Aerial material demand prediction method and system based on cooperation of sequence decomposition and double-feature channel

Abstract

The invention relates to the technical field of aerial material management, and discloses a sequence decomposition and double-feature channel collaborative aerial material demand prediction method and system. The method comprises the following steps: pre-processing aeronautical material historical data and exogenous factors, and performing scale unification by introducing a normalization method of gating weight; adaptively decomposing the normalized sequence into a trend component and a peak component by using a filter capable of learning parameters; based on the trend component, capturing trend prediction features through a bidirectional long-short-term memory network in combination with an external attention mechanism; based on the peak component, capturing peak prediction features through a time-delay attention mechanism based on fast Fourier transform; and carrying out adaptive weighted fusion on the trend and peak prediction features, and carrying out reverse normalization to recover to an original scale to obtain a demand prediction value. According to the method, the problems of tendency and fluctuation feature aliasing, poor small sample learning ability and insufficient prediction stability in the aerial material demand sequence are solved, and the accuracy and robustness of high-value and low-frequency aerial material demand prediction are improved.

Description

Technical Field

[0001] This application relates to the field of aircraft material management technology, and in particular to an aircraft material demand forecasting method and system that combines sequence decomposition and dual-feature channel collaboration. Background Technology

[0002] Forecasting demand for aircraft parts is a core element for airlines and aircraft maintenance (MRO) companies to ensure operational safety and optimize inventory costs. Currently, the industry's aircraft parts supply systems generally rely on statistical forecasting methods based on historical consumption data or simple time series models.

[0003] Existing forecasting techniques can be broadly categorized as follows: First, there are traditional statistical models, such as ARIMA, exponential smoothing, and grey forecasting. These methods effectively characterize linear trends in time series, but their inherent limitation lies in their slow response to sudden fluctuations commonly seen in aircraft material demand, leading to a significant decrease in forecast accuracy. Second, there are machine learning models, such as random forests and XGBoost. They outperform traditional statistical models in handling nonlinear relationships; however, their model architecture inherently lacks the ability to explicitly model time-series dependencies, making it difficult to capture the dynamic patterns of demand over time. Third, there are deep learning models, represented by LSTM and Transformer. While they show potential in capturing complex time dependencies, existing architectures are mostly single-channel learning modes, making it difficult to efficiently handle the complex characteristics of long-term trends, seasonal cycles, short-term peaks, and abnormal fluctuations coexisting in aircraft material demand sequences within the same model.

[0004] Although some improvement methods attempt to introduce frequency domain transformation or attention mechanisms to improve performance, they generally suffer from the following common problems: Poor interpretability: It is impossible to clearly distinguish and decouple the trend components and random perturbation components in the sequence in terms of structure, resulting in an opaque decision-making process.

[0005] Feature aliasing interference is severe: dynamic features at different time scales are mixed and learned in a unified representation space, resulting in mutual interference and gradient competition, which leads to a decrease in prediction stability.

[0006] Insufficient utilization of external factors: Aircraft material demand is closely related to multi-dimensional external factors such as flight plans, component lifespan, inventory strategies, and time semantics (such as months and flight seasons). However, most existing methods are limited to historical consumption data itself and lack the ability to perform in-depth joint modeling of external features.

[0007] Poor adaptability to the characteristics of aviation material data: High-value turnover parts in civil aviation are characterized by low-frequency demand, sparse samples, a large number of zero-value samples, and high volatility. Existing methods have severely insufficient generalization ability under such small sample and non-stationary data conditions, and are prone to learning failure or overfitting.

[0008] Therefore, existing technologies lack a dedicated structured method that can effectively decompose trend and peak components based on the unique data characteristics required for aerospace materials, and fully utilize internal and external information for dual-channel collaborative and accurate prediction. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, this application provides a method and system for predicting aircraft material demand using sequence decomposition and dual feature channels. This method aims to solve the problems of trend and peak feature overlap, insufficient ability to model the temporal structure of aircraft material demand characteristics, inability to fully utilize external influencing factors and temporal semantic features, and low predictive interpretability in existing methods. It also aims to improve the modeling ability and generalization performance under sparse and zero sample conditions.

[0010] In a first aspect, this application provides a method for predicting aircraft material demand using sequence decomposition and dual-feature channel synergy, the method comprising: Step 1: Obtain historical time series data and exogenous influencing factor data related to aircraft material demand, and perform preprocessing; Step 2: Normalize the preprocessed data. The normalization process introduces adaptive gating weights dynamically generated based on time features and exogenous variables. Channel-level adaptive normalization is performed on the preprocessed data to obtain a normalized sequence. Step 3: Decompose the normalized time series into trend components and peak components. The trend components represent long-term smooth change patterns, while the peak components represent short-term fluctuations and sudden demand. Step 4: Input the trend component into the trend feature extraction module, which is processed sequentially by the bidirectional long short-term memory network and the multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features. Step 5: Input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features. Step 6: Input the trend prediction features and peak prediction features into the fusion module, generate normalized prediction values ​​through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

[0011] In conjunction with the first aspect, in the first implementation of the first aspect of this application, step 2 includes: Calculate the mean and standard deviation of the input data along the channel dimension; Perform standard normalization on the preprocessed data to obtain preliminary normalized data; The mean, standard deviation, time feature embeddings, and exogenous variable embeddings are concatenated to form the input features of the gating network; The input features of the gating network are fed into the activation function to generate adaptive gating weights; Adaptive gating weights are used to modulate the initially normalized data to obtain a normalized sequence.

[0012] In conjunction with the first aspect, in the second implementation of the first aspect of this application, in step 3, the normalized sequence is decomposed using a learnable Savitzky-Golay smoothing filter. The window width wi of the Savitzky-Golay smoothing filter is used as a trainable parameter and is adaptively adjusted through backpropagation to achieve optimal trend-peak separation of different aviation material sequences and obtain the trend component and peak component.

[0013] In conjunction with the first aspect, in the third implementation of the first aspect of this application, step 4 includes: The trend component is input into a two-layer bidirectional long short-term memory network, and then the output of the long short-term memory network is added to the input residual and normalized by the layer to obtain the first deep temporal feature. The first deep temporal features are fed into a module consisting of N stacked external attention encoding layers with identical structures. Each external attention encoding layer performs the following operations: Using the input features of this layer as the query vector, and using the first global memory matrix and the second global memory matrix, which are independent of the input sequence, as the key vector and value vector respectively, the external attention features are calculated. The external attention features are added to the input features of this layer for the first time, and then normalized by the layer to obtain intermediate features. The intermediate features are then input into the feedforward network. The output of the feedforward network is added to the intermediate features for a second time using residuals, and then normalized by the layer to obtain the output features of this layer. The output features of the Nth external attention encoding layer are linearly mapped to obtain the trend prediction features.

[0014] In conjunction with the first aspect, in the fourth implementation of the first aspect of this application, the global memory matrix of the multi-layer external attention encoding unit is updated in real time according to the sample loss gradient during the training process, so that the trend extraction continuously adapts to the periodic drift of the distribution of aircraft material demand.

[0015] In conjunction with the first aspect, in the fifth implementation of the first aspect of this application, step 5 includes: Perform a linear mapping on the peak components to obtain the query matrix, key matrix, and value matrix; The cross-correlation spectrum of the query matrix and the key matrix is ​​calculated in the frequency domain to capture the correlation at different time lag positions; Top-k amplitude screening is performed on the cross-correlation spectrum to retain the k largest lag positions, and the sparse lag weights are obtained by inverse transformation back to the time domain. Peak prediction features are obtained by using sparse lag weights to perform weighted summation on the value matrix.

[0016] In conjunction with the first aspect, in the sixth implementation of the first aspect of this application, the k value of Top-k is dynamically adjusted according to the sparsity of the aircraft material samples. When the number of training samples is lower than a preset threshold, the k value decreases to reduce the risk of overfitting. When the number of samples is higher than the preset threshold, the k value increases to improve the ability to capture fluctuations.

[0017] In conjunction with the first aspect, in the seventh implementation of the first aspect of this application, in step 6, the gating weighting coefficient is automatically determined by the peak intensity. First, the energy proportion of the peak prediction feature is calculated, and then the trend prediction feature and the peak prediction feature are linearly weighted using the energy proportion as the weight.

[0018] In conjunction with the first aspect, in the eighth implementation of the first aspect of this application, in step 6, the mean and standard deviation generated in step 2 are called to perform inverse normalization to obtain the original scale of the predicted value of aircraft material demand.

[0019] Secondly, this application provides an aircraft material demand prediction system based on sequence decomposition and dual-feature channel collaboration, the system comprising: The data acquisition module is used to acquire historical time-series data and exogenous influencing factor data related to aircraft material demand, and to perform preprocessing. The normalization module is used to normalize the preprocessed data. The normalization process introduces adaptive gating weights that are dynamically generated based on time features and exogenous variables. Channel-level adaptive normalization is performed on the preprocessed data to obtain a normalized sequence. The sequence decomposition module is used to decompose the normalized time series into trend components and peak components. The trend components represent long-term smooth change patterns, while the peak components represent short-term fluctuations and sudden demand. The trend feature extraction module is used to input trend components into the trend feature extraction module, which is processed sequentially by a bidirectional long short-term memory network and a multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features. The peak feature extraction module is used to input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features. The prediction fusion module is used to input trend prediction features and peak prediction features into the fusion module, generate normalized prediction values ​​through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

[0020] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. Reversible Gated Normalization Preprocessing Mechanism for Aircraft Material Time Series Data: The proposed reversible gated normalization method can adaptively adjust the normalization parameters according to time characteristics and external variables, achieving dynamic standardization processing of data from different aircraft materials and time periods at a unified scale. This mechanism can maintain stable feature representation even with limited or unevenly distributed data, providing a unified and reliable input foundation for subsequent model training.

[0021] 2. Sequence Decomposition Strategy for Separating Trends and Disturbances: By combining smoothing filtering with sliding window differencing, the aircraft material demand sequence is decomposed into trend and disturbance components, thereby achieving structural decoupling between long-term patterns and short-term fluctuations. This strategy effectively reduces interference from features at different time scales, enabling the model to maintain good learning stability even with multi-frequency features and small sample scenarios.

[0022] 3. Dual-feature module collaborative prediction architecture: Two independent prediction channels are set up for trend feature extraction and peak feature extraction, respectively learning long-term stable patterns and short-term fluctuation features, and the results are superimposed in the fusion layer. The dual-channel collaborative structure can effectively alleviate the problem of insufficient model learning under small sample conditions, and improve the overall prediction accuracy and robustness through task decomposition.

[0023] 4. Multi-layer time-dependent modeling algorithm for trend feature extraction: Captures the long-term dependence and periodicity features of the aircraft material demand sequence. Its structure includes a bidirectional long short-term memory network (Bi-LSTM) that enables the model to simultaneously utilize historical state and future information for contextual modeling; a multi-layer external attention encoding unit (MEAformer) introduces a multi-head attention structure to associate trend features with external input features (such as seasonality, aircraft type, flight frequency, etc.) and achieves feature aggregation across time steps under multi-layer residual connections.

[0024] 5. Time-Lag Dependency Attention Modeling Algorithm for Peak Feature Extraction: This algorithm is used to capture short-term fluctuations and sudden changes in the demand sequence of aviation materials. Its core algorithm is the Time Lag Attention (TLA) mechanism. The time-lag feature extraction layer obtains the local time-lag features of the input sequence through convolution and autocorrelation operations, and uses Fast Fourier Transform (FFT) to calculate the correlation under different time delays. The Top-k time-lag filtering layer selects several representative time-lag points based on the correlation strength, retaining the main perturbation dependencies and removing noise features. This algorithm has a strong ability to capture fluctuations and performs particularly stably in scenarios with sparse samples, frequent fluctuations, or many outliers, effectively compensating for the smoothness limitations of trend feature extraction. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 The flowchart is a process for predicting aircraft material demand using sequence decomposition and dual-feature channel synergy as described in this application. Figure 2 This is a schematic diagram of the structure of the aircraft material demand prediction system based on sequence decomposition and dual feature channel collaboration in this application. Detailed Implementation

[0027] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of the aircraft material demand prediction method based on sequence decomposition and dual-feature channel collaboration provided by this invention. The flowchart specifically includes the following steps: Step 1: Obtain historical time series data and exogenous influencing factor data related to aircraft material demand, and perform preprocessing.

[0029] Specifically, the historical time series data related to aircraft material demand is the historical aircraft material consumption data series, and the exogenous influencing factor data includes at least historical data such as flight operation plans, maintenance plans, seasonal factors, aircraft type information, flight hours, and fleet size.

[0030] To achieve aircraft material demand forecasting, historical consumption records, as well as data on exogenous influencing factors such as flight schedules, maintenance work orders, and component basic information, need to be collected. Missing values ​​in historical consumption records are imputed and outliers are smoothed to form a regular demand sequence. Flight schedule data is aggregated daily into flight hours; planned scheduled maintenance events are extracted from maintenance work order data and encoded as binary tags; and aircraft type and part number in component basic information are converted into low-dimensional dense vectors through an embedding layer. Timestamps are decomposed into periodic features such as month and weekday and encoded using sine and cosine. All the above heterogeneous data are left-joined and aligned based on a unified timestamp index to ensure the integrity and consistency of feature vectors at each time step.

[0031] Step 2: Normalize the preprocessed data. The normalization process introduces adaptive gating weights dynamically generated based on time features and exogenous variables. Channel-level adaptive normalization is performed on the preprocessed data to obtain a normalized sequence.

[0032] In one specific embodiment, the process of performing step 2 may specifically include the following steps: Calculate the mean and standard deviation of the input data along the channel dimension; Perform standard normalization on the preprocessed data to obtain preliminary normalized data; The mean, standard deviation, time feature embeddings, and exogenous variable embeddings are concatenated to form the input features of the gating network; The input features of the gating network are fed into the activation function to generate adaptive gating weights; Adaptive gating weights are used to modulate the initially normalized data to obtain a normalized sequence.

[0033] Specifically, in the task of predicting aircraft material demand, the preprocessed data constitutes a tensor U with dimensions [B, T, C], where B represents the batch size, T represents the length of the historical time window, and C represents the total number of feature channels. This tensor U contains various heterogeneous data, including historical aircraft material consumption, flight hours, maintenance event markers, and coded time-period features. These features have different physical meanings and numerical dimensions. Directly inputting them into the model would lead to features with large numerical ranges dominating the training process. At the same time, aircraft material consumption exhibits significantly different distribution characteristics under different seasons or different operational strategies, and a fixed normalization strategy is difficult to adapt to such dynamic changes.

[0034] To address the aforementioned challenges, a reversible normalization mechanism incorporating dynamic gating is implemented. The process begins by calculating statistics for the tensor U along the channel dimension. For each feature channel c of each sample in the batch, its mean μ_c and standard deviation σ_c over the entire time window T are calculated. Subsequently, a standard normalization operation is performed. For the value U_{b, t, c} of channel c at time step t, its initial normalization result is obtained using the formula V_{b, t, c} = (U_{b, t, c} - μ_c) / (σ_c + ε), where ε is a minimal constant introduced to ensure numerical stability. This operation transforms the data distribution of each feature channel into a form with an approximate mean of zero and a standard deviation of one, thus initially solving the problem of unifying the scale of data with different dimensions.

[0035] However, standard normalization fails to consider the impact of the external operating environment on data distribution. For example, during peak summer travel seasons, a surge in flight hours may lead to an upward shift in the baseline level of aircraft material consumption; or when a new aircraft model is introduced into the fleet, the wear pattern of specific components may change. To address the problem that traditional normalization cannot adapt to dynamic changes in data distribution, this method introduces a gated generative network. The input feature vector F of this network is composed of the aforementioned statistics μ_c and σ_c, the temporal feature embedding vector E_time, and the exogenous variable embedding vector E_exo. The temporal feature embedding E_time is derived from the sine and cosine encoded representation of information such as month and day of the week decomposed from the timestamp, which can capture seasonal patterns; the exogenous variable embedding E_exo integrates operational context information such as flight schedule intensity and fleet configuration. The concatenated feature vector F is mapped through a fully connected layer and then input into a sigmoid activation function to generate an adaptive gated weight vector G with a dimension matching the number of feature channels C, where each element g_c corresponds to the modulation weight of a feature channel, and its value ranges between zero and one.

[0036] The adaptive gating weight G is used to modulate the initially normalized data V. The modulation operation is achieved through channel-by-channel weighting, where each element of the normalized sequence H is calculated as H_{b, t, c} = g_c × V_{b, t, c}. This modulation process gives the model the ability to dynamically adjust the importance of each feature channel based on the time context and operational status. When the gating weight g_c is close to one, it indicates that the feature information content of that channel is rich under the current external conditions, and its normalization result is fully preserved; when g_c is close to zero, it may mean that the feature of that channel has weak discriminative power in the current context, and its influence is suppressed accordingly. For example, in months with intensive flight schedules, feature channels directly related to flight intensity may be assigned higher gating weights, while during scheduled maintenance cycles, the weight of maintenance event marker channels may be enhanced.

[0037] This normalization mechanism, which integrates data-driven statistics and context-aware modulation, not only achieves scale unification for multi-source heterogeneous data, but more importantly, it enables adaptive calibration of feature importance under different operational scenarios. By dynamically injecting external factors into the data preprocessing stage, it provides a cleaner and more context-aware feature representation for subsequent trend-peak decomposition, avoiding feature aliasing and gradient competition caused by treating features of different scales and importance equally in subsequent models. This lays a data foundation for improving the prediction accuracy and stability of the model in complex and ever-changing operational environments.

[0038] Step 3: Decompose the normalized time series into trend components and peak components. The trend components represent long-term smooth change patterns, while the peak components represent short-term fluctuations and sudden demand.

[0039] In one specific embodiment, in step 3, the normalized sequence is decomposed using a learnable Savitzky-Golay smoothing filter. The window width wi of the Savitzky-Golay smoothing filter is used as a trainable parameter and is adaptively adjusted through backpropagation to achieve optimal trend-peak separation for different aviation material sequences and obtain the trend component and peak component.

[0040] Specifically, after data normalization, the normalized sequence is input into the sequence decomposition module to separate the trend and disturbance components of the aircraft material demand signal. Through sliding window smoothing filtering, the long-term stable trend component is extracted, and the short-term fluctuations and disturbance components of sudden demand are obtained through differential processing. This decomposition process effectively distinguishes between long-term regular changes and short-term random fluctuations, providing independent input for the subsequent specialized learning of the feature extraction network.

[0041] Sequence decomposition employs a Savitzky-Golay filter, which smooths the signal by performing polynomial least-squares fitting within a local window. Unlike filters with fixed parameters, this scheme sets the filter's window half-width parameter wi as a trainable parameter, with an initial value set to a moderate value and adaptively adjusted through gradient backpropagation during model training. For an input sequence H of length L, at each time point t, the filter selects data points within a local window [t-wi, t+wi] of length (2wi+1) centered on the input sequence, fits the data points with a low-order polynomial, and uses the fitted value of this polynomial at point t as the output of the trend component T_t. This process slides across the entire sequence, generating a complete trend component sequence T. The peak component S is obtained by subtracting the trend component T from the original normalized sequence H, i.e., S = HT.

[0042] In the scenario of aircraft material demand forecasting, the consumption patterns of different components differ significantly. For example, some consumable parts may exhibit a slowly changing baseline demand superimposed with frequent small fluctuations, while high-value reusable parts may remain stable for a long time but experience sharp spikes after specific maintenance events. Filters with fixed window widths struggle to adapt to both patterns simultaneously: a window that is too small will retain too much high-frequency noise in the trend component, causing the peak component to contain a large number of non-sudden fluctuations; a window that is too large may over-smooth the data, submerging the true sudden demand in the trend component, causing the peak component to miss crucial information. By setting wi as a learnable parameter, the model can automatically find the optimal smoothness based on the inherent characteristics of the input sequence. During training, the gradient signal guides the wi value to update in the direction that minimizes the overall prediction loss. For sequences with frequent fluctuations, training may converge to a smaller wi value to preserve finer trend changes; for sequences with long-term stability and occasional spikes, training may tend towards a larger wi value to extract a smoother baseline and highlight significant sudden peaks.

[0043] Step 4: Input the trend component into the trend feature extraction module, which is processed sequentially by the bidirectional long short-term memory network and the multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features.

[0044] In one specific embodiment, the process of performing step 4 may specifically include the following steps: The trend component is input into a two-layer bidirectional long short-term memory network, and then the output of the long short-term memory network is added to the input residual and normalized by the layer to obtain the first deep temporal feature. The first deep temporal features are fed into a module consisting of N stacked external attention encoding layers with identical structures. Each external attention encoding layer performs the following operations: Using the input features of this layer as the query vector, and using the first global memory matrix and the second global memory matrix, which are independent of the input sequence, as the key vector and value vector respectively, the external attention features are calculated. The external attention features are added to the input features of this layer for the first time, and then normalized by the layer to obtain intermediate features. The intermediate features are then input into the feedforward network. The output of the feedforward network is added to the intermediate features for a second time using residuals, and then normalized by the layer to obtain the output features of this layer. The output features of the Nth external attention encoding layer are linearly mapped to obtain the trend prediction features.

[0045] Specifically, the trend component is input into the trend feature extraction module, which includes a bidirectional long short-term memory network (Bi-LSTM) and a multi-layer external attention encoding unit (MEAformer) to capture the long-term dependence characteristics and seasonal variation patterns of aircraft material demand. During the prediction process, the model employs a sliding window strategy for segment-by-segment prediction to ensure computational stability and result continuity when dealing with long-sequence inputs.

[0046] The component first enters a two-layer bidirectional long short-term memory network. The network scans the sequence in both forward and backward directions, selectively retaining and transmitting information through its internal gating mechanism, thereby constructing hidden states containing complete contextual information. The output of this network is residually connected to the original trend component, and the result is processed by layer normalization to stabilize and train dynamics, forming the first deep temporal feature.

[0047] The first deep temporal feature is then fed into a module consisting of N stacked external attention encoding layers. In each encoding layer, the processing begins with the computation of external attention. The input features of this layer are linearly projected into a query vector Q, while the key vector K and value vector V are derived from two trainable global memory matrices M_K and M_V, independent of the input sequence. The attention weights are calculated by comparing the query vector Q with the memory key M_K, as shown in the formula: Where d_k is the dimension of the query vector. The weighted summation operation is performed using the weight matrix A and the memory value M_V to generate external attention features, calculated as follows: This process enables the model to match the current sequence segment with a global trend pattern library learned from the entire training data, rather than focusing solely on local correlations within the sequence.

[0048] Subsequently, the external attention features are added to the input features of this coding layer using a first residual sum, and the output is normalized by layers to produce intermediate features. These intermediate features are fed into a feedforward network, typically composed of two linear transformations and an activation function in between, to introduce nonlinear transformation capabilities. The output of the feedforward network is then added to the intermediate features again using a second residual sum, and the result is normalized through another layer of normalization to finally form the output features of this coding layer. This process is performed sequentially across N stacked layers, with the output of each layer serving as the input to the next, progressively refining and enhancing the trend representation of the sequence through this hierarchical processing.

[0049] The output features of the Nth external attention encoding layer are mapped to the target dimension via a linear projection layer to generate trend prediction features. In this architecture, a bidirectional long short-term memory network is responsible for capturing medium-term cyclical patterns with strong local dependencies in the sequence, while the multi-layer external attention mechanism gives the model the ability to access and fuse global memory, enabling it to identify and model macro-trends and seasonal patterns that span long time spans and are difficult to detect in local sequences. The global memory matrices M_K and M_V are continuously updated during training, thereby encapsulating various typical long-term demand patterns in the dataset, such as the baseline consumption patterns of different aircraft types under different flight seasons.

[0050] This design directly addresses the problem of "lack of time-series structure modeling capability for the characteristics of aircraft material demand". By combining local sequence modeling with global pattern matching, it effectively overcomes the limitations of traditional single-channel models or models that rely solely on internal attention in capturing long-term, stable trends in aircraft material demand. It avoids the model ignoring macro-level patterns due to excessive focus on short-term fluctuations, thus laying the foundation for generating stable and accurate trend predictions.

[0051] In a preferred embodiment, the global memory matrix of the multi-layer external attention encoding unit is updated in real time according to the sample loss gradient during training, so that the trend extraction continuously adapts to the periodic drift of the distribution of aircraft material demand.

[0052] Specifically, in the context of aircraft material support, demand distribution changes slowly due to fleet expansion, seasonal changes, or adjustments to maintenance policies—a phenomenon known as periodic drift. A global memory matrix encapsulates typical long-term trend patterns in the training data. By updating these matrices in real-time based on sample loss gradients, the model can dynamically adjust its internal "experience base." When new training samples reflect changes in demand distribution, the gradient signal drives the memory matrix to fine-tune its stored pattern representations, thereby evolving the prediction logic of the trend feature extraction module rather than adhering to historical patterns from the initial training phase. This mechanism endows the model with the ability to adapt to long-term changes in the operational environment, effectively mitigating the performance degradation problem caused by distribution drift in static models.

[0053] Step 5: Input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features.

[0054] In one specific embodiment, the process of performing step 5 may specifically include the following steps: Perform a linear mapping on the peak components to obtain the query matrix, key matrix, and value matrix; The cross-correlation spectrum of the query matrix and the key matrix is ​​calculated in the frequency domain to capture the correlation at different time lag positions; Top-k amplitude screening is performed on the cross-correlation spectrum to retain the k largest lag positions, and the sparse lag weights are obtained by inverse transformation back to the time domain. Peak prediction features are obtained by using sparse lag weights to perform weighted summation on the value matrix.

[0055] Specifically, the peak component is input to the peak feature extraction module, which is performed simultaneously with the trend feature extraction and prediction. It adopts the time lag attention mechanism (TLA), which obtains the local time lag features of the input sequence through convolution and autocorrelation operations, and uses fast Fourier transform to identify rapid change features such as short-term bursts and peak fluctuations in the aviation material demand sequence, thereby enhancing the model's ability to respond to abnormal fluctuations.

[0056] The component first passes through three independent linear projection layers, which are mapped to a query matrix Q_p, a key matrix K_p, and a value matrix V_p, respectively. These matrices will be used for subsequent time-delay correlation analysis.

[0057] To efficiently capture the dynamic correlation at different time lag positions in the sequence, the autocorrelation function is approximated in the frequency domain. Specifically, a Fast Fourier Transform (FFT) is applied to Q_p and K_p respectively, transforming them back to the frequency domain to obtain F_Q and F_K. Then, a pointwise multiplication operation is performed on these two in the frequency domain, i.e., R_freq = F_Q ⊙ F_K, where ⊙ represents the pointwise product. This operation in the frequency domain is equivalent to cyclic cross-correlation in the time domain. An Inverse Fast Fourier Transform (IFFT) is then performed on R_freq, transforming it back to the time domain to obtain a time-lag correlation matrix R_lag, where each element... This characterizes the query sequence and key sequence in terms of lag step size. The correlation strength.

[0058] Considering that sudden demand for aerospace materials is often only significantly correlated with a limited number of critical historical moments, and that the peak components contain a large amount of noise, directly using the complete R_lag would lead to distraction. Therefore, a Top-k amplitude filtering operation is performed on R_lag, retaining only the k lag positions with the largest correlation amplitudes. The correlations at the remaining positions are set to negative infinity. This sparsified correlation matrix is ​​then normalized using the Softmax function to produce a sparse temporal attention weight matrix α_sparse, which has significant weights only at the selected k lag positions.

[0059] The sparse lag weight α_sparse is used to perform a weighted summation of the value matrix to generate peak prediction features. The weighting process can be implemented in two ways: one is to treat α_sparse as attention weights and directly perform matrix multiplication with the value matrix; the other is to consider the time lag characteristics, roll the value matrix along the time axis according to the selected lag position, and then aggregate it using α_sparse as the weight. , where Roll(·) represents a cyclic shift operation. This mechanism, based on frequency domain transformation and sparse filtering, enables the model to focus on a few of the most indicative historical moments from potentially noisy peak components, such as a specific number of days after a particular maintenance event, thereby accurately capturing the short-term temporal patterns that lead to the current sudden demand.

[0060] This design directly addresses the problems of "ineffective decoupling of trend and fluctuation features" and "sparse sample distribution, significant zero-sample phenomenon, and insufficient model generalization ability." Through time-delay attention and Top-k screening, the peak feature extraction module can resist noise interference and focus on learning sparse but crucial sudden dependencies. This avoids the tendency of traditional models to misclassify short-term peaks as long-term trends or random noise due to feature aliasing when processing high-fluctuation sequences, thus improving the model's recognition accuracy and generalization ability for sudden demands in typical small-sample, high-fluctuation scenarios of aviation materials.

[0061] In a preferred embodiment, the k value of Top-k is dynamically adjusted according to the sparsity of the aircraft material samples. When the number of training samples is lower than a preset threshold, the k value is reduced to reduce the risk of overfitting. When the number of samples is higher than the preset threshold, the k value is increased to improve the ability to capture fluctuations.

[0062] Specifically, in aircraft material demand forecasting, sample sparsity manifests as a shortage of historical consumption records for specific components and a large number of zero-value samples. The k value in the Top-k mechanism, acting as the lag window size, is dynamically adjusted to balance overfitting control and feature capture capability. When the training sample size N is below a preset threshold, it indicates weak data support for the component. Maintaining a large k value forces the model to learn too many lag dependencies in a limited number of samples, leading to overfitting due to memorizing noisy patterns. Reducing k to k_low at this point essentially constrains the model to focus only on the most significant few lag relationships, simplifying the learning task and improving generalization. Conversely, when N exceeds the preset threshold, sufficient data provides statistical support for complex patterns. Increasing k to k_high allows the model to explore a wider range of short-term dependencies, thereby enhancing the granularity of capturing high-frequency fluctuations. This adaptive mechanism, through the negative correlation between k value and data volume, directly addresses the risk of overfitting under small sample conditions and the problem of insufficient feature extraction in high-fluctuation scenarios, ensuring that the peak feature extraction module maintains optimal discriminative capability under different data densities.

[0063] Step 6: Input the trend prediction features and peak prediction features into the fusion module, generate normalized prediction values ​​through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

[0064] In one specific embodiment, in step 6, the gating weighting coefficient is automatically determined by the peak intensity. First, the energy proportion of the peak prediction feature is calculated, and then the trend prediction feature and the peak prediction feature are linearly weighted using the energy proportion as the weight.

[0065] In one specific embodiment, in step 6, the mean and standard deviation generated in step 2 are called to perform inverse normalization to obtain the original scale of the aircraft material demand forecast.

[0066] Specifically, the prediction results from the trend feature extraction module and the peak feature extraction module are input into the fusion module. The fusion module uses a fixed weighting method to linearly superimpose the two sets of prediction results to achieve a comprehensive output of the trend prediction result and the disturbance prediction result.

[0067] The fusion process employs a gated weighting mechanism, the core of which lies in automatically determining the mixing ratio of the two types of features. The gated weighting coefficients are not fixed values ​​but are dynamically generated based on the intensity of the peak prediction feature. Specifically, the energy proportion of the peak prediction feature in the joint feature space is first calculated. This calculation is achieved by taking the ratio of the norm of the peak prediction feature vector to the sum of the norms of the trend prediction feature vector, expressed as γ = ||F_spike|| / (||F_trend|| + ||F_spike||), where F_trend represents the trend prediction feature and F_spike represents the peak prediction feature. This ratio γ reflects the significance of sudden fluctuations at the current prediction time relative to the overall prediction pattern. Subsequently, the two types of features are linearly weighted using this energy proportion γ as the weight to generate a normalized prediction value. Its calculation formula is Where W_t and W_s are trainable linear projection matrices used to map features to a unified prediction space. This adaptive fusion mechanism based on peak intensity ensures that the model output relies more on the reliable baseline provided by the trend feature extraction module during the stationary phase of the sequence, while significantly increasing the weight of the peak feature extraction module when there are sharp fluctuations or sudden peaks, thereby dynamically balancing the influence of long-term patterns and short-term disturbances.

[0068] Normalized predicted values The data is still in a normalized space with a mean of zero and a standard deviation of one, and needs to be transformed back to the original data scale with actual physical meaning. This transformation uses the mean vector and standard deviation vector calculated and stored during the reversible gated normalization in step 2. The denormalization process is performed independently for each feature channel. For the target aerospace material demand prediction channel, the final predicted value is... Through formula The calculations show that σ_target and μ_target represent the standard deviation and mean of the target channel, respectively. This operation restores the model output to a numerical range consistent with historical consumption records, generating a final demand forecast that can be directly used for aircraft material procurement and inventory decisions.

[0069] This fusion and inverse transformation process addresses the issues of "ineffective decoupling of trend and fluctuation features" and "poor model interpretability." Adaptive gated weighting structurally acknowledges that the contributions of trend and peak components to total demand are variable at different times, and achieves dynamic fusion through a data-driven approach. This avoids the shortcomings of fixed-weight fusion, such as introducing unnecessary fluctuation noise during stable periods or insufficient response during periods of sudden changes. Simultaneously, it clearly maps the fusion results back to the original scale, giving the predicted values ​​clear business interpretability. The entire process forms a closed loop with the aforementioned steps, ensuring the continuity and consistency between feature collaboration in the normalized space and the final demand quantification in the physical space, completing the final transformation from multi-dimensional feature extraction to a single demand value output.

[0070] The above describes the aircraft material demand forecasting method based on sequence decomposition and dual-feature channel collaboration in the embodiments of this application. The following describes the aircraft material demand forecasting system based on sequence decomposition and dual-feature channel collaboration in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 The schematic diagram of the aircraft material demand prediction system based on sequence decomposition and dual feature channel collaboration provided in this application is shown. The system includes: The data acquisition module 10 is used to acquire historical time series data and exogenous influencing factor data related to the demand for aviation materials, and to perform preprocessing.

[0071] The normalization module 20 is used to normalize the preprocessed data. The normalization process introduces adaptive gating weights dynamically generated based on time features and exogenous variables to perform channel-level adaptive normalization on the preprocessed data, resulting in a normalized sequence.

[0072] The sequence decomposition module 30 is used to decompose the normalized time series into trend components and peak components, wherein the trend components represent long-term smooth change patterns, and the peak components represent short-term fluctuations and sudden demands.

[0073] The trend feature extraction module 40 is used to input the trend components into the trend feature extraction module, which is then processed sequentially by a bidirectional long short-term memory network and a multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features.

[0074] The peak feature extraction module 50 is used to input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features.

[0075] The prediction fusion module 60 is used to input the trend prediction features and peak prediction features into the fusion module, generate normalized prediction values ​​through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

[0076] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting aircraft material demand based on sequence decomposition and dual-feature channel synergy, characterized in that, The method includes: Step 1: Obtain historical time series data and exogenous influencing factor data related to aircraft material demand, and perform preprocessing; Step 2: Normalize the preprocessed data. The normalization process introduces adaptive gating weights dynamically generated based on time features and exogenous variables. Channel-level adaptive normalization is performed on the preprocessed data to obtain a normalized sequence. Step 3: Decompose the normalized time series into trend components and peak components, wherein the trend components represent long-term smooth change patterns, and the peak components represent short-term fluctuations and sudden demand. Step 4: Input the trend component into the trend feature extraction module, which is processed sequentially by a bidirectional long short-term memory network and a multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features. Step 5: Input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features. Step 6: Input the trend prediction feature and the peak prediction feature into the fusion module, generate a normalized prediction value through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

2. The method according to claim 1, characterized in that, Step 2 includes: Calculate the mean and standard deviation of the input data along the channel dimension; Perform standard normalization on the preprocessed data to obtain preliminary normalized data; The mean, standard deviation, time feature embedding, and exogenous variable embedding are concatenated to form the input features of the gating network. The input features of the gated network are input into the activation function to generate the adaptive gate weights; The adaptive gating weights are used to modulate the initial normalized data to obtain the normalized sequence.

3. The method according to claim 1, characterized in that, In step 3, the normalized sequence is decomposed using a learnable Savitzky-Golay smoothing filter. The window width wi of the Savitzky-Golay smoothing filter is used as a trainable parameter and is adaptively adjusted through backpropagation to achieve optimal trend-peak separation for different aviation material sequences, thereby obtaining the trend component and the peak component.

4. The method according to claim 1, characterized in that, Step 4 includes: The trend component is input into a two-layer bidirectional long short-term memory network, and then the output of the long short-term memory network is added to the input residual and normalized by the layer to obtain the first deep temporal feature. The first deep temporal features are fed into a module consisting of N stacked external attention encoding layers with identical structures, wherein each external attention encoding layer performs the following operations: Using the input features of this layer as the query vector, and using the first global memory matrix and the second global memory matrix, which are independent of the input sequence, as the key vector and value vector respectively, the external attention features are calculated. The external attention features are added to the input features of this layer for the first time, and then normalized by the layer to obtain intermediate features. The intermediate features are then input into the feedforward network. The output of the feedforward network is added to the intermediate features for a second time using residuals, and then normalized by the layer to obtain the output features of this layer. The trend prediction features are obtained by linearly mapping the output features of the Nth external attention encoding layer.

5. The method according to claim 4, characterized in that, The global memory matrix of the multi-layer external attention encoding unit is updated in real time according to the sample loss gradient during training, so that the trend extraction can continuously adapt to the periodic drift of the distribution of aircraft material demand.

6. The method according to claim 1, characterized in that, Step 5 includes: Perform a linear mapping on the peak components to obtain the query matrix, key matrix, and value matrix; The cross-correlation spectrum of the query matrix and the key matrix is ​​calculated in the frequency domain to capture the correlation at different time lag positions; The cross-correlation spectrum is subjected to Top-k amplitude filtering, and the k largest lag positions are retained. The sparse lag weights are obtained by inverse transformation back to the time domain. The peak prediction features are obtained by weighting and summing the value matrix using the sparse lag weights.

7. The method according to claim 6, characterized in that, The k value of Top-k is dynamically adjusted according to the sparsity of the aircraft material samples. When the number of training samples is lower than the preset threshold, the k value is reduced to reduce the risk of overfitting. When the number of samples is higher than the preset threshold, the k value is increased to improve the ability to capture fluctuations.

8. The method according to claim 1, characterized in that, In step 6, the gating weighting coefficient is automatically determined by the peak intensity. First, the energy proportion of the peak prediction feature is calculated, and then the energy proportion is used as the weight to linearly weight the trend prediction feature and the peak prediction feature.

9. The method according to claim 2, characterized in that, In step 6, the mean and standard deviation generated in step 2 are called to perform inverse normalization to obtain the original scale of the aircraft material demand forecast.

10. A system for predicting aircraft material demand using sequence decomposition and dual-feature channel coordination, used to implement the method as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition module is used to acquire historical time-series data and exogenous influencing factor data related to aircraft material demand, and to perform preprocessing. The normalization module is used to normalize the preprocessed data. The normalization process introduces adaptive gating weights dynamically generated based on time features and exogenous variables to perform channel-level adaptive normalization on the preprocessed data to obtain a normalized sequence. The sequence decomposition module is used to decompose the normalized time series into trend components and peak components, wherein the trend components represent long-term smooth change patterns, and the peak components represent short-term fluctuations and sudden demands. The trend feature extraction module is used to input the trend components into the trend feature extraction module, which is processed sequentially by a bidirectional long short-term memory network and a multi-layer external attention encoding unit to capture long-term dependencies and obtain trend prediction features. The peak feature extraction module is used to input the peak component into the peak feature extraction module, and the time-delay attention module based on fast Fourier transform performs Top-k sparse lag selection to capture short-term sudden fluctuations and output peak prediction features. The prediction fusion module is used to input the trend prediction features and the peak prediction features into the fusion module, generate a normalized prediction value through gated weighted fusion, and then restore the original scale through the inverse transformation of the reversible gated normalization module to obtain the final demand prediction value.

Images