CN-121998611-A - Financial self-service terminal predictive maintenance method

CN121998611ACN 121998611 ACN121998611 ACN 121998611ACN-121998611-A

Abstract

The invention provides a financial self-service terminal predictive maintenance method which comprises the following steps of S1, collecting physical state data, intercepting business transaction log data, S2, carrying out alignment slicing according to a unified timestamp, respectively inputting the physical flow branches and the business flow branches in a preset cross-mode space-time fusion network, S3, carrying out fusion processing, S4, projecting the fused multi-mode feature vector into a contrast learning feature space to obtain contrast feature vectors, calculating Euclidean distance between the contrast feature vectors and the center point of each fault prototype, S5, outputting residual service life predicted values of key components of the financial self-service terminal by a regression prediction layer, determining confidence scores based on the residual service life predicted values and the Euclidean distance, and generating maintenance work orders when the residual service life predicted values are lower than a preset threshold and the confidence scores are higher than the confidence threshold. The invention improves the prediction accuracy of the residual life of the financial self-service terminal equipment and remarkably improves the operation and maintenance efficiency.

Inventors

CUI YONGJIE
CAI MING
FANG WEICONG
YE JUNBIN
CHEN HAOWEI

Assignees

杭州易雅通科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251224

Claims (8)

1. The predictive maintenance method for the financial self-service terminal is characterized by being applied to a predictive maintenance system with an end-side sensing component, an edge computing component and a cloud platform component, and comprises the following steps of: S1, acquiring physical state data through a physical sensor deployed at a key part of a financial self-service terminal, and intercepting business transaction log data through a middleware interface of the financial self-service terminal, wherein the physical state data comprises vibration frequency and motor current waveforms, and the business transaction log data comprises transaction type instructions and a module action sequence; S2, carrying out alignment slicing on the physical state data and the business transaction log data according to a uniform timestamp, and respectively inputting the physical state data and the business transaction log data into a physical flow branch and a business flow branch in a preset cross-mode space-time fusion network; s3, utilizing a cross-mode gating attention unit in the cross-mode space-time fusion network to fuse the physical feature vector sequence and the service feature vector sequence, wherein the processing procedure of the cross-mode gating attention unit comprises the following steps: Taking the service feature vector sequence as a query vector, and taking the physical feature vector sequence as a key vector and a value vector at the same time; calculating the product of the query vector and the transpose matrix of the key vector, and obtaining an attention score matrix through scaling and normalization processing; Generating a gating coefficient between 0 and 1 based on the query vector by using a gating generator, multiplying the gating coefficient by the value vector weighted by the attention score matrix element by element, and outputting a fused multi-modal feature vector; S4, projecting the fused multi-mode feature vector to a contrast learning feature space by using a contrast learning projection layer in the cross-mode space-time fusion network to obtain a contrast feature vector, and calculating Euclidean distance between the contrast feature vector and each fault prototype center point in a fault prototype library preset in a cloud platform assembly; S5, inputting the fused multi-mode feature vector to a regression prediction layer of the cross-mode space-time fusion network, outputting a residual service life predicted value of the key component of the financial self-service terminal by the regression prediction layer, determining a confidence coefficient score based on the residual service life predicted value and the Euclidean distance, and generating a maintenance work order when the residual service life predicted value is lower than a preset threshold and the confidence coefficient score is higher than the confidence threshold.
2. The method for predictively maintaining the financial self-service terminal according to claim 1, wherein in the step S1, key components of the financial self-service terminal comprise a banknote discharging module, a card reader module and a receipt printing module, the step of intercepting business transaction log data through a middleware interface of the financial self-service terminal specifically comprises the steps of immediately capturing a bottom hardware instruction response through an XFS interface, analyzing and obtaining the number of deposited and withdrawn banknotes of a current transaction, a card reader throughput action and a module reset instruction, and recording the start time and the end time of instruction execution.
3. The method for predictive maintenance of a financial self-service terminal as set forth in claim 1, wherein in the step S2, the physical state data and the business transaction log data are aligned and sliced according to a uniform timestamp, and the method specifically includes: Taking a starting time point and an ending time point of each complete transaction in the business transaction log data as window boundaries, and intercepting a physical state data stream in the time window; and if the number of the physical state data sampling points in the time window is less than the preset length, zero filling is carried out, and if the number of the physical state data sampling points is more than the preset length, the maximum pooling operation is carried out, so that the length of the physical state data sampling points is matched with the time step of the service feature vector sequence.
4. The method for predictive maintenance of a financial self-service terminal as set forth in claim 1, wherein in the step S3, the calculation formula of the cross-mode gating attention unit is as follows: Wherein Q is a query vector of a business feature vector sequence after linear transformation, K is a key vector of a physical feature vector sequence after linear transformation, V is a value vector of the physical feature vector sequence after linear transformation, sigma is a Sigmoid activation function, W g is a weight matrix of a gate generator, Representing an element-wise multiplication operation, d k is a scaling factor.
5. The method for predictive maintenance of a financial self-service terminal according to claim 1, wherein in the step S4, the fault prototype library is constructed by the following method: Collecting feature vectors of historical fault samples at a cloud platform assembly; Training the model by adopting a supervision contrast loss function, so that sample features belonging to the same fault type are gathered in a feature space, and sample features belonging to different fault types are mutually exclusive in the feature space; And calculating the mass center of each type of fault sample cluster, and taking the mass center as a fault prototype center point of the type of fault.
6. The method for predictive maintenance of a financial self-service terminal as set forth in claim 1, wherein in the step S5, the confidence score is calculated by using the following Monte Carlo Dropout method: keeping a Dropout layer in a network in an open state in an reasoning stage; carrying out forward propagation on the same sample for N times to obtain N residual service life prediction results; And calculating variances of the N residual service life prediction results, and normalizing the reciprocal of the variances to be used as the confidence score.
7. The method for predictively maintaining a financial self-service terminal according to claim 1, further comprising the steps of: and uploading the original physical state data and business transaction log data of the difficult sample to a cloud platform component, re-labeling and training the difficult sample, and updating a fault prototype library.
8. The method for predictively maintaining a financial self-service terminal according to claim 1, further comprising the steps of: Acquiring a historical service peak time table of a website where a financial self-service terminal is located; if the predicted time point corresponding to the residual service life is located in the service peak time period, the trigger threshold for generating the maintenance work order is reduced, and the maintenance work order is generated in advance.

Description

Financial self-service terminal predictive maintenance method Technical Field The invention relates to a financial terminal maintenance method, in particular to a financial self-service terminal predictive maintenance method, and belongs to the technical field of intelligent operation and maintenance of financial equipment. Background With the digital transformation of banking business, financial self-service terminals such as Automatic Teller Machines (ATM) and integrated deposit and withdrawal machines (STM) have become core channels for banking website service clients. The internal structure of the equipment is complex, a large number of mechanical transmission parts are included, and abrasion, banknote clamping or module failure easily occur under high-frequency use. The old financial self-service terminal maintenance mode often generates a repair work order after equipment fails, so that the equipment has longer downtime. The existing predictive maintenance technology generally only depends on the data of a physical sensor, however, the physical state of a financial terminal is strongly related to service actions, the existing technology lacks the capability of combining service logic with the physical state, and normal high-load service operation is easily misjudged as mechanical abnormality, so that ineffective on-door maintenance is caused. Severe faults of the financial terminals occur less frequently than industrial motors, resulting in fewer negative samples, i.e. fault data. Therefore, under the condition that data is extremely unbalanced, the traditional regression-based deep learning model is difficult to train a high-precision prediction model to effectively support a predictive maintenance mode of the financial self-service terminal. In addition, the predictive maintenance mode also faces the contradiction of computational effort and timeliness, if all high-frequency sensor data are uploaded to the cloud for processing, the bandwidth cost is high, privacy risks exist, if the high-frequency sensor data are processed only at the end side, the high-frequency sensor data are limited by the computational effort of an industrial personal computer, and a complex deep learning model is difficult to operate. Disclosure of Invention Based on the background, the invention aims to provide the predictive maintenance method for the financial self-service terminal, which improves the prediction accuracy of the residual service life of the financial self-service terminal equipment, can greatly reduce the false alarm rate by combining with business context and remarkably improves the operation and maintenance efficiency. In order to achieve the above object, the present invention provides the following technical solutions: the predictive maintenance method of the financial self-service terminal is applied to a predictive maintenance system with an end-side sensing component, an edge computing component and a cloud platform component, and comprises the following steps of: S1, acquiring physical state data through a physical sensor deployed at a key part of a financial self-service terminal, and intercepting business transaction log data through a middleware interface of the financial self-service terminal, wherein the physical state data comprises vibration frequency and motor current waveforms, and the business transaction log data comprises transaction type instructions and a module action sequence; S2, carrying out alignment slicing on the physical state data and the business transaction log data according to a uniform timestamp, and respectively inputting the physical state data and the business transaction log data into a physical flow branch and a business flow branch in a preset cross-mode space-time fusion network; s3, utilizing a cross-mode gating attention unit in the cross-mode space-time fusion network to fuse the physical feature vector sequence and the service feature vector sequence, wherein the processing procedure of the cross-mode gating attention unit comprises the following steps: Taking the service feature vector sequence as a query vector, and taking the physical feature vector sequence as a key vector and a value vector at the same time; calculating the product of the query vector and the transpose matrix of the key vector, and obtaining an attention score matrix through scaling and normalization processing; Generating a gating coefficient between 0 and 1 based on the query vector by using a gating generator, multiplying the gating coefficient by the value vector weighted by the attention score matrix element by element, and outputting a fused multi-modal feature vector; S4, projecting the fused multi-mode feature vector to a contrast learning feature space by using a contrast learning projection layer in the cross-mode space-time fusion network to obtain a contrast feature vector, and calculating Euclidean distance between the contrast feature vector and each fault prototype center poi