A machine learning-based method for predicting depreciation of equipment assets

Abstract

The application discloses a kind of equipment asset depreciation prediction methods based on machine learning, comprising the following steps: S1, acquisition asset original value, depreciation code, life, starting date and historical net value, normalization generates the asset state input sequence in uniform time domain;S2, the normalized asset state input neural differential equation model, constructs the differential system of asset state evolution with time;S3, identify the parameter change during the depreciation of assets, extract change time and amplitude, construct disturbance event sequence;S4, according to the disturbance event sequence constructs time gate function and disturbance coupling function;S5, inject the disturbance coupling function into the differential system, and obtain the depreciation evolution path by solving;S6, the predicted path is mapped to actual time interval, and the depreciation trend prediction result is output.The present application realizes the fine prediction of equipment depreciation trend and the dynamic response of strategy change, and improves the intelligent level of asset management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment asset management, and particularly relates to a device asset depreciation prediction method based on machine learning. BACKGROUND

[0002] In the process of enterprise fixed asset management, the depreciation prediction of equipment assets is a key basic work for asset evaluation, financial budget, scrapping planning and revaluation decision. The traditional depreciation calculation method mainly relies on the fixed formula rule in the financial system, and the common methods include the straight-line method, the balance declining method, etc. Usually, the depreciation period, the depreciation rate and the asset original value are taken as parameters, and the linear or nonlinear deduction of the depreciation value is carried out according to the fixed period. This kind of method is simple in calculation and clear in operation, and is suitable for accounting needs, but cannot reflect the influence of the real use state and the strategy change of the equipment on the asset value evolution path.

[0003] In the prior art, most of the depreciation simulation systems are integrated in the ERP platform or the asset management system. The common method is to manually adjust the depreciation parameters on the basis of the asset master data, trigger the batch depreciation recalculation module, and output the simulation results. Some systems can support setting different depreciation code, period or starting date parameters, and form multiple comparable depreciation curves through branch simulation, which are used to assist strategy selection. However, this kind of method relies on static rule logic processing, and does not have the ability to learn the historical depreciation behavior and predict the future trend, and cannot depict the dynamic influence of strategy disturbance on the time axis.

[0004] In recent years, the prediction method based on machine learning has been gradually applied in asset management. Some methods try to fit the historical net book value sequence through multilayer perceptron, recurrent neural network or time series model, so as to output the predicted value in the future period. These methods improve the prediction flexibility to a certain extent, but most of them rely on discrete time modeling structure, which is difficult to process non-equidistant depreciation records, and is also difficult to inject structured strategy change behavior into the model. In addition, different assets have different depreciation starting points, which makes it difficult to align the time series, resulting in inconsistent structure of the model training samples. For the behaviors such as depreciation code switching and period modification commonly seen in actual operation, this kind of model can only be processed as input features, and cannot express its real-time intervention effect on state evolution in the prediction process.

[0005] The current lacks a technical solution capable of modeling the net asset value evolution path in the continuous time domain, supporting multiple depreciation policy disturbance injection, and having a clear structure of the prediction solution mechanism. The existing method cannot achieve fine modeling of policy adjustment at any time in the asset life cycle on the future depreciation trend, and cannot simultaneously meet the modeling needs of time continuity, disturbance response and prediction interpretability. Therefore, there is an urgent need for a continuous prediction method that establishes a dynamic function mapping relationship between the initial state of the asset, the historical depreciation behavior and the policy change, to solve the problem of accurate modeling of time-varying disturbances in asset depreciation trend prediction.

[0006] Therefore, how to provide a machine learning-based equipment asset depreciation prediction method is a problem that those skilled in the art need to solve. SUMMARY

[0007] One object of the present application is to provide a machine learning-based equipment asset depreciation prediction method. The present application combines asset historical state sequence modeling and policy disturbance injection mechanism, constructs a neural ordinary differential system for continuous solution on a unified time axis, and completes the dynamic prediction process of the equipment depreciation trajectory through asset state normalization processing, policy change event identification, time gating function construction and disturbance function coupling. The method has the advantages of high time modeling accuracy, clear policy response structure and continuous smooth prediction results, and is suitable for asset evaluation, depreciation simulation, financial calculation and other business scenarios.

[0008] According to the machine learning-based equipment asset depreciation prediction method of the embodiment of the present application, the following steps are included:

[0009] S1, collect data of the equipment asset, construct an asset state sequence dataset, perform time normalization processing on the asset state sequence of different depreciation starting dates, and generate a normalized asset state input sequence;

[0010] S2, input the normalized asset state input sequence into a neural ordinary differential equation prediction model, construct a differential equation system with asset state change over time as the modeling target, parameterize the right function of the differential equation system through a neural network function, and generate an asset depreciation evolution path using a numerical solver;

[0011] S3, identify the depreciation policy change behavior in the asset life cycle, extract the change time, change type and change amplitude of each policy change operation, and construct a depreciation policy disturbance event sequence;

[0012] S4, construct a time gating function, dynamically control whether to inject a disturbance item into the differential system according to the activation time of the depreciation policy disturbance event sequence, and establish a disturbance coupling function;

[0013] S5. Inject the perturbation coupling function into the neural ordinary differential equation prediction model to form a perturbation coupled differential system. Perform continuous-time numerical solution on the system and output the net asset value prediction trajectory after considering the influence of strategy perturbation.

[0014] S6. Map the net asset value forecast trajectory to the actual time interval to form the corresponding asset depreciation trend forecast result.

[0015] Optionally, the time normalization process in S1 includes: taking the depreciation start date of each asset as the normalization starting point, mapping the corresponding state time series to the standard time interval [0,1] according to its life cycle, and normalizing the time point using the formula. Calculations are performed, in which, The original time point, The starting time for asset depreciation. The normalized timestamp will be used as the end time of the current observation period. The combination of asset state variables corresponding to a given time point forms an input pair. Construct a normalized asset state input sequence.

[0016] Optionally, S2 specifically includes:

[0017] S21. Input each pair of data in the normalized asset state input sequence. The input is fed into the state encoder network, where, Represents a normalized timestamp. Represent the asset state variables at the corresponding point in time, including the original asset value, depreciation code, depreciation period, and net book value, and output the initial hidden state vector. ;

[0018] S22, Based on the initial hidden state Construct a system of ordinary differential equations of the form:

[0019] ;

[0020] in, For normalized time The hidden state below, For trainable neural network parameters, This is a neural network function used to represent the derivative term of asset status changes over time;

[0021] S23, within the normalized time interval Set the set of numerical solution time points above ,by , Using ordinary differential equation solvers as boundaries, continuous-time numerical solutions are performed on the system of ordinary differential equations, calculating at each time point. Corresponding hidden state ;

[0022] S24, each time point Corresponding hidden state The input is fed into a decoder network to decode and generate the corresponding predicted net book value of the assets. To form a sequence of predicted values .

[0023] Optionally, S3 specifically includes:

[0024] S31. Obtain the policy parameter change records of equipment assets within the depreciation period, filter out historical change operations that include depreciation code changes, depreciation period modifications, or depreciation start date adjustments, and extract the time point of each change record. Parameter values ​​before the change Compared with the changed parameter values ;

[0025] S32. For each change operation, calculate the disturbance value based on the change type. When the change involves a depreciation code change, the parameter values ​​before and after the change are encoded into discrete vectors, and the encoding difference is calculated as... When the depreciation period is changed, the calculation is performed. When the depreciation start date is changed to an adjustment, the normalized time offset corresponding to the time difference before and after the adjustment is calculated as... ;

[0026] S33, based on the change of time point With corresponding disturbance value Constituting a disturbance event And arrange them in chronological order to generate a sequence of disturbance events. ,in, Indicates the number of valid strategy changes that occur within the asset's lifecycle;

[0027] S34, the disturbance event sequence Stored in the disturbance management structure to complete the unified encapsulation of asset disturbance events.

[0028] Optionally, the disturbance value in S32 for the depreciation code change type The calculation operation also includes: adjusting the depreciation code parameter value before the change. With the changed depreciation code parameter value The generated encoded vectors are normalized using a vector normalization function. The normalized vector difference is taken as the final perturbation value, denoted as . For all disturbance values The dimensions were uniformly adjusted to a fixed length, and zero padding was used to complete the dimension alignment.

[0029] Optionally, S4 specifically includes:

[0030] S41, within the normalized time interval Set a fixed step size Generate time point series ,in, , , This represents the total number of discrete time steps.

[0031] S42, at each point in time Initialize the perturbation value and traverse the sequence of perturbation events. For each disturbance event Determine whether the conditions are met. If satisfied, then the perturbation value will be... Add the perturbation value corresponding to the current time point. The accumulated result is used as the new perturbation value at the time point;

[0032] S43. Arrange the disturbance values ​​corresponding to all time points into a vector in chronological order. And according to the time sequence The values ​​are stored in the corresponding order to form a time series of perturbation values;

[0033] S44. At each point in time Invoke the hidden state of assets Calculate the output of the right-hand function of the neural network. The disturbance value corresponding to the time point Add them together to generate an update item. ,Will Store the updated vector sequence .

[0034] Optionally, the time series of perturbation values ​​generated in S43 The following processing steps are included before storage: for each perturbation value in the sequence... Perform a numeric type conversion operation to convert all numeric values. Mapped to a fixed-dimensional floating-point vector; if the original perturbation value is a scalar, it is expanded to a length of [value missing] according to a preset rule. The sparse vectors are filled with zeros, and precision normalization is performed on all perturbation vectors, normalizing all elements of each vector to the interval by the maximum absolute value. The processed vectors maintain the original sequence order and form a uniform format vector structure.

[0035] Optionally, S5 specifically includes:

[0036] S51, reading the sequence of normalized time points with the corresponding sequence of update terms , reading the initial hidden state , setting the differential solving time step , constructing the array of hidden state sequences and assigning the initial state to ;

[0037] S52, initializing the loop index , performing the following iteration operation: at each time step , reading the previous state from the array , reading the corresponding update term from the array , performing the Euler integration operation to calculate the hidden state of the current time step

[0038] ;

[0039] writing the result to the th item of the array ; updating the index to , repeating the process until ;

[0040] S53, establishing the array of equity value prediction values , initializing the loop index , in each loop, reading the hidden state from the array , inputting into the decoder network, and performing forward inference to calculate the equity value prediction value , writing the calculation result to the th item of the array , updating the index , repeating until ;

[0041] S54, establishing a dual-channel data structure in the memory structure for storing the sequence of normalized time points and the array of equity value prediction values, respectively, pairing each pair according to the index, and writing each pair to the structure field to complete the timing matching and caching operation of the prediction sequence.

[0042] Optionally, the S6 specifically comprises:

[0043] ​​​​S61, read the normalized time point sequence, the book net worth prediction value sequence, the depreciation start time and the depreciation end time, determine the actual time interval corresponding to the depreciation;

[0044] S62, for each normalized time point in the normalized time point sequence, calculate its corresponding time point in the actual time interval, and write all the calculated actual time points in sequence into the actual time point sequence;

[0045] S63, combine each actual time point in the actual time point sequence with the prediction value at the same index position in the book net worth prediction value sequence into a group of data pairs, and write all the data pairs in index order into the asset net worth prediction mapping structure;

[0046] S64, perform format standardization processing on each data pair in the asset net worth prediction mapping structure, convert each actual time point to a uniform format timestamp, adjust each prediction value to a set floating point precision value, and write it in time order into the depreciation trend prediction result table.

[0047] The beneficial effects of the present application are:

[0048] The continuous time modeling of the equipment asset depreciation process is realized: by introducing the structure of neural differential equation, the present application can gradually integrate the asset state on the normalized time axis, avoiding the shortcomings of traditional discrete models in time precision and interpolation fitting, and can more finely depict the dynamic change process of asset net worth, adapting to the unified modeling needs of non-equidistant data input and multiple starting time assets.

[0049] A disturbance injection mechanism that can respond to strategy changes is constructed: for the behaviors of depreciation code switching, life adjustment, start time correction, etc. in the asset depreciation process, the present application embeds strategy changes in the form of coupling terms into the differential system through disturbance event sequence extraction and time gating function construction, supports the triggering of strategy disturbance at any time point to intervene in the state evolution process, and effectively improves the fitting ability of the model to the real asset depreciation behavior.

[0050] A prediction result process with clear structure and standardized output is formed: by mapping the model solution results with the actual time interval and completing the format processing of the prediction values, the present application maintains the stability of the calculation process while facilitating the interface with existing asset management systems, improving the system compatibility and business landing efficiency of the prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0052] Fig. 1 A machine learning-based equipment asset depreciation prediction method according to the present application;

[0053] Fig. 2 An asset state normalization and time axis mapping diagram of a machine learning-based equipment asset depreciation prediction method according to the present application;

[0054] Fig. 3 A net value prediction trajectory mapping and output diagram of a machine learning-based equipment asset depreciation prediction method according to the present application. DETAILED DESCRIPTION

[0055] The present application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only show the basic structure of the present application in a schematic manner, and thus only show the components related to the present application.

[0056] REFERENCE Figs. 1-3 A machine learning-based equipment asset depreciation prediction method, comprising the following steps:

[0057] S1, collect data of equipment assets, construct an asset state sequence dataset, perform time normalization processing on asset state sequences of different depreciation starting dates, and generate a normalized asset state input sequence;

[0058] S2, input the normalized asset state input sequence into a neural ordinary differential equation prediction model, construct a differential equation system with asset state change over time as the modeling target, parameterize the right-hand function of the differential equation system through a neural network function, and generate an asset depreciation evolution path using a numerical solver;

[0059] S3, identify depreciation strategy change behavior in the asset life cycle, extract the change time, change type and change amplitude of each strategy change operation, and construct a depreciation strategy disturbance event sequence;

[0060] S4, construct a time gating function, dynamically control whether to inject a disturbance term into the differential system according to the activation time of the depreciation strategy disturbance event sequence, and establish a disturbance coupling function;

[0061] S5, inject the disturbance coupling function into the neural ordinary differential equation prediction model to form a disturbance-coupled differential system, perform continuous-time numerical solution on the system, and output an asset net value prediction trajectory considering the influence of strategy disturbance;

[0062] S6, map the asset net value prediction trajectory to the actual time interval to form a depreciation trend prediction result corresponding to the asset.

[0063] This invention introduces a perturbation-coupled neural ordinary differential equation structure for equipment asset depreciation prediction, realizing the fusion modeling of three types of information: asset status, time series, and strategy parameters. This overcomes the structural defects of traditional discrete models that cannot express dynamic strategy changes, and effectively improves the continuity of depreciation trajectory and the accuracy of prediction response.

[0064] In this embodiment, the time normalization process in S1 includes: taking the depreciation start date of each asset as the normalization starting point, mapping the corresponding state time series to the standard time interval [0,1] according to its life cycle, and normalizing the time point using the formula. Calculations are performed, in which, The original time point, The starting time for asset depreciation. The normalized timestamp will be used as the end time of the current observation period. The combination of asset state variables corresponding to a given time point forms an input pair. Construct a normalized asset state input sequence.

[0065] This invention introduces a unified time normalization mechanism in the asset state construction process, standardizes and aligns the state sequence based on the asset start time, solves the problem of training sample time misalignment caused by inconsistent depreciation start times of multiple assets, and improves the consistency of model input structure and expressive power.

[0066] In this embodiment, S2 specifically includes:

[0067] S21. Input each pair of data in the normalized asset state input sequence. The input is fed into the state encoder network, where, Represents a normalized timestamp. Represent the asset state variables at the corresponding point in time, including the original asset value, depreciation code, depreciation period, and net book value, and output the initial hidden state vector. ;

[0068] S22, Based on the initial hidden state Construct a system of ordinary differential equations of the form:

[0069] ;

[0070] in, For normalized time The hidden state below, For trainable neural network parameters, This is a neural network function used to represent the derivative term of asset status changes over time;

[0071] S23, within the normalized time interval Set the set of numerical solution time points above , with , as boundaries, the neural ordinary differential equation system is solved using a continuous-time numerical solver to calculate the corresponding hidden state at each time point ;

[0072] S24, input the corresponding hidden state at each time point to the decoder network to generate the corresponding asset net worth prediction value , forming a prediction value sequence .

[0073] The present application adopts neural ordinary differential equation to model the depreciation process, replaces the traditional discrete structure with time-continuous modeling method, allows the asset net worth to evolve in differential form over time, has higher time resolution and numerical stability, and is suitable for non-equidistant accounting record scenarios.

[0074] In the embodiment, the S3 specifically comprises:

[0075] S31, obtain the strategy parameter change record of the equipment asset in the depreciation period, filter out the historical change operations including depreciation code change, depreciation life modification or depreciation start date adjustment, extract the time point , the parameter value before change and the parameter value after change ;

[0076] S32, for each change operation, calculate the perturbation value , wherein when the change is depreciation code change, the parameter values before and after the change are encoded into a discrete vector and the encoding difference is calculated as , when the change is depreciation life modification, the , when the change is depreciation start date adjustment, the normalized time offset corresponding to the time difference before and after the adjustment is calculated as ;

[0077] S33, the change time point and the corresponding perturbation value constitute a perturbation event pair , and are arranged in chronological order to generate a perturbation event sequence , wherein represents the number of valid strategy changes in the asset life cycle;

[0078] S34, store the perturbation event sequence in the perturbation management structure, and complete the unified packaging of the asset perturbation events.

[0079] The present application represents the depreciation code, the life, the starting date and the like change operation in a structured manner by constructing the disturbance event sequence, forms the clear disturbance injection path, provides the logical entrance for the subsequent time dynamic control and the strategy response mechanism, and improves the strategy change modeling precision.

[0080] In the present embodiment, the calculation operation of the disturbance value of the depreciation code change type in S32 Further includes: the normalized processing of the generated code vector of the depreciation code parameter value before change and the depreciation code parameter value after change Adopt the vector normalization function , the normalized vector difference is taken as the final disturbance value, and is recorded as All disturbance values are uniformly adjusted to the fixed dimension length, and the dimension alignment is completed by using the zero padding mode.

[0081] The present application introduces the disturbance vector normalization and dimension uniform operation in the depreciation code disturbance processing process, so that the code difference value has the numerical stability and the consistency of the input vector, and improves the learnability of the disturbance to the model and the training stability of the neural network.

[0082] In the present embodiment, S4 specifically includes:

[0083] S41, a fixed step is set on the normalized time interval , and a time point sequence is generated, wherein , , is the total number of discrete time steps;

[0084] S42, the disturbance value is initialized at each time point , and the disturbance event sequence is traversed, for each disturbance event pair , whether the condition is met is judged, if met, the disturbance value is added to the disturbance value corresponding to the current time point, and the accumulation result is taken as the new disturbance value of the time point;

[0085] S43, the disturbance values corresponding to all time points are composed into a vector in time sequence, and are stored in the corresponding order of the time point sequence , forming the disturbance value time sequence;

[0086] S44, at each time point , the asset hidden state Calculate the output of the right-hand function of the neural network. The disturbance value corresponding to the time point Add them together to generate an update item. ,Will Store the updated vector sequence .

[0087] This invention designs a time-gated control structure coupled with disturbance logic. By activating the disturbance value under specific time conditions and writing it into the right-hand function of the differential equation, the coupling and fusion of policy change and predicted state path are realized, which effectively improves the model's dynamic response capability to policy switching behavior.

[0088] In this embodiment, the time series of perturbation values ​​generated in S43 The following processing steps are included before storage: for each perturbation value in the sequence... Perform a numeric type conversion operation to convert all numeric values. Mapped to a fixed-dimensional floating-point vector; if the original perturbation value is a scalar, it is expanded to a length of [value missing] according to a preset rule. The sparse vectors are filled with zeros, and precision normalization is performed on all perturbation vectors, normalizing all elements of each vector to the interval by the maximum absolute value. The processed vectors maintain the original sequence order and form a uniform format vector structure.

[0089] This invention performs dimensionality standardization and precision normalization on the perturbation vector, unifies the vector structure of the perturbation input, reduces modeling errors caused by inconsistent input dimensions, and enhances the model's compatibility with multiple types of perturbation values ​​and training convergence performance.

[0090] In this embodiment, S5 specifically includes:

[0091] S51. Read the normalized time point sequence With corresponding update item sequence Read the initial hidden state Set the time step for differential solution Construct a hidden state sequence array and assign the initial state value ;

[0092] S52. Initialize the loop index Perform the following iterative operation: at each time step From array Read the previous state From array Read the update item corresponding to the previous time step. Perform Euler integration to calculate the hidden state at the current time step.

[0093] ;

[0094] Write the results into an array The Item; Update the index to Repeat the process until ;

[0095] S53. Establish an array of predicted book net value values. Initialize the circular index In each loop, from the array Read hidden state ,Will The input is fed into the decoder network to perform forward inference to calculate the predicted net book value. Write the calculation results into an array The Item, update index Repeat until ;

[0096] S54. Establish a dual-channel data structure in the memory structure to store the normalized time point sequence separately. array of predicted net book value Pair them one by one according to the index, and pair each pair Write the data into the structure fields to complete the time-series matching and caching operations for the predicted sequence.

[0097] This invention constructs a hidden state update process based on numerical integration methods, and performs Euler iterative calculations in conjunction with perturbation-coupled update terms, thereby avoiding the state jump problem in the discrete modeling process and ensuring the continuity of state evolution and the accuracy and stability of differential system solutions.

[0098] In this embodiment, S6 specifically includes:

[0099] S61. Read the normalized time point sequence, the book net value prediction sequence, the depreciation start time and the depreciation end time, and determine the actual time interval corresponding to the depreciation.

[0100] S62. For each normalized time point in the normalized time point sequence, calculate its corresponding time point in the actual time interval, and write all the calculated actual time points into the actual time point sequence in order.

[0101] S63. Combine each actual time point in the actual time point sequence with the predicted value at the same index position in the book net value prediction value sequence into a data pair, and write all data pairs into the asset net value prediction mapping structure according to the index order.

[0102] S64, perform format standardization processing on each group of data in the net asset value prediction mapping structure, convert each actual time point to a unified format timestamp representation, adjust each prediction value to a set floating point precision value, and write the depreciation trend prediction result table in chronological order.

[0103] The present application maps the normalized time points in the prediction results to the actual time interval, completes the time alignment and format standardization of the result output by structuring the actual time points and net value prediction values, and facilitates the integration and visualization of the prediction trajectory in the asset management system.

[0104] Embodiment 1:

[0105] In order to verify the feasibility of the present application in implementation, the present application is applied to the depreciation prediction scene of the equipment assets of a large manufacturing enterprise. The enterprise has more than 12000 fixed assets, covering heavy processing equipment, transmission lines, power equipment, intelligent storage equipment and various asset types, with large number of assets, various types, high value and frequent changes. The management department is faced with many realistic problems such as frequent adjustment of depreciation strategy, lagging prediction work and distorted simulation results.

[0106] In actual operation, the depreciation simulation system of the enterprise originally uses the standard depreciation engine integrated with the ERP system, only supports the static simulation function after the depreciation code and the service life are changed, cannot reflect the real-time influence of strategy adjustment on future net value, and all predictions are completed based on fixed time step and rule calculation method, lacking unified processing ability for different equipment depreciation start times. In the budget preparation period of multiple projects, the deviation between the equipment financial prediction results and the final financial clearance results is more than 14% on average, and the deviation of individual high-value equipment is even more than 28%, causing budget misjudgment and financial lag.

[0107] In this embodiment, the method of the present application is deployed in the asset management platform of the enterprise, the neural ordinary differential equation is used to build the depreciation evolution path, the original value, depreciation code, service life, start depreciation time and historical net value record of all equipment assets are converted to a standard format, and time normalization processing is performed. For equipment assets with depreciation strategy adjustment behavior, the historical adjustment record is called in the background, the disturbance time point and parameter difference are extracted, and the disturbance event sequence is automatically generated.

[0108] Take a certain group of typical equipment as an example, select 90 heavy numerical control equipment put into use after 2017, the original value is in the interval of 1.6 million to 3 million, the depreciation period is unified for 10 years, adopt the balance decreasing method, the initial depreciation time span is between January 2017 and December 2019. The historical data shows that 37 devices have modified the depreciation code or depreciation period in the running process due to policy adjustment. For this kind of equipment with depreciation strategy disturbance, the present application dynamically injects the disturbance factor into the prediction model by constructing a time gating function and a disturbance coupling function.

[0109] The present application method is used to predict the future 36 months of the net book value of the equipment cluster, and compared with the traditional static rule method and the LSTM model without disturbance learning. In the prediction accuracy evaluation, the actual depreciation book value is used as a reference, and the average absolute error (MAE) and relative error (MAPE) of the three methods in different time periods are calculated respectively. The results show that: the overall average absolute error of the present application method on all equipment for 36 months is 18,300 yuan, and the relative error is 3.1%; while the relative error of the static rule simulation method is 10.7%, and the relative error of the LSTM model is 6.2%. The error improvement of the present application is the most obvious on the equipment with high frequency of strategy adjustment.

[0110] At the same time, the consistency of the prediction results and the actual depreciation trend is visualized and analyzed, and the net value trajectory curve output by the present application shows a smooth response and a continuous transition before and after the change of the depreciation code, without inflection point mutation or depreciation fault, with high explainability and simulation fidelity. In the project financial simulation, the monthly budget deviation of the present application prediction value is controlled within ±3%, which is about 70% less than the prediction fluctuation of the traditional method.

[0111] Table 1 shows the error comparison data of the original asset parameters, strategy change history, actual depreciation net value and the prediction results of the present application of some heavy equipment, all data are from the business records in the real running environment:

[0112] Table 1: Comparison table of heavy equipment depreciation prediction effect

[0113] ;

[0114] The above Table 1 shows the depreciation prediction effect of the present application on typical heavy equipment assets, covering the original asset parameters of seven groups of equipment, strategy change, comparison data of actual net value and predicted net value, and reflecting the precision level of the prediction results of the present application through mean absolute error (MAE) and relative error indicators. From the asset original value distribution, the sample equipment original value range is between 1.604 million yuan and 3 million yuan, which has a certain representativeness. In terms of depreciation strategy, there are 5 equipment with depreciation code or depreciation period adjustment behavior, and 2 equipment remain unchanged, which can be used for comparative analysis of the response ability of the prediction model to disturbance events.

[0115] Comparing the actual net value with the predicted net value of the present application, it can be seen that the prediction results of all samples are highly close to the actual net value at the 36th month, with the maximum absolute error of 34,900 yuan and the minimum of only 3,600 yuan. The relative error is balanced, and most samples are less than 2%, and the lowest is only 0.23%. In the equipment with strategy change, such as HDX-014 and HDX-050, the present application still accurately tracks the net value change after the adjustment of the depreciation period, showing strong disturbance response ability. In the equipment HDX-035 and HDX-078 with unchanged strategy, the predicted value is highly consistent with the actual value, indicating that the model has stable non-disturbance depreciation trajectory learning ability.

[0116] In addition, for equipment with high value and high depreciation intensity (such as HDX-066), the prediction result of the present application still maintains an error level of 1.96% in the context of large changes, indicating that the control ability on high-value assets is also effective. Overall, the data shown in Table 1 verifies that the present application is superior to the traditional depreciation estimation method in terms of unified time modeling, strategy disturbance processing and prediction accuracy control, and has engineering feasibility for popularization and application in complex industrial asset environment.

[0117] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine learning-based method for predicting equipment asset depreciation, characterized in that, Includes the following steps: S1. Collect equipment asset data, construct an asset status sequence dataset, perform time normalization processing on asset status sequences with different depreciation start dates, and generate a normalized asset status input sequence. S2. Input the normalized asset state input sequence into the neural network differential equation prediction model to construct a differential equation system with the asset state changing over time as the modeling objective. The differential equation system parameterizes the right-hand side function through a neural network function and uses a numerical solver to generate the asset depreciation evolution path. S3. Identify depreciation strategy changes during the asset lifecycle, extract the change time, change type, and change magnitude of each strategy change operation, and construct a sequence of depreciation strategy disturbance events; S4. Construct a time-gating function to dynamically control whether the disturbance term is injected into the differential system based on the activation time of the disturbance event sequence of the depreciation strategy, and establish a disturbance coupling function. S4 specifically includes: S41, within the normalized time interval Set a fixed step size Generate time point series ,in, , , This represents the total number of discrete time steps. S42, at each point in time Initialize the perturbation value and traverse the sequence of perturbation events. For each disturbance event Determine whether the conditions are met. If satisfied, then the perturbation value will be... Add the perturbation value corresponding to the current time point. The accumulated result is used as the new perturbation value at the time point; S43. Arrange the disturbance values ​​corresponding to all time points into a vector in chronological order. And according to the time sequence The values ​​are stored in the corresponding order to form a time series of perturbation values; S44, at each point in time Invoke the hidden state of assets Calculate the output of the right-hand function of the neural network. The disturbance value corresponding to the time point Add them together to generate an update item. ,Will Store the updated vector sequence ; S5. Inject the perturbation coupling function into the neural ordinary differential equation prediction model to form a perturbation coupled differential system. Perform continuous-time numerical solution on the system and output the net asset value prediction trajectory after considering the influence of strategy perturbation. S5 specifically includes: S51. Read the normalized time point sequence With corresponding update item sequence Read the initial hidden state Set the time step for differential solution Construct a hidden state sequence array and assign the initial state value ; S52. Initialize the loop index Perform the following iterative operation: at each time step From array Read the previous state From array Read the update item corresponding to the previous time step. Perform the Euler integration operation to calculate the hidden state at the current time step: ； Write the results into an array The Item; Update the index to Repeat the process until ; S53. Establish an array of predicted book net value values. Initialize the circular index In each loop, from the array Read hidden state ,Will The input is fed into the decoder network to perform forward inference to calculate the predicted net book value. Write the calculation results into an array The Item, update index Repeat until ; S54. Establish a dual-channel data structure in the memory structure to store the normalized time point sequence separately. array of predicted net book value Pair them one by one according to the index, and pair each pair Write the data into the structure fields to complete the time-series matching and caching operations for the predicted sequence; S6. Map the net asset value forecast trajectory to the actual time interval to form the corresponding asset depreciation trend forecast result.

2. The method for predicting equipment asset depreciation based on machine learning according to claim 1, characterized in that, The time normalization process in S1 includes: taking the depreciation start date of each asset as the normalization starting point, mapping the corresponding state time series to the standard time interval [0,1] according to its life cycle, and normalizing the time point using the formula. Calculations are performed, in which, The original time point, The starting time for asset depreciation. The normalized timestamp will be used as the end time of the current observation period. The combination of asset state variables corresponding to a given time point forms an input pair. Construct a normalized asset state input sequence.

3. The method for predicting equipment asset depreciation based on machine learning according to claim 2, characterized in that, S2 specifically includes: S21. Input each pair of data in the normalized asset state input sequence. The input is fed into the state encoder network, where, Represents a normalized timestamp. Represent the asset state variables at the corresponding point in time, including the original asset value, depreciation code, depreciation period, and net book value, and output the initial hidden state vector. ; S22, Based on the initial hidden state Construct a system of ordinary differential equations of the form: ； in, For normalized time The hidden state below, For trainable neural network parameters, This is a neural network function used to represent the derivative term of asset status changes over time; S23, within the normalized time interval Set the set of numerical solution time points above ,by , Using ordinary differential equation solvers as boundaries, continuous-time numerical solutions are performed on the system of ordinary differential equations, calculating at each time point. Corresponding hidden state ; S24, each time point Corresponding hidden state The input is fed into a decoder network to decode and generate the corresponding predicted net book value of the assets. To form a sequence of predicted values .

4. The equipment asset depreciation prediction method based on machine learning according to claim 3, characterized in that, S3 specifically includes: S31. Obtain the policy parameter change records of equipment assets within the depreciation period, filter out historical change operations that include depreciation code changes, depreciation period modifications, or depreciation start date adjustments, and extract the time point of each change record. Parameter values ​​before the change Compared with the changed parameter values ; S32. For each change operation, calculate the disturbance value based on the change type. When the change involves a depreciation code change, the parameter values ​​before and after the change are encoded into discrete vectors, and the encoding difference is calculated as... When the depreciation period is changed, the calculation is performed. When the depreciation start date is changed to an adjustment, the normalized time offset corresponding to the time difference before and after the adjustment is calculated as... ; S33, based on the change of time point With corresponding disturbance value Constituting a disturbance event And arrange them in chronological order to generate a sequence of disturbance events. ,in, Indicates the number of valid strategy changes that occur within the asset's lifecycle; S34, the disturbance event sequence Stored in the disturbance management structure to complete the unified encapsulation of asset disturbance events.

5. The machine learning-based equipment asset depreciation prediction method according to claim 4, characterized in that, The perturbation value for the depreciation code change type in S32 The calculation operation also includes: adjusting the depreciation code parameter value before the change. With the changed depreciation code parameter value The generated encoded vectors are normalized using a vector normalization function. The normalized vector difference is taken as the final perturbation value, denoted as . For all disturbance values The dimensions were uniformly adjusted to a fixed length, and zero padding was used to complete the dimension alignment.

6. The method for predicting equipment asset depreciation based on machine learning according to claim 5, characterized in that, The time series of disturbance values ​​generated in S43 The following processing steps are included before storage: for each perturbation value in the sequence... Perform a numeric type conversion operation to convert all numeric values. Mapped to a fixed-dimensional floating-point vector; if the original perturbation value is a scalar, it is expanded to a length of [value missing] according to a preset rule. The sparse vectors are filled with zeros, and precision normalization is performed on all perturbation vectors, normalizing all elements of each vector to the interval by the maximum absolute value. The processed vectors maintain the original sequence order and form a uniform format vector structure.

7. The method for predicting equipment asset depreciation based on machine learning according to claim 6, characterized in that, S6 specifically includes: S61. Read the normalized time point sequence, the book net value prediction sequence, the depreciation start time and the depreciation end time, and determine the actual time interval corresponding to the depreciation. S62. For each normalized time point in the normalized time point sequence, calculate the corresponding time point in the actual time interval, and write all the calculated actual time points into the actual time point sequence in order. S63. Combine each actual time point in the actual time point sequence with the predicted value at the same index position in the book net value prediction value sequence into a data pair, and write all data pairs into the asset net value prediction mapping structure according to the index order. S64. Perform format standardization processing on each data pair in the net asset value forecast mapping structure, convert each actual time point into a unified format timestamp representation, adjust each forecast value to the set floating-point precision value, and write it into the depreciation trend forecast result table in chronological order.

Images