CN-121997293-A - Frequency-time cooperative enhancement transducer model and prediction method for PEMFC voltage decay prediction

CN121997293ACN 121997293 ACN121997293 ACN 121997293ACN-121997293-A

Abstract

The invention discloses a frequency-time cooperative enhancement transducer model and a prediction method for fuel cell voltage decay prediction, and belongs to the field of fuel cell health management and artificial intelligence. The model comprises data preprocessing, F-Block, time-frequency fusion and enhancement encoder and decoder modules, and is characterized in that the core is a double-branch cooperative framework, wherein F-Block extracts frequency domain global trend characteristics, convAdapter enhances local perception of a time domain, and time-frequency complementation is realized through channel splicing, weighted summation or attention fusion. The prediction method is used for completing prediction through data preprocessing, double-branch feature extraction, time-frequency fusion, enhanced coding and autoregressive generation. The method solves the problems of insufficient long-range trend capture and weak local feature perception of the existing model, has the advantages of high prediction precision, strong generalization, light model and the like, can accurately predict the voltage decay trend, provides support for predictive maintenance of the fuel cell, and is suitable for various fuel cell application scenes.

Inventors

SHI YONG
HU ZHILONG
SU JIANHUI
XIE BAO
LAI JIDONG

Assignees

合肥工业大学

Dates

Publication Date: 20260508
Application Date: 20260112

Claims (10)

1. The frequency-time cooperative enhancement transducer model for PEMFC voltage decay prediction is characterized by comprising a data preprocessing module, a frequency domain feature extraction module, a time domain representation learning module, a time-frequency fusion module, an enhancement encoder module and a decoder module, wherein the modules are connected and cooperate through data streams; The data preprocessing module is used for performing feature screening, data resampling, signal filtering and standardization processing on the original fuel cell monitoring data and outputting the regular input data; The frequency domain feature extraction module is used for executing frequency domain transformation and feature extraction on input data, converting a time domain sequence into a frequency domain representation through the expansion Fourier transformation unit, modeling the correlation among frequencies through the complex self-attention mechanism, and outputting the frequency domain representation rich in global periodicity and trend information through the frequency domain reconstruction unit; The time domain representation learning module is used for carrying out embedding processing and position coding on input data and outputting basic time domain representation; The time-frequency fusion module is used for combining the frequency domain representation and the basic time domain representation in a channel dimension splicing, weighting summation or attention fusion mode, and converting the frequency domain representation and the basic time domain representation into a transform hidden space through linear mapping to form a time-frequency joint representation, wherein the combination mode needs to keep the complementation of time-frequency information; the enhancement encoder module consists of a plurality of layers of Transformer encoders, a time domain local enhancement module is embedded after part of encoder layers selectively, the time domain local enhancement module realizes local feature extraction through one-dimensional convolution, channel compression and recovery and residual connection, and global and local feature modeling is completed in cooperation with a self-attention mechanism; the decoder module generates a fuel cell voltage decay prediction sequence based on the time-frequency joint representation and the autoregressive input.
2. The frequency-time cooperative enhancement converter model for PEMFC voltage decay prediction according to claim 1, wherein the data preprocessing module comprises a feature screening sub-module, a data resampling sub-module, a signal filtering sub-module and a normalizing sub-module, wherein the feature screening sub-module adopts a pearson correlation coefficient and spearman rank correlation coefficient method to screen seven variables most relevant to voltage variation as input features, the data resampling sub-module adopts a fixed-step sliding average method, average values are taken as new samples every 5 continuous samples, the signal filtering sub-module adopts a Savitzky-Golay filter, and the normalizing sub-module adopts a Z-Score method, and parameters are only calculated from training sets to avoid data leakage.
3. The frequency-time co-enhancement fransformer model for PEMFC voltage decay prediction according to claim 2, wherein the feature screening submodule uses the following calculation formula: Wherein, the And (3) with Representing the input variable and the voltage output samples respectively, 、 Is the average value thereof; Represent the first The two variables of the individual samples are sorted by difference, And (3) with The ranking of the samples in the respective variables is respectively, Is the total number of samples; the data resampling submodule adopts the following calculation formula: Wherein, the Representing the average window length of the window, Is the original sample number; The standardized submodule outputs the preprocessed data as input of a subsequent module: Wherein, the To input data, i.e. raw sensor monitor data values, And Respectively represent the mean value and standard deviation of the features in the training set.
4. The frequency-time cooperative enhancement transducer model for PEMFC voltage decay prediction according to claim 1, wherein the extended Fourier transform unit of the frequency domain feature extraction module extends an input sequence to integer times of a prediction window in a time dimension, then performs fast Fourier transform to obtain a frequency domain representation, the complex self-attention mechanism unit generates a query, a key and a value vector through a complex weight matrix, calculates attention weights based on vector modular length, and the frequency domain reconstruction unit restores the weighted frequency domain features to a time domain sequence through inverse Fourier transform.
5. The PEMFC voltage decay prediction oriented frequency-time co-enhancement fransformer model of claim 4, wherein said extended fourier transform unit employs the following calculation formula: ; representing the normalized frequency domain feature matrix for enhancing the stability of the frequency spectrum data; representing the mean value of the frequency domain characteristics; Representing standard deviation of frequency domain characteristics; Is the number of frequency points; The complex self-attention mechanism unit adopts the following calculation formula: representing the first in a multi-head attention mechanism A plurality of attention heads; respectively represent the first A complex linear transformation weight matrix of queries (Query), keys (Key) and values (Value) corresponding to the individual heads; Respectively representing the generated complex query vector, key vector and value vector; Represent the first Output results of the individual attention heads; The frequency domain reconstruction unit adopts the following calculation formula: the output representation of the F-Block module after frequency domain coding is subjected to inverse Fourier transform Recovering; The frequency domain reconstruction sequence representing the finally output prediction interval has the dimension of Wherein To predict window length.
6. The frequency-time co-enhancement transducer model for PEMFC voltage decay prediction according to claim 1, wherein the specific implementation steps of the time domain local enhancement module include: (1) Rearranging the characteristic tensor input by the encoder layer in the channel dimension and the time dimension to adapt to one-dimensional convolution; (2) Channel compression, namely compressing the dimension of a channel by adopting one-dimensional convolution with the convolution kernel size k of 3 or 5, wherein the compression ratio r is 2-4, and ensuring lightweight design; (3) Nonlinear activation, namely enhancing nonlinear expression capacity of a model by applying a ReLU activation function; (4) Channel recovery, namely recovering the channel dimension to the original dimension through one-dimensional convolution, and enhancing generalization capability by applying a Dropout mechanism; (5) Residual connection, namely carrying out residual fusion on the processed characteristics and the original input characteristics and outputting the processed characteristics, so as to ensure that local enhancement does not interfere with a global attention path.
7. The PEMFC voltage decay prediction oriented frequency-time co-enhancement fransformer model of claim 1, wherein the selective insertion strategy of the time domain local enhancement module (ConvAdapter) in the enhancement encoder module is to insert only after the shallow layers of the encoder, the remaining layers maintaining the original fransformer encoder structure; ; Represent the first An output of the layer encoder; Preset insertion layer set if Indicating insertion ConvAdapter only at layer 0; the fused features enter an encoder, and complementary modeling of global and local features is realized through the cooperative action of a self-attention mechanism and ConvAdapter; representing the final output voltage prediction sequence of the model.
8. The frequency-time co-enhancement transducer model for PEMFC voltage decay prediction according to claim 1, wherein the linear mapping is satisfied by the time-frequency fusion module by way of channel dimension stitching, the combined feature dimension is converted into a hidden space dimension consistent with the transducer model dimension, and the mapping formula is Wherein To represent the time-domain embedded feature after truncation, For the frequency domain features from F-Block, As a matrix of weights, the weight matrix, As a result of the bias term, To activate the function.
9. The frequency-time co-enhancement fransformer model for PEMFC voltage decay prediction according to claim 1, wherein the model is implemented based on PyTorch framework, and the super-parameter setting comprises Batchsize of 128, epochs of 100, adam used by an optimizer, MSE used by a loss function, model dimension of 512, attention header number of 8, encoder layer number of 2, and decoder layer number of 1.
10. A fuel cell voltage degradation prediction method based on the model of any one of claims 1 to 9, comprising the steps of: (1) The data input is to acquire multivariable monitoring data including voltage, current and temperature operation parameters in the operation process of the fuel cell; (2) The data preprocessing, namely performing feature screening, sliding average resampling, savitzky-Golay filtering and Z-Score standardization processing on the original monitoring data through a data preprocessing module, and outputting a preprocessed input sequence; (3) The frequency domain branch outputs frequency domain features rich in global trend through expanding Fourier transform, and the time domain branch outputs basic time domain features; (4) The time-frequency characteristic fusion is carried out, namely a time-frequency fusion module is used for combining the frequency domain characteristic and the time domain characteristic in a channel dimension splicing, weighting summation or attention fusion mode, and a time-frequency joint representation is generated through linear mapping; (5) The enhanced coding is that the time-frequency joint representation is input into an enhanced coder module, and the comprehensive modeling of the full-situation length Cheng Yilai and the local dynamic characteristics is completed through the synergistic effect of a self-attention mechanism and a time domain local enhancement module; (6) And (3) degradation prediction, namely generating a fuel cell voltage sequence in a future time period by the decoder module based on the coded characteristics and the autoregressive input to complete the voltage degradation trend prediction.

Description

Frequency-time cooperative enhancement transducer model and prediction method for PEMFC voltage decay prediction Technical Field The invention relates to the technical field of fuel cell health management and artificial intelligence prediction, in particular to a PEMFC voltage decay prediction-oriented frequency-time cooperative enhancement transducer model and a prediction method. Background Proton Exchange Membrane Fuel Cells (PEMFCs) are used as core representatives of new generation clean energy technologies, and show wide application prospects in multiple fields of new energy automobiles, portable power sources, fixed power stations and the like by virtue of the outstanding advantages of high energy density, zero emission, quick start and the like. However, during long-term operation of the PEMFC, the output voltage of the core performance index is irreversibly degraded due to the comprehensive influence of multiple factors such as internal complex electrochemical reaction, aging loss of the core material, fluctuation of external dynamic load working conditions and the like. The voltage decay directly restricts the service life and the operation reliability of the PEMFC, and becomes a key bottleneck for preventing the commercial large-scale popularization of the PEMFC. Therefore, the accurate prediction of the voltage decay trend of the PEMFC is realized, is a core link for constructing a predictive maintenance and health management (PHM) system of the PEMFC, and has important significance for early warning of sudden faults, making scientific and reasonable predictive maintenance strategies and reducing the total life cycle cost of equipment. In the development process of the prior art, the PEMFC voltage decay and remaining service life (RUL) prediction method mainly evolves into two major technical paths: First, the model drives the path. This path attempts to characterize aging mechanisms within the fuel cell, such as catalyst active area decay, proton exchange membrane perforation, electrode flooding, etc., by building an accurate physicochemical mathematical model. Although the method has advantages in mechanism interpretation and can obtain a certain prediction effect under the ideal condition of stable working conditions, the inherent limitation exists that the PEMFC is a strongly coupled nonlinear complex system involving electricity, heat and mass transfer, a universal and accurate physical model is difficult to build in an actual complex and changeable operating environment, and more importantly, the calibration process of model parameters is severely dependent on expensive and time-consuming experimental data support, so that the engineering suitability and the economical efficiency of the PEMFC are poor, and the practical application requirements are difficult to meet. And two, a data driving path. With the rapid development of sensing detection technology and computer computing power, the path deduces future performance trend by analyzing historical monitoring data accumulated in the running process of equipment, and is widely focused because of not relying on complex physical prior knowledge, so that the path has become the main technical direction in the field. The technical development of the method can be subdivided into several stages: In the early stage, the traditional statistical learning methods such as Gaussian process regression, support vector machine and the like are generally adopted. The methods have certain effects when processing small samples and low-dimensional data, but when facing high-frequency, multidimensional and massive monitoring data generated in industrial sites, the model expression capacity and generalization capacity are insufficient, and the prediction precision is difficult to guarantee; in the deep learning emerging phase, recurrent Neural Networks (RNNs) and variants thereof, such as long short term memory networks (LSTM), become the mainstream model in this field by virtue of their modeling capabilities of time series dependencies. However, RNN models have inherent problems of low training efficiency and vanishing gradient due to their sequence-dependent recursive computation structure, making it difficult to effectively capture long-range fading trends spanning hundreds or even thousands of hours, and the prediction results often have hysteresis or bias; Subsequently, convolutional Neural Networks (CNNs) were introduced to capture local features, but their receptive fields were limited by convolutional kernel size, with natural drawbacks in global trend modeling that were unable to compromise long-range dependence with local detail; at the current stage, the transducer model can process sequences in parallel and establish direct long-range dependency relationship due to the self-attention mechanism, and has great potential in a time sequence prediction task, so that the transducer model gradually becomes a research hotspot. However, there are two core t