Retired battery soh trajectory prediction method and system based on hdcn network

Abstract

The application discloses a retired battery SOH trajectory prediction method and system based on an HDCN network. The method first collects time data of the retired battery, performs pretreatment, calculates an SOH curve, obtains a battery SOH dataset, constructs a sliding window dataset, takes historical data in the window as input, takes a value at a next moment at the end of the window as a target, and constructs a training sample. Secondly, an HDCN network including a trend prediction subnetwork and a cycle prediction subnetwork is constructed, feature extraction is performed on the training sample, and a prediction sequence of each component is obtained. Finally, the test set sample is input into the trained HDCN network for rolling prediction, a prediction sequence of each component is obtained, and the prediction sequences of all components are added to reconstruct a battery SOH trajectory prediction result. The application realizes effective decoupling of an SOH signal, adaptive extraction of key timing characteristics, and accurate and efficient battery SOH trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of battery health management technology, and relates to battery SOH prediction technology, specifically to a method and system for predicting the SOH trajectory of retired batteries based on HDCN network. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, energy storage systems, and portable electronic devices due to their high energy density and long cycle life. Irreversible capacity degradation occurs during long-term use, making accurate prediction of their State of Health (SOH) degradation trajectory crucial for battery safety management, maintenance strategy development, and remaining lifespan assessment. Currently, data-driven SOH prediction methods are mainly based on models such as recurrent neural networks (RNN / LSTM) and temporal convolutional networks (TCN). However, battery SOH signals exhibit significant non-stationary characteristics, encompassing both long-term monotonic degradation trends and localized periodic fluctuations caused by capacity regeneration. A single model struggles to simultaneously capture these two different timescales. Existing models often employ fixed-length sliding windows for local modeling, relying solely on recent observations within the window for each prediction, lacking the ability to memorize degradation patterns over longer time spans. This results in the underutilization of valuable information accumulated during early degradation stages. Furthermore, the contribution of each time step within the sliding window to the prediction is uneven; recent capacity abrupt changes often carry more critical degradation cues, but traditional methods frequently apply equal weights or fixed pooling operations to each time step, failing to adaptively identify and highlight the features of key time steps. To address these issues, there is an urgent need for a battery SOH trajectory prediction method that can effectively decouple complex SOH signals, possess long-range degradation memory capabilities, and adaptively extract key temporal features. Summary of the Invention

[0003] To address the limitations of existing data-driven methods in areas such as multi-timescale decoupling of non-stationary battery SOH signals, long-range degradation memory loss within fixed sliding windows, and adaptive identification of key time steps within the window, this invention proposes a method and system for predicting the SOH trajectory of decommissioned batteries based on an HDCN network. This includes battery SOH trajectory prediction based on Peak Guided Fourier Mode Decomposition (PFMD) and Hybrid Decomposition Context Network (HDCN). This method effectively improves the prediction accuracy and robustness of complex degradation trajectories by separately modeling long-term monotonic degradation and capacity regeneration fluctuations, establishing a global dependency across early degradation stages, and adaptively enhancing the feature weights of key time steps such as capacity mutations.

[0004] In one aspect, this invention provides a method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks, comprising the following steps:

[0005] S1. Collect time data such as voltage, current, and temperature of retired batteries, extract the constant current charging stage operation data, and preprocess and calculate the SOH curve to obtain the battery SOH dataset.

[0006] S2. Based on the battery SOH dataset, the SOH curve is decomposed into a trend component and K-1 periodic components using the peak-guided Fourier mode decomposition (PFMD) method. Sliding window datasets are constructed for the trend component and K-1 periodic components respectively. The historical data within the window is used as input and the value at the next moment after the end of the window is used as the target to construct training samples.

[0007] S3. Construct the HDCN network, which includes a trend prediction subnetwork and a period prediction subnetwork. The trend prediction subnetwork is a multi-layer LSTM network that integrates a context library and inter-layer gated attention residuals, while the period prediction subnetwork is a multi-layer TCN network that integrates a context library and inter-layer gated attention residuals.

[0008] S4. After defining the training hyperparameters and network hyperparameters, input the training samples of the trend component into the trend prediction subnetwork in the HDCN network for training; input the training samples of the K-1 period components into the period prediction subnetwork in the HDCN network for training.

[0009] S5. Input the test set samples into the trained HDCN network for rolling prediction to obtain the prediction sequence of each component. Then, add the prediction sequences of all components together to reconstruct the battery SOH trajectory prediction result.

[0010] In another aspect, the present invention also provides a decommissioned battery SOH trajectory prediction system based on HDCN network, comprising the following modules:

[0011] The battery SOH data module is used to collect time data of retired batteries, extract the operating data of the constant current charging stage, perform preprocessing to calculate the SOH curve, and obtain the battery SOH dataset.

[0012] The training sample module, based on the battery SOH dataset, constructs a sliding window dataset, using historical data within the window as input and the value of the next moment after the last position of the window as the target, to build training samples.

[0013] The HDCN network prediction module is used to construct an HDCN network that includes a trend prediction subnetwork and a period prediction subnetwork. It extracts features from the training samples to obtain the prediction sequence of each component.

[0014] The SOH trajectory prediction output module is used to input test set samples into the trained HDCN network for rolling prediction, obtain the prediction sequence of each component, and add the prediction sequences of all components to reconstruct the battery SOH trajectory prediction result.

[0015] The present invention has the following beneficial effects:

[0016] This invention proposes a Peak-Guided Fourier Decomposition (PFMD) method. By detecting local peaks in the spectral amplitude of the SOH signal and using the midpoint of adjacent peaks as adaptive frequency band boundaries, it decouples the non-stationary degradation signal into intrinsic mode components at different frequency scales, avoiding mode aliasing in EMD and the computational overhead of VMD iterative optimization. Based on this, an HDCN network is constructed to perform heterogeneous differentiated modeling of components in different frequency bands: the low-frequency trend components with smooth changes and strong time-dependent characteristics are fed into an LSTM subnetwork, utilizing its gated memory mechanism to capture long-term degradation trends; the high-frequency components containing local fluctuations and capacity regeneration effects are fed into a TCN subnetwork, using dilated causal convolution to extract multi-scale local modes. Each subnetwork only needs to learn the degradation rules within a single frequency band, effectively decoupling complex SOH signals.

[0017] To overcome the information limitations of sliding windows, a context library mechanism is introduced. This mechanism continuously accumulates the degradation feature vectors of historical windows through progressive EMA updates, enabling prediction by retrieving and fusing long-range degradation context using cosine similarity, thus enhancing the perception of degradation patterns throughout the entire lifecycle. Window features are generated through attention-weighted aggregation, adaptively focusing on key time steps to improve the discriminative power of query vectors. Simultaneously, attention residual modules with learnable gating scalars are embedded between adjacent layers of the backbone network, supporting global information interaction at any time step within the window. This compensates for the long-distance decay in LSTM's progressive propagation and the limited receptive field of TCN. The gating value is initialized from near zero, with gradients autonomously adjusting attention contributions. When no benefit is gained, it automatically degrades to protect backbone features. This enhances the global context awareness of intermediate layer features while preserving original representation capabilities, providing long-range degradation memory and adaptively extracting key temporal features, achieving accurate and efficient prediction of battery SOH trajectories. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the network structure in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the context library in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further explained below with reference to the accompanying drawings;

[0021] like Figure 1 and Figure 2 As shown, in one aspect, the present invention provides a battery SOH trajectory prediction method based on HDCN, comprising the following steps:

[0022] S1. Collect voltage, current, capacity and temperature data of retired batteries during constant current charging. After preprocessing such as removing outliers and smoothing filtering, calculate the health status SOH curve using methods such as capacity increment analysis. Then, standardize the SOH curve to obtain the battery SOH dataset.

[0023] S2. Based on the battery SOH dataset, to address the non-stationary characteristics of the battery SOH signal and the superposition of long-term monotonic degradation trends and local periodic fluctuations, this method employs Peak-Guided Fourier Mode Decomposition (PFMD) to adaptively decompose the standardized SOH sequence, separating a trend component and at least one periodic component reflecting fluctuations such as capacity regeneration. Sliding window datasets are then constructed for the trend component and each periodic component, with the window width determined... , will continuously within the window Using historical values ​​as input features and the value of the next time step after the last window as the prediction target, supervised training samples are formed by sliding point by point along the time axis:

[0024] S2-1. Based on battery SOH data, Peak-Guided Fourier Mode Decomposition (PFMD) is used to adaptively decompose the SOH sequence in the battery SOH dataset, separating a trend component and K-1 periodic components reflecting fluctuations such as capacity regeneration:

[0025] S2-1-1, For SOH sequences of length N Perform a Discrete Fourier Transform (DFT) to obtain the frequency domain representation. :

[0026]

[0027] in Here, j represents the complex spectral coefficient at frequency index k, where j is the imaginary unit. The SOH value at point n in the sequence corresponds to the normalized frequency. .

[0028] S2-1-2, Calculate the spectral amplitude :

[0029]

[0030] in, For the frequency domain representation obtained in step S2-1-1, Re(·) and Im(·) are the operations of taking the real part and taking the imaginary part, respectively, and ⌊N / 2⌋ represents rounding down N / 2, covering all discrete frequency points in the non-negative frequency range [0,0.5].

[0031] Secondly, the detection is based on all spectral amplitudes, detecting the set M of local maxima of the spectral amplitudes:

[0032]

[0033] That is, for each internal frequency index k, if its amplitude is simultaneously greater than the amplitudes of its left and right adjacent points, it is assigned to the set of local maxima. The frequency band boundary is determined by detecting the set M of local maxima in the amplitude spectrum, and the frequency band with a frequency range of [0, 0.5] is adaptively divided into K sub-bands:

[0034]

[0035] in This represents the number of local maxima. If the number of local maxima is less than K-1, a uniform partition is used, and the i-th boundary frequency is... If there are enough maxima, select the K-1 points with the largest amplitudes from M, sort them in ascending order of frequency, and the i-th point is the maximum value. After the boundary frequencies are determined, the frequency range [0, 0.5] is finally divided into K frequency bands: .

[0036] S2-1-3. Design FIR filters for the K sub-bands obtained in S2-1-2: a low-pass filter for the lowest frequency band, a band-pass filter for the middle frequency band, and a high-pass filter for the highest frequency band. The coefficients of the three filters are generated differently: the low-pass and high-pass filters are designed with a single cutoff frequency, while the band-pass filter is designed with two cutoff frequencies. However, the filtering execution method is the same for all three: forward and reverse zero-phase filtering (filtfilt) is used. The same set of filter coefficients is used to perform forward and backward filtering on the signal to eliminate phase distortion and extract the corresponding modal components. :

[0037]

[0038]

[0039]

[0040] The above formula describes the decomposition process of a certain component, where... Represents linear convolution. For the filter corresponding to this component, the range of values ​​for m is: , where T is the filter order plus 1. That is, first process the original SOH sequence. Forward pass through filter to obtain Then After time reversal, passing the sample through the same filter again yields... Finally, Time reversal yields modal components without phase distortion. .

[0041] The above operation is performed on all K sub-bands, and the final SOH signal is decomposed into K modal components:

[0042]

[0043] in For the first Reconstruction residual at point, As a trend component, it reflects the long-term monotonic degradation trend of battery capacity; to The periodic component reflects the capacity regeneration and fluctuation characteristics at different frequency scales.

[0044] S2-2. Construct a sliding window dataset for each of the decomposed components, setting the window length to w. For the i-th component sequence of length L... Using a sliding window with a step size of 1, construct input-target pairs: for each starting position Construct a training sample, from which each component sequence is generated. Given samples, the input vector and the prediction target are:

[0045] ,

[0046] ,

[0047] in For continuous within the window One historical observation value, This represents the value at the next time step after the last element of the window. All samples constitute the sliding window dataset corresponding to this component: For input , Normalization is performed, using the global maximum absolute value of each component. Normalization factor The normalized input and target are: , This forms training samples.

[0048] S3. Construct the HDCN network, which includes a trend prediction subnetwork and a period prediction subnetwork. The trend prediction subnetwork is a multi-layer LSTM network that integrates a context library and inter-layer gated attention residuals, while the period prediction subnetwork is a multi-layer TCN network that integrates a context library and inter-layer gated attention residuals.

[0049] S3-1. Constructing a trend and cycle prediction sub-network. The trend prediction sub-network uses a three-layer LSTM as the backbone, with gated attention residual modules and Dropout inserted between layers. After layer normalization, the output of the last layer is weighted and aggregated through a single attention layer to fuse the features of each time step into a window feature vector q. This avoids problems such as ignoring abnormal time steps and can dynamically amplify the weights of effective time steps. After context library lookup and fusion, the input to the prediction head outputs the predicted value, obtaining the predicted sequence of the trend component.

[0050] The periodic prediction subnetwork uses a three-layer TCN residual block as its backbone, with the inflation rate increasing exponentially by 1, 2, and 4 to expand the receptive field layer by layer. Causal convolutions are padded on the left to ensure that future information is not leaked. Gated attention residual modules and Dropout are also inserted between layers. The subsequent attention aggregation, context query fusion, and prediction head structure are the same as those of the trend prediction subnetwork, resulting in the prediction sequence of the periodic components.

[0051] S3-2. To address the issue that existing sliding window models rely solely on recent observations within the window for each prediction, lacking the ability to remember degradation patterns over longer time spans, a context library module is constructed to store degradation features from historical windows, providing long-range contextual information across windows for the current prediction:

[0052] (1) Query: The context library maintains P d-dimensional feature vectors. Using the window feature vector q of the current window as the query feature, calculate the match between each entry and all entries in the database. Cosine similarity, via temperature parameters After softmax normalization and weighted summation, the context vector z is obtained:

[0053]

[0054] in This is the weight matrix.

[0055] (2) Update: During HDCN network training, the exponential moving average (EMA) is used to update the slots in the context library that are most similar to the current query feature q.

[0056]

[0057] in The attenuation coefficient is... The slot vector with the highest cosine similarity to the current feature in the context library is the selected historical feature for updating, and q is the query feature vector (the newly observed degradation pattern) obtained after encoding the current window. During the warm-up phase in the early training stage, the context library is not enabled to allow the basic network to establish stable representations first.

[0058] S3-3. Construct a gated attention residual (GatedAttnRes) module, which is inserted between adjacent layers of the backbone network to capture the global dependencies between time steps within a window. Since different time steps contribute unevenly to the prediction during battery degradation—for example, recent capacity mutations or periodic regeneration locations often contain more critical degradation information—this module adaptively assigns different attention weights to each time step through a self-attention mechanism. Compared to fixed-weight methods such as average pooling, it can more effectively extract key degradation features.

[0059] Specifically, the features output by the LSTM module The input is fed into the gated attention residual module, and then... Layer normalization and multi-head self-attention computation are performed sequentially, with attention contributions controlled by a learnable gating scalar g:

[0060]

[0061] Where A represents the multi-head self-attention output. For the sigmoid function, The output of the gated attention residual module is passed to the Dropout layer. g is initialized to a negative value (in this embodiment, g = -3.0, σ(-3.0) ≈ 0.047) to ensure that the attention contribution is close to zero in the early stages of training, avoiding the introduction of meaningless attention perturbations before the backbone network has established effective representations. As training progresses, the gradient adaptively adjusts the value of g: if the attention information is beneficial to prediction, the value of g increases, increasing the contribution; otherwise, it remains low, achieving a smooth introduction of attention contribution without the need for additional warm-up strategies.

[0062] S4. After defining the training hyperparameters and network hyperparameters, the training samples for the trend component are input into the trend prediction subnetwork of the HDCN network for training; the training samples for the K-1 period components are input into the period prediction subnetwork of the HDCN network for training. The two subnetworks are trained independently, both using the MSE loss function, Adam optimizer, adaptive learning rate descent, gradient pruning, and early stopping mechanism. The context library uses an interval update strategy to reduce computational overhead.

[0063] S5. Input the test set samples into the trained HDCN network for rolling prediction to obtain the prediction sequence of each component. Then, add the prediction sequences of all components together to reconstruct the battery SOH trajectory prediction result.

[0064] S5-1. Perform rolling single-step prediction for each component: Using the most recent w known values ​​as the input window, predict the value at the next time step, append the predicted value to the end of the sequence, slide the window forward one step, and repeat the above process until the target length is predicted. Multiply the prediction result by the normalization factor of the corresponding component. Restored to the original scale.

[0065] S5-2. Add the predicted sequences of each component point by point to reconstruct the complete SOH trajectory prediction result. .

[0066] With the above structure, the method can effectively decouple the trend and periodic components by using peak-guided Fourier mode decomposition, and capture the long-term degradation trend and multi-scale periodic fluctuations of the SOH sequence through the dual-branch heterogeneous prediction subnetwork of the HDCN network, thereby achieving high-precision SOH trajectory prediction.

[0067] In another aspect, the present invention also provides a decommissioned battery SOH trajectory prediction system based on HDCN network, and a decommissioned battery SOH trajectory prediction method, comprising a battery SOH data module, a training sample module, an HDCN network prediction module, and an SOH trajectory prediction output module:

[0068] The battery SOH data module is used to collect time data of retired batteries, extract the constant current charging stage operation data, perform preprocessing to calculate the SOH curve, and obtain the battery SOH dataset.

[0069] The training sample module constructs a sliding window dataset based on the battery SOH dataset, using historical data within the window as input and the value at the next moment after the last position of the window as the target, to construct training samples.

[0070] The HDCN network prediction module is used to construct an HDCN network including a trend prediction subnetwork and a period prediction subnetwork, and to extract features from the training samples to obtain the prediction sequence of each component.

[0071] The SOH trajectory prediction output module is used to input test set samples into the trained HDCN network for rolling prediction, obtain the prediction sequence of each component, and add the prediction sequences of all components to reconstruct the battery SOH trajectory prediction result.

[0072] Example:

[0073] The sliding window datasets of the trend component and the periodic component are used as model inputs; the true values ​​of the next time step after the last window of each component are used as model outputs; the trend prediction subnetwork and the periodic prediction subnetwork are trained independently, both using mean squared error (MSE) as the loss function, and the training process is stabilized through adaptive learning rate descent, gradient pruning, and early stopping mechanisms; the model performance is verified using test set data, and the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) between the reconstructed prediction results and the true SOH are calculated to evaluate the overall SOH trajectory prediction accuracy.

[0074] The network performance was verified using the input test set data. The relevant model parameters and test results are shown in Table 1 below.

[0075] Table 1

[0076]

[0077] Among them, when the number of modalities K=3, the HDCN model achieved the best performance on the test set, with MAE, RMSE and MAPE of 0.70%, 1.06% and 1.18% respectively. All three indicators are significantly better than the configurations of K=2 and K=4.

[0078] When the number of modes increased from K=2 to K=3, MAE decreased significantly from 2.09% to 0.70% (a decrease of approximately 66.5%), RMSE decreased from 2.63% to 1.06% (a decrease of approximately 59.7%), and MAPE decreased from 3.16% to 1.18% (a decrease of approximately 62.7%). These results indicate that when K=2, FMD can only decompose the SOH sequence into one trend component and one periodic component. This coarse decomposition granularity leads to the superposition of fluctuation features at different frequency scales, such as capacity regeneration, within a single periodic component, making it difficult for the prediction network to fully learn its differential degradation patterns. When K increases to 3, FMD further subdivides the periodic component into two components of different frequency bands, achieving a more reasonable decoupling of trend and multi-scale periodic features, and significantly improving prediction accuracy.

[0079] When the number of modes was further increased from K=3 to K=4, the MAE increased from 0.70% to 1.51%, the RMSE increased from 1.06% to 1.79%, and the MAPE increased from 1.18% to 2.06%. This indicates that excessive decomposition introduces a redundant high-frequency modal component with a low signal-to-noise ratio and limited effective degradation information, which increases the fitting difficulty of the prediction network and interferes with the overall prediction performance. The above experimental results verify the rationality of the number of modes K=3 in this method, which can achieve the optimal balance between fully capturing capacity regeneration features and maintaining trend smoothness.

[0080]

[0081]

[0082]

[0083] in, This represents the true SOH value. This represents the estimated SOH value. This represents the total number of samples, i.e., the battery charge-discharge cycle times. MAE and RMSE measure the absolute magnitude and variance level of the prediction error, respectively, while MAPE reflects the magnitude of the relative error of the prediction result.

Claims

1. A method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks, characterized in that, Includes the following steps: S1. Collect time data of retired batteries and extract the constant current charging stage operation data, perform preprocessing to calculate the SOH curve, and obtain the battery SOH dataset; S2. Based on the battery SOH dataset, construct a sliding window dataset, using historical data within the window as input and the value of the next moment after the last position of the window as the target, to construct training samples; S3. Construct an HDCN network including a trend prediction sub-network and a period prediction sub-network. Extract features from the training samples to obtain the prediction sequences of each component. The specific implementation process is as follows: S3-1. Construct a trend prediction subnetwork and a cycle prediction subnetwork; The trend prediction subnetwork predicts the trend components and uses a Long Short-Term Memory (LSTM) network as the backbone, with gated attention residual modules and Dropout inserted between layers. After the output of the last layer is normalized, the features of each time step are fused into a window feature vector q through attention weighted aggregation. After fusion by querying the context library, the q is input into the prediction head to output the predicted value, thus obtaining the predicted sequence of the trend components. The periodic prediction subnetwork predicts the output of the periodic components. It uses a three-layer TCN residual block as the backbone, and the expansion rate increases exponentially layer by layer to expand the receptive field. The causal convolution is padded on the left to ensure that no future information is leaked. Gated attention residual modules and Dropout are inserted between layers. The subsequent attention aggregation, context query fusion and prediction head structure are the same as those of the trend prediction subnetwork to obtain the prediction sequence of the periodic components. The construction process of the context library is as follows: A context library is built to store degradation features from historical windows, providing long-range contextual information across windows for current predictions. Query: The context library maintains P d-dimensional feature vectors. Use the current window's feature vector q as the query feature, and calculate the match between each feature and all entries in the database. Cosine similarity, via temperature parameters After softmax normalization, a weighted sum is obtained to get the context vector z; Update: During HDCN network training, an exponential moving average is used to update the slot in the context library that is most similar to the current query feature q. ,in The attenuation coefficient is... q is the slot vector in the context library with the highest cosine similarity to the current feature, i.e., the historical feature selected for updating, and q is the query feature vector obtained after encoding the current window; S3-2, The gated attention residual module adaptively assigns different attention weights to each time step through a self-attention mechanism: The feature X output from the LSTM module is input into the gated attention residual module, and layer normalization and multi-head self-attention calculation are performed on X sequentially. The attention contribution is controlled by a learnable gated scalar g initialized to a negative value, resulting in the output of the gated attention residual module. This is passed to the Dropout layer; As training progresses, the gradient adaptively adjusts the value of g; S4. Input the test set samples into the trained HDCN network for rolling prediction to obtain the prediction sequence of each component. Then, add the prediction sequences of all components together to reconstruct the battery SOH trajectory prediction result.

2. The method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks according to claim 1, characterized in that, The specific implementation of step S1 is as follows: collecting voltage, current, capacity and temperature data during the constant current charging stage of retired batteries, preprocessing them, calculating the health status SOH curve using capacity increment analysis, and then standardizing the SOH curve to obtain the battery SOH dataset.

3. The method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks according to claim 2, characterized in that, Step S2 is implemented as follows: S2-1. Peak-guided Fourier mode decomposition (PFMD) is used to adaptively decompose the SOH sequence in the battery SOH dataset, separating a trend component and K-1 periodic components reflecting capacity regeneration fluctuations. S2-2. Construct a sliding window dataset for each of the decomposed components, setting the window length to w. For the i-th component sequence of length L... Using a sliding window with a step size of 1, construct input-target pairs: for each starting position Construct a training sample, from which each component sequence is generated. Given samples, the input vector and the prediction target are... and ,in For continuous within the window One historical observation value, This is the value at the next moment after the last position in the window; All samples constitute the sliding window dataset corresponding to this component: ; For input , Normalization is performed, using the global maximum absolute value of each component. The normalization factor is used to normalize the input and target. and This forms training samples.

4. The method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks according to claim 3, characterized in that, The specific implementation of step S2-1 is as follows: Performing a discrete Fourier transform on a SOH sequence of length N yields its frequency domain representation. : Based on frequency domain representation Calculate the spectral amplitude; Secondly, based on all spectral amplitudes, the set M of local maxima of the spectral amplitudes is detected. The frequency band boundary is determined by the set M of local maxima, and the frequency band is adaptively divided into K sub-frequency bands. FIR filters are designed for each of the K sub-bands: a low-pass filter for the lowest frequency band, a band-pass filter for the middle frequency band, and a high-pass filter for the highest frequency band. The coefficients of the three filters are generated differently: the low-pass and high-pass filters are designed with a single cutoff frequency, while the band-pass filter is designed with two cutoff frequencies. However, the filtering execution method is the same for all three: forward and backward zero-phase filtering is used, employing the same set of filter coefficients to perform forward and backward filtering on the signal, extracting the corresponding modal components. : The above operation is performed on all K sub-bands, and the final SOH signal is decomposed into K modal components: in For the first Reconstruction residual at point, As a trend component, it reflects the long-term monotonic degradation trend of battery capacity; to The periodic component reflects the capacity regeneration and fluctuation characteristics at different frequency scales.

5. The method for predicting the SOH trajectory of decommissioned batteries based on HDCN networks according to claim 4, characterized in that, The specific implementation process of step S4 is as follows: S4-1. Perform rolling single-step prediction for each component: Using the most recent w known values ​​as the input window, predict the value at the next time step, append the predicted value to the end of the sequence, slide the window forward one step, and repeat the above process until the target length is predicted; multiply the prediction result by the normalization factor of the corresponding component. Restored to the original scale; S4-2. Add the predicted sequences of each component point by point to reconstruct the complete SOH trajectory prediction result.

6. A decommissioned battery SOH trajectory prediction system based on HDCN network, used to implement the decommissioned battery SOH trajectory prediction method according to any one of claims 1 to 5, characterized in that, Includes the following modules: The battery SOH data module is used to collect time data of retired batteries, extract the constant current charging stage operation data, perform preprocessing to calculate the SOH curve, and obtain the battery SOH dataset. The training sample module, based on the battery SOH dataset, constructs a sliding window dataset, using historical data within the window as input and the value of the next moment after the last position of the window as the target, to construct training samples; The HDCN network prediction module is used to construct an HDCN network that includes a trend prediction subnetwork and a period prediction subnetwork. It extracts features from the training samples to obtain the prediction sequence of each component. The SOH trajectory prediction output module is used to input test set samples into the trained HDCN network for rolling prediction, obtain the prediction sequence of each component, and add the prediction sequences of all components to reconstruct the battery SOH trajectory prediction result.

Images