Vibration-based method and device for early fault identification of steam turbine generators

Abstract

This application relates to the field of generator fault identification technology, and discloses a method and apparatus for early fault identification of steam turbine generators based on vibration. The method includes: acquiring training samples; extracting multiple continuous vibration signal segments from each vibration signal sequence sample, constructing grayscale images based on each continuous vibration signal segment, and generating a grayscale image set; training an initial convolutional neural network classifier using an unsupervised sequence segmentation method on each grayscale image set to obtain a target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each grayscale image set, and uses the pseudo-labels to train the initial convolutional neural network classifier; acquiring the vibration signal to be identified in real time, dynamically constructing the vibration signal segment to be identified based on the acquired vibration signal, constructing the grayscale image to be identified based on it, and inputting the grayscale image to be identified into the target convolutional neural network classifier to determine whether a fault exists.

Description

Technical Field

[0001] This application relates to the field of generator fault identification technology, and in particular to a method and apparatus for early fault identification of steam turbine generators based on vibration. Background Technology

[0002] Against the backdrop of the accelerated construction of new power systems, thermal power units are undergoing a significant transformation from power-guaranteed main power sources to regulation-oriented power sources. With deep peak shaving to 20% of rated load becoming the norm, the operational safety of these units faces unprecedented challenges. As a core power equipment, the vibration state of the steam turbine generator directly reflects the health of the unit; therefore, vibration analysis-based fault detection methods have become a key technical means to ensure the safe and stable operation of the power system. Currently, technological development in this field has mainly evolved from traditional signal processing to data-driven methods.

[0003] Existing fault identification methods based on traditional signal processing heavily rely on expert experience and prior knowledge, requiring the manual creation of fault feature databases and parameter readjustment under different operating conditions, resulting in poor adaptability. Furthermore, although data-driven methods have gained attention in recent years due to their powerful feature extraction capabilities, existing methods typically train models directly using full lifecycle data. This data often lacks accurate labels, making it impossible to determine the specific time points of early fault occurrences in historical data, leading to unsatisfactory early fault identification results. Summary of the Invention

[0004] In view of this, this application provides a vibration-based method and apparatus for early fault identification of steam turbine generators, a storage medium, and a computer device. It fully utilizes the temporal characteristics of vibration signals from historical full lifecycle data, transforming the early fault state identification problem into a vibration signal sequence segmentation problem. An intelligent optimization algorithm automatically determines the optimal early fault occurrence time, achieving high-precision early fault identification without relying on manual labeling, reducing the cost and subjective error of manual labeling. Simultaneously, the use of convolutional neural networks to process grayscale images constructed based on vibration signals effectively captures complex patterns and subtle changes in the vibration signals, improving the accuracy and robustness of early fault identification. Furthermore, by acquiring vibration signals in real time and dynamically constructing grayscale images, the state of the steam turbine generator to be identified can be monitored and faults identified in a timely manner, helping to detect problems in their early stages, take preventative measures, reduce damage to the steam turbine generator, and improve the operational reliability and safety of the steam turbine generator.

[0005] According to one aspect of this application, a vibration-based method for early fault identification of a steam turbine generator is provided, comprising:

[0006] Acquire training samples, wherein the training samples include vibration signal sequence samples of multiple steam turbine generators throughout their historical life cycle;

[0007] For each vibration signal sequence sample, multiple continuous vibration signal segments are extracted from the vibration signal sequence sample, and grayscale images are constructed based on each continuous vibration signal segment. Based on each grayscale image, a set of grayscale images of the vibration signal sequence sample is generated.

[0008] Using each set of grayscale images, an unsupervised sequence segmentation method is used to train an initial convolutional neural network classifier to obtain a target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each set of grayscale images, and uses the generated pseudo-labels to train the initial convolutional neural network classifier.

[0009] The vibration signal of the turbine generator to be identified is acquired in real time, and a vibration signal segment to be identified is dynamically constructed based on the acquired vibration signal. A grayscale image to be identified is constructed based on the vibration signal segment to be identified, and the grayscale image to be identified is input into the target convolutional neural network classifier to determine whether the turbine generator to be identified has a fault.

[0010] According to another aspect of this application, a vibration-based early fault identification device for steam turbine generators is provided, comprising:

[0011] The sample acquisition module is used to acquire training samples, wherein the training samples include vibration signal sequence samples of multiple steam turbine generators throughout their historical life cycle.

[0012] The image construction module is used to extract multiple continuous vibration signal segments from each vibration signal sequence sample, construct grayscale images based on each continuous vibration signal segment, and generate a set of grayscale images of the vibration signal sequence sample based on each grayscale image.

[0013] The classifier training module is used to train an initial convolutional neural network classifier using an unsupervised sequence segmentation method on each set of grayscale images to obtain a target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each set of grayscale images and uses the generated pseudo-labels to train the initial convolutional neural network classifier.

[0014] The fault identification module is used to acquire the vibration signal of the turbine generator to be identified in real time, dynamically construct the vibration signal segment to be identified based on the acquired vibration signal, construct the grayscale image to be identified based on the vibration signal segment to be identified, and input the grayscale image to be identified into the target convolutional neural network classifier to determine whether the turbine generator to be identified has a fault.

[0015] According to another aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described vibration-based early fault identification method for steam turbine generators.

[0016] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described vibration-based early fault identification method for steam turbine generators.

[0017] By employing the aforementioned technical solutions, this application provides a vibration-based method and apparatus for early fault identification of steam turbine generators, along with a storage medium and computer equipment. This method fully utilizes the temporal characteristics of vibration signals from historical full-lifecycle data, transforming the early fault state identification problem into a vibration signal sequence segmentation problem. An intelligent optimization algorithm automatically determines the optimal early fault occurrence time, achieving high-precision early fault identification without relying on manual labeling. This reduces the cost and subjective errors of manual labeling. Furthermore, the use of convolutional neural networks to process grayscale images constructed from vibration signals effectively captures complex patterns and subtle changes in the vibration signals, improving the accuracy and robustness of early fault identification. In addition, the real-time acquisition of vibration signals and dynamic construction of grayscale images allows for timely monitoring and fault identification of the steam turbine generator, facilitating early detection of problems and enabling proactive measures to reduce damage to the generator and improve its operational reliability and safety.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1A schematic flowchart of a vibration-based early fault identification method for steam turbine generators provided in an embodiment of this application is shown.

[0021] Figure 2 A flowchart illustrating another vibration-based early fault identification method for steam turbine generators provided in an embodiment of this application is shown.

[0022] Figure 3 A schematic diagram of the structure of a vibration-based early fault identification device for a steam turbine generator provided in an embodiment of this application is shown.

[0023] Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0024] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0025] This embodiment provides a vibration-based method for early fault identification of steam turbine generators, such as... Figure 1 As shown, the method includes:

[0026] Step 101: Obtain training samples, wherein the training samples include vibration signal sequence samples of multiple steam turbine generators throughout their historical life cycle.

[0027] Step 102: For each vibration signal sequence sample, extract multiple continuous vibration signal segments from the vibration signal sequence sample, construct a grayscale image based on each continuous vibration signal segment, and generate a grayscale image set of the vibration signal sequence sample based on each grayscale image.

[0028] Step 103: Using each grayscale image set, an unsupervised sequence segmentation method is used to train the initial convolutional neural network classifier to obtain the target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each grayscale image set, and uses the generated pseudo-labels to train the initial convolutional neural network classifier.

[0029] Step 104: Real-time acquisition of vibration signals of the turbine generator to be identified, dynamic construction of vibration signal segments to be identified based on the acquired vibration signals, construction of grayscale images to be identified based on the vibration signal segments to be identified, input of the grayscale images to be identified into the target convolutional neural network classifier, and determination of whether the turbine generator to be identified has a fault.

[0030] In this embodiment, training samples are first acquired. These training samples may include multiple vibration signal sequence samples, each containing a vibration signal sequence of a steam turbine generator throughout its entire historical lifecycle. These samples cover vibration signal data from the start of operation to failure and shutdown of the steam turbine generator, ensuring the comprehensiveness of the training data and providing data support for subsequent analysis and classifier training.

[0031] Next, for each vibration signal sequence sample, multiple continuous vibration signal segments are extracted from it. Then, a grayscale image is constructed for each continuous vibration signal segment. These grayscale images can visualize the temporal characteristics of the vibration signal, thereby capturing potential fault modes. Finally, all grayscale images are combined into a grayscale image set to represent the vibration history of the turbine generator. It is important to note that when extracting continuous vibration signal segments, every stage of the entire life cycle should be captured to ensure that the extracted multiple continuous vibration signal segments cover all stages of the turbine generator's entire life cycle.

[0032] Then, using these grayscale image sets, an unsupervised sequence segmentation method is employed to train the initial convolutional neural network classifier. The unsupervised sequence segmentation method automatically determines the optimal early fault occurrence time for each grayscale image set by analyzing the variation patterns of each grayscale image. Based on this, pseudo-labels are generated for each grayscale image in the set. These pseudo-labels, in the absence of real labels, generate supervisory signals for training the initial convolutional neural network classifier using the optimal early fault occurrence time found by the unsupervised sequence segmentation method itself. This allows the classifier to gradually learn how to identify early fault characteristics of the turbine generator, ultimately leading to the target convolutional neural network classifier.

[0033] In the real-time identification phase, vibration signals from the turbine generator to be identified are continuously acquired, and vibration signal segments to be identified are dynamically constructed based on these acquired signals. Here, the length of the vibration signal segment to be identified can be consistent with the length of the continuous vibration signal segments in the aforementioned classifier training process, and the construction frequency of the vibration signal segment to be identified can be determined according to actual needs. Next, using the same method, the constructed vibration signal segment to be identified is converted into a grayscale image to be identified, and this grayscale image is input into the trained target convolutional neural network classifier. Based on the learned features, the target convolutional neural network classifier determines whether there is a fault in the turbine generator to be identified within the time period corresponding to the vibration signal segment to be identified, thereby achieving early fault detection.

[0034] By applying the technical solution of this embodiment, the temporal characteristics of vibration signals from historical full lifecycle data are fully utilized to transform the problem of early fault state identification into a vibration signal sequence segmentation problem. An intelligent optimization algorithm automatically determines the optimal early fault occurrence time, achieving high-precision early fault identification without relying on manual labeling. This reduces the cost and subjective error of manual labeling. Simultaneously, the use of convolutional neural networks to process grayscale images constructed based on vibration signals effectively captures complex patterns and subtle changes in the vibration signals, improving the accuracy and robustness of early fault identification. Furthermore, the method of acquiring vibration signals in real-time and dynamically constructing grayscale images allows for timely monitoring and fault identification of the turbine generator's status, facilitating early detection of problems and enabling proactive measures to reduce damage to the turbine generator and improve its operational reliability and safety.

[0035] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, another vibration-based method for early fault identification of steam turbine generators is provided, such as... Figure 2 As shown, the method includes:

[0036] Step 201: Obtain training samples, wherein the training samples include vibration signal sequence samples of multiple steam turbine generators throughout their historical life cycle.

[0037] In this embodiment, a highly sensitive piezoelectric accelerometer can be installed on the turbine generator bearing housing, frame, or other parts to collect vibration signals from the turbine generator, obtaining a vibration signal sequence sample throughout its entire historical lifecycle. Setting a reasonable sampling frequency ensures the integrity and accuracy of the collected vibration signals. Since the main vibration frequency components of the turbine generator are distributed in the range of 10Hz to 2kHz, the sampling frequency needs to meet the requirements of the Nyquist sampling theorem, i.e., the sampling frequency should not be less than twice the highest frequency of the vibration signal. Therefore, the sampling frequency can be determined accordingly.

[0038] Step 202: For each vibration signal sequence sample, extract multiple continuous vibration signal segments from the vibration signal sequence sample.

[0039] Step 203: For each continuous vibration signal segment, perform a fast Fourier transform on the continuous vibration signal segment to obtain the corresponding frequency domain signal segment, extract the amplitude corresponding to each frequency component from the frequency domain signal segment to obtain a frequency domain feature sequence, and perform normalization processing on the frequency domain feature sequence to obtain a standardized frequency domain feature sequence.

[0040] Step 204: For each feature value in the standardized frequency domain feature sequence, calculate the corresponding gray value based on the feature value using a preset rule, and determine the target position of the gray value in the grayscale image based on the position of the feature value in the standardized frequency domain feature sequence, and map the gray value to the target position.

[0041] Step 205: When the gray values ​​corresponding to each feature value in the standardized frequency domain feature sequence are mapped to the grayscale image, the grayscale image corresponding to the continuous vibration signal segment is obtained.

[0042] In this embodiment, each extracted continuous vibration signal segment can be processed using a Fast Fourier Transform. This transforms the continuous vibration signal segment from the time domain to the frequency domain, because many mechanical faults manifest as amplitude variations in specific frequency components (such as rotational frequency, harmonics, bearing throughput frequency, etc.). Through this transformation, a frequency domain signal segment that clearly reflects the intensity of each frequency component can be obtained.

[0043] Next, the amplitude corresponding to each frequency component is extracted from the aforementioned frequency domain signal segment, and these amplitudes are arranged in frequency order to form a frequency domain feature sequence. Subsequently, this frequency domain feature sequence is normalized, specifically by scaling all amplitudes to a fixed range (e.g., between 0 and 255). This normalization process can eliminate the influence of overall amplitude fluctuations, making the generated grayscale images consistent and comparable.

[0044] Furthermore, for each feature value in the standardized frequency domain feature sequence, it is converted into a grayscale value according to a preset rule. Simultaneously, based on the position of this feature value in the standardized frequency domain feature sequence (which can be its corresponding frequency component index), its target position in the final grayscale image is determined (specifically, its pixel coordinates in the horizontal and vertical directions). After the calculation is complete, this grayscale value can be filled into the corresponding target position in the grayscale image.

[0045] Once all feature values ​​in the entire normalized frequency domain feature sequence have been successfully converted into grayscale values ​​and mapped to the corresponding target positions in the grayscale image according to their frequency components, a grayscale image representing the frequency domain features of the continuous vibration signal segment is obtained.

[0046] In one specific embodiment, in order to convert a continuous vibration signal segment into a two-dimensional grayscale image of size N×N, a segment of length 2×N can be randomly selected. 2 The continuous vibration signal segment is converted into a frequency domain signal segment using Fast Fourier Transform (FFT), mainly including characteristic parameters such as frequency components and amplitude. Normalization is then applied to the frequency domain feature sequence, unifying the characteristic parameters to the interval [0, 1], resulting in a signal of length N.2 The standardized frequency domain feature sequence.

[0047] In a specific embodiment, the preset rules are as follows:

[0048] ;

[0049] in, Represents the first grayscale image. m Line number n The grayscale values ​​of the column pixels, m, n=1,...,N. , Representing the normalized frequency domain feature sequences respectively i The maximum and minimum values ​​in the range. Represents the normalized frequency domain feature sequence i , Represents the normalized frequency domain feature sequence i The Middle x The eigenvalues ​​of each frequency component.

[0050] For example, from the vibration signal sequence sample of a steam turbine generator throughout its entire historical life cycle, a continuous vibration signal segment of length 512 is extracted. Using Fast Fourier Transform, the continuous vibration signal segment is converted into a frequency domain signal segment, and after normalization, a standardized frequency domain feature sequence of length 256 can be obtained. A 16×16 two-dimensional grayscale image can then be obtained using the aforementioned preset rules.

[0051] This application transforms a one-dimensional, abstract, standardized frequency domain feature sequence into an intuitive two-dimensional grayscale image. This allows for full utilization of the powerful image feature recognition capabilities of convolutional neural networks, automatically learning frequency patterns associated with early faults. Compared to directly processing standardized frequency domain feature sequences, the image format better preserves the spatial (i.e., frequency) distribution of features. Furthermore, due to the normalization process, this method is insensitive to changes in the absolute magnitude of amplitudes, focusing more on relative changes in frequency distribution. This enhances the robustness of the target convolutional neural network classifier, enabling it to more effectively capture weak spectral features indicating early faults.

[0052] Step 206: Generate a set of grayscale images of the vibration signal sequence samples based on each grayscale image.

[0053] Step 207: Determine the optimal early fault occurrence time for each grayscale image set based on the unsupervised sequence segmentation method and the initial convolutional neural network classifier.

[0054] Step 208: For each grayscale image set, based on the optimal early fault occurrence time corresponding to the grayscale image set and the sampling time of the continuous vibration signal segment corresponding to each grayscale image in the grayscale image set, the grayscale images in the grayscale image set are divided into a normal subset and a fault subset. A pseudo-label indicating a normal state is set for each grayscale image in the normal subset, and a pseudo-label indicating a fault state is set for each grayscale image in the fault subset.

[0055] Step 209: Train the initial convolutional neural network classifier based on the grayscale images with pseudo-labels in each grayscale image set to obtain the target convolutional neural network classifier.

[0056] In this embodiment, firstly, an unsupervised sequence segmentation algorithm and an initial convolutional neural network classifier can be combined to analyze and determine an optimal early fault occurrence time for each set of grayscale images of a steam turbine generator. Here, each grayscale image in the set can be arranged in the order of the sampling time of the corresponding continuous vibration signal segment.

[0057] Next, for each set of grayscale images for which the optimal early fault occurrence time has been determined, the entire grayscale image set can be divided according to this optimal early fault occurrence time and the actual sampling time of the continuous vibration signal segment corresponding to each grayscale image in the set. All grayscale images with sampling times earlier than the optimal early fault occurrence time are assigned to the normal subset, while all grayscale images with sampling times later than the optimal early fault occurrence time are assigned to the fault subset. Subsequently, pseudo-labels are automatically generated for these grayscale images; that is, each grayscale image in the normal subset is labeled with a pseudo-label indicating a normal state, and each grayscale image in the fault subset is labeled with a pseudo-label indicating a fault state.

[0058] Then, all grayscale images with pseudo-labels (including all grayscale images labeled as normal and fault states) in the entire grayscale image set are merged to form a large-scale, labeled training dataset. This training dataset is then used to perform supervised training on the initial convolutional neural network classifier. The initial convolutional neural network classifier learns and distinguishes the differences in visual features between grayscale images of normal and fault states, adjusting its internal parameters to ultimately become a target convolutional neural network classifier capable of effectively identifying early fault states.

[0059] This application's embodiments bypass the strong reliance on precise manual annotation of data through an unsupervised approach. It achieves the automatic construction of a large-scale labeled dataset and the training of a powerful target convolutional neural network classifier even with only historical full-lifecycle vibration data of a steam turbine generator and no precise fault labels. This not only significantly saves on the costs of professional knowledge and manual annotation, making the method more applicable to real-world industrial scenarios, but also, by analyzing the macroscopic changes across the entire grayscale image set to pinpoint the optimal early fault occurrence time, it may capture more reliable early fault signs than analysis of a single grayscale image, thereby improving the accuracy and generalization ability of the target convolutional neural network classifier.

[0060] Step 210: The vibration signal of the turbine generator to be identified is acquired in real time, and the vibration signal segment to be identified is dynamically constructed based on the acquired vibration signal. According to the vibration signal segment to be identified, a grayscale image to be identified is constructed. The grayscale image to be identified is input into the target convolutional neural network classifier to determine whether the turbine generator to be identified has a fault.

[0061] In this embodiment of the application, step 207 includes:

[0062] Step 207-1: For each set of grayscale images, initialize multiple candidate early fault occurrence times.

[0063] Step 207-2: For each candidate early fault occurrence time, based on the early fault occurrence time and the sampling time corresponding to each grayscale image in the grayscale image set, set pseudo-labels for each grayscale image in the grayscale image set, and train the initial convolutional neural network classifier based on the grayscale images with pseudo-labels to obtain the convolutional neural network classifier corresponding to the early fault occurrence time.

[0064] Step 207-3: Based on the objective function, calculate the objective function value of the convolutional neural network classifier corresponding to each early fault occurrence time.

[0065] Step 207-4: Through the optimization algorithm, the candidate early fault occurrence times are iteratively updated, and the objective function value of the convolutional neural network classifier corresponding to each early fault occurrence time is repeatedly calculated until the iteration termination condition is met. The early fault occurrence time with the smallest objective function value is taken as the optimal early fault occurrence time.

[0066] In this embodiment, firstly, for each set of grayscale images of a steam turbine generator, multiple candidate early fault occurrence times can be generated. These candidate early fault occurrence times can be randomly initialized or obtained by uniform sampling based on the total operating time of the steam turbine generator, thus enabling the exploration of situations where faults may occur at different time points.

[0067] Next, for each candidate early fault occurrence time, a temporary pseudo-label assignment and classifier training process can be performed. Specifically, based on this early fault occurrence time, all grayscale images in the grayscale image set sampled earlier than this early fault occurrence time can be labeled as normal, and all grayscale images sampled later than this early fault occurrence time can be labeled as faulty. Then, using this dataset with temporary pseudo-labels, the initial convolutional neural network classifier is trained independently, resulting in a trained convolutional neural network classifier specific to this candidate early fault occurrence time.

[0068] Then, the performance of the convolutional neural network classifier for each candidate early fault occurrence time obtained in the above steps can be evaluated according to the objective function. The purpose of this objective function is that the closer a candidate early fault occurrence time is to the actual early fault occurrence time, the clearer the grayscale images of the normal state and the fault state should be distinguished in terms of features, so that the trained convolutional neural network classifier performs optimally under a certain evaluation criterion.

[0069] Finally, an optimization algorithm, such as gradient descent or a genetic algorithm, is initiated to generate a new batch of candidate early fault occurrence times that are more likely to approximate the true early fault occurrence time, based on the results of all current candidate early fault occurrence times and their corresponding objective function values. Then, the previous steps are repeated to generate pseudo-labels for these new candidate early fault occurrence times, train a classifier, and calculate the objective function value. This iterative process can continue until a preset termination condition is met, such as reaching the maximum number of iterations or the improvement in the objective function value becoming negligible. Ultimately, among all the evaluated candidate early fault occurrence times, the early fault occurrence time that minimizes the objective function value is determined as the optimal early fault occurrence time for the set of grayscale images of the turbine generator.

[0070] This application integrates the tasks of finding the optimal early failure occurrence time and training the classifier into a unified, automated optimization framework. It does not rely on any prior, manually labeled fault tags, but rather infers the most reasonable early failure occurrence time through the generalization performance of the classifier. This data-driven approach significantly reduces reliance on domain expert knowledge, making it possible to achieve fully automated early failure diagnosis classifier training, and ensuring the statistical reasonableness of the found early failure occurrence time through a rigorous optimization process.

[0071] Optionally, in this embodiment, step 207-2, "training an initial convolutional neural network classifier based on grayscale images with pseudo-labels to obtain a convolutional neural network classifier corresponding to the early fault occurrence time," includes: inputting each grayscale image in the grayscale image set into the initial convolutional neural network classifier to obtain a predicted label corresponding to each grayscale image; calculating the model loss value based on the pseudo-labels and predicted labels of each grayscale image; adjusting the model parameters of the initial convolutional neural network classifier using a backpropagation algorithm based on the model loss value; and calculating the model loss value again based on the parameter-adjusted initial convolutional neural network classifier and the grayscale image set, until a preset training stopping condition is reached, thereby obtaining the trained convolutional neural network classifier corresponding to the early fault occurrence time.

[0072] In this embodiment, firstly, all grayscale images in the current grayscale image set are sequentially input into the initial convolutional neural network classifier. For each input grayscale image, the initial convolutional neural network classifier performs forward propagation calculations, ultimately producing a prediction result at the output layer. Specifically, this prediction result can be a probability distribution representing whether the image belongs to a normal or faulty state. This prediction result is the predicted label corresponding to the grayscale image, reflecting the initial convolutional neural network classifier's judgment of the grayscale image content.

[0073] Next, a pre-created loss function, such as the cross-entropy loss function, can be used to quantify the prediction error of the initial convolutional neural network (CNN) classifier. This loss function compares the predicted labels generated by the initial CNN classifier for each grayscale image with the pseudo-labels previously assigned based on the candidate early failure occurrence time. By calculating the difference between these two labels, the loss function can summarize and output a numerical value representing the overall prediction error of the current initial CNN classifier, i.e., the model loss value. The larger this value, the greater the deviation between the initial CNN classifier's predictions and the pseudo-label settings.

[0074] Furthermore, the backpropagation algorithm can be initiated to calculate the gradient of the model loss value relative to each model parameter of the convolutional neural network classifier. These gradients indicate the direction and magnitude in which each parameter should be adjusted to reduce the loss. Subsequently, optimization algorithms such as gradient descent are used to fine-tune the parameters in the initial convolutional neural network classifier based on the calculated gradients. This cyclical process of calculating loss-backpropagation-adjusting parameters can continue, and the initial convolutional neural network classifier can be iterated multiple times using the same set of grayscale images with pseudo-labels until a preset training stopping condition is met, thus obtaining the trained convolutional neural network classifier corresponding to the time of the early failure.

[0075] Optionally, in this embodiment, step 207-4, "iteratively updating the candidate early fault occurrence times using an optimization algorithm," includes: dividing the initialized multiple candidate early fault occurrence times into a first set and a second set based on the objective function value corresponding to each early fault occurrence time, wherein the objective function value corresponding to each early fault occurrence time in the first set is less than the objective function value corresponding to each early fault occurrence time in the second set, and the ratio between the number of times contained in the first set and the number of times contained in the second set is a preset ratio; for each early fault occurrence time in the first set, updating it based on the early fault occurrence time, the current iteration number, and a random factor to obtain an updated early fault occurrence time; for each early fault occurrence time in the second set, updating it based on the early fault occurrence time and the early fault occurrence time with the smallest objective function value in the first set to obtain an updated early fault occurrence time; generating a random number and determining the relationship between the random number and a preset safety threshold; if the random number is greater than the preset safety threshold, randomly perturbing a preset proportion of the updated early fault occurrence times and regenerating the updated early fault occurrence times.

[0076] In this embodiment, firstly, the optimization algorithm sorts and divides the candidate early failure occurrence times based on the objective function values ​​corresponding to all current candidate early failure occurrence times. Specifically, all candidate early failure occurrence times can be divided into two sets: the first set contains early failure occurrence times with smaller objective function values ​​(i.e., better performance), and the second set contains early failure occurrence times with larger objective function values ​​(i.e., relatively poor performance). The criterion for division is to ensure that the ratio of the number of candidate times in the first set to the number of candidate times in the second set is equal to a pre-set ratio.

[0077] Next, for each candidate early failure occurrence time in the first partitioned set, the optimization algorithm can update it. Specifically, it can comprehensively consider the candidate early failure occurrence time itself, the current iteration number, and a random factor. The iteration number is introduced to gradually reduce the adjustment step size as the search progresses, achieving a fine search; while the random factor ensures the diversity of update directions, thereby conducting an exploratory local search in the surrounding area to avoid prematurely getting trapped in local optima.

[0078] For each candidate early failure time in the second set, the optimization algorithm can employ a different update strategy. These poorly performing candidate early failure times can be moved closer to and updated towards the candidate early failure time in the first set that minimizes the objective function value, i.e., the current globally optimal candidate early failure time. This strategy simulates a learning behavior, aiming to guide search resources to the most promising region and accelerate the overall convergence speed.

[0079] Finally, to further enhance the optimization algorithm's ability to escape local optima, it can generate a random number and compare it with a preset safety threshold. If the random number is greater than the threshold, a preset percentage (e.g., 10%) of the newly updated candidate early failure times is randomly perturbed. This perturbation introduces a significant random change to these candidate early failure times, essentially creating a mutation in the search space, thus offering an opportunity to explore entirely new and potentially better regions.

[0080] This application's embodiments balance the relationship between exploration and exploitation in the optimization algorithm by combining multiple mechanisms such as learning from the best individual and random perturbation. This allows for rapid convergence to a promising search region while maintaining sufficient diversity to avoid getting trapped in local optima, thereby greatly increasing the likelihood of finding the global optimum or near-optimal early failure time.

[0081] In this embodiment of the application, step 207-3 optionally includes: for each early fault occurrence time, using the corresponding convolutional neural network classifier to predict the verification grayscale images in the verification grayscale image set to obtain the predicted label of each verification grayscale image; based on the predicted label of each verification grayscale image and the pseudo-label generated for each verification grayscale image based on the early fault occurrence time, calculating the classification error rate through the first sub-function of the objective function, wherein the verification grayscale image is reconstructed based on the vibration signal sequence samples used when training the convolutional neural network classifier; determining a first number of grayscale images in the grayscale image set, and determining a second number of grayscale images with pseudo-labels of normal state according to the early fault occurrence time; calculating the ratio of the first number to the second number through the second sub-function of the objective function to obtain a time penalty value; and weightedly fusing the classification error rate and the time penalty value through the fusion sub-function of the objective function to obtain the objective function value of the convolutional neural network classifier corresponding to the early fault occurrence time.

[0082] In this embodiment, firstly, for each candidate early fault occurrence time and its corresponding trained convolutional neural network (CNN) classifier, a set of validation grayscale images can be prepared to test the CNN classifier. This set of validation grayscale images can be constructed from additional vibration signal sequence samples that were not used in the training of the CNN classifier. The CNN classifier can predict the label for each validation grayscale image in this set. Subsequently, the first sub-function in the objective function is called to calculate a classification error rate by comparing the predicted labels of the validation grayscale images with the pseudo-labels generated based on the current candidate early fault occurrence time. This metric directly reflects the generalization ability and classification accuracy of the CNN classifier trained under the assumption of the candidate early fault occurrence time.

[0083] In one specific embodiment, the verification grayscale image set can be constructed as follows: For the grayscale image set used in training the convolutional neural network classifier, vibration signal sequence samples on which the grayscale image set is based are obtained. Then, multiple continuous vibration signal segments are extracted again from the vibration signal sequence samples. The sampling time corresponding to this extraction is not exactly the same as or completely different from the sampling time corresponding to the previous extraction, so as to obtain continuous vibration signal segments that were not learned during the training of the convolutional neural network classifier. Afterwards, multiple verification grayscale images are constructed based on the multiple newly extracted continuous vibration signal segments, and these verification grayscale images are combined into a verification grayscale image set. It should be noted that when performing classifier performance verification, the pseudo-labels of the verification grayscale images are labeled with the same candidate early fault occurrence time.

[0084] Next, we introduce an evaluation metric that is independent of the performance of the convolutional neural network classifier but related to the timing of candidate early faults. Specifically, we can count the total number of grayscale images in the grayscale image set used to train the convolutional neural network classifier (i.e., the first quantity) and calculate the number of grayscale images marked as normal based on the current timing of candidate early faults (i.e., the second quantity). Then, we calculate the ratio of these two quantities using the second sub-function of the objective function, which serves as the time penalty value.

[0085] Finally, the two metrics mentioned above—classification error rate reflecting classifier performance and time penalty value reflecting the reasonableness of the timing—are weighted and combined through the fusion sub-function of the objective function. The weight of each metric determines whether more emphasis is placed on the classification accuracy of the classifier or on the reasonableness of the early failure occurrence time. Through this calculation, a final objective function value representing the overall quality of the trained convolutional neural network classifier (i.e., the candidate early failure occurrence time) is obtained.

[0086] In a specific embodiment, the objective function can be expressed as:

[0087]

[0088] in, This indicates the time when an early fault occurred (which can be represented by the number of the grayscale image, that is, the pseudo-label of the h-th grayscale image is the fault state). It is the first quantity. , This is the adjustment coefficient. Indicates the first i Predicted labels for a valid grayscale image. Indicates the first i A pseudo-label for a grayscale image. If... =0, otherwise =1. It should be noted that the number of grayscale images to be verified can be the first number or not; this is not limited here. In the formula above, the number of grayscale images to be verified is the first number N.

[0089] It's important to note that for each candidate early fault occurrence time within each grayscale image set, a fixed initial convolutional neural network (CNN) classifier is used. For each candidate early fault occurrence time *h*, this initial CNN classifier is trained using pseudo-labels generated from *h*. However, for different *h* values, the same fixed training strategy (e.g., a fixed number of iterations) is used, and early stopping is employed to prevent overfitting. Thus, when *h* is inconsistent with the true early fault occurrence time, the initial CNN classifier will struggle to learn a good decision boundary due to label noise (i.e., incorrect pseudo-labels), resulting in a higher subsequent classification error rate. Conversely, when *h* is consistent with the true early fault occurrence time, the initial CNN classifier can quickly learn the correct pattern, leading to a lower classification error rate. Therefore, by calculating the classification error rate, we can effectively identify the better-performing trained CNN classifier, i.e., find candidate early fault occurrence times that are closer to the true early fault occurrence time.

[0090] In this embodiment, the training phase involves: extracting continuous vibration signal segments from the vibration signal sequence sample → generating a grayscale image → labeling pseudo-labels with h → training an initial convolutional neural network classifier; the verification phase involves: extracting different continuous vibration signal segments from the vibration signal sequence sample again → generating a verification grayscale image → labeling pseudo-labels with the same h → evaluating the trained convolutional neural network classifier.

[0091] The training and validation sets are derived from the same vibration signal sequence samples but different continuous vibration signal segments. This ensures that the trained convolutional neural network classifier is evaluated on unseen data, preventing overfitting to specific training samples. Using the same h to generate training and validation pseudo-labels ensures consistent evaluation. The evaluation assesses whether the convolutional neural network classifier can learn meaningful patterns for this hypothetical h. If h is close to the actual early failure time, the pattern should have generalization ability.

[0092] Furthermore, as Figure 1 In terms of specific implementation, this application provides a vibration-based early fault identification device for steam turbine generators, such as... Figure 3 As shown, the device includes:

[0093] The sample acquisition module is used to acquire training samples, wherein the training samples include vibration signal sequence samples of multiple steam turbine generators throughout their historical life cycle.

[0094] The image construction module is used to extract multiple continuous vibration signal segments from each vibration signal sequence sample, construct grayscale images based on each continuous vibration signal segment, and generate a set of grayscale images of the vibration signal sequence sample based on each grayscale image.

[0095] The classifier training module is used to train an initial convolutional neural network classifier using an unsupervised sequence segmentation method on each set of grayscale images to obtain a target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each set of grayscale images and uses the generated pseudo-labels to train the initial convolutional neural network classifier.

[0096] The fault identification module is used to acquire the vibration signal of the turbine generator to be identified in real time, dynamically construct the vibration signal segment to be identified based on the acquired vibration signal, construct the grayscale image to be identified based on the vibration signal segment to be identified, and input the grayscale image to be identified into the target convolutional neural network classifier to determine whether the turbine generator to be identified has a fault.

[0097] Optionally, the classifier training module is used for:

[0098] Based on the unsupervised sequence segmentation method and the initial convolutional neural network classifier, the optimal early fault occurrence time corresponding to each grayscale image set is determined;

[0099] For each set of grayscale images, the grayscale images in the set are divided into a normal subset and a fault subset based on the optimal early fault occurrence time corresponding to the set of grayscale images and the sampling time of the continuous vibration signal segment corresponding to each grayscale image in the set of grayscale images. Each grayscale image in the normal subset is assigned a pseudo-label indicating a normal state, and each grayscale image in the fault subset is assigned a pseudo-label indicating a fault state.

[0100] The initial convolutional neural network classifier is trained based on the grayscale images with pseudo-labels in each grayscale image set to obtain the target convolutional neural network classifier.

[0101] Optionally, the classifier training module is further configured to:

[0102] For each set of grayscale images, initialize multiple candidate early fault occurrence times;

[0103] For each candidate early fault occurrence time, based on the early fault occurrence time and the sampling time corresponding to each grayscale image in the grayscale image set, a pseudo-label is set for each grayscale image in the grayscale image set, and based on the grayscale images with pseudo-labels, the initial convolutional neural network classifier is trained to obtain the convolutional neural network classifier corresponding to the early fault occurrence time.

[0104] Based on the objective function, the objective function value of the convolutional neural network classifier corresponding to each early fault occurrence time is calculated.

[0105] The algorithm iteratively updates the candidate early failure occurrence times and repeatedly calculates the objective function value of the convolutional neural network classifier corresponding to each early failure occurrence time until the iteration termination condition is met. The early failure occurrence time with the smallest objective function value is taken as the optimal early failure occurrence time.

[0106] Optionally, the classifier training module is further configured to:

[0107] Each grayscale image in the grayscale image set is input into the initial convolutional neural network classifier to obtain the predicted label corresponding to each grayscale image;

[0108] The model loss value is calculated based on the pseudo-labels and predicted labels of each grayscale image;

[0109] Based on the model loss value, the model parameters of the initial convolutional neural network classifier are adjusted using the backpropagation algorithm. Based on the initial convolutional neural network classifier with adjusted parameters and the grayscale image set, the model loss value is calculated again until the preset training stopping condition is reached, thus obtaining the convolutional neural network classifier that has been trained and is located at the time of the early fault occurrence.

[0110] Optionally, the classifier training module is further configured to:

[0111] Based on the objective function value corresponding to each early fault occurrence time, the initial multiple candidate early fault occurrence times are divided into a first set and a second set. The objective function value corresponding to each early fault occurrence time in the first set is less than the objective function value corresponding to each early fault occurrence time in the second set, and the ratio between the number of times contained in the first set and the number of times contained in the second set is a preset ratio.

[0112] For each early fault occurrence time in the first set, the early fault occurrence time is updated based on the early fault occurrence time, the current iteration number, and the random factor to obtain the updated early fault occurrence time.

[0113] For each early fault occurrence time in the second set, the early fault occurrence time is updated based on the early fault occurrence time and the early fault occurrence time with the smallest objective function value in the first set.

[0114] A random number is generated, and the relationship between the random number and a preset security threshold is determined. If the random number is greater than the preset security threshold, the updated early fault occurrence time of the preset ratio is randomly perturbed, and the updated early fault occurrence time is regenerated.

[0115] Optionally, the classifier training module is further configured to:

[0116] For each early fault occurrence time, the corresponding convolutional neural network classifier is used to predict the verification grayscale images in the verification grayscale image set to obtain the predicted label of each verification grayscale image. Based on the predicted label of each verification grayscale image and the pseudo label generated for each verification grayscale image based on the early fault occurrence time, the classification error rate is calculated through the first sub-function of the objective function. The verification grayscale image is reconstructed based on the vibration signal sequence samples used when training the convolutional neural network classifier.

[0117] A first number of grayscale images in the grayscale image set is determined, and a second number of grayscale images with pseudo-labels as normal is determined based on the early fault occurrence time. The ratio of the first number to the second number is calculated through the second sub-function of the objective function to obtain the time penalty value.

[0118] The classification error rate and the time penalty value are weighted and fused by the fusion sub-function of the objective function to obtain the objective function value of the convolutional neural network classifier corresponding to the early failure occurrence time.

[0119] Optionally, the image construction module is used for:

[0120] For each continuous vibration signal segment, a fast Fourier transform is performed on the continuous vibration signal segment to obtain the corresponding frequency domain signal segment, and the amplitude corresponding to each frequency component is extracted from the frequency domain signal segment to obtain a frequency domain feature sequence. The frequency domain feature sequence is then normalized to obtain a standardized frequency domain feature sequence.

[0121] For each feature value in the standardized frequency domain feature sequence, the corresponding gray value is calculated based on the feature value according to a preset rule, and the target position of the gray value in the grayscale image is determined based on the position of the feature value in the standardized frequency domain feature sequence, and the gray value is mapped to the target position.

[0122] When the gray values ​​corresponding to each feature value in the standardized frequency domain feature sequence are mapped to a grayscale image, the grayscale image corresponding to the continuous vibration signal segment is obtained.

[0123] It should be noted that other corresponding descriptions of the functional units involved in the vibration-based early fault identification device for steam turbine generators provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0124] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0125] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0127] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0128] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0129] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying early faults in a turbogenerator based on vibrations, characterized in that, The method comprises the following steps: acquiring training samples, wherein the training samples comprise historical full-life vibration signal sequence samples of a plurality of steam turbine generators; for each vibration signal sequence sample, a plurality of continuous vibration signal segments are intercepted from the vibration signal sequence sample, a grayscale image is constructed according to each continuous vibration signal segment, and a grayscale image set of the vibration signal sequence sample is generated according to the grayscale images; using the grayscale image sets, an initial convolutional neural network classifier is trained by using an unsupervised sequence segmentation method to obtain a target convolutional neural network classifier, wherein the unsupervised sequence segmentation method generates pseudo labels by automatically determining an optimal early fault occurrence time for each grayscale image set, and the initial convolutional neural network classifier is trained by using the generated pseudo labels; real-time acquisition of vibration signals of a steam turbine generator to be identified, dynamic construction of a to-be-identified vibration signal segment based on the acquired vibration signals, construction of a to-be-identified grayscale image according to the to-be-identified vibration signal segment, input of the to-be-identified grayscale image into the target convolutional neural network classifier, and determination of whether the steam turbine generator to be identified has a fault; the training of the initial convolutional neural network classifier by using the grayscale image sets and the unsupervised sequence segmentation method to obtain the target convolutional neural network classifier comprises: for each grayscale image set, a plurality of candidate early fault occurrence times are initialized; for each candidate early fault occurrence time, pseudo labels are set for the grayscale images in the grayscale image set according to the early fault occurrence time and the sampling times of the grayscale images corresponding to the grayscale image set, and the initial convolutional neural network classifier is trained based on the grayscale images with the pseudo labels to obtain a convolutional neural network classifier corresponding to the early fault occurrence time; the target function values of the convolutional neural network classifiers corresponding to each early fault occurrence time are calculated based on a target function; the candidate early fault occurrence times are iteratively updated by using an optimization algorithm, and the target function values of the convolutional neural network classifiers corresponding to each early fault occurrence time are repeatedly calculated until an iteration termination condition is met, and the early fault occurrence time with the minimum target function value is taken as the optimal early fault occurrence time; for each grayscale image set, the grayscale images in the grayscale image set are divided according to the optimal early fault occurrence time corresponding to the grayscale image set and the sampling times of the continuous vibration signal segments corresponding to the grayscale images in the grayscale image set to obtain a normal subset and a fault subset, and each grayscale image in the normal subset is set with a pseudo label indicating a normal state, and each grayscale image in the fault subset is set with a pseudo label indicating a fault state; the initial convolutional neural network classifier is trained according to the grayscale images with the pseudo labels in the grayscale image sets to obtain the target convolutional neural network classifier.

2. The method of claim 1, wherein, the training of the initial convolutional neural network classifier based on the grayscale images with the pseudo labels to obtain the convolutional neural network classifier corresponding to the early fault occurrence time comprises: input each gray image in the set of gray images into an initial convolutional neural network classifier respectively to obtain a predicted label corresponding to each gray image; calculate a model loss value based on the pseudo label and the predicted label of each gray image; adjust the model parameters of the initial convolutional neural network classifier through a back propagation algorithm based on the model loss value, and calculate the model loss value again based on the initial convolutional neural network classifier after the parameter adjustment and the set of gray images, until a preset training stop condition is reached, to obtain a trained convolutional neural network classifier corresponding to the early failure occurrence time.

3. The method of claim 1, wherein, The iterative update of the candidate early failure occurrence time through the optimization algorithm comprises: According to the target function value corresponding to each early failure occurrence time, the initialized multiple candidate early failure occurrence times are divided into a first set and a second set, wherein the target function value corresponding to each early failure occurrence time in the first set is less than the target function value corresponding to each early failure occurrence time in the second set, and the ratio between the number of times contained in the first set and the number of times contained in the second set is a preset ratio; For each early failure occurrence time in the first set, the early failure occurrence time, the current iteration number and a random factor are used to update, to obtain an updated early failure occurrence time; For each early failure occurrence time in the second set, the early failure occurrence time and the early failure occurrence time with the minimum target function value in the first set are used to update, to obtain an updated early failure occurrence time; A random number is generated, and the relationship between the random number and a preset safety threshold is determined, if the random number is greater than the preset safety threshold, a preset proportion of the updated early failure occurrence times are randomly disturbed, and the updated early failure occurrence times are regenerated.

4. The method according to claim 2 or 3, characterized in that, The target function is used to calculate the target function value of the convolutional neural network classifier corresponding to each early failure occurrence time respectively, comprising: For each early failure occurrence time, the corresponding convolutional neural network classifier is used to predict the verification gray image in the verification gray image set to obtain the predicted label of each verification gray image, and the pseudo label generated for each verification gray image based on the early failure occurrence time is used to calculate the classification error rate through the first sub-function of the target function, wherein the verification gray image is reconstructed based on the vibration signal sequence sample used when training the convolutional neural network classifier; Determine the first number of gray images in the set of gray images, and determine the second number of gray images with pseudo labels of normal state according to the early failure occurrence time, and calculate the ratio of the first number to the second number through the second sub-function of the target function to obtain a time penalty value; The classification error rate and the time penalty value are weighted and fused through the fusion sub-function of the target function to obtain the target function value of the convolutional neural network classifier corresponding to the early failure occurrence time.

5. The method of claim 1, wherein, The constructing a gray-scale image according to each continuous vibration signal segment respectively comprises: For each continuous vibration signal segment, a fast Fourier transform is performed on the continuous vibration signal segment to obtain a corresponding frequency domain signal segment, and an amplitude corresponding to each frequency component is extracted from the frequency domain signal segment to obtain a frequency domain feature sequence; and the frequency domain feature sequence is normalized to obtain a normalized frequency domain feature sequence; For each feature value in the normalized frequency domain feature sequence, a corresponding gray value is calculated based on the feature value through a preset rule, and a target position of the gray value in the gray-scale image is determined based on a position of the feature value in the normalized frequency domain feature sequence, and the gray value is mapped to the target position; When the gray values corresponding to the feature values in the normalized frequency domain feature sequence are all mapped to the gray-scale image, the gray-scale image corresponding to the continuous vibration signal segment is obtained.

6. A vibration-based early fault identification device for a turbogenerator, characterized by The method comprises: a sample acquisition module configured to acquire training samples, wherein the training samples comprise historical full-life cycle vibration signal sequence samples of a plurality of steam turbine generators; an image construction module configured to, for each vibration signal sequence sample, extract a plurality of continuous vibration signal segments from the vibration signal sequence sample, construct a gray-scale image according to each continuous vibration signal segment respectively, and generate a gray-scale image set of the vibration signal sequence sample according to the gray-scale images; a classifier training module configured to train an initial convolutional neural network classifier by using the gray-scale image sets and adopting an unsupervised sequence segmentation method to generate a target convolutional neural network classifier, wherein the unsupervised sequence segmentation method generates pseudo labels by automatically determining an optimal early fault occurrence time for each gray-scale image set, and trains the initial convolutional neural network classifier by using the generated pseudo labels; a fault identification module configured to acquire vibration signals of a steam turbine generator to be identified in real time, dynamically construct a to-be-identified vibration signal segment based on the acquired vibration signals, construct a to-be-identified gray-scale image according to the to-be-identified vibration signal segment, input the to-be-identified gray-scale image into the target convolutional neural network classifier, and determine whether the steam turbine generator to be identified has a fault; The classifier training module is configured to: for each gray-scale image set, initialize a plurality of candidate early fault occurrence times; for each candidate early fault occurrence time, set pseudo labels for the gray-scale images in the gray-scale image set according to the early fault occurrence time and sampling times corresponding to the gray-scale images in the gray-scale image set, train an initial convolutional neural network classifier based on the gray-scale images with the pseudo labels to obtain a convolutional neural network classifier corresponding to the early fault occurrence time; calculate a target function value of the convolutional neural network classifier corresponding to each early fault occurrence time based on a target function; and iteratively update the candidate early fault occurrence times by an optimization algorithm, and repeatedly calculate the target function value of the convolutional neural network classifier corresponding to each early fault occurrence time until an iteration termination condition is met, and take the early fault occurrence time with the minimum target function value as an optimal early fault occurrence time. For each gray image set, according to the optimal early failure time corresponding to the gray image set and the sampling time of each continuous vibration signal segment corresponding to each gray image in the gray image set, the gray images in the gray image set are divided to obtain a normal subset and a failure subset, and each gray image in the normal subset is set with a pseudo-label indicating a normal state, and each gray image in the failure subset is set with a pseudo-label indicating a failure state; According to the gray images in each gray image set set with the pseudo-labels, the initial convolutional neural network classifier is trained to obtain a target convolutional neural network classifier.

7. A storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements the method of any one of claims 1 to 5.

8. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, The processor, when executing the computer program, implements the method of any one of claims 1 to 5.

Images