Low-power-consumption Bluetooth automatic vending and real-time accounting system based on edge computing

Abstract

The invention discloses a low-power-consumption Bluetooth automatic vending and real-time accounting system based on edge computing, and the system comprises an operation data input module which is used for generating a multi-channel control path data structure on a unified time axis; the NCDE modeling module is used for generating a continuous time hidden state vector sequence; the service state prediction module is used for generating an extrapolation hidden state vector sequence and a service state prediction result set; the control decision generation module is used for generating a control parameter set; the control execution module is used for generating transaction records and account division result records and writing the transaction records and the account division result records into the edge data structure; and the model updating module is used for constructing a joint loss data set and updating parameters of the improved NCDE model. According to the invention, low-power-consumption communication control, transaction processing optimization and real-time account division collaboration of the automatic vending equipment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent vending equipment, and particularly relates to a low-power Bluetooth automatic vending and real-time account distribution system based on edge computing. BACKGROUND

[0002] With the development of automatic vending equipment towards unmanned, low maintenance cost and high transaction concurrency, terminals generally rely on low-power Bluetooth for near-field communication to realize user identification, product selection and payment triggering. However, the existing automatic vending equipment adopts a fixed communication cycle and a static communication power consumption strategy, which cannot dynamically adjust the Bluetooth broadcast parameters according to the user arrival intensity and transaction load in different time periods, resulting in high power consumption and limited standby time. At the same time, the account distribution process after transaction completion of the automatic vending equipment usually relies on cloud processing, which has the problems of large link delay, high failure rate and insufficient real-time performance, and cannot meet the real-time account distribution needs of multiple participants.

[0003] Although the existing technology attempts to execute part of the transaction processing logic through an edge gateway or a local micro control unit, it still lacks the ability to model and predict the time continuity according to the device running state, communication load and transaction behavior changes, and cannot realize the dynamic closed loop from running data to control decision. Especially in the scene where the user arrival intensity, transaction load, energy consumption and account distribution delay risk present significant time-varying characteristics, the traditional analysis method based on discrete sampling or fixed window cannot capture the continuous time variation law, so it cannot provide accurate feedforward information for power consumption control, transaction processing and account distribution execution.

[0004] In addition, the existing system lacks a mechanism for using running data, historical transactions and account distribution results for continuous optimization, and cannot update the model through real-time feedback, which easily leads to the control strategy lagging behind the device running state, affecting the communication stability, transaction continuity and account distribution timeliness.

[0005] Therefore, how to provide a low-power Bluetooth automatic vending and real-time account distribution system based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a low-power Bluetooth automatic vending and real-time account distribution system based on edge computing. The present application introduces an improved NCDE model, uses multi-channel running data to build a continuous time business state representation, realizes dynamic prediction of user arrival intensity, transaction load, energy consumption and account distribution delay risk, and generates control parameters for Bluetooth communication, transaction processing and account distribution execution. The present application realizes the fusion mechanism of continuous time modeling, business prediction and control decision, so that the automatic vending equipment can complete the communication power consumption adjustment, transaction processing optimization and account distribution acceleration on the edge side, and has the advantages of lower power consumption, faster response and stronger real-time account distribution.

[0007] According to an embodiment of the present application, a low-power Bluetooth vending and real-time account distribution system based on edge computing comprises:

[0008] A running data input module is configured to collect a running data set in an edge computing node, perform time alignment on a unified time axis, and generate a multi-channel control path data structure.

[0009] An NCDE modeling module is configured to load an improved NCDE model, take the multi-channel control path data structure as a control input path, and generate a continuous-time hidden state vector sequence.

[0010] A business state prediction module is configured to perform time evolution reasoning and continuous-time extrapolation calculation based on the continuous-time hidden state vector sequence, generate an extrapolated hidden state vector sequence, and generate a business state prediction result set.

[0011] A control decision generation module is configured to splice the continuous-time hidden state vector sequence and the business state prediction result set to generate a decision feature vector sequence, and generate a control parameter set according to the decision feature vector sequence.

[0012] A control execution module is configured to perform Bluetooth communication control, transaction processing control, and account distribution execution control according to the control parameter set, generate transaction records and account distribution result records, and write the transaction records and account distribution result records into an edge data structure.

[0013] A model updating module is configured to construct a joint loss data set, update parameters of the improved NCDE model, and call the updated improved NCDE model in a subsequent running cycle.

[0014] Optionally, the modules are implemented through the following methods:

[0015] A running data set is collected in an edge computing node, time alignment processing is performed, and a multi-channel control path data structure is generated.

[0016] An improved NCDE model is constructed in an edge computing node, the multi-channel control path data structure is taken as a model control input path, and a continuous-time hidden state vector sequence is generated.

[0017] Time evolution reasoning and state extrapolation calculation are performed on the continuous-time hidden state vector sequence, and a business state prediction result set is generated.

[0018] Based on the continuous-time hidden state vector sequence and the business state prediction result set, a prediction quantity in the business state prediction result set is converted into a control parameter set through control decision generation calculation.

[0019] According to the control parameter set, the Bluetooth communication control, the transaction processing control and the account distribution execution control are performed in the edge computing node, and the transaction record and the account distribution result record are generated;

[0020] Based on the business state prediction result set, the transaction record and the account distribution result record, a joint loss data set is constructed, the parameters of the improved NCDE model are updated according to the joint loss data set, an updated improved NCDE model is obtained, and the updated improved NCDE model is called in the subsequent running cycle.

[0021] Optionally, the generation of the multi-channel control path data structure comprises:

[0022] Data collection is performed in the edge computing node, original running data is obtained from the edge computing node in the running process, and is written into a running data cache structure to generate an initial running data set;

[0023] The original running data in the initial running data set is parsed, and the parsed Bluetooth communication record, energy consumption record and account distribution business record are arranged in timestamp order respectively to generate Bluetooth communication data, energy consumption data and account distribution business data;

[0024] A unified time axis is constructed, fixed time intervals are adopted, the timestamps in the Bluetooth communication data, the energy consumption data and the account distribution business data are mapped, and the missing value filling method is used for the positions where there is no observation value, and the time-aligned Bluetooth communication data, the time-aligned energy consumption data and the time-aligned account distribution business data are generated;

[0025] At each time position of the unified time axis, the time-aligned Bluetooth communication data, the time-aligned energy consumption data and the time-aligned account distribution business data are combined to generate a multi-channel control path data structure.

[0026] Optionally, the generation of the continuous time hidden state vector comprises:

[0027] In the edge computing node, an input encoding structure is configured for the improved NCDE model, the Bluetooth communication data, the energy consumption data and the account distribution business data corresponding to each time position of the multi-channel control path data structure on the unified time axis are spliced to generate a control input vector sequence;

[0028] The dimension parameter of the continuous time hidden state vector is set, a hidden state initialization module is constructed, the continuous time hidden state vector is initialized, and the continuous time hidden state vector corresponding to the starting time position is obtained;

[0029] The control path encoding sub-network receives a control input vector sequence, generates a control feature representation at each time position of a unified time axis, and the vector field representation network receives the control feature representation at each time position of the unified time axis and a continuous time hidden state vector corresponding to the time position, generates a hidden state change amount representation corresponding to the time position.

[0030] A time step parameter is set, the continuous time hidden state vector is updated at each time position on the unified time axis according to the time step parameter, the continuous time hidden state vector corresponding to the current time position is calculated, and a continuous time hidden state vector sequence covering all time positions of the unified time axis is obtained.

[0031] Optionally, the generation of the service state prediction result set comprises:

[0032] A prediction time window parameter is set based on the continuous time hidden state vector sequence, a target prediction time position set corresponding to the unified time axis is determined, a continuous time hidden state vector adjacent to the target prediction time position set is selected from the continuous time hidden state vector sequence, and a prediction input hidden state sequence is generated;

[0033] The time evolution structure of the improved NCDE model is called, the prediction input hidden state sequence is executed for time evolution reasoning according to the prediction time window parameter, the prediction input hidden state sequence is continuously time extrapolated on the unified time axis, and an extrapolated hidden state vector sequence corresponding to the target prediction time position set is generated;

[0034] Based on the extrapolated hidden state vector sequence, a user arrival intensity mapping structure, a transaction load mapping structure, an energy consumption mapping structure and a split time delay risk mapping structure are respectively constructed, the extrapolated hidden state vector sequence is input into the structures, and a user arrival intensity prediction result, a transaction load prediction result, an energy consumption prediction result and a split time delay risk prediction result are respectively generated;

[0035] A service state prediction result set containing the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the split time delay risk prediction result is obtained.

[0036] Optionally, the generation of the control parameter set comprises:

[0037] At each target time position of the unified time axis, the continuous time feature amount of the corresponding time position in the continuous time hidden state vector sequence is spliced with the service state prediction result set to generate a decision feature vector corresponding to the target time position, and the decision feature vectors are sequentially arranged to form a decision feature vector sequence.

[0038] The weight parameter set in the control decision generation calculation is constructed based on the decision feature vector sequence, the weight value of each weight parameter at the target time position is determined according to the component value in the decision feature vector, and the weight parameter set covering all target time positions is obtained;

[0039] At each target time position of the unified time axis, the user arrival intensity prediction result and the energy consumption prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for power consumption control in the weight parameter set, and a Bluetooth communication control parameter sequence is generated;

[0040] The user arrival intensity prediction result, the transaction load prediction result and the account distribution delay risk prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for transaction processing in the weight parameter set, and a transaction processing control parameter sequence is generated;

[0041] The transaction load prediction result and the account distribution delay risk prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for account distribution execution in the weight parameter set, and an account distribution execution control parameter sequence is generated;

[0042] The continuous time feature quantity related to edge computing resource usage in the continuous time hidden state vector sequence and the service state prediction result set are weighted and calculated according to the weight parameter for edge computing power allocation in the weight parameter set, and an edge computing power allocation control parameter sequence is generated;

[0043] The Bluetooth communication control parameter sequence, the transaction processing control parameter sequence, the account distribution execution control parameter sequence and the edge computing power allocation control parameter sequence are combined to generate a control parameter set.

[0044] Optionally, the generation of the transaction record and the account distribution result record comprises:

[0045] The Bluetooth communication control parameter, the transaction processing control parameter, the account distribution execution control parameter and the edge computing power allocation control parameter are extracted from the control parameter set to generate a control execution parameter sequence;

[0046] At each target time position of the unified time axis, the Bluetooth communication control is executed according to the Bluetooth communication control parameter, the data generated in the Bluetooth communication control process is recorded, and the data is written into the transaction record;

[0047] The transaction processing control is executed according to the transaction processing control parameter, the transaction amount data, the payment result data and the related timestamp data are written into the transaction record;

[0048] The account distribution execution control is executed according to the account distribution execution control parameter, the account distribution amount data, the account distribution proportion data and the account distribution completion time data are written into the account distribution result record;

[0049] Write the transaction records and the account distribution result records corresponding to each target time position of the unified time axis in time sequence into the edge data structure.

[0050] Optionally, the model parameter updating comprises:

[0051] At each target time position of the unified time axis, based on the business state prediction result set, the actual transaction number and the actual transaction amount corresponding to the target time position are counted from the transaction records, and the account distribution time delay between the target time position corresponding to the account distribution result record and the transaction generation time is calculated, to generate an alignment sample set;

[0052] In each alignment sample record of the alignment sample set, a user arrival intensity deviation measurement field is constructed based on the user arrival intensity prediction result and the actual user arrival intensity data, a transaction load deviation measurement field is constructed based on the transaction load prediction result and the actual transaction load data, an energy consumption deviation measurement field is constructed based on the energy consumption prediction result, and an account distribution time delay risk deviation measurement field is constructed based on the account distribution time delay risk prediction result and the actual account distribution time delay data, and is written into the corresponding alignment sample record, to generate a joint loss data set;

[0053] According to the joint loss data set, the parameters of the improved NCDE model are updated, the joint loss data set is taken as parameter update driving data, parameter optimization calculation based on the joint loss data set is performed, and an updated improved NCDE model is obtained;

[0054] In the subsequent running period, the multi-channel control path data structure generated by the new round of running data set is input into the updated improved NCDE model, so that the updated improved NCDE model can continuously output a business state prediction result set with higher consistency with the transaction records and the account distribution result records in the subsequent running period.

[0055] The beneficial effects of the present application are:

[0056] Firstly, by introducing the improved NCDE model, the Bluetooth communication data, the energy consumption data and the account distribution business data are continuously time modeled, the present application can capture the dynamic running law of the automatic vending equipment in different time periods, so that the equipment can adjust the communication strategy based on the real-time prediction result, thereby effectively reducing the Bluetooth communication power consumption and prolonging the equipment endurance time.

[0057] Secondly, by completing the prediction of the user arrival intensity, the transaction load, the energy consumption and the account distribution time delay risk in the edge computing node, the present application realizes the feedforward judgment of the business state, so that the transaction processing and the account distribution execution have higher timeliness and stability, the link time delay and the failure risk caused by the cloud dependence are reduced, and the transaction continuity and the real-time of multi-party account distribution are improved.

[0058] In addition, by constructing a joint loss data set based on transaction records and account distribution result records and continuously updating and improving the NCDE model, the application constructs an adaptive optimization mechanism on the edge side, so that the model can continuously improve the prediction accuracy and control decision quality as the device operating state changes, thereby forming a closed-loop optimization system for low-power control, transaction processing and account distribution execution, and significantly improving the intelligentization and autonomous operation capability of the vending device. BRIEF DESCRIPTION OF DRAWINGS

[0059] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0060] Figure 1 A whole flowchart of a low-power Bluetooth vending and real-time account distribution system based on edge computing is proposed for the application;

[0061] Figure 2 A structure diagram for the improved NCDE model in the application to perform multi-channel control path modeling on the edge side and generate a continuous-time hidden state vector sequence is shown;

[0062] Figure 3 A control decision generation structure diagram for the application to generate a business state prediction result set based on the continuous-time hidden state vector sequence and construct a control parameter set is shown. DETAILED DESCRIPTION

[0063] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0064] REFERENCE Figures 1-3 A low-power Bluetooth vending and real-time account distribution system based on edge computing includes:

[0065] The running data input module is used to obtain Bluetooth communication data, energy consumption data and account distribution business data from the Bluetooth communication processing unit, the energy consumption monitoring unit and the account distribution business processing unit, perform timestamp extraction and field extraction, perform time alignment processing based on a unified time axis, generate time-aligned Bluetooth communication data, time-aligned energy consumption data and time-aligned account distribution business data, and combine the three types of aligned data on the unified time axis into a multi-channel control path data structure to provide a continuous-time control input path for the subsequent improved NCDE model;

[0066] an NCDE modeling module, configured to load the improved NCDE model and initialize an input encoding structure, a hidden state structure, a control path encoding subnetwork, and a vector field representation network, receive a multi-channel control path data structure as a control input path, generate a continuous-time hidden state vector sequence covering a unified time axis according to a continuous-time hidden state evolution mechanism, and use the continuous-time hidden state vector sequence to represent a service state of an edge computing node;

[0067] a service state prediction module, configured to set a prediction time window parameter based on the continuous-time hidden state vector sequence and generate a predicted input hidden state sequence, call a time evolution structure of the improved NCDE model to perform time evolution reasoning and continuous-time extrapolation calculation, generate an extrapolated hidden state vector sequence, and convert the extrapolated hidden state vector sequence into a user arrival intensity prediction result, a transaction load prediction result, an energy consumption prediction result, and a split-time delay risk prediction result through a user arrival intensity mapping structure, a transaction load mapping structure, an energy consumption mapping structure, and a split-time delay risk mapping structure, respectively, and generate a service state prediction result set;

[0068] a control decision generation module, configured to splice a continuous-time feature quantity of the continuous-time hidden state vector sequence and four types of prediction results in the service state prediction result set according to a unified time axis to generate a decision feature vector sequence, determine a weight parameter set for power consumption control, transaction processing control, split-time execution control, and edge computing power allocation control according to the decision feature vector sequence, and generate a Bluetooth communication control parameter, a transaction processing control parameter, a split-time execution control parameter, and an edge computing power allocation control parameter based on the weight parameter set, and combine the parameters to form a control parameter set for driving a subsequent control execution process;

[0069] a control execution module, configured to perform Bluetooth communication control, transaction processing control, and split-time execution control according to the control parameter set, write timestamp data, connection state data, and Bluetooth communication result data generated by the Bluetooth communication control into a transaction record, write transaction amount data, payment result data, and timestamp data generated by the transaction processing control into the transaction record, write split-time amount data, split-time proportion data, and split-time completion time data generated by the split-time execution control into a split-time result record, and write the transaction record and the split-time result record into an edge data structure in chronological order;

[0070] a model updating module, configured to construct an aligned sample set and a joint loss data set based on the service state prediction result set, the transaction record, and the split-time result record, update parameters of the control path encoding subnetwork and the vector field representation network in the improved NCDE model according to the joint loss data set, obtain an updated improved NCDE model, and call the updated improved NCDE model to regenerate the continuous-time hidden state vector sequence and the service state prediction result set in a subsequent running period, so as to realize continuous updating of the model.

[0071] In this embodiment, the modules are implemented by the following method:

[0072] A set of running data is collected in the edge computing node, including Bluetooth communication data, energy consumption data, and accounting service data. Time alignment processing is performed on the set of running data to generate a multi-channel control path data structure, which is used to represent the joint change of the set of running data on the time axis;

[0073] An improved NCDE model is constructed in the edge computing node, and the multi-channel control path data structure is used as the model control input path to generate a continuous-time hidden state vector sequence, which is used to represent the service state of the edge computing node;

[0074] Based on the continuous-time hidden state vector sequence, time evolution reasoning and state extrapolation calculation are performed on the continuous-time hidden state vector sequence to generate a set of service state prediction results, including user arrival intensity prediction results, transaction load prediction results, energy consumption prediction results, and accounting delay risk prediction results;

[0075] Based on the continuous-time hidden state vector sequence and the set of service state prediction results, control decision generation calculation is performed to convert the prediction quantities in the set of service state prediction results into Bluetooth communication control parameters, transaction processing control parameters, accounting execution control parameters, and edge computing power allocation control parameters, and generate a set of control parameters;

[0076] Bluetooth communication control, transaction processing control, and accounting execution control are performed in the edge computing node according to the set of control parameters, transaction records and accounting result records are generated, and the transaction records and accounting result records are written into the edge data structure;

[0077] A joint loss data set is constructed based on the set of service state prediction results, transaction records, and accounting result records. The parameters of the improved NCDE model are updated according to the joint loss data set to obtain an updated improved NCDE model, and the updated improved NCDE model is called in the subsequent running period to regenerate the continuous-time hidden state vector and the set of service state prediction results.

[0078] In this embodiment, the generation of the multi-channel control path data structure includes:

[0079] Data collection is performed in the edge computing node. In the running process, the original running data is obtained from the Bluetooth communication processing unit, the energy consumption monitoring unit, and the accounting service processing unit of the edge computing node, and the original running data is written into the running data cache structure to generate an initial running data set;

[0080] The original running data from the Bluetooth communication processing unit in the initial running data set is parsed to extract the timestamp, broadcast state value, connection state value, received signal strength value and link holding time length value of each Bluetooth communication record, and the parsed Bluetooth communication record is arranged in timestamp order to generate Bluetooth communication data;

[0081] The original running data from the energy consumption monitoring unit in the initial running data set is parsed to extract the timestamp, battery remaining capacity value of the modular lattice unit, power consumption value and processor load value of each energy consumption record, and the parsed energy consumption record is arranged in timestamp order to form energy consumption data;

[0082] The original running data from the sub-account service processing unit in the initial running data set is parsed to extract the timestamp, transaction amount value, number of sub-account participating roles, sub-account ratio parameter and sub-account state value of each sub-account service record, and the parsed sub-account service record is arranged in timestamp order to form sub-account service data;

[0083] A unified time axis is constructed in the edge computing node, a fixed time interval is adopted, the timestamps in the Bluetooth communication data, energy consumption data and sub-account service data are mapped, the Bluetooth communication data, energy consumption data and sub-account service data are associated to the corresponding positions on the unified time axis, and a missing filling method is used for positions on the unified time axis that do not exist observation values, to generate time-aligned Bluetooth communication data, time-aligned energy consumption data and time-aligned sub-account service data;

[0084] At each time position of the unified time axis, the time-aligned Bluetooth communication data, time-aligned energy consumption data and time-aligned sub-account service data are combined into a multi-channel time series record, the multi-channel time series record is arranged in order of the unified time axis to generate a multi-channel control path data structure, and the multi-channel control path data structure is used as a control input path of the improved NCDE model in the edge computing node.

[0085] In this embodiment, the generation of the continuous-time hidden state vector includes:

[0086] In the edge computing node, an input encoding structure is configured for the improved NCDE model, the Bluetooth communication data, energy consumption data and sub-account service data corresponding to each time position of the unified time axis in the multi-channel control path data structure are spliced to generate a control input vector sequence corresponding one-to-one to each time position of the unified time axis, and the control input vector sequence is taken as a control input path of the improved NCDE model;

[0087] The dimension parameter of the continuous time hidden state vector is set for the improved NCDE model, a hidden state initialization module is constructed, the continuous time hidden state vector is initialized according to the preset hidden state initial value vector at the starting time position of the unified time axis, and the continuous time hidden state vector corresponding to the starting time position is obtained;

[0088] The control path encoding subnetwork and the vector field representation network are configured for the improved NCDE model, the control path encoding subnetwork receives the control input vector sequence, generates the control feature representation at each time position of the unified time axis, the vector field representation network receives the control feature representation and the continuous time hidden state vector corresponding to each time position of the unified time axis at each time position of the unified time axis, and generates the hidden state change amount representation corresponding to the time position;

[0089] The time step parameter is set, the continuous time hidden state vector is updated at each time position on the unified time axis according to the time step parameter, the continuous time hidden state vector corresponding to the starting time position of the unified time axis is taken as the update starting point, the continuous time hidden state vector corresponding to the current time position is calculated according to the continuous time hidden state vector corresponding to the previous time position and the hidden state change amount representation corresponding to the current time position at each time position of the unified time axis, and a continuous time hidden state vector sequence covering all time positions of the unified time axis is obtained, and the continuous time hidden state vector sequence is used to represent the business state of the edge computing node at each time position on the unified time axis.

[0090] In the embodiment, the generation of the business state prediction result set comprises:

[0091] The prediction time window parameter is set based on the continuous time hidden state vector sequence, the target prediction time position set corresponding to the unified time axis is determined, the continuous time hidden state vector adjacent to the target prediction time position set is selected from the continuous time hidden state vector sequence, and the prediction input hidden state sequence is generated;

[0092] The time evolution structure of the improved NCDE model is called, the time evolution reasoning is performed on the prediction input hidden state sequence according to the prediction time window parameter, the continuous time extrapolation calculation is performed on the prediction input hidden state sequence on the unified time axis, and the extrapolation hidden state vector sequence corresponding to the target prediction time position set is generated;

[0093] Specifically, the time evolution reasoning refers to a vector field representation network based on an improved NCDE model, calculating the continuous time derivative representation of each time position corresponding to the prediction input hidden state sequence on the unified time axis, and calculating the hidden state change amount according to the control feature representation of each time position in the control input path, so that the prediction input hidden state sequence is solved on the unified time axis according to the continuous time dynamic law, thereby obtaining the hidden state evolution trajectory covering the prediction time window range;

[0094] The continuous time extrapolation calculation refers to solving the prediction input hidden state sequence on the target prediction time position set based on the hidden state evolution trajectory formed by the time evolution reasoning, to obtain the corresponding extrapolated hidden state vector sequence. The continuous time extrapolation calculation extends the hidden state from the evolution trajectory to the target prediction time position set according to the continuous derivative representation provided by the vector field representation network, realizes the continuous time solving of the hidden state at the future time position, and enables the extrapolated hidden state vector sequence to represent the future business behavior trend;

[0095] Based on the extrapolated hidden state vector sequence, a user arrival intensity mapping structure is constructed. The extrapolated hidden state vector sequence is input into the user arrival intensity mapping structure, and through the execution of the fixed dimension mapping calculation corresponding to the extrapolated hidden state vector sequence, the continuous time feature quantity in the extrapolated hidden state vector sequence is mapped into the user arrival intensity value, to generate the user arrival intensity prediction result. The user arrival intensity prediction result is written into the business state prediction result set.

[0096] Based on the extrapolated hidden state vector sequence, a transaction load mapping structure is constructed. The extrapolated hidden state vector sequence is input into the transaction load mapping structure, and through the execution of the fixed dimension mapping calculation corresponding to the extrapolated hidden state vector sequence, the continuous time feature quantity in the extrapolated hidden state vector sequence is mapped into the transaction load value, to generate the transaction load prediction result. The transaction load prediction result is written into the business state prediction result set.

[0097] Based on the extrapolated hidden state vector sequence, an energy consumption mapping structure is constructed. The extrapolated hidden state vector sequence is input into the energy consumption mapping structure, and through the execution of the fixed dimension mapping calculation corresponding to the extrapolated hidden state vector sequence, the continuous time feature quantity in the extrapolated hidden state vector sequence is mapped into the energy consumption value, to generate the energy consumption prediction result. The energy consumption prediction result is written into the business state prediction result set.

[0098] The extrapolated hidden state vector sequence is input into the account distribution delay risk mapping structure, and by performing fixed dimension mapping calculation corresponding to the extrapolated hidden state vector sequence, the continuous time characteristic quantity in the extrapolated hidden state vector sequence is mapped into an account distribution delay risk value, an account distribution delay risk prediction result is generated, and the account distribution delay risk prediction result is written into the service state prediction result set;

[0099] A service state prediction result set containing the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the account distribution delay risk prediction result is obtained.

[0100] In the embodiment, the generation of the control parameter set comprises:

[0101] At each target time position on the unified time axis, the continuous time characteristic quantity at the corresponding time position in the continuous time hidden state vector sequence is spliced with the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the account distribution delay risk prediction result at the corresponding time position in the service state prediction result set, a decision feature vector corresponding to the target time position is generated, and all decision feature vectors are arranged in sequence according to the unified time axis to form a decision feature vector sequence;

[0102] The weight parameter set in the control decision generation calculation is constructed based on the decision feature vector sequence, the weight parameter set includes weight parameters for power consumption control, weight parameters for transaction processing, weight parameters for account execution and weight parameters for edge computing power allocation, and at each target time position on the unified time axis, the weight values of each weight parameter at the target time position are determined according to the component values corresponding to the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the account distribution delay risk prediction result in the decision feature vector, and the weight parameter set covering all target time positions is obtained.

[0103] At each target time position on the unified time axis, the user arrival intensity prediction result and the energy consumption prediction result in the service state prediction result set are calculated by the weight parameter for power consumption control in the weight parameter set, a Bluetooth communication control parameter corresponding to the target time position is generated, and the Bluetooth communication control parameters corresponding to each target time position are arranged in time sequence to generate a Bluetooth communication control parameter sequence.

[0104] At each target time position on the unified time axis, the user arrival intensity prediction result, the transaction load prediction result and the account distribution delay risk prediction result in the business state prediction result set are weighted and calculated according to the weight parameters in the weight parameter set for transaction processing, to generate a transaction processing control parameter corresponding to the target time position, and the transaction processing control parameters corresponding to each target time position are arranged in time sequence to generate a transaction processing control parameter sequence;

[0105] At each target time position on the unified time axis, the transaction load prediction result and the account distribution delay risk prediction result in the business state prediction result set are weighted and calculated according to the weight parameters in the weight parameter set for account distribution execution, to generate an account distribution execution control parameter corresponding to the target time position, and the account distribution execution control parameters corresponding to each target time position are arranged in time sequence to generate an account distribution execution control parameter sequence;

[0106] At each target time position on the unified time axis, the continuous time feature quantity related to the edge computing resource usage in the continuous time hidden state vector sequence and the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the account distribution delay risk prediction result in the business state prediction result set are weighted and calculated according to the weight parameters in the weight parameter set for edge computing power allocation, to generate an edge computing power allocation control parameter corresponding to the target time position, and the edge computing power allocation control parameters corresponding to each target time position are arranged in time sequence to generate an edge computing power allocation control parameter sequence;

[0107] The Bluetooth communication control parameter sequence, the transaction processing control parameter sequence, the account distribution execution control parameter sequence and the edge computing power allocation control parameter sequence are combined, and the corresponding Bluetooth communication control parameter, transaction processing control parameter, account distribution execution control parameter and edge computing power allocation control parameter are extracted at each target time position on the unified time axis to generate a control parameter set covering the target time position set on the unified time axis, and the control parameter set is used to provide control input when the Bluetooth communication control, transaction processing control and account distribution execution control are executed subsequently.

[0108] In the embodiment, the generation of the transaction record and the account distribution result record comprises:

[0109] The control parameter set is analyzed, the Bluetooth communication control parameter, the transaction processing control parameter, the account distribution execution control parameter and the edge computing power allocation control parameter are extracted from the control parameter set, the four types of control parameters are corresponded to the target time position on the unified time axis, and a control execution parameter sequence is generated;

[0110] At each target time position on the unified time axis, the Bluetooth communication control is performed according to the Bluetooth communication control parameter in the control execution parameter sequence, the timestamp data, the connection state data and the Bluetooth communication result data generated in the Bluetooth communication control process are recorded, and the Bluetooth communication related fields in the transaction record are written;

[0111] At each target time position on the unified time axis, the transaction processing control is performed according to the transaction processing control parameter in the control execution parameter sequence, the transaction amount data, the payment result data and the related timestamp data are written into the transaction processing related fields in the transaction record;

[0112] At each target time position on the unified time axis, the transaction processing control is performed according to the transaction processing control parameter in the control execution parameter sequence, the transaction amount data, the payment result data and the related timestamp data are written into the transaction processing related fields in the transaction record;

[0113] The transaction record and the distribution result record corresponding to each target time position on the unified time axis are written into the edge data structure in chronological order, and the time index and the type index are established in the edge data structure, so that the transaction record and the distribution result record can be accessed when the joint loss data set is constructed subsequently.

[0114] In the embodiment, the model parameter updating includes:

[0115] Based on the business state prediction result set, the transaction record and the distribution result record, an aligned sample set is constructed on a unified time axis, at each target time position on the unified time axis, the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the distribution time delay risk prediction result are extracted from the business state prediction result set, the actual transaction number and the actual transaction amount corresponding to the target time position are counted from the transaction record, the distribution time delay between the distribution completion time and the transaction generation time corresponding to the target time position is calculated from the distribution result record, the aligned sample record containing the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result, the distribution time delay risk prediction result, the actual user arrival intensity data, the actual transaction load data and the actual distribution time delay data is generated, and all the aligned sample records are arranged in the order of the unified time axis to generate the aligned sample set;

[0116] In each alignment sample record of the alignment sample set, a user arrival intensity deviation metric field is constructed based on the user arrival intensity prediction result and the actual user arrival intensity data, a transaction load deviation metric field is constructed based on the transaction load prediction result and the actual transaction load data, an energy consumption deviation metric field is constructed based on the energy consumption prediction result, a split time delay risk deviation metric field is constructed based on the split time delay risk prediction result and the actual split time delay data, and is written into the corresponding alignment sample record, forming a joint loss data set simultaneously representing four types of deviation quantities at each target time position on the unified time axis;

[0117] According to the joint loss data set, the parameters of the improved NCDE model are updated, the user arrival intensity deviation metric field, the transaction load deviation metric field, the energy consumption deviation metric field and the split time delay risk deviation metric field at each target time position on the unified time axis in the joint loss data set are taken as parameter update driving data, parameter optimization calculation based on the joint loss data set is performed, the parameter values of the control path encoding subnetwork and the vector field representation network in the improved NCDE model are adjusted, and an updated improved NCDE model is obtained.

[0118] In the subsequent running period, the multi-channel control path data structure generated by the new round of running data set is input into the updated improved NCDE model, a new continuous time hidden state vector sequence covering the unified time axis is obtained, and a new business state prediction result set is regenerated based on the new continuous time hidden state vector sequence, so that the updated improved NCDE model continuously outputs a business state prediction result set with higher consistency with the transaction record and the split result record in the subsequent running period.

[0119] Embodiment 1:

[0120] In order to verify the feasibility of the application in implementation, the application is applied to an automatic vending device using low-power Bluetooth for transaction triggering, and through the edge computing node, the Bluetooth communication data, the energy consumption data and the split business data are accessed to realize continuous time modeling, business state prediction, control decision generation and real-time processing of the whole process of transaction and split. In this scenario, the vending device is placed in a public area with continuous human flow, the device generates more short-time close behavior, Bluetooth connection request and transaction operation every day, the user arrival intensity presents strong fluctuation characteristics, the transaction load increases significantly in the periodic peak period, and the energy consumption also fluctuates continuously with the use frequency of the device. The traditional automatic vending device generally uses fixed Bluetooth broadcast interval, fixed transmission power and static split execution strategy, and in the high load scenario, problems such as high energy consumption, increasing Bluetooth connection delay, slow transaction response and unstable split time delay may occur, and the application aims to solve these key problems affecting the stable operation of the device.

[0121] In this scenario, the edge computing node deployed by the application continuously receives the number of broadcast packets generated by the Bluetooth communication processing unit, the connection establishment time, the channel occupation during the communication process, and the data interaction result, while recording the battery power change, power consumption curve and current fluctuation generated by the energy consumption monitoring unit, and combining the transaction amount, account distribution ratio, account distribution trigger time and account distribution completion time generated by the account distribution business processing unit. The three types of data are time-aligned and combined into a multi-channel control path data structure by the running data input module. Then, the improved NCDE model performs continuous time modeling on this structure to generate a continuous time hidden state vector sequence that can represent the change of device business state. The business state prediction module uses this sequence to generate user arrival intensity prediction results, transaction load prediction results, energy consumption prediction results and account distribution delay risk prediction results, and the control decision generation module generates communication power consumption adjustment parameters, transaction processing parameters and account distribution execution parameters based on these prediction results, so that the device can dynamically adjust between Bluetooth broadcast period, transmission power, transaction process and account distribution process. The control execution module controls the device operation according to the parameters, and the model update module continuously updates the NCDE model based on the device operation results, so that the device gradually obtains better prediction ability and control stability in the real environment.

[0122] In actual application process, the device records about 19200 times of Bluetooth broadcast in 24 hours, the user proximity event in peak period is about 3.4 times of that in off-peak period, the number of transactions increases obviously in the evening, and the energy consumption change trend presents continuous time correlation characteristics with user behavior. The application can adjust the Bluetooth broadcast period according to the predicted user arrival intensity, dynamically adjust from a fixed 200 milliseconds to between 150 and 300 milliseconds; the transmission power can also be automatically switched between 3dBm and 6dBm according to the transaction load; the transaction processing delay can be preset to be higher in time period with higher load, so that the transaction response is more stable; the account distribution business still maintains a shorter delay when the load is high. The application can also continuously update the model parameters under long-time running conditions, so that the business state prediction ability is gradually improved, and the control effect is further improved.

[0123] The application collects data on three key indicators, including Bluetooth communication average power consumption, transaction average response time, and account distribution average completion time. The traditional method and the application method are run simultaneously during the experiment, and each data is recorded under the same device and load conditions. The comparison data obtained is shown in the following table.

[0124] Table 1 Comparison of performance data between the application method and the traditional method

[0125] Indicator name Conventional method value Invention method value Bluetooth communication average power consumption mW 182.4 149.7 Transaction average response time length ms 412 355 Average completion time length of distribution ms 1280 1096 Daily average number of Bluetooth connection failures 27 19 Distribution failure rate percentage 2.17 1.58

[0126] As can be seen from Table 1, the application has stable improvement in three aspects of Bluetooth communication power consumption, transaction response time and account distribution delay. The average power consumption of Bluetooth communication is reduced by about 17.9% compared with the traditional method, mainly relying on the continuous time prediction of the change of user arrival intensity, so that the broadcast period and transmission power are adjusted in time when the load changes, avoiding long period fixed power consumption operation. The average transaction response time is reduced by about 13.8%, which benefits from the business state prediction, so that the transaction processing module can obtain more processing resources before the load rises, thereby reducing the blocking in the transaction process. In the account distribution link, the average completion time is reduced by about 14.4%, and the account distribution failure rate is reduced by about 0.59 percentage points, because the application can predict the account distribution delay risk in advance, so that the account distribution execution control is more stable. The number of Bluetooth connection failures is reduced, which shows that the device can better maintain the communication link under high load. The improved NCDE model of the application continuously learns the running data on the edge side, so that the prediction result and the control parameter are continuously optimized, and the overall operation stability and energy efficiency performance of the system are improved.

[0127] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. An edge computing based low power Bluetooth vending and real-time revenue sharing system, characterized in that, Comprise: Running data input module for collecting running data set in edge computing node, performing time alignment on unified time axis, generating multi-channel control path data structure; NCDE modeling module for loading improved NCDE model, taking multi-channel control path data structure as control input path, generating continuous time hidden state vector sequence; Business state prediction module for performing time evolution reasoning and continuous time extrapolation calculation based on continuous time hidden state vector sequence, generating extrapolated hidden state vector sequence, and generating business state prediction result set; Control decision generation module for splicing continuous time hidden state vector sequence and business state prediction result set to generate decision feature vector sequence, and generating control parameter set according to decision feature vector sequence; Control execution module for executing Bluetooth communication control, transaction processing control and account distribution execution control according to control parameter set, generating transaction record and account distribution result record, and writing transaction record and account distribution result record into edge data structure; Model update module for constructing joint loss data set, updating parameters of improved NCDE model, and calling updated improved NCDE model in subsequent running period.

2. The low-power Bluetooth automatic vending and real-time account distribution system based on edge computing according to claim 1, characterized in that, The modules are realized by the following methods: Collecting running data set in edge computing node, performing time alignment and generating multi-channel control path data structure; Building improved NCDE model in edge computing node, taking multi-channel control path data structure as model control input path, and generating continuous time hidden state vector sequence; Performing time evolution reasoning and state extrapolation calculation on continuous time hidden state vector sequence, and generating business state prediction result set; Based on continuous time hidden state vector sequence and business state prediction result set, the predicted quantity in business state prediction result set is converted into control parameter set through control decision generation calculation; According to control parameter set, Bluetooth communication control, transaction processing control and account distribution execution control are executed in edge computing node to generate transaction record and account distribution result record; Based on business state prediction result set, transaction record and account distribution result record, joint loss data set is constructed, parameters of improved NCDE model are updated according to joint loss data set, updated improved NCDE model is obtained, and updated improved NCDE model is called in subsequent running period.

3. The low-power Bluetooth vending and real-time accounting system based on edge computing according to claim 2, characterized in that, The generation of the multi-channel control path data structure includes: Data acquisition in edge computing node, obtaining original running data from edge computing node during running process, and writing into running data cache structure to generate initial running data set; The original running data in the initial running data set is parsed, and the parsed Bluetooth communication record, energy consumption record and account distribution business record are arranged in time stamp order respectively to generate Bluetooth communication data, energy consumption data and account distribution business data; A unified time axis is constructed, fixed time intervals are adopted, time stamps in Bluetooth communication data, energy consumption data and account sharing business data are mapped, and a missing data filling method is used for positions without observation values to generate time-aligned Bluetooth communication data, time-aligned energy consumption data and time-aligned account sharing business data; On each time position of the unified time axis, the time-aligned Bluetooth communication data, the time-aligned energy consumption data and the time-aligned account sharing business data are combined to generate a multi-channel control path data structure.

4. The low-power Bluetooth automatic vending and real-time account sharing system based on edge computing according to claim 2, characterized in that, The generation of the continuous-time hidden state vector includes: In the edge computing node, an input encoding structure is configured for the improved NCDE model, the Bluetooth communication data, the energy consumption data and the account sharing business data corresponding to each time position of the multi-channel control path data structure on the unified time axis are spliced to generate a control input vector sequence; A dimension parameter of the continuous-time hidden state vector is set, a hidden state initialization module is constructed, and the continuous-time hidden state vector is initialized to obtain a continuous-time hidden state vector corresponding to a starting time position; A control path encoding subnetwork and a vector field representation network are configured, the control path encoding subnetwork receives the control input vector sequence to generate a control feature representation at each time position of the unified time axis, and the vector field representation network receives the control feature representation and the continuous-time hidden state vector corresponding to each time position of the unified time axis to generate a hidden state change amount representation corresponding to the time position; A time step parameter is set, the continuous-time hidden state vector is updated at each time position on the unified time axis according to the time step parameter, a continuous-time hidden state vector corresponding to a current time position is calculated, and a continuous-time hidden state vector sequence covering all time positions of the unified time axis is obtained.

5. The low-power Bluetooth vending and real-time accounting system based on edge computing according to claim 2, characterized in that, The generation of the service state prediction result set includes: A prediction time window parameter is set based on the continuous-time hidden state vector sequence, a target prediction time position set corresponding to the unified time axis is determined, a continuous-time hidden state vector adjacent to the target prediction time position set is selected from the continuous-time hidden state vector sequence to generate a prediction input hidden state sequence; The time evolution structure of the improved NCDE model is called, time evolution reasoning is performed on the prediction input hidden state sequence according to the prediction time window parameter, continuous-time extrapolation calculation is performed on the prediction input hidden state sequence on the unified time axis to generate an extrapolation hidden state vector sequence corresponding to the target prediction time position set; Based on the extrapolation hidden state vector sequence, a user arrival intensity mapping structure, a transaction load mapping structure, an energy consumption mapping structure and an account sharing time delay risk mapping structure are respectively constructed, and the extrapolation hidden state vector sequence is input into the structures to respectively generate a user arrival intensity prediction result, a transaction load prediction result, an energy consumption prediction result and an account sharing time delay risk prediction result. A service state prediction result set containing the user arrival intensity prediction result, the transaction load prediction result, the energy consumption prediction result and the account sharing time delay risk prediction result is obtained.

6. The low-power Bluetooth automatic vending and real-time account sharing system based on edge computing according to claim 2, characterized in that, The generation of the control parameter set includes: The continuous time feature quantity in the continuous time hidden state vector sequence corresponding to each target time position on the unified time axis is spliced with the service state prediction result set to generate a decision feature vector corresponding to the target time position, and the decision feature vectors are sequentially arranged to form a decision feature vector sequence; A weight parameter set is constructed in the control decision generation calculation based on the decision feature vector sequence, and the weight values of the weight parameters at the target time position are determined according to the component values in the decision feature vector, so as to obtain a weight parameter set covering all target time positions; At each target time position on the unified time axis, the user arrival intensity prediction result and the energy consumption prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for power consumption control in the weight parameter set, to generate a Bluetooth communication control parameter sequence; The user arrival intensity prediction result, the transaction load prediction result and the account sharing delay risk prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for transaction processing in the weight parameter set, to generate a transaction processing control parameter sequence; The transaction load prediction result and the account sharing delay risk prediction result in the service state prediction result set are weighted and calculated according to the weight parameter for account sharing execution in the weight parameter set, to generate an account sharing execution control parameter sequence; The continuous time feature quantity related to the use of edge computing resources in the continuous time hidden state vector sequence is weighted and calculated with the service state prediction result set according to the weight parameter for edge computing power allocation in the weight parameter set, to generate an edge computing power allocation control parameter sequence; The Bluetooth communication control parameter sequence, the transaction processing control parameter sequence, the account sharing execution control parameter sequence and the edge computing power allocation control parameter sequence are combined to generate a control parameter set.

7. The low-power Bluetooth automatic vending and real-time account sharing system based on edge computing according to claim 2, characterized in that, The generation of the transaction record and the account sharing result record includes: Bluetooth communication control parameters, transaction processing control parameters, account sharing execution control parameters and edge computing power allocation control parameters are extracted from the control parameter set to generate a control execution parameter sequence; At each target time position on the unified time axis, Bluetooth communication control is performed according to the Bluetooth communication control parameters, data generated in the Bluetooth communication control process is recorded, and the data is written into the transaction record; Transaction processing control is performed according to the transaction processing control parameters, transaction amount data, payment result data and related timestamp data are written into the transaction record; Account sharing execution control is performed according to the account sharing execution control parameters, account sharing amount data, account sharing proportion data and account sharing completion time data are written into the account sharing result record; The transaction record and the account sharing result record corresponding to each target time position on the unified time axis are written into the edge data structure in chronological order.

8. The low-power Bluetooth automatic vending and real-time account sharing system based on edge computing according to claim 2, characterized in that, The model parameter updating includes: At each target time position on the unified time axis, an aligned sample set is generated based on the service state prediction result set, the actual number of transactions and the actual transaction amount corresponding to the target time position are counted from the transaction record, and the account sharing delay between the account sharing completion time and the transaction generation time corresponding to the target time position is calculated from the account sharing result record. In each alignment sample record of the alignment sample set, a user arrival intensity deviation measure field is constructed based on the user arrival intensity prediction result and the actual user arrival intensity data, a transaction load deviation measure field is constructed based on the transaction load prediction result and the actual transaction load data, an energy consumption deviation measure field is constructed based on the energy consumption prediction result, a split time delay risk deviation measure field is constructed based on the split time delay risk prediction result and the actual split time delay data, and the corresponding alignment sample record is written to generate a joint loss data set; According to the joint loss data set, the parameters of the improved NCDE model are updated, the joint loss data set is used as parameter update driving data, parameter optimization calculation based on the joint loss data set is performed, and an updated improved NCDE model is obtained; In the subsequent running period, the multi-channel control path data structure generated by the new round of running data set is input into the updated improved NCDE model, so that the updated improved NCDE model continuously outputs a business state prediction result set with higher consistency with the transaction record and the split result record in the subsequent running period.

Images