Aerial material demand intelligent prediction method and system based on dynamic weight adaptive adjustment

Abstract

The invention discloses an intelligent aviation material demand prediction method and system based on dynamic weight adaptive adjustment, and belongs to the field of civil aviation aviation material intelligent scheduling. The method comprises the following steps: constructing a multi-source feature matrix fusing historical maintenance data features, model parameter features and environment factor features; performing weighted fusion on the multi-source features based on a dynamic weighting function comprising a seasonal adjustment component and a gradient feedback component; inputting the weighted and fused features into a hybrid prediction model comprising a time sequence feature extraction layer and a long-term correlation modeling layer for aerial material demand prediction; and training a hybrid prediction model based on a weight loss function of the embedded dynamic weight vector. According to the method, the adaptive adjustment of the multi-source feature weight is realized through the dynamic weighting function, the seasonal fluctuation of the aerial material demand can be accurately captured, the sudden change can be quickly responded, and the prediction precision and the inventory turnover efficiency are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent scheduling of civil aviation materials, specifically involving an intelligent forecasting method and system for aviation material demand based on dynamic weight adaptive adjustment, which is applicable to accurate inventory management and demand forecasting in scenarios such as airport aviation material warehouses and airline maintenance centers. Background Technology

[0002] Aircraft material demand forecasting is a crucial link in the civil aviation maintenance support system, and its accuracy directly affects aircraft material inventory levels and maintenance support efficiency. With the continuous expansion of civil aviation transport and the increasing complexity of fleet structure, the technical challenges of aircraft material demand forecasting are also growing.

[0003] Existing aircraft material demand forecasting technologies have the following main shortcomings: First, regarding feature weight allocation, existing methods typically employ a static weight strategy. This means that after determining the weights of each feature through an optimization algorithm during model training, the weights remain fixed during actual operation. For example, existing techniques use improved particle swarm optimization to optimize the initial weights and thresholds of a backpropagation (BP) neural network. While this improves prediction accuracy to some extent, the weights are fixed after training, failing to respond to the seasonal fluctuations in aircraft material demand. The civil aviation maintenance field exhibits significant seasonal patterns—increased flight density during the Spring Festival travel rush leads to accelerated aircraft material consumption; high summer temperatures exacerbate the failure rate of specific components; and low winter temperatures affect the reliability of precision components such as hydraulic systems. These seasonal factors require feature weights to be dynamically adjusted over time, a requirement that static weight methods struggle to meet.

[0004] Secondly, regarding the predictive model architecture, existing methods mostly employ a single type of neural network model. Traditional BP neural networks are shallow feedforward networks; while they possess some nonlinear fitting capability, their limited depth makes it difficult to effectively capture temporal dependencies in historical maintenance data. While single Long Short-Term Memory (LSTM) networks excel at processing time-series data, they are insufficient for modeling long-distance correlations between multidimensional features such as aircraft parameters and environmental factors and aircraft material requirements. A single model architecture cannot simultaneously address the dual needs of short-term temporal feature extraction and long-term correlation feature modeling, thus limiting prediction accuracy.

[0005] Third, regarding safety assurance, existing aircraft material demand forecasting methods typically treat all aircraft materials as equally important, failing to implement differentiated forecasting strategies for materials with different safety levels. However, the civil aviation sector has much higher requirements for the assurance of critical flight components (such as engines, landing gear, and other core components that directly affect flight safety) than for general aircraft materials. Therefore, it is necessary to introduce a safety redundancy mechanism based on the forecasting results to ensure that the supply of critical flight components is not interrupted under any circumstances. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent forecasting method and system for aircraft material demand based on dynamic weight adaptive adjustment, which realizes dynamic adaptive adjustment of multi-source feature weights in aircraft material demand forecasting, improves forecasting accuracy and inventory turnover efficiency in complex scenarios, and meets the redundancy requirements for civil aviation safety.

[0007] To achieve the above-mentioned objectives, the technical solution provided by this invention includes: A smart forecasting method for aircraft material demand based on dynamic weight adaptive adjustment includes: Construct a multi-source feature matrix, which integrates historical maintenance data features, machine model parameter features, and environmental factor features; The various features in the multi-source feature matrix are weighted and fused based on a dynamic weighting function, which includes a seasonal adjustment component and a gradient feedback component. The weighted and fused features are input into a hybrid prediction model for predicting aircraft material demand. The hybrid prediction model includes a time-series feature extraction layer and a long-term correlation modeling layer. The hybrid prediction model is trained based on a weighted loss function, and the dynamic weight vector output by the dynamic weight function is embedded in the weighted loss function.

[0008] Preferably, the historical maintenance data features include component failure frequency and maintenance cycle time series data; the aircraft model parameter features include flight hours, number of takeoffs and landings, and aircraft model code data; and the environmental factor features include temperature, humidity, and airport altitude data.

[0009] Preferably, the dynamic weighting function is expressed as: ; in, These are the initial weight base values; For the seasonal adjustment component, t is the current time step, T is the period parameter, and λ is the fluctuation amplitude coefficient; Here, β represents the gradient feedback component, and β is the learning rate parameter. This is the gradient of the predicted loss from the previous period.

[0010] Preferably, the temporal feature extraction layer is a long short-term memory network, the long-term association modeling layer is a Transformer network, and the hybrid prediction model adopts an LSTM-Transformer concatenated architecture, using the output of the long short-term memory network as the input of the Transformer network.

[0011] Preferably, the Transformer network employs a multi-head attention mechanism with 8 attention heads, each with a feature dimension of 64.

[0012] Preferably, the initial weight base value is allocated to various features according to a preset ratio, and the initial weight ratio of the historical maintenance data feature, the model parameter feature and the environmental factor feature is 0.45∶0.30∶0.25.

[0013] Preferably, before constructing the multi-source feature matrix, the method further includes data preprocessing of the historical maintenance data features, the model parameter features, and the environmental factor features, respectively. The data preprocessing includes data normalization and missing value imputation.

[0014] Preferably, it further includes: when the predicted object is a critical flight component, applying a safety factor to the predicted demand output by the hybrid prediction model to obtain a safety redundancy demand, and setting an emergency procurement threshold, triggering an emergency procurement process when the inventory of the critical flight component is lower than the emergency procurement threshold.

[0015] Preferably, the safety factor ranges from 1.2 to 1.5, and the emergency procurement threshold is set as the predicted usage of the critical flight components within a preset number of days.

[0016] This invention also discloses an intelligent forecasting system for aircraft material demand based on dynamic weight adaptive adjustment, comprising: The multi-source data acquisition module is used to collect historical maintenance data characteristics, machine model parameter characteristics, and environmental factor characteristics, and to construct a multi-source feature matrix; The dynamic weight calculation module is used to calculate the dynamic weights of various features based on a dynamic weight function that includes seasonal adjustment components and gradient feedback components, and to perform weighted fusion on the multi-source feature matrix. The hybrid prediction model module includes a time-series feature extraction layer and a long-term correlation modeling layer, which are used to receive weighted and fused features and perform aircraft material demand prediction. The hybrid prediction model module is trained based on a weighted loss function with embedded dynamic weight vectors. The prediction output module is used to output the prediction results of aircraft material demand.

[0017] Beneficial effects The intelligent forecasting method and system for aircraft material demand based on dynamic weight adaptive adjustment provided by this invention has the following beneficial effects: 1. By designing a dynamic weighting function that includes seasonal adjustment components and gradient feedback components, the weights of multi-source features can be automatically adjusted with time period and prediction error, thereby achieving accurate capture of seasonal fluctuations in aircraft material demand and timely correction of prediction deviations. Compared with the traditional method using static weights, the prediction accuracy is significantly improved.

[0018] 2. By adopting a hybrid prediction model architecture that includes a temporal feature extraction layer and a long-term correlation modeling layer, the advantages of the temporal feature extraction layer in capturing short-term temporal dependencies of historical maintenance data and the advantages of the long-term correlation modeling layer in modeling long-distance correlations of multi-dimensional features are fully utilized to achieve collaborative modeling of multi-scale temporal features of aircraft material demand, overcoming the limitations of a single model architecture in feature extraction capabilities.

[0019] 3. By embedding dynamic weight vectors into the weighted loss function, end-to-end collaborative optimization of feature importance and model training process is achieved, enabling the model to adaptively enhance the contribution of key features during training, thereby improving inventory turnover efficiency and reducing the cost of aircraft material backlog.

[0020] 4. By setting up a safety redundancy module for critical flight components, a safety factor is automatically applied and an emergency procurement threshold is set based on the demand forecast results. This enables a differentiated guarantee strategy for aircraft materials of different safety levels, effectively meeting the stringent requirements of civil aviation safety operations for sufficient inventory of critical flight components. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a preferred embodiment of the intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment. Figure 2 This is a schematic diagram of the structure of an intelligent prediction system for aircraft material demand based on dynamic weight adaptive adjustment, provided in a preferred embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] Example 1 like Figure 1 As shown in the figure, this embodiment provides an intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment. This method is applied to the civil aviation aircraft material inventory management scenario. In a specific implementation scenario, the method is deployed in the aircraft material warehouse management system of a large hub airport, accessing the airport's historical maintenance records, fleet operation data, and meteorological environmental data for the past three years, providing aircraft material inventory management personnel with a forecast list of aircraft material demand for a certain period in the future and inventory optimization suggestions.

[0024] The hardware environment required for this method includes a server cluster with GPU computing capabilities for training and inference operations of the hybrid prediction model. The data acquisition terminal interfaces with airport maintenance information systems, airline operation control systems, and meteorological data services to achieve automated collection and aggregation of multi-source data.

[0025] Step S1: Construct a multi-source feature matrix, which integrates historical maintenance data features, machine model parameter features, and environmental factor features.

[0026] The purpose of this step is to structure and encode heterogeneous data from different data sources to form a unified multi-source feature matrix, providing standardized input for subsequent dynamic weighted fusion and model prediction.

[0027] In aircraft material demand forecasting, factors influencing demand come from multiple dimensions, and a single data source cannot fully depict the changing patterns of demand. Therefore, this invention constructs a multi-source feature matrix from three dimensions: historical maintenance data, aircraft parameters, and environmental factors. .

[0028] Specifically, characteristics of historical maintenance data This includes component failure frequency and maintenance cycle time series data. Component failure frequency reflects the failure density of various types of aircraft parts in historical operation, while maintenance cycle time series data records the time interval sequence between the last maintenance and the next failure for each aircraft part. In this embodiment, maintenance work order data from a certain airport over the past three years were collected, and the monthly failure frequency and maintenance interval time series of each type of aircraft part were extracted to form a historical maintenance data feature vector.

[0029] Model parameters and features This includes flight hours, number of takeoffs and landings, and aircraft type code data. Flight hours and number of takeoffs and landings directly reflect the aircraft's usage intensity and are important drivers of wear and tear and consumption of aircraft parts. Aircraft type codes are used to differentiate the different parts requirements of different aircraft types (such as A320, B737, etc.) because different aircraft types have different structural designs and component configurations, resulting in significantly different parts consumption patterns. In this embodiment, monthly flight hours and number of takeoffs and landings for each aircraft type are obtained from the airline's operation control system, and the aircraft type codes are processed using one-hot encoding.

[0030] Environmental Factor Characteristics This includes data on temperature, humidity, and airport altitude. Environmental factors directly impact the lifespan and failure rate of aircraft materials—high temperatures accelerate the aging of rubber seals, high humidity promotes corrosion of metal components, and the low-pressure environment of high-altitude airports places higher demands on the reliability of hydraulic systems. In this embodiment, monthly average temperature, average humidity, and airport altitude data are obtained from a meteorological data service interface.

[0031] The three types of features mentioned above are concatenated column-wise to form a multi-source feature matrix. Each row of this matrix corresponds to a time step (month), and each column corresponds to a feature dimension.

[0032] In one embodiment of the present invention, before constructing the multi-source feature matrix, data preprocessing is performed on historical maintenance data features, aircraft parameter features, and environmental factor features respectively. Because the dimensions and numerical ranges of the three types of features differ significantly (for example, flight hours may be on the order of thousands of hours, while temperature values ​​range from below zero to over forty degrees Celsius), if the raw data is directly input into the model, features with larger values ​​will dominate during model training, preventing the model from fairly learning the contributions of each type of feature. Therefore, data preprocessing includes data normalization—scaling each feature value to a uniform numerical range (e.g., the [0,1] interval)—and missing value imputation—interpolating and imputing missing data caused by data acquisition interruptions or system failures to ensure the integrity of the input matrix. It should be understood that, in addition to the above normalization methods, other data normalization methods such as standardization (z-score normalization) can also be used, as long as they can eliminate the dimensional differences between different features.

[0033] Step S2: Weight and fuse various features in the multi-source feature matrix based on a dynamic weight function, wherein the dynamic weight function includes a seasonal adjustment component and a gradient feedback component.

[0034] This step is one of the core innovations of this invention. Traditional aircraft material demand forecasting methods typically determine and fix the weights of each feature during the model training phase. This static weighting strategy cannot adapt to the seasonal fluctuations in aircraft material demand. For example, during the high temperatures of summer, the impact of environmental factors on aircraft material failure rates is significantly enhanced, and the weight of environmental factor features should be increased accordingly. Conversely, during periods of stable flight scheduling, historical maintenance data has relatively higher reference value, and its weight should be increased. This invention designs a dynamic weighting function, enabling the weights of various features to automatically adjust with time and prediction errors.

[0035] The dynamic weighting function is expressed as follows: ; in, The initial weight base value represents the basic weight allocation of various features in the absence of seasonal fluctuations. In this embodiment, the initial weight ratio of historical maintenance data features, aircraft parameter features, and environmental factor features is set to 0.45:0.30:0.25. This ratio is set based on the experience of domain experts, reflecting that historical maintenance data usually has the highest reference value in aircraft material demand forecasting.

[0036] Seasonal Adjustment Components This is the first driving factor of the dynamic weighting function. Here, T is the periodic parameter, which is set to 12 in this embodiment, corresponding to a monthly cycle to adapt to the annual seasonality of civil aviation maintenance. λ is the fluctuation amplitude coefficient, controlling the adjustment range of weights with seasonal changes; in this embodiment, its value ranges from 0.1 to 0.3. When λ=0.2, the weights of each feature can fluctuate by 20% above or below the baseline value. This range is sufficient to reflect the impact of seasonal factors while avoiding excessively drastic weight adjustments that could lead to model instability. The periodicity of the sine function allows the weight adjustment to smoothly track the annual cyclical pattern of civil aviation maintenance, resulting in larger weight shifts during the summer and winter seasons when failure rates change significantly, while the weights tend to approach the baseline value during the spring and autumn transition periods.

[0037] Gradient feedback components This is the second driving factor of the dynamic weighting function. Among them, The gradient represents the prediction loss gradient of the previous period, reflecting the sensitivity of the model's prediction bias to the weights of each feature in the previous prediction period. β is the learning rate parameter, controlling the influence of gradient feedback on weight adjustment; in this embodiment, its value ranges from 0.05 to 0.2. The role of the gradient feedback component is to adjust the gradient direction of each feature weight through the loss function when the model's prediction result in the previous period deviates, so that features that contribute more to the prediction error receive higher weights, thus receiving more attention in the next prediction period. This error-driven weight optimization mechanism enables the dynamic weight function not only to track periodic patterns but also to respond quickly to sudden events (such as temporary flight schedule changes, extreme weather, etc.).

[0038] The collaborative mechanism of the seasonal adjustment component and the gradient feedback component is as follows: the seasonal adjustment component provides the baseline for weight adjustment, that is, predicting the relative importance of each feature at the current time point based on historical patterns; the gradient feedback component then fine-tunes the weights based on the latest prediction feedback information. The combination of the two forms a dual-driven dynamic weight adjustment mechanism that balances regularity and adaptability.

[0039] It should be understood that, in addition to the sine function, the seasonal adjustment component can also be modeled using the cosine function, Fourier series, or other periodic functions to model seasonal fluctuations. The core of this invention lies in capturing the seasonal patterns of aircraft material demand through periodic functions, rather than being limited to a specific function form. Similarly, the gradient feedback component can also use other gradient optimization strategies such as the momentum method and adaptive learning rate to replace simple first-order gradient feedback, as long as it can achieve adaptive adjustment of weights based on prediction error.

[0040] The dynamic weight vector at the current time step is calculated based on the dynamic weight function. Then, the various features in the multi-source feature matrix are weighted and fused. Specifically, the dynamic weight vector is... The weighted components of each type of feature are multiplied by their corresponding feature vectors to obtain the weighted fused features.

[0041] To better understand the actual effect of dynamic weight adjustment, a specific numerical example is given below. Assume we are currently in July (summer), with a time step t=7, a period parameter T=12, a fluctuation amplitude coefficient λ=0.2, and an initial weight base value... At this time, the seasonal adjustment component is... Assume the gradient of the predicted loss in the previous period. Learning rate parameter Then the gradient feedback component is The final dynamic weight is As can be seen, at this time step, the weights of the three types of features have all been adjusted, and the magnitude and direction of the adjustment take into account both seasonal patterns and error feedback information.

[0042] The update frequency of dynamic weights can be configured according to actual business needs. In this embodiment, the dynamic weights are updated monthly, consistent with the granularity of aircraft material demand forecasting. In application scenarios requiring higher time resolution (such as daily forecasting), the period parameter T can be adjusted to 365 (corresponding to a daily period), and the fluctuation amplitude coefficient λ can be appropriately reduced to avoid excessive weight fluctuations. In addition, in the initial stage of system operation (when historical data accumulation is insufficient), the learning rate parameter β of the gradient feedback component can be set to a small value (such as 0.05), and gradually increased to the normal range of 0.1 to 0.2 after the system operation stabilizes.

[0043] Step S3: Input the weighted and fused features into the hybrid prediction model to predict the demand for aircraft materials. The hybrid prediction model includes a time-series feature extraction layer and a long-term correlation modeling layer.

[0044] A core challenge in aircraft material demand forecasting is the need to simultaneously capture both short-term temporal dependencies (such as periodic patterns of component failures and gradual trends in maintenance intervals) and long-term correlations between multidimensional features (such as the material consumption patterns of specific aircraft models under specific environmental conditions) in historical maintenance data. Traditional single-model architectures—whether recurrent neural networks, which excel at temporal modeling, or attention networks, which excel at global correlation modeling—struggle to meet both of these requirements at the same time.

[0045] The hybrid prediction model of this invention adopts a series architecture of temporal feature extraction layer and long-term correlation modeling layer, giving full play to the complementary advantages of the two types of networks.

[0046] The temporal feature extraction layer is used to extract short-term temporal dependencies from historical maintenance data features. In one embodiment of the invention, the temporal feature extraction layer is a Long Short-Term Memory (LSTM) network. LSTM, through its unique gating mechanism (input gate, forget gate, and output gate), can selectively remember and forget information in temporal data, thereby effectively capturing short-term fluctuation patterns in aircraft material failure frequency and gradual trends in maintenance cycles. In this embodiment, the number of hidden units in the LSTM layer is set to 128 dimensions, the input is a weighted and fused feature sequence, and the output is a feature vector containing temporal dependency information. .

[0047] The long-term correlation modeling layer is used to capture the long-term correlation features between aircraft parameter features, environmental factor features, and aircraft material requirements. In one embodiment of the invention, the long-term correlation modeling layer is a Transformer network. The Transformer, through a multi-head attention mechanism, can compute the correlation strength between any two positions in the input sequence in parallel, thus overcoming the limitations of recurrent neural networks in long-distance dependency modeling. In this embodiment, the Transformer employs a multi-head attention mechanism with 8 attention heads, each with a feature dimension of 64, and 4 encoder layers. The Transformer layer receives the output of the LSTM layer. As input, the final predicted feature representation is output through multi-head attention computation and feedforward network transformation.

[0048] The overall data flow of the hybrid prediction model is as follows: the weighted and fused feature sequence first passes through an LSTM layer to extract temporal dependency features, then the output of the LSTM is fed into a Transformer layer for long-term correlation modeling, and finally, the predicted value of aircraft material demand is generated through a fully connected output layer. The advantage of this concatenated architecture is that the LSTM layer filters and compresses temporal information in advance, reducing the input complexity of the Transformer layer and allowing it to focus more on modeling long-term associated features; at the same time, the Transformer layer retains the temporal information extracted by the LSTM layer through residual connections, achieving effective fusion of short-term and long-term features.

[0049] It should be understood that, in addition to LSTM, the temporal feature extraction layer can also employ other network structures with temporal modeling capabilities, such as Gated Recurrent Units (GRUs) and Temporal Convolutional Networks (TCNs). Similarly, the long-term correlation modeling layer can employ other variant networks based on attention mechanisms, such as Transformer variants with relative position encoding, or hybrid networks based on convolutional attention. The core of this invention lies in employing a serial hybrid architecture that processes temporal feature extraction and long-term correlation modeling in layers and allows them to work collaboratively, rather than being limited to a specific network implementation.

[0050] Regarding the training strategy for the hybrid prediction model, this embodiment adopts a phased training approach. In the first phase, an LSTM layer is pre-trained separately, using historical maintenance data features as input and aircraft material demand as the supervision signal to train the LSTM layer to learn temporal dependency patterns. Pre-training uses the standard MSE loss function, with 100 training epochs. The initial learning rate is set to 0.001 and gradually reduced using a cosine annealing strategy. In the second phase, a Transformer layer is superimposed on the pre-trained LSTM layer, using complete weighted fused features as input and a weighted loss function... As an optimization objective, end-to-end joint training was performed. The joint training consisted of 200 epochs, with an initial learning rate of 0.0005 and a batch size of 32. The advantage of this phased training strategy is that the pre-training phase enables the LSTM layer to acquire good temporal feature extraction capabilities, avoiding the training instability caused by mutual interference between the LSTM and Transformer layers when joint training begins from random initialization.

[0051] During the model inference phase, the system uses a monthly forecast cycle, automatically triggering a complete forecast process at the beginning of each month: collecting the latest multi-source data, updating dynamic weights, running the hybrid forecast model, and generating a 30-day forecast list of aircraft material demand. Simultaneously, the system supports on-demand, real-time forecasting. When unexpected events are detected (such as extreme weather warnings or large-scale temporary flight adjustments), dynamic weight updates and forecast refreshes can be completed within one hour, providing timely decision support for aircraft material management personnel.

[0052] Step S4: Train the hybrid prediction model based on the weighted loss function, and embed the dynamic weight vector output by the dynamic weight function into the weighted loss function.

[0053] In traditional model training, the loss function is usually expressed as the standard mean squared error (MSE), i.e. The standard loss function assigns the same penalty weight to all prediction biases. However, in aircraft material demand forecasting, different feature dimensions contribute differently to the prediction results, and this difference changes dynamically over time. If a standard loss function is used, the model cannot perceive the dynamic changes in feature importance, leading to a disconnect between the training process and the dynamic weight adjustment mechanism.

[0054] This invention designs a weighted loss function. The dynamic weight vector w is embedded into the calculation process of the loss function. Here, y represents the actual demand for aircraft materials. Let w be the predicted demand for aviation materials by the model, and w be the current dynamic weight vector output by the dynamic weight function in step S2. By using the square of the dynamic weight vector as a weighting factor in the loss function, the following effect is achieved: when a certain feature receives a high dynamic weight at the current time step, the prediction bias corresponding to that feature will receive a larger penalty coefficient in the loss function, thereby guiding the model to pay more attention to the fitting accuracy of that feature during backpropagation.

[0055] This design achieves end-to-end collaboration between the dynamic weighting mechanism and the model training process: the dynamic weighting function adjusts the weights of each feature according to seasonal patterns and error feedback, and the weighted loss function passes this weighting information to the model's gradient update process, so that the direction of the model's parameter optimization is consistent with the dynamic importance of the features.

[0056] The model training process employs a gradient descent-based optimization algorithm (such as the Adam optimizer). In each training batch, forward propagation calculates the weighted loss value, and backpropagation calculates the gradients of each parameter and updates the model parameters. During training, the dynamic weight vector is updated at each time step, allowing the model to focus on different features at different training stages.

[0057] In some preferred embodiments, it further includes: Step S5: When the predicted object is a critical flight component, apply a safety factor to the predicted demand output by the hybrid prediction model to obtain a safety redundancy demand, and set an emergency procurement threshold.

[0058] Civil aviation safety is the primary consideration in aircraft material management. Flight-critical components refer to core components that directly affect flight safety, such as engines, landing gear, and flight control systems. Supply disruptions to these components can directly threaten normal flight operations and even flight safety. Therefore, inventory management of flight-critical components cannot rely solely on forecasts; a safety redundancy mechanism must also be introduced.

[0059] In one embodiment of the present invention, the safety redundancy module is automatically activated when the predicted object is identified as a flight-critical component. This module first multiplies the predicted demand output by the hybrid prediction model by a safety factor k to obtain the safety redundancy demand. The safety factor ranges from 1.2 to 1.5, and its specific value is determined based on the safety level of the flight-critical component and the historical supply chain reliability. For example, for core engine components, the safety factor can be set to 1.5, meaning an additional 50% redundancy is added to the predicted demand; for landing gear components, the safety factor can be set to 1.3.

[0060] Meanwhile, the safety redundancy module sets an emergency procurement threshold. When the current inventory of critical flight components falls below the predicted usage for a preset number of days (e.g., 3 days), the system automatically triggers an emergency procurement process, sends an emergency procurement notification to the procurement department, and recommends the highest priority supplier and the fastest logistics solution.

[0061] The identification of flight-critical components is based on a pre-established aircraft material classification database. This database categorizes all aircraft materials into three safety levels—flight-critical, important, and general—according to airworthiness directives and maintenance protocols issued by the Civil Aviation Administration of China. The safety redundancy module is activated only for flight-critical and important components; for general components, the original prediction results from the hybrid prediction model are used directly.

[0062] Example 2 like Figure 2 As shown, the present invention also provides an intelligent forecasting system for aircraft material demand based on dynamic weight adaptive adjustment. The system includes a multi-source data acquisition module, a dynamic weight calculation module, a hybrid forecasting model module, and a forecasting output module.

[0063] The multi-source data acquisition module interfaces with the airport maintenance information system, airline operation control system, and meteorological data service to automatically collect historical maintenance data characteristics (component failure frequency, maintenance cycle time series), aircraft parameter characteristics (flight hours, number of takeoffs and landings, aircraft type code), and environmental factor characteristics (temperature, humidity, airport altitude), and construct a multi-source feature matrix according to a preset data format. This module is also responsible for data preprocessing, including data normalization and missing value imputation.

[0064] The dynamic weight calculation module calculates weights based on the predicted loss gradient of the current time step and the previous period, using a dynamic weight function. This module calculates the dynamic weights of various features and performs weighted fusion on the multi-source feature matrix. It internally maintains historical weight records and supports visualization of weight change trends and alerts for abnormal fluctuations.

[0065] The hybrid prediction model module receives features after dynamic weighted fusion, extracts short-term temporal dependencies through a temporal feature extraction layer (LSTM network, 128-dimensional hidden units), and captures long-term correlation features through a long-term association modeling layer (Transformer network, 8-head attention mechanism, 64-dimensional feature dimension, 4-layer encoder), finally outputting the predicted value of aircraft material demand. This module is continuously trained and updated based on a weighted loss function with embedded dynamic weight vectors.

[0066] The forecast output module receives the forecast results from the hybrid forecast model module and generates a forecast list of aircraft material demand for a preset future period (e.g., 30 days). For flight-critical components, the forecast output module calls the safety redundancy submodule to apply a safety factor and monitor emergency procurement thresholds. Finally, this module outputs a forecast list of aircraft material demand including safety redundancy and inventory optimization suggestions for aircraft material warehouse managers to make informed decisions.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent forecasting of aircraft material demand based on dynamic weight adaptive adjustment, characterized in that, include: Construct a multi-source feature matrix, which integrates historical maintenance data features, machine model parameter features, and environmental factor features; The various features in the multi-source feature matrix are weighted and fused based on a dynamic weighting function, which includes a seasonal adjustment component and a gradient feedback component. The weighted and fused features are input into a hybrid prediction model for predicting aircraft material demand. The hybrid prediction model includes a time-series feature extraction layer and a long-term correlation modeling layer. The hybrid prediction model is trained based on a weighted loss function, and the dynamic weight vector output by the dynamic weight function is embedded in the weighted loss function.

2. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 1, characterized in that, The historical maintenance data features include component failure frequency and maintenance cycle time series data; the aircraft model parameter features include flight hours, number of takeoffs and landings, and aircraft model code data; and the environmental factor features include temperature, humidity, and airport altitude data.

3. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 1, characterized in that, The dynamic weighting function is expressed as follows: ； in, These are the initial weight base values; For the seasonal adjustment component, t is the current time step, T is the period parameter, and λ is the fluctuation amplitude coefficient; Here, β represents the gradient feedback component, and β is the learning rate parameter. This is the gradient of the predicted loss from the previous period.

4. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 1, characterized in that, The temporal feature extraction layer is a Long Short-Term Memory (LSTM) network, the long-term association modeling layer is a Transformer network, and the hybrid prediction model adopts an LSTM-Transformer concatenated architecture, using the output of the LSTM network as the input of the Transformer network.

5. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 4, characterized in that, The Transformer network employs a multi-head attention mechanism with 8 attention heads, each with a feature dimension of 64.

6. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 3, characterized in that, The initial weight base values ​​are allocated to various features according to a preset ratio, and the initial weight ratio of the historical maintenance data feature, the model parameter feature, and the environmental factor feature is 0.45∶0.30∶0.

25.

7. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 1, characterized in that, Before constructing the multi-source feature matrix, the data preprocessing is performed on the historical maintenance data features, the model parameter features, and the environmental factor features, respectively. The data preprocessing includes data normalization and missing value imputation.

8. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 1, characterized in that, Also includes: When the predicted object is a critical flight component, a safety factor is applied to the predicted demand output by the hybrid prediction model to obtain a safety redundancy demand, and an emergency procurement threshold is set. When the inventory of the critical flight component is lower than the emergency procurement threshold, the emergency procurement process is triggered.

9. The intelligent forecasting method for aircraft material demand based on dynamic weight adaptive adjustment according to claim 8, characterized in that, The safety factor ranges from 1.2 to 1.5, and the emergency procurement threshold is set as the predicted usage of the critical flight components within a preset number of days.

10. An intelligent forecasting system for aircraft material demand based on dynamic weight adaptive adjustment, characterized in that, include: The multi-source data acquisition module is used to collect historical maintenance data characteristics, machine model parameter characteristics, and environmental factor characteristics, and to construct a multi-source feature matrix; The dynamic weight calculation module is used to calculate the dynamic weights of various features based on a dynamic weight function that includes seasonal adjustment components and gradient feedback components, and to perform weighted fusion on the multi-source feature matrix. The hybrid prediction model module includes a time-series feature extraction layer and a long-term correlation modeling layer, which are used to receive weighted and fused features and perform aircraft material demand prediction. The hybrid prediction model module is trained based on a weighted loss function with embedded dynamic weight vectors. The prediction output module is used to output the prediction results of aircraft material demand.

Images