Ship key equipment vibration signal data enhancement method, system, medium and terminal

Abstract

The application provides a ship key equipment vibration signal data enhancement method, system, medium and terminal, comprising: collecting original vibration signals of a ship key equipment under different operating states and different working conditions and preprocessing; performing fast Fourier transform on the preprocessed original vibration signals to decompose the signals into low-frequency structural components and high-frequency detail components; constructing a frequency domain enhancement module to differentially process the low-frequency structural components and the high-frequency detail components to obtain low-frequency branch features and high-frequency branch features respectively; constructing a conditional diffusion model, inputting the low-frequency branch features, the high-frequency branch features and working condition information into the pre-constructed conditional diffusion model, and generating an enhanced vibration signal through a reverse denoising process. The application has the beneficial effects of excellent physical fidelity, strong small sample generation capability and interpretable generation process.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance and signal processing technology for ship equipment, and in particular to a method for enhancing vibration signal data of key ship equipment, which is especially applicable to the generation of vibration signals and expansion of fault diagnosis samples for key equipment such as ship diesel engines, main lubricating oil pumps, oil separators, and turbines under data scarcity conditions. Background Technology

[0002] Key shipboard equipment (such as diesel engines, main lubricating pumps, turbines, etc.) are core power units that ensure the safety of navigation and operational efficiency. In the complex marine environment, these devices are subjected to multiple harsh factors such as alternating loads, salt spray corrosion, and temperature shocks over long periods, making them prone to impact-type failures such as bearing wear, gear tooth breakage, and piston knocking. These failures are sudden and highly destructive; if not detected and diagnosed in time, they can lead to equipment downtime or even loss of ship control, causing significant economic losses and safety risks. Therefore, developing high-precision condition monitoring and fault diagnosis technologies for key shipboard equipment is of significant engineering importance.

[0003] Vibration signal-based intelligent fault diagnosis methods have become a mainstream research direction due to their sensitivity to changes in the internal state of equipment and their non-invasive measurement capabilities. However, the performance of these methods is highly dependent on sufficient and balanced labeled fault data. In actual ship operation and maintenance scenarios, acquiring a large amount of vibration data covering different fault types, severity levels, and operating conditions faces many difficulties: First, fault samples are scarce, as equipment is in normal operation most of the time; second, full life-cycle data acquisition is costly and poses safety risks to fault reproduction; third, the data distribution varies greatly among different ships, speeds, and loads, resulting in insufficient generalization ability of the model across operating conditions. This "small sample dilemma" makes many advanced deep learning models (such as convolutional neural networks and Transformer models) prone to overfitting during training, making it difficult to achieve satisfactory diagnostic results in practical engineering.

[0004] To address the problem of data scarcity, data augmentation techniques have been widely adopted. Traditional data augmentation methods, such as adding Gaussian noise, time shifting, amplitude scaling, and spectral amplitude modulation, while simple to implement, essentially involve limited perturbation within the neighborhood of the original data, failing to generate truly diverse new samples. More importantly, these operations often disrupt the inherent physical structure of vibration signals, such as disrupting the integer multiple relationship between the fundamental frequency and harmonics, distorting the envelope shape of the impact response, and disrupting the frequency spacing of the modulation sidebands, leading to "distorted" augmented data that may mislead diagnostic models.

[0005] In recent years, Generative Adversarial Networks (GANs) have been explored for generating mechanical vibration signals. GANs, through an adversarial game between the generator and discriminator, can produce relatively realistic samples. However, GANs suffer from inherent problems such as training instability and mode collapse, which are particularly pronounced in industrial scenarios with small sample sizes, making it difficult to guarantee the diversity of generated samples. Furthermore, signals generated by GANs often exhibit physical inconsistencies in their frequency domain characteristics, such as anomalous harmonic energy distribution and missing spectral lines, limiting their practical application in safety-critical fields.

[0006] The diffusion model, as an emerging deep generative model, demonstrates superior fidelity and versatility in image and audio generation by progressively adding noise to data and then learning an inverse denoising process, opening up new avenues for vibration signal generation. However, directly applying the standard diffusion model to ship vibration signals still has the following significant shortcomings:

[0007] First, standard diffusion models typically operate in the raw time domain, failing to explicitly utilize the inherent frequency domain structure information of vibration signals. Ship vibration signals exhibit clear physical laws; for example, energy is mainly concentrated near the fundamental frequency and its harmonics, and impact faults can excite resonances in specific frequency bands. Without guidance from frequency domain priors, diffusion models require a large amount of data to learn these laws, resulting in low generation efficiency and a tendency to deviate from physical constraints under small sample conditions.

[0008] Second, existing methods handle the high-frequency components of signals coarsely. The high-frequency components of vibration signals often contain detailed information about impact faults (such as the damped oscillations of transient impacts), but these high-frequency components are usually discarded or simply treated as noise. The key to improving the quality of generated signals is to preserve and utilize fault information in the high frequencies without increasing modeling complexity.

[0009] Third, there is a lack of effective control over operating conditions. The amplitude and frequency components of vibration signals from ship equipment vary significantly under different speeds and loads. Most existing data augmentation methods cannot generate signals under specific operating conditions as needed, resulting in a mismatch between the augmented data and the target operating conditions, which affects the adaptability of downstream diagnostic models.

[0010] The Fourier Neural Operator (FNO) is a neural operator that learns global mappings in the frequency domain and has been shown to have a natural advantage in modeling signals with physical laws. However, existing technologies have not yet provided an effective solution for how to organically integrate the frequency domain modeling capabilities of FNO with the powerful generative capabilities of diffusion models, while simultaneously achieving differentiated processing of low-frequency structures and high-frequency details and introducing operating condition control.

[0011] In summary, developing a vibration signal data enhancement method for key ship equipment that can maintain physical constraints, achieve controllable operating conditions, and is applicable to small sample conditions is of urgent need and great significance for solving the data scarcity problem in actual operation and maintenance and improving the engineering implementation capability of intelligent fault diagnosis technology. Summary of the Invention

[0012] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a method, system, medium and terminal for enhancing vibration signal data of key ship equipment, in order to solve the technical problems existing in the existing methods for enhancing vibration signal data of key ship equipment, such as poor physical fidelity, low generation quality under small sample conditions, insufficient utilization of frequency domain information, and uncontrollable and uninterpretable generation process.

[0013] To achieve the above and other related objectives, the first aspect of this application provides a method for enhancing vibration signal data of key ship equipment, comprising: acquiring raw vibration signals of key ship equipment under different operating states and conditions and preprocessing them; performing a fast Fourier transform on the preprocessed raw vibration signals to decompose the signals into low-frequency structural components and high-frequency detail components; constructing a frequency domain enhancement module to differentiate the low-frequency structural components and high-frequency detail components to obtain low-frequency branch features and high-frequency branch features respectively; constructing a conditional diffusion model, inputting the low-frequency branch features, high-frequency branch features and operating condition information into the pre-constructed conditional diffusion model, and generating enhanced vibration signals through a reverse denoising process.

[0014] In some embodiments of the first aspect of this application, the method of processing low-frequency structural components by constructing a frequency domain enhancement module includes: retaining several low-frequency components at the beginning of the sequence after Fourier transform, and using Fourier neural operators to extract features from the low-frequency components.

[0015] In some embodiments of the first aspect of this application, low-frequency branch features are obtained by linearly transforming the retained low-frequency components using a learnable complex weight matrix.

[0016] In some embodiments of the first aspect of this application, the method of processing high-frequency detail components by constructing a frequency domain enhancement module includes: calculating the amplitude spectrum and phase spectrum of the remaining high-frequency components that are not retained, and then concatenating the amplitude spectrum and phase spectrum and inputting them into a multilayer perceptron for dimensionality reduction processing to obtain high-frequency branch features.

[0017] In some embodiments of the first aspect of this application, the construction process of the conditional diffusion model includes: building a generator framework based on the denoising diffusion probability model and constructing a noise prediction network as the core execution unit; wherein the denoising diffusion probability model includes a forward diffusion process and a reverse denoising process; and the noise prediction network is used to predict noise components in the reverse denoising process.

[0018] In some embodiments of the first aspect of this application, the noise prediction network is composed of multiple stacked Fourier neural operator blocks, each Fourier neural operator block including: a frequency domain convolutional layer, a time domain convolutional layer with a kernel size of 1, an instance normalization layer, a Dropout layer, and a GELU activation function; the frequency domain convolutional layer and the time domain convolutional layer receive the same input feature in parallel, and the outputs after processing the two paths are fused element-wise; the fused feature is input to the instance normalization layer; the normalized feature is regularized by the Dropout layer and then fed into the GELU activation function for nonlinear transformation.

[0019] In some embodiments of the first aspect of this application, after the conditional diffusion model is constructed, the frequency domain enhancement module is embedded in the noise prediction network so that the denoising diffusion probability model and the noise prediction network embedded with the frequency domain enhancement module form an end-to-end model; in the end-to-end model, the low-frequency branch features, high-frequency branch features, corresponding operating condition information, noisy signals and time step information are used together as conditional guidance information input into the noise prediction network; wherein, before being input into the noise prediction network, the low-frequency branch features and high-frequency branch features are spliced ​​along the frequency domain dimension to obtain enhanced frequency domain features.

[0020] In some embodiments of the first aspect of this application, the end-to-end model is trained by predicting the mean square error loss function between the noise and the actual noise; the trained end-to-end model is then used to iteratively perform a reverse denoising process to recover the enhanced vibration signal.

[0021] To achieve the above and other related objectives, a second aspect of this application provides a vibration signal data enhancement system for key ship equipment, comprising: a data acquisition module for acquiring and preprocessing raw vibration signals of key ship equipment under different operating states and conditions; a frequency domain decomposition module for performing a fast Fourier transform on the preprocessed raw vibration signal to decompose the signal into low-frequency structural components and high-frequency detail components; a frequency domain enhancement construction module for differentially processing the low-frequency structural components and high-frequency detail components by constructing a frequency domain enhancement module to obtain low-frequency branch features and high-frequency branch features respectively; and a diffusion model construction module for constructing a conditional diffusion model, inputting the low-frequency branch features, high-frequency branch features, and operating condition information into the pre-constructed conditional diffusion model, and generating enhanced vibration signals through a reverse denoising process.

[0022] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for enhancing vibration signal data of key ship equipment.

[0023] To achieve the above and other related objectives, a fourth aspect of this application provides a computer program product comprising computer program code that, when executed on a computer, enables the computer to implement the method for enhancing vibration signal data of key ship equipment.

[0024] To achieve the above and other related objectives, a fifth aspect of this application provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the vibration signal data enhancement method for key ship equipment.

[0025] As described above, the vibration signal data enhancement method, system, medium, and terminal for key ship equipment of this application have the following beneficial effects:

[0026] (1) Excellent physical fidelity: By decoupling the physical structure (low-frequency periodic components) of the vibration signal from random details (high-frequency impact / noise) through frequency domain decomposition, and by using Fourier neural operators to explicitly model low-frequency harmonic relationships with clear physical meaning, the generated signal is fundamentally consistent with the real signal in terms of spectral structure, harmonic energy distribution, and impact envelope morphology.

[0027] (2) Strong small sample generation capability: High-frequency parameterized encoding compresses high-dimensional high-frequency information into low-dimensional feature vectors, which greatly reduces the modeling complexity of the diffusion model; at the same time, the diffusion model itself has a stable training process and superior pattern coverage capability. The combination of the two enables the present invention to generate a large amount of diverse and high-fidelity augmented data even with only a small number of real samples (dozens of each class), effectively solving the "small sample dilemma" of downstream diagnostic tasks.

[0028] (3) The generation process is interpretable: The use of frequency domain decomposition and FNO makes the model no longer a black box. Researchers can analyze the frequency domain weight matrix learned by the low-frequency branch (reflecting the model's attention to the fundamental frequency and each harmonic) and the output distribution of the high-frequency branch encoder separately, so as to understand how the model captures and reconstructs fault features and provide a physical explanation for the diagnostic results. Attached Figure Description

[0029] Figure 1 The diagram shown is a flowchart illustrating a method for enhancing vibration signal data of critical ship equipment according to an embodiment of this application.

[0030] Figure 2 The diagram shown is a schematic representation of the internal structure of a frequency domain enhancement module in one embodiment of this application.

[0031] Figure 3 The diagram shown is a schematic representation of a multilayer sensor in one embodiment of this application.

[0032] Figure 4 The diagram shown is a schematic representation of the structure of a noise prediction network in one embodiment of this application.

[0033] Figure 5 This is a time-domain comparison diagram of the original real signal and the model-generated signal in one embodiment of this application.

[0034] Figure 6 The image shown is a comparison of the frequency domain power spectra of the original real signal and the model-generated signal in one embodiment of this application.

[0035] Figure 7 The diagram shown is a structural schematic of a vibration signal data enhancement system for key ship equipment according to an embodiment of this application.

[0036] Figure 8 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of this application. Detailed Implementation

[0037] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0038] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" do not necessarily imply that they are different.

[0039] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0040] This application aims to address the following technical problems existing in current methods for enhancing vibration signal data of critical ship equipment:

[0041] (1) Poor physical fidelity: Traditional data augmentation methods (such as adding noise, time-domain cropping, etc.) and signals generated by generative adversarial networks often destroy the inherent frequency domain physical laws of vibration signals (such as the integer multiple relationship between fundamental frequency and harmonics, the envelope shape of impact response, the frequency interval of modulation sidebands, etc.), resulting in "distorted" generated data and misleading downstream diagnostic models.

[0042] (2) Low quality of generated samples under small sample conditions: In actual ship operation and maintenance, fault samples are extremely scarce. Existing generative models (especially GANs) are prone to pattern collapse or overfitting when trained with small samples, and cannot generate sufficiently diverse and high-fidelity augmented samples, making it difficult to effectively alleviate the data scarcity dilemma.

[0043] (3) Insufficient utilization of frequency domain information: Existing diffusion models mostly operate in the time domain and fail to explicitly utilize the frequency domain structure prior of the vibration signal; at the same time, the high-frequency components containing details of the impact fault are processed roughly (usually directly discarded or simply denoised), resulting in the generated signal lacking transient impact characteristics.

[0044] (4) The generation process is uncontrollable and unexplainable: It is impossible to generate signals corresponding to the user-specified operating parameters (such as speed and load) as needed, and the generation process is a black box, making it difficult to verify whether the generated signals conform to the physical mechanism, which reduces the credibility in safety-critical areas.

[0045] Furthermore, it is worth noting that the technical solutions proposed in this application have good versatility and practicality, and can effectively adapt to the operation monitoring needs of various key power equipment, transmission equipment and precision electromechanical equipment in ship navigation and maintenance scenarios. They can be widely used in data augmentation-related work such as vibration signal acquisition optimization, feature mining and sample expansion of key ship equipment, helping to improve the training effect and actual generalization ability of equipment fault identification, status judgment and health assessment models.

[0046] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 This document illustrates a flowchart of a method for enhancing vibration signal data of critical ship equipment according to an embodiment of the present invention. The method in this embodiment mainly includes the following steps:

[0047] Step S11: Collect raw vibration signals of key ship equipment under different operating conditions and working conditions and perform preprocessing.

[0048] In the embodiments of this application, all kinds of core operating machinery in the ship's navigation and power system are the focus of vibration monitoring. Key ship equipment includes, but is not limited to, core operating components such as diesel engine cylinder head, bearing housing, main lubricating oil pump housing, main engine crankcase, generator set base, reduction gearbox housing, stern shaft support, air compressor body, hydraulic pump station base, ventilator and cooling water pump housing, etc., which are all key monitoring points that are prone to mechanical wear and abnormal operation.

[0049] In the embodiments of this application, the operating status of key ship equipment includes normal and stable operating status and fault status. Fault status includes typical fault modes such as bearing wear and spalling, gear tooth breakage and pitting, shaft misalignment, rotor imbalance, loose anchor bolts, air gap eccentricity, and oil circuit blockage. The monitoring conditions of key ship equipment include different rated speed ranges of the main engine, full-range load changes, ship navigation and port entry / exit speed changes, no-load and full-load operation, engine start / stop switching, and continuous long-term steady-state operation, etc., and the collection of vibration signals for the entire scenario is completed by relying on multi-dimensional data acquisition.

[0050] In the embodiments of this application, the original vibration signals x(t) under different operating states and working conditions are collected by vibration sensors installed on key equipment of the ship. The original vibration signals x(t) are then detrended and standardized, and divided into discrete samples x[n] of fixed length L, n=0,1,…,L-1.

[0051] The entire processing procedure for the original vibration signal is as follows: First, the acquired original vibration signal is detrended and filtered for correction. Then, a global data standardization numerical transformation is performed. Finally, the preprocessed stationary signal is segmented into ordered segments according to a preset fixed length L, resulting in a discrete sample sequence x[n] with sequence numbers from 0 to L-1, thus completing the sample dataset regularization construction. It should be noted that detrending processing removes low-frequency linear or nonlinear trend components in the original vibration signal caused by equipment temperature rise, gradual changes in operating conditions, and baseline drift, eliminating overall signal offset interference and retaining the true fault characteristic fluctuation information. Standardization processing performs mean-zeroing and variance-normalization conversion on the signal amplitude, unifying the data distribution range of vibration data of different magnitudes and measurement points, and eliminating the influence of differences in dimensions and amplitudes.

[0052] Step S12: Perform a fast Fourier transform on the preprocessed original vibration signal to decompose the signal into low-frequency structural components and high-frequency detail components.

[0053] The Fast Fourier Transform (FFT) is a highly efficient discrete Fourier transform algorithm that can rapidly convert time-domain signals to frequency-domain signals, significantly reducing computational complexity and is widely used for signal feature decomposition. Specifically, performing a Fast Fourier Transform on each time-domain sample yields the following frequency-domain representation:

[0054] ; Formula (1)

[0055] ; Formula (2)

[0056] Where x[n] represents the discrete time-domain sample; L represents the length of each sample (number of sampling points); X[k] represents the frequency domain signal obtained after FFT; This represents the complex exponential basis function, which projects a time-domain signal onto a sine / cosine basis of different frequencies to achieve time-frequency conversion; Indicates the sampling frequency. Indicates frequency resolution. This indicates the actual frequency corresponding to the frequency point number k.

[0057] It should be noted that low-frequency structural components correspond to the portion with smaller k values ​​(close to k=0). These components represent the overall trend of slow changes in the signal, the fundamental frequency operating characteristics of the equipment, and the main frequency components of major faults; they are the skeleton information of the signal. High-frequency detail components correspond to the portion with larger k values ​​(close to k=L / 2). These components represent the details of rapidly changing signals, impulse noise, and the sidebands and harmonic components of components such as bearings / gears; they are the detailed information of the signal.

[0058] Step S13: By constructing a frequency domain enhancement module, the low-frequency structural components and high-frequency detail components are differentiated to obtain low-frequency branch features and high-frequency branch features respectively.

[0059] Figure 2 The diagram illustrates the internal structure of the frequency domain enhancement module in this embodiment. First, the original time-domain signal is converted into a frequency-domain representation through a Fast Fourier Transform (FFT). Then, through a "mode-segmentation" operation, the frequency-domain signal is divided into two independent branches: low-frequency and high-frequency. The processing procedures for the low-frequency and high-frequency branches will be described in detail below.

[0060] For low-frequency structural components, several low-frequency components at the beginning of the sequence after Fourier transform are retained, and Fourier neural operators are used to extract features from the low-frequency components.

[0061] Specifically, for the low-frequency processing branch, the first M low-frequency components X are retained. low [k] = X[k], k = 0, ..., M−1. Fourier neural operator (FNO) is used to extract features from the retained low-frequency components, preserving the main periodic structure of the signal, i.e., through a learnable complex weight matrix. The following linear transformation is performed on the retained low-frequency components:

[0062] ;Formula (3)

[0063] in, , These represent the number of input channels and the number of output channels, respectively; M represents the number of low-frequency components retained. The transformed output features have a dimension of . , where b represents the batch size; The instructions represent the Einstein summation convention and define the operational rules for two input tensors; This indicates the submapping relationship of operations, representing the multiplication and summation operations of two tensors according to their corresponding subscripts; The input low-frequency feature tensor has a dimension of . ; Represents a learnable complex weight matrix with dimension . Each element in the matrix is ​​a trainable complex parameter.

[0064] It should be noted that the learnable complex weight matrix , is the core trainable parameter in the low-frequency processing branch of this Fourier neural operator, and its dimension is determined by the number of input channels. Number of output channels The number of low-frequency components retained, M, is jointly determined; each element in the matrix is ​​a complex number, as described in this invention. , This means it can map low-frequency features from a single channel to high-dimensional features from 64 channels. During model training, the elements of this matrix are continuously updated through backpropagation to learn the effective transformation methods of low-frequency components in the vibration signal, ultimately achieving modeling of the signal's periodicity and long-range dependency structure.

[0065] Furthermore, it is worth noting that processing low-frequency structural components is equivalent to performing global convolution on low-frequency components in the Fourier domain, which can efficiently capture periodic and long-range dependent structures in vibration signals. Traditional temporal convolution, limited by fixed-size local convolution kernels, can only capture local dependencies in signals, making it difficult to efficiently model long-range correlations across multiple periods in vibration signals. It also cannot directly model key features such as the periodicity and harmonic components of the signal. In contrast, global convolution in the Fourier domain, by applying a learnable complex weight matrix transformation to the retained low-frequency components in the frequency domain, essentially achieves global interactive modeling of all frequency components of the signal. It can capture long-range dependencies across the time domain in a single operation without stacking multiple layers of local convolution. Furthermore, the low-frequency components naturally correspond to the fundamental frequency and main harmonic components of the vibration signal. Directly performing linear transformations on these components in the frequency domain is equivalent to targeted enhancement and feature mapping of the signal's periodicity. This enables efficient extraction of periodic patterns and trend features crucial for tasks such as fault diagnosis and condition monitoring, significantly improving the model's ability to represent key information in vibration signals and its computational efficiency. It also provides a more physically meaningful and globally informative feature foundation for subsequent tasks.

[0066] For the high-frequency detail components, the amplitude spectrum and phase spectrum are calculated for the remaining high-frequency components. The amplitude spectrum and phase spectrum are then concatenated and input into a multilayer perceptron for dimensionality reduction processing to compress the high-frequency detail information into a low-dimensional feature vector, thus obtaining the high-frequency branch features.

[0067] Specifically, for the remaining high-frequency component X high [k]=X[k], k=M, ..., [L / 2] Calculate the amplitude spectrum A[k]=|X high [k]| and phase spectrum The two are then concatenated and input into a multilayer perceptron for dimensionality reduction processing as follows:

[0068] ;Formula (4)

[0069] in, The high-frequency components of the input are all the high-frequency components remaining after removing the first M low-frequency components following the Fourier transform; |X high[k]| denotes the complex form of X high [k] is used for modulo operation, and each A[k] corresponds to the amplitude of a frequency component, reflecting the energy strength of the vibration signal at that frequency. ∠X high [k] denotes the complex form of X high [k] performs phase angle calculation, each It represents the phase information of the corresponding frequency component, reflecting the temporal position and phase shift of the vibration signal at that frequency. This indicates the operation of splicing the amplitude spectrum and the phase spectrum. This represents a learnable multilayer perceptron. Represents all trainable parameters (weights and biases) of the network, and its function is to perform dimensionality reduction and nonlinear transformation on the spliced ​​high-frequency features. These are the output features of the high-frequency processing branch, which are low-dimensional feature vectors after dimensionality reduction by a multilayer perceptron. Their dimension is typically the same as the output of the low-frequency branch. Alignment is necessary for subsequent feature fusion.

[0070] It should be noted here that concatenating the amplitude spectrum and phase spectrum involves merging two one-dimensional sequences of the same shape along the feature dimension or channel dimension to form a new two-dimensional feature. Taking a common tensor shape as an example: if the shape of the amplitude spectrum A is (b, N) (where b is the batch size and N is the number of high-frequency points), the phase spectrum... The shape is also (b, N), and after splicing, a feature vector with the shape of (b, 2, N) or (b, 2N) will be obtained. The former is spliced ​​according to the channel dimension, and the latter is directly flattened and spliced ​​according to the feature dimension. Both methods can be applied to the embodiments of this application.

[0071] Furthermore, the reason for concatenating the amplitude and phase spectra is due to the need for information integrity and the efficiency requirements of dimensionality reduction and feature learning. The need for information integrity involves high-frequency components in complex form. The signal contains both amplitude and phase information. Amplitude reflects the energy of the vibration signal at the corresponding frequency, while phase determines how different frequency components are superimposed in the time domain, directly affecting the waveform details. Inputting only the amplitude spectrum will lose phase information, making it impossible to fully reconstruct the characteristics of high-frequency components; however, directly inputting the complex number form... This increases computational complexity, and since most multilayer perceptron frameworks process real-valued data by default, extracting amplitude and phase as real numbers before concatenation satisfies the information integrity requirement. Regarding the efficiency requirements of dimensionality reduction and feature learning, high-frequency components are numerous and have high dimensionality, making direct processing extremely computationally costly. Concatenating amplitude and phase before inputting them into a multilayer perceptron allows, on the one hand, the network's nonlinear transformations to learn the correlation patterns between amplitude and phase, uncovering hidden fault features in high-frequency components (such as phase shifts caused by weak impacts or early damage); on the other hand, the dimensionality reduction function of the multilayer perceptron significantly compresses the dimensionality of high-frequency features, matching it with the output dimension of low-frequency branches, while filtering out redundant information from high-frequency noise, providing more efficient and robust feature representations for subsequent feature fusion and modeling tasks.

[0072] In the embodiments of this application, the structure of the multilayer perceptron is as follows: Figure 3 As shown, a two-layer fully connected structure is adopted. First, the spliced ​​amplitude spectrum and phase spectrum features are used as input (input dimension In_dim). The first linear layer (Linear1) maps its dimension to Cout / 2, completing the initial dimension compression and feature extraction. Then, a nonlinear transformation is introduced through the ReLU activation function to enhance the model's ability to fit complex patterns in high-frequency impact faults. Next, the second linear layer (Linear2) maps the feature dimension to Cout, which is consistent with the output of the low-frequency branch, to achieve the final dimension alignment and feature integration. The multilayer perceptron transforms the originally high-dimensional high-frequency detail information into low-dimensional feature vectors through layer-by-layer compression, effectively controlling the number of model parameters and computational complexity, and avoiding the redundant overhead caused by directly processing high-frequency components. Furthermore, through nonlinear activation and dimension mapping, the key high-frequency detail features corresponding to impact faults in the vibration signal are accurately preserved, achieving effective modeling of fault information with low computational cost. At the same time, it provides dimension-matched and information-complementary feature representations for subsequent fusion with low-frequency branch features.

[0073] Step S14: Construct a conditional diffusion model. Input the low-frequency branch features, high-frequency branch features, and operating condition information into the pre-constructed conditional diffusion model, and generate an enhanced vibration signal through a reverse denoising process.

[0074] In the embodiments of this application, the Conditional Diffusion Model adds constraints such as text, tags, and images to the basic diffusion model to achieve directional and controllable generation and accurate output of specified content. The construction process of the Conditional Diffusion Model involves first building a generator framework based on the Denoising Diffusion Probability Model (DDPM), and then constructing a noise prediction network as the core execution unit. The Denoising Diffusion Probability Model (DDPM) defines the underlying operating logic of the entire diffusion model, including the gradual addition of noise to the original signal during forward diffusion and the overall process constraints for the reverse denoising process. The noise prediction network is the execution unit, based on a stacked Fourier neural operator structure, responsible for accurately predicting noise components during the reverse process.

[0075] The Denoising Diffusion Probability Model (DDPM) consists of two parts: a forward diffusion process and a reverse denoising process. During the forward diffusion process, the model progressively adds Gaussian noise to the original signal, and its transition probability distribution is defined by the following formula:

[0076] ;Formula (5)

[0077] Where t=1,…,T,T=1000 (T is the diffusion time step); β t For linear noise scheduling, it is represented as:

[0078] ;Formula (6)

[0079] in, Indicates the initial noise intensity. Indicating the termination noise intensity, in the embodiments of this application... , A smooth transition from clear signal to pure noise is achieved through linearly increasing noise scheduling.

[0080] Noise prediction network Structure such as Figure 4 As shown, noise prediction network It consists of four stacked Fourier Neural Operator Blocks (FNO Blocks), each integrating multi-dimensional feature processing capabilities in the frequency and time domains. Specifically, each FNO Block includes a frequency domain convolutional layer (SpectralConv1d), a time domain convolutional layer (Conv1d) with a kernel size of 1, an instance normalization layer (InstanceNorm), a Dropout layer, and a GELU activation function. The input features are first processed synchronously in two parallel branches: the frequency domain convolutional layer... With temporal 1×1 convolutional layer Receiving the same input feature x, the system performs frequency domain global periodic feature extraction, time domain local detail feature extraction, and channel dimension transformation. The processed feature tensors are then fused element-wise to achieve complementary integration of global long-range dependencies and local detail information. The fused features are then input into an instance normalization layer to standardize the single-sample feature distribution and stabilize the network training process. The normalized features then pass through a Dropout layer to randomly deactivate some neurons, thereby suppressing overfitting and improving generalization ability. Finally, the features processed by Dropout regularization are fed into the GELU activation function, which fits complex fault feature patterns through nonlinear transformation, ultimately outputting the complete feature result of the FNO block. The entire block constitutes a hierarchical connection logic of parallel feature extraction, feature addition and fusion, layer-by-layer sequential normalization, regularization, and nonlinear activation. Therefore, the output of the FNO block is represented as:

[0081] ;Formula (7)

[0082] in, K represents the frequency domain transformation operation performed by the Fourier neural operator on the input feature x, and K is the Fourier neural operator (FNO) in the aforementioned step S13. This indicates that a one-dimensional temporal convolution operation is performed on the input feature x, with a kernel size of 1. It is mainly used to perform local linear transformation and channel mapping on the feature in the temporal domain to extract local detail information, which complements the frequency domain transformation. represents the instance normalization layer, used to normalize the summed feature vectors; GELU represents the activation function.

[0083] After constructing the conditional diffusion model, the frequency domain enhancement module constructed in step S13 is embedded into the noise prediction network to achieve end-to-end joint optimization of the entire conditional diffusion model. Therefore, in addition to the noisy signal... In addition to time step information t, low-frequency branch features High-frequency branching characteristics The corresponding operating condition information, along with other relevant information, are input into the noise prediction network as condition-guided information. middle.

[0084] Specifically, firstly, the low-frequency branch features and high-frequency branching characteristics By concatenating along the frequency domain dimension, the enhanced frequency domain feature Y is obtained. combined Subsequently, the enhanced frequency domain features and operating condition information (such as load, speed, equipment type, etc.) are encoded through an embedding layer and combined with the noisy signal. The time step information t is input into the noise prediction network. The noise prediction network predicts the noise component of the current time step by stacking Fourier neural operator blocks and combining condition guidance information. Finally, it realizes controllable signal generation based on signal characteristics and operating condition constraints, ensuring that the generated result matches the frequency domain characteristics of the original signal and conforms to the vibration mode under the specified operating conditions.

[0085] In the specific implementation process, low-frequency branch characteristics will be used. and high-frequency branching characteristics By concatenating along the frequency domain dimension, the enhanced frequency domain feature Y is obtained. combined [k], and then perform inverse Fourier transform to recover the time-domain enhanced signal:

[0086] ;Formula (8)

[0087] in, Y represents the time-domain enhanced signal after inverse Fourier transform; combined [k] represents the enhanced frequency domain feature obtained by concatenating the low-frequency branch feature and the high-frequency branch feature; The complex exponential basis functions represent the inverse Fourier transform, which map the complex components in the frequency domain to the sine / cosine components in the time domain, thus reconstructing the signal from the frequency domain to the time domain. It should be understood that the inverse fast fourier transform (IFFT) is the reverse operation of the FFT, restoring the processed frequency domain data to the original time domain signal, achieving signal restoration after frequency domain editing.

[0088] Furthermore, sampling is performed from the real data x0, and time steps t and noise are randomly selected. Generate noisy signals The training objective is to minimize the mean squared error loss between the predicted noise and the actual noise.

[0089] Noisy signal Represented as:

[0090] ;Formula (8)

[0091] in ) indicates forward-spreading noise scheduling The calculated cumulative coefficient is used to control the proportion of the original signal and noise in the noisy sample. It is the noisy signal at step t; The noise dispatch coefficient representing forward diffusion is the noise intensity added in step s, from... =10 -4 linear growth to =0.02, which controls the noise growth rate throughout the noise addition process. This represents standard Gaussian noise, which follows a distribution. .

[0092] The mean square error loss function between predicted noise and actual noise is expressed as:

[0093] ;Formula (9)

[0094] in, It is the objective function that needs to be minimized during model training, measuring the error between the prediction results of the noise prediction network and the real noise. It is the mathematical expectation operator, which means taking the average of the error term within the parentheses over the joint distribution of the three variables, i.e., traversing all true samples. Real noise The average error is calculated at the diffusion time step t to ensure that the model can learn stably in all cases. This represents the output of the noise prediction network. This represents all trainable parameters of the network. This is the mean square error term, used to calculate the true noise. and network prediction noise The smaller the error, the more accurate the network prediction, which is the square of the Euclidean distance between them.

[0095] In some examples, the Adam optimizer is used during the end-to-end joint training of the model, with a learning rate of 10. -4 With a batch size of 32 and 100 training rounds, the network parameters are updated through backpropagation, enabling the model to gradually acquire the ability to accurately predict noise and complete signal denoising under the guidance of frequency domain enhancement features.

[0096] After training, from standard Gaussian noise Begin by iteratively executing the reverse denoising process:

[0097] ;Formula (10)

[0098] The clear time-domain vibration signal x^0 is gradually recovered; that is, the clear time-domain vibration signal finally recovered by the reverse denoising process is an enhanced sample consistent with the original data distribution, which can be used to expand the fault diagnosis training set. Multiple sampling can obtain a large number of enhanced samples consistent with the original data distribution, which can be used to expand the fault diagnosis training set. This represents the signal generated in step t-1 during the reverse denoising process; This represents the signal generated in step t during the reverse denoising process; This represents the scaling factor for the signal term, used to scale the current noisy signal. Scale adjustments are made to offset the signal attenuation caused by noise scheduling during forward diffusion; This represents the noise dispatch coefficient at step t during the forward diffusion process; This represents the cumulative coefficient during the forward diffusion process; This represents the output of the noise prediction network, which uses the current noisy signal as an example. Using time step t as input, the noise component contained in the signal is predicted; This represents the scaling factor for the random noise term.

[0099] The foregoing has provided a detailed explanation of the implementation process and principle of the vibration signal data enhancement method for key ship equipment in the embodiments of this application. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only for explaining the invention and do not constitute a limitation on the scope of protection of this invention.

[0100] This embodiment uses the normal vibration signal from a university bearing dataset as the raw data and employs the method of this invention for data augmentation. The specific process is as follows.

[0101] Step 1: Vibration signal acquisition and preprocessing.

[0102] The experimental data came from the university bearing dataset, selecting normal-state vibration signals (filenamed normal_0.mat) collected by the accelerometer at the drive end. The sampling frequency was fs = 12000 Hz. Each original signal sample was divided into fixed-length L = 1000 points, resulting in 500 samples. Each sample was detrended and standardized to ensure a mean of 0 and a variance of 1.

[0103] Step 2, frequency domain decomposition.

[0104] For each preprocessed time-domain sample Where K = [L / 2] + 1 = 501. Frequency resolution Δf = f s / L=12Hz. Set the number of low-frequency modes M=100 (corresponding cutoff frequency fc=M×Δf=1200Hz). Divide X along the frequency dimension into low-frequency components X. low and high-frequency part X high .

[0105] Step 3: Extraction of low-frequency components using Fourier neural operator (FNO) features.

[0106] The low-frequency component is achieved through a learnable complex weight matrix. Perform a linear transformation using Einstein's summation operation: The output shape is (B, 64, 100). This operation is the core of FNO, performing global convolution on low-frequency modes in the frequency domain to extract periodic structural features such as the fundamental frequency and harmonics of vibration signals. It should be noted that the Fourier Neural Operator (FNO) is a neural network operator built based on the Fourier transform, adept at fitting partial differential equations and efficiently capturing long-distance spatial physical field features.

[0107] Step 4: Parameterized encoding of high-frequency components.

[0108] For the high-frequency part X high Calculate A[k] = |X high [k]| and phase spectrum Both have a shape of (B, 1, 401). Concatenating them along their last dimension yields a feature vector of shape (B, 1, 802). This vector is then input into a two-layer multilayer perceptron: Layer 1: a linear layer with an 802-dimensional input and a 32-dimensional output, activated by ReLU; Layer 2: a linear layer with a 32-dimensional input and a 64-dimensional output. The output is the high-frequency feature Y. high The shape is (B, 64, 401). Among them, the Multi-Layer Perceptron (MLP) is a basic fully connected feedforward neural network, which is composed of multiple layers of neurons and can perform general nonlinear data fitting and classification / regression tasks.

[0109] Step 5: Feature fusion and inverse transformation.

[0110] Low-frequency output Y low With high frequency output Y high By splicing along the frequency dimension, we obtain the combined frequency domain representation Y. combined The shape is (B, 64, 501). Perform an inverse real-part Fourier transform with parameters n = L = 1000 to restore the time-domain signal with shape (B, 64, L). This output is the result of a single SpectralConv1d module.

[0111] Step 6: Stack multiple FNO Blocks.

[0112] In this embodiment, the noise prediction network The model consists of four identical Fourier Neural Operator (FNO) blocks. Each FNO block includes a frequency-domain convolutional layer (SpectralConv1d), a time-domain convolutional layer (Conv1d) with a kernel size of 1, and the sum of the outputs of the frequency-domain and time-domain convolutional layers; an instance normalization layer; a GELU activation function; and Dropout (with a dropout rate of 0.1). The input to the entire model is a noisy signal x. tAt time step t, after dimensionality upscaling through the input projection layer (Conv1d(1, 64, kernel_size=1)), it passes through four FNO blocks sequentially, and finally is reduced to the original channel number through the output projection layer (Conv1d(64,64,1) + GELU + Conv1d(64,1,1)), outputting the prediction noise. .

[0113] Step 7: Diffusion process parameters and training.

[0114] The Denoising Diffusion Probabilistic Model (DDPM) framework is adopted. This model is a mainstream generative model that stably generates high-resolution realistic images, time series, and other high-quality data through a progressive noise addition and reverse denoising process. The total time step T = 1000. Noise scheduling is linear.

[0115] ;

[0116] in =10 -4 , =0.02.

[0117] definition .

[0118] During training, samples are taken from the real data x0, and time steps are randomly selected. and noise Generate noisy signals x t The predicted noise is obtained by inputting t into the model. Calculate the mean squared error loss: The AdamW optimizer is used, with a learning rate of... Weight decay The gradient pruning threshold is 1.0. The batch size is 32, and the number of training epochs is 100. The learning rate is scheduled using cosine annealing.

[0119] Step 8: Generate new samples.

[0120] After training, from standard Gaussian noise Begin by executing the reverse denoising process. : Set the current signal x t Input the time step t into the model to obtain the predicted noise. Update as follows:

[0121] ;

[0122] When t=1, z=0.

[0123] The final result is x0, which is the generated vibration signal. By sampling independently multiple times, any number of enhanced samples can be obtained.

[0124] Step 9: Verify the implementation effect.

[0125] To verify the effectiveness of this invention, 500 new samples were generated using the trained model. The generated signals were compared with the original real signals in both the time and frequency domains, and the results are shown in the attached figure.

[0126] In the time domain, as shown in the appendix Figure 5 (Time-domain signal comparison diagram) As shown, the upper diagram is the original real signal, and the lower diagram is the signal generated by the model. The horizontal axis represents time (sampling points), and the vertical axis represents amplitude. From the time-domain waveforms, it can be seen that the generated signal is highly consistent with the original signal in terms of amplitude range, fluctuation trend, and impulse characteristics, indicating that this invention can maintain the time-domain dynamic characteristics of the signal. In the frequency domain, as shown in the attached diagram... Figure 6 As shown in the (frequency domain power spectrum comparison diagram), the horizontal axis is frequency (Hz) and the vertical axis is power spectral density. This invention effectively preserves the frequency domain structure of the vibration signal and the spectral characteristics of the impact component, proving that the generated signal is highly similar to the real signal in frequency domain distribution.

[0127] Figure 7 This is a schematic block diagram of a vibration signal data enhancement system for critical ship equipment provided in an embodiment of this application. Figure 7 As shown, the system includes a data acquisition module 701, a frequency domain decomposition module 702, a frequency domain enhancement construction module 703, and a diffusion model construction module 704.

[0128] The acquisition module 701 is used to acquire and preprocess the raw vibration signals of key ship equipment under different operating states and conditions. The frequency domain decomposition module 702 performs a Fast Fourier Transform on the preprocessed raw vibration signals to decompose them into low-frequency structural components and high-frequency detail components. The frequency domain enhancement module 703 differentiates the low-frequency structural components and high-frequency detail components by constructing a frequency domain enhancement module to obtain low-frequency branch features and high-frequency branch features respectively. The diffusion model construction module 704 constructs a conditional diffusion model, inputting the low-frequency branch features, high-frequency branch features, and operating condition information into the pre-constructed conditional diffusion model, and generating enhanced vibration signals through a reverse denoising process.

[0129] It should be understood that the specific processes by which each module performs the corresponding steps described above have been detailed in the above method embodiments, and will not be repeated here for the sake of brevity. It should also be understood that the module division in the embodiments of this application is illustrative and merely a logical functional division; other division methods may exist in actual implementation. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or have two or more modules integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0130] Figure 8 This is a schematic block diagram of the electronic terminal provided in an embodiment of this application. Figure 8 As shown, the electronic terminal 800 includes at least one processor 801, a memory 802, at least one network interface 803, and a user interface 805. The various components in the electronic terminal 800 are coupled together via a bus system 804. It is understood that the bus system 804 is used to implement communication between these components. In addition to a data bus, the bus system 804 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 8 The general will label all buses as bus systems.

[0131] The user interface 805 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0132] It is understood that memory 802 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0133] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for enhancing vibration signal data of key ship equipment, characterized in that, include: Collect raw vibration signals of key ship equipment under different operating conditions and working conditions and perform preprocessing. The preprocessed original vibration signal is subjected to a fast Fourier transform to decompose the signal into low-frequency structural components and high-frequency detail components. By constructing a frequency domain enhancement module, low-frequency structural components and high-frequency detail components are processed differently to obtain low-frequency branch features and high-frequency branch features respectively. A conditional diffusion model is constructed by inputting the low-frequency branch features, high-frequency branch features, and operating condition information into the pre-constructed conditional diffusion model, and generating an enhanced vibration signal through a reverse denoising process.

2. The method for enhancing vibration signal data of key ship equipment according to claim 1, characterized in that, The methods for processing low-frequency structural components by constructing a frequency domain enhancement module include: retaining the first few low-frequency components of the sequence after Fourier transform, and using Fourier neural operators to extract features from the low-frequency components.

3. The method for enhancing vibration signal data of key ship equipment according to claim 1, characterized in that, The method of processing high-frequency detail components by constructing a frequency domain enhancement module includes: calculating the amplitude spectrum and phase spectrum of the remaining high-frequency components that are not retained, and then concatenating the amplitude spectrum and phase spectrum and inputting them into a multilayer perceptron for dimensionality reduction to obtain high-frequency branch features.

4. The method for enhancing vibration signal data of key ship equipment according to claim 1, characterized in that, The construction process of the conditional diffusion model includes: A generator framework based on a denoising diffusion probability model was built, and a noise prediction network was constructed as the core execution unit. The denoising diffusion probability model includes a forward diffusion process and a reverse denoising process; the noise prediction network is used to predict noise components during the reverse denoising process.

5. The method for enhancing vibration signal data of key ship equipment according to claim 4, characterized in that, The noise prediction network consists of multiple stacked Fourier neural operator blocks. Each Fourier neural operator block includes: a frequency domain convolutional layer, a time domain convolutional layer with a kernel size of 1, an instance normalization layer, a Dropout layer, and a GELU activation function. The frequency domain convolutional layer and the time domain convolutional layer receive the same input feature in parallel, and the outputs of the two processed paths are fused element-wise. The fused feature is input to the instance normalization layer. The normalized feature is regularized by the Dropout layer and then fed into the GELU activation function for nonlinear transformation.

6. The method for enhancing vibration signal data of key ship equipment according to claim 4, characterized in that: After constructing the conditional diffusion model, the frequency domain enhancement module is embedded in the noise prediction network so that the denoising diffusion probability model and the noise prediction network with the embedded frequency domain enhancement module form an end-to-end model. In the end-to-end model, the low-frequency branch features, high-frequency branch features, corresponding operating condition information, noisy signals, and time step information are used together as conditional guidance information input into the noise prediction network. Before being input into the noise prediction network, the low-frequency branch features and high-frequency branch features are concatenated along the frequency domain dimension to obtain enhanced frequency domain features.

7. The method for enhancing vibration signal data of key ship equipment according to claim 6, characterized in that, The end-to-end model is trained by using the mean square error loss function between the predicted noise and the actual noise; the trained end-to-end model is then used to iteratively perform a reverse denoising process to recover the enhanced vibration signal.

8. A vibration signal data enhancement system for key ship equipment, characterized in that, include: The data acquisition module is used to collect and preprocess the raw vibration signals of key ship equipment under different operating conditions and working conditions. The frequency domain decomposition module is used to perform a fast Fourier transform on the preprocessed original vibration signal to decompose the signal into low-frequency structural components and high-frequency detail components. The frequency domain enhancement module is used to differentiate the low-frequency structural components and high-frequency detail components by constructing a frequency domain enhancement module, so as to obtain low-frequency branch features and high-frequency branch features respectively. The diffusion model construction module is used to construct a conditional diffusion model. The low-frequency branch features, high-frequency branch features, and operating condition information are input into the pre-constructed conditional diffusion model, and an enhanced vibration signal is generated through a reverse denoising process.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the vibration signal data enhancement method for key ship equipment as described in any one of claims 1 to 7.

10. An electronic terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the vibration signal data enhancement method for critical ship equipment as described in any one of claims 1 to 7.

Images