An engine accessory demand prediction method and device based on repair data

Abstract

The application discloses a kind of engine accessory demand prediction method and device based on repair data, it is related to supply chain management technical field, the method includes: obtaining engine three package repair data from enterprise business system, generates fault accessory demand matrix by data cleaning, fault mode clustering, establishes the mapping relationship of fault mode and accessory model;Based on demand matrix statistics, obtain each accessory historical demand data, according to demand characteristics, accessory is divided into low-frequency intermittent demand accessory and medium-high frequency demand accessory;Two kinds of accessories are carried out demand prediction by respectively using improved Croston model of introducing Bootstrap, LSTM model;Through self-adapting selector, corresponding model prediction result is selected according to accessory demand characteristics, or fuses the results of two types of models of same accessory, outputs final demand prediction value.The application solves the problem of low accuracy of engine accessory demand prediction in the prior art.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to a method and apparatus for forecasting the demand for engine parts based on repair data. Background Technology

[0002] Current engine parts demand forecasting relies heavily on historical sales data or simple time series models, ignoring the causal relationship between failure modes and parts demand. Demand forecasting cannot adaptively handle different intermittent periods, resulting in large errors and ultimately low forecast accuracy. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method and apparatus for predicting the demand for engine parts based on repair data, which aims to solve the problem of low accuracy in predicting the demand for engine parts in the prior art.

[0004] This invention proposes a method for predicting engine parts demand based on repair data, the method comprising: Engine warranty repair data is obtained from at least one business system of the enterprise. The data is then cleaned and clustered by failure mode to generate a failure parts demand matrix and establish a mapping relationship between failure modes and parts models. Based on the demand matrix of faulty parts, historical demand data for each part is generated. Based on the historical demand data of the parts, the demand is divided into low-frequency intermittent demand parts and medium- and high-frequency demand parts. For components with low-frequency intermittent demand, an improved Crosston model incorporating the Bootstrap method is used to generate demand forecasts. For components with medium- to high-frequency demand, an LSTM model is used to generate demand forecasts. The adaptive selector selects the corresponding model's prediction result output based on the demand characteristics of the accessory, or merges the prediction results of two models for the same accessory and outputs the final demand prediction value.

[0005] Furthermore, the above-mentioned engine parts demand forecasting method based on repair data includes the following steps: generating a fault parts demand matrix through data cleaning and fault mode clustering: Perform data cleaning on the warranty repair data to remove records with missing fields and abnormal usage, and obtain a valid repair data set; Standardize the coding of fault codes in valid repair data, map the original fault codes to a preset multi-level fault classification system, and unify the fault coding rules; K-Shape or DBSCAN algorithms are used to cluster the fault occurrence time series to generate several fault mode clusters; The demand frequency and average single usage of each part model under each failure mode cluster are statistically analyzed, and a failure part demand matrix is ​​constructed with failure mode as the row and part model as the column. The mapping accuracy is verified by random sampling and manual verification. When the accuracy is lower than a preset threshold, the mapping rules are updated and optimized.

[0006] Furthermore, the aforementioned engine parts demand forecasting method based on repair data includes the following steps: classifying parts into low-frequency intermittent demand parts and medium-to-high-frequency demand parts based on historical demand data: Calculate the demand density and zero-value ratio for each component, where demand density is the ratio of the number of demand occurrence cycles to the total number of statistical cycles, and zero-value ratio is the ratio of the number of zero demand cycles to the total number of statistical cycles. Set a first threshold and a second threshold. If the demand density of a component is less than the first threshold or the proportion of zero values ​​is greater than the second threshold, it is determined to be a low-frequency intermittent demand component, and the rest are determined to be high-frequency demand components.

[0007] Furthermore, the above-mentioned engine parts demand forecasting method based on repair data includes the following steps: for low-frequency intermittent demand parts, an improved Crosston model incorporating the Bootstrap method is used to generate forecast results; for medium- and high-frequency demand parts, an LSTM model is used to generate forecast results. For low-frequency intermittent demand components, demand interval sequences and single demand quantity sequences are generated based on the corresponding historical demand data. Exponential smoothing is used to smooth the two types of sequences respectively. The smoothing parameters are optimized and determined by grid search algorithm. Based on the estimated demand intervals and estimated demand quantities obtained through smoothing calculations, the average demand forecast for the future preset period is calculated and output. The Bootstrap resampling method is used to resample the historical demand data several times with replacement. Based on each group of resampled data, the Croston forecast value is calculated to construct an empirical distribution of the demand forecast value. Based on the preset confidence level, the corresponding upper and lower quantile values ​​are extracted from the empirical distribution to generate the demand forecast confidence interval. The demand forecast mean and the corresponding confidence interval are used together as the forecast result for low-frequency intermittent demand components. For parts with medium to high demand, historical demand data is extracted and matched to obtain multi-dimensional covariate feature data for the corresponding statistical period. The covariate features cover vehicle operation, season, maintenance hours and usage area. Construct an LSTM prediction model that integrates covariate inputs, using historical demand sequences of a preset length and covariate features of the corresponding periods as model inputs, and outputting the predicted value of parts demand for the next single period. Historical data is divided into training and validation sets to iteratively train the LSTM prediction model. During the training process, an early stopping strategy is used to suppress model overfitting. After the model converges, the optimal model parameters are saved, and the prediction results of high-frequency demand parts are output.

[0008] Furthermore, the above-mentioned engine parts demand forecasting method based on repair data includes the following steps for outputting the final demand forecast value after fusing the prediction results of two models for the same part: Obtain the demand forecast results of the same component using the improved Croston model and the LSTM prediction model respectively; Based on the validation dataset, a linear weighted fusion method is used to fuse the two types of prediction results. With the goal of achieving the best prediction accuracy, the fusion weights corresponding to the two types of models are determined by a grid search algorithm. The final demand forecast value for the accessory is calculated and output based on the optimal fusion weight obtained through optimization.

[0009] Furthermore, the above-mentioned engine parts demand forecasting method based on repair data, after the step of selecting the corresponding model's prediction result output through an adaptive selector based on the parts' demand characteristics, or merging the prediction results of two models for the same parts and outputting the final demand forecast value, also includes: Using the final demand forecast as the core demand input, a multi-level inventory optimization model integrating a horizontal transfer mechanism is constructed for a multi-level inventory architecture consisting of a central inventory node and multiple forward inventory nodes. Configure (s,Q,S) inventory strategies for each model of parts at each inventory node. Calculate the replenishment trigger threshold, replenishment batch and target inventory level for each node based on the final demand forecast. The optimization objective is to minimize the sum of inventory holding cost, stockout cost and horizontal transfer cost, with preset service level and preset capital occupation limit as constraints. The Benders decomposition algorithm is used to iteratively solve the multi-level inventory optimization model. After iterating until the objective function meets the preset convergence condition, the inventory strategy parameters and horizontal transfer rules of each node are output.

[0010] Furthermore, the above-mentioned engine parts demand forecasting method based on repair data, after the step of selecting the corresponding model's prediction result output through an adaptive selector based on the parts' demand characteristics, or merging the prediction results of two models for the same parts and outputting the final demand forecast value, also includes: The strategy update process is automatically triggered according to the preset update cycle. The newly added warranty repair data since the previous update cycle is extracted incrementally and merged with the historical dataset to obtain the updated full basic data. Based on the updated full set of basic data, the fault mode and part model mapping relationship update, demand forecasting model retraining, and multi-level inventory strategy optimization are performed in sequence to generate inventory strategy parameters and horizontal transfer rules for the new cycle. The parameters of the new cycle's inventory strategy are compared with the parameters of the previous cycle's strategy. If the difference exceeds the preset difference threshold, a strategy update instruction is automatically sent to the enterprise inventory execution system.

[0011] Another object of the present invention is to provide an engine parts demand prediction device based on repair data, the device comprising: The acquisition module is used to obtain engine warranty repair data from at least one business system of the enterprise, and generate a fault parts demand matrix by sequentially performing data cleaning and fault mode clustering, and establish a mapping relationship between fault modes and parts models. The segmentation module is used to generate historical demand data for each part based on the demand matrix of faulty parts, and to segment the parts based on the historical demand data, so as to divide the parts into low-frequency intermittent demand parts and medium- and high-frequency demand parts. The forecasting module is used to generate forecast results for low-frequency intermittent demand components by using an improved Croston model that incorporates the Bootstrap method, and for medium- and high-frequency demand components by using an LSTM model. The generation module is used to select the prediction results of the corresponding model based on the demand characteristics of the accessory through an adaptive selector, or to output the final demand prediction value after fusing the prediction results of two models of the same accessory.

[0012] Another object of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0013] Another object of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0014] This invention obtains engine warranty repair data from at least one business system of an enterprise, and generates a faulty parts demand matrix through data cleaning and fault mode clustering, establishing a mapping relationship between fault modes and parts models. Based on the faulty parts demand matrix, historical demand data for each part is statistically generated, and parts are divided into low-frequency intermittent demand parts and medium-to-high-frequency demand parts according to demand characteristics. Differential demand forecasting is performed for the two types of parts using an improved Crosston model incorporating the Bootstrap method and an LSTM model, respectively. An adaptive selector selects the corresponding model prediction result for output based on the part demand characteristics, or merges the prediction results of the two models for the same part to output the final demand forecast value. This avoids the shortcomings of existing technologies that rely on historical sales data and simple time series models, ignore the causal relationship between fault modes and parts demand, and cannot adaptively adapt to different intermittent demand characteristics with a single prediction model. It effectively reduces prediction errors and ultimately achieves accurate engine parts demand forecasting. This solves the problems of low accuracy and large errors in existing engine parts demand forecasting. Attached Figure Description

[0015] Figure 1 This is a flowchart of the engine parts demand prediction method based on repair data in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the engine parts demand prediction device based on repair data in the third embodiment of the present invention.

[0016] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0018] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] Example 1 Please see Figure 1 The figure shows the engine parts demand prediction method based on repair data in the first embodiment of the present invention, the method including steps S10 to S13.

[0021] Step S10: Obtain engine warranty repair data from at least one business system of the enterprise, and generate a fault parts demand matrix by sequentially performing data cleaning and fault mode clustering, and establish a mapping relationship between fault modes and parts models.

[0022] The internal business systems that handle warranty repair records include three main types: ERP (Enterprise Resource Planning) systems, DMS (Dealer Management System), and EOL (Electronic Online Vehicle Calibration) systems. Individual data sources can be selected from a single system, or multiple systems can be integrated simultaneously for data collection. The system collects core fields such as the 17-digit Vehicle Identification Number (VIN), 5-digit standard fault code, repair start and end dates, parts and material models, quantity of parts used in a single repair, vehicle's cumulative mileage, vehicle's average annual operating temperature, and repair man-hours. After raw data collection, a unified data cleaning operation is performed to remove invalid and corrupted data. Then, clustering operations are performed on the cleaned, valid fault time series to classify standardized fault mode categories. The frequency and quantity of various parts consumed under each fault mode are statistically analyzed, forming a two-dimensional matrix structure of fault part demand. Each set of values ​​in the matrix corresponds to the average quantity of corresponding parts consumed when a certain type of fault occurs. This establishes a stable quantitative mapping relationship between the cause of the fault and the demand for parts, breaking the limitations of traditional prediction methods that only use historical parts sales and ignore the causes of the fault.

[0023] Specifically, a batch cleaning operation is performed on the original warranty repair data obtained from the enterprise's business system. The cleaning is divided into two categories of abnormal data removal logic. The first category is records with missing fields. The program automatically traverses all collected fields. If a single repair record has any of the following: missing repair date, blank fault code, empty part model, or no mileage recorded, it is directly marked as an invalid record and removed entirely. The second category is records with abnormal usage. A reasonable threshold range for part usage is set. The minimum number of parts used per repair is 0, and the maximum is 10. If the part usage in a single record is negative or exceeds 10, it is judged as an error in data entry and is removed simultaneously. After all the original data has been traversed, the remaining records without missing or abnormal data are integrated to form a unified and valid repair dataset, which serves as the basic data source for subsequent fault coding and clustering operations. In an exemplary implementation scenario, the total amount of original repair data for 36,000 monthly cycles is 120,000 records. After cleaning, dirty data with a missing rate of 1.2% and a usage abnormality rate of 2% is removed, leaving 115,000 valid repair records. The original fault code coding rules for vehicles manufactured by enterprises are inconsistent, and the number of digits and coding logic of fault codes vary across different engine systems, making direct clustering calculations impossible. Therefore, a five-level standardized fault classification system was established, with the levels from top to bottom as follows: vehicle system level, engine subsystem level, core component level, standardized fault mode level, and unified fault code level. The program has a built-in code mapping table that converts the original fault codes of varying lengths into a unified five-digit fixed-length standard code. Multiple sets of original fault codes corresponding to the same fault mechanism are uniformly mapped to the same classification node, completing the standardized conversion of fault codes for all valid repair records and eliminating clustering calculation biases caused by inconsistent original codes. After all standardized fault codes have been converted, time-series samples of fault occurrences within the warranty period of each engine are extracted at fixed time intervals. Two feasible clustering algorithms are then selected for clustering operations. The first algorithm uses K-Shape clustering, which is adapted to time-series shape similarity matching. It uses SBD shape distance as the sequence similarity metric, and the total number of clusters is pre-set manually. The algorithm automatically groups fault samples with consistent time-series changes into the same cluster. The second algorithm uses DBSCAN density clustering, which does not require pre-setting the number of clusters. It automatically divides fault clusters based on sample density, automatically identifying and classifying discrete, isolated fault samples separately. After the clustering operation, multiple independent fault mode clusters are output. Each cluster corresponds to a fixed fault occurrence mechanism. In the example scenario, fifteen clusters are set, corresponding to typical engine fault modes such as high-pressure common rail faults and thermostat faults. The entire clustering process iterates through all fault mode clusters generated by the clustering. For all valid maintenance records within each cluster, two core statistical indicators are calculated based on the assigned part types. The first indicator is demand frequency, which is the total number of times the corresponding part is consumed when the fault mode occurs. The second indicator is average usage per instance, which is the average number of parts replaced in a single maintenance scenario for the fault mode. After the statistics are completed, a two-dimensional numerical matrix is ​​constructed. The matrix row index is the number of all fault mode clusters generated by the clustering, and the matrix column index is the material type of all engine parts in the enterprise. The values ​​in the cells of the matrix are filled with the average usage per instance of the corresponding part under the corresponding fault mode, forming a complete fault part demand matrix. The matrix dimensions are adaptively adjusted synchronously with the number of fault clusters and the total number of part types. In an exemplary scenario, fifteen fault clusters are paired with two hundred and thirty engine parts, generating a standard demand matrix of fifteen rows and two hundred and thirty columns. After the matrix is ​​constructed, the mapping accuracy verification process is automatically initiated. The program randomly selects a fixed number of samples from all valid repair records. For example, 500 repair records are selected and pushed to after-sales personnel for manual verification of the matching relationship between fault modes and parts consumption. The ratio of the number of correctly matched samples to the total number of samples is used as the mapping accuracy. Enterprises pre-set a minimum accuracy threshold, with a commonly used threshold of 90%. If the accuracy value obtained by manual verification is lower than this threshold, the system automatically triggers a full-process update and optimization instruction, re-executing all steps of fault code standardization, fault sequence clustering, and matrix statistics, and regenerating a new fault parts demand matrix until the mapping accuracy meets the minimum standard. In the example scenario, 462 out of 500 sampled records are correctly matched, with a mapping accuracy of 92.4%, which meets the threshold requirement and does not require re-iteration.

[0024] Step S11: Based on the faulty parts demand matrix, generate historical demand data for each part, and classify the parts into low-frequency intermittent demand parts and medium-to-high-frequency demand parts based on the historical demand data.

[0025] Based on the established faulty parts demand matrix and combined with the company's full-cycle warranty repair records, the actual consumption quantity of each engine part is summarized within a continuous statistical period according to a fixed statistical cycle, forming independent time series historical demand data for each part. After the generation of the single-part time series data, two core demand characteristic indicators are calculated. The magnitude of these two indicators distinguishes the fluctuation characteristics of part demand, completing the automatic classification of two types of parts. Low-frequency intermittent demand parts correspond to niche parts such as fuel injectors and thermostats, which have no long-term maintenance consumption and only generate a small amount of demand in sporadic periods. Medium- and high-frequency demand parts correspond to conventional parts such as filters and seals, which are continuously and stably consumed by repair stations in various regions. The demand characteristics of the two types of parts are significantly different, and a single prediction model cannot simultaneously adapt to both types of sequences. Therefore, a hierarchical classification operation is performed to provide a classification basis for subsequent differentiated modeling and prediction.

[0026] Specifically, time-series demand data for each engine part is extracted separately for the entire statistical period, with a fixed total statistical period set. For example, 36 consecutive months of warranty data are selected as the total period range. Two core indicators are calculated for each part's time-series data: the first is demand density, calculated by dividing the total number of cycles in which the part is actually consumed by the total number of statistical periods, with a value between 0 and 1; the second is the zero-value ratio, calculated by dividing the total number of cycles with zero demand by the total number of statistical periods, also with a value between 0 and 1. Both indicators are automatically calculated based on historical maintenance demand data, requiring no manual intervention, and all part indicator values ​​are output in batches.

[0027] Enterprises pre-determine two sets of judgment thresholds based on the characteristics of their engine parts business. The first threshold corresponds to the demand density judgment standard, exemplified by a value of 0.2. The second threshold corresponds to the zero-value ratio judgment standard, exemplified by a value of 0.6. The program reads the calculated demand density and zero-value ratio values ​​for each part, executes dual judgment logic, classifying it as a low-frequency, intermittent demand part if either indicator meets the judgment condition, and classifying it as a medium- to high-frequency demand part if neither condition is met. The classification result directly serves as the input identifier for the subsequent differentiated prediction model. The threshold values ​​can be flexibly adjusted according to different automakers' engine parts categories and warranty periods, and are not limited to the example values, as long as the judgment logic of the two thresholds remains unchanged.

[0028] Step S12: For low-frequency intermittent demand components, an improved Crosston model incorporating the Bootstrap method is used to generate demand forecasting results; for medium- and high-frequency demand components, an LSTM model is used to generate demand forecasting results.

[0029] Specifically, dedicated prediction models were built for the two categories of parts, with each model adapted to the specific demand sequence characteristics of its respective component. For low-frequency, intermittent parts, which have numerous periods of zero demand, traditional time-series models exhibit extremely high prediction bias. Therefore, an improved Crosston segmented prediction model was chosen, along with a Bootstrap resampling algorithm to quantify the uncertainty of the prediction results, outputting complete prediction results with confidence intervals. For mid-to-high-frequency parts, the demand sequence is continuous with few zero values, exhibiting clear time-series dependencies. Furthermore, demand fluctuations are significantly influenced by external factors such as vehicle operating conditions and seasons. Therefore, an LSTM (Long Short-Term Memory) neural network model, capable of fusing multi-dimensional external features, was selected. This model relies on time-series features and covariates to jointly complete demand predictions, independently outputting predicted values ​​for each type of part. The two models operate in parallel without interfering with each other.

[0030] Specifically, complete historical demand time-series data for low-frequency intermittent spare parts is retrieved, and the original one-dimensional demand sequence is split into two independent sub-series. The first sub-series is the demand interval sequence, recording the number of cycles between two spare part demand occurrences; the second sub-series is the single demand quantity sequence, recording the quantity of spare parts consumed each time a demand occurs. After splitting, an exponential smoothing algorithm is used to smooth both sub-series to eliminate random fluctuation noise within the time series. The core hyperparameter of the exponential smoothing operation is the smoothing coefficient, with a value range of 0.05 to 0.3. A grid search algorithm is used to traverse all candidate parameter combinations within the range, using the minimum mean absolute error of the validation set prediction as the optimization objective to select the optimal smoothing parameter. In the exemplary scenario, the optimal smoothing parameter value of 0.1 is determined after grid search and is used for smoothing calculations of both the demand interval and demand quantity sequences.

[0031] Subsequently, after completing the exponential smoothing calculations for the two sets of sequences, the estimated values ​​of the next cycle demand interval and the single demand quantity are obtained respectively. The single demand quantity estimate is divided by the demand interval estimate to obtain the average demand forecast per unit cycle. Based on this average value, the total demand for future multiple cycles is extrapolated. The preset forecast period can be set to four weeks, twelve weeks, or other durations according to the enterprise's business needs. In the exemplary scenario, the average demand forecast for each week of the next four weeks is output as the basic forecast value.

[0032] Furthermore, to quantify the uncertainty of low-frequency component forecast results, a Bootstrap resampling operation with replacement is introduced, setting a fixed number of resampling times, exemplarily selecting one thousand repeated samplings. Each sampling randomly selects an equal number of samples from all historical demand data for the component. The sampling process allows the same historical data to be selected repeatedly multiple times, and each set of sampling generates a completely new simulation demand dataset. Each simulation dataset is then individually processed using the complete improved Crosston model, outputting a set of corresponding predicted means. After one thousand samplings, one thousand independent predicted values ​​are obtained. All values ​​are integrated to form an empirical distribution set of the component's demand forecast values, fully covering all possible fluctuation ranges of the forecast results.

[0033] Next, the company pre-determines the required confidence level for its business, with a commonly used confidence level of 90%. The corresponding quantiles are the 2.5% lower quantile and the 97.5% upper quantile. The program sorts the 1,000 predicted values ​​in ascending order, extracts the values ​​at the corresponding quantile positions as the upper and lower boundaries of the interval, and combines them to form a complete prediction confidence interval. The final low-frequency parts prediction output includes two items: the average predicted demand per unit period and the upper and lower boundaries of the confidence interval. For example, the predicted average for injectors is 0.8 units per week, with a 90% confidence interval of 0.2 to 1.6 units per week. The complete results are simultaneously transmitted to the downstream inventory optimization module.

[0034] For mid-to-high frequency demand parts, a complete historical demand time series of mid-to-high frequency parts is retrieved. Simultaneously, external auxiliary features are matched one-to-one according to the time period. Covariates are divided into four dimensions: first, vehicle operation dimension, including average vehicle mileage and ambient temperature within the period; second, seasonal dimension, dividing into spring, summer, autumn, and winter, converted into digital features using one-hot encoding; third, maintenance time dimension, the average time per maintenance within the period; and fourth, usage area dimension, dividing into urban, suburban, and mountainous usage scenarios, also digitized using one-hot encoding. All covariate features are bound to the corresponding periodic demand data, forming a complete training dataset with multi-dimensional auxiliary features, compensating for the deficiency of a single time series in capturing the influence of external operating conditions.

[0035] Finally, a covariate fusion LSTM neural network prediction model was constructed. The overall model structure consists of four layers. The first layer is the input layer, where the input feature dimension is determined by the length of the historical demand sequence and the number of covariate features. For example, the historical sequence input length is set to twelve weeks, and the covariate features are eight-dimensional, resulting in an overall input feature dimension of eight. The second layer consists of two stacked LSTM hidden layers, each with sixty-four neurons, used to extract long-term dependencies within the time series. The third layer is a random deactivation layer with a deactivation ratio of 0.2 to reduce the risk of model overfitting. The fourth layer is a fully connected output layer that outputs a single value, namely the predicted demand value for the next cycle. The model input is fixed as a continuous twelve-week historical demand sequence combined with all covariate features of the corresponding cycle, and a single operation outputs the predicted demand value for the next single cycle.

[0036] The entire historical dataset with covariates is split chronologically. Data from earlier time periods is used as the training set for iterative model parameter updates, while data from later time periods is used as an independent validation set to evaluate the model's generalization ability. For example, data from the first two years are used for training, and data from the first half of the following year is used for validation. An early stopping strategy is implemented during model iteration, with a patience value of five. If the prediction error on the validation set does not decrease after five consecutive iterations, training is terminated to avoid overfitting. After training, the set of network weights with the lowest validation set error is retained as the optimal model. This model is then directly called to predict the demand for mid-to-high-frequency components, outputting a single, definitive cycle demand prediction value as the standard prediction result for mid-to-high-frequency components.

[0037] Furthermore, in some optional embodiments of the present invention, since the proportion of component consumption samples corresponding to high-probability frequent failures in the sampling is exactly the same as the weight of samples for low-probability occasional failures, the empirical distribution generated by resampling excessively amplifies the demand fluctuations caused by rare failures, resulting in distorted and overly wide prediction confidence intervals. Therefore: Before performing the resampling operation, a fault mode hierarchical weighted sampling constraint mechanism is introduced, abandoning the uniform and equal probability random sampling method commonly used in this field. The specific implementation process is as follows: First, retrieve the faulty parts demand matrix, count the total number of occurrences of each fault mode cluster in the entire statistical period, and calculate the global occurrence probability of each fault mode. Second, traverse all historical repair samples corresponding to the low-frequency parts, match the fault mode cluster to which each sample belongs, and use the global occurrence probability of the corresponding fault mode as the basic sampling weight of the sample. Third, set a dynamic weight truncation threshold. For extremely rare fault samples with a global occurrence probability of less than 0.05, uniformly compress their basic sampling weight to the truncation threshold of 0.05 to prevent extremely low-frequency fault samples from excessively interfering with the overall sampling distribution. Fourth, perform polynomial weighted sampling with replacement based on the corrected weights of all samples. The sampling probability of each sample is proportional to its own weight, and one thousand weighted samplings are performed to generate one thousand sets of differentiated simulation demand datasets.

[0038] Step S13: Based on the demand characteristics of the accessory, the adaptive selector selects the prediction result of the corresponding model for output, or merges the prediction results of two models of the same accessory and outputs the final demand prediction value.

[0039] The adaptive selector is an independently running program module, internally storing two sets of execution logic: the judgment rules for component demand classification and the multi-model fusion calculation logic. During program execution, it first reads two feature parameters—demand density and zero-value ratio—to determine the demand category of the component. For transitional components with ambiguous boundaries and demand characteristics between low-frequency and mid-to-high-frequency, instead of selecting the output of a single model, it retrieves the predicted values ​​output by two different models and performs a weighted fusion calculation to obtain a comprehensive predicted value. For standard low-frequency and standard mid-to-high-frequency components with clearly distinguishable characteristics, the adaptive selector directly matches the corresponding dedicated model and directly outputs the predicted value generated by that model. Finally, it uniformly outputs a single standardized final demand predicted value, which serves as the basic input data for subsequent inventory optimization calculations.

[0040] For example, in this embodiment, the adaptive selector is an independent scheduler module. Specifically, it can be implemented using any independent module among threshold rule determination subroutine, small logistic regression classification scheduling subroutine, and multi-model API microservice middleware. The module runs decoupled from the two types of prediction models. Internally, it integrates demand category determination logic and multi-model weighted fusion calculation logic. By reading the component demand density and zero value ratio feature parameters, it distinguishes component types and executes the corresponding branch logic of directly outputting the single model result or weighted fusion outputting the dual model prediction results. Finally, it outputs a unified and standardized component demand prediction value.

[0041] Specifically, for transitional components whose demand characteristics fall between low-frequency and mid-to-high-frequency and cannot be directly categorized, the program synchronously calls two independent prediction models to perform parallel operations. It retrieves the predicted mean with confidence intervals from the improved Croston model and the single-cycle demand prediction value from the LSTM covariate model, respectively. The two sets of prediction values ​​are stored in the cache module as the two sets of basic input data for fusion calculation. The two sets of prediction values ​​correspond to the same prediction period to ensure a unified calculation benchmark.

[0042] Linear weighted fusion is selected as the basic fusion algorithm. The fusion calculation formula is: the final predicted value equals the Crosston model prediction value multiplied by the first weight, plus the LSTM model prediction value multiplied by the second weight. The sum of the two weight values ​​is fixed at one. The weight values ​​are set from 0.1 to 0.9. A grid search algorithm is used to traverse all weight combinations. All weight combinations are substituted into all transitional accessory samples in the validation set to calculate the overall average absolute error. The weight set with the smallest overall error is selected as the globally optimal fusion weight. In the exemplary scenario, the grid search yields a Crosston model weight of 0.7 and an LSTM model weight of 0.3, which are used as fixed fusion weight parameters. In addition to linear weighted fusion, those skilled in the art can also replace the Stacking fusion algorithm and use a multi-level learner to fuse the prediction results of the two models, which also falls within the scope of this technical solution.

[0043] The two optimal weights determined by grid search are retrieved and substituted into the linear weighted fusion calculation formula to complete the numerical calculation, yielding the comprehensive final demand forecast value for transitional parts. This comprehensive numerical forecast takes into account both the low-frequency prediction logic driven by failure mechanisms and the time-series prediction logic of multi-condition covariates. Compared with the single model, the prediction error is significantly reduced. After the calculation is completed, the standardized final demand forecast value is directly output and simultaneously pushed to the downstream multi-level inventory optimization calculation module as basic input data. In the exemplary scenario, the average absolute error of the fused model test set is 0.3 units per week, which is a 26.3% reduction in error compared to the single Croston model.

[0044] In summary, the engine parts demand forecasting method based on repair data in the above embodiments of the present invention obtains engine warranty repair data from at least one business system of an enterprise, and generates a fault parts demand matrix through data cleaning and fault mode clustering, establishing a mapping relationship between fault modes and parts models. Based on the fault parts demand matrix, historical demand data for each part is statistically generated, and parts are divided into low-frequency intermittent demand parts and medium-to-high-frequency demand parts according to demand characteristics. Differential demand forecasting is performed for the two types of parts using an improved Crosston model incorporating the Bootstrap method and an LSTM model, respectively. An adaptive selector selects the corresponding model prediction result based on the part demand characteristics for output, or merges the prediction results of the two models for the same part to output the final demand forecast value. This avoids the shortcomings of existing technologies that rely on historical sales data and simple time series models, ignore the causal relationship between fault modes and parts demand, and cannot adaptively adapt to different intermittent demand characteristics with a single prediction model. It effectively reduces prediction errors and ultimately achieves accurate engine parts demand forecasting. This solves the problems of low accuracy and large errors in existing engine parts demand forecasting.

[0045] Example 2 This embodiment also proposes a method for predicting engine parts demand based on repair data. The difference between the engine parts demand prediction method based on repair data in this embodiment and the engine parts demand prediction method based on repair data in Embodiment 1 is as follows: The step of selecting the prediction result of the corresponding model based on the demand characteristics of the accessory using an adaptive selector, or merging the prediction results of two models for the same accessory and then outputting the final demand prediction value, further includes: Using the final demand forecast as the core demand input, a multi-level inventory optimization model integrating a horizontal transfer mechanism is constructed for a multi-level inventory architecture consisting of a central inventory node and multiple forward inventory nodes. Configure (s,Q,S) inventory strategies for each model of parts at each inventory node. Calculate the replenishment trigger threshold, replenishment batch and target inventory level for each node based on the final demand forecast. The optimization objective is to minimize the sum of inventory holding cost, stockout cost and horizontal transfer cost, with preset service level and preset capital occupation limit as constraints. The Benders decomposition algorithm is used to iteratively solve the multi-level inventory optimization model. After iterating until the objective function meets the preset convergence condition, the inventory strategy parameters and horizontal transfer rules of each node are output.

[0046] The model uses the final cycle demand forecast of the output parts as the sole demand input parameter, adapting to the mainstream two-tier inventory architecture of enterprises. The architecture includes a regional central inventory node (RDC) and multiple forward inventory nodes (FDC), with an example configuration of one central warehouse and three forward warehouses. A new horizontal transfer mechanism constraint is added during the modeling process, allowing parts to be transferred between the central warehouse and forward warehouses, and between different forward warehouses. This overcomes the limitation of traditional multi-tier inventory models that only allow one-way replenishment between upper and lower levels. The transfer process is set with a fixed single transfer cost and a fixed transfer delay. An example single transfer cost is 500 yuan, and the transfer waiting time is 24 hours. A complete mixed-integer stochastic programming multi-tier inventory optimization model is constructed. The model internally configures a three-stage (s, Q, S) inventory management strategy for each engine part within each inventory node. The three parameters represent: s (representing the replenishment trigger threshold, i.e., safety stock), Q (representing the standard replenishment batch size), and S (representing the maximum target inventory level). All parameters are calculated based on the predicted demand for parts throughout the cycle. The model's optimization objective is to minimize the sum of three costs: the first cost is inventory holding cost, calculated as a fixed percentage of the annual value of the parts (e.g., 15%); the second cost is the stockout loss cost, with a fixed loss of 2,000 yuan per stockout; and the third cost is the logistics and labor cost incurred by horizontal transfers between warehouses at the same level, at 500 yuan per transfer. The model sets two sets of hard constraints: the first is a service level constraint, meaning the stockout rate for parts throughout the entire cycle cannot exceed a preset upper limit (e.g., 2%); the second is a total inventory capital constraint, meaning the total value of all parts inventory across all warehouses cannot exceed a preset capital limit (e.g., 3 million yuan). All optimization solutions must not exceed these two sets of constraints. This model belongs to the NP-hard mixed integer programming problem. Directly solving the whole problem is too computationally time-consuming. Therefore, the Benders decomposition algorithm is used to decompose the problem into two independent iterative subproblems. The first layer is the main problem, which specifically solves the (s,Q,S) inventory strategy parameters of all parts in each warehouse. The branch and bound method is used in conjunction with the commercial solver Gurobi to complete the numerical solution. The second layer is the lateral transfer subproblem, which transforms warehouse inventory, goods in transit, and predicted demand into Markov Decision Process (MDP). The Q-learning reinforcement learning algorithm is used to solve for the optimal transfer action. The state variables include the real-time inventory quantity of each warehouse, the quantity of goods in transit that have not been completed, and the periodic demand prediction value. The action variable is the quantity of parts transferred between warehouses in a single transaction. The reward function is set to the negative of the transfer cost and the stockout loss. The hyperparameters of the algorithm are set as follows: learning rate of 0.1, discount factor of 0.9, and exploration rate linearly decaying from 0.5 to 0.05. Iterative training is conducted for 1,000 rounds. The two-layer problem is iterated alternately. After the sub-problem is solved, Benders cut is generated and fed back to the main problem for re-optimization. The convergence condition for the iteration is set to be that the change in the objective function value between two adjacent iterations is less than 0.1%. The iteration stops after the convergence standard is reached. The (s,Q,S) strategy parameters of all parts of all inventory nodes and the complete rules for horizontal transfer triggering and quantity allocation between warehouses are uniformly output. The total time for a single complete solution is controlled within 20 seconds, which meets the near real-time calculation requirements of the enterprise inventory system.

[0047] Furthermore, in some optional embodiments of the present invention, after the step of selecting the prediction result of the corresponding model based on the demand characteristics of the accessory using an adaptive selector, or merging the prediction results of two types of models for the same accessory and then outputting the final demand prediction value, the method further includes: The strategy update process is automatically triggered according to the preset update cycle. The newly added warranty repair data since the previous update cycle is extracted incrementally and merged with the historical dataset to obtain the updated full basic data. Based on the updated full set of basic data, the fault mode and part model mapping relationship update, demand forecasting model retraining, and multi-level inventory strategy optimization are performed in sequence to generate inventory strategy parameters and horizontal transfer rules for the new cycle. The parameters of the new cycle's inventory strategy are compared with the parameters of the previous cycle's strategy. If the difference exceeds the preset difference threshold, a strategy update instruction is automatically sent to the enterprise inventory execution system.

[0048] Enterprises can pre-configure a fixed automatic update cycle, setting either a weekly or monthly update frequency. For example, an update program can be set to automatically start at 2:00 AM every Monday. After the program starts, it executes incremental data retrieval logic, extracting only newly generated warranty repair records after the previous update cycle, avoiding repeated reading of all historical data to reduce computational overhead. For example, 800 to 1200 repair records are added per week. After the incremental data extraction is complete, the new data is merged with the historical data stored in the system's sliding window. The sliding window retains all repair data from the past 24 months, automatically removing outdated data outside the window's time frame. The merged data generates a completely new full-scale basic dataset, serving as the unified data source for this iteration's update. Based on the newly merged full set of basic data, a three-layer iterative update operation is performed in a fixed sequence. The first layer updates the mapping relationship of faulty parts, and re-executes the entire process of fault code standardization, K-Shape time-series clustering, and demand matrix statistics. If the new clustering result shows a change of more than 30% in the composition of parts within more than two fault clusters compared to the previous version, the system automatically pushes a review email to remind business personnel to manually verify. The second layer updates the demand forecasting model. For low-frequency parts, the smoothing parameters of the Crosston model are directly updated and improved. For medium and high-frequency LSTM models, incremental fine-tuning training is used. The weights of the first layer network are frozen, and only the parameters of the last layer network are retrained, which greatly shortens the training time. The third layer re-executes the multi-level inventory optimization solution, and recalculates the parameters (s,Q,S) of all warehouse parts and the horizontal transfer rules using Benders decomposition combined with reinforcement learning algorithms. After all three layers of operations are completed, a new set of complete inventory control strategies for the next cycle is generated. The program automatically reads the parameters of the new inventory strategy for this round and compares them with the parameters of the old strategy currently being implemented in the warehouse system. It then compares the values ​​of the two sets of core parameters item by item across different parts and warehouses, using preset difference thresholds: a 20% threshold for the safety stock parameter (S) and a 15% threshold for the target inventory parameter (S). If the change in either parameter value for any part in any warehouse exceeds the corresponding threshold, the system automatically generates a standardized strategy update instruction and pushes it to the enterprise WMS warehouse management system via a RESTful application interface, synchronously updating the warehouse-side control parameters. If the change in all part parameters does not exceed the threshold, only an iteration log is recorded on the local server, and no update instruction is pushed to the warehouse system, reducing the operational pressure caused by frequent parameter adjustments. After each update, the system automatically records key indicators such as inventory capital occupation, overall warehouse stockout rate, and total number of monthly horizontal transfers before and after the update, retaining the iteration history logs long-term for business personnel to track the effectiveness of strategy optimization.

[0049] In summary, the engine parts demand forecasting method based on repair data in the above embodiments of the present invention obtains engine warranty repair data from at least one business system of an enterprise, and generates a fault parts demand matrix through data cleaning and fault mode clustering, establishing a mapping relationship between fault modes and parts models. Based on the fault parts demand matrix, historical demand data for each part is statistically generated, and parts are divided into low-frequency intermittent demand parts and medium-to-high-frequency demand parts according to demand characteristics. Differential demand forecasting is performed for the two types of parts using an improved Crosston model incorporating the Bootstrap method and an LSTM model, respectively. An adaptive selector selects the corresponding model prediction result based on the part demand characteristics for output, or merges the prediction results of the two models for the same part to output the final demand forecast value. This avoids the shortcomings of existing technologies that rely on historical sales data and simple time series models, ignore the causal relationship between fault modes and parts demand, and cannot adaptively adapt to different intermittent demand characteristics with a single prediction model. It effectively reduces prediction errors and ultimately achieves accurate engine parts demand forecasting. This solves the problems of low accuracy and large errors in existing engine parts demand forecasting.

[0050] Example 3 Please see Figure 2 The image shows an engine parts demand prediction device based on repair data proposed in the third embodiment of the present invention. The device includes: The acquisition module 100 is used to acquire engine warranty repair data from at least one business system of the enterprise, and generate a fault parts demand matrix by sequentially performing data cleaning and fault mode clustering, and establish a mapping relationship between fault modes and parts models. The segmentation module 200 is used to generate historical demand data for each part based on the faulty parts demand matrix, and to segment the parts based on the historical demand data, so as to divide the parts into low-frequency intermittent demand parts and medium- and high-frequency demand parts. The prediction module 300 is used to generate prediction results by using an improved Crosston model that incorporates the Bootstrap method for low-frequency intermittent demand parts, and by using an LSTM model for demand prediction parts for medium- and high-frequency demand parts. The generation module 400 is used to select the prediction result of the corresponding model based on the demand characteristics of the accessory through an adaptive selector, or to output the final demand prediction value after fusing the prediction results of two types of models for the same accessory.

[0051] Example 4 In another aspect, the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in any one of Embodiments 1 to 2 above.

[0052] Example 5 In another aspect, the present invention provides an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of any one of the methods described in Embodiments 1 to 2 above.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0055] More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0056] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0057] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0058] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting engine parts demand based on repair data, characterized in that, The method includes: Engine warranty repair data is obtained from at least one business system of the enterprise. The data is then cleaned and clustered by failure mode to generate a failure parts demand matrix and establish a mapping relationship between failure modes and parts models. Based on the demand matrix of faulty parts, historical demand data for each part is generated. Based on the historical demand data of the parts, the demand is divided into low-frequency intermittent demand parts and medium- and high-frequency demand parts. For components with low-frequency intermittent demand, an improved Crosston model incorporating the Bootstrap method is used to generate demand forecasts. For components with medium- to high-frequency demand, an LSTM model is used to generate demand forecasts. The adaptive selector selects the corresponding model's prediction result output based on the demand characteristics of the accessory, or merges the prediction results of two models for the same accessory and outputs the final demand prediction value.

2. The engine parts demand forecasting method based on repair data according to claim 1, characterized in that, The steps involved in generating the faulty parts demand matrix, through data cleaning and fault mode clustering, include: Perform data cleaning on the warranty repair data to remove records with missing fields and abnormal usage, and obtain a valid repair data set; Standardize the coding of fault codes in valid repair data, map the original fault codes to a preset multi-level fault classification system, and unify the fault coding rules; K-Shape or DBSCAN algorithms are used to cluster the fault occurrence time series to generate several fault mode clusters; The demand frequency and average single usage of each part model under each failure mode cluster are statistically analyzed, and a failure part demand matrix is ​​constructed with failure mode as the row and part model as the column. The mapping accuracy is verified by random sampling and manual verification. When the accuracy is lower than a preset threshold, the mapping rules are updated and optimized.

3. The engine parts demand forecasting method based on repair data according to claim 1, characterized in that, The steps for classifying parts into low-frequency, intermittent demand parts and medium-to-high-frequency demand parts based on historical demand data include: Calculate the demand density and zero-value ratio for each component, where demand density is the ratio of the number of demand occurrence cycles to the total number of statistical cycles, and zero-value ratio is the ratio of the number of zero demand cycles to the total number of statistical cycles. Set a first threshold and a second threshold. If the demand density of a component is less than the first threshold or the proportion of zero values ​​is greater than the second threshold, it is determined to be a low-frequency intermittent demand component, and the rest are determined to be high-frequency demand components.

4. The engine parts demand forecasting method based on repair data according to claim 1, characterized in that, For components with low-frequency intermittent demand, an improved Crosston model incorporating the Bootstrap method is used to generate demand forecasts. For components with medium- to high-frequency demand, an LSTM model is used to generate demand forecasts. The steps include: For low-frequency intermittent demand components, demand interval sequences and single demand quantity sequences are generated based on the corresponding historical demand data. Exponential smoothing is used to smooth the two types of sequences respectively. The smoothing parameters are optimized and determined by grid search algorithm. Based on the estimated demand intervals and estimated demand quantities obtained through smoothing calculations, the average demand forecast for the future preset period is calculated and output. The Bootstrap resampling method is used to resample the historical demand data several times with replacement. Based on each group of resampled data, the Croston forecast value is calculated to construct an empirical distribution of the demand forecast value. Based on the preset confidence level, the corresponding upper and lower quantile values ​​are extracted from the empirical distribution to generate the demand forecast confidence interval. The demand forecast mean and the corresponding confidence interval are used together as the forecast result for low-frequency intermittent demand components. For parts with medium to high demand, extract the corresponding historical demand data and match it to obtain multi-dimensional covariate feature data for the corresponding statistical period. The covariate features cover vehicle operation, season, maintenance hours and usage area dimensions. Construct an LSTM prediction model that integrates covariate inputs, using historical demand sequences of a preset length and covariate features of the corresponding periods as model inputs, and outputting the predicted value of parts demand for the next single period. Historical data is divided into training and validation sets to iteratively train the LSTM prediction model. During the training process, an early stopping strategy is used to suppress model overfitting. After the model converges, the optimal model parameters are saved, and the prediction results of high-frequency demand parts are output.

5. The engine parts demand forecasting method based on repair data according to claim 4, characterized in that, The steps for merging the prediction results of two models for the same component and outputting the final demand forecast include: Obtain the demand forecast results of the same component using the improved Croston model and the LSTM prediction model respectively; Based on the validation dataset, a linear weighted fusion method is used to fuse the two types of prediction results. With the goal of achieving the best prediction accuracy, the fusion weights corresponding to the two types of models are determined by a grid search algorithm. The final demand forecast value for the accessory is calculated and output based on the optimal fusion weight obtained through optimization.

6. The engine parts demand forecasting method based on repair data according to claim 1, characterized in that, The process of selecting the corresponding model's prediction result based on the demand characteristics of the component using an adaptive selector, or merging the prediction results of two models for the same component to output the final demand prediction value, also includes: Using the final demand forecast as the core demand input, a multi-level inventory optimization model integrating a horizontal transfer mechanism is constructed for a multi-level inventory architecture consisting of a central inventory node and multiple forward inventory nodes. Configure (s,Q,S) inventory strategies for each model of parts at each inventory node. Calculate the replenishment trigger threshold, replenishment batch and target inventory level for each node based on the final demand forecast. The optimization objective is to minimize the sum of inventory holding cost, stockout cost and horizontal transfer cost, with preset service level and preset capital occupation limit as constraints. The Benders decomposition algorithm is used to iteratively solve the multi-level inventory optimization model. After iterating until the objective function meets the preset convergence condition, the inventory strategy parameters and horizontal transfer rules of each node are output.

7. The engine parts demand forecasting method based on repair data according to claim 6, characterized in that, The process of selecting the corresponding model's prediction result based on the demand characteristics of the component using an adaptive selector, or merging the prediction results of two models for the same component to output the final demand prediction value, also includes: The strategy update process is automatically triggered according to the preset update cycle. The newly added warranty repair data since the previous update cycle is extracted incrementally and merged with the historical dataset to obtain the updated full basic data. Based on the updated full set of basic data, the fault mode and part model mapping relationship update, demand forecasting model retraining, and multi-level inventory strategy optimization are performed in sequence to generate inventory strategy parameters and horizontal transfer rules for the new cycle. The parameters of the new cycle's inventory strategy are compared with the parameters of the previous cycle's strategy. If the difference exceeds the preset difference threshold, a strategy update instruction is automatically sent to the enterprise inventory execution system.

8. An engine parts demand prediction device based on repair data, characterized in that, The device includes: The acquisition module is used to obtain engine warranty repair data from at least one business system of the enterprise, and generate a fault parts demand matrix by sequentially performing data cleaning and fault mode clustering, and establish a mapping relationship between fault modes and parts models. The segmentation module is used to generate historical demand data for each part based on the demand matrix of faulty parts, and to segment the parts based on the historical demand data, so as to divide the parts into low-frequency intermittent demand parts and medium- and high-frequency demand parts. The forecasting module is used to generate forecast results for low-frequency intermittent demand components by using an improved Croston model that incorporates the Bootstrap method, and for medium- and high-frequency demand components by using an LSTM model. The generation module is used to select the prediction results of the corresponding model based on the demand characteristics of the accessory through an adaptive selector, or to output the final demand prediction value after fusing the prediction results of two models of the same accessory.

Images