A blade crack fault feature mining and early warning method

Abstract

The present application relates to the technical field of rotating machinery fault diagnosis and condition monitoring, and discloses a kind of blade crack fault feature mining and early warning method, comprising: obtaining blade simulation crack measured signal, source domain dataset is constructed, target domain dataset is constructed, and source domain dataset and target domain dataset form input tensor;Utilize one-dimensional deep residual network as feature extractor, and utilize source domain dataset to pre-train feature extractor, obtain pre-trained feature extractor;Unified framework containing pre-trained feature extractor, fault classifier and domain discriminator is constructed, and knowledge transfer from source domain to target domain is realized by joint optimization classification loss, domain adversarial loss and maximum mean difference, and the trained cross-domain transfer diagnosis model is obtained;Real-time monitoring signal is input into the trained cross-domain transfer diagnosis model, and fault probability vector is obtained, multi-level early warning is carried out in combination with adaptive threshold, which can provide technical support for unit safe service.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis and condition monitoring technology, and in particular to a method for identifying and warning of blade crack fault characteristics. Background Technology

[0002] As the core working component of rotating machinery, blades operate under high temperature, high pressure, high speed, and complex load coupling for extended periods, making them prone to cracking failures due to fatigue accumulation and stress concentration. Early crack characteristics are extremely subtle and are masked by strong background noise in industrial environments such as electromagnetic interference, airflow disturbances, and sensor noise, making them difficult to identify directly. When cracks extend to a detectable extent, the equipment faces serious safety risks and is highly susceptible to major accidents such as blade breakage and complete machine shutdown.

[0003] Currently, blade crack fault diagnosis methods mainly rely on traditional signal processing techniques and machine learning algorithms. Traditional methods require manual feature extraction through Fourier transform, wavelet analysis, and other means, which not only demands a high level of operator experience but also struggles to capture the nonlinear, non-stationary, and subtle characteristics of early-stage cracks. While machine learning algorithms can achieve a certain degree of automatic classification, they heavily depend on a large number of labeled real fault samples for model training. However, in industrial settings, real blade crack fault samples are scarce and extremely costly to obtain, resulting in a homogeneous distribution of model training data. Consequently, when faced with blade monitoring data from different operating conditions and types of units, the generalization ability is extremely poor, making it difficult to adapt to the complex and ever-changing needs of on-site monitoring.

[0004] Traditional deep learning models are still limited by the assumption that the distribution of training data and test data is consistent. When a model trained on laboratory-simulated crack samples is directly applied to the field, the diagnostic accuracy of the model drops significantly due to the difference in feature distribution between the source domain (simulated crack) and the target domain (field operation) (i.e., the domain offset problem). Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for identifying and providing early warning of blade crack fault features. This method is applicable to the identification and real-time warning of early crack faults in rotating blades of key equipment such as steam turbines and aero engines, providing technical support for the safe operation of the unit.

[0006] This invention provides a method for identifying and predicting blade crack fault features, comprising the following steps: Acquire measured signals of simulated cracks in blades under multiple operating conditions and different crack parameters to construct a source domain dataset. Acquire blade monitoring signals of actual units in service in industrial sites to construct an unlabeled target domain dataset. The source domain dataset and the target domain dataset form an input tensor. A one-dimensional deep residual network is used as a feature extractor. The high-dimensional fault features in the input tensor are extracted through the residual learning unit. The feature extractor is pre-trained using the source domain dataset to obtain the pre-trained feature extractor. A unified framework is constructed, comprising a pre-trained feature extractor, a fault classifier, and a domain discriminator. The fault classifier predicts fault categories in the source domain dataset and calculates the classification loss. The domain discriminator performs domain classification on the features of the source and target domain datasets and calculates the domain adversarial loss to force the feature extractor to extract domain-invariant features. At the same time, the maximum mean difference is used to explicitly bring the feature distribution centers of the source and target domain datasets closer. Knowledge transfer from the source domain to the target domain is achieved by jointly optimizing the classification loss, the domain adversarial loss, and the maximum mean difference, resulting in a trained cross-domain transfer diagnostic model. The real-time monitoring signals from the site are input into the trained cross-domain transfer diagnostic model to obtain the fault probability vector, and multi-level early warning is performed by combining adaptive thresholds.

[0007] Optional, the construction of the source domain dataset includes the following steps: Select specimens with the same material as the blades of the actual unit, and pre-induce cracks of different severity in the key stress concentration area; Define a set of crack severity, construct a multi-field coupled working condition simulation environment, collect blade vibration or acoustic emission signals through sensors, and construct a source domain dataset with fault labels. The set of crack severity is represented as follows: , in, This is a set of crack severity. For the first i Each crack length sample value For the first j Each crack depth sample value This is the minimum critical value for monitoring crack length. This represents the maximum critical value for monitoring crack length. This is the minimum critical value for monitoring crack depth. This is the maximum critical value for monitoring crack depth.

[0008] Optionally, the measured and monitored signals are preprocessed to form source domain datasets and target domain datasets. The preprocessing includes the following steps: Kalman filtering is used to remove Gaussian white noise and power frequency interference from the measured and monitored signals to form a filtered signal. The filtered signal is Z-score normalized to form a normalized filtered signal; An enhancement operation is applied to the signal in the source domain dataset. The enhancement operation includes at least one of amplitude scaling, time stretching, random noise superposition, and feature domain enhancement.

[0009] Optionally, the feature extractor is a temporal feature extraction network based on a one-dimensional deep residual network, and the mathematical expression of the residual learning unit is: , in, For the first The input of each residual block, For the first The output of each residual block Indicates the convolution kernel weights. It is the ReLU activation function; This is the residual mapping function.

[0010] Optionally, the feature extractor also includes an attention module, which is used to calculate the weight vector of each channel and multiply it with the intermediate feature map channel by channel to obtain a weighted feature map.

[0011] Optionally, before pre-training the feature extractor using the source domain dataset, the following steps are also included: The SMOTE algorithm is introduced to perform linear interpolation synthesis in the feature space where the high-dimensional fault features are located, generating synthetic samples of microcrack categories. The synthetic samples are added to the source domain dataset to expand the number of samples of microcrack categories, resulting in a balanced source domain dataset. The balanced source domain dataset is then input into a one-dimensional deep residual network and optimized end-to-end with a weighted loss function. After issuing multi-level warnings, the following steps are also included: Calculate the prediction entropy of the target domain dataset, select samples with entropy values ​​higher than a preset threshold for manual review and labeling, add the newly labeled samples to the training set, and incrementally fine-tune the cross-domain transfer diagnostic model. During incremental fine-tuning, the shallow parameters of the feature extractor are fixed, and the parameters of the deep residual block and the fault classifier are updated only.

[0012] Optionally, the classification loss uses weighted cross-entropy loss, expressed as: , in, Weighting coefficients set for microcrack categories. ,in, N This represents the total number of samples; K Total number of fault categories; n i For the first i Number of samples with microcrack-like structures; Indicates the total number of samples in the source domain. This represents the nth source domain input sample. Indicates the first The true fault category label corresponding to each source domain input sample This indicates that for the input sample The model predicts that it belongs to the true category. The conditional probability.

[0013] Optionally, a gradient inversion layer is set between the domain discriminator and the feature extractor. The gradient inversion layer is an identity mapping during forward propagation and multiplies the gradient by a negative weight coefficient during backward propagation.

[0014] Optionally, the adaptive threshold is set using a statistical process control-based method, expressed as: , in, This represents the mean of the predicted probabilities of healthy samples within the sliding time window. The standard deviation of the predicted probability of healthy samples within the sliding time window. This is for the safety factor.

[0015] Optionally, multi-level early warning includes: when the fault probability is less than the adaptive threshold, it is judged as normal; when the fault probability is greater than or equal to the adaptive threshold and less than the preset secondary threshold, it is judged as a first-level early warning; when the fault probability is greater than or equal to the preset secondary threshold, it is judged as a second-level alarm.

[0016] The technical solution provided by the embodiments of the present invention has the following advantages compared with the prior art: This invention provides a method for feature mining and early warning of blade crack faults. By acquiring measured signals of simulated blade cracks under multiple operating conditions and different crack parameters, a high-fidelity source domain dataset is constructed. A one-dimensional deep residual network is then used to pre-train the feature extractor, enabling the model to learn basic crack feature representation capabilities without requiring a large number of real fault samples. This overcomes the dependence of traditional machine learning algorithms on a large number of labeled real fault samples and effectively improves the model's generalization ability under different operating conditions and monitoring data of different types of units. Using a one-dimensional deep residual network as the feature extractor, the short-circuit connection structure combining identity mapping and residual mapping in the residual learning unit allows weak crack impact features to be preserved in the deep network without dissipating with increasing layer count. This effectively overcomes the gradient vanishing problem in deep network training, achieving accurate mapping from the original signal to high-dimensional fault features and significantly improving the ability to mine early, weak crack features. By constructing a unified framework comprising a feature extractor, a fault classifier, and a domain discriminator, the domain discriminator calculates domain adversarial loss, forcing the feature extractor to extract domain-invariant features and removing redundant information related to sensor installation deviations, environmental noise, and operating condition differences. Simultaneously, the maximum mean difference is used to explicitly narrow the feature distribution centers of the source and target domains, aligning the higher-order statistical moments of the feature distributions. Knowledge transfer from the source to the target domain is achieved through joint optimization of the classification loss, domain adversarial loss, and maximum mean difference. This enables the trained cross-domain transfer diagnostic model to possess fault identification capabilities in the target domain comparable to those in the source domain, effectively overcoming the domain offset problem and achieving high-precision cross-scenario transfer diagnosis from simulated environments to actual field conditions. This invention inputs real-time monitoring signals from the field into the trained cross-domain transfer diagnostic model to obtain a fault probability vector. Combined with adaptive thresholds, multi-level early warnings are provided, avoiding the false alarms easily caused by fluctuations in field conditions with traditional fixed thresholds. This provides a reliable technical guarantee for the health management of rotating machinery blades. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for identifying and warning of blade crack fault features, provided in the first embodiment of the present invention. Figure 2 A flowchart of a blade crack fault feature mining and early warning method provided in the second part of the embodiments of the present invention; Figure 3 A flowchart illustrating the construction process of the multi-condition simulated crack measured signal dataset provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the depth residual feature extraction mesh structure provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the domain adversarial transfer learning model structure provided in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0020] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Traditional deep learning models are still limited by the assumption that training and test data distributions are consistent. When a model trained on simulated crack samples in a laboratory is directly applied to the field, the diagnostic accuracy drops significantly due to the difference in feature distribution between the source domain (simulated crack) and the target domain (field operation) (i.e., the domain shift problem). In recent years, deep learning technology has been widely used in fault diagnosis due to its powerful automatic feature extraction capabilities. Deep transfer learning, by aligning the feature distributions between different domains, can effectively utilize the knowledge from the source domain data to transfer to the target domain task, providing a new approach to solving the two core challenges of scarce fault samples and domain shift.

[0022] Therefore, there is an urgent need for a method for feature mining and early warning of blade cracks. By constructing a high-fidelity simulated fault sample library, designing an efficient feature extraction network and a cross-domain transfer model, this method can accurately mine weak features of early cracks and provide cross-scenario early warning. This overcomes the shortcomings of traditional methods in terms of dependence on real fault samples and insufficient generalization ability, and provides a reliable guarantee for the health management of rotating machinery blades.

[0023] First, embodiments of the present invention provide a method for identifying and predicting blade crack fault features, specifically as follows: Figure 1 and Figure 2 As shown, it includes the following steps: Acquire measured signals of simulated cracks in blades under multiple operating conditions and different crack parameters to construct a source domain dataset. Acquire blade monitoring signals of actual units in service in industrial sites to construct an unlabeled target domain dataset. The source domain dataset and the target domain dataset form an input tensor. A one-dimensional deep residual network is used as a feature extractor. The high-dimensional fault features in the input tensor are extracted through the residual learning unit. The feature extractor is pre-trained using the source domain dataset to obtain the pre-trained feature extractor. A unified framework is constructed, comprising a pre-trained feature extractor, a fault classifier, and a domain discriminator. The fault classifier predicts fault categories in the source domain dataset and calculates the classification loss. The domain discriminator performs domain classification on the features of the source and target domain datasets and calculates the domain adversarial loss to force the feature extractor to extract domain-invariant features. At the same time, the maximum mean difference is used to explicitly bring the feature distribution centers of the source and target domain datasets closer. Knowledge transfer from the source domain to the target domain is achieved by jointly optimizing the classification loss, the domain adversarial loss, and the maximum mean difference, resulting in a trained cross-domain transfer diagnostic model. The real-time monitoring signals from the site are input into the trained cross-domain transfer diagnostic model to obtain the fault probability vector, and multi-level early warning is performed by combining adaptive thresholds.

[0024] This invention provides a method for feature mining and early warning of blade crack faults. By acquiring measured signals of simulated blade cracks under multiple operating conditions and different crack parameters, a high-fidelity source domain dataset is constructed. A one-dimensional deep residual network is then used to pre-train the feature extractor, enabling the model to learn basic crack feature representation capabilities without requiring a large number of real fault samples. This overcomes the dependence of traditional machine learning algorithms on a large number of labeled real fault samples and effectively improves the model's generalization ability under different operating conditions and monitoring data of different types of units. Using a one-dimensional deep residual network as the feature extractor, the short-circuit connection structure combining identity mapping and residual mapping in the residual learning unit allows weak crack impact features to be preserved in the deep network without dissipating with increasing layer count. This effectively overcomes the gradient vanishing problem in deep network training, achieving accurate mapping from the original signal to high-dimensional fault features and significantly improving the ability to mine early, weak crack features. By constructing a unified framework comprising a feature extractor, a fault classifier, and a domain discriminator, the domain discriminator calculates domain adversarial loss, forcing the feature extractor to extract domain-invariant features and removing redundant information related to sensor installation deviations, environmental noise, and operating condition differences. Simultaneously, the maximum mean difference is used to explicitly narrow the feature distribution centers of the source and target domains, aligning the higher-order statistical moments of the feature distributions. Knowledge transfer from the source to the target domain is achieved through joint optimization of the classification loss, domain adversarial loss, and maximum mean difference. This enables the trained cross-domain transfer diagnostic model to possess fault identification capabilities in the target domain comparable to those in the source domain, effectively overcoming the domain offset problem and achieving high-precision cross-scenario transfer diagnosis from simulated environments to actual field conditions. This invention inputs real-time monitoring signals from the field into the trained cross-domain transfer diagnostic model to obtain a fault probability vector. Combined with adaptive thresholds, multi-level early warnings are provided, avoiding the false alarms easily caused by fluctuations in field conditions with traditional fixed thresholds. This provides a reliable technical guarantee for the health management of rotating machinery blades.

[0025] The prerequisite for deep transfer learning is the construction of source and target domains that are correlated but have different distributions. To address the scarcity of real-world crack samples from industrial settings, this invention constructs a high-fidelity source domain dataset by acquiring measured signals under multiple operating conditions and different crack parameters, providing sufficient labeled samples for transfer learning. The specific implementation is as follows:

[0026] 1) Source domain dataset acquisition ( ) To address the challenge of obtaining full lifecycle crack data in industrial settings, this invention first establishes a high-fidelity simulated crack test bench.

[0027] Physical Model Construction: Specimens consistent with the actual turbine blade material (such as titanium alloy TC4 or high-temperature alloy GH4169) are selected. Using wire EDM or fatigue prefabrication techniques, cracks of varying severity are prefabricated in key stress concentration areas of the blade (blade root, blade middle, and blade tip). Traditional fault diagnosis often uses readily available public datasets or general-purpose materials. However, the dynamic characteristics of turbine blades are highly correlated with material properties. This invention, through physical model construction, minimizes the original distribution differences between the source domain (simulation) and the target domain (field) at the physical level, ensuring physical consistency for subsequent feature mining and making the model more valuable for industrial reference.

[0028] Crack parameterization definition: The crack severity set is defined as follows: , in, This is a set of crack severity. For the first i Each crack length sample value For the first j Each crack depth sample value This is the minimum critical value for monitoring crack length. This represents the maximum critical value for monitoring crack length. This is the minimum critical value for monitoring crack depth. This is the maximum critical value for monitoring crack depth.

[0029] In this embodiment, Take 0.1mm, Take 10mm; Take 0.1mm, Take 5mm. 0.1mm is a typical dividing point for the blade from microscopic initiation to macroscopic expansion. When the crack exceeds 10mm, the blade has entered the unstable expansion stage, which is a serious fault that requires immediate shutdown.

[0030] In this embodiment, the preferred crack length step is 0.5 mm and the depth step is 0.5 mm to cover the entire process from microcrack initiation to macrocrack propagation.

[0031] Construction of a multi-field coupled operating condition simulation environment: In order to simulate real operating conditions, a set of rotational speeds is set. and load set .

[0032] Signal acquisition model: Let the source domain signal acquired by the sensor be... To capture weak acoustic emissions and high-frequency vibration characteristics, the sampling frequency... It must strictly satisfy Shannon's sampling theorem, and preferably... .

[0033] Construct source domain dataset ,in, This represents the nth source domain signal sample acquired by the sensor, and the blade specimen corresponding to this sample has a crack severity. ; For the corresponding fault labels, 0 represents healthy, 1- K These represent different levels of crack categories. K The maximum index value representing the crack condition category. In this embodiment, the total number of samples in the source domain is [number]. .

[0034] The complete construction process of the above source domain dataset is as follows: Figure 3 As shown. Figure 3 It demonstrates the entire process from specimen preparation, crack pre-fabrication, working condition simulation to signal acquisition and dataset integration, and achieves efficient construction of high-fidelity labeled samples.

[0035] 2) Obtaining the target domain dataset ( ) Blade monitoring signals from actual service units in industrial settings were collected. Due to the lack of clearly labeled field data, the target domain dataset was defined as follows: ,in The number of samples in the target domain.

[0036] Based on the above problems, the blade crack fault feature mining and early warning method provided in this embodiment of the invention further includes 3) signal preprocessing and tensor construction, the preprocessing including the following steps: To improve the signal-to-noise ratio and standardize the input, the blade signals acquired by the sensor, i.e., all the raw signals, are processed. That is, the measured signal and the monitored signal, and perform the following operations: Adaptive filtering: A Kalman filter is used to estimate and filter out Gaussian white noise and power frequency interference. The state equation and observation equation are defined as follows:

[0037] , in This represents the actual signal state. For observing signals, For process and observation noise, This represents the state transition matrix, used to describe the evolution of the system state from the previous time step to the current time step. The observation matrix is ​​used to describe the mapping relationship between the real state and sensor observations.

[0038] Normalization: To eliminate the influence of dimensions, the acquired signal is first denoised and then normalized, i.e., Z-Score normalization is performed on the signal segments. , in, This represents the original signal to be processed. This represents the normalized signal. The mean of the signal. The standard deviation of the signal. To prevent the division by zero of minute quantities.

[0039] Data augmentation: augmenting the source domain dataset The following enhancements are applied to simulate potential operating condition fluctuations in the target domain, thereby improving the robustness of the model. Data augmentation includes: Amplitude scaling: Randomly adjust the signal amplitude to 0.8×~1.2× of the original value to simulate sensitivity fluctuations caused by sensor contamination and installation deviations; Time stretching: The signal length is adjusted to 0.9×~1.1× of the original length using linear interpolation to simulate the signal timing changes caused by small fluctuations in rotational speed; Random noise superposition: Gaussian white noise with a signal-to-noise ratio of 10dB~20dB is added to simulate the uncertainty of random interference in the field; Feature domain enhancement: Wavelet packet decomposition of the signal is performed, and the energy distribution of each frequency band is randomly adjusted to simulate the characteristic changes of different crack propagation stages.

[0040] The preprocessing provided in this embodiment of the invention is not a simple application of conventional preprocessing. Instead, it addresses the issues of weak early-stage crack signals in blades, complex field noise, and inconsistent source / target domain distributions by specifically designing preprocessing and enhancement steps. Specifically, on the one hand, Kalman filtering is used to improve the signal-to-noise ratio of weak crack features, and Z-score standardization is used to unify the input scale under different operating conditions and sensor conditions. On the other hand, amplitude scaling, time stretching, random noise superposition, and wavelet packet feature domain enhancement are used to specifically simulate sensitivity fluctuations, speed disturbances, random interference, and crack propagation stage changes in the target domain, thereby reducing the input distribution differences between the source and target domains and improving the stability, robustness, and early-stage microcrack identification accuracy of the subsequent cross-domain transfer model.

[0041] To deeply mine the subtle early crack features hidden in signals, this invention designs a feature extractor based on a deep residual network (ResNet). By using residual connections, it solves the gradient vanishing problem during deep network training, achieving a precise mapping from the original signal to high-dimensional fault features. The network structure is as follows: Figure 4 As shown.

[0042] Figure 4 The diagram shows the structure of a deep residual feature extraction network. This structure consists of an input convolutional layer, stacked residual blocks, and a global pooling layer. The residual blocks are combined with the identity mapping short-circuit connection through the residual mapping Conv1D→BN→ReLU→Conv1D→BN. After ReLU activation, weak fault features are retained. Finally, a high-dimensional and compact fault feature vector is output through global pooling, providing high-quality feature input for cross-domain transfer.

[0043] To address the high-dimensional and non-stationary characteristics of blade vibration signals, this invention provides a one-dimensional deep residual network (1D-ResNet) as a feature extractor. .

[0044] 1) Network topology The network consists of input convolutional layers, stacked residual blocks, and global pooling layers.

[0045] Input layer: Large convolutional kernels are used to perform preliminary sampling and feature mapping on the original high-frequency signal.

[0046] Residual learning units: To address the vanishing gradient problem in deep networks, residual connections are introduced. Assume the... The input of each residual block is , No. The output of each residual block is Its mathematical expression is:

[0047] , in, It is the ReLU activation function; This is the residual mapping function, which includes the serialization operation: Conv1D->BN->ReLU->Conv1D->BN.

[0048] In this structure, This represents the convolution kernel weights. Residual connections enable... Include The identity mapping component ensures that weak crack impact features are preserved in deep networks without being dissipated as the number of layers increases.

[0049] High-dimensional feature output: after The residual blocks are processed by compressing the time dimension through global average pooling, and the output feature vector is then obtained. .

[0050] 2) Pre-training strategy Before performing transfer learning, a labeled source domain dataset is used. The feature extractor is pre-trained. The optimization objective is to minimize the source domain classification loss:

[0051] , in, This represents the source domain classification loss function during the pre-training phase. Feature extractor Network parameters, Fault classifier Network parameters, This represents a labeled source domain dataset. Indicates the source domain input sample. This represents the fault category label corresponding to the input sample in the source domain. This indicates that the expected value is calculated for sample pairs in the source domain dataset. This represents the deep feature representation extracted by the feature extractor from the source domain input sample; This represents the category prediction probability output by the fault classifier based on deep features. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. Indicates category index, This represents the upper bound of the total number of crack failure categories, where 0 corresponds to a healthy state, and 1 to... K Corresponding to different levels of crack categories.

[0052] Through this step, the network learns the basic ability to represent crack features.

[0053] Optimizer and Training Strategy: Optimizer: The Adam optimizer is used, with reasonable initial learning rate and weight decay coefficients set to prevent model overfitting.

[0054] Learning rate scheduling: A cosine annealing learning rate strategy is adopted, and an early stopping mechanism is introduced to improve the model's ability to escape local optima.

[0055] Training parameters: Set the appropriate batch size and number of training rounds, and use mixed precision training to accelerate the training process.

[0056] Model regularization: Adding Dropout layers and label smoothing techniques suppresses overfitting and improves the model's generalization ability.

[0057] In actual operation and simulation, the number of microcrack samples is far less than that of normal or severe crack samples. This "class imbalance" causes the model to tend to ignore early weak features.

[0058] To address the aforementioned issues, this embodiment of the invention further includes class imbalance optimization: The SMOTE algorithm is introduced to generate synthetic data of microcrack samples, expanding the number of scarce microcrack samples, improving the class distribution of the training data, and further enhancing the accuracy of microcrack identification. During the construction of the source domain dataset, to address the class imbalance problem caused by the scarcity of microcrack class samples, the SMOTE algorithm is introduced to perform linear interpolation synthesis in the feature space to expand the decision samples for microcrack classes. The balanced dataset is then input into 1D-ResNet, and end-to-end optimization is performed using a weighted loss function, thereby significantly improving the model's accuracy in detecting early, weak fault features.

[0059] Model performance validation: After training, the feature extractor performance is validated using the source domain test set to meet the accuracy requirements for health / crack classification and microcrack identification. Specifically, the binary classification accuracy for health / crack should ideally reach above 99%. The accuracy of identifying the first-level microcrack category needs to reach over 95%. The shallow parameters of the feature extractor (convolutional layers and the first 1 / 3 of the residual learning units) are fixed, while the deep parameters (the last 2 / 3 of the residual learning units and the global average pooling layer) are retained for subsequent transfer learning fine-tuning to ensure the model adapts to the target domain data features.

[0060] To adapt to dynamic changes in on-site operating conditions (such as unit aging, operating condition adjustments, and sensor performance degradation), this embodiment of the invention constructs an online model update and optimization mechanism to ensure long-term stable operation of the model, specifically implemented as follows: Uncertainty Sampling: Calculating the Prediction Entropy of Samples in the Target Domain Samples with high entropy values ​​(i.e., the least certain model information) are selected, prompting manual verification or calibration during downtime maintenance. The model can automatically identify operational data it is unsure about and evolve through minimal manual calibration, ensuring long-term stability throughout the entire lifecycle monitoring process. Samples with high entropy values ​​are defined as the top 5%–10% of all target domain samples collected over a period (e.g., an inspection cycle), sorted by entropy value from highest to lowest.

[0061] Incremental Fine-tuning: Newly labeled samples are added to the training set, the shallow parameters of the feature extractor are fixed, and only the deep residual blocks and classifier are fine-tuned, allowing the model to continuously adapt to new operating conditions. 1. Differentiated parameter freezing strategy: During incremental fine-tuning, the model fixes the shallow parameters of the feature extractor and updates only the deep residual blocks and classifier. By freezing this fundamental knowledge, the core diagnostic logic of the model is preserved, allowing only the deep network to fit the subtle shifts brought about by the new operating conditions. 2. Hybrid training mechanism based on sample replay: This embodiment of the invention does not simply use new samples for overlay training, but adds high-entropy new samples that have been manually reviewed or calibrated during shutdown to the original training set. This means that the incremental process is carried out on a combination of historical core samples and newly added samples. In this way, the model maintains the predicted probability distribution of the old operating conditions (such as laboratory standard operating conditions) at the mathematical level through the loss function.

[0062] Model degradation monitoring and version rollback system: Once the system detects that the model's accuracy drops below a threshold when handling certain older scenarios, it will determine that model degradation has occurred and automatically trigger a global retraining process based on the complete dataset. Simultaneously, by establishing a model version management mechanism, the system records parameters and performance metrics at different stages, supporting version rollback in extreme cases.

[0063] Model degradation monitoring: Regularly evaluate the model's diagnostic performance in the target domain. When the performance index drops below the threshold, the model is determined to have degraded, and the retraining process is automatically triggered to ensure that the model maintains high diagnostic accuracy in the long term.

[0064] Model version management: Establish a model version management mechanism to record model parameters and performance indicators at different stages, support model version backtracking, and facilitate the analysis of the impact of changes in operating conditions on model performance.

[0065] In this invention, the pre-training process using labeled source domain data is specifically improved for the early-stage crack identification scenario in blades. First, by minimizing the source domain classification loss, the feature extractor possesses basic crack representation capabilities before entering cross-domain transfer, avoiding training instability caused by direct domain alignment under random initialization conditions. Second, addressing the significant scarcity of micro-crack samples and the severe imbalance in class distribution, SMOTE oversampling is introduced during pre-training, coupled with a weighted loss function for end-to-end optimization, enhancing the model's sensitivity to early, weak crack features. Third, after pre-training, shallow parameters are fixed while deep parameters are retained for subsequent transfer fine-tuning, thus balancing the preservation of general basic features with target domain adaptability. Through these improvements, pre-training is no longer just a routine initialization step, but provides a stable, separable feature base sensitive to micro-cracks for subsequent cross-domain transfer, thereby improving the accuracy and robustness of target domain fault identification.

[0066] Due to differences in environmental noise, sensor installation locations, and unit operating conditions, the source domain distribution varies. Distribution with target domain There are significant differences, namely .

[0067] To achieve knowledge transfer from the source domain to the target domain, embodiments of this invention construct a feature extractor. Fault classifier Domain discriminant A unified framework is proposed, and a dual constraint mechanism is introduced. While simple adversarial learning can align distributions, the training process is often unstable; and simple statistical alignment struggles to handle complex nonlinear features. Adversarial loss is used to address this issue. This forces the network to extract domain-invariant features while utilizing the maximum mean difference. By explicitly bringing the distribution centers of the two domains closer together, this double-layered protection significantly enhances the robustness and diagnostic accuracy of cross-condition migration. The model structure is as follows: Figure 5 As shown.

[0068] 1) Model Architecture Definition Feature extractor Mapping the input signal into a feature vector After incorporating the attention mechanism, it can output domain-invariant and fault-sensitive features. The input signal refers to the one-dimensional temporal tensor that has been preprocessed and directly fed into the 1D-ResNet feature extractor, including the source domain signal. and target domain signal .

[0069] The specific attention fusion mechanism is as follows: Feature extractor First, the input signal is convolutionally processed and residual features are extracted using 1D-ResNet to obtain intermediate feature maps. ; Then, the intermediate feature map is input into the attention module to calculate the weight vector for each channel. ; Then, multiply the weights of each channel by the intermediate feature map channel by channel to obtain the weighted feature map. ; Finally, global pooling or flattening is performed on the weighted feature map to obtain the feature vector. .

[0070] Feature vector This refers to the high-dimensional representation of the input signal after nonlinear mapping through a deep residual network. Domain invariance refers to eliminating the distributional differences between simulation and field conditions. Through adversarial game theory implemented by the gradient inversion layer and MMD statistical distance constraints, background noise caused by sensor installation deviations, environmental noise differences, or varying degrees of unit aging is forcibly eliminated. Constraints. Fault-sensitive features refer to those components that can accurately capture the nonlinear, non-stationary, weak characteristics caused by early-stage blade cracks, distinguishing between healthy and cracked types. and constraint.

[0071] Fault classifier : The feature vector This is mapped to fault category probabilities. The fault classifier uses a fully connected layer to process the input feature vector. Mapped to a predefined fault label space The classifier uses a softmax function at the end to transform the output into a probability vector that sums to 1. Each of the components The representative model's confidence level in considering the current blade to belong to the i-th state.

[0072] Domain discriminator Determine the feature vector The determination is whether the data originates from the source domain or the target domain. The logic is essentially a binary classification. First, explicit domain labels are pre-defined for the data (source domain simulated data is labeled 0, target domain field data is labeled 1). The domain discriminator calculates the feature vector by minimizing the domain classification error. The probability of belonging to the target domain; if the output is close to 1, then the feature is determined to originate from the scene.

[0073] 2) Dual-constraint loss function The overall objective function constructed in this invention It consists of three parts: , in, and These are the adversarial loss weights and the statistical distance weights, respectively. Adversarial loss weights ,in p This represents the current training progress (current iteration count / total iteration count).

[0074] The statistical distance weights are usually given through empirical verification and are typically a small scaling factor (e.g., 0.1 or 0.05).

[0075] (1) Fault classification constraints ) To ensure the separability of features in the source domain, a weighted cross-entropy loss is used to handle class imbalance.

[0076] , in, A high weighting coefficient is set for the microcrack category. ,in, N The total number of samples, K This represents the total number of fault categories. n i For the first i The number of samples in the microcrack class. The effective signal samples in the microcrack initiation stage are often far fewer than the healthy data during stable operation. Therefore, through this calculation logic, the weight of the microcrack class will be automatically increased. Indicates the total number of samples in the source domain. This represents the nth source domain input sample. Indicates the first The true fault category label corresponding to each source domain input sample This indicates that for the input sample The model predicts that it belongs to the true category. The conditional probability.

[0077] (2) Domain adversarial constraints ) A gradient inversion layer is introduced to realize the minimax game between the feature extractor and the domain discriminator.

[0078] The gradient reversal layer is placed between the feature extractor and the domain discriminator. During forward propagation, the gradient reversal layer acts as an identity mapping function, without modifying the input feature vector to ensure that the domain discriminator receives the original feature distribution and performs domain classification. During backpropagation, when the error is propagated back from the domain discriminator to the feature extractor, the gradient reversal layer truncates the gradient and multiplies it by a negative weight coefficient. This transforms the gradient direction, which originally reduced the domain classification error, into an instruction that increases the domain classification error.

[0079] By introducing a gradient inversion layer, the feature extractor no longer works with the domain discriminator to differentiate features. Instead, the extracted features are so similar in distribution that the discriminator cannot distinguish their origins, thus maximizing the domain classification error.

[0080] This game-theoretic process ultimately forces the feature extractor to strip away redundant information that carries specific operating condition labels, sensor biases, or environmental noise. The remaining features are domain-invariant features, which preserve the common physical characteristics of blade crack failures, ensuring that the model can still accurately identify them even when faced with unseen field target domain data, thanks to the knowledge learned from the source domain.

[0081] Domain discriminator The goal is to minimize the domain classification error and distinguish between the source and target domains.

[0082] Feature extractor It attempts to maximize the domain classification error to deceive the discriminator.

[0083] The mathematical expression is: , in, The domain labels are (source domain = 0, target domain = 1). The gradient is multiplied by during backpropagation using a gradient reversal layer. Thus forcing Extract domain-invariant features. Indicates the total number of samples in the source domain. Indicates the total number of samples in the target domain. Represents the source domain dataset. Represents the target domain dataset. Indicates input sample The feature vector extracted by the feature extractor The domain discriminator outputs the predicted probability that the sample comes from the target domain. This represents the domain label corresponding to the input sample. It is 0 when the sample comes from the source domain and 1 when the sample comes from the target domain.

[0084] (3) Statistical distribution constraints ) To enhance training stability, the maximum mean difference is explicitly introduced as a statistical constraint to bring the distribution centers of the two domains closer in the regenerating kernel Hilbert space (RKHS).

[0085] , in, Gaussian kernel mapping function This constraint can effectively align the higher-order statistical moments of the characteristic distribution. Indicates the total number of samples in the source domain. Indicates the total number of samples in the target domain. This represents the feature vector extracted from the nth source domain sample by the feature extractor. This represents the feature vector extracted by the feature extractor from the nth target domain sample.

[0086] 3) Optimization Algorithm A combination of mini-batch stochastic gradient descent and the Adam optimizer is employed. The parameter update strategy is as follows:

[0087] , By analyzing the parameters of the domain discriminator Fault classifier parameters and feature extractor parameters The process involves iterative, alternating updates. Specifically, in each round of training, the domain discriminator parameters are first updated. Minimize the domain adversarial loss, then update the fault classifier parameters. To minimize the fault classification loss, the feature extractor parameters are finally updated. To simultaneously minimize fault classification loss and statistical distribution loss, and to adversarially increase the difficulty of domain discrimination through gradient inversion mechanism. The domain discriminator gradually improves its ability to distinguish features between the source domain and the target domain. Under the combined effect of fault classification constraints, domain adversarial constraints, and statistical distribution constraints, the feature extractor gradually learns feature representations that combine fault sensitivity and domain invariance, ultimately enabling the model to have fault recognition capabilities in the target domain comparable to those in the source domain.

[0088] This invention provides a method for identifying and warning of blade crack fault features. The method inputs real-time monitoring signals from the field into a trained cross-domain transfer diagnostic model to obtain a fault probability vector. Multi-level warnings are then performed using an adaptive threshold. The specific steps are as follows: 1) Fault probability calculation and adaptive threshold Signal Acquisition: The blade operation signals are acquired in real time through fiber optic sensors, accelerometers, and acoustic emission sensors installed on the unit casing at the field site. The sensor layout is consistent with the test platform (to ensure characteristic consistency), the sampling frequency is maintained at 1MHz, and industrial-grade data acquisition modules are used to store the raw data, supporting continuous operation around the clock. Real-time monitoring signals on site Input the trained network to obtain the fault probability vector Define the crack probability. .

[0089] Industrial environments are complex, and environmental noise and fluctuations in operating conditions can cause natural drift in prediction probabilities, making fixed thresholds highly susceptible to false alarms. Dynamically adjusting the threshold based on background noise and unit status effectively distinguishes between normal fluctuations and abnormal cracks, significantly reducing the false alarm rate. Considering the volatility of on-site operating conditions, this invention abandons fixed thresholds and adopts an adaptive threshold based on statistical process control (SPC). :

[0090] , in, This represents the mean of the predicted probabilities of healthy samples within the sliding time window. for t Adaptive threshold at time, The standard deviation of the predicted probability of healthy samples within the sliding time window. This is for the safety factor.

[0091] like This is considered normal. like It was determined to be a Level 1 warning (early microcracks). like The condition was determined to be a Level 2 alarm (severe crack).

[0092] 2) Visualization of t-SNE manifolds To enable maintenance personnel to intuitively understand the equipment status, the t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm is used to map the 768-dimensional feature vector to a two-dimensional plane.

[0093] , On the monitoring interface, the cluster centers (healthy clusters, cracked clusters) of the source domain data are plotted as a background, and the coordinates of the current monitoring point are projected in real time. If the current point gradually deviates from the center of the healthy cluster and drifts towards the cracked cluster, a visually intuitive early warning can be issued.

[0094] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for identifying and predicting blade crack fault features, characterized in that, Includes the following steps: Acquire measured signals of simulated cracks in blades under multiple operating conditions and different crack parameters to construct a source domain dataset. Acquire blade monitoring signals of actual units in service in industrial sites to construct an unlabeled target domain dataset. The source domain dataset and the target domain dataset form an input tensor. A one-dimensional deep residual network is used as a feature extractor. The high-dimensional fault features in the input tensor are extracted through the residual learning unit in the network. The feature extractor is pre-trained using the source domain dataset to obtain the pre-trained feature extractor. A unified framework is constructed, comprising the pre-trained feature extractor, fault classifier, and domain discriminator. The fault classifier predicts fault categories and calculates classification loss for the source domain dataset. The domain discriminator performs domain classification on features of the source and target domain datasets and calculates domain adversarial loss to force the feature extractor to extract domain-invariant features. Simultaneously, the maximum mean difference is used to explicitly bring the feature distribution centers of the source and target domain datasets closer. Knowledge transfer from the source domain to the target domain is achieved by jointly optimizing the classification loss, domain adversarial loss, and maximum mean difference, resulting in a trained cross-domain transfer diagnostic model. The real-time monitoring signal from the site is input into the trained cross-domain migration diagnostic model to obtain a fault probability vector, which is then combined with an adaptive threshold for multi-level early warning.

2. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The construction of the source domain dataset includes the following steps: Select specimens with the same material as the blades of the actual unit, and pre-induce cracks of different severity in the key stress concentration area; Define a set of crack severity, construct a multi-field coupled working condition simulation environment, collect blade vibration or acoustic emission signals through sensors, and construct a source domain dataset with fault labels. The set of crack severity is represented as follows: ， in, This is a set of crack severity. For the first i Each crack length sample value For the first j Each crack depth sample value This is the minimum critical value for monitoring crack length. This represents the maximum critical value for monitoring crack length. This is the minimum critical value for monitoring crack depth. This is the maximum critical value for monitoring crack depth.

3. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The measured and monitored signals are preprocessed to form source domain datasets and target domain datasets. The preprocessing includes the following steps: Kalman filtering is used to remove Gaussian white noise and power frequency interference from the measured and monitored signals to form a filtered signal. The filtered signal is Z-score normalized to form a normalized filtered signal; An enhancement operation is applied to the signal in the source domain dataset, the enhancement operation including at least one of amplitude scaling, time stretching, random noise superposition, and feature domain enhancement.

4. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The feature extractor is a temporal feature extraction network based on a one-dimensional deep residual network, and the mathematical expression of the residual learning unit is as follows: ， in, For the first The input of each residual block, For the first The output of each residual block Indicates the convolution kernel weights. It is the ReLU activation function; This is the residual mapping function.

5. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The feature extractor also includes an attention module, which is used to calculate the weight vector of each channel and multiply it with the intermediate feature map channel by channel to obtain a weighted feature map.

6. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, Before pre-training the feature extractor using the source domain dataset, the following steps are also included: The SMOTE algorithm is introduced to perform linear interpolation synthesis in the feature space where the high-dimensional fault features are located, generating synthetic samples of microcrack categories. The synthetic samples are added to the source domain dataset to expand the number of samples of microcrack categories, resulting in a balanced source domain dataset. The balanced source domain dataset is then input into a one-dimensional deep residual network and optimized end-to-end with a weighted loss function. The process of issuing multi-level early warnings also includes the following steps: Calculate the prediction entropy of the target domain dataset, select samples with entropy values ​​higher than a preset threshold for manual review and calibration, add the newly labeled samples to the training set, and incrementally fine-tune the cross-domain transfer diagnostic model. During incremental fine-tuning, the shallow parameters of the feature extractor are fixed, and the parameters of the deep residual block and the fault classifier are updated only.

7. The method for identifying and warning of blade crack fault features as described in claim 6, characterized in that, The classification loss uses weighted cross-entropy loss, expressed as: ， in, Weighting coefficients set for microcrack categories. ,in, N This represents the total number of samples; K Total number of fault categories; n i For the first i Number of samples with microcrack-like structures; Indicates the total number of samples in the source domain. This represents the nth source domain input sample. Indicates the first The true fault category label corresponding to each source domain input sample This indicates that for the input sample The model predicts that it belongs to the true category. The conditional probability.

8. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, A gradient inversion layer is provided between the domain discriminator and the feature extractor. The gradient inversion layer is an identity mapping during forward propagation and multiplies the gradient by a negative weight coefficient during backward propagation.

9. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The adaptive threshold is set using a statistical process control-based method, expressed as follows: ， in, This represents the mean of the predicted probabilities of healthy samples within the sliding time window. The standard deviation of the predicted probability of healthy samples within the sliding time window. This is for the safety factor.

10. The method for identifying and warning of blade crack fault features as described in claim 1, characterized in that, The multi-level early warning includes: when the fault probability is less than the adaptive threshold, it is determined to be normal; when the fault probability is greater than or equal to the adaptive threshold and less than the preset secondary threshold, it is determined to be a first-level early warning; when the fault probability is greater than or equal to the preset secondary threshold, it is determined to be a second-level alarm.

Images