Method, system and equipment for refilling cash box based on machine learning and medium

Abstract

On the basis of testing of an entity machine by the inventor, the invention provides a special framework which takes two parallel running models as an integrated model method. The integrated model method is provided as a physical system that runs two models simultaneously, both of which are trained as candidate models. The two models are respectively used in the inference phase to optimize the loss function (e.g., using mean absolute error representation, MAE representation), and the model with better MAE representation is used to control the generation of ATM refill control signals in the deduction phase. The two models are used together, including a first model, a fully connected neural network data architecture, and a second model, a decision tree-based learning algorithm, used as a gradient boosted framework (e.g., a lightweight gradient boosted framework, also referred to as LightGBM). From the perspective of practical application, the integrated model can be configured with a prediction buffer area so as to carry out specific parameter adjustment.

Description

Technical Field

[0001] The present invention relates to machine learning / artificial intelligence, and more particularly to a method, system, device and medium for refilling a banknote box based on machine learning, wherein the technical solution uses physical sensor data and improved prediction to control the physical refilling operation. Background Art

[0002] Currently, the management of physical cash box refilling (refilling includes the operation of removing cash from the cash box and putting cash into the cash box) operations is inefficient and inaccurate, which has affected the availability of cash in automatic teller machines (i.e., ATMs). There are deficiencies in using manual forecasting methods to predict ATM cash demand because such forecasting methods are inaccurate for predicting current cash holdings and do not accurately consider dynamic factors that affect cash deposit and withdrawal rates, such as seasonality, holidays, public events, and recent withdrawal trends.

[0003] This results in ATMs either running low on cash, impacting customer service, or holding too much cash, which is not cost-effective. This results in ATMs sometimes needing emergency refills or having to undergo more maintenance than necessary, incurring unnecessary operating costs. These deficiencies lead to increased refilling times and rising costs associated with third-party cash delivery services. In some cases, cash refill delivery cycles can take up to 36 hours, which can affect the overall availability of the ATM if the changing cash needs cannot be quickly responded to. Summary of the invention

[0004] These existing technical problems have prompted us to develop an improved machine learning / artificial intelligence system that is configured as a specific solution for controlling the refilling box. The improved technical solution provides a physical control tool for optimizing the cash dispensing process of an automatic teller machine (ATM), solving the problem of predicting the cash demand of the ATM. There are many specific types of machines similar to ATMs. Although ATMs are mentioned in the present invention, they can also be cash deposit machines (CDMs), certain ATMs that can deposit and withdraw cash at the same time, known as multi-function machines (MFMs), and multi-currency machines that can handle multiple currencies (for example, a machine used in airports). The so-called "out of service" refers to the moment when the customer cannot interact with the machine. At this time, the box either has too much cash (i.e., no deposits can be made) or too little cash (i.e., no withdrawals can be made). One solution to out of service is to request a real-time refill. However, one goal of the present invention is to minimize the number of times real-time refills are required, thereby reducing the total number of transports.

[0005] The physical control tool is coupled to a real-time ATM sensor feedback stream and sends control information to the logistics controller and scheduling system to control the refilling activity. At the same time, the refilling operation can be tracked in real time in the ATM sensor feedback stream. In some embodiments, the present invention can generate a specific path for the cash delivery vehicle responsible for the refilling operation. The present invention provides an artificial intelligence / machine learning-based system for tracking and analyzing captured real-time ATM data, for example, such tracking and analysis can be based on physical sensor data inputs and data sets obtained from users' physical interactions with ATMs. These data inputs are processed by machine learning algorithms, which can be suitable for processing factors such as seasonality, holidays, public events, and recent withdrawal trends, so as to accurately predict the amount of cash required for each ATM. Using machine learning can achieve a more dynamic and responsive cash distribution strategy. The real-time data stream of the ATM may include physical sensor data, which, after processing, provides a prediction solution by updating the current withdrawal pattern to a trained machine learning model.

[0006] The present invention proposes a specific architecture that utilizes two models running in parallel as an integrated model solution based on testing of a physical machine. The integrated model solution is a physical system that can run two models at the same time, and both models are trained as candidate models. In the deduction stage, the two models are each used to optimize the loss function (for example, using the mean absolute error performance, MAE performance), and in the deduction stage, the model with better MAE performance is used to control the generation of ATM refill control signals. The two models used together include a first model, which is a fully connected neural network data architecture, and a second model, which is a decision tree-based learning algorithm, used as a gradient boosting framework (for example, a lightweight gradient boosting framework, also known as LightGBM). From a practical application perspective, the integrated model can be configured with a prediction buffer for specific parameter adjustments.

[0007] At runtime, the forecasting system can be configured to run periodically (e.g., every night) to forecast cash deposits and generate settlement instructions based on the forecast data output, and upload small batches of data at a higher frequency (e.g., every 15 minutes). In addition, the model predictions and retraining can be used to generate cash settlement instructions, which can be configured to control one or more cash-in-transit logistics operations. In a variation of this scheme, the AI / ML-based system is configured to generate machine outputs to directly control and provide cash refills for cash boxes by generating and submitting cash refill logistics requests instead of or in addition to the graphical user interface.

[0008] The predictive system can be optimized for different uses and operations, such as to improve cassette utilization, reduce the total number of clearing times, and / or reduce outages (thereby improving service availability). The system can be configured to synchronize operations with real-time generated data, thereby acting as an automatic monitoring system to automatically or semi-automatically operate to predictively control refill operations.

[0009] In some embodiments, entropy can be introduced during the generation of refill control commands, by injecting noise to modify the path and operation timing, thereby reducing the possibility of physical attack on cash in transit by increasing unpredictability. However, this noise injection will also reduce the certainty of the optimal refill timing. As a physical output, a graphical user interface, such as a dashboard, can intuitively visualize real-time withdrawal patterns, allowing users to make informed decisions and respond quickly to cash needs. In the present invention, this solution has been demonstrated to reduce cash refill delivery time from 36 hours to 15 minutes.

[0010] In some embodiments, there may be multiple ATM groups, wherein multiple ATMs in a particular location may be selectively refilled. For example, an ATM group (e.g., five ATMs) provides services to customers coming out of a stadium entrance and exit at the same location. If one of the ATMs is out of money, the customer can just use another ATM that has money. In this scenario, each ATM in the ATM group is considered to be a member of the entire group, in which case deactivation is tracked only when the total amount of money or other money output of all ATMs in the ATM group drops below a certain threshold. That is, the entire ATM group is considered to be one ATM.

[0011] In one aspect, the present invention provides a machine learning-based cash box refilling method, comprising:

[0012] Instantiate a machine learning data model architecture using an integrated model, the integrated model includes a fully connected neural network framework architecture and a gradient boosting framework architecture, the gradient boosting framework architecture uses a decision tree-based algorithm and applies a leaf-node growing scheme, the fully connected neural network architecture is coupled to the gradient boosting framework architecture; receive a training data set and train the integrated model using a loss function to establish a trained integrated model, the training data set obtains one or more historical data sets, the loss function is suitable for reducing the total number of banknote box refills; receive a deduction data set based on a real-time ATM sensor feedback stream and provide it to the trained integrated model; use the trained integrated model to generate one or more predicted data outputs of each ATM coupled to the real-time ATM sensor feedback stream; for each predicted data output of the one or more predicted data outputs that exceeds a threshold, automatically generate one or more data information for controlling a refill or removal operation associated with the corresponding coupled ATM; confirm one or more refill or removal operations based on physical sensor readings from the real-time ATM sensor feedback stream; and generate an updated deduction data set for subsequent operation of the trained integrated model.

[0013] In a further embodiment, downstream usage of the refilled ATM is used to update the trained ensemble model.

[0014] In a further embodiment, the fully connected neural network includes a ReLU layer network combined with a sigmoid layer.

[0015] In a further embodiment, a reduce-increase-reduce framework is used to determine the number of nodes in each layer.

[0016] In a further embodiment, the gradient boosting framework is an LightGBM model.

[0017] In a further embodiment, batch normalization and dropout methods are used to control the activation function.

[0018] In a further embodiment, the mean absolute error (MAE) is used as the loss function.

[0019] In a further embodiment, the one or more data messages for controlling refill or removal operations associated with corresponding coupled ATMs are sequenced based on generated routes of regularly scheduled cash-in-transit vehicles.

[0020] In a further embodiment, the correspondingly coupled ATMs include a group of ATMs, wherein a group of ATMs located at the same location are identified as a single ATM.

[0021] In a further embodiment, the ensemble model employs batch normalization and dropout methods.

[0022] The present invention provides a machine learning-based cash box refilling system, comprising: a data storage module, configured to instantiate and store a machine learning data model architecture using an integrated model, the integrated model comprising a fully connected neural network framework architecture and a gradient boosting framework architecture, the gradient boosting framework architecture using a decision tree-based algorithm and applying a leaf node growth scheme, the fully connected neural network architecture coupled to the gradient boosting framework architecture; a model training module, configured to receive a training data set and train the integrated model using a loss function to establish a trained integrated model, the training data set acquiring one or more historical data sets, the loss function being suitable for reducing the total number of cash box refills; receiving a machine learning data model architecture based on real-time ATM sensor feedback; A deductive data set of the feed is provided to the trained integrated model; a model prediction module is configured to use the trained integrated model to generate one or more predicted data outputs for each ATM coupled to the real-time ATM sensor feedback stream; a logistics control module is configured to automatically generate one or more data information for controlling a refill or removal operation associated with the corresponding coupled ATM for each predicted data output exceeding a threshold value in the one or more predicted data outputs; based on physical sensor readings from the real-time ATM sensor feedback stream, confirm one or more refill or removal operations; and the model training module is further configured to generate an updated deductive data set for subsequent operation of the trained integrated model.

[0023] In a further embodiment, downstream usage of the refilled ATM is used to update the trained ensemble model.

[0024] In a further embodiment, the fully connected neural network includes a ReLU layer network combined with a sigmoid layer.

[0025] In a further embodiment, a reduce-increase-reduce framework is used to determine the number of nodes in each layer.

[0026] In a further embodiment, the gradient boosting framework is an LightGBM model.

[0027] In a further embodiment, batch normalization and dropout methods are used to control the activation function.

[0028] In a further embodiment, the mean absolute error (MAE) is used as the loss function.

[0029] In a further embodiment, the one or more data messages for controlling refill or removal operations associated with corresponding coupled ATMs are sequenced based on generated routes of regularly scheduled cash-in-transit vehicles.

[0030] In a further embodiment, the correspondingly coupled ATMs include a group of ATMs, wherein a group of ATMs located at the same location are identified as a single ATM.

[0031] In a further embodiment, the ensemble model employs batch normalization and dropout methods.

[0032] In another aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored therein and executable on the processor, wherein the processor implements any of the steps of the machine learning-based cash box refilling method described above when executing the computer program. Due to space limitations, no further description is given here.

[0033] In another aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the machine learning-based cash box refilling method as described above. Due to space limitations, no further description is given here.

[0034] To better understand the specific embodiments below, the above content lists the features and technical advantages. Other features and technical advantages will be discussed below. Those skilled in the art will appreciate that the concepts and specific embodiments disclosed in the present invention can be used as a basis for modifying or designing other architectures at any time. Those skilled in the art can also recognize that such equivalent architectures do not violate the spirit and scope of the embodiments described herein. In the specific embodiments below, the novel features of the present invention in terms of organization and operating methods, as well as further purposes and advantages can be better understood in conjunction with the accompanying drawings. However, it should be clearly understood that each drawing is used for illustration and description purposes only and is not intended to be a definition of limitations on the embodiments described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In the accompanying drawings, the embodiments are described by way of example. It should be clearly understood that the embodiments and the accompanying drawings are only for illustration and to aid understanding.

[0036] The specific embodiments will be described below in conjunction with the accompanying drawings. The specific embodiments described below are only described as examples, wherein:

[0037] Figure 1 is a block diagram of an example model architecture and system for machine learning based cash box refilling according to some embodiments of the present invention.

[0038] Figure 2 is a logic flow diagram of a model selection scheme according to some embodiments of the present invention, and a list of example features of two models used in conjunction.

[0039] Figure 3 is an example data flow diagram illustrating an end-to-end scheme of data flow and model feedback flow according to some embodiments of the present invention.

[0040] Figure 4 is an example flow chart for illustrating cash refill logic according to some embodiments of the present invention.

[0041] Figure 5 is an example of a cash instruction that may be generated by the proposed system according to some embodiments of the present invention.

[0042] Figure 6 It is an example table that illustrates the experimental result output during the operation of the technical solution provided by the present invention under different buffer amounts, and compares it with the benchmark reference model.

[0043] Figure 7 is an example diagram of cash deposit data according to some embodiments of the present invention.

[0044] Figure 8 is an example diagram showing cash deposits illustrating potential periodicity according to some embodiments of the present invention.

[0045] Fig. 9 is a schematic diagram of an example fully connected neural network according to some embodiments of the present invention.

[0046] Fig.10 are example screenshots showing example code for implementing the system according to some embodiments of the present invention.

[0047] Fig.11 is an example schematic diagram of a gradient boosting framework using a decision tree-based algorithm and a decision leaf-based approach according to some embodiments of the present invention.

[0048] Fig.12 is an example connectivity graph showing an intermediate training diagram of LightGBM according to some embodiments of the present invention.

[0049] Fig.13 is a schematic diagram showing an example CDM-level deposit model according to some embodiments of the present invention.

[0050] Fig.14 is a schematic diagram of feature importance values ​​generated using an ARIMA model according to some embodiments of the present invention.

[0051] Fig.15is a chart showing cash deposits versus forecasts according to some embodiments of the present invention.

[0052] Fig.16 is a schematic diagram showing the relationship between CDM performance and the total number of deactivations according to some embodiments of the present invention.

[0053] Fig.17 2 is an example of a multi-ATM site according to some embodiments of the present invention, wherein each ATM has a different cash box containing cash of different amounts.

[0054] Fig.18 is an example graph showing the relationship between the number of refills and the number of days in a specified time period according to some embodiments of the present invention.

[0055] Fig.19 is an example chart showing the relationship between the optimal number of refills and the total cost and interest rate according to some embodiments of the present invention.

[0056] Fig. 20 is an example graph showing cash availability and total cost for each experimental subject when the improved experimental subject is expanded to all remote ATM experimental data for one year.

[0057] Fig.21 is an example user interface example with a simulation feature that enables a user to optimize a cash refill strategy according to some embodiments of the present invention. DETAILED DESCRIPTION

[0058] The present invention proposes an improved machine learning / artificial intelligence system and a corresponding method for controlling physical cash refilling. The improved machine learning / artificial intelligence system is an integrated model data architecture that utilizes two machine learning model data architectures that are trained and run simultaneously and in parallel in the deduction phase. Both models are used to generate prediction outputs optimized based on mean absolute error (MAE) performance scores. During the testing phase, it was found that different models can be used for different physical ATMs due to their different characteristics, thus better mapping to different model data architectures. Therefore, the use of an integrated model approach can bring better results. During the experimental phase, it was found that the scheme proposed by the present invention has achieved significant progress compared to the reference baseline scheme based on rule logic. The technical solution provided by the present invention is expressed as an optimization problem, the goal of which is to determine when a settlement instruction should be executed based on a time series prediction of the cash deposit amount in the next few days.

[0059] The system includes an ATM controller coupled to a plurality of ATMs at different locations, which reports sensory information through real-time ATM feedback based on physical sensors coupled to physical cash boxes. The data includes data objects, such as JSON files or XML files with field structures, and the fields include fields such as the terminal ID of each cash deposit machine (CDM), the CDM site location and the terminal ID associated with each CDM, the capacity of each cash box in each CDM (i.e., the maximum number of banknotes that can be accommodated in each cash box), the historical CDM cash balance of each CDM within every 15 minutes, the historical cash instructions of each settlement process of each CDM, and the historical cash box utilization percentage when each CDM is settled.

[0060] Figure 1 is a block diagram of an example model architecture and system for machine learning-based physical box refilling according to some embodiments of the present invention. Figure 1 In the embodiment, the system 100 is a physical control server configured to run in a data center, coupled to real-time ATM sensor feedback 102 from physical ATM sensors coupled to the cassettes 106 of individual ATMs 104 to measure the available capacity (e.g., by measuring spring pressure, size of the cavity) to estimate the number of banknotes in the cassettes 106. Thus, the real-time ATM sensor feedback 102 can represent the current state of the banknote holdings in any ATM cassette 106 at a given point in time. The real-time ATM sensor data feedback 102 can be arranged in a variety of electronic communications, such as by periodically polling it to obtain a data set for training / deduction, or in other embodiments, push-based communication based on a request message transmitted on an interrupt signal.

[0061] These data are processed as input to the machine learning model and may include variable information as shown in Table 1 below.

[0062]

[0063] Table 1

[0064] The system 100 includes a model training module 108, a model prediction module 110, and a data storage module (database) 112, which stores prediction model weights and filter parameters of an integrated model having at least a first model and a second model running together. A model selector 114 is configured to control which model output the model prediction module 110 uses and provides to generate logistics control information sent to a logistics controller 116. The logistics controller 116 is coupled to a scheduling system to control refilling behavior. In some embodiments, the logistics controller 116 can be configured to generate cash delivery or cash withdrawal instructions, and / or a path for cash delivery or cash withdrawal. The logistics controller 116 can be coupled to one or more cash-in-transit vehicles through the scheduling system to issue instruction sets and / or generate specific paths and waypoints for cash delivery or withdrawal as needed. Refilling can be tracked in real time in ATM sensor feedback, and in some embodiments, a specific route can be generated for the refilling vehicle. The ML system 100 is configured to generate a prediction output, i.e., a predicted cash withdrawal / deposit for the next duration (e.g., X hours later). Based on the predicted value and the current cash balance, the system 100 is configured to generate cash instructions to calculate the cash balance for the next duration, and the logistics controller 116 is configured to generate Cash-in-Transit (CIT) instructions to modify the cash balance so that sufficient holdings (for ATMs) or space (for cash deposit machines) are maintained in the corresponding cash boxes. In some embodiments, the logistics controller 116 is used to control a single cash box for a specific type of banknotes.

[0065] The logistics controller 116 controls the CIT trip scheduling according to a predetermined plan, and ideally refills are performed during non-working hours so that there is enough capacity to handle all transactions in a period. Each visit to the ATM site is counted to calculate the total number of trips and a fixed score can be assigned to the model to reward / penalize points. For the model, the objective function is to optimize the MAE of withdrawals, adjust the model, cash instruction logic, and cash box configuration to control the level of inactivity based on the number of CIT trips. In some embodiments, these parameters are configurable options that administrators can use to define which thresholds and configurations are the best choices. In another embodiment, the ML system is configured to reward / penalize points only based on the error terms corresponding to the actual deposits / withdrawals, and this information is used to control scheduling capabilities.

[0066] If the anticipated demand exceeds the physical capacity of the existing cassettes, in some embodiments, the logistics controller 116 is configured to control unplanned refill trips in advance.

[0067] The purpose of the time series forecasting model is to predict the amount of cash deposits in the next few days in order to recommend the corresponding cash settlement instructions to make the settlement trip at the appropriate time.

[0068] The model prediction module 110 is set by the scheduler to run at 00:00 every day to predict cash deposits and generate settlement instructions based on the prediction results. The scheduler uploads small batches of data every 15 minutes, triggers model prediction, triggers model retraining, etc. The model training module 108 is used to upgrade the model. The model prediction module 110 generates estimated deposit / withdrawal behavior information and provides it as output to the downstream cash settlement instruction generation module, so as to automatically generate cash instructions based on the model prediction results, and automatically calculate performance indicators using the cyclic report generation module, and summarize them into reports on a monthly and weekly basis.

[0069] like Figure 2 As shown, Figure 2 A logical flow is shown to show how the model prediction module 110 operates as a time series prediction module. An ensemble model based on LightGBM and fully connected neural network is adopted to model time series data. The project adopts one of the LightGBM model or the fully connected neural network model for each CDM, with the one with lower prediction error first. Time series prediction assumes that the future variance and mean are similar to the training data. However, this assumption may fail after an unexpected event. Users can configure the prediction buffer to reduce the risk of the model. If the external environment (such as COVID) causes a sudden change in deposits, users can adjust the buffer to prevent outage events. The operations team evaluates the model based on business indicators in weekly and monthly regular reports, including overall service availability, total outage hours, cash box utilization, and CIT trip times. If the model performance is not satisfactory, the development team will retrain the model. The project sponsor can review the model and further verify it in the group review meeting. The model can be optimized by adding or removing specific functions and using integrated models in production. The model can be enhanced by adding functions such as refill windows and cash costs.

[0070] Figure 2 The model output can include the following:

[0071] Forecast cash deposits

[0072] The output of the time series forecasting model is the cash deposits of all CDMs in the next 8 days. The predicted cash deposit amount is used to decide whether a specific CDM needs to be settled.

[0073] Cash Instruction Report

[0074] The configuration and manual logic defined by the user in the cash instruction generation module further adjusts the cash instruction results. The module output is a cash instruction that the business user can send to the supplier for cash settlement. An example of a cash instruction is shown below.

[0075] Cycle Report

[0076] The recurring report summarizes the model performance based on the business metrics defined by the user. Business users evaluate the model based on the results in the recurring report. The program generates a weekly recurring report at the end of each week and a monthly recurring report at the end of the month. An example of a recurring report is shown below.

[0077] Figure 2 is a simplified diagram illustrating a model selection scheme and an example feature list of two models used together in coordination.

[0078] As shown in system 200, the integrated approach is used to select the model that performs better for each CDM transaction, so both models are used simultaneously, so that the architecture is able to exploit the different characteristics of the two models during each game generation process.

[0079] In the process of determining the model approach, the first step in selecting a model is to try several possible models. The inventors did not try linear regression and exponential smoothing models because they are sensitive to outliers. In the cash deposit data, the inventors expected a large number of spikes, and the model was designed to predict these values ​​as accurately as possible.

[0080] During the experiment with different model structures, the LightGBM model and the fully connected neural network model performed better in terms of prediction error than the seasonal autoregressive integrated moving average model (SARIMA) and other linear models, so the inventors also abandoned the SARIMA model. Therefore, the models tried at the beginning included LightGBM and the fully connected neural network model. The inventors then carried out performance analysis and model selection. The overall model selection process includes two stages:

[0081] 1. Use statistical evaluation indicators to select models

[0082] 2. Verify through business performance indicators

[0083] During the experiment, different methods (MAE, MSE, RMSE, MAPE and R2) were first used to determine the configuration of each model. Then, the method included using business performance indicators in simulation experiments to verify the prediction results. Based on the results of these two stages, the inventors selected the model of the present invention.

[0084] Both the LightGBM model and the fully connected neural network model are very efficient because both use only one model to predict all CDMs. The inventors further investigated the performance of the models and found that for some machines, the LightGBM performed better, while for other machines, the fully connected neural network model was able to provide a more solid prediction.

[0085] In summary, the integrated model of the LightGBM model and the fully connected neural network model is a better model because it not only has the best statistical performance, but is also more efficient. The model performance is based on the current rule-based CDM settlement logic. Therefore, no other challenger models are needed in this usage scenario.

[0086] Through analysis, it is found that the deposit pattern during holidays and before holidays is different from the cash deposit pattern during normal times. During important holidays such as the Spring Festival and Christmas, the difference in deposit patterns is more obvious (some peaks or valleys appear). Therefore, these features are selected to make the model perform better during holidays to reduce deactivation events.

[0087] The fully connected neural network consists of a series of fully connected layers. Each output dimension depends on each input dimension. To prevent overfitting, the inventors used a random split of 0.75, 0.25 in the training and validation datasets. The neural network architecture started with a base network of three rectified linear units (ReLU) layers, and after several rounds of experiments, the inventors added a neural network layer. The inventors used a "reduce-increase-reduce" framework similar to convolutional neural networks to choose the number of nodes in each layer. The idea behind this network is to encode first, then decode to avoid too much information loss, and then encode again. Batch normalization ensures that there are no activations that are too high or too low, which allows for higher learning rates. Therefore, things that were previously untrainable can be trained. Batch normalization has a slight regularization effect, so it can reduce overfitting. Similar to the random dropout regularization technique, it also adds some noise to the activations of each hidden layer. Therefore, if batch normalization is used, less dropout can be used, which is technically beneficial because this method helps reduce information loss.

[0088] For some layers, a value of 0.1 is used as dropout. In addition, early stopping and a specific number of iterations can be used to find the convergence point of the model and avoid overfitting. Designing the FC model with a ReLU function in the output layer ensures that any predicted value is greater than 0. This ReLU layer can also avoid and solve the gradient vanishing problem.

[0089] The reason why the inventors use MAE instead of RMSE and MAPE as the loss function is that RMSE has too high a penalty for extreme points (driven by special events) and is not stable enough in normal times, while MAPE does not estimate the error well when the actual value is very large. In other words, when the actual cash deposit is low, the penalty for wrong prediction is high, but when the actual cash deposit is high, the penalty for wrong prediction is not enough. As for MAE, in the proposed technical solution, the inventors optimize the absolute value of the difference between the predicted value and the actual cash deposit. All machines are trained using one model, so the loss function will give higher penalties to machines with higher cash deposits.

[0090] LightGBM is a gradient boosting framework that uses a decision tree-based algorithm, using a leaf node growing method, while other algorithms use a hierarchical growing method. It is designed to have faster training speed and higher efficiency. It outperforms all other algorithms when the data set is extremely large. Compared with other algorithms, LightGBM takes less time to run on large data sets. Therefore, in this scenario, LightGBM is efficient because the inventors use the same model to train all machines.

[0091] With several parameters of the model, the inventors used grid search to adjust the hyperparameters, and finally selected a learning rate of 0.15, 2400 iterations, 150 estimators, and a maximum depth of 17. The inventors also selected 0.8 as the feature score, which means that LightGBM will randomly select 80% of the parameters for building a decision tree in each iteration. The combination of learning rate, tree structure, and feature score will improve accuracy while avoiding overfitting.

[0092] In view of the fact that the inventors found that some machines performed better in LightGBM, while others performed better in fully connected neural networks, the inventors proposed an integrated model with the best MAE performance for the two models, so as to consciously train and maintain the two models until a model with a shorter running time is selected. When the model code is executed as machine code on the computer CPU, MAE is used not only to maintain the consistency of the loss functions of the two models, but also to penalize machines with higher deposit values, because these machines are more likely to cause serious deactivation problems. The process of integrated model modeling is the process of using multiple different basic models (such as LightGBM and fully connected neural networks in the present technical solution) to predict results. The integrated model can reduce the generalization error of the results by integrating the prediction results of different models. The use of integrated models can not only improve accuracy, but also combat uncertainty in the data.

[0093] The inventor used the prediction results of the fully connected neural network and LightGBM as the integrated training period from August 2019 to October 2020, and calculated the MAE of the two models based on each terminal ID, and took the model with lower MAE for the terminal ID. After making the selection, the inventor obtained a list of terminals and the model corresponding to the terminal.

[0094] After a preliminary study of several modeling approaches, namely ARIMA / SARIMA, LightGBM, XGBoost, Catboost, and Fully Connected Neural Network (FCNN), LightGBM and FCNN were selected as the ensemble modeling scheme. The modeling scheme was tested on the same training and test sets and used the same feature set.

[0095] Table 2 shows the test results of 500 randomly selected machines, demonstrating the actual experimental conditions of various models.

[0096] Model Name RMSE MAE Training time (minutes) LightGBM 118911 77419 96 LinearRegressor 132438 90050 46 RandomForestRegressor 133078 89655 282 XGBRegressor 121490 79466 185 Keras Neural Network 129149 77303 202

[0097] Table 2

[0098] LightGBM and fully connected neural network models performed best in terms of MAE and were therefore chosen as ensemble models. For ARIMA / SARIMA, the model library is difficult to maintain (since there is only one model per machine), and training these models individually is slow and the performance of the models is not satisfactory. Compared to testing LightGBM / XGboost / Catboost, LightGBM requires the least training resources and is the fastest to train. The test dataset has thousands of machines, so training with XGboost requires a GPU. Catboost did not show significantly better results when training high-dimensional categorical features (geographic location / machine ID). Therefore, for the decision tree-based model, LightGBM was ultimately chosen because it only needs to train one model for all machine time series, which is easy to manage. In addition, the model does not need to be retrained frequently. The fully connected neural network model requires fewer resources to build and tune. Although other neural network architectures (such as RNN-based models) can be used, this will require a more extensive investigation of the architecture and hyperparameters.

[0099] Figure 3 300 is an example data flow diagram according to some embodiments of the present invention, which shows an end-to-end solution of data flow and model feedback flow. As shown in data flow 300, different periodic training can be performed for different types of queries and training. In this example, training can be performed in real time.

[0100] Figure 4is an example logic flow diagram for illustrating cash refill logic according to some embodiments of the present invention.

[0101] In the logic flow 400, Figure 3 A periodic scheme is used to control operations for different time periods, which also shows a scheme for generating cash instructions and controlling settlement operations, taking into account the combination of remaining capacity and cash deposit amount prediction.

[0102] Figure 5 is an example of a cash instruction that may be generated by a system according to some embodiments of the present invention. In cash instruction 500, a specific list of ATMs that need to be refilled is indicated, as well as specific amounts required for different currency banknotes and denominations.

[0103] Figure 6 is an example table showing the experimental results of the technical solution of the present invention during operation under different buffer amounts and compared with the baseline reference model. As shown in table 600, compared with the baseline reference model, the utilization of the machine cash box and the total number of CIT trips are reduced. However, it can also be found between the integrated model data architectures using different buffer amounts that optimizing the machine cash box utilization or reducing the total number of CIT trips will also have a corresponding impact on reducing the service availability, and also affect the total number of downtime hours.

[0104] Figure 7 700 is a schematic diagram of cash deposit data according to some embodiments of the present invention. In chart 700, the underlying regularity of ATM feedback data relative to external data (e.g., weather, sports events) begins to emerge, indicating that the underlying seasonality may be somewhat abnormal. Since periodicity may be non-linearly related and may further include entropy or regular differences, etc., it may be difficult for people to understand the concept of seasonality.

[0105] Figure 8 is an example diagram illustrating that cash deposits exhibit potential periodicity according to some embodiments of the present invention.

[0106] Fig. 9 is a diagram of an example fully connected neural network according to some embodiments of the present invention.

[0107] Fig.10 are example screenshots showing example code for implementing the system according to some embodiments of the present invention.

[0108] Fig.11 is a schematic diagram of a gradient boosting framework using a decision tree-based algorithm and a decision leaf-based approach according to some embodiments of the present invention.

[0109] Fig.12is an example connectivity graph illustrating a LightGBM midline training diagram according to some embodiments of the present invention, showing a LightGBM intermediate training graph.

[0110] Fig.13 is a schematic diagram showing an example CDM-level deposit model according to some embodiments of the present invention.

[0111] Fig.14 is a schematic diagram of feature importance values ​​generated using an ARIMA model according to some embodiments of the present invention.

[0112] Fig.15 is a chart showing cash deposits versus forecasts according to some embodiments of the present invention.

[0113] Fig.16 is a schematic diagram showing the relationship between CDM performance and the total number of deactivations according to some embodiments of the present invention.

[0114] Fig.17 2 is an example of a multi-ATM site according to some embodiments of the present invention, wherein each ATM has a different cash box containing cash of different amounts.

[0115] Fig.18 is an example graph showing the number of refills versus days for a specified time period according to some embodiments of the present invention.

[0116] As interest rates rise, funds held in the ATM incur significant opportunity costs. The higher the interest rate, the higher the cost of these funds. However, ATM costs come primarily from two sources: supplier costs and interest costs. A key variable that can affect both supplier costs and interest costs (cash costs) is the refill volume, as it affects the number of refill trips and the end-of-day balance. To reduce cash costs, the optimal refill volume can be determined for a given interest rate.

[0117] like Fig.18 As shown, the refill amount can be modeled based on two assumptions. The first assumption is that the remaining cash is 25% of the refill amount. This assumption can be modeled using the following formula.

[0118] R T =p*R T +W T =R*N(1)

[0119] The second assumption is that the withdrawal amount is the same every day, which means that the balance continues to decrease at a constant rate every day. This assumption can be modeled using the following equation.

[0120]

[0121] Among them, B0, B1, ..., Bn represents the end-of-day balance, and d represents the frequency of the trip. T Represents the number of days in that specific time period, N represents D T The number of mid-strokes, W T Representative D T The total amount of withdrawals, R T Representative D T The total amount of refills, R represents the amount of each refill, and P represents the remaining cash ratio (assuming it is 25%).

[0122] The total cost depends on the supplier cost and interest cost, and the specific relationship is as follows:

[0123] C T =N*(C V +C I ) (3)

[0124] Among them C V represents the supplier's cost, C I Represents interest cost.

[0125] The cost of one trip for the supplier C V It can be expressed using the following formula:

[0126]

[0127] Among them, C b This is the base fee for each trip, taking USD 751.80 as an example. is the cost rate for each additional 1,000 when R exceeds 700,000. For example, when b1 equals 700,000 and b2 equals 1,000, The value should be 0.61. s represents the distance of the trip, is the cost rate of s, which may be 10.86, for example.

[0128] The cost ratio between trips can be expressed as follows:

[0129]

[0130] The following conditions must also be met:

[0131] B0=R

[0132] B d =p*R

[0133] B j+1 =B j -W

[0134]

[0135] Here Bj represents the end-of-day balance between trips on the jth day, W is the daily withdrawal amount, i d represents the daily interest rate, and i y is the annual interest rate.

[0136] The objective cost function with respect to refill and annual interest rate is:

[0137]

[0138] Then, for a given interest rate, the optimal refill can be expressed as:

[0139]

[0140] Fig.19 The optimal refill amount is shown as a function of total cost and interest rate. The goal is to find a value of R that can achieve the desired refill amount for a given i without significantly affecting cash availability. y Optimize cash costs to reduce total costs. Given an interest rate, the optimal refill amount for a single machine can be determined. Fig.19 The relationship between the optimal refill amount and the interest rate is also shown, with the higher the interest rate, the lower the corresponding optimal refill amount for that machine.

[0141] The above mathematical model can be verified in an example experiment: Over a period of about 3 months, such as from March to mid-June, there are 475 remote ATMs, and the annual interest rate at this time is 11.25%.

[0142] Table 3 shows the experimental results:

[0143] Group ID Number of machines start date End Date Refill times Cash availability Cash cost Trip cost Total Cost original 475 2023-03-01 2023-06-18 5821 99.29％ 19.499.975.40 10.357.363.30 29.857.338.7 Optimization_V1 475 2023-03-01 2023-06-18 16144 96.80％ 7，740，983.78 15，816，649.68 23,557,633.5

[0144] Table 3

[0145] Here the original row shows what would happen based on current settings, while the optimized V1 row saves $370,000, but cash availability drops from 99.29% to 96.80%, a decrease of about 3%.

[0146] However, in some cases, the optimal R is too small to meet daily demand. The optimal refill amount must take into account the number of days that the refill amount is sufficient to support demand while increasing cash availability. To address this problem, a restriction can be imposed on R to create an improved mathematical model:

[0147] R refined =max(R,n*W DailyUpperLimit ) (9)

[0148] Here N represents the number of days, W DailyUpperLimit is the upper limit of daily withdrawal amount, and R refinedIt is an improved refill amount that can support n days of demand.

[0149] This improved model is tested in an example experiment involving a sample of 451 remote ATMs for a period of one month with an annual interest rate of 11.25%.

[0150] Table 4 shows the experimental results as follows:

[0151] Group ID Number of machines start date End Date Refill times Cash availability Cash cost Trip cost Total Cost original 451 2022-03-01 2023-03-31 17268 99.62％ 64,332,249 29,941,508 94,273,757 Optimization_2w 451 2022-03-01 2023-03-31 28591 99.16％ 41,982,103 32,433,836 74,415,939 Optimization_2.5w 451 2022-03-01 2023-03-31 24567 99.34％ 46,965,300 32,433,836 79,399,136 Optimization_3w 451 2022-03-01 2023-03-31 21641 99.48％ 51,231,999 32,433,836 83,665,835

[0152] Table 4

[0153] As shown in the optimal_2w line, when n is equal to 2, approximately $1.2 million is saved and the cash availability rate is 99.16%. In the optimal_2.5w line, when n is equal to 2.5, approximately $880,000 is saved and the cash availability rate is 99.34%. In the optimal_3w line, when n is equal to 3, approximately $630,000 is saved and the cash availability rate is 99.48%. As n increases, a large amount of money is still saved at a higher cash availability rate, which is an improvement on the original mathematical model.

[0154] Table 5 shows the experimental results of the improved model extended to all remote ATM experimental data within one year:

[0155]

[0156] Table 5

[0157] Among all remote_2_wd, when n is equal to 2, the savings are about $6 million, and the cash availability rate is 98.9%. Among all remote_2_5_wd, when n is equal to 2.5, the savings are about $5.2 million, and the cash availability rate is 99.12%. Among all remote_3_wd, when n is equal to 3, the savings are about $4.5 million, and the cash availability rate is 99.21%.

[0158] Fig. 20 It is a diagram showing the cash availability rate and total cost of each experimental object when the improved experimental object is expanded to all remote ATM experimental data within one year. Fig. 20 The more days of support, the higher the cash availability, allowing users to choose the refill amount within a certain range according to their needs.

[0159] Fig.21 is an example user interface example of a having a simulation feature that enables a user to optimize a cash refill strategy.

[0160] In some embodiments, the system of the present invention includes a self-service feature that can use machine learning-based cash withdrawal prediction to simulate the impact of adjusting key thresholds on performance indicators. Users can experiment with different settings and analyze the results before making changes to the production environment.

[0161] For example, if the previous trigger threshold was 20% and the base threshold was 50,000, the final threshold would be 60,000. The user adjusts the refill amount trigger threshold to 10%, which will make the final threshold 55,000, meaning that a cash instruction will only be generated if the ATM balance is less than 55,000.

[0162] This simulation function can help users analyze whether the cash availability of the machine can be reduced to save travel costs. Users can start a simulation run for a period of time by selecting a machine and adjusting the threshold to a specific value. Fig.21 As shown, there is a slider for the user to adjust the threshold. Once the simulation is completed, the system proposed by the present invention can compare the simulation results with the actual performance statistics for cash availability, trip cost and cash cost during the same period. This comparison will help the system decide whether to apply the simulation threshold in the production environment for the next iteration. Providing this simulation feature to the user can enable the user to interact with the system and self-service the generation of cash instructions in the production environment.

[0163] The present invention proposes an AI / ML-based system that can track and analyze ATM data captured in real time, for example, based on physical sensor inputs and data sets obtained from users' physical interactions with ATMs. These inputs are processed by a machine learning algorithm that can adapt to processing factors such as seasonality, holidays, public events, and recent withdrawal trends to accurately predict the amount of cash required for each ATM. The use of machine learning can achieve a more dynamic and responsive cash allocation strategy. The real-time data feed from the ATM can include physical sensor data, which is processed and updated by a trained machine learning model to provide a prediction solution for the current withdrawal pattern.

[0164] The present invention proposes a specific architecture, which uses two parallel running models as an integrated model scheme based on the inventor's testing of physical machines in actual and application environments. The integrated model scheme is provided as a physical system that runs two models simultaneously, and both models are trained as candidate models. The two models are used and trained separately in the deduction stage to optimize the loss function (e.g., MAE performance). And, in the deduction stage, the model with better MAE performance is used to control the generation of ATM refill control signals. In a practical scenario, the better MAE performance is based on the performance of the model in actual operation, because until the deduction process, it is not clear which of the two models will perform better. This scheme does not rely on selecting one of the two models in advance, but intentionally maintains the two models until the runtime. The two models are used together, including a first model, which is a fully connected neural network data architecture, and a second model, which is a decision tree-based learning algorithm provided as a gradient boosting framework (e.g., a lightweight gradient boosting machine, also known as LightGBM). From a practical application perspective, the integrated model can be configured with a prediction buffer for specific parameter adjustments.

[0165] Although the particular dual architecture proposed is not intuitive, it is a special combination of two models that produces results and benefits that are greater than the sum of their parts. Such ensemble models can reduce the generalization error of the results by combining the predictions of different models. Using ensemble models not only improves accuracy, but also resists uncertainty and bias in the data.

[0166] During operation, the prediction system can be configured to run periodically (e.g., every night) to predict cash deposits and generate settlement instructions based on the prediction output data, and small batches of data will be uploaded at a higher frequency (e.g., every 15 minutes). In addition, the model predictions and retraining can be used to generate cash settlement instructions, which can be configured to control one or more logistics operations of cash in transit. In a variation of the method, the artificial intelligence / machine learning based system is configured to generate machine outputs to directly control and provide cash refill of the cash box by generating and submitting logistics requests for cash refill, as an alternative to or in addition to the graphical user interface.

[0167] The predictive system can be optimized for different uses and operations, such as to improve the utilization of the cassettes, reduce the total number of settlement trips, and / or reduce outages (thereby improving service availability). The system can be configured to operate in sync with real-time generated data, thereby acting as an automatic monitoring system capable of operating autonomously or semi-autonomously to predictively control refill operations.

[0168] In some embodiments, entropy can be introduced into the generation process of the generated refill control commands to modify the path and running time by injecting noise, thereby reducing the possibility of physical attack on cash in transit by increasing unpredictability. However, this noise injection will also reduce the determination of the optimal refill timing. As a physical output, a graphical user interface, such as a dashboard, can intuitively visualize the real-time withdrawal pattern, allowing users to make informed decisions and respond quickly to cash needs. In the present invention, this solution has been demonstrated to reduce cash refill delivery time from 36 hours to 15 minutes.

[0169] In some embodiments, there may be multiple ATM groups, wherein multiple ATMs in a particular location may be selectively refilled. For example, an ATM group (e.g., five ATMs) provides services to customers coming out of a stadium entrance and exit at the same location. If one of the ATMs is out of banknotes, the customer can just use another ATM that has banknotes. In this scenario, each ATM in the ATM group is considered a member of the group. That is, the entire ATM group is considered a single ATM. In this case, deactivation is tracked only when the total amount of banknotes or other banknote output of all ATMs in the ATM group drops below a certain threshold.

[0170] One of the purposes of the present invention is to introduce the design, development and implementation of a machine learning model for CDM cash deposit prediction to reduce manual workload and optimize KPIs in existing CDM operations. Specifically, CDM cash settlement instructions are automatically or semi-automatically generated to replace the original rule-based manual logic. That is, settlement operations are performed only when the ATM at the same site is refilled or the CDM has not been settled for a certain number of days. The improvement of the manual process is expressed as an optimization problem, and its technical goal is to decide when to implement the clearing instruction based on the time series prediction of the cash deposit amount in the next few days.

[0171] As described in the present invention, in one variation, the scheme semi-automatically or automatically controls the refill process based on a combination of real-time ATM physical sensor feedback, coupled machine learning models that generate cash box refill instructions, and tracking sensor inputs that record physical refills and subsequent user withdrawal activities. Therefore, the system can operate in a feedback loop indicated by physical sensor readings and continuously improve the prediction method and automatic refill method. As described above, the system can also be coupled to one or more Oracle systems that provide a pipeline of data sets related to third-party / environmental data, which data sets can be combined with real-time ATM physical sensor feedback information to run the machine learning model in a deduction mode to generate outputs that are ultimately used to generate cash box refill instructions.

[0172] In some embodiments, a computer device according to some embodiments of the present invention has a corresponding computer-readable medium. The computer device may include: at least one processor, a memory, a computer-readable storage medium, and a computer program stored on the computer-readable storage medium in the memory and can be run on at least one processor. When the processor executes the computer program, the steps and operations in the embodiments of the method and system for refilling a banknote box based on machine learning are implemented. Due to the limited length of the specification, the inventors will not repeat the steps of the relevant methods and operations here. It can be understood by those skilled in the art that the above description is only an example of a computer device and does not constitute a limitation on the computer device. The computer device may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include a network interface, a display screen, and an input device.

[0173] As will be understood by those of ordinary skill in the art, in this application, "ATM" and "CDM" refer to the same type of self-service machines for cash deposits and withdrawals and can be used interchangeably. Various terms are used only to describe specific embodiments and are not intended to limit the embodiments. For example, when ordinal numbers (e.g., "first," "second," "third," etc.) used in the text are used to modify an element, such as a structure, component, operation, etc., they do not themselves indicate the priority or order of the element relative to another element, but are only used to distinguish it from another element of the same name (but with ordinal numbers). The term "coupled" is defined as connected, although not necessarily directly or mechanically; the two items of "coupled" may be integral to each other. "One" is defined as one or more, unless otherwise expressly required by the present disclosure. The term "substantially" means that the specified criteria are met in the vast majority of cases, but not completely consistent - it covers the specified criteria; for example, "substantially 90 degrees" includes 90 degrees, and "substantially parallel" includes parallel - as understood by those of ordinary skill in the art. In any disclosed embodiment, the word "substantially" may be replaced with the designation "within [percentage]," where percentages include 0.1%, 1%, 5%, and 10%; the word "approximately" may be replaced with the designation "within 10%." "And / or" means in conjunction with or. For example, A, B, and / or C includes: A alone, B alone, C alone, a combination of A and B, a combination of A and C, a combination of B and C, or a combination of A, B, and C. In other words, "and / or" operates as an inclusive or. In addition, "A, B, C, or a combination thereof" or "A, B, C, or any combination thereof" includes: A alone, B alone, C alone, a combination of A and B, a combination of A and C, a combination of B and C, or a combination of A, B, and C.

[0174] The terms "comprising" and any of its forms, such as "including" and "comprising", "having" and any of its forms, and "include" and any of its forms are open-ended linking verbs. Thus, an apparatus that "comprising", "having" or "includes" one or more elements has those one or more elements, but is not limited to having only those elements. Similarly, a method that "comprising", "having" or "including" one or more steps has those one or more steps, but is not limited to having only those one or more steps.

[0175] Any embodiment of any apparatus, system, and method may include or consist essentially of - rather than include / comprise / have - any of the steps, elements, and / or features described. Thus, in any claim, the term "consisting of" or "consisting essentially of" may replace any of the above open-ended linking verbs to alter the scope of a given claim from that when the open-ended linking verb is used. Furthermore, it is understood that the term "wherein" may be used interchangeably with "wherein".

[0176] Furthermore, a device or system configured in a certain way is configured in at least that way, but may also be configured in other ways than those specifically described. Some aspects of one example may be applied to other examples even if not described or illustrated, unless expressly prohibited by the disclosure or the nature of the particular example.

[0177] It will be appreciated by those skilled in the art that various illustrative logic blocks, modules, circuits and algorithm steps related to the disclosure herein can be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, the functions of various illustrative components, modules, modules, circuits and steps are generally described above. As for whether these functions are implemented in the form of hardware or software, it depends on the specific application and the design restrictions on the entire system. Those skilled in the art can implement the functions in different ways for each specific application, but this implementation decision should not be interpreted as departing from the scope of the present disclosure. It will also be appreciated by those skilled in the art that the order or combination of components, methods or interactions described herein is merely an example, and the components, methods or interactions of various aspects of the present disclosure can be combined or performed in a manner other than illustrated and described herein.

[0178] Various illustrative logical blocks, modules, and circuits related to the disclosure herein may be implemented or executed by a processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or a combination thereof to implement the functions described herein. The processor may be a microprocessor, but may also be other forms of processors, controllers, microcontrollers, or state machines. The processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, a combination of one or more microprocessors and a DSP core, or any other such configuration.

[0179] The steps of the method or algorithm disclosed herein may be directly embodied in hardware, or in a software module executed by a processor, or in a combination of the two. The software module may be present in a random access memory (RAM), flash memory, read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Alternatively, the storage medium may be an integral part of the processor. The processor and the storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Alternatively, the processor and the storage medium may be located in a user terminal as separate components.

[0180] In one or more exemplary designs, the functions described may be implemented by hardware, software, firmware, or any combination thereof. If implemented by software, these functions may be stored or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include any medium that facilitates the transfer of a computer program from one place to another, including both computer storage media and communication media. Computer-readable storage media may be any medium that can be accessed by a computer. By way of example but not limitation, these computer-readable media may include random RAM, ROM, EEPROM, CD-ROM, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of instructions or data structures and can be accessed by a computer or processor. In addition, a connection may also be appropriately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, or digital subscriber line (DSL), then the coaxial cable, fiber optic cable, twisted pair, or DSL is also included in the definition of the medium. As used herein, "disk" and "disc" include compact disk (CD), laser disk, optical disk, digital versatile disk (DVD), hard disk, solid state disk and Blu-ray disk, where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0181] The above description and examples provide a complete description of the structure and use of illustrative embodiments. Although some of the above examples are described in a particular level of detail, or with reference to one or more separate examples, those skilled in the art may make many modifications to the disclosed embodiments without departing from the scope of the invention. Therefore, the various illustrative embodiments of the method and system are not limited to the specific forms disclosed. Instead, they include all modifications and alternatives that fall within the scope of the claims, and other examples may include some or all of the features of the illustrated examples. For example, some elements may be omitted, or they may be combined into a unit structure, and / or the connecting parts may be replaced. In addition, when appropriate, some aspects of any of the examples described above may be combined with some aspects of any of the other examples described to form further examples with similar or different properties and / or functions, and to solve the same or different problems. Similarly, it should be understood that the benefits and advantages described above may be related to one embodiment or to several embodiments.

[0182] The claims are not intended to include, nor should they be interpreted as including, means-plus-function or step-plus-function limitations, unless such limitations are explicitly recited using phrases such as "means-by" or "step-by" in a particular claim.

[0183] Although various aspects of the present disclosure and their advantages have been described in detail, it should be understood that changes, substitutions and modifications may be made herein without departing from the spirit of the disclosure as defined by the appended claims. In addition, the scope of the present application is not intended to be limited to the specific embodiments of the processes, machines, manufactures, material compositions, means, methods and steps described in the specification. As will be readily understood by those of ordinary skill in the art from this disclosure, processes, machines, manufactures, material compositions, means, methods or steps currently existing or developed in the future that perform substantially the same functions or achieve substantially the same results as the corresponding embodiments described herein may be utilized in accordance with the present disclosure. Therefore, the appended claims are intended to include such processes, machines, manufactures, material compositions, means, methods or steps within their scope.

Claims

1. A machine learning-based cash box refilling method, characterized in that: include: Instantiating a machine learning data model architecture utilizing an integrated model, the integrated model comprising a fully connected neural network framework architecture and a gradient boosting framework architecture, the gradient boosting framework architecture utilizing a decision tree-based algorithm and applying a leaf node growing scheme, the fully connected neural network architecture coupled with the gradient boosting framework architecture; Receiving a training data set and training the ensemble model using a loss function to establish a trained ensemble model, the training data set being derived from one or more historical data sets, the loss function being adapted to reduce a total number of refills of the banknote box; Receiving a deduced data set based on a real-time ATM sensor feedback stream and providing it to the trained integrated model; generating one or more predicted data outputs for each ATM coupled to the real-time ATM sensor feedback stream using the trained ensemble model; for each predicted data output of the one or more predicted data outputs that exceeds a threshold, automatically generating one or more data messages for controlling a refill or removal operation associated with a corresponding coupled ATM; confirming one or more refill or removal operations based on physical sensor readings from the real-time ATM sensor feedback stream; as well as An updated inference dataset is generated for subsequent operation of the trained ensemble model.

2. The method according to claim 1, characterized in that Downstream usage of the refilled ATM is used to update the trained ensemble model.

3. The method according to claim 1, characterized in that The fully connected neural network includes a ReLU layer network combined with a sigmoid layer.

4. The method according to claim 1, characterized in that A reduce-increase-reduce framework is used to determine the number of nodes in each layer.

5. The method according to claim 1, characterized in that The gradient boosting framework architecture is the LightGBM model.

6. The method according to claim 1, characterized in that Batch normalization and dropout methods are used to control the activation function.

7. The method according to claim 1, characterized in that The mean absolute error (MAE) is used as the loss function.

8. The method according to claim 1, characterized in that The one or more data messages for controlling refill or removal operations associated with corresponding coupled ATMs are sequenced based on the generated paths of the regularly scheduled cash-in-transit vehicles.

9. The method according to claim 1, characterized in that The correspondingly coupled ATMs comprise an ATM group, wherein a group of ATMs located at the same location is identified as a single ATM.

10. The method according to claim 1, characterized in that The ensemble model uses batch normalization and dropout methods.

11. A machine learning-based cash box refilling system, characterized in that: include: A data storage module configured to instantiate and store a machine learning data model architecture using an integrated model, wherein the integrated model includes a fully connected neural network framework architecture and a gradient boosting framework architecture, wherein the gradient boosting framework architecture uses a decision tree-based algorithm and applies a leaf node growth scheme, wherein the fully connected neural network architecture is coupled with the gradient boosting framework architecture; A model training module configured to receive a training data set and train the integrated model using a loss function to establish a trained integrated model, wherein the training data set is derived from one or more historical data sets, and the loss function is suitable for reducing the total number of banknote box refills; Receiving a deduction data set based on a real-time ATM sensor feedback stream and providing it to the trained integrated model; a model prediction module configured to generate one or more predicted data outputs for each ATM coupled to the real-time ATM sensor feedback stream using the trained ensemble model; a logistics control module configured to, for each predicted data output of the one or more predicted data outputs that exceeds a threshold value, automatically generate one or more data messages for controlling a refill or removal operation associated with a corresponding coupled ATM; confirming one or more refill or removal operations based on physical sensor readings from the real-time ATM sensor feedback stream; as well as The model training module is further configured to generate an updated deduction data set for subsequent operation of the trained integrated model.

12. The system according to claim 11, characterized in that Downstream usage of the refilled ATM is used to update the trained ensemble model.

13. The system according to claim 11, characterized in that The fully connected neural network includes a ReLU layer network combined with a sigmoid layer.

14. The system according to claim 11, characterized in that A reduce-increase-reduce framework is used to determine the number of nodes in each layer.

15. The system of claim 11, wherein: The gradient boosting framework architecture is the LightGBM model.

16. The system of claim 11, wherein: Batch normalization and dropout methods are used to control the activation function.

17. The system of claim 11, wherein: The mean absolute error (MAE) is used as the loss function.

18. The system of claim 11, wherein: The one or more data messages for controlling refill or removal operations associated with corresponding coupled ATMs are sequenced based on the generated paths of the regularly scheduled cash-in-transit vehicles.

19. The system of claim 11, wherein: The correspondingly coupled ATMs comprise an ATM group, wherein a group of ATMs located at the same location is identified as a single ATM.

20. The system of claim 11, wherein: The ensemble model uses batch normalization and dropout methods.

21. A computer device comprising a memory, a processor and a computer program stored therein and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the machine learning-based cash box refilling method according to any one of claims 1 to 10 are implemented.

22. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the machine learning-based cash box refilling method according to any one of claims 1 to 10 are implemented.

Images