A rotating machinery fault diagnosis method based on time-frequency domain joint modeling and self-supervised learning

Abstract

The application discloses a rotating machinery fault diagnosis method based on time-frequency domain joint modeling and self-supervised learning, and belongs to the technical field of rotating machinery state monitoring and fault diagnosis, and comprises a vibration signal preprocessing module, a time-frequency domain joint feature modeling module, a self-supervised learning diagnosis module, a confidence check closed loop module and an incremental iteration optimization module; the vibration signal preprocessing module is used for receiving a raw vibration signal of rotating machinery, performing signal standardization preprocessing, generating three types of input data in a time domain, a frequency domain and a time-frequency domain, adaptively weighting fault sensitive dimensions through a channel attention mechanism, comprehensively representing fault evolution information, significantly improving fault feature distinguishability in a variable working condition and a strong noise scene, and guaranteeing the identification ability of early weak faults.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery condition monitoring and fault diagnosis, specifically to a rotating machinery fault diagnosis method based on joint time-frequency domain modeling and self-supervised learning. Background Technology

[0002] The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning is an intelligent diagnostic system that relies on vibration signal analysis technology and deep learning algorithms to extract features and identify faults from vibration data of rotating machinery such as motors, gearboxes, and fans, while taking into account diagnostic accuracy, suitability for small sample sizes, and reliability of results.

[0003] As industrial equipment becomes increasingly high-speed and complex, rotating machinery experiences diverse fault causes and frequent fluctuations in operating conditions. Vibration signals exhibit non-stationary and high-noise characteristics, making traditional manual threshold diagnosis insufficient for on-site needs. Deep learning has gradually become the mainstream technology for fault diagnosis. Currently, fault diagnosis schemes based on convolutional neural networks exist, achieving fault classification by extracting single features in the time or frequency domain and relying on a large number of labeled samples for supervised training. However, typical deep learning-based rotating machinery fault diagnosis systems often have a single dimension in fault feature representation, relying solely on independent analysis in the time or frequency domain. They fail to achieve deep fusion of multi-dimensional information in both time and frequency domains, and the weights of fault-sensitive features cannot be adaptively adjusted. This results in low fault feature discrimination under complex operating conditions and high-noise environments, easily masking early, subtle faults, ultimately leading to insufficient diagnostic accuracy and a high rate of missed diagnoses.

[0004] Meanwhile, most deep learning-based rotating machinery fault diagnosis systems rely heavily on a large number of well-labeled fault samples. However, fault samples in industrial settings are scarce and labeling costs are high. Purely supervised learning models have weak generalization capabilities and lack confidence verification and iterative optimization loops for diagnostic results. Low-confidence results are directly output, and features cannot be automatically optimized for difficult samples. Ultimately, this results in poor model adaptability and insufficient reliability of diagnostic results.

[0005] In summary, existing deep learning-based fault diagnosis technologies for rotating machinery suffer from drawbacks such as low fusion of time and frequency features, insufficient adaptation to scenarios with few samples, and lack of automated verification and iteration mechanisms, making it difficult to meet the current needs for high-precision and high-reliability fault diagnosis in industrial settings. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing rotating machinery faults based on joint time-frequency domain modeling and self-supervised learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: including a vibration signal preprocessing module, a time-frequency domain joint feature modeling module, a self-supervised learning diagnosis module, a confidence verification closed-loop module, and an incremental iterative optimization module;

[0008] The vibration signal preprocessing module is used to receive the original vibration signal of the rotating machinery, perform signal standardization preprocessing, and generate three types of input data: time domain, frequency domain, and time-frequency domain.

[0009] The time-frequency domain joint feature modeling module interacts with the vibration signal preprocessing module to construct a multi-dimensional fault characterization system that integrates time-domain statistical features, frequency-domain spectral features, and time-frequency domain deep features.

[0010] The self-supervised learning diagnostic module is connected to the time-frequency domain joint feature modeling module, and includes a self-supervised pre-training unit and a fine-tuning diagnostic unit. The self-supervised pre-training unit is used to complete the pre-training of the feature encoder based on unlabeled samples. The fine-tuning diagnostic unit is used to complete fault classification and identification based on a small number of labeled samples on the basis of the pre-trained encoder, and output the initial diagnostic results.

[0011] The confidence verification closed-loop module interacts with the self-supervised learning diagnosis module to perform confidence and feature matching degree verification on the initial diagnosis results, and sends low-confidence samples back to the self-supervised learning diagnosis module for feature optimization iteration, and outputs the final diagnosis results.

[0012] The incremental iterative optimization module is used to mark the high-confidence samples that have passed the verification as pseudo-label samples and store them in the sample library. Based on the newly added samples, the self-supervised learning diagnosis module is periodically subjected to incremental training to optimize the model performance.

[0013] As a further preferred embodiment of this technical solution: the vibration signal preprocessing module sequentially performs power frequency denoising, outlier removal, equal-time resampling and minimum-maximum normalization processing, then generates a frequency domain spectrum through fast Fourier transform and a time-frequency spectrum through continuous wavelet transform, and finally outputs three types of standardized data: time domain signal sequence, frequency domain spectrum vector, and time-frequency grayscale spectrum.

[0014] As a further preferred embodiment of this technical solution: the time-frequency domain joint feature modeling module extracts time-domain statistical features, frequency-domain structural features, and time-frequency domain depth features respectively; the time-domain statistical features include four types of time-domain indicators: mean, kurtosis, margin, and impulse factor; the frequency-domain structural features include the main frequency position, sideband energy ratio, and harmonic distribution features; the three types of features are weighted and fused through a channel attention mechanism to generate a multi-dimensional fault representation vector, with the weight of the fault sensitivity dimension automatically increased;

[0015] As a further preferred embodiment of this technical solution: the self-supervised learning diagnostic module adopts a two-level architecture of pre-training-fine-tuning; the self-supervised pre-training unit performs mask reconstruction tasks based on a large number of unlabeled samples to train a general feature extraction encoder; the fine-tuning diagnostic unit freezes the low-level parameters of the encoder and only fine-tunes the top-level classification network, relying on a small number of labeled samples to achieve classification and recognition of fault type and severity.

[0016] As a further preferred embodiment of this technical solution: the self-supervised pre-training unit adopts a fault-sensitive weighted masking mechanism: randomly selects 15% to 30% of the input time-frequency domain features for masking, while increasing the masking probability of fault-sensitive regions such as impact components and fault side frequencies; with the goal of completely reconstructing the original time-frequency features, the encoder is forced to learn deep fault-related features;

[0017] As a further preferred embodiment of this technical solution: the feature encoder adopts a multi-scale convolutional Transformer structure. The bottom layer extracts local fault features of different frequency bands through multi-scale convolutional kernels, and the upper layer learns long-term fault evolution correlation through the Transformer attention mechanism. It is also suitable for feature extraction of three types of inputs: time domain, frequency domain, and time-frequency domain.

[0018] As a further preferred embodiment of this technical solution: pre-training and fine-tuning adopt a step-by-step joint training mechanism: in the pre-training stage, the reconstruction loss is used as the optimization target to train all encoder parameters; in the fine-tuning stage, the bottom feature extraction layer of the encoder is frozen, and only the top attention layer and classification head are fine-tuned, with cross-entropy loss as the optimization target; the fine-tuning loss is backpropagated to the top layer of the encoder to achieve the synergistic optimization of feature extraction and fault diagnosis.

[0019] As a further preferred embodiment of this technical solution: the confidence verification closed-loop module is set with dual triggering conditions: the confidence of the diagnostic result is lower than a preset threshold, or the feature does not match the standard fault template well enough; if either condition is met, the sample is sent back to the self-supervised pre-training unit for secondary feature optimization; at the same time, the maximum number of iterations is set to 2 to 3 times, and if the limit is exceeded, the current result is output with a low confidence mark.

[0020] As a further preferred embodiment of this technical solution: the incremental iterative optimization module marks diagnostic samples with qualified confidence and feature matching as pseudo-label high-quality samples and stores them in the experience sample library; after each set number of new samples are accumulated, incremental pre-training is performed once, and the encoder high-level parameters are fine-tuned with a small learning rate. The learning rate is 1 / 5 to 1 / 10 of the initial pre-training, so as to retain the original fault feature extraction capability and avoid catastrophic forgetting.

[0021] As a further preferred embodiment of this technical solution, the system performs the following steps:

[0022] S1, the vibration signal preprocessing module performs standardized preprocessing on the original vibration signal to generate three types of input data: time domain, frequency domain, and time-frequency domain;

[0023] S2, the time-frequency domain joint feature modeling module extracts three types of fault features and weights them together to construct a multi-dimensional fault characterization system;

[0024] S3, the self-supervised learning diagnostic module first pre-trains the feature encoder using unlabeled samples, then fine-tunes the classification network using a small number of labeled samples, and outputs the initial diagnostic results;

[0025] S4, the confidence verification closed-loop module performs double verification on the initial diagnosis results, and low-confidence samples are returned for feature optimization iteration until the standard is met or the maximum number of iterations is reached;

[0026] S5 outputs the final fault diagnosis results and stores high-confidence, high-quality samples into the sample library, and periodically performs incremental model updates.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. To address the issues of single feature representation dimensions, insufficient time-frequency information fusion, and fixed weights for fault sensitivity dimensions in general deep learning-based rotating machinery fault diagnosis systems, which lead to low fault discrimination and high missed rates for minor faults under complex operating conditions, this solution constructs a multi-dimensional joint feature modeling system in the time-frequency domain. This system integrates three types of information: time-domain statistical features, frequency-domain structural features, and time-frequency domain deep features. Through a channel attention mechanism, it adaptively weights the fault sensitivity dimensions to comprehensively represent fault evolution information, significantly improving the fault feature discrimination under varying operating conditions and strong noise scenarios, and ensuring the ability to identify early minor faults.

[0029] 2. To address the issues of high dependence on labeled samples, weak generalization ability with few samples, and lack of confidence verification loops in general deep learning-based rotating machinery fault diagnosis systems, which lead to high implementation costs and insufficient reliability of results, this solution adopts a two-level architecture of self-supervised pre-training and fine-tuning. It designs a fault-sensitive weighted mask pre-training task, learns general fault features based on a large number of unlabeled samples, and achieves high-precision diagnosis with few labeled samples. A dual confidence verification loop is built, with low-confidence samples used for iterative feature optimization. An incremental iteration mechanism using pseudo-labeled samples is also included, allowing the model to continuously evolve with data accumulation. This effectively reduces labeling costs, improves cross-condition adaptability and diagnostic reliability, and strengthens the system's engineering implementation capabilities. Attached Figure Description

[0030] Figure 1 This is a system architecture block diagram of a rotating machinery fault diagnosis method based on joint time-frequency domain modeling and self-supervised learning according to the present invention.

[0031] Figure 2 This invention presents a self-supervised learning diagnostic module architecture for a rotating machinery fault diagnosis method based on joint time-frequency domain modeling and self-supervised learning. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example

[0034] Please see Figures 1-2 This is a schematic diagram of some embodiments of a rotating machinery fault diagnosis method based on joint time-frequency domain modeling and self-supervised learning, as described in this application.

[0035] Example 1

[0036] See Figure 1 The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning provided by the present invention includes a vibration signal preprocessing module, a joint time-frequency domain feature modeling module, a self-supervised learning diagnosis module, a confidence verification closed-loop module, and an incremental iterative optimization module.

[0037] The vibration signal preprocessing module receives the original vibration signal of the rotating machinery, performs standardized preprocessing to generate three types of time-frequency domain input data, and sends the data to the time-frequency domain joint feature modeling module.

[0038] The time-frequency domain joint feature modeling module extracts multi-dimensional fault features and weights them together to establish a multi-source fault characterization system, and sends the feature data to the self-supervised learning diagnosis module.

[0039] The self-supervised learning diagnostic module adopts a two-level architecture of pre-training and fine-tuning. It relies on unlabeled samples to complete the encoder pre-training, and then fine-tunes the classification network with a small number of labeled samples to output the initial fault diagnosis results and send the data to the confidence verification closed-loop module.

[0040] The confidence verification closed-loop module performs dual verification of confidence and feature matching degree on the initial diagnosis results, anchors low-confidence samples and triggers feature iterative optimization, and sends the iterated diagnosis results to the incremental iterative optimization module.

[0041] The incremental iterative optimization module outputs the final diagnostic results and adds high-confidence, high-quality samples to the database to achieve incremental model updates.

[0042] Example 2

[0043] See Figure 1 This embodiment is based on the above embodiment. The vibration signal preprocessing module performs multi-dimensional standardization processing on the original vibration signal to provide standardized data input for subsequent feature modeling and diagnosis; specifically, it includes:

[0044] The signal undergoes basic preprocessing, including power frequency notch denoising, outlier removal, and equal-time resampling, which convert the original non-uniform sampled signal into a uniform time-domain sequence with a fixed sampling rate. The sampling rate is adaptively adjusted according to the equipment rotation speed, with a range of 1kHz to 20kHz.

[0045] Signal normalization processing uses minimum-maximum normalization to map the time-domain signal amplitude to the [0,1] interval, eliminating differences in amplitude dimensions. The calculation formula is as follows: ;

[0046] Where x is the original time-domain signal point, , These are the maximum and minimum values ​​of the signal sequence, respectively. This is the normalized time-domain signal.

[0047] Frequency domain data generation involves performing a Fast Fourier Transform on the normalized time-domain signal to obtain the frequency domain amplitude spectrum. The calculation formula is as follows: ;

[0048] Where N is the number of time-domain signal sampling points, For frequency points, This represents the amplitude at the corresponding frequency.

[0049] Time-frequency domain data generation: Continuous wavelet transform is used to convert the time-domain signal into a time-frequency grayscale image. Morlet wavelet is used as the mother wavelet, and the calculation formula is as follows:

[0050] ;

[0051] in, For the mother wavelet function, The time-frequency domain coefficient matrix is ​​generated by gray-level normalization to produce a time-frequency spectrum.

[0052] The final output consists of three types of standardized data: time-domain signal sequence, frequency-domain spectrum vector, and time-frequency grayscale image, which are simultaneously input into the downstream feature modeling module.

[0053] Example 3

[0054] See Figure 1 This embodiment, based on the above embodiment, constructs a three-category complementary fault feature system using a time-frequency domain joint feature modeling module. Adaptive fusion is achieved through an attention mechanism, replacing traditional manual single-domain features. Specifically, it includes:

[0055] Time-domain statistical feature extraction: Extract time-domain indicators that reflect the impact characteristics and distribution characteristics. The core indicators include kurtosis and impulse factor.

[0056] Kurtosis characterizes the signal impact intensity, and its calculation formula is: ;

[0057] in, The mean of the signal. The standard deviation of the signal is represented by the kurtosis value, and a larger kurtosis value indicates a stronger fault impact.

[0058] The pulse factor characterizes the relative intensity of the impact pulse, and its calculation formula is as follows: ;

[0059] In addition, six time-domain statistical indicators, including mean, margin, and waveform factor, are added to form a time-domain feature vector.

[0060] Frequency domain structural features are extracted, including four frequency domain indicators: main frequency position, side frequency energy ratio, harmonic energy distribution, and centroid frequency. These indicators reflect the modulation characteristics of the fault characteristic frequency and form a frequency domain feature vector.

[0061] Time-frequency domain deep feature extraction uses a convolutional network to extract deep time-frequency joint features from the time-frequency spectrum, capture the time-varying evolution of faults, and construct a time-frequency deep feature vector.

[0062] Multi-feature adaptive fusion employs a channel attention mechanism to weight and fuse three types of features. First, global average pooling is performed on each dimension of the features. Then, a two-layer fully connected network generates weight coefficients for each feature channel, with automatic weighting of the fault-sensitive dimension. The formula for calculating the fused multi-dimensional fault representation vector is as follows:

[0063] ,in, Let i be the feature vector of the i-th class. For the attention weights of the corresponding channels, satisfying M represents the total number of feature categories.

[0064] Example 4

[0065] See Figure 1 and Figure 2 This embodiment is based on the above embodiment. The self-supervised learning diagnostic module adopts a two-level architecture of pre-training-fine-tuning, separating the training stages of feature learning and fault classification, and achieving high-precision diagnosis with few samples through the collaboration of the two-level units; specifically including:

[0066] The self-supervised pre-training unit, serving as the upper-layer general feature learning unit, is responsible for pre-training the feature encoder based on a large number of unlabeled vibration samples. The input consists of three types of time-frequency domain standardized data, and the output is the pre-trained general feature extraction encoder. The pre-training stage focuses on mask reconstruction, forcing the encoder to learn the essential features related to the fault, rather than fitting noise and operational disturbances.

[0067] The fine-tuning diagnostic unit, serving as the downstream fault classification unit, fine-tunes the fault classification network based on a pre-trained encoder and a small number of labeled samples. The input is the fused feature vector of the labeled samples, and the output is the classification result of the fault type and severity. During the fine-tuning stage, the encoder's low-level feature extraction parameters are frozen, and only the top-level attention layer and classification head are optimized to avoid overfitting with small sample sizes.

[0068] The feature encoder employs a multi-scale convolutional Transformer structure. The bottom layer uses three different sized convolutional kernels to extract local fault features in the low-frequency, mid-frequency, and high-frequency bands, respectively. The upper layer uses a self-attention mechanism to capture fault evolution correlations over long time periods. This structure can simultaneously adapt to feature extraction from three types of inputs: time-domain sequences, frequency-domain vectors, and time-frequency maps.

[0069] Example 5

[0070] See Figure 1 and Figure 2 This embodiment, based on the above embodiments, employs a fault-sensitive weighted mask pre-training task and a tiered joint training mechanism in the self-supervised learning diagnostic module to ensure diagnostic performance in scenarios with few samples; specifically, it includes:

[0071] The fault-sensitive weighted masking mechanism randomly selects regions of the input time-frequency domain features for masking during the pre-training phase, with a masking ratio of 15% to 30%. At the same time, a fault-sensitive weighting strategy is introduced to increase the masking probability of fault-sensitive locations such as signal impact areas and fault feature frequency bands, with the masking weight of sensitive areas being twice that of ordinary areas. The pre-training goal is to completely reconstruct the original time-frequency features, forcing the encoder to learn deep features associated with faults.

[0072] During the pre-training phase, the loss function is reconstructed using mean squared error, and the calculation formula is as follows: ,in, For the original feature truth value, The encoder reconstructs the output feature values, where N is the total number of feature dimensions; pre-training aims to minimize the reconstruction loss and updates all parameters of the encoder.

[0073] The fine-tuning phase uses the cross-entropy classification loss function, calculated as follows:

[0074] Where C represents the total number of fault categories. For category real labels, The output class probabilities of the model; during the fine-tuning stage, 70% of the parameters of the encoder's bottom layer are frozen, and only the parameters of the top attention layer and the classification head are optimized to prevent overfitting on small samples.

[0075] A tiered joint training mechanism is used: first, pre-training is completed on a large number of unlabeled samples, and then fine-tuning is performed on a small number of labeled samples; the fine-tuning loss is backpropagated to the top layer of the encoder to simultaneously optimize the top-level feature extraction capability, thereby achieving synergistic optimization of general feature learning and fault classification; the training discount factor ranges from 0.9 to 0.99.

[0076] Example 6

[0077] See Figure 1 This embodiment, based on the above embodiment, performs dual verification on the initial diagnostic results using the confidence verification closed-loop module, and improves diagnostic reliability through iterative optimization using low-confidence samples; specifically including:

[0078] The dual-validation rule engine incorporates two standards: confidence level validation and feature matching degree validation, to validate each item of the initial diagnostic results.

[0079] Confidence verification: The maximum class probability output by the model is taken as the diagnostic confidence score, calculated using the following formula: ,in, The predicted probability of class c is given. Samples with a confidence level lower than a preset threshold are marked as low-confidence samples.

[0080] Feature matching degree verification: Calculate the cosine similarity between the features of the current sample and the features of the standard fault template. The calculation formula is: ,in, This is the fused feature vector of the current sample. The standard fault template features are used for the corresponding category; samples with similarity below the threshold are marked as insufficiently matched.

[0081] Low-confidence region anchoring locates the feature distribution deviation of all unqualified samples, generates feature optimization constraint range, and performs feature re-extraction and optimization only on the deviation dimension, while the qualified dimension remains unchanged.

[0082] During iterative optimization, when a non-compliant sample is found during verification, the sample is sent back to the self-supervised pre-training unit. Only the bias dimension is optimized for feature secondary optimization, while the other features remain unchanged.

[0083] Triggering and Termination Mechanism: Set dual triggering conditions, and start iterative optimization when either one is met—confidence is lower than the threshold or feature matching degree is lower than the threshold; set the maximum number of iterations to 2 to 3 times. If the requirements are still not met after reaching the upper limit of the number of iterations, output the current best result with a low confidence mark to avoid infinite iteration and loss of computing power.

[0084] Example 7

[0085] See Figure 1 This embodiment is based on the above embodiment, and the incremental iterative optimization module completes the final diagnostic result output and continuous model optimization; specifically including:

[0086] The diagnostic results output includes the fault type, severity, and confidence score of the final diagnosis, and simultaneously generates a diagnostic analysis report listing the scores of each characteristic indicator and the reliability conclusion.

[0087] High-quality pseudo-label samples are selected by marking diagnostic samples that meet the confidence level and feature matching criteria as high-quality pseudo-label samples. The selection ratio is the top 20% of the diagnostic samples in the same period.

[0088] The model is incrementally updated by storing high-quality pseudo-label samples in the experience sample library. Every 100 newly added high-quality samples are used to perform incremental pre-training on the self-supervised learning diagnostic module. The incremental training adopts a small learning rate fine-tuning strategy, with the learning rate set to 1 / 5 to 1 / 10 of the initial pre-training learning rate. Only the high-level attention parameters of the encoder are fine-tuned to retain the low-level general feature extraction capability and avoid catastrophic forgetting in the model.

[0089] See Figure 1 This embodiment is based on the above embodiment, and the corresponding diagnostic method of this system performs the following steps:

[0090] S1. The vibration signal preprocessing module performs standardized preprocessing on the original vibration signal to generate three types of input data: time domain, frequency domain, and time-frequency domain.

[0091] S2. The time-frequency domain joint feature modeling module extracts three types of fault features and weights them through an attention mechanism to construct a multi-dimensional fault representation system;

[0092] S3. The self-supervised learning diagnostic module first pre-trains the feature encoder using unlabeled samples, then fine-tunes the classification network using a small number of labeled samples, and outputs the initial diagnostic results.

[0093] S4. The confidence verification closed-loop module performs double verification on the initial diagnostic results. If there are low-confidence samples, they are sent back to the feature encoder to perform local optimization iterations until the target is met or the maximum number of iterations is reached.

[0094] S5. The system outputs the final fault diagnosis results and stores high-confidence, high-quality samples into the experience sample library, and periodically performs incremental model updates.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art, inspired by this description, design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the scope of protection of this invention.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning, characterized in that: It includes a vibration signal preprocessing module, a time-frequency domain joint feature modeling module, a self-supervised learning diagnosis module, a confidence verification closed-loop module, and an incremental iterative optimization module; The vibration signal preprocessing module is used to receive the original vibration signal of the rotating machinery, perform signal standardization preprocessing, and generate three types of input data: time domain, frequency domain, and time-frequency domain. The time-frequency domain joint feature modeling module interacts with the vibration signal preprocessing module to construct a multi-dimensional fault characterization system that integrates time-domain statistical features, frequency-domain spectral features, and time-frequency domain deep features. The self-supervised learning diagnostic module is connected to the time-frequency domain joint feature modeling module, and includes a self-supervised pre-training unit and a fine-tuning diagnostic unit; the self-supervised pre-training unit is used to complete the pre-training of the feature encoder based on unlabeled samples; The fine-tuning diagnostic unit is used to complete fault classification and identification based on a small number of labeled samples on the basis of a pre-trained encoder, and output the initial diagnostic results. The confidence verification closed-loop module interacts with the self-supervised learning diagnosis module to perform confidence and feature matching degree verification on the initial diagnosis results, and sends low-confidence samples back to the self-supervised learning diagnosis module for feature optimization iteration, and outputs the final diagnosis results. The incremental iterative optimization module is used to mark qualified high-confidence samples as pseudo-label samples and store them in the sample library. Based on the newly added samples, the self-supervised learning diagnostic module is periodically trained to optimize the model performance.

2. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 1, characterized in that: The vibration signal preprocessing module sequentially performs power frequency denoising, outlier removal, equal-time resampling, and minimum-maximum normalization. Then, it generates a frequency domain spectrum through fast Fourier transform and a time-frequency spectrum through continuous wavelet transform. Finally, it outputs three types of standardized data: time domain signal sequence, frequency domain spectrum vector, and time-frequency grayscale spectrum.

3. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 2, characterized in that: The time-frequency domain joint feature modeling module extracts time-domain statistical features, frequency-domain structural features, and time-frequency domain depth features, respectively. The time-domain statistical features include four types of time-domain indices: mean, kurtosis, margin, and impulse factor. The frequency-domain structural features include the dominant frequency position, sideband energy proportion, and harmonic distribution features. The three types of features are weighted and fused through a channel attention mechanism to generate a multi-dimensional fault representation vector, with the weight of the fault-sensitive dimension automatically increased.

4. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 3, characterized in that: The self-supervised learning diagnostic module adopts a two-level architecture of pre-training-fine-tuning; the self-supervised pre-training unit performs mask reconstruction tasks based on a large number of unlabeled samples to train a general feature extraction encoder. The diagnostic unit freezes the encoder's underlying parameters and only fine-tunes the top-level classification network, relying on a small number of labeled samples to achieve the classification and identification of fault types and severity.

5. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 4, characterized in that: The self-supervised pre-training unit adopts a fault-sensitive weighted masking mechanism: randomly selects 15% to 30% of the input time-frequency domain features for masking, and increases the masking probability of fault-sensitive regions such as impact components and fault side frequencies; with the goal of completely reconstructing the original time-frequency features, the encoder is forced to learn deep fault-related features.

6. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 5, characterized in that: The feature encoder adopts a multi-scale convolutional Transformer structure. The bottom layer extracts local fault features in different frequency bands through multi-scale convolutional kernels, and the upper layer learns long-term fault evolution correlations through the Transformer attention mechanism. It is also suitable for feature extraction of three types of inputs: time domain, frequency domain, and time-frequency domain.

7. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 6, characterized in that: Pre-training and fine-tuning adopt a step-by-step joint training mechanism: in the pre-training stage, the full encoder parameters are trained with reconstruction loss as the optimization target; in the fine-tuning stage, the bottom feature extraction layer of the encoder is frozen, and only the top attention layer and classification head are fine-tuned with cross-entropy loss as the optimization target; the fine-tuning loss is backpropagated to the top layer of the encoder to achieve collaborative optimization of feature extraction and fault diagnosis.

8. The rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 7, characterized in that: The confidence verification closed-loop module is set with dual triggering conditions: the confidence of the diagnostic result is lower than the preset threshold, or the feature does not match the standard fault template well enough. If one of these conditions is met, the sample is fed back to the self-supervised pre-training unit for secondary feature optimization; at the same time, the maximum number of iterations is set to 2 to 3 times, and if the limit is exceeded, the current result is output with a low confidence label.

9. A rotating machinery fault diagnosis system based on joint time-frequency domain modeling and self-supervised learning according to claim 8, characterized in that: The incremental iterative optimization module marks diagnostic samples that meet the confidence level and feature matching as pseudo-label high-quality samples and stores them in the experience sample library. After each set number of new samples are accumulated, incremental pre-training is performed. The encoder high-level parameters are fine-tuned with a small learning rate of 1 / 5 to 1 / 10 of the initial pre-training rate. This preserves the original fault feature extraction capability and avoids catastrophic forgetting.

10. A method for fault diagnosis of rotating machinery based on joint time-frequency domain modeling and self-supervised learning, characterized in that: The system according to any one of claims 1 to 9 performs the following steps: S1, the vibration signal preprocessing module performs standardized preprocessing on the original vibration signal to generate three types of input data: time domain, frequency domain, and time-frequency domain; S2, the time-frequency domain joint feature modeling module extracts three types of fault features and weights them together to construct a multi-dimensional fault characterization system; S3, the self-supervised learning diagnostic module first pre-trains the feature encoder using unlabeled samples, then fine-tunes the classification network using a small number of labeled samples, and outputs the initial diagnostic results; S4, the confidence verification closed-loop module performs double verification on the initial diagnosis results, and low-confidence samples are returned for feature optimization iteration until the standard is met or the maximum number of iterations is reached; S5 outputs the final fault diagnosis results and stores high-confidence, high-quality samples into the sample library, and periodically performs incremental model updates.

Images