting a blade tip original pulse signal of a rotating machine during operation; inputting the original pulse signal into a pre-trained adaptive wavelet denoising network for multi-scale decomposition; performing point-by-point nonlinear transformation on a feature map after decomposition through a parameterized soft threshold activation function to obtain sparse features; simulating a wavelet inverse transform through one-dimensional transpose convolution operation, upsampling and summing the sparse features to reconstruct a denoised time-domain waveform; and inputting the denoised time-domain waveform into a pre-trained end-to-end time regression network to directly output continuous blade tip arrival time offset prediction values. The method realizes the collaborative innovation of adaptive wavelet basis deep learning denoising and end-to-end time regression, and significantly improves the reliability and accuracy of rotating machine blade vibration non-contact monitoring.

Description

Technical Field

[0001] This invention relates to the field of non-contact vibration monitoring and signal processing technology for rotating blades, specifically to a method, system, and device for extracting the arrival time of the blade tip. Background Technology

[0002] As a core power-working component in rotating machinery such as steam turbines and aero engines, the operating condition of blades under extreme conditions such as high temperature, high pressure, and high speed directly determines the safety, reliability, and service life of the entire unit. In actual operation, blades are subjected to multiple load coupling effects, including airflow excitation, centrifugal force, and rotor imbalance. This not only generates conventional vibration stress caused by excitation but may also lead to flutter or asynchronous resonance, which can then cause high-cycle fatigue fracture accidents. Therefore, real-time and accurate online monitoring of blade vibration is crucial to ensuring the safe operation of the unit.

[0003] Currently, non-contact blade tip timing measurement technology has gradually replaced traditional contact strain gauge measurement methods and become the mainstream method for blade vibration monitoring because its sensors do not need to rotate with the rotor, do not damage the blade flow field structure, and have good durability. The core premise of blade tip timing technology is to accurately obtain the "arrival time" of the blade passing the sensor. The accuracy of this moment directly determines the accuracy of subsequent vibration displacement reconstruction and parameter identification. However, the electromagnetic environment in industrial sites is complex, and factors such as changes in blade tip clearance, airflow disturbance, and blade dust accumulation often cause the raw pulse signals collected by the sensor to be mixed with a large amount of background noise, and the waveform is prone to amplitude fluctuations and asymmetric distortion.

[0004] Existing methods for extracting blade tip arrival times mainly rely on hardware analog circuits or traditional digital signal processing algorithms. While hardware methods offer fast response times, they are sensitive to signal amplitude fluctuations. When significant blade vibrations cause changes in the signal envelope, substantial timing errors can easily occur. Traditional digital methods offer improvements, but usually require pre-processing to denoise the signal. Traditional wavelet denoising or filtering algorithms heavily rely on manual experience in selecting wavelet basis functions and decomposition levels. If the selected basis functions do not match the actual waveform characteristics, effective signals may be lost or denoising may be incomplete. These methods are highly subjective and limited, making them unsuitable for the demands of online monitoring under varying operating conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as loss of effective signals or incomplete noise reduction, which makes them unsuitable for online monitoring under varying operating conditions, this invention proposes a method, system, and device for extracting the arrival time of the blade tip, thereby solving the problems existing in the prior art.

[0006] A method for extracting the arrival time of a leaf tip includes the following steps: Acquire raw pulse signals from the blade tips during the operation of rotating machinery; The original pulse signal is input into a pre-trained adaptive wavelet denoising network. The adaptive decomposition module performs multi-scale decomposition of the input signal using multi-layer convolution kernels and downsampling operations to generate feature maps. The parameterized soft thresholding activation function in the nonlinear sparse thresholding module performs point-by-point nonlinear transformation on the decomposed feature maps to obtain sparse features. The adaptive reconstruction module uses one-dimensional transposed convolution operation to simulate inverse wavelet transform, upsampling and summing the sparse features to reconstruct the denoised time-domain waveform. The denoised time-domain waveform is input into a pre-trained end-to-end time regression network. Deep feature extraction is performed on the denoised time-domain waveform through stacked one-dimensional convolutional layers in the deep feature extraction module to obtain a time-series feature map. The extracted time-series feature map is then subjected to global average pooling by the global feature aggregation module to obtain a fixed-length global feature vector. The global feature vector is then nonlinearly mapped by the nonlinear regression module to directly output the continuous leaf tip arrival time offset prediction values.

[0007] Furthermore, the point-by-point nonlinear transformation of the decomposed feature map using the parameterized soft thresholding activation function in the nonlinear sparse thresholding module to obtain sparse features is specifically represented as follows: ; in, The decomposed feature map, , The kernel length is 1. For the first The weight vector of each convolutional kernel. For network parameters, An independent threshold is learned for each feature channel.

[0008] Furthermore, the process of simulating the inverse wavelet transform using one-dimensional transpose convolution to upsample and sum sparse features and reconstruct the denoised time-domain waveform is specifically represented as follows: ; in, These are sparse features after thresholding. It involves reconstructing the convolutional kernel.

[0009] Furthermore, the training process of the adaptive denoising network specifically includes the following steps: The simulated blade tip pulse signal with arrival time label is obtained during the operation of rotating machinery, and the measured noise library is constructed based on static and dynamic background noise in industrial field. A mixed training dataset containing noisy signal samples, corresponding clean signal labels and arrival time true value labels is generated. Using noisy signal samples as input to a one-dimensional convolutional neural network and corresponding clean signal labels as training targets, the reconstruction error between the network output and the clean signal labels is minimized. By jointly applying sparsity constraints on the intermediate coefficients of the network, the convolutional kernel weights in the network are adaptively updated to approximate the optimal wavelet basis function and the threshold parameters are adaptively updated to filter out noise coefficients, resulting in a high signal-to-noise ratio waveform.

[0010] Furthermore, the process of constructing the hybrid training dataset specifically includes the following steps: Based on the photoelectric sensing principle and blade vibration model, a simulated pulse signal with a precise arrival time label is generated; Through static and dynamic experiments, measured background noise of rotating machinery under different operating conditions was collected, and a measured noise database was constructed. The simulated pulse signal is mixed with a noise segment randomly selected from the measured noise library according to a preset signal-to-noise ratio distribution to generate a noisy synthetic signal sample. The synthesized signal samples are subjected to data augmentation operations including amplitude scaling, time shifting, and waveform distortion simulation, and are then subjected to standardization preprocessing to finally form a training dataset containing the input signal and its corresponding labels. The labels include a clean simulated signal for training the denoising network and an accurate arrival time value for training the regression network.

[0011] Furthermore, it also includes employing a composite loss function that incorporates physical constraints. The parameters of the adaptive wavelet denoising network are trained; the composite loss function Represented as: ; Among them, reconstruction fidelity loss To measure network output using mean square error With the ideal noise-free label Differences between them; sparsity-induced loss For the feature tensor output of the threshold layer Apply L1 norm penalty.

[0012] Furthermore, the training process of the end-to-end regression network specifically includes: using a high signal-to-noise ratio waveform as the input of the deep regression network and the arrival time ground truth label as the target, training the network to learn the mapping from waveform features to arrival time ground truth, so as to output the predicted value of the leaf tip arrival time, thereby completing the training of the leaf tip arrival time extraction model.

[0013] The present invention also proposes a leaf tip arrival time extraction system, comprising: The acquisition module is used to collect the raw pulse signals from the blade tips during the operation of rotating machinery; The reconstruction module is used to input the original pulse signal into a pre-trained adaptive wavelet denoising network. The adaptive decomposition module uses multi-layer convolution kernels and downsampling operations to decompose the input signal at multiple scales to generate feature maps. The parameterized soft thresholding activation function in the nonlinear sparse thresholding module performs a point-by-point nonlinear transformation on the decomposed feature maps to obtain sparse features. The adaptive reconstruction module uses one-dimensional transposed convolution operations to simulate the inverse wavelet transform, upsampling and summing the sparse features to reconstruct the denoised time-domain waveform. The output module is used to input the denoised time-domain waveform into the pre-trained end-to-end time regression network. The deep feature extraction module performs deep feature extraction on the denoised time-domain waveform through stacked one-dimensional convolutional layers to obtain a time-series feature map. The global feature aggregation module performs global average pooling on the extracted time-series feature map to obtain a fixed-length global feature vector. The nonlinear regression module performs nonlinear mapping on the global feature vector to directly output the continuous leaf tip arrival time offset prediction values.

[0014] The present invention also proposes a computer device for extracting the leaf tip arrival time, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the leaf tip arrival time extraction method.

[0015] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the leaf tip arrival time extraction method.

[0016] This invention provides a method for extracting the arrival time of leaf tips, which has the following beneficial effects: This invention proposes an adaptive wavelet denoising network, whose adaptive decomposition module, nonlinear sparse thresholding module, and adaptive reconstruction module explicitly simulate the three-stage physical process of "decomposition-thresholding-reconstruction" in classical wavelet analysis. Multi-layer convolutional kernels and downsampling operations physically correspond to wavelet basis functions and multi-scale decomposition; the learnable threshold parameter physically corresponds to the noise truncation threshold; and transposed convolution physically corresponds to the inverse wavelet transform. Through the synergistic adaptation of parameterized soft thresholding and the learnable wavelet basis, optimal separation of noise and signal is achieved. The global feature aggregation module employs a global average pooling layer for deep feature extraction. The mean of the complete time-series feature map output by the block is calculated along the time dimension to completely eliminate the interference of waveform translation and amplitude fluctuation on time estimation. This method automatically learns the optimal wavelet basis function that matches the tip pulse shape by constructing an adaptive wavelet denoising network with embedded learnable convolutional kernels and learnable thresholds. At the same time, it automatically optimizes the best truncation threshold that distinguishes the signal and noise coefficients. This mechanism completely eliminates the dependence on expert experience, making the denoising process completely data-driven and adaptive to waveform differences under different units, different speeds, and different sensor models, which significantly improves the versatility and generalization ability of the method. Attached Figure Description

[0017] Figure 1 This is a flowchart of the leaf tip timing arrival time extraction method based on adaptive wavelet basis denoising in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction of the hybrid data leaf tip timing training dataset in an embodiment of the present invention. Figure 3 This is a schematic diagram of the adaptive wavelet denoising network structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the leaf tip moment extraction network structure in an embodiment of the present invention; Figure 5 This is a hardware architecture diagram of the online deployment system in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] This invention proposes a method for extracting the timing arrival time of blade tips based on adaptive wavelet basis denoising. This method overcomes the subjectivity of manual basis selection and the lag of timing algorithms in traditional methods, and can achieve real-time and high-precision processing of massive blade tip pulse signals through online deployment at the edge, thereby providing high-quality data support for the health monitoring of rotating machinery blades.

[0020] The method specifically includes the following steps: S1. Based on a hybrid data-driven strategy that combines virtual and real data, a timing training dataset for leaf tips is constructed, which includes high-fidelity simulation ground truth and measured noise characteristics.

[0021] To obtain accurate labels required for neural network training, this invention first constructs an idealized blade tip timing signal generator in digital space based on the photoelectric conversion principle and rotor dynamics equations. This invention simulates a commonly used fiber optic blade tip timing sensor, assuming that the energy of the laser beam emitted by the sensor has a Gaussian distribution. When a rotating blade sweeps across the end face of the fiber optic probe, the reflected light intensity... This can be described as the convolution integral of the light intensity density function of the light spot with the effective reflective area of ​​the leaf tip in the time domain. The mathematical model is constructed as follows: The thickness of the leaf tip is set to... The blade linear velocity is (in For rotational speed, (where the blade tip radius is 1), and the corresponding spot diameter is the numerical aperture of the optical fiber. Simulated signal The simulation is a Gaussian-like pulse or a hyperbolic secant pulse, and its mathematical expression is defined as follows:

[0022] ; In the formula, The signal amplitude, The theoretical truth value for the time of leaf tip arrival. The pulse width parameter is inversely proportional to the blade rotation speed and thickness. t It is a time variable.

[0023] To enable the network to recognize temporal changes containing vibration information, the blade vibration equation is explicitly introduced into the simulation model. It is assumed that the blade primarily undergoes first-order bending vibration, and its vibration displacement... It obeys the laws of simple harmonic motion:

[0024] ; In the formula, This represents the amplitude of blade vibration. The frequency of blade vibration. The phase angle is the initial position of the blade.

[0025] At this point, the true value of the actual blade tip arrival time. From the moment of rigid body rotation and additional time difference caused by vibration composition: ; During the data generation process, vibration frequencies are randomly selected. (Range 50Hz~500Hz) and vibration amplitude (Range 0.05mm~2.0mm), generating a massive amount of pure pulse signals with different vibration states and their corresponding precise time tags. .

[0026] Networks trained solely on simulated signals exhibit extremely poor generalization ability when faced with complex industrial environments. This invention collects real background noise using a physical test bench to construct a "measured noise feature library." To obtain pure "environmental noise" and "system noise," this invention employs the following two strategies for extraction:

[0027] (1) Static background noise acquisition: The signal is continuously acquired for a period of time while the motor is stationary and the laser is on. This part mainly includes the dark current noise of the photodetector, the thermal noise of the amplifier circuit, and the power frequency interference.

[0028] (2) Dynamic flow field and scattering noise acquisition: The drive rotor rotates at the operating speed, but the signal segment between two adjacent blades is captured. Since the sensor is facing the shaft surface or gap at this time, theoretically there should be no strong pulse signal. The signal acquired in this stage includes airflow disturbance noise caused by the high speed of rotor rotation, baseline drift caused by casing vibration, and diffuse reflection speckle noise caused by blade surface roughness.

[0029] The long-sequence noise signals collected above were truncated and classified to establish a noise library containing three typical operating conditions: (1) Low speed operation: electronic noise dominated by Gaussian white noise.

[0030] (2) Medium speed operating conditions: mechanical interference noise including periodic baseline fluctuations.

[0031] (3) High-speed operation: aerodynamic / electromagnetic coupling noise including non-stationary impact components.

[0032] Based on the above-mentioned "simulated signal" and "measured noise", this invention designs an automated data synthesis pipeline to generate the final training dataset through a random mixing strategy.

[0033] To enable neural networks to extract features from weak signals, the signal-to-noise ratio distribution is strictly controlled during the synthesis process. The synthesized signal is defined as follows: for:

[0034] ; in These are noise segments randomly selected from a measured noise database, with coefficients... Based on the target signal-to-noise ratio Dynamic adjustment: ; In the formula, For simulation signals power, For actual noise The power.

[0035] The invention sets Follows uniform distribution This covers everything from extremely harsh conditions where “the signal is almost drowned out by noise” to ideal conditions where “the signal is clear”, forcing the adaptive wavelet network to learn the deep signal structure rather than simple amplitude features.

[0036] To further improve the robustness of the model and prevent overfitting, the following data augmentation operations were performed: (1) Amplitude scaling: Randomly adjust the amplitude of the simulated pulse (0.5×~1.5×) to simulate the decrease in sensitivity caused by sensor contamination.

[0037] (2) Time shift: Randomly shift the pulse position left and right within the sampling window to ensure that the extraction network is not sensitive to the position of the pulse within the window.

[0038] (3) Waveform distortion simulation: By superimposing an asymmetric skew factor on a standard Gaussian pulse, the waveform asymmetry caused by blade leading edge wear or installation angle deviation is simulated.

[0039] After signal synthesis, the data undergoes standardization preprocessing to adapt to the input requirements of the neural network. To accelerate network convergence, each synthesized waveform is preprocessed. Perform Z-Score standardization:

[0040] ; in and These are the mean and standard deviation of the signal within the sample window, respectively.

[0041] The final training set is stored in the form of tensor pairs, in the format of ,in, It is a one-dimensional time series vector with dimension 1. This includes targeting denoising networks labeled as clean simulated waveforms. Extracting network labels as scalar time-to-time truth values ​​for specific time points. .

[0042] This invention generates a total of 10,000 training samples, 2,000 validation samples, and 1,000 test samples.

[0043] S2. Through an adaptive wavelet denoising network with embedded sparsity constraints, the frequency domain difference between noise and signal is automatically learned without the need for preset basis functions, and a high signal-to-noise ratio waveform is output.

[0044] In this invention, the adaptive wavelet denoising network employs a one-dimensional convolutional neural network, structurally simulating the classic signal processing flow of "wavelet decomposition—threshold truncation—wavelet reconstruction," as shown in the neural network structure below. Figure 3 As shown, it mainly includes three functional modules: an adaptive decomposition module, a nonlinear sparse threshold module, and an adaptive reconstruction module.

[0045] The adaptive decomposition module in this invention simulates multi-layer wavelet decomposition, employing a multi-layer stacked convolutional structure. The first convolutional layer is followed by downsampling to halve the signal length and extract high-frequency details; subsequent layers continue to convolve and downsample the low-frequency components, thereby capturing signal features at different time scales.

[0046] The encoder's first layer is set to a one-dimensional convolutional layer, containing... There are n convolutional kernels, and the kernel length is 1. For the input signal , No. Output feature map of each channel The calculation is as follows: ; in, For the first The weight vector of each convolutional kernel. The convolutional kernel is the "wavelet basis function". Unlike the fixed basis functions of traditional wavelets, the weights of the convolutional kernels in this layer are adaptively updated during training to match the specific shape of the blade pulse.

[0047] In this invention, a parameterized soft threshold activation function is designed for the nonlinear sparse threshold module. The characteristic coefficients output by the decomposition layer (Feature map matrix) A point-by-point nonlinear transformation is performed on a certain element in the data to obtain sparse features. : ; In traditional methods, Typically determined by the Donoho general threshold formula This was obtained through static calculation. However, in this invention, It is defined as a learnable network parameter. The network learns an independent threshold for each feature channel. During backpropagation, the network automatically finds the optimal cutoff point based on the loss function:

[0048] when When the coefficient is less than the threshold, it is judged as noise. .

[0049] when When the coefficient is greater than the threshold, it is determined to be a signal. .

[0050] In this invention, the adaptive reconstruction module uses one-dimensional transposed convolution to simulate the inverse wavelet transform, reconstructing sparse features back into time-domain waveforms. This layer performs upsampling and convolution summation operations:

[0051] ; in, These are sparse features after thresholding. It involves reconstructing the convolutional kernel. Similarly, It is also automatically learned and updated during training. This ensures that the reconstruction process can smoothly fill in signal details and eliminate the "Gibbs phenomenon" or artifacts that may be caused by traditional hard thresholding denoising.

[0052] In order to train the above network parameters ( , , This invention constructs a composite loss function that includes physical constraints. : ; Where: Reconstruction fidelity loss To measure network output using mean square error With the ideal noise-free label Differences between them: ; In the formula, For batch size during training, The time dimension length of the sampling window. To reconstruct a clean output waveform, The label for an ideal noise-free signal.

[0053] sparsity-induced loss For the feature tensor output of the threshold layer Apply L1 norm penalty: ; In the formula, The number of convolutional feature channels. These are the feature tensor elements output by the threshold layer.

[0054] The hybrid dataset constructed in step 1 is divided into training, validation, and test sets in a 7:2:1 ratio. The AdamW optimizer is used for parameter iteration, with an initial learning rate set to... The learning rate is decayed using a cosine annealing strategy. During training, the signal-to-noise ratio (SNR) of the input data is dynamically changed. High SNR data is initially used to allow the network to converge quickly and learn the basic shape of the pulse; subsequently, the SNR is gradually reduced to -5dB, forcing the network to optimize the threshold parameters under strong noise conditions. and convolution kernel This results in extremely strong robustness.

[0055] S3. Input the denoised clean waveform into the end-to-end regression network, and directly calculate the leaf tip arrival time through global feature aggregation and nonlinear mapping.

[0056] The end-to-end deep regression network in this invention mainly consists of three functional modules: a deep feature extraction module, a global feature aggregation module, and a nonlinear regression module. Its network structure is as follows: Figure 4 As shown. Its input is the clean waveform signal output from step S2, and the output is the precise time offset of the leaf tip pulse relative to the starting point of the sampling window.

[0057] In this invention, the deep feature extraction module is responsible for deconstructing the time-domain waveform into a high-dimensional feature map, employing a 3-4 layer stacked one-dimensional convolutional neural network structure. Let the first layer be... The input feature map of the convolutional block is (wherein, the input feature map of the first layer) This is the network output of step 2. ), No. The weight tensors of each convolutional kernel are denoted as... , bias is Then the first intermediate variables of convolutional layer output The Each channel at time The value is calculated as follows:

[0058] ; In the formula, The total number of channels in the input feature map. The length of the convolution kernel. This is the stride of the convolution kernel.

[0059] To accelerate convergence and improve generalization ability, intermediate variables... Normalize: ; In the formula, Let j be the mean value of the j-th feature channel in the current training batch. Let be the variance of the j-th feature channel in the current training batch. A learnable scaling factor. A learnable offset factor. It is a very small constant.

[0060] Final output feature map Generated via the ReLU function: ; In this invention, the global feature aggregation module uses a global average pooling layer to solve the problem of excessive parameters and sensitivity to waveform translation in traditional fully connected layers. This operation maps the feature map containing temporal information into a fixed-length global feature descriptor, realizing the transformation from "temporal space" to "semantic space".

[0061] Let the input feature tensor be For the first Each feature channel, and its aggregated eigenvalues for: ; in, In terms of time dimension, For the c-th channel in the feature tensor output by the last convolutional layer t The value at any given moment.

[0062] In this invention, the nonlinear regression module converts the feature vector The input consists of a multilayer perceptron composed of fully connected layers, with intermediate layers incorporating a Dropout mechanism (dropout rate of 0.2) to prevent overfitting. The final layer contains only a single neuron, directly outputting a continuous scalar.

[0063] Suppose the regression network contains two fully connected layers with weights respectively. , , bias is , Final prediction time The calculation formula is: ; ; Finally, the absolute arrival time is obtained. : ; In the formula, This represents the total number of signal sampling points within the sampling window. This represents the system's sampling frequency.

[0064] This invention uses Huber Loss as the regression loss function. Its definition is as follows: ; In the formula, This represents the total number of samples in the current training batch.

[0065] The single-sample loss is a piecewise function: .

[0066] In the formula, The absolute error between the predicted value and the true value. This is a hyperparameter used to switch between squared error and linear error.

[0067] S4. Based on model quantization and edge computing acceleration technology, the trained network model is deployed on an embedded monitoring device to realize online real-time inference of the algorithm.

[0068] This invention employs a strategy of "offline lightweight compilation + online heterogeneous acceleration" to construct an embedded edge computing monitoring device based on an FPGA+ARM architecture. This device achieves a fully closed-loop processing from raw sensor signal input to millisecond-level monitoring result output. Its hardware architecture is as follows: Figure 5 As shown.

[0069] This invention utilizes a quantization toolchain and a dynamic range calibration algorithm to map FP32 format model parameters to 8-bit fixed-point integer (INT8) format. During this process, KL divergence analysis is performed on the validation set to determine the optimal quantization threshold, ensuring that the quantized fixed-point model, while reducing its size to 1 / 4 of the original, maintains inference accuracy loss within acceptable engineering error ranges (e.g., less than 0.5%).

[0070] To eliminate compatibility barriers between deep learning training frameworks and underlying hardware inference engines, this invention exports the quantized model in the Open Neural Network Exchange (ONNX) universal format. ONNX, as a standard intermediate representation layer, can fully preserve the topology and weight information of the neural network, providing a standardized interface for subsequent cross-platform deployment.

[0071] Using a deep learning compiler optimized for the target FPGA chip, the ONNX model is compiled into a hardware-executable instruction stream or bitstream file. At this stage, the compiler performs deep graph optimization, including operator fusion (e.g., merging convolutional layers, batch normalization layers, and activation layers into a single hardware computation block to reduce intermediate data read / write operations) and instruction pipeline scheduling (optimizing the data stream loading order based on the distribution of DSP slices and on-chip memory resources within the FPGA to maximize the utilization of hardware parallel computing units).

[0072] The embedded online monitoring device designed in this invention adopts an FPGA+ARM heterogeneous system-on-chip architecture. The hardware core of the device is divided into two independent but tightly coupled subsystems: the first is a programmable logic unit (FPGA), which serves as the underlying parallel acceleration unit of the system. It directly interfaces with the raw signal interface of the turbine sensor and utilizes the FPGA's superior streaming processing capabilities to directly perform high-load operations such as convolution and pooling on the high-speed sampled raw signal stream without CPU intervention. The second is an application processing unit (ARM), which communicates with the FPGA through the on-chip high-speed AXI bus, enabling the feature data or intermediate results calculated by the FPGA to be mapped to the ARM's memory space in a "zero-copy" manner.

[0073] Once the device is deployed at the turbine monitoring site and started, the system enters online real-time inference mode: First, the light intensity pulse signal captured by the fiber optic sensor installed on the turbine casing is digitized by a high-speed A / D converter and directly injected into the on-chip buffer of the FPGA; then, the hardware execution engine inside the FPGA is triggered, loading the offline-compiled adaptive wavelet denoising network and time extraction network instructions. The adaptive wavelet layer filters out environmental noise in real time, the full waveform regression layer quickly extracts features, and the FPGA outputs a normalized predicted value. Finally, the ARM processor reads the FPGA output from shared memory. Mapping the dimensionless prediction ratio to the actual physical tip arrival time. After the calculation is completed, the ARM processor packages the monitoring results with absolute timestamps and uploads them to the host server in real time.

[0074] The present invention has the following advantages: By employing a hybrid driving strategy of "high-fidelity simulated pulses + measured industrial noise," the model is exposed to real non-stationary noise characteristics during the training phase. Compared to traditional hard or soft thresholding denoising methods, this invention can maintain waveform integrity even under extreme conditions with low signal-to-noise ratios (e.g., SNR < 5dB), without the severe signal amplitude reduction or phase drift caused by traditional filtering.

[0075] This invention breaks free from the limitations of traditional "zero-crossing detection" or "peak alignment" methods that rely on waveform symmetry. The adaptive wavelet basis can automatically capture the deep physical characteristics of the pulse, and even when the waveform undergoes asymmetric distortion or large amplitude fluctuations, it can still accurately calculate the pulse center position through nonlinear regression, significantly improving the accuracy of subsequent vibration displacement calculations.

[0076] The convolutional layer designed in this invention has the ability to adaptively evolve into the optimal wavelet basis. This means that the algorithm does not require tedious manual parameter tuning for different blade shapes or sensor types. The network can automatically adjust the internal parameters of "decomposition-thresholding-reconstruction" according to the characteristics of the input signal, achieving intelligent adaptation to "one machine, one strategy".

[0077] By introducing INT8 quantization and FPGA+ARM heterogeneous computing technology, this invention successfully solves the pain points of high computational cost and difficulty in online application of deep learning models. While ensuring an accuracy loss of less than 0.5%, the model size is reduced to 1 / 4 of its original size, achieving millisecond-level single inference response.

[0078] Based on the same inventive concept, this invention also proposes a leaf tip arrival time extraction system, comprising: The acquisition module is used to collect the raw pulse signals from the blade tips during the operation of rotating machinery.

[0079] The reconstruction module is used to input the original pulse signal into a pre-trained adaptive wavelet denoising network. The adaptive decomposition module uses multi-layer convolution kernels and downsampling operations to decompose the input signal at multiple scales to generate feature maps. The parameterized soft thresholding activation function in the nonlinear sparse thresholding module performs a point-by-point nonlinear transformation on the decomposed feature maps to obtain sparse features. The adaptive reconstruction module uses one-dimensional transposed convolution operations to simulate the inverse wavelet transform, upsampling and summing the sparse features to reconstruct the denoised time-domain waveform.

[0080] The output module is used to input the denoised time-domain waveform into the pre-trained end-to-end time regression network. The deep feature extraction module performs deep feature extraction on the denoised time-domain waveform through stacked one-dimensional convolutional layers to obtain a time-series feature map. The global feature aggregation module performs global average pooling on the extracted time-series feature map to obtain a fixed-length global feature vector. The nonlinear regression module performs nonlinear mapping on the global feature vector to directly output the continuous leaf tip arrival time offset prediction values.

[0081] The present invention also proposes a computer device for extracting the leaf tip arrival time, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the leaf tip arrival time extraction method.

[0082] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the leaf tip arrival time extraction method.

[0083] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting the arrival time of a leaf tip, characterized in that, Includes the following steps: Acquire raw pulse signals from the blade tips during the operation of rotating machinery; The original pulse signal is input into a pre-trained adaptive wavelet denoising network. The adaptive decomposition module uses multi-layer convolutional kernels and downsampling operations to decompose the input signal at multiple scales to generate feature maps. The decomposed feature map is subjected to a point-by-point nonlinear transformation by the parameterized soft threshold activation function in the nonlinear sparse threshold module to obtain sparse features. The adaptive reconstruction module uses one-dimensional transposed convolution to simulate wavelet inverse transform, upsampling and summing sparse features to reconstruct the denoised time-domain waveform. The denoised time-domain waveform is input into a pre-trained end-to-end time regression network. Deep feature extraction is performed on the denoised time-domain waveform through stacked one-dimensional convolutional layers in the deep feature extraction module to obtain a time-series feature map. The extracted time-series feature map is then subjected to global average pooling by the global feature aggregation module to obtain a fixed-length global feature vector. The global feature vector is then nonlinearly mapped by the nonlinear regression module to directly output the continuous leaf tip arrival time offset prediction values.

2. The method for extracting the leaf tip arrival time according to claim 1, characterized in that, The decomposed feature map is subjected to a point-by-point nonlinear transformation using the parameterized soft thresholding activation function in the nonlinear sparse thresholding module to obtain sparse features. The transformation process is specifically represented as follows: ； in, The decomposed feature map, , The kernel length is 1. For the first The weight vector of each convolutional kernel. For network parameters, An independent threshold is learned for each feature channel.

3. The method for extracting the leaf tip arrival time according to claim 1, characterized in that, The process of using one-dimensional transposed convolution to simulate wavelet inverse transform, upsampling and summing sparse features to reconstruct the denoised time-domain waveform is specifically represented as follows: ； in, These are sparse features after thresholding. It involves reconstructing the convolutional kernel.

4. The method for extracting the leaf tip arrival time according to claim 1, characterized in that, The training process of the adaptive denoising network specifically includes the following steps: The simulated blade tip pulse signal with arrival time label is obtained during the operation of rotating machinery, and the measured noise library is constructed based on static and dynamic background noise in industrial field. A mixed training dataset containing noisy signal samples, corresponding clean signal labels and arrival time true value labels is generated. Using noisy signal samples as input to a one-dimensional convolutional neural network and corresponding clean signal labels as training targets, the reconstruction error between the network output and the clean signal labels is minimized. By jointly applying sparsity constraints on the intermediate coefficients of the network, the convolutional kernel weights in the network are adaptively updated to approximate the optimal wavelet basis function and the threshold parameters are adaptively updated to filter out noise coefficients, resulting in a high signal-to-noise ratio waveform.

5. The method for extracting the leaf tip arrival time according to claim 4, characterized in that, The process of constructing the hybrid training dataset specifically includes the following steps: Based on the photoelectric sensing principle and blade vibration model, a simulated pulse signal with a precise arrival time label is generated; Through static and dynamic experiments, measured background noise of rotating machinery under different operating conditions was collected, and a measured noise database was constructed. The simulated pulse signal is mixed with a noise segment randomly selected from the measured noise library according to a preset signal-to-noise ratio distribution to generate a noisy synthetic signal sample. The synthesized signal samples are subjected to data augmentation operations including amplitude scaling, time shifting, and waveform distortion simulation, and are then subjected to standardization preprocessing to finally form a training dataset containing the input signal and its corresponding labels. The labels include a clean simulated signal for training the denoising network and an accurate arrival time value for training the regression network.

6. The method for extracting the leaf tip arrival time according to claim 1, characterized in that, It also includes using a composite loss function that incorporates physical constraints. The parameters of the adaptive wavelet denoising network are trained; the composite loss function Represented as: ； Among them, reconstruction fidelity loss To measure network output using mean square error With the ideal noise-free label Differences between them; sparsity-induced loss For the feature tensor output of the threshold layer Apply L1 norm penalty.

7. The method for extracting the leaf tip arrival time according to claim 1, characterized in that, The training process of the end-to-end regression network specifically includes: using a high signal-to-noise ratio waveform as the input of the deep regression network and the arrival time ground truth label as the target, training the network to learn the mapping from waveform features to arrival time ground truth, so as to output the predicted value of the leaf tip arrival time, thereby completing the training of the leaf tip arrival time extraction model.

8. A system for extracting the arrival time of a leaf tip, characterized in that, include: The acquisition module is used to collect the raw pulse signals from the blade tips during the operation of rotating machinery; The reconstruction module is used to input the original pulse signal into a pre-trained adaptive wavelet denoising network. The adaptive decomposition module uses multi-layer convolutional kernels and downsampling operations to decompose the input signal at multiple scales to generate feature maps. The decomposed feature map is subjected to a point-by-point nonlinear transformation by the parameterized soft threshold activation function in the nonlinear sparse threshold module to obtain sparse features. The adaptive reconstruction module uses one-dimensional transposed convolution to simulate wavelet inverse transform, upsampling and summing sparse features to reconstruct the denoised time-domain waveform. The output module is used to input the denoised time-domain waveform into the pre-trained end-to-end time regression network. The deep feature extraction module performs deep feature extraction on the denoised time-domain waveform through stacked one-dimensional convolutional layers to obtain a time-series feature map. The global feature aggregation module performs global average pooling on the extracted time-series feature map to obtain a fixed-length global feature vector. The nonlinear regression module performs nonlinear mapping on the global feature vector to directly output the continuous leaf tip arrival time offset prediction values.

9. A computer device for extracting the arrival time of leaf tips, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the leaf tip arrival time extraction method according to any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, are used to perform the steps of the leaf tip arrival time extraction method according to any one of claims 1-7.

Images