Self-adaptive water turbine bearing temperature prediction method, system and equipment based on multi-layer LSTM (Long Short Term Memory), and medium

Abstract

The invention discloses a self-adaptive water turbine bearing temperature prediction method, system and device based on multilayer LSTM and a medium, and belongs to the technical field of water turbine bearing temperature prediction.The method comprises the steps that historical temperature data of a water turbine bearing are obtained, and abnormal values in the temperature data are recognized and replaced through a Hampel filtering method; constructing a time sequence sample based on the processed data; the time sequence sample is input into a multi-layer LSTM network model, the multi-layer LSTM network model comprises a plurality of LSTM layers, a Dropout layer and a full connection layer, and optimization training is carried out on related parameters; and predicting the target time sequence data based on the trained multilayer LSTM network model, and outputting a future temperature value of the water turbine bearing. According to the method, processing windows and threshold parameters are dynamically adjusted, effective fluctuation characteristics of data are reserved, data quality is improved, dynamic adjustment and efficient parallel calculation of the multi-layer LSTM structure are achieved, and deployment is simple and rapid.

Description

Technical Field

[0001] The present invention relates to the technical field of turbine bearing temperature prediction, and in particular to a multi-layer LSTM-based adaptive turbine bearing temperature prediction method, system, equipment and medium. Background Art

[0002] The operation of hydro turbines typically involves complex mechanical and electrical components, whose performance is affected by a variety of factors, including water quality, temperature fluctuations, wear, and corrosion. Due to the widespread distribution of equipment and the harsh operating environment, regular manual inspections are time-consuming and labor-intensive, and it is difficult to cover all critical points. Furthermore, sudden equipment failures can lead to unplanned downtime, resulting in significant economic losses and negatively impacting grid stability. To overcome these challenges, the introduction of AI-based predictive maintenance systems is crucial. Long Short-Term Memory (LSTM) networks are a key component.

[0003] Traditional temperature prediction methods primarily include those based on statistical regression and simple neural network prediction. These methods struggle to effectively capture the temporal characteristics of turbine operating data. Furthermore, their fixed network structures lack adaptability, making them inaccurate in capturing long-term trends and slow to respond to sudden changes. Furthermore, parameter adjustment in traditional temperature prediction models relies primarily on manual trial-and-error, which not only requires significant time and computing resources but also places high demands on researchers' experience in parameter adjustment. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to solve the problem that traditional methods use simple threshold filtering or moving average techniques to process outliers, which easily delete valid data by mistake or retain noise data, and the fixed processing parameters cannot adapt to the data characteristics under different working conditions. The existing technology mostly uses a single neural network structure or a traditional statistical model. The expression ability of this type of fixed structure model is limited. It is difficult to accurately capture the complex nonlinear characteristics of turbine temperature changes, and it is easy to overfit, resulting in unstable prediction accuracy. Traditional prediction models often require experienced researchers to manually adjust the parameters, which requires extremely high experience from the researchers. As the data is updated, the model often needs to be readjusted, but the on-site personnel of most projects do not have the ability to adjust the model. The existing technology lacks an effective model management mechanism and computing resource scheduling strategy, resulting in low system operation efficiency and difficulty in meeting the needs of real-time prediction.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: an adaptive turbine bearing temperature prediction method based on a multi-layer LSTM, which comprises the following steps: obtaining historical temperature data of turbine bearings, identifying and replacing abnormal values ​​in the temperature data using a Hampel filtering method; constructing time series samples based on the processed data; inputting the time series samples into a multi-layer LSTM network model, the multi-layer LSTM network model comprising multiple LSTM layers, a Dropout layer and a fully connected layer, and optimizing and training relevant parameters; predicting target time series data based on the trained multi-layer LSTM network model, and outputting future temperature values ​​of the turbine bearings.

[0007] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the construction of time series samples includes generating continuous time window data based on the processed temperature data and extracting time series features.

[0008] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the optimization training of relevant parameters includes performing parameter search on the structural parameters of the multi-layer LSTM network model to obtain a parameter combination that meets the preset evaluation indicators.

[0009] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the multi-layer LSTM network model includes LSTM layers, Dropout layers and fully connected layers, the number and neuron scale of which are adjustable and serve as the object of parameter optimization.

[0010] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the method uses the Hampel filtering method to identify and replace outliers in the temperature data, including constructing a sliding subsequence for the original temperature data with a preset window length, calculating the median and median absolute deviation of each subsequence respectively, identifying data points that deviate from the median by more than a set multiple threshold as outliers, and replacing them with the median of the corresponding subsequence.

[0011] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the number and neuron scale of the LSTM layers, Dropout layers and fully connected layers are adjustable and serve as objects of parameter optimization, including setting the value range of the number of LSTM layers, the number of fully connected layers, the number of Dropout layers and the number of neurons in each layer, using a grid search strategy to traverse each parameter combination, and using the mean absolute percentage error as an evaluation indicator to select the model structure corresponding to the optimal parameter combination.

[0012] As a preferred solution of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM described in the present invention, the method predicts the target time series data based on the trained multi-layer LSTM network model, including using the trained multi-layer LSTM network model to receive input temperature time series data of a fixed length, and outputting the predicted turbine bearing temperature value based on the state transition and gating structure of the model, where the length of the input sequence is set by the Sequence_length parameter of the model.

[0013] Another object of the present invention is to provide an adaptive turbine bearing temperature prediction system based on multi-layer LSTM.

[0014] To solve the above technical problems, the present invention provides the following technical solutions: an adaptive turbine bearing temperature prediction system based on a multi-layer LSTM, comprising: a data acquisition and processing module, a sample construction module, a model construction optimization module and a prediction module; the data acquisition and processing module is used to obtain historical temperature data of turbine bearings, and use the Hampel filtering method to identify and replace abnormal values ​​in the temperature data; the sample construction module is used to construct time series samples based on the processed data; the model construction optimization module is used to input the time series samples into a multi-layer LSTM network model, the multi-layer LSTM network model includes multiple LSTM layers, Dropout layers and fully connected layers, and optimizes and trains relevant parameters; the prediction module is used to predict the target time series data based on the trained multi-layer LSTM network model, and output the future temperature value of the turbine bearing.

[0015] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the method for adaptive turbine bearing temperature prediction based on multi-layer LSTM are implemented.

[0016] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM are implemented.

[0017] The beneficial effects of the present invention are as follows: the present invention adopts Hampel filtering to realize adaptive outlier detection, can dynamically adjust the processing window and threshold parameters, retain the effective fluctuation characteristics of the data, improve data quality, dynamically adjust the multi-layer LSTM structure, adaptively optimize according to data characteristics, automate parameter tuning mechanism, and improve prediction efficiency through model caching mechanism, achieve efficient parallel computing, simple and fast deployment, and is friendly to inexperienced personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 An overall flow chart of an adaptive turbine bearing temperature prediction method based on multi-layer LSTM is provided for one embodiment of the present invention.

[0020] Figure 2 A before-and-after comparison diagram of the data processing effect of an adaptive turbine bearing temperature prediction method based on a multi-layer LSTM provided in one embodiment of the present invention.

[0021] Figure 3 A grid search strategy diagram is provided for the number of neurons in an adaptive turbine bearing temperature prediction method based on a multi-layer LSTM according to an embodiment of the present invention.

[0022] Figure 4 A fitting effect diagram of an adaptive turbine bearing temperature prediction method based on multi-layer LSTM provided in one embodiment of the present invention.

[0023] Figure 5 An API interface structure framework diagram of an adaptive turbine bearing temperature prediction method based on multi-layer LSTM is provided in one embodiment of the present invention.

[0024] Figure 6 A system solution module diagram of a multi-layer LSTM-based adaptive turbine bearing temperature prediction system provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0026] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides an adaptive turbine bearing temperature prediction method based on a multi-layer LSTM, comprising:

[0027] S1. Obtain historical temperature data of turbine bearings and use the Hampel filtering method to identify and replace abnormal values ​​in the temperature data.

[0028] S2. Construct a time series sample based on the processed data.

[0029] S3. Input the time series samples into the multi-layer LSTM network model, which includes multiple LSTM layers, Dropout layers, and fully connected layers, and optimize the relevant parameters for training.

[0030] S4. Predict the target time series data based on the trained multi-layer LSTM network model and output the future temperature value of the turbine bearing.

[0031] The present invention proposes an adaptive turbine bearing temperature prediction system based on a multi-layer LSTM network and its implementation method. The system innovatively combines Hampel filtering, dynamic sliding windows, and multi-layer LSTM deep learning technology to construct an end-to-end intelligent prediction framework. The system first processes outliers in the raw temperature data using Hampel filtering to ensure data quality. It then uses dynamic sliding window technology to construct multi-scale time series features to effectively capture the temporal dependencies of temperature changes. Predictive modeling is then performed using an innovative multi-layer LSTM network architecture that can adaptively adjust the number of network layers, the number of neurons, and learning parameters based on data characteristics. Finally, the model's data processing, training, and inference functions are encapsulated as an API interface that can provide external Web services and deployed on a Linux server using Gunicorn and Flask as WSGI (Web Server Gateway Interface) to quickly handle high-concurrency requests.

[0032] Example 2, reference Figures 1 to 5 , which is the second embodiment of the present invention, provides an adaptive turbine bearing temperature prediction method based on a multi-layer LSTM based on the above embodiment.

[0033] In the embodiment of the present application, in step S1, a sliding subsequence is constructed for the original temperature data with a preset window length, the median and median absolute deviation of each subsequence are calculated respectively, and the data points that deviate from the median by more than a set multiple threshold are identified as outliers and replaced with the median of the corresponding subsequence.

[0034] Specifically, when building machine learning and deep learning models, dataset quality is a key factor in determining model performance and predictive accuracy. A high-quality dataset should be complete, consistent, accurate, and noise-free. However, in real-world applications, data often contains a variety of noise and outliers. These anomalies may be caused by measurement errors, sensor failures, data entry errors, or other unforeseen factors. Outliers and noise in datasets not only increase the difficulty of model training but also mislead the model's learning direction, preventing it from correctly capturing the true patterns of the data. For example, if outliers are present in turbine bearing temperature data, the model may mistakenly interpret them as part of normal behavior, resulting in biased predictions of future trends. If these outliers are not addressed, they can severely impact the model's learning process, leading to reduced generalization ability and even overfitting. Therefore, ensuring dataset quality—that is, removing or correcting outliers and noise—is a prerequisite for building reliable predictive models.

[0035] To address the problem of outliers in datasets, the Hampel filter is widely used as an effective anomaly detection technique. The Hampel filter uses the median and median absolute deviation (MAD) to identify and replace outliers in a data series. Compared to traditional mean and standard deviation methods, the Hampel filter is more robust and less susceptible to extreme values. By setting an appropriate threshold, the Hampel filter can accurately detect and replace outliers that deviate significantly from the center of the data distribution, thereby improving the overall data quality.

[0036] In the operating data of turbine bearing temperature, applying the Hampel filter to preprocess the data can significantly improve the accuracy of the prediction model. This is because the filtered data is closer to the actual situation, and the model can better learn the normal operation mode of the equipment, thereby providing more reliable and accurate results when predicting the future operating status of the equipment. The comparison of the data before and after processing is shown in the attached figure. Figure 2 shown.

[0037] In the embodiment of the present application, in step S2, continuous time window data is generated based on the processed temperature data to extract time series features.

[0038] Time series samples are constructed based on the turbine bearing temperature data after outlier processing, and a dynamic sliding window mechanism is adopted.

[0039] This method can adaptively adjust the modeling granularity while ensuring data integrity, improve the model's modeling ability for non-stationary sequences, and improve the generalization and stability of predictions.

[0040] In one alternative implementation, a fixed-length sliding window mechanism is used to construct time series samples. This mechanism sets a static window length, sequentially partitions the data series into equal steps, and generates equal-length samples for model training. This method is simple to implement and suitable for scenarios with relatively stable data. However, it is prone to information truncation or redundancy when faced with periodic or trending changes, limiting the model's ability to extract multi-scale time series patterns.

[0041] In another alternative implementation, time series samples are constructed using an event-driven segmentation approach. For example, when the temperature change rate exceeds a preset threshold or a trend reversal point occurs, a data segment is automatically captured as a sample. This approach is suitable for anomaly detection or alarm triggering models, but is not suitable for building continuous prediction models and may result in uneven sample distribution.

[0042] The implementation method of the present application constructs time series samples through a dynamic sliding window mechanism, which can flexibly adjust the sample structure based on the fluctuation characteristics of actual operating data, thereby effectively extracting the multi-scale temporal dependencies of temperature changes, and providing high-quality, structured input samples for the modeling and prediction of the multi-layer LSTM model, thereby improving the prediction accuracy and enhancing the model's sensitivity to changes in operating status.

[0043] In the embodiment of the present application, in step S3, a parameter search is performed on the structural parameters of the multi-layer LSTM network model to obtain a parameter combination that meets the preset evaluation criteria. The number and neuron size of the LSTM layers, Dropout layers, and fully connected layers of the multi-layer LSTM network model are adjustable and serve as the object of parameter optimization. The range of values ​​for the number of LSTM layers, the number of fully connected layers, the number of Dropout layers, and the number of neurons in each layer is set, and a grid search strategy is used to traverse each parameter combination. The mean absolute percentage error is used as the evaluation metric to select the model structure corresponding to the optimal parameter combination.

[0044] Specifically, after obtaining a relatively complete dataset, the next step is to build a suitable network model to predict temperature trends. Among the various prediction models, Long Short-Term Memory networks (LSTMs) stand out due to their excellent ability to process sequential data and capture long-term dependencies, making them an ideal choice for temperature trend prediction.

[0045] The LSTM network is a specialized form of recurrent neural network (RNN), specifically designed to address the vanishing or exploding gradient problem encountered by RNNs when processing long sequences of data. By introducing a "gate" mechanism—an input gate, a forget gate, and an output gate—LSTM effectively controls the flow of information, enabling the model to remember information over long time spans while ignoring irrelevant historical data. This property is particularly important for predicting the operating parameters of power equipment, as the equipment's operating status is often influenced by multiple factors over a period of time, and these factors can operate across different timescales.

[0046] The LSTM network has the ability to filter information through its unique gate mechanism. The forget gate and input gate are used to control the content of the cell state C. Its calculation formula is expressed as:

[0047] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0048] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0049] Among them, f t 、i t are the outputs of the forget gate and input gate respectively, W f 、W i The weight matrices of the forget gate and input gate, respectively, [h t-1 ,x t ] means connecting two vectors into a longer vector, b f 、b i are the bias terms of the forget gate and input gate respectively, and σ is the sigmoid function.

[0050] Through the output gate, the LSTM network controls the cell state C t How many outputs are there to the current output value h t , and its calculation formula is expressed as:

[0051] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0052]

[0053] Among them, t 、 Represent the output of the current moment and the output of the candidate memory unit, b o 、b C Represent the bias of the output gate and the bias of the candidate memory unit, W o 、W C are the weight matrices of the output gate and candidate memory unit respectively.

[0054] Based on the above gating mechanism, LSTM updates the cell state and calculates the cell state at the current moment, which is expressed as:

[0055]

[0056] Among them, C t Represents the cell state at the current moment, C t-1 Represents the cell state at the previous moment.

[0057] For traditional artificial neural networks, hyperparameter tuning typically requires experienced researchers to continuously optimize network hyperparameters through trial and error. However, on-site service personnel often lack the ability to adjust these parameters. Grid optimization strategies can address this shortcoming.

[0058] The grid search strategy is a systematic approach for traversing a predefined hyperparameter space to find the optimal hyperparameter combination. The key advantages of the grid search strategy are: standardization and automation. Grid search provides a standardized approach to model tuning. It does not require users to have advanced machine learning knowledge or intuition to manually adjust parameters. Instead, it automatically performs parameter exploration, which is particularly attractive for non-experts because it lowers the threshold for model tuning. It also ensures model performance. By traversing all specified parameter combinations, grid search can guarantee to find the parameter configuration that performs best on the validation set. It also reduces trial and error costs. For inexperienced workers, randomly trying different parameter combinations can consume a lot of time and computing resources, and the results are unstable. Grid search ensures that each parameter is fairly evaluated, reducing the number of invalid experiments and making resource utilization more efficient.

[0059] The present invention uses a grid search strategy for the number of LSTM layers, Dense layers, Dropout layers, and neurons in the LSTM network, such as Figure 3 shown.

[0060] The setting ranges of various parameters are shown in Table 1.

[0061] Table 1 Optimization strategy range settings for each hyperparameter of the LSTM network

[0062] Number of LSTM layers Number of Dense layers Number of neurons 1-3 1-3 16，32，64，128

[0063] Then, we set the batch_size for model training to 32 and the sequence_length to 5, and divided the training and test sets into a ratio of 8:2. In addition, we used MSE (Mean Squared Error) and MAPE (Mean Absolute Percentage Error) as the loss function and evaluation function, respectively.

[0064] Finally, for the dataset, the grid strategy determined a model structure with a MAPE of 0.68. The model used 2 LSTM layers, 1 Dense layer, 2 Dropout layers, and the number of neurons was determined to be 32. Figure 4 The figure shows the fitting effect of 200 data selected from the test set of the present invention, where the horizontal axis is the index value of the 200 data and the vertical axis is the temperature value.

[0065] In one optional implementation, the parameters of the neural network structure are manually adjusted based on expert experience. Specifically, the operator gradually fine-tunes key parameters such as the number of LSTM layers and neurons based on model training error, training duration, and overfitting trends. This approach can produce superior results when the operator has sufficient experience, but it relies heavily on the operator's skills and is difficult to standardize. It is therefore unsuitable for batch deployment or on-site adaptive training scenarios.

[0066] In another alternative implementation, parameter optimization employs a Bayesian optimization strategy. This strategy establishes a surrogate model to predict the relationship between hyperparameters and model performance, and guides the next round of parameter selection based on an expected improvement criterion. This method outperforms grid search in search efficiency, but is more complex to implement, requiring the construction of a high-dimensional function estimation model and carrying the risk of convergence instability.

[0067] The implementation method of the present application adopts a grid search strategy that traverses the entire parameter space, combined with clear performance evaluation indicators, to obtain the optimal LSTM model structure without relying on parameter adjustment experience, ensuring that the model training results are highly stable and reproducible, thereby improving the versatility and reliability of the prediction system in industrial field deployment.

[0068] In an embodiment of the present application, in step S4, a trained multi-layer LSTM network model is used to receive input temperature time series data of a fixed length, and based on the state transition and gating structure of the model, the predicted turbine bearing temperature value is output. The length of the input sequence is set by the Sequence_length parameter of the model.

[0069] To implement remote calls to the model service, this paper uses Gunicorn and Flask to build a web service. Flask is a lightweight web framework that allows developers to quickly build web applications, while Gunicorn is a WSGI (Web Server Gateway Interface) HTTP server that receives web requests and passes them to web applications (such as Flask applications) for processing.

[0070] The advantage of Gunicorn is that it uses a pre-forked worker process model, which means it can create multiple worker processes to handle incoming requests, helping to improve concurrent processing capabilities and response speed. Each worker process can handle client requests independently, thereby enhancing the robustness of the server. In addition, Gunicorn provides some more flexible request strategies, such as max_requests (the process automatically restarts to free up memory after reaching the maximum number of requests to prevent memory leaks) and max_requests_jitter (a process restart jitter parameter to prevent all processes from restarting at the same time).

[0071] Based on this, the present invention designs five API interfaces that can be called remotely, and the structural framework diagram is as follows: Figure 5 shown.

[0072] The init interface is used to process GET requests, return a random uuid to the client, and generate a folder corresponding to the uuid locally to store the model files saved by model training.

[0073] The data interface processes POST requests. The request body must carry two parameters: the uuid generated by the init interface and the data required for model training. The data interface saves the processed data as a dictionary to the local data.pkl file for subsequent use. The data interface returns a string containing the client's execution results.

[0074] The train interface handles POST requests, which require a single parameter in the request body: the uuid generated by the init interface. Based on this uuid, the trian interface deserializes the data dictionary from the data.pkl file and retrieves the required data using the uuid as the key. It then uses a grid search strategy to train and optimize the model. The model generated during the training phase is saved locally for subsequent use. The return value of the train interface is the MAPE value of the optimal model.

[0075] The predict API handles POST requests. The request body must carry two parameters: the uuid generated by the init API and the amount of data required for model inference, determined by the model's Sequence_length. The model is then inferred based on the passed uuid and data, and the predicted value is returned.

[0076] The clear API handles POST requests. The request body must carry one parameter: the uuid generated by the init API. The clear API is used to clear the model cache and locally stored model files to facilitate retraining.

[0077] Experiments have shown that the Gunicorn+Flask framework used can stably handle high concurrent requests according to the number of workers set.

[0078] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0079] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0080] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0081] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0082] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, it can be implemented using a combination of any of the following technologies known in the art: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0083] Example 4, reference Figure 6 , which is the fourth embodiment of the present invention, provides an adaptive turbine bearing temperature prediction system based on multi-layer LSTM, including a data acquisition and processing module, a sample construction module, a model construction optimization module and a prediction module.

[0084] The data acquisition and processing module is used to obtain the historical temperature data of the turbine bearings and use the Hampel filtering method to identify and replace abnormal values ​​in the temperature data.

[0085] The sample construction module is used to construct time series samples based on the processed data.

[0086] The model construction optimization module is used to input time series samples into a multi-layer LSTM network model, which includes multiple LSTM layers, Dropout layers, and fully connected layers, and optimize the relevant parameters for training.

[0087] The prediction module is used to predict the target time series data based on the trained multi-layer LSTM network model and output the future temperature value of the turbine bearing.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An adaptive turbine bearing temperature prediction method based on multi-layer LSTM, characterized by: include, Obtain historical temperature data of turbine bearings and use the Hampel filter method to identify and replace abnormal values ​​in the temperature data; Construct a time series sample based on the processed data; Input the time series samples into a multi-layer LSTM network model, which includes multiple LSTM layers, Dropout layers, and fully connected layers, and optimize the relevant parameters for training. Based on the trained multi-layer LSTM network model, the target time series data is predicted and the future temperature value of the turbine bearing is output.

2. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 1, characterized in that: The constructing of the time series samples includes generating continuous time window data based on the processed temperature data and extracting time series features.

3. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 2, characterized in that: The optimization training of relevant parameters includes: Parameter search is performed on the structural parameters of the multi-layer LSTM network model to obtain a parameter combination that meets the preset evaluation indicators.

4. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 3, characterized in that: The multi-layer LSTM network model includes LSTM layers, Dropout layers, and fully connected layers, the number and neuron scale of which are adjustable and serve as objects for parameter optimization.

5. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 1, characterized in that: The Hampel filtering method is used to identify and replace abnormal values ​​in temperature data. include, A sliding subsequence is constructed for the original temperature data with a preset window length. The median and median absolute deviation of each subsequence are calculated respectively. Data points that deviate from the median by more than a set multiple threshold are identified as outliers and replaced with the median of the corresponding subsequence.

6. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 4, characterized in that: The number and neuron size of the LSTM layer, Dropout layer and fully connected layer are adjustable and are the objects of parameter optimization, including: The number of LSTM layers, fully connected layers, Dropout layers, and the number of neurons in each layer are set. A grid search strategy is used to traverse each parameter combination. The mean absolute percentage error is used as the evaluation indicator to select the model structure corresponding to the optimal parameter combination.

7. The method for adaptive turbine bearing temperature prediction based on multi-layer LSTM according to claim 4, characterized in that: The target time series data is predicted based on the trained multi-layer LSTM network model, including: A trained multi-layer LSTM network model is used to receive fixed-length input temperature time series data and output predicted turbine bearing temperature values ​​based on the model's state transition and gating structure. The length of the input sequence is set by the model's Sequence_length parameter.

8. An adaptive turbine bearing temperature prediction system based on a multi-layer LSTM, applying the adaptive turbine bearing temperature prediction method based on a multi-layer LSTM according to any one of claims 1 to 7, characterized in that: include: Data acquisition and processing module, sample construction module, model construction and optimization module and prediction module; The data acquisition and processing module is used to obtain historical temperature data of the turbine bearing and identify and replace abnormal values ​​in the temperature data using the Hampel filtering method; The sample construction module is used to construct a time series sample based on the processed data; The model construction optimization module is used to input time series samples into a multi-layer LSTM network model, which includes multiple LSTM layers, Dropout layers, and fully connected layers, and optimize and train relevant parameters; The prediction module is used to predict the target time series data based on the trained multi-layer LSTM network model and output the future temperature value of the turbine bearing.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of an adaptive turbine bearing temperature prediction method based on multi-layer LSTM according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the adaptive turbine bearing temperature prediction method based on multi-layer LSTM according to any one of claims 1 to 7 are implemented.

Images