Generator remaining life prediction method, system, electronic device and storage medium

Abstract

The present invention relates to a method, system, electronic device, and storage medium for predicting the remaining life of a generator, comprising: obtaining historical data of a generator device; identifying the operating condition of the generator device at each moment using a preset operating condition identification method based on the historical data; preprocessing the operating condition at each moment to obtain a data sample set; training the data in the data sample set using an improved CNN module and an improved LSTM module to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate; and predicting the remaining life of the generator based on the hybrid fault preset model. The present invention overcomes the problems of existing prediction methods, such as insensitivity to fault characteristics, multiple parameters, and slow training speed, by effectively extracting multi-parameter features. Furthermore, the multi-feature data can be used to accurately predict equipment faults, reducing long-term downtime costs, optimizing maintenance cycles, and improving early fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and more particularly to a method, system, electronic device and storage medium for predicting the remaining life of a generator. Background Art

[0002] In terms of equipment failure prediction methods, prediction models can be categorized as physics-based, data-driven, and hybrid model-based approaches. Physics-based approaches use a physical model of the machine to perform reliability prediction, requiring a combination of physical systems and human expertise to build a specific model. Data-driven approaches leverage condition monitoring data to analyze and predict current and future health conditions. While these approaches can be applied to nonlinear predictions, they require a relatively high computational load compared to physics-based approaches. Hybrid model-based approaches combine the strengths of physics and data-driven approaches to improve prediction accuracy. However, their implementation is complex and, therefore, they are rarely used in practical scenarios. With the advancement of technologies such as sensors and data storage, data-driven RUL (Remaining Useful Life) prediction has become the predominant approach. Data-driven approaches primarily rely on neural networks or statistical methods, such as regression models and state-space models, most of which originate from pattern recognition theory. Data-driven approaches that utilize artificial intelligence tools do not require in-depth knowledge of system physics and offer the possibility of using several types of data as input.

[0003] Generators have complex internal structures, and manual fault diagnosis requires extensive prior knowledge, making it difficult to obtain large amounts of labeled data. The large-scale data accumulated by sensors is multidimensional and unlabeled data cannot be utilized. Neural networks, with their powerful nonlinear fitting, feature extraction, and feature mapping capabilities, are ideally suited to processing this massive amount of state data. Therefore, CNNs (Convolutional Neural Networks) were chosen as the prediction model. However, the convergence speed of neural networks is significantly affected by the selection of initial parameters. Appropriate initial parameters can effectively improve network training efficiency. Therefore, it is necessary to combine optimization algorithms with CNNs to achieve better prediction results.

[0004] However, the current hybrid fault prediction model has the following problems:

[0005] (1) Some algorithms still rely on manual feature extraction. An efficient algorithm is needed to automatically extract features. Existing prediction models do not take into account the temporal dependence of sensor monitoring data. It is necessary to consider how to use the time series characteristics of data to make more accurate predictions. In addition, the problem of limited labeled data in industrial big data needs to be solved.

[0006] (2) The traditional Long Short-Term Memory Network (LSTM) has defects such as too many parameters and slow training speed. The cell state at the current moment has no decision-making power on the cell state at the next moment, which breaks the contextual relevance of LSTM for time series prediction.

[0007] (3) When using traditional CNN for prediction, the obtained prediction curve is insensitive to small fluctuations of the actual curve, and the prediction algorithm shows "inertia" in fitting the actual curve. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a method, system, electronic device and storage medium for predicting the remaining life of a generator in response to the above-mentioned defects in the prior art.

[0009] The technical solution adopted by the present invention to solve the technical problem is to construct a method for predicting the remaining life of a generator, including the following steps:

[0010] Obtain historical data of generator equipment;

[0011] Based on the historical data, a preset operating condition identification method is used to identify the operating condition of the generator equipment at each moment;

[0012] Preprocessing the operating condition data at each moment to obtain a data sample set;

[0013] An improved CNN module and an improved LSTM module are used to train the data in the data sample set to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate;

[0014] The remaining life of the generator is predicted based on the hybrid fault preset model.

[0015] In the generator remaining life prediction method of the present invention, identifying the operating condition of the generator equipment at each moment using a preset operating condition identification method based on the historical data includes:

[0016] Based on the historical data, the operating condition identification method of K-means is used to identify the operating condition of the generator equipment at each moment.

[0017] In the generator remaining life prediction method of the present invention, the preprocessing of the operating condition data at each moment to obtain a data sample set includes:

[0018] Performing outlier processing, missing value filling, data normalization, and signal filtering on the operating condition data at each moment;

[0019] The signal filtering process includes: using a finite impulse response data filter to filter the operating condition data at each moment.

[0020] In the generator remaining life prediction method described in the present invention, the improved CNN module is a multi-layer one-dimensional CNN convolutional layer;

[0021] The improved LSTM is a superimposed multi-layer LSTM.

[0022] In the generator remaining life prediction method of the present invention, the use of the improved CNN module and the improved LSTM module to train the data in the data sample set to obtain a hybrid fault prediction model includes:

[0023] Extracting generator fault monitoring time series features from the data sample set based on the improved CNN module;

[0024] Inputting the generator fault monitoring time series features into the improved LSTM module;

[0025] The improved LSTM module is used to extract and train long-term and short-term time dependency features of the generator fault monitoring time series to obtain the hybrid fault prediction model.

[0026] The present invention also provides a generator remaining life prediction system, comprising:

[0027] An acquisition unit, used for acquiring historical data of the generator equipment;

[0028] an identification unit, configured to identify the operating condition of the generator equipment at each moment using a preset operating condition identification method based on the historical data;

[0029] A preprocessing unit, configured to preprocess the operating condition data at each moment to obtain a data sample set;

[0030] A model unit, configured to train the data in the data sample set using an improved CNN module and an improved LSTM module to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate;

[0031] A prediction unit is used to predict the remaining life of the generator based on the hybrid fault preset model.

[0032] In the generator remaining life prediction system of the present invention, the identification unit includes:

[0033] The operating condition identification module is used to identify the operating condition of the generator equipment at each moment based on the historical data using a K-means operating condition identification method.

[0034] In the generator remaining life prediction system of the present invention, the preprocessing unit includes:

[0035] A preprocessing module, configured to perform outlier processing, missing value filling, data normalization, and signal filtering on the operating condition data at each moment;

[0036] The signal filtering process includes: using a finite impulse response data filter to filter the operating condition data at each moment.

[0037] The present invention also provides an electronic device, comprising: a memory and a processor;

[0038] The memory is used to store computer programs;

[0039] The processor is configured to execute the computer program to implement the above-mentioned method for predicting the remaining life of a generator.

[0040] The present invention also provides a storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor processes the steps of the above-mentioned method for predicting the remaining life of a generator.

[0041] The generator remaining life prediction method, system, electronic device, and storage medium of the present invention have the following beneficial effects: comprising: obtaining historical data of the generator equipment; identifying the operating condition of the generator equipment at each moment using a preset operating condition identification method based on the historical data; preprocessing the operating condition at each moment to obtain a data sample set; training the data in the data sample set using an improved CNN module and an improved LSTM module to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate; and predicting the remaining life of the generator based on the hybrid fault preset model. The present invention overcomes the problems of existing prediction methods such as insensitivity to fault characteristics, large number of parameters, and slow training speed, thereby effectively extracting multi-parameter features and accurately predicting equipment faults using multi-feature data, reducing long-term downtime costs, optimizing maintenance cycles, and improving early fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0043] Figure 1 1 is a flow chart of a method for predicting the remaining life of a generator provided by an embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of the structure of an attention-based CNN module provided by an embodiment of the present invention;

[0045] Figure 3 Schematic diagram of the structure of an LSTM module with a peephole and an update gate provided by an embodiment of the present invention;

[0046] Figure 4 This is a principle block diagram of a generator remaining life prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0048] The present invention provides a method for predicting the remaining life of a generator. First, the traditional convolutional neural network (CNN) is improved, and a CNN based on the attention mechanism is proposed to extract the features of the generator fault monitoring time series. Then, the input gate of the long short-term memory network (LSTM) is removed and the forget gate is replaced with an update gate to obtain an LSTM module with a peephole and an update gate. The LSTM module with a peephole and an update gate is used to generate feature descriptions. The improved CNN module and the improved LSTM module are combined to form a hybrid fault prediction module. Finally, the hybrid fault prediction module is used to predict the life of the generator. By hybridizing the prediction algorithms of the CNN and LSTM models, the present invention focuses "attention" on the fault features, overcoming the problems of the existing prediction "laziness" (insensitivity to fault features), multiple parameters, and slow training speed, thereby effectively extracting multi-parameter features. The use of multi-feature data can accurately predict equipment failures, reduce long-term downtime costs, optimize maintenance cycles, and improve early fault detection.

[0049] Specifically, such as Figure 1 As shown, the generator remaining life prediction method includes the following steps:

[0050] Step S101: Acquire historical data of the generator equipment.

[0051] In the embodiment of the present invention, the historical data of the generator device includes but is not limited to the condition monitoring data of the generator device, such as pressure data, vibration data, speed data, temperature data, etc. These condition monitoring data can be obtained through corresponding sensors and / or database storage.

[0052] Step S102: Based on historical data, a preset operating condition identification method is used to identify the operating condition of the generator equipment at each moment.

[0053] Understandably, as the complexity and intelligent programming of generator equipment continue to increase, the corresponding equipment operating modes or operating conditions are also becoming more complex and variable. In the embodiments of the present invention, by first identifying the equipment operating conditions and then performing data preprocessing based on the identification results, the data preprocessing stage can eliminate the adverse effects of complex operating conditions.

[0054] Optionally, in an embodiment of the present invention, identifying the operating condition of the generator device at each moment using a preset operating condition identification method based on historical data includes: identifying the operating condition of the generator device at each moment using a K-means operating condition identification method based on historical data. Specifically, the present invention uses the K-means operating condition identification method to identify the device operating condition, and after unsupervised clustering training, can identify the operating condition of the device at each moment in the multidimensional sensor sequence.

[0055] In the embodiment of the present invention, during the K-means working condition identification process, the input historical data is the original multi-dimensional sensor sequence, which is a time series, and the K-means algorithm is used to perform working condition identification on the data of each moment of the equipment.

[0056] In the embodiment of the present invention, the distance metric in the K-means working condition identification algorithm adopts Euclidean distance, and the cluster center is initialized by random sampling. The specific steps are as follows:

[0057] a1. Randomly select K samples as the initial cluster centers, as shown in formula (1):

[0058]

[0059] Where, Z in formula (1) j k represents k initial cluster centers and Xj represents the jth sample.

[0060] a2. Classify the samples according to their distance from the cluster center, as shown in formula (2):

[0061]

[0062] Where, in formula (2), d(X1, Z j k ) represents the distance between the lth sample and the jth cluster center.

[0063] a3. Recalculate the cluster center, as shown in formula (3):

[0064]

[0065] Among them, in formula (3), Z j k+1 , which means that after clustering is completed and the average value is calculated, the average value is used as the new cluster center.

[0066] a4. Determine whether the stopping condition is met, that is, the cluster center hardly changes, as shown in formula (9). If the stopping condition is met, the result is returned directly; otherwise, continue with steps a2 and a3.

[0067] a5. Obtain clustering results and get cluster centers.

[0068] Furthermore, the generator equipment is defined as i The sensor monitoring data at time x i , then {x} represents the entire sensor sequence, the set of K types of operating conditions is Ω = {1, 2, ..., K}, and the operating condition identification method f(x i ) Identify the operating conditions of the equipment at each moment and record Indicates device t i The moment belongs to working condition k. Each operating condition is subjected to z-score standardization, and the corresponding mean and standard deviation are shown in formula (4) and formula (5) respectively.

[0069]

[0070] s (k) =Std({x} (k) ) (5).

[0071] Step S103: pre-process the operating condition data at each moment to obtain a data sample set.

[0072] Optionally, in an embodiment of the present invention, the operating condition data at each moment is preprocessed to obtain a data sample set, including: performing outlier processing, missing value filling, data normalization processing, and signal filtering processing on the operating condition data at each moment; the signal filtering processing includes: using a finite impulse response data filter to filter the operating condition data at each moment.

[0073] In the embodiment of the present invention, the operating condition data at each moment is preprocessed to achieve "cleaning" of the original data of the operating condition at each moment, thereby reducing interference from abnormal data.

[0074] Specifically, as in step S103, at any moment x in the original sensor monitoring data i After the working condition identification and data preprocessing process, it is converted into y i The specific process is as follows:

[0075]

[0076] (6) In the formula, y i Indicates t i The result of the working condition identification at the moment, x ni 、 and S n (k) They are x i 、 and s (k) The nth element of .

[0077] Furthermore, in the embodiment of the present invention, a finite impulse response digital filter (FIRDC) is used. Due to its all-zero structure, the FIRDC digital filter has no stability issues, can ensure arbitrary amplitude-frequency characteristics while having strict linear phase-frequency characteristics, and can use fast Fourier transform to speed up filtering operations. Therefore, it plays a very important role in digital filtering. The FIRDC filter has the following form:

[0078]

[0079] (7) where N is the filter length, x(k) is the value of the sensor measurement signal at time k, y(k) is the filtered signal value at time k, and {b(i), i = 1, 2, ..., N} represents the weighted coefficient sequence that determines the filter characteristics, whose sum is 1. When {b(i), i = 1, 2, ..., N} have the same value, it is an averaging filter.

[0080] Step S104: Use the improved CNN module and the improved LSTM module to train the data in the data sample set to obtain a hybrid fault prediction model.

[0081] In some embodiments, an improved CNN module and an improved LSTM module are used to train data in a data sample set to obtain a hybrid fault prediction model, including: extracting generator fault monitoring time series features in the data sample set based on the improved CNN module; inputting the generator fault monitoring time series features into the improved LSTM module; and the improved LSTM module is used to extract and train long-term and short-term time dependency features of the generator fault monitoring time series to obtain a hybrid fault prediction model.

[0082] Optionally, in an embodiment of the present invention, the improved CNN module is an attention-based CNN module. Specifically, the improved CNN module in an embodiment of the present invention is a multi-layer one-dimensional CNN convolution layer. The multi-layer one-dimensional CNN convolution layer performs convolution calculation along the time axis, the time window size is in the range of 15 to 30, and the convolution kernel size is set to 3 to 7. The stacked multi-layer one-dimensional CNN convolution layer uses zero padding to keep the time window size unchanged. The output format of the improved CNN module is: N window ×N ID-CNN , which can be directly input into the improved LSTM module.

[0083] Specifically, such as Figure 2 , which is a structural diagram of a CNN module based on the attention mechanism according to an embodiment of the present invention.

[0084] like Figure 2 As shown in the figure, assuming the input is m time series of length n, the size can be expressed as (n×m). The CNN module of the attention mechanism consists of two parts: feature aggregation and scale restoration. The feature aggregation part uses a stack of multiple convolutional and pooling layers to extract key features from cross-scale subsequences. The final layer uses a 1×1 convolution kernel to exploit linear relationships. The scale restoration part restores the key features to (n×m), that is, to the same size as the CNN module output features. The sigmoid function is then used to constrain the values between 0 and 1. The extracted contextual features are used as the saliency feature map of the basic features. The inputs of each parallel module have adjustable overlap. Adjusting the input overlap step size can better adapt the model to data with different patterns and further increase the overall input length of the joint model to capture more accurate long-term features. The non-dense overlap of the inputs between parallel modules can avoid overfitting.

[0085] The CNN module based on the attention mechanism, the attention module branch input sequence covers the CNN input, and stacks deep convolutional networks and pooling layers. The input perception corresponding to the features is also expanded, and the model obtains more comprehensive contextual information, thereby learning the importance of the current local sequence features. The attention module can increase the influence of important time series features on the model and suppress the interference of unimportant features on the model, effectively solving the problem that the model cannot distinguish the differences in the importance of time series data. At the same time, the standard CNN branch and the attention branch use sequences of different lengths as input. Multi-scale input can effectively extract richer short sequence features and prevent the problem of input cutting too small or too large in single-scale CNN, resulting in missing matching features during prediction.

[0086] The higher the importance of a feature output by a CNN module, the closer its corresponding attention module output is to 1. Conversely, the lower the importance of a feature output by a CNN module, the closer its corresponding attention module output is to 0. The numerical value reflects the importance of a feature, thereby enabling the identification of key features. The fused result serves as the input to a node in an LSTM module, which performs sequence modeling to obtain the final prediction result. The LSTM module extracts coarse-grained features from the significant fine-grained features extracted by the attention-based CNN. This refines the processing of features across various dimensions while preventing memory loss and gradient vanishing caused by excessive step sizes. This architecture captures the temporal dependencies of the effective features extracted by the attention-based optimized convolution operation, enabling the fusion of coarse and fine-grained features and a comprehensive characterization of time series data.

[0087] In the embodiment of the present invention, the improved LSTM module is an LSTM module with a peephole and an update gate, wherein the improved LSTM is a superimposed multi-layer LSTM.

[0088] like Figure 3 As shown in FIG, an LSTM module with a peephole and an update gate is provided in an embodiment of the present invention. t-1 The cell state is the state of the cell at the previous moment. The cell state is the memory state of the neuron at the past moment. The LSTM neural network completes the memory and transmission of time series features through it. t-1 is the output at the previous moment, x t is the input at the current moment, C t is the cell state at the current moment, h t is the output at the current moment.

[0089] In the embodiment of the present invention, the improved LSTM removes the input gate and replaces the forget gate with an update gate. The update gate u t Input x at the current time t 、The output of the last moment is h t-1 and the cell state C at the previous moment t-1 The update gate weight is multiplied by the augmented matrix composed of these three parameters plus the bias matrix, and the Sigmoid activation function outputs a number between 0 and 1. "0" means completely abandoning the cell state at the previous moment, and "1" means completely retaining the cell state at the previous moment. The update gate formula is as follows:

[0090] u t =σ(W u ·[C t-1 , h t-1 , x t ]+b u ) (8).

[0091] Where, σ in (8) is the sigmoid activation function, b u It is bias.

[0092] The function of the traditional LSTM "input gate" is now replaced by the update gate, which determines how much input the cell state receives at the current moment, and its value is 1-u t The current input status as follows:

[0093]

[0094] Among them, in formula (9), W i is the weight matrix of the input gate, b i is the bias corresponding to the input gate.

[0095] The current cell state is determined by the update gate. The formula for the current cell state is as follows:

[0096]

[0097] The output gate is similar to the input gate, and the current input x t 、The output of the last moment is h t-1 and the cell state C at the previous moment t-1 Together, the output gate controls how much of the cell state at the current moment flows as output to the next moment. The output gate formula is as follows:

[0098] o t =σ(W0·[h t-1 , x t ]+b0) (11).

[0099] Where, in formula (11), W0 is the weight matrix of the output gate, σ is the sigmoid activation function, and b0 is the corresponding input bias.

[0100] Finally, the cell state at the current moment passes through the tanh activation function and is output at the output gate o t Under the control of t , the current output is as follows:

[0101] h t =tanh(C t )*o t (12).

[0102] The embodiment of the present invention adopts superimposed multi-layer LSTM modules, which are mainly responsible for time-dependent feature extraction. In order to prevent overfitting of the training sample set, the Dropout regularization technology is adopted for each LSTM layer, that is, there is a certain probability of ignoring some neurons during training, and the parameter size is set to 0.2-0.5. After the LSTM layer, the feature vector is connected to the fully connected layer for regression prediction of the RUL value. All neural network layers use the ReLU function as the activation function. ReLU is a newer activation function. This activation function can more effectively perform gradient descent and error back propagation to avoid the problem of gradient disappearance. In addition, ReLU generalizes faster and can accelerate the convergence speed of gradient descent. The formula of the activation function is as follows:

[0103] f ReLU(x) =Max(0,x) (13).

[0104] The model loss is calculated using formula (12), and the Adam algorithm is used to optimize the network parameters based on the error.

[0105] L total (θ)=αL score (θ)+(1-α)L MSE (θ)+J(θ) (14).

[0106] In the formula, θ represents the network weight parameter, J(θ) represents the regularization term, and α represents L score The weight coefficient of (θ) is used as a hyperparameter of the model.

[0107] Step S105: predicting the remaining life of the generator based on the hybrid fault preset model.

[0108] The generator remaining life prediction provided by the embodiment of the present invention includes two stages: online application and offline modeling. The online and offline parts complement and support each other. In the offline stage, the test data and historical monitoring data of the equipment system are fully utilized to establish a generator remaining life prediction model; in the online stage, after necessary data preprocessing and other operations are performed on the data collected by real-time monitoring, the remaining life of the generator is predicted with the help of an intelligent algorithm based on the actual operating environment of the generator equipment (such as the external environment, load conditions, etc.). In this framework, offline training and learning are fully combined with online actual predictions, and the model can be adjusted offline with reference to the prediction effect in actual online applications, thereby constructing a closed-loop system from offline modeling, training, and learning to online actual application, and then dynamically optimizing the algorithm model based on the actual online application effect.

[0109] Furthermore, in the embodiment of the present invention, a time window sliding is used to construct training samples during training, with a step size of 1 and a time window size of N.window , the sensor data dimension is recorded as N sensor , then the input format is N window ×N sensor ,After slicing along the time axis, the original monitoring data is directly input into the ,hybrid model, without relying on domain knowledge or ,signal processing techniques to manually extract features.

[0110] In a specific embodiment, the improved CNN module can be designed with three layers, with the number of convolution kernels in each layer being 32, 48, and 64, and the size being 2. The improved CNN module can extract more temporal features by changing the number of total area kernels. Furthermore, increasing the number of LSTM network layers can increase the depth of the network, helping to improve the predictive ability of the prediction model. Alternatively, the present invention can establish a two-layer LSTM, with the number of neurons in each layer being 32 or 48.

[0111] The prediction method of the embodiment of the present invention utilizes the advantages of CNN's powerful feature self-extraction capability and LSTM's time series sensitivity, and can fully extract the characterization capabilities of data in both spatial and temporal dimensions, better fit the nonlinear relationship between the remaining service life and monitoring variables in the short term, and the prediction results are closer to the measured results. It is a more advanced generator remaining life prediction model.

[0112] Furthermore, based on the traditional LSTM model, we propose an update gate mechanism with a "peephole." This update gate, combining a forget gate and an output gate, effectively reduces parameter size and improves training speed. This "peephole" mechanism allows the current cell state to contribute to the determination of the next cell state, further strengthening the context of time series data and improving prediction accuracy.

[0113] Furthermore, the attention-based CNN can increase the algorithm's sensitivity to fault signatures, focusing attention on local sequences. Furthermore, the improved CNN can perform necessary filtering on most normal data, allowing the algorithm to focus more on the fluctuation patterns of local sequences, effectively extracting their features.

[0114] At the same time, the trend in detecting anomalies in time series data is to use an anomaly detection algorithm that combines two models. Hybrid fault prediction models apply deep learning concepts to industrial big data, fully utilizing unlabeled data and improving the predictive capabilities of fault prediction algorithms by learning about generator equipment failures.

[0115] like Figure 4 FIG2 is a block diagram of the principle of the generator remaining life prediction system provided by the present invention. The generator remaining life prediction system can be used to implement the generator remaining life prediction method disclosed in the embodiment of the present invention.

[0116] like Figure 4 As shown, the generator remaining life prediction system includes:

[0117] The acquisition unit 401 is used to acquire historical data of the generator equipment.

[0118] The identification unit 402 is used to identify the operating condition of the generator equipment at each moment based on historical data using a preset operating condition identification method.

[0119] The pre-processing unit 403 is used to pre-process the operating condition data at each moment to obtain a data sample set.

[0120] The model unit 404 is configured to train the data in the data sample set using an improved CNN module and an improved LSTM module to obtain a hybrid fault prediction model. The improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate.

[0121] The prediction unit 405 is configured to predict the remaining life of the generator based on the hybrid fault preset model.

[0122] Specifically, in some embodiments, the identification unit 402 includes: an operating condition identification module, which is used to identify the operating condition of the generator equipment at each moment based on historical data using a K-means operating condition identification method.

[0123] In some embodiments, the preprocessing unit 403 includes a preprocessing module configured to perform outlier processing, missing value filling, data normalization, and signal filtering on the operating condition data at each moment. The signal filtering includes filtering the operating condition data at each moment using a finite impulse response (FIR) data filter.

[0124] The present invention also provides an electronic device, including a memory and a processor.

[0125] The memory is used to store computer programs.

[0126] The processor is used to execute a computer program to implement the generator remaining life prediction method disclosed in an embodiment of the present invention.

[0127] The present invention also provides a storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor processes the steps of the method for predicting the remaining life of a generator disclosed in an embodiment of the present invention.

[0128] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the present invention and implement it accordingly. They are not intended to limit the scope of protection of the present invention. All equivalent variations and modifications within the scope of the claims of the present invention are intended to be covered by the claims of the present invention.

Claims

1. A method for predicting the remaining life of a generator, characterized in that: The following steps are involved: Acquire historical data of the generator equipment; wherein the historical data includes pressure data, vibration data, speed data, and temperature data; Based on the historical data, a preset operating condition identification method is used to identify the operating condition of the generator equipment at each moment; Preprocessing the operating condition data at each moment to obtain a data sample set; An improved CNN module and an improved LSTM module are used to train the data in the data sample set to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate; Predicting the remaining life of the generator based on the hybrid fault preset model; Among them, the attention-based CNN module also includes a parallel standard CNN branch and an attention branch. The standard CNN branch and the attention branch perform multi-scale feature extraction with input sequences of different lengths; in the CNN module based on the attention mechanism, the attention branch input sequence covers the CNN input, and stacks deep convolutional networks and pooling layers. The input perception corresponding to the features is also expanded, and more comprehensive context information is obtained, thereby learning the importance of the current local sequence features. Through the CNN module with the attention mechanism, the influence of key temporal features on the model can be improved; The formula for the update gate is as follows: ； Where σ is the sigmoid activation function, b u is bias; The function of the traditional LSTM "input gate" is now replaced by the update gate, which determines how much input the cell state receives at the current moment. Its value is , the input status at the current moment as follows: ； Where, Wi is the weight matrix of the input gate, b i is the bias corresponding to the input gate; The current cell state is determined by the update gate. The formula for the current cell state is as follows: ； The output gate is similar to the input gate, and is input by the current moment , the last moment output and the cell state at the previous moment Together, the output gate controls how much of the cell state at the current moment flows as output to the next moment. The output gate formula is as follows: about t =σ(W0·[h t-1 ,x t ]+b0); Where, W0 is the weight matrix of the output gate, σ is the sigmoid activation function, and b0 is the corresponding input bias; Finally, the cell state at the current moment passes through the tanh activation function and is activated at the output gate. Under the control of , the current output is as follows: 。 2. The method for predicting the remaining life of a generator according to claim 1, characterized in that: The method of identifying the operating condition of the generator equipment at each moment using a preset operating condition identification method based on the historical data includes: Based on the historical data, the operating condition identification method of K-means is used to identify the operating condition of the generator equipment at each moment.

3. The method for predicting the remaining life of a generator according to claim 1, characterized in that: The preprocessing of the operating condition data at each moment to obtain a data sample set includes: Performing outlier processing, missing value filling, data normalization, and signal filtering on the operating condition data at each moment; The signal filtering process includes: using a finite impulse response data filter to filter the operating condition data at each moment.

4. The method for predicting the remaining life of a generator according to claim 1, characterized in that: The improved CNN module is a multi-layer one-dimensional CNN convolutional layer; The improved LSTM is a superimposed multi-layer LSTM.

5. The method for predicting the remaining life of a generator according to claim 1, characterized in that: The method of using the improved CNN module and the improved LSTM module to train the data in the data sample set to obtain a hybrid fault prediction model includes: Extracting generator fault monitoring time series features from the data sample set based on the improved CNN module; Inputting the generator fault monitoring time series features into the improved LSTM module; The improved LSTM module is used to extract and train long-term and short-term time dependency features of the generator fault monitoring time series to obtain the hybrid fault prediction model.

6. A generator remaining life prediction system, characterized in that: include: An acquisition unit, configured to acquire historical data of the generator device; wherein the historical data includes pressure data, vibration data, speed data, and temperature data; an identification unit, configured to identify the operating condition of the generator equipment at each moment using a preset operating condition identification method based on the historical data; A preprocessing unit, configured to preprocess the operating condition data at each moment to obtain a data sample set; A model unit, configured to train the data in the data sample set using an improved CNN module and an improved LSTM module to obtain a hybrid fault prediction model; the improved CNN module is an attention-based CNN module, and the improved LSTM module is an LSTM module with a peephole and an update gate; A prediction unit, configured to predict the remaining life of the generator based on the hybrid fault preset model; Among them, the attention-based CNN module also includes a parallel standard CNN branch and an attention branch. The standard CNN branch and the attention branch perform multi-scale feature extraction with input sequences of different lengths; in the CNN module based on the attention mechanism, the attention branch input sequence covers the CNN input, and stacks deep convolutional networks and pooling layers. The input perception corresponding to the features is also expanded, and more comprehensive context information is obtained, thereby learning the importance of the current local sequence features. Through the CNN module with the attention mechanism, the influence of key temporal features on the model can be improved; The formula for the update gate is as follows: ； Where σ is the sigmoid activation function, b u is bias; The function of the traditional LSTM "input gate" is now replaced by the update gate, which determines how much input the cell state receives at the current moment. Its value is , the input status at the current moment as follows: ； Where, Wi is the weight matrix of the input gate, b i is the bias corresponding to the input gate; The current cell state is determined by the update gate. The formula for the current cell state is as follows: ； The output gate is similar to the input gate, and is input by the current moment , the last moment output and the cell state at the previous moment Together, the output gate controls how much of the cell state at the current moment flows as output to the next moment. The output gate formula is as follows: about t =σ(W0·[h t-1 ,x t ]+b0); Where, W0 is the weight matrix of the output gate, σ is the sigmoid activation function, and b0 is the corresponding input bias; Finally, the cell state at the current moment passes through the tanh activation function and is activated at the output gate. Under the control of , the current output is as follows: 。 7. The generator remaining life prediction system according to claim 6, characterized in that: The identification unit includes: The operating condition identification module is used to identify the operating condition of the generator equipment at each moment based on the historical data using a K-means operating condition identification method.

8. The generator remaining life prediction system according to claim 6, characterized in that: The pre-processing unit comprises: A preprocessing module, configured to perform outlier processing, missing value filling, data normalization, and signal filtering on the operating condition data at each moment; The signal filtering process includes: using a finite impulse response data filter to filter the operating condition data at each moment.

9. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program to implement the generator remaining life prediction method according to any one of claims 1 to 5.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the processor processes the steps of the method for predicting the remaining life of a generator according to any one of claims 1 to 5.

Images