System and method for dynamic allocation of aircraft fixed costs

Abstract

The application discloses an aircraft fixed cost dynamic allocation system and method, wherein the method comprises a data fusion module, a causal modeling module, a responsibility quantification module, an allocation optimization module and a voucher generation module; the data fusion module is used for obtaining fixed costs and multi-source aviation business data; the causal modeling module is used for constructing a fixed cost cause and effect diagram based on a domain constraint causal discovery algorithm; the responsibility quantification module is used for calculating the marginal responsibility value of each flight to each fixed cost based on counterfactual inference and SHAP value, and generating initial allocation weights; the allocation optimization module is used for optimizing and adjusting the initial allocation weights to obtain effective allocation weights; and the voucher generation module is used for performing flight-level allocation calculation of the fixed costs by using the effective allocation weights, and generating financial voucher files. The application can improve the accuracy, rationality, dynamic adaptability and interpretability of aircraft fixed cost attribution allocation.

Description

Technical Field

[0001] This invention relates to the field of aviation settlement technology, and in particular to a dynamic allocation system and method for aircraft fixed costs. Background Technology

[0002] Airlines incur numerous fixed costs during operations, such as aircraft depreciation, lease payments, hull insurance, engine overhaul reserves, APU maintenance costs, fixed allocations for takeoff and landing services, and basic crew salaries, totaling hundreds of expense categories. These fixed costs need to be allocated reasonably and with evidence to each flight, each route, and each aircraft to support route profitability analysis, aircraft asset management, capacity allocation decisions, and financial compliance accounting.

[0003] Existing technology includes an aviation cost allocation system based on activity-based costing. This system pre-assigns several "cost drivers" to each fixed cost through interviews with business experts. For example, the driver for aircraft depreciation is "flight hours," the driver for engine maintenance reserves is "engine cycle count," and the driver for crew fixed salaries is "number of days on duty." Each month, the system extracts the actual occurrence of these drivers from the operational system for each flight and allocates the total cost to each flight proportionally. The fundamental flaw of this approach is its reliance on manual experience and judgment, leading to poor accuracy and adaptability in the allocation results, specifically as follows: (1) Inability to handle redundancy and spurious correlations between drivers. Business personnel may simultaneously set both "flight hours" and "passenger numbers" as drivers for a certain cost, but these two variables are highly correlated, leading to double counting of costs. More seriously, "passenger numbers" may have a statistical correlation with engine maintenance costs, but the real drivers of maintenance costs are "takeoff thrust" and "outside temperature," and "passenger numbers" only show a spurious correlation because of their correlation with "route distance." Current technology cannot identify this distinction and will incorrectly allocate maintenance costs to flights with more passengers, resulting in distorted benefit analysis.

[0004] (2) The selection of driving factors is fixed and cannot cope with dynamic changes. When the fleet ages (maintenance consumption of the same flight hour accelerates) or new high-altitude routes are opened (consumption of the same flight cycle is greater), the old driving factor system immediately becomes ineffective, but manual revision is often seriously delayed.

[0005] (3) It cannot answer the counterfactual question of “how much cost was actually caused by a certain flight”. The apportionment is just a proportional division rather than causal attribution.

[0006] Existing technology also includes a semi-automated cost allocation system based on a fixed coefficient table. This system automatically extracts flight data from multiple business systems, but uses a set of allocation coefficient tables maintained manually on a regular basis (e.g., annually) for calculation. The coefficient table specifies the percentage of fixed costs each flight should bear under a certain aircraft type and route type, and the system automatically calculates according to the table monthly. While this solution automates data extraction and calculation, it still has a key drawback: (1) The coefficient table is outdated and lacks dynamism. After the aircraft maintenance plan is adjusted or the route network is changed, the old coefficients quickly become invalid.

[0007] (2) The basis for apportionment is inexplicable. Only a string of numbers can be given, for example, it cannot explain "why this flight pays more and that flight pays less".

[0008] Therefore, existing technologies suffer from drawbacks such as allocation based on human experience, rigid cost driver allocation logic, and lack of interpretability of allocation results. Summary of the Invention

[0009] This invention provides a dynamic allocation system for aircraft fixed costs to improve the accuracy, rationality, dynamic adaptability, and interpretability of aircraft fixed cost attribution allocation. The system includes: a data fusion module, a causal modeling module, a responsibility quantification module, an allocation optimization module, and a voucher generation module. The data fusion module is used to: acquire fixed cost and multi-source aviation business data, and establish an element pool; the element pool includes candidate cost drivers formed based on multi-source aviation business data; The causal modeling module is used to: construct observation data from an element pool, introduce aviation domain knowledge constraints, and construct a directed acyclic causal graph using a causal discovery algorithm; the causal graph represents the causal relationships between variables through nodes and edges; the variables are derived from the observation data; The responsibility quantification module is used to: obtain the target cost drivers for fixed costs through the causal graph; input the target cost drivers into the fixed cost prediction model, call the machine learning interpretability tool to attribute the predicted values ​​to each target cost driver, and obtain the marginal contribution value of each target cost driver to the predicted value; determine the initial allocation weight of fixed cost consumption for each flight using the marginal contribution values ​​of all target cost drivers to the predicted value; the fixed cost prediction model is used to predict the fixed cost expenses allocated at the flight level; the machine learning interpretability tool is used to quantify the degree of influence of input features on the model's predicted values; The allocation optimization module is used to: optimize and adjust the initial allocation weight of each flight to obtain the effective allocation weight of each flight; The voucher generation module is used to: calculate the fixed cost allocation at the flight level using the effective allocation weight of each flight, and obtain the allocation result; generate a financial voucher file using the allocation result; and store the allocation result in the financial voucher file in the form required by financial regulations.

[0010] This invention provides a method for dynamically allocating fixed costs for aircraft, aimed at improving the accuracy, rationality, dynamic adaptability, and interpretability of attribution and allocation of fixed costs for aircraft. The method includes: Acquire fixed cost and multi-source aviation business data, and establish a factor pool; the factor pool includes candidate cost drivers formed based on multi-source aviation business data; The observation data is constructed using an element pool, and knowledge constraints from the aerospace field are introduced. A directed acyclic causal graph is constructed using a causal discovery algorithm. The causal graph represents the causal relationships between variables through nodes and edges. The variables are derived from the observation data. The target cost drivers for fixed costs are obtained through the causal graph; these target cost drivers are input into the fixed cost prediction model, and machine learning interpretability tools are used to attribute the predicted values ​​to each target cost driver, obtaining the marginal contribution value of each target cost driver to the predicted value; the initial allocation weight of fixed cost consumption for each flight is determined using the marginal contribution values ​​of all target cost drivers to the predicted value; the fixed cost prediction model is used to predict fixed cost expenses allocated at the flight level; the machine learning interpretability tools are used to quantify the degree of influence of input features on the model's predicted values; The initial allocation weight for each flight is optimized and adjusted to obtain the effective allocation weight for each flight; Fixed costs are allocated at the flight level using the effective allocation weight of each flight to obtain the allocation result; financial voucher files are generated using the allocation result; the allocation result is stored in the financial voucher file in accordance with financial requirements.

[0011] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for dynamically allocating fixed costs of aircraft.

[0012] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for dynamically allocating fixed costs of aircraft.

[0013] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for dynamically allocating fixed costs of aircraft.

[0014] This invention proposes a dynamic allocation system and method for aircraft fixed costs. The system includes a data fusion module, a causal modeling module, a responsibility quantification module, an allocation optimization module, and a voucher generation module. The causal modeling module constructs observation data using an element pool, introduces aviation domain knowledge constraints, and constructs a directed acyclic causal graph through a causal discovery algorithm. This invention automatically identifies the true cost drivers through causal inference, avoiding the false associations and cross-subsidies problems caused by manual selection of drivers in traditional methods, and improving the accuracy and dynamic adaptability of aircraft fixed cost attribution allocation. In this embodiment of the invention, the responsibility quantification module obtains the target cost drivers for fixed costs through the causal graph; the target cost drivers are input into the fixed cost prediction model, and machine learning interpretability tools are called to attribute the predicted values ​​to each target cost driver, obtaining the marginal contribution value of each target cost driver to the predicted value, and then determining the initial allocation weight of each flight's fixed cost consumption. This embodiment of the invention introduces machine learning interpretability tools into the field of aviation cost management, enabling financial and business personnel to clearly understand the driving reasons for each cost allocation, improving the rationality, reliability, and interpretability of aircraft fixed cost attribution allocation; furthermore, the allocation optimization module optimizes and adjusts the initial allocation weight of each flight to obtain the effective allocation weight of each flight; finally, the voucher generation module performs allocation calculations and generates financial vouchers. This invention utilizes a causal inference technology system to automate and data-drivenly discover cost drivers, eliminate false correlations and redundant variables, and accurately locate the direct causal drivers of fixed costs. This ensures that the allocation results truly match the underlying logic of cost occurrence. From multi-source data fusion, driver discovery, responsibility quantification, allocation optimization to voucher entry, the entire process is unattended and fully automated. Furthermore, through interpretable machine learning technology, the allocation results are fully interpretable and traceable across all dimensions, clearly quantifying the marginal impact of each business element on cost allocation, making the allocation logic transparent and the allocation basis verifiable. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic diagram of the aircraft fixed cost dynamic allocation system in an embodiment of the present invention; Figure 2 This is a specific example diagram of the aircraft fixed cost dynamic allocation system in an embodiment of the present invention; Figure 3This is a flowchart illustrating the dynamic allocation method for aircraft fixed costs in an embodiment of the present invention. Figure 4 This is a specific example diagram of the aircraft fixed cost dynamic allocation method in the embodiments of the present invention; Figure 5 This is another specific example of the aircraft fixed cost dynamic allocation method in the embodiments of the present invention; Figure 6 This is a schematic diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0017] Existing fixed cost allocation schemes for aircraft generally adopt a fixed percentage allocation method based on a single indicator (such as flight hours, wheel chute hours, or available seat kilometers). This method has serious flaws: the allocation indicator is selected by operational personnel based on experience, which cannot reflect the true causes of cost occurrence; once the allocation logic is determined, it remains static and fixed for a long period, unable to adapt to changes in aircraft condition, operating environment, or route network adjustments. This leads to severely distorted cost information, with a large amount of cross-subsidization, i.e., profitable routes subsidizing loss-making routes, which may mislead important decisions such as capacity deployment and route selection.

[0018] With the enhancement of aviation data acquisition capabilities, multi-source data, including flight operations (QAR / ACARS, QAR: Quick Access Recorder; ACARS: Aircraft Communications Addressing and Reporting System), aircraft maintenance, meteorology, ground services, crew scheduling, and financial ledgers, are becoming increasingly comprehensive, providing a data foundation for more scientific allocation methods. In actual operations, there are hundreds of factors that can be considered as allocation candidates, covering aspects such as flight dynamics, aircraft attributes, carrying information, and service usage. Simultaneously, there may be hundreds of fixed cost items to be allocated. However, how to automatically identify the drivers with a real and direct causal relationship to each cost from these massive amounts of data, how to dynamically adapt to business changes, how to support on-demand allocation scope delineation for specialized analysis, and how to provide interpretable allocation results are currently unresolved key technical challenges.

[0019] This invention addresses the core pain points of aircraft fixed cost allocation management in civil aviation transportation enterprises, as well as the systemic and fundamental technical deficiencies of existing allocation technologies. It aims to provide a dynamic aircraft fixed cost allocation system and method based on causal inference and multi-objective adaptive optimization. This fundamentally overcomes the technical bottlenecks of existing technologies in five dimensions: cost driver identification, dynamic adaptation of allocation logic, model self-learning iteration, interpretability of allocation results, and flexibility in adapting to business scenarios. It achieves accurate attribution, compliant accounting, dynamic adaptation, and traceable control of aircraft fixed costs, providing airlines with real, reliable, and interpretable cost data support for route benefit analysis, aircraft asset management, capacity deployment decisions, and internal operational assessments. Specifically, this invention addresses the following key technical issues: (1) Solve the core technical problem that the cost drivers in the existing technology rely on human experience and cannot actively identify the real causal driving relationship from a large number of business elements, resulting in false correlation, double calculation and cross-subsidy in the allocation results.

[0020] Existing technologies, relying on manual designation of single or a few cost drivers by business personnel, cannot handle redundant and spurious correlations among hundreds of candidate business elements. They also cannot distinguish between statistical correlations and genuine causal relationships between variables, easily misjudging indirectly related elements as cost drivers. This leads to double-counting of costs, misallocation of cost responsibilities, and ultimately creates the illusion of unreasonably subsidizing loss-making routes on profitable routes, resulting in severely distorted route benefit analysis and misleading crucial operational decisions such as capacity deployment and route selection. The embodiments of this invention aim to automate and data-drivenly discover cost drivers through a causal inference technology system, eliminating spurious correlations and redundant variables, accurately locating the direct causal driving elements of each fixed cost, and ensuring that the allocation results truly match the underlying logic of cost occurrence.

[0021] (2) Solve the technical problem that the allocation logic in the existing technology is static and fixed, which cannot adapt to dynamic business scenarios such as fleet aging, changes in operating environment, and route network adjustment, and the allocation system fails and lags behind.

[0022] Existing cost allocation coefficients and rules, once determined, remain fixed for a long period. Manual revisions suffer from significant lag and cannot respond to dynamic business changes such as aircraft age and maintenance status, the opening of high-altitude / special routes, fluctuations in operating environment parameters, and fleet restructuring. This leads to the rapid invalidation of allocation rules and a continuous deviation between the allocation results and actual cost consumption levels. Furthermore, existing technologies fail to balance the core contradiction between "accurately tracking business changes" and "maintaining smooth and stable financial statements." Either the rigid rules lead to data distortion, or frequent rule adjustments cause drastic fluctuations in financial data, failing to meet the basic requirements of financial accounting. This invention aims to achieve periodic iteration of the allocation logic and real-time response to business changes through a multi-objective adaptive optimization mechanism. Simultaneously, through a decoupled smooth network design, it balances the accuracy of allocation results with the continuity and stability of financial data.

[0023] (3) Solve the technical problem that the existing allocation system has no closed-loop self-learning capability, cannot absorb qualitative feedback from business experts for continuous optimization, and the problem of repeated incorrect allocation occurs.

[0024] Existing semi-automated allocation systems can only automate the data extraction and calculation processes. They cannot transform qualitative feedback from business experts—such as objections, corrections, and judgments on the reasonableness of allocation results—into effective bases for system optimization. Allocation strategies cannot iteratively evolve, and similar abnormal allocation problems repeatedly occur, always relying on manual post-event corrections, failing to address the issue at the underlying strategy level. This invention aims to construct a human-machine closed-loop self-learning mechanism, transforming the qualitative judgments of business experts into high-weighted reward signals for model optimization. Through a reinforcement learning framework, it achieves continuous iteration and autonomous optimization of allocation strategies, making the system increasingly aligned with the actual business management needs of enterprises.

[0025] (4) Solve the technical problems of the existing technology where the allocation process is a "black box mode", which cannot provide a traceable and quantifiable explanation of the motivation for the allocation result, resulting in high communication costs between the finance and business sides and insufficient credibility of the allocation result.

[0026] Existing technologies can only output the final allocation value, failing to quantify the marginal contribution of each business element to the allocation result, and unable to explain the core driving reasons for differences in allocation amounts across different flights, routes, and aircraft. This does not meet the end-to-end traceability requirements of management accounting and is insufficient to support refined cost control decisions. Furthermore, unexplainable allocation results easily lead to discrepancies in understanding between the finance and business departments, significantly increasing communication costs and undermining the management credibility of the allocation results. This invention aims to achieve end-to-end explainability and full-dimensional traceability of allocation results through explainable AI technology, clearly quantifying the marginal impact of each business element on cost allocation, making the allocation logic transparent and the allocation basis verifiable.

[0027] (5) Solve the technical problem that the existing allocation system only supports unified allocation of all data, which cannot flexibly adapt to diversified management scenarios such as special cost analysis and independent accounting unit assessment, and has insufficient scenario adaptability.

[0028] Existing cost allocation systems with fixed allocation criteria cannot support users in defining the allocation data range as needed. A uniform allocation model across the board cannot meet the independent accounting and specialized cost analysis needs of specific fleets, route networks, or operating entities. A single allocation system cannot simultaneously satisfy the dual requirements of compliant financial accounting and refined management analysis, requiring enterprises to invest significant manpower in offline data recalculation and adjustment. This invention aims to achieve flexible customization of allocation criteria through configurable allocation range filtering and linked cost allocation mechanisms, adapting to diverse management accounting analysis scenarios while ensuring the accuracy and compliance of financial data consistency.

[0029] First, let me introduce the technical terms proposed or involved in the embodiments of this invention.

[0030] Fixed costs refer to indirect operating costs that do not change linearly with flight volume or flight hours, such as aircraft depreciation, lease payments, airframe / engine insurance, maintenance reserve accrual, and parking fees. Each month, the total amount F_total is obtained from the ERP (Enterprise Resource Planning) system, and the amount of the k-th item is recorded as F_total^k, satisfying ∑_k F_total^k = F_total.

[0031] Cost Driver: The driving factor that causes fixed costs to occur. In this embodiment of the invention, the driver is automatically identified from hundreds of candidate business elements by the causal discovery algorithm without the need for manual preset. It can also be called the allocation driver.

[0032] Causal Graph: A directed acyclic graph model that uses nodes and edges to represent causal relationships between variables. It is generated by PC algorithm combined with domain knowledge constraints and is used to identify direct cost drivers.

[0033] Marginal Responsibility Value: Based on counterfactual reasoning logic, this quantifies the causal contribution of each flight to a certain fixed cost. It can be positive or negative. A positive value indicates that the combination of operational characteristics of that flight results in higher-than-average cost consumption.

[0034] SHAP (SHapley Additive exPlanations): An interpretable machine learning method used to fairly attribute model predictions to individual input features. In this embodiment of the invention, it is used to calculate flight-level cost responsibility values.

[0035] PPO (Proximal Policy Optimization): A reinforcement learning algorithm used to optimize multi-objective amortization policies.

[0036] Smoothing Network (NN_η): An independent neural network that shares the feature extraction layer with the PPO Actor Network but has an independent output head. It is specifically designed for dynamically deciding the mixing coefficient η of the new and old weight allocation and the amount of error smoothing pool release, decoupling and optimizing "precise tracking" and "financial smoothing".

[0037] Smoothing Pool: An independent persistent account used to temporarily store the distributive differences caused by subsequent corrections to template data. The smoothing network autonomously decides the rate at which these differences are gradually absorbed, avoiding cross-period cost confusion.

[0038] Flight-level wide table: A flat data table with flights as the primary key, which links all relevant business elements such as aircraft, routes, transportation, services, crew, and weather through entity mapping rules.

[0039] Allocation Scope Filter: A functional module that allows users to define a subset of allocated data by combining various business conditions before an automated process is started.

[0040] Figure 1 This is a schematic diagram of the aircraft fixed cost dynamic allocation system in an embodiment of the present invention, such as... Figure 1 As shown, the system 100 includes: a data fusion module 101, a causal modeling module 102, a responsibility quantification module 103, an allocation optimization module 104, and a voucher generation module 105; The data fusion module 101 is used to: acquire fixed cost and multi-source aviation business data, and establish an element pool; the element pool includes candidate cost drivers formed based on multi-source aviation business data; The causal modeling module 102 is used to: construct observation data using an element pool, introduce aviation domain knowledge constraints, and construct a directed acyclic causal graph using a causal discovery algorithm; the causal graph represents the causal relationship between variables through nodes and edges; the variables are derived from the observation data; The responsibility quantification module 103 is used to: obtain the target cost drivers for fixed costs through the causal graph; input the target cost drivers into the fixed cost prediction model, call the machine learning interpretability tool to attribute the predicted value to each target cost driver, and obtain the marginal contribution value of each target cost driver to the predicted value; determine the initial allocation weight of fixed cost consumption for each flight using the marginal contribution values ​​of all target cost drivers to the predicted value; the fixed cost prediction model is used to predict the fixed cost expenses allocated at the flight level; the machine learning interpretability tool is used to quantify the degree of influence of input features on the model's predicted value; The allocation optimization module 104 is used to: optimize and adjust the initial allocation weight of each flight to obtain the effective allocation weight of each flight; The voucher generation module 105 is used to: calculate the fixed cost allocation at the flight level using the effective allocation weight of each flight, and obtain the allocation result; generate a financial voucher file using the allocation result; and store the allocation result in the financial voucher file in the form required by financial requirements.

[0041] Figure 1 The necessary modules of the aircraft fixed cost dynamic allocation system are shown. In this embodiment of the invention, the aircraft fixed cost dynamic allocation system adopts a microservice architecture, and the modules exchange data through event-driven and API-based communication.

[0042] Figure 2 This is a specific example diagram of the aircraft fixed cost dynamic allocation system in an embodiment of the present invention, as shown below. Figure 2 The diagram shows the overall system architecture and data flow of the aircraft fixed cost dynamic allocation system. The data flow path is as follows: [External Business System Integration] → (Daily Scheduled ETL, Raw Data Input) → [Data Fusion Layer] → (Data Quality Script, Output Flight-Level Wide Table) → [Causal Modeling Layer] → (Causal Driver Set Output) → [Responsibility Quantification Layer] → (Initial Allocation Weight Output) → [Allocation Optimization Layer] → (Final Effective Weight Output) → [Voucher Generation Layer] → (Voucher Posting Status Feedback) → [Monitoring Feedback Layer] → [ERP Financial System]; [ERP Financial System] → (Closing Completion Event) → [Amount Trigger Scheduler]; [Business Expert] → (Abnormal Feedback Button) → [Feedback Log Table] → (Reward Signal Generator) → [Reinforcement Learning Experience Pool].

[0043] The following is combined Figure 1 , Figure 2 This invention provides a detailed description of the aircraft fixed cost dynamic allocation system and method in its embodiments.

[0044] Module 1: Data Fusion Module

[0045] The data fusion module is responsible for acquiring raw data from multiple external systems, performing cleaning, standardization, missing value handling, outlier interception, and constructing a flight-level wide table based on entity mapping rules. This module is the system's sole data entry point.

[0046] External business systems that can be integrated include flight operation systems, aircraft master data systems, maintenance management systems, ERP financial systems, and business expert interactive terminals.

[0047] The data fusion layer corresponds to the data fusion module, which may include a multi-source data ETL unit, a flight-level wide table construction unit, a three-level missing value processing unit, amortization range filter, amortization error smoothing pool, cold start adaptation unit, etc. The specific functional implementation is described below.

[0048] Multi-source aviation business data obtained from external business systems includes at least one of the following elements: aircraft, routes, transportation, services, crew, and weather.

[0049] (i) Data source accurate to the field level

[0050] Referring to Tables 1(1)-1(5) below, the system obtains data from the following external systems. All elements are divided into domains according to business entities and labeled with primary keys and relationships.

[0051] Table 1 (1)

[0052] Table 1 (2)

[0053] Table 1 (3)

[0054] Table 1 (4)

[0055] Table 1 (5)

[0056] (II) Data Cleaning and Standardization Script Logic

[0057] In one embodiment, the data fusion module is specifically used for: Collect total fixed costs and full volume of multi-source aviation business data within a preset time frame (e.g., 24 months, 36 months, etc.); The full volume of multi-source aviation business data is preprocessed to obtain preprocessed full data. A flight-level wide table is established using the preprocessed data; the flight-level wide table uses the flight as the primary key and associates at least one of the following elements through entity mapping rules: aircraft, route, transportation, service, crew, and weather.

[0058] The data fusion module is specifically used to perform the following data preprocessing on the full amount of multi-source aviation business data: Flight number unification, timestamp alignment, numerical verification, missing data filling, and cold start processing.

[0059] Flight numbers are standardized: spaces and special characters are removed, and all are uppercase. If flight_id is empty, a temporary ID is generated using ac_reg + actual_dep_time, and data_quality_flag is set to 'ID_GENERATED'. data_quality_flag is a data quality flag field used to record any anomalies or special processing states that occur during the data cleaning process (such as automatically generated IDs, time anomalies, out-of-bounds values, etc.) for subsequent data traceability and quality monitoring.

[0060] Timestamp alignment: All timestamps are converted to UTC time. If actual_arr_time < actual_dep_time, it is considered a cross-day flight, and the duration is calculated by directly subtracting the timestamps. If the difference is less than 0 or greater than 24 hours, data_quality_flag is set to 'TIME_ERROR', and an attempt is made to retrieve the original ACARS gate opening / closing message from the operating system for repair.

[0061] Numerical verification: `block_hour` must be greater than 0, and the difference in minutes between `actual_arr_time` and `actual_dep_time` must be within ±5%. `cycle` (number of takeoff and landing cycles) must be 1; if it is 0, check the ACARS air-to-ground switch signal for repair; if it is still 0, mark it as abnormal. `pax_count` and `fuel_kg` cannot be negative. The air-to-ground switch signal is used to determine whether the aircraft is currently in the air or on the ground, and is a reliable data source for verifying whether takeoff and landing cycles have actually occurred.

[0062] Next, a static master data table is created. Using `ac_reg` as the primary key, it maintains immutable attributes such as aircraft type, engine type, and manufacturing date, with join completion performed during data cleanup. `ac_reg` must be able to be associated with the aircraft master data; `dep_airport` and `arr_airport` must form a valid route combination; otherwise, the process is suspended and an alert is issued.

[0063] When constructing the flight-level wide table, the static master data table is linked via ac_reg to complete the aircraft's inherent attributes (aircraft type, engine model, manufacturing date, etc.) into each flight record. The static master data table is one of the upstream dimension tables of the flight-level wide table.

[0064] (III) Missing Data Handling Strategy

[0065] In one embodiment, the missing data imputation process includes three strategies: causal inference imputation, autoencoder multimodal imputation, and baseline template replacement.

[0066] During implementation, the system automatically routes to one of three strategies based on the type of missing field, the missing rate, and the completeness of other fields for the flight.

[0067] Causal inference imputation: Based on the currently active causal graph, obtain the parent node data of the missing field. If the parent node data is not empty, input the parent node data into the online ridge regression model, output the estimated value of the missing field, add a random perturbation term to the estimated value, and obtain the imputed value of the missing field. Autoencoder multimodal imputation: If the parent node data is empty or there is no effective causal graph, all the non-missing variable values ​​of the current flight are used as input, and the complete vector is reconstructed through a pre-trained deep autoencoder. Missing fields are taken from the reconstructed complete vector. Baseline template replacement: When multiple specified core fields are missing data, template data is used to fill the gaps; the template data is a pre-calculated three-dimensional matrix of aircraft type, route and season, which includes the median of each characteristic variable of all flights under the combination of aircraft type, route and season.

[0068] Strategy A: Causal Inference Filling

[0069] Triggering condition: Triggered when all parent node data of the missing field X exists.

[0070] Implementation steps: Query the set of parent nodes Pa(X) of node X from the currently active causal graph G.

[0071] If X is a continuous variable, the system maintains an online ridge regression model M_X for X. M_X is retrained monthly using complete historical data, and its feature is Pa(X). For a currently missing record, the current value of Pa(X) is input into M_X to obtain the point estimate x_fill. In this embodiment, one node corresponds to one ridge regression model M_X, and the system maintains a set of models so that each continuous variable to be filled has its own dedicated online ridge regression model.

[0072] Add a random perturbation term to x_fill, the perturbation follows N(0, σ²), where σ² is the residual variance of M_X on historical data, to preserve uncertainty.

[0073] Write "CAUSAL_PREDICT" in the data quality tracking column fill_method_X. fill_method_X is the data imputation method marker column corresponding to field X, used to record the specific source of imputation for missing values ​​in this field, so as to facilitate subsequent data quality audits, anomaly tracing, and model iteration analysis.

[0074] Example: missing eng1_vibration(engine vibration value, IPS, continuous variable).

[0075] The cause-effect graph shows that Pa(eng1_vibration) = {csn_hour_eng1, takeoff_oat}.

[0076] M_X is a pre-trained linear model with the formula: vibration_fill = 1.2 + 0.0002 × csn_hour_eng1 + 0.01 × (takeoff_oat - 15), and the training residual variance σ² = 0.05.

[0077] Current record parameters: csn_hour_eng1 = 8000, takeoff_oat = 35.

[0078] Calculation process: x_fill = 1.2 + 0.0002 × 8000 + 0.01 × (35 - 15) = 1.2 +1.6 + 0.2 = 3.0 (IPS).

[0079] After adding the perturbation term, the final fill value might be 3.1. The flag `fill_method_eng1_vibration = "CAUSAL_PREDICT"` is set.

[0080] Strategy B: Autoencoder Multimodal Completion

[0081] Triggering conditions: Triggered when the parent node of the missing field is also missing, or when the cause-effect graph has not yet been established.

[0082] Model: A pre-trained deep autoencoder. The input layer covers all continuous variables in the flight wide table, and the output layer is the reconstructed complete vector.

[0083] Training data: All complete flight records from the past 24 months (missing rate <1%). During training, a portion of the input was randomly masked to allow the model to learn to reconstruct the full picture from partial observations.

[0084] Imputation process: All non-missing variable values ​​for the current flight are input into the encoder to obtain the latent spatial embedding, and then the complete vector is reconstructed through the decoder. The imputation values ​​for missing positions are directly taken from the reconstructed values. The flag is fill_method_X = "AE_RECONSTRUCT".

[0085] Strategy C: Baseline Template Replacement (Flight-Level Downgrade)

[0086] Triggering condition: When multiple core fields of a flight are missing simultaneously (core fields can be specified and configured by the user, such as the entire QAR file not being downloaded), the first two strategies cannot run reliably.

[0087] Template: A pre-calculated 3D matrix of "aircraft type-route-season". Key: (ac_model, dep_airport, arr_airport, month); Value: The median vector of all historical flight feature variables under this combination, excluding cost.

[0088] Application: This function replaces all missing and non-missing fields of the current flight with the template values, creating a standard "typical flight" dataset. The fill_method_X value for all fields in this flight is set to "TEMPLATE_REPLACE".

[0089] Post-processing correction mechanism: When the actual data is completed after T+n days (where "T" refers to the specific flight date), the system generates an adjustment record (adjustment_record). original_flight_id: Original flight identifier; original_period: The original period of allocation; template_allocation_amount: The allocated amount calculated using template data; actual_allocation_amount: The allocated amount calculated using complete real data; difference: the difference between the two (actual - template); This difference is not directly rolled over to the next month's total fixed cost F_total, but is instead recorded in a separate "smoothing_pool_balance" account. This account is a persistent state variable, and its update formula is as follows: smoothing_pool_balance_new = smoothing_pool_balance_old + Σ(actual_allocation_amount - template_allocation_amount); where actual_allocation_amount represents the allocation amount calculated for this flight using the completed actual data, template_allocation_amount represents the allocation amount previously calculated for this flight using template data, smoothing_pool_balance_old represents the smoothing pool balance carried over from the previous period to the current period (note: this balance is the accumulated value of historical differences, not the current period's total fixed cost F_total, the two are independent of each other), and smoothing_pool_balance_new represents the new balance of the smoothing pool after adjusting for all differences in the current period, which is used to carry over to the next allocation period.

[0090] A positive `smoothing_pool_balance_new` indicates an additional amount of over-allocation that needs to be absorbed. `smoothing_pool_balance_new > 0` means the historically accumulated over-allocation amount is positive; that is, the previously calculated allocation amount `template_allocation_amount` using template data is less than the subsequent calculated `actual_allocation_amount` using real data, and the accumulated difference is positive. This indicates that some costs were previously under-allocated, and this difference accumulates in the smoothing pool. "Absorption" means incorporating this historically under-allocated amount into the current deferred cost pool for distribution, thereby gradually reducing the balance in the pool and bringing it back to near zero. Conversely, if `smoothing_pool_balance_new < 0`, it means that there was previously over-allocation, and the pool has a negative balance, which needs to be offset by deducting deferred costs in the future. In each allocation period, a dedicated pool absorption amount `pool_absorption` is released from the pool and added to the current deferred cost pool (i.e., the total fixed costs actually allocated in this period). The amount released is dynamically determined by the smoothing network `NN_η`. This avoids confusion of inter-period responsibilities and keeps the fixed cost accounts clean for each period. Here, `pool_absorption` is the amount released from the smoothing pool and added to the current period's deferred costs (a positive value indicates an amount added to deferred costs, and a negative value indicates an amount deducted from deferred costs). It is one of the decision outputs of the smoothing network NN_η, used to control the rate at which the pool balance is absorbed.

[0091] (iv) Cold start treatment

[0092] In one embodiment, the cold start process includes: Missing fields for new aircraft are replaced with historical averages for the same aircraft type and engine model. Missing fields for new routes are replaced with data from similar older routes; these similar older routes are determined through feature similarity assessment of the takeoff and landing airports.

[0093] In practice, for new aircraft: the static attributes (age 0, aircraft type, engine type) are known. For dynamic attributes (TSN=0), simply enter 0. TSN (Total Service Hours) represents total service hours divided by self-renewal hours. Dynamic attributes also include CSN (Cycles Since New, total service cycles divided by self-renewal cycles) and TSO (Time Since Overhaul, service hours since overhaul). For fields requiring historical averages, use the average of the sub-fleet for the same aircraft type and engine type.

[0094] New Route: Obtain data on the airport pairs for the new route, such as the elevation and runway length of the two airports. Calculate the feature vector of the new route's airport pair (ICAO), denoted as the airport feature vector. Calculate the Euclidean distance between the airport feature vector of the new route and the airport feature vectors of all known routes. Select the K=3 closest existing routes. Use the union of the historical flight data of these 3 existing routes to construct a "pseudo-historical dataset" for the new route, used for calculating the causal and liability models for cold start. Mark route_cold_start = 'MIGRATED_FROM:US001,US002'.

[0095] In this embodiment of the invention, the definition and switching criteria for cold start state are as follows: For a route L, if the total number of actual flights executed by the route L is less than 200, then the route is in a cold start state.

[0096] Transition Strategy: As count_of_own_flights grows from 0 to 200, the system uses an asymptotic transition formula to generate the final liability value R_final for each flight on the route (R_final is the mixed marginal liability value, which needs to be translated and normalized before it can be converted into the initial allocated weight W_init for the flight): R_final=λ×R_pseudo + (1-λ)×R_real: Where R_pseudo is the responsibility value calculated based on transfer learning "pseudo-historical data", and R_real is the responsibility value calculated based on the actual data of the flight route itself. The mixing coefficient λ = max(0, 1 - count_of_own_flights / 200).

[0097] When the total number of real flights count_of_own_flights ≥ 200, λ = 0, completely exiting the cold start and using only real data.

[0098] (v) Element pool, dimension mapping and association logic

[0099] The system organizes hundreds of raw elements from various external business systems into a pool of candidate attributable drivers and maps them by business entity dimension to support automated analysis for subsequent causal discovery.

[0100] In one embodiment, candidate cost drivers in the factor pool are categorized into three types: Operational variables (which vary with the flight, such as the number of passengers, cargo and mail weight, flight hours, number of takeoff and landing cycles, and number of various services), aircraft route attributes (which are relatively fixed, such as aircraft type, aircraft age, maximum takeoff weight, and flight distance), and environmental variables (such as outside atmospheric temperature, density, and altitude).

[0101] The dimension-entity mapping relationship is shown in Table 2 below. Table 2 lists some of the candidate cost drivers for the feature pool.

[0102] Table 2

[0103] Combining Tables 1(1)-1(5) and Table 2, the association logic is as follows: Using `flight_id` as the core, a snapshot of the aircraft master data for the flight execution day is obtained via `ac_reg` (ensuring time accuracy). `dep_airport` and `arr_airport` are used to locate the route master data and obtain flight distances, etc. Detailed data such as passenger load, crew, and service usage are aggregated by `flight_id`. If a flight contains multiple flight sections, flight section-level information is retained using `flight_id + leg_code`. All entity associations are implemented through metadata configuration; adding new features or entities only requires updating the metadata, without modifying the core calculation logic.

[0104] (vi) Distributed Scope Filter

[0105] In one embodiment, the data fusion module is specifically used for: After preprocessing the full amount of multi-source aviation business data to obtain the preprocessed full data, the system receives the filtering conditions input by the user, filters the preprocessed full data according to the filtering conditions, and constructs a flight-level wide table subset. The data of the flight-level width table subset is self-checked; the data self-check includes empty set check, skewness check, and cost segmentation linkage; the skewness check includes checking that the ratio of any indicator data obtained by filtering to the total data of that indicator is lower than a set ratio; the cost segmentation linkage means that the amount corresponding to the flight-level width table subset is segmented from the total fixed cost.

[0106] The data fusion module includes an allocation range filter, which enables the aforementioned functions. The allocation range filter allows users to define a subset of data to participate in the allocation calculation before the automated process begins, resolving the issue that full allocation cannot meet the needs of specific analyses.

[0107] Filter configuration

[0108] The filtering criteria include any combination of flight attributes, route attributes, aircraft attributes, time attributes, and business tags.

[0109] For example, users can freely combine filtering criteria from the following dimensions through the system interface or configuration files: Flight attributes: Flight nature = "Regular flight" (excluding ferry flights, charter flights, and test flights), Flight type = "Passenger"; Route attributes: Route range = "Departure from Beijing", Route network = "European routes", Immigration / Exit identification = "International"; Aircraft attributes: Company = "Subsidiary A", Aircraft type = "B787"; Time attribute: Actual flight date BETWEEN '2026-03-01' AND '2026-03-15'; Business tags: Is it free? = Y, Is it a newly opened route? = N.

[0110] Filter execution process

[0111] The filter is placed after the data cleaning task and before the generation of the official flight-level wide table. After the full data cleaning and repair is completed, the system applies filtering conditions to filter the full data and generates an "amounted subset wide table". Automatic data self-checking is performed after execution. Empty set check: If the number of flights after screening is 0, the process will be immediately interrupted and an alarm will be issued.

[0112] Skewness check: Calculates the percentage of a subset of core statistical indicators such as flight volume, passenger volume, and flight hours relative to the total. If the percentage of a certain core indicator is lower than the set percentage, such as 5%, a warning is generated indicating that the user range may be incorrectly set, but the process is not interrupted.

[0113] Cost allocation linkage: After using a data subset, the costs to be allocated also need to be allocated synchronously. The system uses the "Deferred Cost Pool Calculator" to allocate the amount belonging to that subset from the total fixed costs based on the ratio of the sum of the basic responsibility values ​​of the subset flights to the sum of the basic responsibility values ​​of all flights in the entire month. The formula is: F_total_subset k =F_total k × ; in, F_total_subset k This represents the k-th fixed cost amount belonging to the current data subset after range partitioning. F_total k This is the total amount of the k-th fixed cost item for the current month, obtained from the ERP system. k This indexes fixed-cost items (such as engine overhaul reserve funds, fuselage insurance premiums, etc.), where i is the flight number in the current data subset, and j is the flight number in the full dataset. subset This represents a subset of flight data filtered by the allocated range filter, while allflight represents the full dataset of all flights for the current month. This represents the marginal liability value (i.e., the positive liability value after translation mapping) of flight i for the k-th fixed cost in the subset. This represents the marginal liability value (i.e., the positive liability value after translation and mapping) of flight j for the k-th fixed cost in the full dataset of Alflight.

[0114] The multi-dimensional allocation results only reflect the business situation within a subset, ensuring complete consistency in the analysis scope.

[0115] Business application scenarios: Supports specialized cost analysis (such as separate accounting for engine repair costs of the B787 fleet), independent accounting for specific route networks (separate accounting for international and domestic routes), and delineation of internal assessment boundaries (operating cost accounting for a subsidiary), etc.

[0116] Module 2: Causal Modeling Module

[0117] The causal modeling module consumes a historical flight-level wide table, trains cost-driven causal models offline, and checks the causal structure monthly to determine whether to update the causal graph.

[0118] The causal modeling layer, which corresponds to the causal modeling module, can include a variable standardization unit, a PC causal discovery unit with aeronautical constraints, a causal graph storage unit, and a causal structure drift detection unit. The implementation functions are described in detail below.

[0119] (a) Variable definition

[0120] The observation data is constructed using an element pool. The set of input variables V for the causal model comes from all candidate variables in the element pool, without prior screening, and specifically includes: Runtime variables: Examples include cycle, flight_hour, block_hour, pax_count, cargo_kg, fuel_kg, apu_runtime, and the number of times and duration of use for various services.

[0121] Aircraft / Route Attributes: For example, ac_age_year, ac_model, max_takeoff_weight, max_seats, land_distance, is_intl, etc.

[0122] Environment variables: Examples include takeoff_oat and density_alt_ft.

[0123] Cost variables (endogenous): For example, maint_cost_total (which can be attributed to the maintenance cost of an aircraft) can be used as a proxy target variable.

[0124] All variables are treated equally, and the causal model will automatically handle the correlations and redundancies between them. The causal model is used to construct a directed acyclic causal graph through causal discovery algorithms.

[0125] (II) Details of the Causal Discovery Algorithm

[0126] The causal discovery algorithm described above is the PC algorithm (Peter-Clark), and the algorithm selection is: PC algorithm + conditional independence extension.

[0127] The causal modeling module is specifically used to: perform Fisher's Z test for conditional independence for continuous variables and G2 test for conditional independence for discrete variables. The significance level α = 0.01.

[0128] Constraint Filtering: Hard aviation domain knowledge constraints are introduced to prohibit causal edges in specific directions. These constraints include temporal and hierarchical constraints; temporal constraints prohibit future variables from influencing past events; and hierarchical constraints prohibit flight-level variables from pointing to aircraft-level attributes, such as passenger numbers not affecting maximum takeoff weight.

[0129] The causal modeling module is specifically used to: based on the PC algorithm and superimposed with aviation domain knowledge constraints, traverse all variable pairs, determine the causal relationship between variables, iterate repeatedly until the number of edges converges, and obtain a directed acyclic causal graph.

[0130] Execution process: Starting from a fully connected undirected graph, traverse all variable pairs. If a d-separation set exists such that the pairs are conditionally independent given that set, delete the edges between them. Iterate repeatedly until the number of edges converges. In the orientation phase, the direction of the edges is determined using the V-structure (A1→A3←A2) and constraint rules. A d-separation set is defined as follows: given a set of condition variables Z, if all paths between variables A and B are blocked by nodes in Z, then Z is called a d-separation set of A and B. In the PC algorithm, the system traverses and searches in ascending order of the size of the condition variable set; the first set of variables that makes A and B conditionally independent is the d-separation set of that variable pair.

[0131] Output: A complete pairwise complete partially directed acyclic graph (PDAG), i.e., a complete causal graph G. All edges have causal effect coefficients (estimated through subsequent linear structural equations or nonparametric regression). These coefficients quantify the strength and direction of the influence of parent node variables on child node variables, and are estimated through subsequent linear structural equations or nonparametric regression.

[0132] Automatic feature filtering: The cause-effect graph directly reveals which factors have a direct, non-redundant impact on costs. All nodes pointing to costs and their upstream nodes constitute the cost driver feature set for that cost item. The problem of manually debating "number of people or weight" in traditional methods is automatically resolved by the algorithm here—for example, the algorithm will find that the real drivers of engine maintenance costs are csn_hour_eng1 and takeoff_oat, while the number of passengers is automatically excluded only because it is related to the route distance.

[0133] (III) Causal Structure Drift Detection

[0134] In one embodiment, the causal modeling module is specifically used for: The historical element pool is segmented to create new and old datasets; Construct two causal graphs using the new dataset and the old dataset respectively, and calculate the Structural Hamming Distance (SHD) of the edge sets of the two causal graphs. Compare the SHD with a set threshold to determine whether the causal structure has changed statistically significantly; If a statistically significant change occurs, the currently active causal graph will be updated.

[0135] During implementation, after new data (the entire previous month) is appended to the database each month, a stability test is performed: Windowing: Slide a window over historical data, using the most recent 6 months as "new data" D_new and the previous 18 months as "old data baseline" D_old.

[0136] Difference calculation: Run the PC algorithm independently on D_old and D_new to obtain causal graphs G_old and G_new.

[0137] Edge set comparison: Calculate the Structural Hamming Distance (SHD) between the edge sets of two graphs = number of newly added edges + number of deleted edges + number of reversed edges.

[0138] The set threshold is updated periodically in the following manner: Block resampling is performed on the feature pool data from the most recent months to generate a specified number (e.g., 24) of pseudo-samples of the same size; For each pair of pseudo-samples, run the PC algorithm separately to calculate the SHD between them and form the zero distribution of SHD; The 95th percentile of this distribution is used as the set threshold.

[0139] For example, during system initialization, a block-based Bootstrap resampling of the historical 24-month dataset is performed to generate 100 pseudo-samples of the same size. A PC algorithm is run on each pair of pseudo-samples to calculate the relative SHD (Signal Variable Hierarchy) between them, forming a null distribution of SHD (i.e., a natural fluctuation distribution under the same data generation process). The 95th percentile of this distribution is taken as the initial threshold SHD_threshold. After the system goes live, the above simulation is repeated every quarter, updating the threshold with the latest 24-month data to adapt to changes in fleet size, data quality, and operational patterns.

[0140] Decision-making rules: If SHD == 0, the structure is completely stable and no update is triggered.

[0141] If 0 < SHD <= SHD_threshold, it indicates a slight drift, an alarm is recorded, but G_old is still used.

[0142] If SHD > SHD_threshold, it indicates that the causal structure has undergone statistically significant changes, triggering offline causal graph retraining.

[0143] Retraining: The PC algorithm is retrained using the full dataset from the past specified number of months. A new causal graph, denoted as G_vNext, is generated using the retrained PC algorithm. After machine learning engineers review the new edges for validity, G_vNext is marked as current, indicating the currently active causal graph. The old version of G is archived to the causal graph archive repository causal_graph_archive.

[0144] Module 3: Responsibility Quantification Module

[0145] Based on the currently active cause-effect graph, this responsibility quantification module automatically determines the driver feature set for each fixed cost and calculates the "basic marginal responsibility value" for each flight for each fixed cost.

[0146] The responsibility quantification layer corresponds to the responsibility quantification module, which specifically includes a cost-motivation matching unit, a cost prediction model training unit, a SHAP marginal responsibility value calculation ternary unit, and an initial weight generation unit. The implementation functions are described in detail below.

[0147] (a) Automatic matching mechanism of cost and motivation

[0148] In one embodiment, fixed costs include multiple fixed expense items; the causal modeling module is used to: construct observation data using an element pool, introduce aviation domain knowledge constraints, and construct a directed acyclic causal graph for each fixed expense item using a causal discovery algorithm; the responsibility quantification module is used to: obtain the target cost driver for each fixed expense item through the causal graph of each fixed expense item, input the target cost driver for each fixed expense item into the fixed cost prediction model, call machine learning interpretability tools to attribute the predicted value to each target cost driver, and obtain the marginal contribution value of each target cost driver to the predicted value; and determine the initial allocation weight of each flight's consumption of each fixed expense item using the marginal contribution values ​​of all target cost drivers to the predicted value.

[0149] Specifically, when faced with hundreds of fixed cost items, the system does not require manual configuration of allocation rules for each item; instead, it automatically completes the matching through a unified intelligent pipeline. Cost clustering and grouping: First, all fixed cost items are grouped into several cost pools based on accounting categories, such as "Airframe Maintenance Reserve Fund", "Engine Overhaul Reserve Fund", "APU Maintenance Fee", "Aircraft Insurance Fee", "Landing and Takeoff Service Amortized Fixed Fee", and "Flight Crew Fixed Salary". For cost items with large amounts or special characteristics, they can be separated into an independent cost pool.

[0150] The causal discovery algorithm runs independently for each cost pool: the system uses maintenance costs attributable to flights as proxy target variables, and feeds them into the causal model along with all candidate variables in the factor pool. The algorithm automatically selects a subset of factors from hundreds of factors that have a direct causal impact on that cost pool.

[0151] Output the cost driver feature set for each fixed cost item: for example, the cost driver feature set for "engine overhaul reserve fund" might be {csn_hour_eng1, takeoff_oat, avg_takeoff_n1, csn_cycle_eng1}; while the cost driver feature set for "landing service apportionment fixed fee" might be {block_hour, cycle, apu_runtime, is_intl}. These feature sets are entirely data-driven and not manually preset.

[0152] Unified computational framework: Each cost pool has its own independent prediction model and SHAP interpreter, but shares the same computational framework.

[0153] (ii) Feature Engineering

[0154] Based on the causal graph G and the matching results described above, the system automatically collects its feature set for each cost pool. Example: Engine maintenance costs: Total hours since overhaul of engine 1 (csn_hour_eng1, i.e., Engine-1 CyclesSince Overhaul in Hours), Total number of cycles since overhaul of engine 1 (csn_cycle_eng1, i.e., Engine-1 Cycles Since Overhaul in Cycles), Takeoff ambient temperature (takeoff_oat), and Average Takeoff N1 Speed ​​(avg_takeoff_n1, i.e., Average Takeoff N1 Speed).

[0155] Aircraft maintenance costs include: total service cycles of aircraft (tsn_cycle), aircraft age (ac_age_year), and landing airport pressure altitude (landing_pressure_alt).

[0156] APU maintenance costs include: APU runtime (apu_runtime, i.e., Auxiliary Power Unit Runtime in Minutes) and dep_airport_pressure_alt (i.e., Departure Airport Pressure Altitude).

[0157] In one embodiment, the fixed cost prediction model is a tree-based machine learning model, such as an XGBoost regressor; the marginal contribution value includes a SHAP value; and the machine learning interpretability tool is a SHAP interpreter. The responsibility quantification module is specifically used for: For multiple fixed cost items actually incurred in a single flight in batch multi-source aviation business data, the target cost driver for each fixed cost item is obtained through the cause-effect graph. Input the target cost drivers of a single fixed cost item into the fixed cost forecasting model, and call the SHAP interpreter to calculate the SHAP value of each input target cost driver to the forecast value. The sum of the SHAP values ​​of all input target cost drivers to the predicted values ​​is determined as the marginal responsibility value for a single flight for the single fixed cost item; A translation mapping is applied to the original marginal responsibility values ​​of all flights in the batch multi-source aviation business data to obtain the translated marginal responsibility values; the marginal responsibility values ​​are used to quantify the causal contribution of a single flight to the consumption of a single fixed cost item; The initial allocation weight of fixed cost consumption for each flight is determined by the ratio of the marginal responsibility value after a single flight translation mapping to the sum of the marginal responsibility values ​​after all flights translation mapping.

[0158] (III) Training of the cost prediction model

[0159] Model: illustrated with an example of the XGBoost regressor.

[0160] Target variable Y: maint_cost_total. This is the actual maintenance cost obtained from maintenance work orders linked by flight_id, which can be directly attributed to the flight. It serves as a proxy variable for the "actual consumption of fixed costs" that cannot be directly obtained.

[0161] Feature X: The set of cost driver features obtained in the previous step.

[0162] Training: The model is trained using data from the past 24 months, minimizing the MSE. Five-fold cross-validation is used to prevent overfitting, and the final output is a model M_k.

[0163] (iv) Calculation of Flight Liability Value

[0164] For each flight i executed last month, input its cost driver feature value x_i into model M_k.

[0165] Call SHAP's TreeExplainer to calculate the Shapley value Φ_j,i of feature j for the predicted value f(x_i). Here, x_i represents the cost driver feature vector of flight i (i.e., the complete set of all cost driver feature values ​​for that flight), for example, x_i = [csn_hour_eng1=8200, csn_cycle_eng1=2100, takeoff_oat=28, avg_takeoff_n1=92.5]. Feature j represents the j-th feature component in the vector x_i, for example, j=1 corresponds to csn_hour_eng1, j=2 corresponds to csn_cycle_eng1, and so on.

[0166] The original marginal liability value of flight i for fixed cost item k is defined as follows: ; in, Features This represents the set of cost driver characteristics.

[0167] This invention proposes a marginal responsibility value: based on counterfactual inference logic, it quantifies the causal contribution of each flight to a certain fixed cost, which can be positive or negative. A positive value indicates that the combination of operational characteristics of that flight causes cost consumption to be higher than average. The marginal responsibility value is the initial allocation weight of fixed cost consumption for each flight.

[0168] This value can be positive or negative, representing the increase or decrease in cost relative to the baseline forecast, reflecting the marginal cost responsibility. A positive value indicates that the combination of operational characteristics of the flight results in the expected consumption of this fixed cost being higher than average, while a negative value indicates it being lower than average.

[0169] To obtain a strictly positive allocation base, a translation mapping is applied to the original marginal liability values ​​of all flights within the batch: ;in, Let i be the marginal liability value after shifting and mapping fixed cost item k. Let i be the original marginal liability value for fixed cost item k. The minimum value is taken from the original marginal responsibility values ​​of all flights on the same route within the batch of multi-source aviation business data. `batch` represents the set of all flights on the same route within the batch of multi-source aviation business data. p This indicates the index for traversing all flights within a batch.

[0170] This translation mapping does not change the relative differences between flights.

[0171] The initial allocation weights are calculated using the marginal liability values ​​of flights after the translation mapping is applied, according to the following formula: ; in, This represents the initial allocation weight of flight i to fixed cost item k. allflight_j This subset, obtained by removing flights with no available data for cost item j from the allflight dataset, ensures that the sum of the initial allocation weights for all flights consuming the same fixed cost item is 1. Different fixed cost items have different cost driver feature sets; for example, "engine maintenance cost" requires data for engine #1, while "APU maintenance cost" requires APU operational data. Some flights (such as those without APUs) may lack cost driver data for APU maintenance costs and should not participate in the allocation of this cost. Therefore, allflight_j It is a subset obtained by removing flights for which no data is available for cost item j, based on allflight.

[0172] Module 4: Amount Allocation Optimization Module

[0173] The allocation optimization module is used to optimize and adjust the initial allocation weight of each flight to obtain the effective allocation weight of each flight.

[0174] The amortization optimization layer, or amortization optimization module, can include a state space construction unit, a PPO reinforcement learning engine, an independent smoothing network, a multi-objective reward function calculation unit, and a final weight generation unit. The implementation functions are described in detail below.

[0175] In one embodiment, the allocation optimization module uses the initial allocation weights as a basis, and through the reinforcement learning agent guided by the multi-objective reward function, outputs adjustment actions and uses an independent smoothing network to dynamically determine the mixing coefficient of the old and new weights to generate the final allocation weights.

[0176] The allocation optimization module is specifically used to: obtain the initial allocation weight of the current period's flights, the remaining amount in the smoothing pool, the statistical data after allocation in the previous period, and historical allocation feedback to construct the current system state; optimize and adjust the initial allocation weight of each flight using the current system state to obtain the effective allocation weight of each flight; the smoothing pool stores the difference between the first allocation amount calculated using template data and the second allocation amount calculated using completed real data; the template data is the data after filling in the missing fields using template data.

[0177] The amortization optimization module is specifically used for: Based on the current state of the system, the pre-trained reinforcement learning agent outputs adjustment actions under the guidance of a multi-objective reward function, thereby obtaining the weights after the reinforcement learning actions are adjusted. The adjustment actions include a uniform multiplier of the initial allocation weights of all flights within the same route. The multi-objective reward function includes at least one of the following: financial compliance reward, management decision usefulness reward, route-level allocation smoothness reward, and human-machine feedback reward. Based on the current state of the system, a pre-trained smoothing network is used to predict the output smoothing mixing coefficient; this smoothing mixing coefficient is used to balance the change in response speed and the stability of financial allocation; the smoothing network is a supervised training neural network. The weights adjusted for reinforcement learning actions are combined with the effective allocation weights from the previous period using a smoothing mixing coefficient to obtain the effective allocation weights for each flight in the current period; the previous period refers to the previous allocation cycle, which is, for example, once a month.

[0178] (a) Design of state space S_t

[0179] For example, before the start of each amortization period t (month), the state vector received by the agent is concatenated from the following parts: Flight aggregation statistics (aggregated by route): mean and standard deviation of the basic responsibility (i.e., the marginal responsibility value applied by translation mapping) for all flights on the same route in the current period; final allocation amount for the previous period; revenue, variable costs, and marginal contribution for the previous period; ranking of total revenue profit margin for the previous period (percentile). Among these, revenue for the previous period represents the net revenue from passenger and cargo transport generated by the flight, taken from the revenue_total field of the settlement system; variable costs represent fuel costs, landing fees, and route fees directly incurred with the flight, taken from the variable_cost_fuel / airport / route fields of the settlement system; and marginal contribution is the balance after subtracting variable costs from revenue, used to measure the route's ability to cover fixed costs. Here, "marginal contribution" is a financial indicator and is unrelated to the feature contribution value Φ output by the SHAP interpreter in the responsibility quantification module.

[0180] Global statistics: Total fixed costs F_total for the current period and F_total^k for each item; Kendall Tau-b coefficients for the marginal contribution ranking and natural profit ranking (ranked by revenue after deducting variable costs) of all routes after the previous period's allocation; mean Z-Score for the smoothness of the previous period's allocation. The "marginal contribution ranking of routes" is based on the financial marginal contribution amount from largest to smallest, not on the sum of SHAP values; the "natural profit ranking" refers to the profit level ranking excluding allocated fixed costs, i.e., ranking only by the marginal contribution (revenue - variable costs) of the routes. The Kendall Tau-b coefficient is a statistical indicator that measures the consistency between two ranking sequences, with a value range of [-1, 1]. The Z-Score is a standard score calculated for the year-on-year change rate of the allocated amount for each route, used to measure whether the overall fluctuation of the previous period's allocation results is abnormal.

[0181] Financial feedback statistics: the frequency of allocation errors reported for each route in the previous period, such as the frequency of the "mark error" button being pressed.

[0182] Smoothing pool state: smoothing_pool_balance, which is the current balance of the smoothing pool for amortized error. This feature allows the agent to perceive the amount of historical error that needs to be digested.

[0183] (ii) Design of Action Space A_t

[0184] The action is a vector with a dimension equal to the number of routes involved, N_routes. Each element δ_l, in the range (0.8, 1.2), represents a uniform multiplier for the weights of all flights on that route.

[0185] (III) Detailed calculation of the reward function R_t

[0186] After each action is executed and a share of the results is generated, the environment calculates the following reward: (1) Financial compliance reward r_fin: If abs(Σallocated - F_total) > 1 yuan: r_fin = -1.0. Here, alllocated represents the total allocated amount for all flights (i.e., the sum of the allocated amounts Alloc_flight_i for each flight); If the deviation is ≤ 1 yuan: r_fin = +0.5.

[0187] (2) Management decision-useful reward r_mgt: Calculate the Kendall Tau-b coefficients for the ranking of "marginal contribution rate" and "naturally reasonable" ranking (revenue ranking after deducting variable costs) of all routes under the new allocation results.

[0188] r_mgt = 1.0 × (Tau - 0.5). The theoretical range for r_mgt is [-0.5, 0.5]. Here, Tau is the calculated Kendall-Tau-b correlation coefficient.

[0189] (2) Smoothness reward r_smooth: For each route L, calculate its month-on-month change rate of allocated amount: change_L = |A_L,t - A_L,t-1| / A_L,t-1. Where A_L,t and A_L,t-1 are the allocated amounts for the current period t and the previous period t-1, respectively.

[0190] For all routes, calculate the sum of the rates of change exceeding the threshold (15%): penalty = sum(max(0, change_L - 0.15)). The penalty is the sum of the excess fluctuations of all routes exceeding the tolerance threshold, used to quantify the overall volatility of the current period's allocation result relative to the previous period. The larger the penalty, the more routes or the greater the jump in allocation amount.

[0191] r_smooth = -0.4 × penalty.

[0192] Additional rule: No penalty will be imposed on cold start routes (first 3 months), allowing their market share to quickly reach a stable state.

[0193] (3) Human-computer feedback reward r_human: If an expert marks the route as "disagreeing," the route receives an additional penalty of -2.0.

[0194] Total reward: R_t = r_fin + r_mgt + r_smooth + r_human. All components are designed to be of similar magnitude (approximately [-2, 2]) to avoid any one objective dominating the optimization.

[0195] (iv) Training and Execution Algorithms

[0196] The Actor-Critic structure (a concrete implementation of reinforcement learning agents): The Actor network takes the state S_t as input and outputs the mean action μ and standard deviation σ, used to sample A_t. The Critic network takes S_t as input and outputs an estimate of the state value V(S_t). μ and σ are "distribution parameters," and the unified multiplier is "the specific action value sampled / taken from this distribution." The former determines the range and tendency of the latter's possible values; both are "parameter layer" and "business layer" representations of the same process.

[0197] Offline training: Using rolling historical data from the past 36 months, a simulated environment is constructed month by month. The agent runs in the simulated environment cycle by cycle, collecting trajectories, and using the PPO algorithm to update network parameters to maximize long-term cumulative discount rewards. In this embodiment of the invention, the PPO algorithm is implemented based on the Actor-Critic architecture, where the agent = Actor network + Critic network + PPO optimization algorithm + experience replay mechanism.

[0198] Online inference: When the actual billing is completed each month, the pre-trained Actor network is loaded, the actual state S_t of the month is input, and the Actor outputs a deterministic action A_t (using the mean μ directly without sampling), which is used to generate the final allocation voucher.

[0199] Online learning (human-machine closed loop): "Objection" feedback submitted by business experts will be converted into a high negative reward -2.0 in the next cycle. This experience is stored in a high-priority experience replay pool. During the monthly PPO offline fine-tuning update, the proportion of samples from this high-priority pool is increased to 50%, forcing the policy network to quickly learn and correct the action patterns that lead to unsatisfactory allocation.

[0200] (v) Dynamic smoothing factor and final weight formula

[0201] Concise definition of weight transition relationship: Let the initial weight of subject k for the current month obtained through SHAP be: The target weights after action adjustment in reinforcement learning are: .

[0202] ; Where i is the flight index, j represents the index when traversing all flights (in the same domain as i, used to sum all flights in the denominator), and L(i) is the route to which flight i belongs (used to map the flight to the dimension of the route it flies). It is a unified multiplier (range 0.8~1.2) output by the reinforcement learning agent for route L(i), which is applied to the initial allocation weight of all flights within the route.

[0203] The smoothing network NN_η is a small neural network that shares the underlying state feature extraction layer with the PPO Actor, but has an independent output head. Its output is compressed to the [0,1] interval by the sigmoid function and serves as the smoothing mixing coefficient η.

[0204] In one embodiment, the amortization optimization module is specifically used for: The effective weights for each flight are obtained by fusing the weights adjusted for reinforcement learning actions with the initial allocated weights using the following formula and smoothing mixing coefficient: W eff =η×W target +(1-η)×W old ; Among them, W eff For the weighting to take effect, W target To reinforce the weights of the learned actions after adjustment, W old The final weights actually used in the previous period (W when used for the first time) old Degenerate into W init η is the smoothing mixing coefficient.

[0205] In this embodiment, when the system services undergo structural changes, the smoothing network predicts a smoothing mixing coefficient close to 1; when the system services operate smoothly, the smoothing network predicts a smoothing mixing coefficient close to 0.

[0206] For example, when the smoothing network NN_η detects a structural change in the business (such as the introduction of a new route or aircraft replacement), it outputs an η close to 1 to allow the weights to quickly respond to the current cost drivers; while when the business is running smoothly, it outputs an η close to 0 to maintain the continuity of the reports. When the smoothing pool balance is high, NN_η learns to appropriately increase pool_absorption to gradually absorb historical errors.

[0207] In one embodiment, the amount released and absorbed by the smoothing pool in the current period is predicted and output by the auxiliary output head of the smoothing network; the amount released and absorbed in the current period will be added to the fixed cost of the current period for allocation.

[0208] In practice, the amount of absorption released by the smoothing pool in this period, pool_absorption, is also determined by the auxiliary output head of NN_η (a non-negative value). This amount will be added to the total fixed cost pool for this period for allocation, and deducted from smoothing_pool_balance. This gives the network the ability to simultaneously decide on the "degree of weight smoothing" and the "speed of error digestion", achieving decoupled optimization of "exploration-stabilization".

[0209] Module 5: Voucher Generation Module

[0210] The voucher generation module is used to: calculate the fixed cost allocation at the flight level using the effective allocation weight of each flight, and obtain the allocation result; generate a financial voucher file using the allocation result; and store the allocation result in the financial voucher file in the form required by financial regulations.

[0211] The voucher generation module includes a voucher generation and push module, a multi-dimensional allocation calculation unit, a financial reconciliation verification unit, a standard voucher generation unit, and an ERP interface integration unit. The functions are described in detail below.

[0212] In one embodiment, the voucher generation module includes the following two triggers: the first trigger is used to receive closing dynamics sent by an external financial business system and generate a first trigger signal; the second trigger uses an internal timed task to periodically check whether fixed cost and multi-source aviation business data cover all executed flights of the previous month, and if the coverage exceeds a specified ratio, it generates a second trigger signal. When the first and second trigger signals of the month are present simultaneously, the allocation task workflow is started; the allocation task workflow automatically mobilizes the data fusion module, causal modeling module, responsibility quantification module, allocation optimization module, and voucher generation module to start running.

[0213] In one embodiment, the credential generation module is specifically used for: After obtaining the allocation results, the allocation results are aggregated in multiple dimensions. Fixed costs are used to perform financial reconciliation verification on the multi-dimensional aggregated allocation results. After the financial reconciliation verification is passed, financial voucher files are generated. The multi-dimensional aggregation includes any combination of flight dimension, route dimension, flight season dimension, aircraft dimension, and aircraft type dimension.

[0214] (a) Triggering Mechanism

[0215] Two independent triggers exist simultaneously: Trigger A (Strong Dependency): After the monthly closing process is completed and the fixed cost account balance is determined, the external financial business system, such as an ERP system, sends an HTTP POST message to this system via its workflow engine: {"event":"COST_CLOSED", "period_month":"2026-03"}. After the system API gateway verifies the signature, it writes the first trigger signal to the trigger_signal table.

[0216] Trigger B (weak dependency): An internal scheduled task that checks at 2:00 AM on the 5th of each month whether the data in systems such as QAR and maintenance have covered all executed flights of the previous month. If the coverage exceeds 99.5%, a second trigger signal is written to the trigger_signal table.

[0217] Execution Conditions: The Orchestrator scheduling engine will only start the allocation pipeline after both triggers A and B for the current month have "ready" records in the trigger_signal table. The Orchestrator is the workflow scheduling engine in this embodiment of the invention, responsible for managing and coordinating the execution order and dependencies of each task module in the allocation pipeline. It acts as the "commander-in-chief" of the entire allocation process: when the trigger conditions are met, it sequentially calls modules such as data fusion, causal modeling, responsibility quantification, allocation optimization, and voucher generation according to a predefined DAG (Directed Acyclic Graph) definition, monitors the status of each task, and supports failure retries and timeout alarms.

[0218] (II) Automated Production Line Steps (Airflow DAG Definition)

[0219] Task 1: Data Quality Verification and Repair

[0220] Run the Spark job in parallel to execute cleaning scripts and missing value handling strategies on the raw, multi-source aviation business data. Output qualified data to the Flights_Wide_Table. Write a data quality report; if the percentage of flights replaced by templates exceeds 1%, immediately push an alert.

[0221] Task 2: Splitting the Flight Class Width Table Subset

[0222] Execute the allocation range filter, apply user-preset filtering conditions, and generate a "flight-level wide table subset". Perform empty set and skewness checks; if they fail, interrupt or issue an alarm.

[0223] Task 3: Cause-effect graph structure drift detection

[0224] Input last month's data and run the causal graph structure drift detection script. If retraining is not triggered, skip directly; if it is triggered, automatically submit a model training task to the GPU cluster and suspend subsequent tasks until the new model is ready. Increment the model version number by 1.

[0225] Task 4: Marginal Responsibility Value Calculation

[0226] Load the current effective version of the cause-effect graph G_current and determine the cost driver characteristics. Read last month's data from a subset of the flight-level width table and calculate the basic responsibility value R_pos (marginal responsibility value after translation mapping) and initial allocation weight W_init for each flight.

[0227] Task 5: Optimizing Reinforcement Learning Weights

[0228] Load the trained PPO Actor model and NN_η smoothing network model. Construct the state S_t. Perform one forward propagation to obtain the action A_t and the smoothing factor η. Calculate the effective amortized weight W_eff.

[0229] Task 6: Allocation Calculation and Reconciliation Check

[0230] Input W_eff and the total fixed cost F_total^k for each account retrieved from the ERP (or F_total_subset^k after subset splitting if a range filter is used). Calculate the three-tier allocation: Flight: Alloc_flight_i = W_eff_i × F_total; Reconciliation check: Execute an assertion: abs((sum(Alloc_flight) - F_total) / F_total) < 0.001. If the assertion fails, immediately terminate the pipeline and report a fatal error.

[0231] Task 7: Voucher Generation and Push

[0232] Read the financial voucher template configuration (debit accounts, credit accounts, auxiliary accounting dimensions, etc.). Populate the allocation results into the template, generate journal entries line by line, and package them into a standard XML file. Upload the file to the ERP's voucher interface directory via SFTP (SSH file transfer protocol) and call the ERP's "Import Standard Vouchers" API.

[0233] Task 8: Monitor ERP Status

[0234] Poll the ERP's batch_job_id (the batch_job_id is the asynchronous task identifier returned by the ERP system after receiving the voucher push request, used to subsequently query the execution status of the voucher import task) status every 60 seconds, with a timeout of 30 minutes. If successful, send a notification to the finance team stating "Allocation complete, voucher number XXXX"; if it fails, automatically download the failure log and parse the error code. For data format errors, attempt to correct and retry once; for business rule errors, trigger a manual alert.

[0235] (III) Generation of Multi-Dimensional Allocation Results

[0236] The final weight W_eff is directly assigned to each flight_id, and the system automatically outputs the amortization results across the following dimensions through aggregation: Flight-level: Fixed costs for each flight are directly recorded in the accounts.

[0237] Flight segment dimension: If a flight includes multiple flight segments, the flight allocation is split to each flight segment identifier (leg_code) according to the flight segment operating time or take-off and landing cycle ratio.

[0238] Route dimension: The sum of the cost-sharing for all flights under the same route code (route_code), used for route benefit analysis.

[0239] Aircraft / Registration Number Dimension: The sum of the allocated costs for all flights operated by the same aircraft registration number (ac_reg) constitutes the single-aircraft fixed cost card.

[0240] Aircraft type dimension: Summarized by aircraft type (ac_model) for fleet cost comparison.

[0241] All dimensions are subject to mandatory cross-checking, and the sum equals the total fixed cost. Any allocated amount can be drilled down to trace back to the specific flight's contributing factors and original operational data.

[0242] Module Six: Monitoring and Online Feedback

[0243] The monitoring and feedback layer may include a monitoring and online feedback module, a dynamic confidence differentiation engine, a multi-dimensional anomaly detector, a distributed explanation report generation unit, a high-priority expert feedback experience pool, a model fine-tuning scheduling unit, etc. The specific implementation functions are described below.

[0244] In one embodiment, the system further includes a monitoring feedback module for monitoring one or any combination of the following: Monitoring for discrepancies in total allocated costs: If abs(sum of allocated costs for all flights - total fixed costs) > 1.0, immediately block the generation of voucher files and issue an alert to the designated terminal; the total fixed costs refer to the total fixed costs in the ERP general ledger. Single route cost allocation jump: If the cost allocation of a single route exceeds the dynamic confidence interval and there is no major event label, an alert is issued to the designated terminal; The monitoring and feedback module: By calculating the exponentially weighted moving average and exponentially weighted moving variance, a dynamic confidence interval is maintained for the fixed cost of each route each month; Cost driver structure drift: Calculate the percentage of the marginal contribution of a single target cost driver in the sum of the marginal contributions of all target cost drivers. If the difference between this month's percentage and the average of the past few months exceeds the set percentage and there is no business explanation, issue an alert to the designated terminal. Reversal of profitability indicators: Monitor changes in profitability indicators after allocation for a single route. If the profitability indicator of a single route suddenly changes from profit to loss after several consecutive months, issue an early warning to the designated terminal.

[0245] (a) Dynamic Confidence Interval Engine

[0246] Objective: To maintain a dynamic confidence interval for the fixed costs allocated to each route each month. For example, to establish a confidence interval for the "fixed cost allocation per 10,000 yuan of transport turnover" (Unit_Cost_L,t) for each route each month. This indicator (Unit_Cost_L,t) is the "fixed cost per unit of transport turnover," which is the total fixed cost allocation for route L in period t divided by the transport turnover (e.g., ton-kilometers or revenue ton-kilometers) of that route during the same period. It is used to measure the cost efficiency of the route and facilitates comparisons between routes with different transport volumes.

[0247] Method: For each route L, a tracking window of the past 12 months is maintained. The exponentially weighted moving average EWMA_t and the exponentially weighted moving variance EWMVar_t are calculated online, with a smoothing coefficient α=0.3. The upper and lower limits are [EWMA_t - 3 × sqrt(EWMVar_t), EWMA_t + 3 × sqrt(EWMVar_t)].

[0248] Adaptive bandwidth adjustment: Before performing the comparison, the system first checks whether there are any major event tags for the route in the previous month (such as aircraft type change, flight volume change exceeding 30% month-on-month, or any major maintenance closure record at the route's departure and arrival airports). If a major event exists, the multiplier factor in the upper and lower limit formula is temporarily increased from 3 to 8 to avoid misjudging normal business adjustments as errors in the allocation system itself.

[0249] (II) Multidimensional Anomaly Detector

[0250] Detector 1: Total allocated amount discrepancy. Logic: abs(sum of allocated amounts for all flights - total fixed costs) > 1.0. Level: BLOCKER. Action: Immediately block voucher generation and notify core developers.

[0251] Detector 2: Single-route cost sharing jump. Logic: Calculate Unit_Cost_L; if it exceeds the confidence interval and there are no major event labels, the level is CRITICAL. Action: Generate an alarm, including a trend chart of the route over the past 12 months.

[0252] Detector 3: Cost Driver Structure Shift. Logic: Calculate the percentage contribution of each cost driver (e.g., cycle number) to the total allocation across all flights. If the difference between this month's percentage and the average of the past 6 months exceeds 20%, and there is no business explanation, the warning level is WARNING. Action: Indicates potential overfitting in the causal model; data scientists are advised to review it.

[0253] Detector 4: Profitability Indicator Reversal. Logic example: If a route has been among the top 10 most profitable routes after accounting for four consecutive months, but suddenly falls into the bottom 10 most unprofitable routes this month, simultaneously check its operating data (load factor, ticket price, variable cost ratio). If there is no significant deterioration, the level is: WARNING. Action: Push a message to the route manager.

[0254] (III) Human-Machine Closed-Loop Feedback and Distributed Explanation

[0255] In one embodiment, the monitoring feedback module is specifically used to: provide an allocation explanation report while issuing an early warning to a designated terminal; the allocation explanation report includes Top N positive impact drivers and Top N negative impact drivers statistically determined by marginal contribution values; N is a positive integer.

[0256] When the system pushes an allocation anomaly alert to business experts, it simultaneously provides an allocation explanation report, automatically generated by a pre-trained model explanation engine. The report may include the following: (1) The total fixed cost allocation for this route this month, and the change compared to last month.

[0257] (2) Top 3 positive impact factors: For example, “1. The number of engine cycles is higher than the average of the route, resulting in a 12% increase in the maintenance reserve fund allocation” and “2. The take-off and landing airports are at higher elevations, increasing the responsibility for engine hot end consumption.”

[0258] Top 3 factors with the greatest negative impact: For example, "1. The aircraft age is lower than the fleet average, resulting in an 8% reduction in fuselage insurance coverage."

[0259] This explanation directly utilizes the intermediate calculation results of SHAP values: after flights are aggregated to routes, the Shapley values ​​of each driver are summed and sorted to make the allocation basis transparent and verifiable.

[0260] There are two key buttons on the alarm card: [Accept, reasonable allocation]: After clicking, the system records a feedback event, type='ACCEPT', with a reward value of 0, which enhances the model's confidence in the current strategy.

[0261] [Objection, Marked as Error Allocation]: Clicking this will bring up a form asking for the reason. After submission, the system records a feedback_event with type='REJECT'. The reward calculator will convert this into a high negative reward of r_human = -2.0 and use it as a high-priority sample (sampling weight 50%) in the next cycle's PPO experience replay pool for targeted learning, exploring new strategies to avoid such errors.

[0262] In summary, this invention provides a dynamic allocation system for aircraft fixed costs based on causal inference and multi-objective adaptive optimization. The system includes six core modules that are sequentially connected: a data fusion module, a causal modeling module, a responsibility quantification module, an allocation optimization module, a voucher generation module, and a monitoring and feedback module. The data fusion module is used to complete ETL processing of multi-source aviation business data, construction of flight-level wide tables, missing value imputation, cold start adaptation, and has built-in configurable amortization range filter and amortization error smoothing pool. The causal modeling module is used to construct a fixed-cost driver causal graph based on a domain-constrained PC causal discovery algorithm, and is equipped with a causal structure drift detection and dynamic update mechanism. The responsibility quantification module is used to calculate the marginal responsibility value of each flight for each fixed cost based on counterfactual inference and SHAP value, and generate an initial allocation weight with causal explanatory power. The allocation optimization module has a built-in PPO reinforcement learning optimization engine and an independent smoothing network NN_η, which is used to complete the optimization of multi-objective allocation strategies and the smooth transition between old and new weights, so as to achieve decoupled optimization of "precise tracking" and "financial smoothing". The voucher generation module is used to complete the multi-dimensional aggregation of allocation results, financial reconciliation verification, standard voucher generation, and automatic push to the ERP system; The monitoring and feedback module has a built-in dynamic confidence interval engine, multi-dimensional anomaly detector and expert feedback entry point, which are used for anomaly monitoring of the distribution results and generation of interpretable reports. It is also equipped with a high-priority expert feedback experience pool to realize human-machine closed-loop model self-learning iteration.

[0263] This invention addresses the fundamental pain points in the field of fixed cost allocation for aircraft, breaking through the logical limitations of traditional allocation technologies and forming six complementary and inseparable core technical points. Each key point addresses the systemic defects of existing technologies: Key technical point 1: A fully automated and dynamic update mechanism for fixed cost-driver discovery based on constraint causal inference in the aviation field.

[0264] This invention introduces a complete causal inference system into the field of aviation fixed cost allocation for the first time, fundamentally solving the core pain points of existing technologies that rely on artificially preset motivations, cannot distinguish between statistical correlation and real causal relationships, and cannot handle hundreds of candidate business elements.

[0265] Core Technology: Based on the PC causal discovery algorithm, and superimposed with hard knowledge constraints in the aviation field (temporal constraints prevent future variables from affecting the past, and hierarchical constraints prevent flight-level variables from affecting aircraft-level fixed attributes), it automatically constructs a directed acyclic causal graph of cost drivers. From hundreds of candidate elements in dimensions such as flight operation, aircraft status, environment, and transportation, it automatically filters out direct and non-redundant causal drivers for each fixed cost, completely eliminating false associations, redundant variables, and the risk of double-counting. At the same time, it is equipped with a causal structure drift detection mechanism, which dynamically monitors statistically significant changes in the causal structure through structural Hamming distance (SHD) and Bootstrap adaptive threshold, automatically triggering causal graph retraining and updating, adapting to dynamic business scenarios such as fleet aging, route network adjustments, and changes in the operating environment.

[0266] Results: Completely eliminates the limitations of human experience in selecting drivers, and for the first time achieves fully automated, data-driven optimal matching between hundreds of fixed costs and hundreds of business elements, fundamentally solving the problems of false correlation, cross-subsidy, and mismatch of responsibilities in traditional allocation methods.

[0267] Key technical point 2: A flight-level marginal cost responsibility quantification system based on the fusion of counterfactual inference and SHAP value.

[0268] This invention breaks through the limitations of the underlying logic of the traditional allocation of "fixed proportions" and for the first time achieves accurate quantification of the marginal cost responsibility of a single flight based on counterfactual inference. It solves the core problem of economic responsibility attribution that existing technologies cannot answer: "How much would the consumption of this fixed cost be reduced if this flight were not there?"

[0269] Core technology: Based on a causal driver set, a fixed cost prediction model is constructed. Through the SHAP interpretable AI method, the model predictions are fairly attributed to the various driver features of each flight, and the marginal responsibility value (which can be positive or negative) of each flight to the corresponding fixed cost is calculated. This accurately quantifies the incremental / reduced contribution of the combination of operational features of a single flight to the consumption of fixed costs. At the same time, the responsibility value is normalized through translation mapping. Without changing the relative responsibility differences between flights, a compliant initial allocation weight is generated, providing an accurate base with causal explanatory power for subsequent optimization.

[0270] Effect: The traditional arithmetic allocation of "dividing the cake proportionally" is upgraded to economic attribution of "determining the share based on causal responsibility", so that the allocation results truly match the actual cost consumption responsibility of each flight, and completely solve the problem that the traditional allocation results have no economic meaning and cannot support refined cost control.

[0271] Key technical point 3: "Exploration-Stability" dual-objective decoupling multi-objective adaptive optimization architecture.

[0272] This invention addresses the core contradiction of existing technologies in balancing "accurately tracking business changes" and "maintaining smooth and stable financial statements" by designing a proprietary decoupled optimization architecture. This solves the industry problem that traditional allocation methods either lead to data distortion due to rigid rules or cause drastic fluctuations in financial statements due to frequent adjustments.

[0273] Core Technology: Based on the PPO proximal strategy optimization reinforcement learning framework, a multi-objective allocation strategy optimization system is constructed. A composite reward function covering four dimensions—financial compliance, management decision usefulness, data smoothness, and expert feedback—is designed to achieve global optimization of the allocation strategy. Simultaneously, an independent smoothing network NN_η is pioneered. While sharing the underlying feature extraction layer with the PPO Actor network, it has an independent output head specifically used for dynamically determining the mixing coefficient η of the old and new allocation weights, as well as the release amount of the allocation error smoothing pool. This completely decouples and optimizes the two strongly conflicting objectives of "accurate tracking of cost drivers" and "smooth and stable financial statements." It responds quickly to structural changes in business and ensures the continuity of reports when business is stable.

[0274] Results: For the first time in the field of aviation cost allocation, both "accuracy" and "smoothness" have been achieved, ensuring that the allocation results can dynamically adapt to changes in business while fully complying with the requirements of continuity, stability and compliance of financial accounting.

[0275] Key technical point 4: A human-machine closed-loop self-learning mechanism that integrates a high-priority experience pool with qualitative feedback from experts.

[0276] The embodiments of the present invention change the static mode of the traditional allocation system of "one-time configuration and no iterative evolution", and build a continuous self-learning system with human-machine closed loop. It solves the problems of existing technologies being unable to absorb feedback from business experts, repeated erroneous allocation, and inability to adapt to the iterative needs of enterprise management.

[0277] Core technology: In the PPO reinforcement learning framework, an independent high-priority expert feedback experience pool is designed to transform the "acceptance / dissent" binary qualitative feedback from business experts on the allocation results into high-weighted reward / penalty signals. During the monthly offline model fine-tuning phase, the sampling ratio of this experience pool is increased to 50%, forcing the policy network to quickly learn and correct the deep policy biases that lead to expert objections. This changes the traditional reinforcement learning paradigm that relies solely on numerical rewards, enabling the continuous autonomous evolution of the allocation policy.

[0278] Results: The system is equipped with the ability to continuously iterate and optimize, and can be deeply adapted to the management standards and business habits of different airlines. The more it is used, the more it fits the actual needs of enterprises, and completely gets rid of the limitations of traditional allocation systems that are "fixed as soon as they go online and rely on manual optimization".

[0279] Key technical point 5: Pre-positioned multi-dimensional allocation range filter and linkage cost precise segmentation mechanism.

[0280] This invention addresses the shortcomings of existing technologies that only support uniform allocation across the entire amount and cannot adapt to specific management and analysis scenarios. It designs a flexible and configurable allocation standard customization system, solving the problem that a single standard in traditional systems cannot simultaneously meet the dual needs of financial compliance accounting and refined management analysis.

[0281] Core technology features: After data cleaning and before core computing begins, a pre-emptive allocation range filter is set up, allowing users to freely combine filtering conditions based on five dimensions: flight attributes, route attributes, aircraft attributes, time attributes, and business tags, to define the allocation data subset. Simultaneously, a linked cost allocation logic based on the proportion of responsibility value is provided, automatically and accurately allocating the deferred costs of the corresponding subset from the total fixed costs according to the ratio of the sum of the basic responsibility values ​​of the subset flights to the sum of the responsibility values ​​of all flights. This ensures the complete compliance of the financial reconciliation of subset allocation and achieves consistency between full accounting and special analysis.

[0282] Results: Without the need for additional offline data processing, it can support diverse management scenarios such as specific fleet accounting, independent assessment of specific route networks, and special cost analysis, significantly reducing the implementation cost of management accounting and expanding the business adaptability boundaries of the system.

[0283] Key technical point 6: A full-link data fusion and anomaly handling system adapted to aviation scenarios.

[0284] This invention addresses the challenges of handling multi-source, heterogeneous data in aviation operations by constructing a complete data governance and anomaly handling system. It solves the practical problems that existing technologies cannot handle, such as data loss, cold start of new businesses, and cross-period cost confusion caused by subsequent data correction.

[0285] Core technologies include: constructing a flight-level wide table with flights as the primary key to achieve data fusion and entity mapping from more than 10 source systems such as flight operations, aircraft master data, QAR / ACARS, maintenance, meteorology, settlement, and ERP; designing a three-level missing value handling strategy (causal inference imputation, autoencoder multimodal imputation, and baseline template replacement) to ensure the compliance and usability of incomplete data; pioneering an independent account for the amortized error smoothing pool to specifically handle the amortized differences caused by the post-correction of template data, with the smoothing network autonomously deciding the rate of difference digestion to completely avoid cross-period cost confusion; and designing a cold start transition scheme based on transfer learning for new aircraft and new routes, achieving a smooth transition from cold start to steady state through asymptotic hybrid coefficients.

[0286] Results: It fully covers data anomalies and cold start issues across all aviation business scenarios, ensuring the system's stability and availability in real business environments, and solving the shortcomings of traditional data sharing systems that can only handle ideal, complete data and cannot adapt to complex real business scenarios.

[0287] This invention also provides a method for dynamically allocating fixed costs for aircraft, as described in the following embodiments. Since the principle underlying this method is similar to that of the dynamic allocation system for fixed costs of aircraft, its implementation can be found in the implementation of the dynamic allocation system for fixed costs of aircraft; details that are repeated will not be repeated.

[0288] Figure 3 This is a flowchart illustrating the dynamic allocation method for aircraft fixed costs in an embodiment of the present invention, as shown below. Figure 3 As shown, this method is applied to the aforementioned aircraft fixed cost dynamic allocation system, and the method includes: Step 301: Obtain fixed cost and multi-source aviation business data, and establish an element pool; the element pool includes candidate cost drivers formed based on multi-source aviation business data; Step 302: Construct observation data using an element pool, introduce aviation domain knowledge constraints, and construct a directed acyclic causal graph using a causal discovery algorithm; the causal graph represents the causal relationships between variables through nodes and edges; the variables are derived from the observation data; Step 303: Obtain the target cost drivers for fixed costs through the causal graph; input the target cost drivers into the fixed cost prediction model, and call the machine learning interpretability tool to attribute the predicted values ​​to each target cost driver, obtaining the marginal contribution value of each target cost driver to the predicted value; use the marginal contribution values ​​of all target cost drivers to the predicted value to determine the initial allocation weight of fixed cost consumption for each flight; the fixed cost prediction model is used to predict the fixed cost expenses allocated at the flight level; the machine learning interpretability tool is used to quantify the degree of influence of input features on the model's predicted values; Step 304: Optimize and adjust the initial allocation weight for each flight to obtain the effective allocation weight for each flight; Step 305: Calculate the fixed cost allocation at the flight level using the effective allocation weight of each flight to obtain the allocation result; generate a financial voucher file using the allocation result; store the allocation result in the financial voucher file in the form required by financial regulations.

[0289] Specifically, the aircraft fixed cost dynamic allocation method in this embodiment of the invention can be implemented according to the following steps: 1. Data fusion and preprocessing steps: Acquire multi-source raw data of aviation business, complete data cleaning, standardization, and missing value handling, and construct a flight-level wide table with flight as the primary key; support users to delineate the amortized data subset through multi-dimensional conditions, and simultaneously complete the accurate allocation of the corresponding deferred costs; 2. Causal Driver Discovery and Dynamic Update Steps: Based on the PC causal discovery algorithm with constraints in the aviation field, a directed acyclic graph of causal drivers is automatically constructed for each fixed cost, and the set of direct causal drivers is screened; through the causal structure drift detection mechanism, the causal graph is dynamically monitored and updated to adapt to dynamic changes in business. 3. Flight-level marginal responsibility quantification steps: Train a fixed cost prediction model based on the causal driver set, complete the fair attribution of the model prediction value through SHAP value, calculate the marginal cost responsibility value of each flight, and normalize to generate the initial allocation weight; 4. Multi-objective adaptive optimization steps: Based on the PPO reinforcement learning framework, financial compliance, management decision usefulness, smoothness, and expert feedback are used as composite reward objectives to optimize and generate objective allocation weights; through an independent smoothing network NN_η, the mixing coefficients of new and old weights are dynamically decided to generate the final effective allocation weights, and the difference release amount of the decision allocation error smoothing pool is synchronized. 5. Voucher generation and push steps: Based on the final allocation weight, complete the flight-level allocation calculation of fixed costs, perform financial reconciliation verification, generate standard financial vouchers and push them to the ERP system; 6. Monitoring Feedback and Self-Learning Steps: Perform multi-dimensional anomaly monitoring on the allocation results and automatically generate an interpretable allocation motivation report; receive qualitative feedback from business experts, transform it into high-weight reward signals and store them in a high-priority experience pool for iterative optimization of the allocation strategy model, and realize continuous self-learning in a human-machine closed loop.

[0290] Figure 4 This is a specific example diagram of the aircraft fixed cost dynamic allocation method in an embodiment of the present invention, as shown below. Figure 4 As shown, the method includes: 1. Begin; 2. Obtain multi-source aviation business data and ERP fixed cost general ledger data; 3. Data cleaning, standardization, and missing value imputation to construct a wide table for each flight; 4. Perform allocation range filtering, define the allocation data subset, and link the deferred costs to the allocation; 5. Has the data quality verification passed? 6. If so, load the currently active cost driver cause-effect graph and perform cause-effect structure drift detection; 7. If not, issue a data anomaly alarm, re-fetch the data, then load the currently effective cost driver cause-effect graph and perform causal structure drift detection; 8. Has the causal structure undergone statistically significant drift? 9. If so, trigger causal graph retraining and update the version; 10. If not, automatically match the driver feature set for each fixed cost based on the cause-effect graph; 11. Calculate the marginal responsibility value for each flight based on the SHAP value, and generate the initial allocation weight; 12. Based on the PPO reinforcement learning engine, the target allocation weights are optimized using a multi-objective composite reward function; 13. Calculate the mixing coefficients using an independent smoothing network to generate the final effective smoothing weights; 14. The final weighted allocation of the basic-level funds for each flight will be subject to financial reconciliation verification. 15. Has the cross-checking been passed? 16. If the reconciliation check fails, a fatal error alarm will be issued, and the process will be interrupted; 17. If the reconciliation verification passes, the multi-dimensional allocation results and standard financial vouchers are generated and pushed to the ERP system for accounting. 18. Detect anomalies in allocation results, automatically generate an allocation explanation report, and push it to business experts; 19. Have you received any feedback from experts regarding their objections? 20. If expert objections are received, they are converted into high negative reward signals and added to the high-quality experience pool; 21. If no expert objection feedback is received, record the positive feedback and write it into the training dataset; 22. Perform monthly fine-tuning of the model based on experience pool data to complete strategy iteration and optimization; 23. End.

[0291] The technical solutions of the present invention will be described in detail below with reference to specific embodiments.

[0292] Figure 5 This is another specific example diagram of the aircraft fixed cost dynamic allocation method in the embodiments of the present invention, with reference to... Figure 5 Scenario: An airline operates over 100 domestic and international routes using both B787 and A320 aircraft, with over 120 fixed cost items per month. The airline needs to allocate the "Engine Overhaul Reserve Fund" (item 510301) for March 2026, and the user only wants to accurately calculate the cost of "international passenger scheduled flights operated by B787 aircraft" to assess the engine utilization efficiency of the 787 fleet.

[0293] Step 1: Data Preparation and Scope Customization

[0294] The system obtains all flight data for March 2026 from systems such as flight operations, QAR, maintenance, and meteorology. At the same time, it obtains the total engine overhaul reserve fund for March, F_total^510301 = 80 million yuan, from the ERP system.

[0295] The data fusion module performs cleaning, standardization, and missing value imputation. For a flight (flight data ID: CA987_20260315) operated by a B787 aircraft (registration number B-1234) from city B to city L, the system automatically associates the data through entity mapping: Aircraft: Model B787-9, maximum takeoff weight 254 tons, age 5.2 years, engine has accumulated 8200 hours and 2100 cycles since major overhaul.

[0296] Route details: 10,000 km distance, international route.

[0297] Flight status: Flight hours 12.5, outside air temperature at takeoff 28℃.

[0298] Cargo information: 245 adult passengers, 3,500 kg of cargo and mail.

[0299] Users configure the "Assignment Range Filter" as follows: Flight Type = Regular Flight, Aircraft Type = B787-9, Route Entry / Exit Identifier = International, Actual Flight Dates = "2026-03-01 to 2026-03-31". The filter selects 350 flights that meet the criteria from 12,000 flights throughout the month. The system self-checks and generates a subset wide table.

[0300] Step 2: Running the causal model

[0301] The causal modeling module loads the currently active causal graph G_current. This graph shows the direct causal drivers of the engine overhaul reserve fund: csn_hour_eng1, csn_cycle_eng1, takeoff_oat, and avg_takeoff_n1. Factors such as passenger numbers are automatically excluded because they are only indirectly related.

[0302] Step 3: Quantify Responsibilities

[0303] The responsibility quantification module loads the corresponding XGBoost prediction model and SHAP interpreter. For each subset of flights, the following calculations are performed: The sum of SHAP values ​​(original responsibility value) for flight CA987_20260315 is +18500, because it has completed 2100 cycles since the overhaul and the takeoff temperature is relatively high.

[0304] Another flight that had just completed a major overhaul had an original liability value of -12,000.

[0305] The minimum value within the batch is -15000, which is normalized after translation mapping to obtain W_init. The initial weight of CA987_20260315 is 0.0082.

[0306] Step 4: Amount Allocation Optimization

[0307] The allocation optimization module loads the PPO Actor and NN_η. Based on the subset state (350 flights, 35 routes, smoothing pool balance of 1.2 million yuan), the Actor outputs the multiplier for each route, and NN_η outputs η=0.3 and decides to release 180,000 yuan from the smoothing pool. The multiplier for the US West Coast route where CA987_20260315 is located is 1.05, and the final weight W_eff after smoothing and mixing is 0.0084.

[0308] The system allocates 18.5 million yuan of deferred costs from the 80 million yuan based on the proportion of the responsibility value of the subset flights.

[0309] Step 5: Voucher Generation

[0310] The system calculates the allocated amount for CA987_20260315 as 18.5 million × 0.0084 = 155,400 yuan. Simultaneously, it generates aggregated amounts at the route level (US West Coast route 3.42 million yuan), aircraft level (B-1234 aircraft 2.16 million yuan), and aircraft type level (B787-9 fleet 18.5 million yuan). After reconciliation verification, the voucher XML is pushed to the ERP system.

[0311] Step 6: Monitoring and Feedback

[0312] The monitoring feedback module detected an 8% month-on-month increase in the cost-sharing expense for the US West Coast route. An investigation revealed that several older aircraft with high flight cycles were replaced on this route last month, which was flagged as a major event, and the confidence interval was automatically widened, resulting in no system alerts. Another newly opened B787LD route is in its cold start transition period (80 flights, λ=0.6), and the system generated an explanation report. The route manager reviewed it and clicked "Accept." The system provided a detailed explanation of the driving forces behind this turnaround route, which the route manager confirmed was reasonable after review.

[0313] The embodiments of the present invention achieve the following functions: Causal inference addresses "counterfactual reasoning": Traditional cost allocation only uses statistical relevance to assign weights, failing to answer the economically relevant question of cost responsibility: "How much would the fixed cost be reduced if this flight were not there?" This invention introduces causal discovery and counterfactual inference to quantify each flight's "marginal responsibility" for fixed costs, a first in the field of cost allocation.

[0314] The specific network design for “exploration-stability” decoupling: Unlike conventional reinforcement learning that directly outputs weights and relies on hard environmental constraints for smoothing, this embodiment of the invention specifically designs an independent smoothing network NN_η, which autonomously learns when to respond quickly (η→1) and when to be conservatively smooth (η→0) during interaction with the environment, thus decoupling and optimizing the contradictory goals of “accurate tracking” and “financial smoothing”.

[0315] Human-machine closed-loop system changes the fundamental learning paradigm: In the field of aviation finance, the qualitative judgment of business experts is the final verification of the rationality of allocation. This embodiment of the invention introduces a high-priority expert feedback experience pool into the PPO framework, enabling the model to learn not only from numerical rewards, but also to correct deep policy biases from the binary feedback of "right / wrong" from experts. This changes the reinforcement learning paradigm that is solely based on reward signals.

[0316] The embodiments of the present invention have the following technical effects: A revolutionary improvement in allocation accuracy: By automatically identifying the true cost drivers through causal inference, the false associations and cross-subsidies caused by manual selection of drivers in traditional methods are eliminated, giving the allocation results the explanatory power of "cost responsibility" in an economic sense.

[0317] End-to-end automation and continuous intelligent evolution: From multi-source data fusion, driver discovery, responsibility quantification, and allocation optimization to voucher entry, the entire process is unattended. The system can continuously learn and optimize from the binary feedback of "right" or "wrong" from business experts, becoming increasingly aligned with business expectations with use.

[0318] Excellent dynamic adaptive capability: The allocated weights are automatically recalculated monthly based on aircraft aging, changes in the operating environment, and adjustments to the route network. The unique smooth network achieves decoupling optimization between "responding to business changes" and "maintaining report stability," avoiding drastic fluctuations in the allocation results.

[0319] Fully automated optimal matching of costs and drivers: Faced with hundreds of costs and hundreds of candidate factors, the system can automatically select the optimal combination of drivers for each cost, completely freeing up the huge workload of manually configuring allocation rules item by item, and the accuracy far exceeds human experience.

[0320] Ultimate flexibility in analysis and accounting: The unique "allocation range filter" allows users to define allocation boundaries by any combination of flight attributes, routes, aircraft types, and times, linking costs to intelligent segmentation, easily supporting diverse management accounting scenarios such as specialized fleet analysis and independent route network accounting.

[0321] Completely transparent and explainable: For the first time, the SHAP-based “allocation explanation report” is introduced into the field of aviation cost management, enabling finance and business personnel to clearly understand the driving reasons for each cost allocation, significantly reducing communication costs and improving the quality and credibility of management decisions.

[0322] Figure 6 This is a schematic diagram of a computer device in an embodiment of the present invention, such as... Figure 6 As shown, this embodiment of the invention also provides a computer device 600, including a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601. When the processor 601 executes the computer program 603, it implements the above-mentioned method for dynamic allocation of fixed costs of aircraft.

[0323] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for dynamically allocating fixed costs of aircraft.

[0324] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for dynamically allocating fixed costs of aircraft.

[0325] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0326] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0327] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0328] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0329] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic allocation system for fixed costs of aircraft, characterized in that, include: Data fusion module, causal modeling module, responsibility quantification module, allocation optimization module, and voucher generation module; The data fusion module is used to: acquire fixed cost and multi-source aviation business data, and establish an element pool; the element pool includes candidate cost drivers formed based on multi-source aviation business data; The causal modeling module is used to: construct observation data from an element pool, introduce aviation domain knowledge constraints, and construct a directed acyclic causal graph using a causal discovery algorithm; the causal graph represents the causal relationships between variables through nodes and edges; the variables are derived from the observation data; The responsibility quantification module is used to: obtain the target cost drivers for fixed costs through the causal graph; input the target cost drivers into the fixed cost prediction model, call machine learning interpretability tools to attribute the predicted values ​​to each target cost driver, and obtain the marginal contribution value of each target cost driver to the predicted value; determine the initial allocation weight of fixed cost consumption for each flight using the marginal contribution values ​​of all target cost drivers to the predicted value; the fixed cost prediction model is used to predict the fixed cost expenses allocated at the flight level; The machine learning interpretability tool is used to quantify the degree of influence of input features on the model's predicted values; The allocation optimization module is used to: optimize and adjust the initial allocation weight of each flight to obtain the effective allocation weight of each flight; The voucher generation module is used to: calculate the fixed cost allocation at the flight level using the effective allocation weight of each flight, and obtain the allocation result; generate a financial voucher file using the allocation result; and store the allocation result in the financial voucher file in the form required by financial regulations.

2. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The multi-source aviation business data includes at least one of the following elements: aircraft, routes, transportation, services, crew, and weather. The data fusion module is specifically used for: Collect total fixed costs and full volume of multi-source aviation business data within a preset time frame; The full volume of multi-source aviation business data is preprocessed to obtain preprocessed full data. Receive user-input filtering criteria, filter based on preprocessed full data according to the filtering criteria, and construct a flight-level wide table subset; The data of the flight-level width table subset is self-checked; the data self-check includes empty set check, skewness check, and cost segmentation linkage; the skewness check includes checking that the ratio of any indicator data obtained by filtering to the total data of that indicator is lower than a set ratio; the cost segmentation linkage means that the amount corresponding to the flight-level width table subset is segmented from the total fixed cost.

3. The aircraft fixed cost dynamic allocation system as described in claim 2, characterized in that, The data fusion module is specifically used for: The following data preprocessing is performed on the full amount of multi-source aviation business data: Flight number unification, timestamp alignment, numerical verification, missing data filling, and cold start processing.

4. The aircraft fixed cost dynamic allocation system as described in claim 3, characterized in that, The missing data imputation process includes the following: Causal inference imputation: Based on the currently active causal graph, obtain the parent node data of the missing field. If the parent node data is not empty, input the parent node data into the online ridge regression model, output the estimated value of the missing field, add a random perturbation term to the estimated value, and obtain the imputed value of the missing field. Autoencoder multimodal imputation: If the parent node data is empty or there is no effective causal graph, all the non-missing variable values ​​of the current flight are used as input, and the complete vector is reconstructed through a pre-trained deep autoencoder. Missing fields are taken from the reconstructed complete vector. Baseline template replacement: When multiple specified core fields are missing data, template data is used to fill the gaps; the template data is a pre-calculated three-dimensional matrix of aircraft type, route and season, which includes the median of each characteristic variable of all flights under the combination of aircraft type, route and season.

5. The aircraft fixed cost dynamic allocation system as described in claim 3, characterized in that, The cold start process includes: Missing fields for new aircraft are replaced with historical averages for the same aircraft type and engine model. Missing fields for new routes are replaced with data from similar older routes; these similar older routes are determined through feature similarity assessment of the takeoff and landing airports.

6. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The aviation domain knowledge constraints include temporal constraints and hierarchical constraints; the temporal constraints prohibit future variables from affecting the past; the hierarchical constraints prohibit flight-level variables from pointing to aircraft-level attributes. The causal discovery algorithm is the PC algorithm; The causal modeling module is specifically used to: based on the PC algorithm and superimposed with aviation domain knowledge constraints, traverse all variable pairs, determine the causal relationship between variables, iterate repeatedly until the number of edges converges, and obtain a directed acyclic causal graph.

7. The aircraft fixed cost dynamic allocation system as described in claim 6, characterized in that, The causal modeling module is specifically used for: Fisher's Z-test is used for conditional independence of continuous variables, and G-test is used for discrete variables. 2 Test the independence of conditions.

8. The aircraft fixed cost dynamic allocation system as described in claim 6, characterized in that, The causal modeling module is specifically used for: The historical element pool is segmented to create new and old datasets; Construct two causal graphs using the new dataset and the old dataset respectively, and calculate the structural Hamming distance (SHD) between the edge sets of the two causal graphs. Compare the SHD with a set threshold to determine whether the causal structure has changed statistically significantly; If a statistically significant change occurs, the currently active causal graph will be updated.

9. The aircraft fixed cost dynamic allocation system as described in claim 8, characterized in that, The set threshold is updated periodically in the following manner: Block resampling is performed on the feature pool data from the most recent months to generate a specified number of pseudo-samples of the same size; For each pair of pseudo-samples, run the PC algorithm separately to calculate the SHD between them and form the zero distribution of SHD; The 95th percentile of this distribution is used as the set threshold.

10. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The fixed cost prediction model is a tree-based machine learning model; the machine learning interpretability tool is the SHAP interpreter; the fixed cost includes multiple fixed expense items; the marginal contribution value includes the SHAP value; The responsibility quantification module is specifically used for: For multiple fixed cost items actually incurred in a single flight in batch multi-source aviation business data, the target cost driver for each fixed cost item is obtained through the cause-effect graph. Input the target cost drivers of a single fixed cost item into the fixed cost forecasting model, and call the SHAP interpreter to calculate the SHAP value of each input target cost driver to the forecast value. The sum of the SHAP values ​​of all input target cost drivers to the predicted values ​​is determined as the original marginal liability value for a single flight for the single fixed cost item; A translation mapping is applied to the original marginal responsibility values ​​of all flights in the batch multi-source aviation business data to obtain the translated marginal responsibility values; the marginal responsibility values ​​are used to quantify the causal contribution of a single flight to the consumption of a single fixed cost item; The initial allocation weight of fixed cost consumption for each flight is determined by the ratio of the marginal responsibility value after a single flight translation mapping to the sum of the marginal responsibility values ​​after all flights translation mapping.

11. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The allocation optimization module is specifically used to: obtain the initial allocation weight of the current period's flights, the remaining amount in the smoothing pool, the statistical data after allocation in the previous period, and historical allocation feedback to construct the current system state; optimize and adjust the initial allocation weight of each flight using the current system state to obtain the effective allocation weight of each flight; the smoothing pool stores the difference between the first allocation amount calculated using template data and the second allocation amount calculated using completed real data; the template data is the data after filling in the missing fields using template data.

12. The aircraft fixed cost dynamic allocation system as described in claim 11, characterized in that, The amortization optimization module is specifically used for: Based on the current state of the system, the pre-trained reinforcement learning agent outputs adjustment actions under the guidance of a multi-objective reward function, thereby obtaining the weights after the reinforcement learning actions are adjusted. The adjustment actions include a uniform multiplier of the initial allocation weights of all flights within the same route. The multi-objective reward function includes at least one of the following: financial compliance reward, management decision usefulness reward, route-level allocation smoothness reward, and human-machine feedback reward. Based on the current state of the system, a pre-trained smoothing network is used to predict the output smoothing mixing coefficient; this smoothing mixing coefficient is used to balance the change in response speed and the stability of financial allocation; the smoothing network is a supervised training neural network; By using a smoothing mixing coefficient, the weights adjusted for reinforcement learning actions are combined with the effective allocation weights from the previous period to obtain the effective allocation weights for each flight in the current period.

13. The aircraft fixed cost dynamic allocation system as described in claim 12, characterized in that, The amortization optimization module is specifically used for: The effective weights for each flight are obtained by fusing the weights adjusted for reinforcement learning actions with the initial allocated weights using the following formula and smoothing mixing coefficient: W eff =η×W target +(1-n)×W old ; Among them, W eff For the weighting to take effect, W target To reinforce the weights of the learned actions after adjustment, W old η is the initial allocated weight, and η is the smoothing mixing coefficient.

14. The aircraft fixed cost dynamic allocation system as described in claim 13, characterized in that, When the system's services undergo structural changes, the smoothing network predicts a smoothing mixing coefficient close to 1; when the system's services operate smoothly, the smoothing network predicts a smoothing mixing coefficient close to 0.

15. The aircraft fixed cost dynamic allocation system as described in claim 12, characterized in that, The amount released and absorbed by the smoothing pool in the current period is predicted and output by the auxiliary output head of the smoothing network; this amount released and absorbed in the current period will be added to the fixed cost of the current period for allocation.

16. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The voucher generation module includes the following two triggers: the first trigger is used to receive the closing dynamics sent by the external financial business system and generate the first trigger signal; the second trigger uses an internal timed task to periodically check whether the fixed cost and multi-source aviation business data cover all executed flights in the previous month. If the coverage exceeds a specified ratio, a second trigger signal is generated. When the first and second trigger signals of the month are present simultaneously, the allocation task workflow is started; the allocation task workflow automatically mobilizes the data fusion module, causal modeling module, responsibility quantification module, allocation optimization module, and voucher generation module to start running.

17. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, The voucher generation module is specifically used for: After obtaining the allocation results, the allocation results are aggregated in multiple dimensions. Fixed costs are used to perform financial reconciliation verification on the multi-dimensional aggregated allocation results. After the financial reconciliation verification is passed, financial voucher files are generated. The multi-dimensional aggregation includes any combination of flight dimension, route dimension, flight season dimension, aircraft dimension, and aircraft type dimension.

18. The aircraft fixed cost dynamic allocation system as described in claim 1, characterized in that, It also includes a monitoring feedback module for monitoring one or any combination of the following: Monitoring for discrepancies in total allocated costs: If abs(sum of allocated costs for all flights - total fixed costs) > 1.0, immediately block the generation of voucher files and issue an alert to the designated terminal; Single route cost allocation jump: If the cost allocation of a single route exceeds the dynamic confidence interval and there is no major event label, an alert is issued to the designated terminal; The monitoring and feedback module: By calculating the exponentially weighted moving average and exponentially weighted moving variance, a dynamic confidence interval is maintained for the fixed cost of each route each month; Cost driver structure drift: Calculate the percentage of the marginal contribution of a single target cost driver in the sum of the marginal contributions of all target cost drivers. If the difference between this month's percentage and the average of the past few months exceeds the set percentage and there is no business explanation, issue an alert to the designated terminal. Reversal of profitability indicators: Monitor changes in profitability indicators after allocation for a single route. If the profitability indicator of a single route suddenly changes from profit to loss after several consecutive months, issue an early warning to the designated terminal.

19. The aircraft fixed cost dynamic allocation system as described in claim 18, characterized in that, The monitoring feedback module is specifically used for: While issuing an early warning to the designated terminal, an allocation explanation report is provided; the allocation explanation report includes the Top N positive impact drivers and the Top N negative impact drivers statistically analyzed through marginal contribution values; N is a positive integer.

20. A method for dynamically allocating fixed costs of aircraft, characterized in that, Applied to the aircraft fixed cost dynamic allocation system according to any one of claims 1 to 19, the method includes: Acquire fixed cost and multi-source aviation business data, and establish a factor pool; the factor pool includes candidate cost drivers formed based on multi-source aviation business data; The observation data is constructed using an element pool, and knowledge constraints from the aerospace field are introduced. A directed acyclic causal graph is constructed using a causal discovery algorithm. The causal graph represents the causal relationships between variables through nodes and edges. The variables are derived from the observation data. The target cost drivers for fixed costs are obtained through the causal graph; these target cost drivers are input into the fixed cost prediction model, and machine learning interpretability tools are used to attribute the predicted values ​​to each target cost driver, obtaining the marginal contribution value of each target cost driver to the predicted value; the initial allocation weight of fixed cost consumption for each flight is determined using the marginal contribution values ​​of all target cost drivers to the predicted value; the fixed cost prediction model is used to predict fixed cost expenses allocated at the flight level; the machine learning interpretability tools are used to quantify the degree of influence of input features on the model's predicted values; The initial allocation weight for each flight is optimized and adjusted to obtain the effective allocation weight for each flight; Fixed costs are allocated at the flight level using the effective allocation weight of each flight to obtain the allocation result; financial voucher files are generated using the allocation result; the allocation result is stored in the financial voucher file in accordance with financial requirements.

21. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of claim 20.

22. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of claim 20.

23. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of claim 20.

Images