A time-frequency representation method and device of a non-stationary signal and a storage medium

Abstract

The application discloses a kind of non-stationary signal time-frequency representation method, device and storage medium, belong to mechanical equipment signal processing field.Method includes: with the original time-frequency diagram generated by short-time Fourier transform as input, time-frequency diagram is decoupled into amplitude component and phase component by two-dimensional Fourier transform;Design phase correlation attention, amplitude modulation attention and hybrid attention, respectively capture instantaneous frequency structure, energy distribution and the correlation characteristics of both, relieve the time-frequency ambiguity caused by amplitude and phase entanglement;Multi-scale frequency domain-space domain joint learning is constructed, adaptive selection of different frequency band characteristics and dynamic calibration of amplitude-phase are realized;And through the cross-stage interaction of input time-frequency diagram and coding feature completes feature modulation and reconstruction.The time-frequency representation ridge line output by the application is clear, energy is gathered, noise suppression is significant, and can be effectively applied to non-stationary signal fault diagnosis and condition monitoring of rotating machinery key components such as rolling bearing.

Description

Technical Field

[0001] This invention relates to a time-frequency characterization method, apparatus, and storage medium for non-stationary signals, belonging to the field of signal processing technology for mechanical equipment. Background Technology

[0002] Rotating machinery typically operates under complex conditions, exhibiting non-stationary vibration signals. Rolling bearings, as critical load-bearing components, are prone to damage. Failure to detect and address faults promptly directly threatens the safe and stable operation of the equipment, potentially leading to major accidents. Therefore, efficient and accurate fault diagnosis and condition monitoring of critical components in rotating equipment are of significant practical importance. Time-frequency analysis effectively reveals the joint time and frequency domain distribution characteristics of signals, serving as a powerful tool for fault feature extraction and condition monitoring. However, traditional time-frequency analysis methods lack adaptive intelligent optimization mechanisms, and the resulting time-frequency representations are prone to energy divergence and feature ambiguity. Furthermore, while existing data-driven time-frequency methods possess some adaptive capabilities, their physical mechanisms are unclear, and their effectiveness in processing noisy cross-frequency signals is limited. Therefore, constructing a high-precision time-frequency analysis method that combines clear physical mechanisms with deep learning-driven principles is of great value for achieving efficient fault diagnosis of critical components in rotating machinery under complex operating conditions. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a time-frequency characterization method, device and storage medium for non-stationary signals. The output time-frequency characterization has clear ridges, concentrated energy and significant noise suppression, and can be effectively applied to fault diagnosis and condition monitoring of non-stationary signals of key components of rotating machinery such as rolling bearings.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a time-frequency characterization method for non-stationary signals, comprising:

[0006] Obtain raw vibration data of non-stationary signals;

[0007] Perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram;

[0008] The original time-frequency graph is input into a pre-trained Fourier domain decoupled time-frequency refinement network, and the following steps are performed:

[0009] The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features;

[0010] A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component;

[0011] The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features;

[0012] The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features;

[0013] Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map.

[0014] Furthermore, before inputting the original time-frequency graph into the pre-trained Fourier domain decoupled time-frequency refinement network, the following steps are also included:

[0015] The Fourier domain decoupled time-frequency refined network is trained by constructing a simulation dataset. The simulation dataset consists of data samples composed of sinusoidal signals, linear signals and quadratic signals. Noise is added to all data samples to simulate background interference.

[0016] Based on the instantaneous frequency theoretical formulas of the sinusoidal signal, linear signal, and quadratic signal, corresponding ideal time-frequency representations are generated as labels;

[0017] The original time-frequency graph obtained by short-time Fourier transform of the data samples is used as network input, and the ideal time-frequency representation is used as a supervision label to participate in network training.

[0018] Furthermore, the original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features, including:

[0019] The dimensions of the original time-frequency plot were unified to 256×256;

[0020] The time-frequency graph after unifying the dimensions is then processed by block embedding, with a block size of [missing value]. Embedding dimension is To obtain multi-scale embedding features The embedding process is represented as:

[0021] ;

[0022] in, For the first Time-frequency plot at scale Corresponding to the original scale, 1 / 2 scale, and 1 / 4 scale; This is the function for the block expansion operation; Implement dimension mapping for 1×1 convolutions; and Indicates the first The height and width of the time-frequency plot at the scale;

[0023] Perform a two-dimensional Fourier transform on the multi-scale embedded features to separate the amplitude components. and phase components :

[0024] ;

[0025] Where FFT(·) represents Fourier transform, e is the base of the natural logarithm, i is the imaginary unit, and arg represents finding its phase angle;

[0026] The amplitude component carries the energy intensity and distribution of the time-frequency ridge, while the phase component carries the structural location information of the time-frequency ridge.

[0027] Furthermore, the frequency domain decoupling attention mechanism is used to decouple and enhance the amplitude and phase components, resulting in frequency domain decoupling enhancement features, including:

[0028] The amplitude component and the phase component are respectively input into the frequency domain decoupling attention mechanism, which includes phase-correlated attention, amplitude modulation attention and hybrid attention.

[0029] The phase-related attention Using only the phase components, the ridge position-related features are captured through phase gradient calculation and self-attention weight allocation, as follows:

[0030] ;

[0031] Where iFFT(·) represents the inverse Fourier transform, This is the amplitude correlation value. It is a phase-related query value. This is key information regarding the phase.

[0032] The amplitude modulation attention Using only the amplitude component, the spatial distribution and intensity variation of ridge energy features are extracted by weighting the amplitude feature channels through a channel attention mechanism, and are expressed as follows:

[0033] ;

[0034] in, It is an amplitude-related query value. It represents the Hadamah accumulation. For amplitude-related key information, It is the phase correlation value;

[0035] The hybrid attention Simultaneously considering both phase and amplitude components, an interactive attention matrix for phase-amplitude features is constructed to achieve adaptive fusion of instantaneous frequency structure information and energy characteristics, expressed as:

[0036] ;

[0037] The outputs of the three types of attention features are concatenated and then convolved with a 1×1 convolution to obtain the frequency domain decoupling enhancement feature. :

[0038] ;

[0039] in, For feature splicing operations, Represents spatial characteristics.

[0040] Furthermore, the step of performing complementary learning on the frequency domain decoupling enhancement features through frequency branching and spatial branching to obtain multi-branch fusion features includes:

[0041] The frequency domain decoupling enhancement features are input into the frequency branch and the spatial branch, respectively;

[0042] The frequency branch uses learnable filters to perform frequency selection and shifting on the amplitude and phase of the feature, respectively, to obtain the frequency branch feature, represented as:

[0043] ;

[0044] ;

[0045] ;

[0046] in, For grouping and block operations, Representing scale Input features on and For learnable parameters, and These are the updated phase correlation features and amplitude correlation features, respectively. This indicates the calculation of its phase value. This is a group merging operation. Indicates frequency branching characteristics;

[0047] The spatial branch is used to learn local information to supplement the representation of frequency domain information, resulting in spatial branch features. ;

[0048] The frequency branch features and spatial branch features are fused to obtain multi-branch fused features. :

[0049] ;

[0050] in, This indicates a gating mechanism.

[0051] Furthermore, the step of performing feature modulation and calibration through cross-stage interaction between multi-branch fusion features and the original time-frequency map to output a high-resolution time-frequency map includes:

[0052] The multi-branch fusion features are interacted with the original time-frequency map across stages, and phase modulation, amplitude modulation and spatial modulation are executed sequentially.

[0053] The phase modulation is achieved by learning the original time-frequency diagram. The phase deviation from the encoded features is used to dynamically calibrate the encoded features, which is expressed as:

[0054] ;

[0055] in, For the updated phase value, For learnable weights of phase modulation;

[0056] The amplitude modulation is based on the amplitude distribution of the original time-frequency diagram, and the encoded amplitude features are adaptively adjusted pixel by pixel, as follows:

[0057] ;

[0058] in, For the updated amplitude, For learnable weights of amplitude modulation;

[0059] The frequency domain characteristics after phase and amplitude modulation are represented by the inverse transform:

[0060] ;

[0061] in, For the updated feature representation, Represents the normalization layer. Indicates original features;

[0062] The spatial modulation is used to achieve adaptive alignment of amplitude and phase representations, expressed as:

[0063] ;

[0064] in, This represents a 3×3 convolution. Representative scale is The output characteristics;

[0065] The features, after being modulated across stages, are reconstructed into the final high-resolution time-frequency map output.

[0066] Furthermore, the pre-trained Fourier domain decoupling time-frequency refinement network is trained by minimizing a total loss function, which includes pixel difference loss. and frequency loss The pixel difference loss is used to constrain the reconstruction of the time-frequency map. Compared with the actual time-frequency diagram The pixel difference is represented as:

[0067] ;

[0068] in, and Here are the height and width of the time-frequency plot; h and w are the height and width indices, respectively.

[0069] The frequency loss is used to constrain the consistency of the distribution of amplitude and phase in the Fourier frequency domain output, and is expressed as:

[0070] ;

[0071] The total loss function is defined as follows: ;

[0072] in, This is the loss weight.

[0073] Secondly, the present invention provides a time-frequency characterization apparatus for non-stationary signals, used to implement the time-frequency characterization method for non-stationary signals as described in any one of the preceding claims, comprising:

[0074] The acquisition module is used to acquire the raw vibration data of non-stationary signals;

[0075] The transformation module is used to perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram;

[0076] The input module is used to input the original time-frequency graph into a pre-trained Fourier domain decoupled time-frequency refinement network and perform the following steps:

[0077] The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features;

[0078] A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component;

[0079] The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features;

[0080] The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features;

[0081] Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map.

[0082] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0083] Fourthly, the present invention provides an electronic device, comprising:

[0084] Memory, used to store computer programs / instructions;

[0085] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0086] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0087] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0088] This invention provides a time-frequency characterization method, apparatus, and storage medium for non-stationary signals. By introducing two-dimensional Fourier transform into the network architecture, it effectively decouples the amplitude and phase components of the time-frequency graph, clarifying the physical meaning of amplitude carrying ridge energy and phase dominating ridge structure, thus providing a theoretical basis for time-frequency enhancement of non-stationary signals. By designing three mechanisms—phase-related attention, amplitude modulation attention, and hybrid attention—it captures the time-frequency ridge structure, energy distribution, and their correlation characteristics, respectively, alleviating the time-frequency ambiguity problem caused by amplitude and phase entanglement. By constructing multi-scale frequency-spatial domain joint learning, it achieves adaptive selection of features in different frequency bands and dynamic amplitude-phase calibration, ultimately achieving the technical effects of noise suppression, ridge energy focusing, and improved time-frequency resolution. Attached Figure Description

[0089] Figure 1 This is a flowchart of a time-frequency characterization method for non-stationary signals provided in an embodiment of the present invention;

[0090] Figure 2 These are time-domain waveforms and rotational speed diagrams of experimental signals provided in embodiments of the present invention, wherein: (a) is a time-domain waveform diagram, and (b) is a rotational speed change diagram;

[0091] Figure 3 This is a framework diagram of the non-stationary signal time-frequency characterization method based on Fourier domain decoupled time-frequency refinement network provided in the embodiments of the present invention;

[0092] Figure 4This is a visualization distribution diagram of the three types of attention provided in the embodiments of the present invention;

[0093] Figure 5 This is a schematic diagram comparing the refined time-frequency representation provided by the embodiment of the present invention with other methods, wherein: (a) is a schematic diagram of the refined time-frequency representation output by the method proposed in this embodiment, (b) is a schematic diagram of the time-frequency representation obtained by short-time Fourier transform, and (c) is a schematic diagram of the time-frequency representation obtained by generalized linear frequency modulation transform. Detailed Implementation

[0094] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0095] Example 1, such as Figure 1 As shown in the figure, this embodiment introduces a time-frequency characterization method for non-stationary signals, including:

[0096] Obtain raw vibration data of non-stationary signals;

[0097] Perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram;

[0098] The original time-frequency graph is input into a pre-trained Fourier domain decoupled time-frequency refinement network, and the following steps are performed:

[0099] The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features;

[0100] A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component;

[0101] The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features;

[0102] The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features;

[0103] Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map.

[0104] The invention will now be described in detail with reference to actual experimental data.

[0105] This embodiment uses the variable-speed bearing vibration dataset from the University of Ottawa to validate the time-frequency characterization method for non-stationary signals based on a Fourier domain decoupled time-frequency refined network. The dataset was collected on a SpectraQuest MFS-PK5M fault simulation test bench, driven by a variable-frequency motor and supported by two ER16K ball bearings. The left bearing was healthy, while the right bearing had an inner race fault. The experiment was set with a sampling frequency of 200kHz, the shaft speed increasing from 14.8Hz to 21.7Hz, and then decreasing to 13.6Hz. The inner race fault characteristic coefficient was 5.43. The original vibration signal was downsampled, with a final sampling frequency of 1250Hz and a duration of 2.5s. The time-domain waveform of the signal is shown below. Figure 2 As shown in (a), the rotational speed changes as follows Figure 2 As shown in (b).

[0106] like Figure 3 As shown, the time-frequency characterization method for non-stationary signals provided in this embodiment involves the following steps in its application:

[0107] Step 1: Construct a simulation dataset to train a Fourier domain decoupled time-frequency refined network.

[0108] In this embodiment, a sufficient simulation dataset is constructed, which consists of data samples composed of sinusoidal signals, linear signals and quadratic signals. All data samples are given a certain amount of noise to simulate background interference in the actual scene.

[0109] Simultaneously, based on the instantaneous frequency theoretical formulas of the sinusoidal signal, linear signal, and quadratic signal, the corresponding ideal time-frequency representations are directly generated as labels.

[0110] The original time-frequency graph obtained by short-time Fourier transform of the data samples is used as network input, and the ideal time-frequency representation is used as a supervision label to participate in network training.

[0111] Step 2: Obtain the raw vibration data of the non-stationary signal. In this embodiment, the raw vibration data is the bearing inner ring fault vibration signal collected from the above-mentioned experimental platform. Perform a short-time Fourier transform on the raw vibration data to generate the raw time-frequency diagram.

[0112] Step 3: Perform block embedding processing on the input time-frequency map to obtain multi-scale embedding features.

[0113] In this embodiment, the dimensions of the original time-frequency graph are unified to 256×256 to meet the network input requirements.

[0114] The model employs an asymmetric encoder-decoder architecture, performing block embedding processing on the input time-frequency representation, with a block size of [missing information]. Embedding dimension is To obtain multi-scale embedding features The embedding process is represented as:

[0115] (1);

[0116] in, For the first Time-frequency plot at scale Corresponding to the original scale, 1 / 2 scale, and 1 / 4 scale; This is the function for the block expansion operation; Implement dimension mapping for 1×1 convolutions; and Indicates the first The height and width of the time-frequency plot at scale.

[0117] Perform a two-dimensional Fourier transform on the multi-scale embedded features to separate the amplitude components. and phase components :

[0118] (2);

[0119] Where FFT(·) represents Fourier transform, e is the base of the natural logarithm, i is the imaginary unit, and arg represents finding its phase angle;

[0120] Among them, the amplitude component carries the energy intensity and distribution of the time-frequency ridge, and the phase component carries the structural location information of the time-frequency ridge.

[0121] Step 4: Decouple and enhance the amplitude component and phase component through a frequency domain decoupling attention mechanism to obtain frequency domain decoupling enhancement features.

[0122] In the Fourier domain, amplitude represents the magnitude of each frequency component in the time-frequency graph. The larger the amplitude, the higher the proportion of these frequency components in the time-frequency graph. Phase indicates the displacement of each frequency component.

[0123] This embodiment designs three attention mechanisms to capture key feature location and energy information. The amplitude and phase components are respectively input into the frequency domain decoupling attention mechanism, which includes phase-correlated attention, amplitude modulation attention, and hybrid attention.

[0124] Phase-related attention Using only the phase components, the ridge position-related features are captured through phase gradient calculation and self-attention weight allocation, as follows:

[0125] (3);

[0126] Where iFFT represents the inverse Fourier transform. This is the amplitude correlation value. It is a phase-related query value. This is key information regarding the phase. It contains only phase components, and and The amplitude is independent and is used to capture ridge position-related features.

[0127] Amplitude Modulation Attention Using only the amplitude component, the spatial distribution and intensity variation of ridge energy features are extracted by weighting the amplitude feature channels through a channel attention mechanism, and are expressed as follows:

[0128] (4);

[0129] in, It is an amplitude-related query value. It represents the Hadamah accumulation. For amplitude-related key information, It is the phase correlation value. It is used to capture amplitude characteristics while avoiding focusing on structural features; It can be viewed as a filter to selectively amplify or attenuate. A specific frequency range.

[0130] Mixed attention Simultaneously considering both phase and amplitude components, an interactive attention matrix for phase-amplitude features is constructed to achieve adaptive fusion of instantaneous frequency structure information and energy characteristics, expressed as:

[0131] (5);

[0132] in, Simultaneously considering phase and amplitude, it retains the structural constraints of the phase component while strengthening the energy characterization of the amplitude component.

[0133] The outputs of the three types of attention features are concatenated and then convolved with a 1×1 convolution to obtain the frequency domain decoupling enhancement feature. :

[0134] (6);

[0135] in, For feature splicing operations, Represents spatial characteristics.

[0136] This embodiment provides a visual analysis of three attention characteristics, such as... Figure 4 As shown, the three do not overlap significantly, indicating that they focus on different features, thus verifying the rationality and effectiveness of the Fourier domain amplitude-phase decoupling design. The frequency domain decoupling enhancement feature is then used as input for the next step.

[0137] Step 5: The frequency domain decoupling enhancement features are learned complementaryly through frequency branch and spatial branch to obtain multi-branch fusion features.

[0138] Ridge structure and energy information are represented as phase and amplitude at different frequency bands. After decoupling attention, frequency and spatial branches are designed to achieve complementary learning of frequency domain features and local details. The frequency domain decoupling enhancement features are then input into the frequency and spatial branches, respectively.

[0139] Frequency branching adaptively learns features from different frequency bands in the Fourier domain, matching the characteristic patterns of ridge structures and energy in different frequency bands. Learnable filters are used to perform frequency selection and shifting on the amplitude and phase of the features, respectively, to obtain the frequency branching features, expressed as:

[0140] (7);

[0141] (8);

[0142] (9);

[0143] in, For grouping and block operations, Representing scale Input features on and For learnable parameters, and These are the updated phase correlation features and amplitude correlation features, respectively. This indicates the calculation of its phase value. This is a group merging operation. This indicates frequency branching characteristics.

[0144] Spatial branching is used to learn local information to supplement the representation of frequency domain information. Spatial and frequency features are fused by addition, and a nonlinear transformation is introduced based on a gating mechanism to obtain spatial branching features. ;

[0145] The frequency branch features and spatial branch features are fused to obtain multi-branch fused features. :

[0146] (10);

[0147] in, This indicates a gating mechanism.

[0148] Step 6: Perform feature modulation and calibration through cross-stage interaction between multi-branch fusion features and the original time-frequency map to output a high-resolution time-frequency map.

[0149] The multi-branch fusion features are interacted with the original time-frequency graph across stages, and phase modulation, amplitude modulation and spatial modulation are executed in sequence to realize feature modulation of amplitude and phase and spatial adaptive weight allocation, thereby completing feature calibration and fusion.

[0150] Phase modulation is achieved by learning the original time-frequency diagram. With coding features The phase deviation is used to quantify the relative ambiguity between the two, and the encoded features are dynamically calibrated to reduce the impact of input ambiguity on feature representation, expressed as:

[0151] (11);

[0152] in, For the updated phase value, The learnable weights for phase modulation enable adaptive adjustment of phase deviation.

[0153] Amplitude modulation is based on the amplitude distribution of the original time-frequency map. The encoded amplitude features are adaptively adjusted pixel by pixel. By adaptively differentiating the energy feature intensity at different locations, the feature response of the high-discrimination ridge region is enhanced, as shown below:

[0154] (12);

[0155] in, For the updated amplitude, For learnable weights of amplitude modulation, This represents the Hadamard product, enabling pixel-by-pixel adaptive adjustment of the amplitude.

[0156] The frequency domain characteristics after phase and amplitude modulation are represented by the inverse transform:

[0157] (13);

[0158] in, For the updated feature representation, Represents the normalization layer. Indicates original features.

[0159] Spatial modulation is used to achieve adaptive alignment of amplitude and phase representations, expressed as:

[0160] (14);

[0161] in, This represents a 3×3 convolution. Representative scale is The output characteristics;

[0162] The features, after being modulated across stages, are reconstructed into the final high-resolution time-frequency map output.

[0163] Step 7: Construct a loss function to constrain the direction of network optimization.

[0164] In this embodiment, network training is performed by minimizing a total loss function, which includes pixel difference loss. and frequency loss .

[0165] Pixel difference loss is used to constrain the reconstruction of the time-frequency map. Compared with the actual time-frequency diagram The pixel difference is represented as:

[0166] (15);

[0167] in, and Here are the height and width of the time-frequency plot; h and w are the height and width indices, respectively.

[0168] The frequency loss is used to constrain the uniformity of the distribution of amplitude and phase in the Fourier frequency domain output, and is expressed as:

[0169] (16);

[0170] The total loss function is defined as: (17);

[0171] in, This is the loss weight.

[0172] In this embodiment, the enhancement of time-frequency representation in terms of ridge sharpness and noise suppression is achieved by minimizing the total loss function.

[0173] After multiple rounds of training, the model achieved optimal loss performance on the test set, demonstrating certain generalization ability and applicability to bearing vibration signal analysis. For example... Figure 5 As shown in (a), in the refined time-frequency characterization output by the method proposed in this embodiment, the energy of the three instantaneous frequency ridges is highly concentrated and the background noise is effectively suppressed; while Figure 5 The traditional short-time Fourier transform shown in (b) and Figure 5 The generalized linear frequency-modulated wavelet transform shown in (c) exhibits problems such as weak time-frequency resolution and insufficient anti-interference capability. The above results fully demonstrate the effectiveness and superiority of the method proposed in this invention.

[0174] This embodiment provides a time-frequency characterization method for non-stationary signals. By introducing two-dimensional Fourier transform into the network architecture, it effectively decouples the amplitude and phase components of the time-frequency graph, clarifies the physical meaning of amplitude carrying ridge energy and phase dominating ridge structure, and provides a theoretical basis for time-frequency enhancement of non-stationary signals. By designing three mechanisms—phase-related attention, amplitude modulation attention, and hybrid attention—it captures the time-frequency ridge structure, energy distribution, and their correlation characteristics, respectively, alleviating the time-frequency ambiguity problem caused by amplitude and phase entanglement. By constructing multi-scale frequency-spatial domain joint learning, it achieves adaptive selection of features in different frequency bands and dynamic amplitude-phase calibration, ultimately achieving the technical effects of suppressing noise, focusing ridge energy, and improving time-frequency resolution.

[0175] Example 2: This example provides a time-frequency characterization device for non-stationary signals, comprising:

[0176] The acquisition module is used to acquire the raw vibration data of non-stationary signals;

[0177] The transformation module is used to perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram;

[0178] The input module is used to input the original time-frequency graph into a pre-trained Fourier domain decoupled time-frequency refinement network and perform the following steps:

[0179] The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features;

[0180] A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component;

[0181] The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features;

[0182] The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features;

[0183] Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map.

[0184] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0185] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0186] Example 4: This example provides an electronic device, including:

[0187] Memory, used to store computer programs / instructions;

[0188] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0189] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.

[0190] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0191] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A time-frequency characterization method for non-stationary signals, characterized in that, include: Obtain raw vibration data of non-stationary signals; Perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram; The original time-frequency graph is input into a pre-trained Fourier domain decoupled time-frequency refinement network, and the following steps are performed: The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features; A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component; The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features; The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features; Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map. Before inputting the original time-frequency graph into the pre-trained Fourier domain decoupled time-frequency refinement network, the process also includes: The Fourier domain decoupled time-frequency refined network is trained by constructing a simulation dataset. The simulation dataset consists of data samples composed of sinusoidal signals, linear signals, and quadratic signals. Noise is added to all data samples to simulate background interference. Based on the instantaneous frequency theoretical formulas of the sinusoidal signal, linear signal, and quadratic signal, corresponding ideal time-frequency representations are generated as labels; The original time-frequency graph obtained by short-time Fourier transform of the data samples is used as network input, and the ideal time-frequency representation is used as a supervision label to participate in network training. The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features, including: The dimensions of the original time-frequency plot were unified to 256×256; The time-frequency graph after unifying the dimensions is then processed by block embedding, with a block size of [missing value]. Embedding dimension is To obtain multi-scale embedding features The embedding process is represented as: ； in, For the first Time-frequency plot at scale Corresponding to the original scale, 1 / 2 scale, and 1 / 4 scale; This is the function for the block expansion operation; Implement dimension mapping for 1×1 convolutions; and Indicates the first The height and width of the time-frequency plot at the scale; Perform a two-dimensional Fourier transform on the multi-scale embedded features to separate the amplitude components. and phase components : ； Where FFT(·) represents Fourier transform, e is the base of the natural logarithm, i is the imaginary unit, and arg represents finding its phase angle; The amplitude component carries the energy intensity and distribution of the time-frequency ridge, while the phase component carries the structural location information of the time-frequency ridge.

2. The time-frequency characterization method for a non-stationary signal according to claim 1, characterized in that, The frequency domain decoupling attention mechanism is used to decouple and enhance the amplitude and phase components, resulting in frequency domain decoupling enhancement features, including: The amplitude component and the phase component are respectively input into the frequency domain decoupling attention mechanism, which includes phase-correlated attention, amplitude modulation attention and hybrid attention. The phase-related attention Using only the phase components, the ridge position-related features are captured through phase gradient calculation and self-attention weight allocation, as follows: ； Where iFFT(·) represents the inverse Fourier transform, This is the amplitude correlation value. It is a phase-related query value. This is key information regarding the phase. The amplitude modulation attention Using only the amplitude component, the spatial distribution and intensity variation of ridge energy features are extracted by weighting the amplitude feature channels through a channel attention mechanism, and are expressed as follows: ； in, It is an amplitude-related query value. It represents the Hadamah accumulation. For amplitude-related key information, It is the phase correlation value; The hybrid attention Simultaneously considering both phase and amplitude components, an interactive attention matrix for phase-amplitude features is constructed to achieve adaptive fusion of instantaneous frequency structure information and energy characteristics, expressed as: ； The outputs of the three types of attention features are concatenated and then convolved with a 1×1 convolution to obtain the frequency domain decoupling enhancement feature. : ； in, For feature splicing operations, Represents spatial characteristics.

3. The time-frequency characterization method for a non-stationary signal according to claim 2, characterized in that, The process of obtaining multi-branch fusion features by complementary learning of frequency domain decoupling enhancement features through frequency branching and spatial branching includes: The frequency domain decoupling enhancement features are input into the frequency branch and the spatial branch, respectively; The frequency branch uses learnable filters to perform frequency selection and shifting on the amplitude and phase of the feature, respectively, to obtain the frequency branch feature, represented as: ； ； ； in, For grouping and block operations, Representing scale Input features on and For learnable parameters, and These are the updated phase correlation features and amplitude correlation features, respectively. This indicates the calculation of its phase value. This is a group merging operation. Indicates frequency branching characteristics; The spatial branch is used to learn local information to supplement the representation of frequency domain information, resulting in spatial branch features. ; The frequency branch features and spatial branch features are fused to obtain multi-branch fused features. : ； in, This indicates a gating mechanism.

4. The time-frequency characterization method for a non-stationary signal according to claim 3, characterized in that, The process of feature modulation and calibration through cross-stage interaction between multi-branch fusion features and the original time-frequency map to output a high-resolution time-frequency map includes: The multi-branch fusion features are interacted with the original time-frequency map across stages, and phase modulation, amplitude modulation and spatial modulation are executed sequentially. The phase modulation is achieved by learning the original time-frequency diagram. The phase deviation from the encoded features is used to dynamically calibrate the encoded features, which is expressed as: ； in, For the updated phase value, For learnable weights of phase modulation; The amplitude modulation is based on the amplitude distribution of the original time-frequency diagram, and the encoded amplitude features are adaptively adjusted pixel by pixel, as follows: ； in, For the updated amplitude, For learnable weights of amplitude modulation; The frequency domain characteristics after phase and amplitude modulation are represented by the inverse transform: ； in, For the updated feature representation, Represents the normalization layer. Indicates original features; The spatial modulation is used to achieve adaptive alignment of amplitude and phase representations, expressed as: ； in, This represents a 3×3 convolution. Representative scale is The output characteristics; The features, after being modulated across stages, are reconstructed into the final high-resolution time-frequency map output.

5. The time-frequency characterization method for a non-stationary signal according to claim 1, characterized in that, The pre-trained Fourier domain decoupling time-frequency refinement network is trained by minimizing the total loss function, which includes pixel difference loss. and frequency loss ; The pixel difference loss is used to constrain the reconstruction of the time-frequency map. Compared with the actual time-frequency diagram The pixel difference is represented as: ； in, and Here are the height and width of the time-frequency plot; h and w are the height and width indices, respectively. The frequency loss is used to constrain the consistency of the distribution of amplitude and phase in the Fourier frequency domain output, and is expressed as: ； The total loss function is defined as follows: ; in, This is the loss weight.

6. A time-frequency characterization device for a non-stationary signal, used to implement the time-frequency characterization method for a non-stationary signal according to any one of claims 1-5, characterized in that, include: The acquisition module is used to acquire the raw vibration data of non-stationary signals; The transformation module is used to perform a short-time Fourier transform on the original vibration data to generate the original time-frequency diagram; The input module is used to input the original time-frequency graph into a pre-trained Fourier domain decoupled time-frequency refinement network and perform the following steps: The original time-frequency map is subjected to block embedding processing to obtain multi-scale embedding features; A two-dimensional Fourier transform is performed on the multi-scale embedded features to separate the amplitude component and the phase component; The amplitude component and the phase component are decoupled and enhanced by a frequency domain decoupling attention mechanism, resulting in frequency domain decoupling enhancement features; The frequency domain decoupling enhancement features are learned through complementary learning of frequency branch and spatial branch to obtain multi-branch fusion features; Feature modulation and calibration are performed through cross-stage interaction between multi-branch fusion features and the original time-frequency map, resulting in a high-resolution time-frequency map.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-5.

8. An electronic device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-5.

Images