Financial self-service terminal predictive maintenance method
By combining business transaction logs and physical sensor data through a cross-modal spatiotemporal fusion network, high-precision predictive maintenance of financial self-service terminals has been achieved, solving the problems of high false alarm rate and low prediction accuracy in existing technologies, and improving operation and maintenance efficiency and the accuracy of equipment life prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU YIYATONG TECH CO LTD
- Filing Date
- 2025-12-24
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies for predictive maintenance of financial self-service terminals struggle to effectively combine business logic with physical status, resulting in high false alarm rates, low prediction accuracy under unbalanced data conditions, and a conflict between computing power and timeliness.
A cross-modal spatiotemporal fusion network is adopted, which performs real-time data processing through edge sensing components and edge computing components. Combined with business transaction logs and physical sensor data, fault prediction is performed using cross-modal gating attention units and contrastive learning projection layers to generate maintenance work orders.
It improves the accuracy of remaining life prediction for financial self-service terminal equipment, significantly reduces false alarm rate, improves operation and maintenance efficiency, and avoids resource waste and privacy risks.
Smart Images
Figure CN121998611A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a financial terminal maintenance method, specifically a predictive maintenance method for financial self-service terminals, belonging to the field of intelligent operation and maintenance technology for financial equipment. Background Technology
[0002] With the digital transformation of banking operations, self-service financial terminals such as ATMs and STMs have become core channels for bank branches to serve customers. These devices have complex internal structures and contain a large number of mechanical transmission components, making them prone to wear and tear, banknote jams, or module failures under high-frequency use.
[0003] The outdated maintenance model for financial self-service terminals often generates repair orders only after equipment failure, resulting in prolonged downtime. Existing predictive maintenance technologies typically rely solely on data from physical sensors. However, the physical state of financial terminals is strongly correlated with business operations, and current technologies lack the ability to integrate business logic with physical status. This can easily lead to misjudging normal, high-load business operations as mechanical malfunctions, resulting in ineffective on-site maintenance. Compared to industrial motors, serious failures in financial terminals occur less frequently, resulting in very few negative samples (failure data). Therefore, traditional regression-based deep learning models struggle to train high-accuracy predictive models to effectively support predictive maintenance for financial self-service terminals under conditions of extreme data imbalance. Furthermore, predictive maintenance faces a trade-off between computing power and timeliness. Uploading all high-frequency sensor data to the cloud for processing incurs high bandwidth costs and poses privacy risks. Processing only on the terminal side is limited by the computing power of industrial control computers, making it difficult to run complex deep learning models. Summary of the Invention
[0004] Based on the above background, the purpose of this invention is to provide a predictive maintenance method for financial self-service terminals, which improves the accuracy of predicting the remaining lifespan of financial self-service terminal equipment, and can significantly reduce the false alarm rate by combining business context, thereby significantly improving operation and maintenance efficiency.
[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: A predictive maintenance method for financial self-service terminals, applied to a predictive maintenance system with edge sensing components, edge computing components, and cloud platform components, includes the following steps: S1. Physical status data is collected by physical sensors deployed at key components of the financial self-service terminal, and business transaction log data is captured through the middleware interface of the financial self-service terminal; the physical status data includes vibration frequency and motor current waveform, and the business transaction log data includes transaction type instructions and module action sequences. S2. Align and slice the physical state data and the business transaction log data according to a unified timestamp, and input them into the physical flow branch and business flow branch in the preset cross-modal spatiotemporal fusion network respectively; the physical flow branch uses a one-dimensional convolutional layer to process the physical state data to generate a physical feature vector sequence, and the business flow branch uses an embedding layer to map the business transaction log data into a business feature vector sequence; S3. The physical feature vector sequence and the business feature vector sequence are fused using the cross-modal gating attention unit in the cross-modal spatiotemporal fusion network; the processing of the cross-modal gating attention unit includes: The business feature vector sequence is used as the query vector, and the physical feature vector sequence is used as both the key vector and the value vector. The product of the query vector and the transpose of the key vector is calculated, and the attention score matrix is obtained after scaling and normalization. A gating generator is used to generate gating coefficients between 0 and 1 based on the query vector. The gating coefficients are then multiplied element-wise with the value vector after weighting the attention score matrix to output the fused multimodal feature vector. S4. Using the contrastive learning projection layer in the cross-modal spatiotemporal fusion network, the fused multimodal feature vector is projected onto the contrastive learning feature space to obtain a contrastive feature vector. The Euclidean distance between the contrastive feature vector and the center point of each fault prototype in the fault prototype library pre-installed in the cloud platform component is calculated. S5. Input the fused multimodal feature vector into the regression prediction layer of the cross-modal spatiotemporal fusion network, and output the remaining service life prediction value of the key components of the financial self-service terminal by the regression prediction layer; determine the confidence score based on the remaining service life prediction value and the Euclidean distance; when the remaining service life prediction value is lower than a preset threshold and the confidence score is higher than the confidence threshold, generate a maintenance work order.
[0006] Preferably, in step S1, the key components of the financial self-service terminal include a cash dispensing module, a card reader module, and a receipt printing module; the step of intercepting business transaction log data through the middleware interface of the financial self-service terminal specifically includes: capturing the underlying hardware instruction response in real time through the XFS interface, parsing and obtaining the number of banknotes deposited or withdrawn in the current transaction, the card reader throughput action and the module reset instruction, and recording the start and end times of instruction execution.
[0007] Preferably, in step S2, the physical status data and the business transaction log data are aligned and sliced according to a unified timestamp, specifically including: Using the start and end times of each complete transaction in the business transaction log data as window boundaries, extract the physical status data stream within that time window; If the number of physical state data sampling points within the time window is less than the preset length, zero padding is performed; if it is more than the preset length, max pooling is performed to match its length with the time step of the business feature vector sequence.
[0008] Preferably, in step S3, the calculation formula for the cross-modal gated attention unit is as follows: In the formula, Q is the query vector after linear transformation of the business feature vector sequence, K is the key vector after linear transformation of the physical feature vector sequence, V is the value vector after linear transformation of the physical feature vector sequence, σ is the Sigmoid activation function, and W... g This is the weight matrix of the gated generator. d represents the element-wise multiplication operation. k This is the scaling factor.
[0009] Preferably, in step S4, the fault prototype library is constructed using the following method: The cloud platform component collects feature vectors from historical failure samples; The model is trained using a supervised contrastive loss function, which makes the features of samples belonging to the same fault type cluster in the feature space, while the features of samples belonging to different fault types exclude each other in the feature space. Calculate the centroid of each type of fault sample cluster, and use the centroid as the fault prototype center point of that type of fault.
[0010] Preferably, in step S5, the confidence score is calculated using the Monte Carlo Dropout method: Keep the Dropout layer in the network enabled during the inference phase; Perform N forward propagations on the same sample to obtain N remaining lifetime prediction results; Calculate the variance of the N remaining useful life prediction results, and normalize the inverse of the variance to obtain the confidence score.
[0011] Preferably, the method further includes the following steps: When the calculated confidence score is lower than the preset confidence threshold, the current data is determined to be a difficult sample; the original physical state data and business transaction log data of the difficult sample are uploaded to the cloud platform component, the difficult sample is re-labeled and trained, and the fault prototype library is updated.
[0012] Preferably, the method further includes the following steps: Obtain the historical peak business time schedule of the branch where the financial self-service terminal is located; If the predicted remaining service life corresponds to a time point within the peak business period, the trigger threshold for generating maintenance work orders will be lowered to generate maintenance work orders in advance.
[0013] Compared with the prior art, the present invention has the following advantages: The present invention provides a predictive maintenance method for financial self-service terminals. By constructing a cloud-edge collaborative system that integrates edge perception, edge reasoning, and cloud evolution, and utilizing a cross-modal spatiotemporal fusion network, it can proactively predict the remaining lifespan of financial self-service terminal equipment, improve prediction accuracy, and significantly reduce false alarm rate by combining business context, thereby significantly improving operation and maintenance efficiency. This invention introduces a cross-modal gating attention unit. Unlike existing technologies that analyze sensor data separately, this solution uses XFS service logs as query vectors to dynamically gate physical sensor signals. This mechanism effectively filters out environmental noise caused by normal service operations and solves the problem of false alarms caused by service load fluctuations. To address the issue of limited financial fault samples, this invention introduces a prototype-based contrastive learning projection layer into the network. By learning the distance between data points and the center of the fault prototype in the feature space, rather than simply relying on regression fitting, the model can keenly capture the trend of equipment status evolving towards a certain fault mode even when there are only a few fault samples, thereby improving the ability to identify rare faults. This invention calculates the confidence score of the prediction using the Monte Carlo Dropout mechanism while outputting the prediction results. A work order is triggered only when the model is very certain about the prediction results. For uncertain predictions, the results are uploaded to the cloud for review and model retraining. This mechanism effectively avoids the waste of operation and maintenance resources caused by model illusion and forms a closed loop for continuous model optimization. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0015] Figure 1 This is a schematic diagram of the overall architecture of the predictive maintenance system in this invention; Figure 2 This is a schematic diagram of the model structure of the cross-modal spatiotemporal fusion network (MSTF-Transformer) in this invention; Figure 3This is a schematic diagram of the computation process of the cross-modal gated attention unit (CM-GAU) in this invention. Detailed Implementation
[0016] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.
[0017] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.
[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In this detailed description, numerous specific details are set forth to facilitate explanation and provide a thorough understanding of the embodiments of the present invention. However, one or more embodiments may be practiced by those skilled in the art without these specific details.
[0019] An embodiment of the present invention discloses a predictive maintenance method for a financial self-service terminal. This method is applied to a predictive maintenance system having an edge sensing component, an edge computing component, and a cloud platform component. The method includes the following steps: S1. Physical status data is collected by physical sensors deployed at key components of the financial self-service terminal, and business transaction log data is captured through the middleware interface of the financial self-service terminal; the physical status data includes vibration frequency and motor current waveform, and the business transaction log data includes transaction type instructions and module action sequences. S2. Align and slice the physical state data and business transaction log data according to a unified timestamp, and input them into the physical flow branch and business flow branch in the pre-built cross-modal spatiotemporal fusion network respectively; the physical flow branch uses a one-dimensional convolutional layer to process the physical state data to generate a sequence of physical feature vectors, and the business flow branch uses an embedding layer to map the business transaction log data into a sequence of business feature vectors. S3. The physical feature vector sequence and the business feature vector sequence are fused using the cross-modal gated attention unit in the cross-modal spatiotemporal fusion network; the processing of the cross-modal gated attention unit includes: Use the business feature vector sequence as the query vector, and use the physical feature vector sequence as both the key vector and the value vector. The product of the query vector and the transpose of the key vector is calculated, and the attention score matrix is obtained after scaling and normalization. The gating generator generates gating coefficients between 0 and 1 based on the query vector. The gating coefficients are then multiplied element-wise with the value vector weighted by the attention score matrix to output the fused multimodal feature vector. S4. Using the contrastive learning projection layer in the cross-modal spatiotemporal fusion network, the fused multimodal feature vectors are projected onto the contrastive learning feature space to obtain contrastive feature vectors. The Euclidean distance between the contrastive feature vectors and the center points of each fault prototype in the fault prototype library pre-installed in the cloud platform component is calculated. S5. Input the fused multimodal feature vector into the regression prediction layer of the cross-modal spatiotemporal fusion network. The regression prediction layer outputs the remaining service life prediction value of the key components of the financial self-service terminal. Determine the confidence score based on the remaining service life prediction value and Euclidean distance. When the remaining service life prediction value is lower than the preset threshold and the confidence score is higher than the confidence threshold, a maintenance work order is generated.
[0020] like Figure 1 As shown, the predictive maintenance system employs a typical cloud-edge-device collaborative architecture. The edge-side sensing components include miniature IoT sensor arrays installed at key locations such as the cash dispensing guide rail, card reader transmission wheel assembly, and receipt printer cutter in the financial self-service terminal. These sensors can employ high-frequency piezoelectric vibration sensors or non-invasive Hall effect current sensors, acquiring vibration frequency data at a sampling rate of 2kHz-10kHz and motor current waveform data at 50Hz-100Hz. Furthermore, a lightweight agent is deployed on the industrial control computer of the financial self-service terminal. This agent uses hook technology or log analysis to capture XFS middleware interface logs in real time, parsing and obtaining the number of banknotes deposited or withdrawn in the current transaction, card reader throughput actions, and module reset commands, while recording the start and end times of command execution.
[0021] Edge computing components are deployed in the aggregation servers or high-performance industrial control computers of bank branches, running a pruned and quantized cross-modal spatiotemporal fusion network as a real-time inference engine responsible for executing real-time inference tasks. The edge computing components are also responsible for data cleaning, anonymization, and the screening and uploading of difficult samples.
[0022] The cloud platform component is deployed on the bank's private cloud, maintaining digital health records and a fault prototype library, and running the model evolution center. The cloud platform component is responsible for retraining the model using data transmitted from across the network, updating the centroid parameters of the fault prototype library, and periodically distributing updated model weights to the edge computing component.
[0023] Steps S2-S5 all revolve around the core cross-modal spatiotemporal fusion network (MSTF-Transformer), which will be discussed below. Figure 2and Figure 3 Steps S2-S5 will be explained in detail.
[0024] The purpose of step S2 is to perform dual-stream feature extraction and time-series alignment. Since physical signals are continuous high-frequency waveforms, while business logs are discrete event sequences, they must be aligned. Specifically, using the start and end times of each complete transaction in the business transaction log data as window boundaries, the physical state data stream within this time window is extracted. If the number of physical state data sampling points within the time window is less than a preset length, zero-padding is performed; if it exceeds the preset length, max-pooling is performed to match its length with the time step of the business feature vector sequence. After processing, the physical state data stream is input into a one-dimensional convolutional neural network (1D-CNN) to extract the physical feature vector sequence. For the business transaction log data stream, the transaction type and action sequence are one-hot encoded and then mapped to a high-dimensional, dense business feature vector sequence through an embedding layer.
[0025] The purpose of step S3 is cross-modal feature fusion inference. To improve the model's understanding of business logic, this invention constructs a cross-modal gated attention unit (CM-GAU). The calculation formula for the cross-modal gated attention unit is as follows: In the formula, Q is the query vector after linear transformation of the business feature vector sequence, K is the key vector after linear transformation of the physical feature vector sequence, V is the value vector after linear transformation of the physical feature vector sequence, σ is the Sigmoid activation function, and W... g This is the weight matrix of the gated generator. d represents the element-wise multiplication operation. k This is the scaling factor.
[0026] The calculation process of the cross-modal gating attention unit is as follows: taking the business feature vector sequence and the physical feature vector sequence as input, after QKV mapping, the attention, i.e., the correlation between the current business action and the physical state, is calculated. Then, using a gating generator, the gating coefficient representing the importance weight of the physical signal in the current business scenario is calculated. Finally, the attention is multiplied element by element by the gating coefficient to obtain the output.
[0027] The significance of setting up a cross-modal gating attention unit is that, for example, when the business transaction log data stream shows that it is idle, the gating coefficient will approach 1, amplifying what may be abnormally small physical vibrations. However, when the business transaction log data stream shows that it is dispensing cash at full speed, the gating coefficient will suppress specific frequency bands and filter out normal mechanical noise.
[0028] The purpose of step S4 is to perform anomaly comparison based on prototypes. The fault prototype library is constructed by: collecting feature vectors of historical fault samples in a cloud platform component; training the model using a supervised contrastive loss function so that features of samples belonging to the same fault type cluster in the feature space, while features of samples belonging to different fault types are mutually exclusive in the feature space; calculating the centroid of each fault sample cluster and using the centroid as the fault prototype center point for that type of fault.
[0029] Calculate the Euclidean distance between the current feature vector and the center point of each fault prototype. If the distance between the current feature vector and a fault prototype is less than the preset alarm radius, it indicates that the device state is evolving towards that fault, even if the current RUL prediction value may not have reached the critical point.
[0030] The purpose of step S5 is to maintain the decision and confidence assessment. The feature vector directly outputs the remaining lifetime through the fully connected layer. To prevent the model from being blindly confident in unknown data, the confidence score is calculated using the following Monte Carlo Dropout method: Keep the Dropout layer in the network enabled during the inference phase; Perform N forward propagations on the same sample to obtain N remaining lifetime prediction results; Calculate the variance of the N remaining useful life prediction results, and normalize the inverse of the variance to obtain the confidence score.
[0031] When the calculated confidence score is lower than the preset confidence threshold, the current data is determined to be a difficult sample; the original physical state data and business transaction log data of the difficult sample are uploaded to the cloud platform component, the difficult sample is re-labeled and trained, and the fault prototype library is updated.
[0032] Furthermore, the method includes a step of dynamically adjusting based on peak business hours. It obtains the historical peak business schedule of the branch where the self-service terminal is located. If the predicted remaining lifespan corresponds to a time point within the peak business period, the trigger threshold for generating maintenance work orders is lowered, allowing maintenance work orders to be generated earlier. This ensures that equipment maintenance is completed before peak business hours, avoiding downtime during peak periods.
[0033] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
Claims
1. A predictive maintenance method for a financial self-service terminal, characterized in that: This method is applied to a predictive maintenance system with edge sensing components, edge computing components, and cloud platform components. The method includes the following steps: S1. Physical status data is collected by physical sensors deployed at key components of the financial self-service terminal, and business transaction log data is captured through the middleware interface of the financial self-service terminal; the physical status data includes vibration frequency and motor current waveform, and the business transaction log data includes transaction type instructions and module action sequences. S2. Align and slice the physical state data and the business transaction log data according to a unified timestamp, and input them into the physical flow branch and business flow branch in the preset cross-modal spatiotemporal fusion network respectively; the physical flow branch uses a one-dimensional convolutional layer to process the physical state data to generate a physical feature vector sequence, and the business flow branch uses an embedding layer to map the business transaction log data into a business feature vector sequence; S3. The physical feature vector sequence and the business feature vector sequence are fused using the cross-modal gating attention unit in the cross-modal spatiotemporal fusion network; the processing of the cross-modal gating attention unit includes: The business feature vector sequence is used as the query vector, and the physical feature vector sequence is used as both the key vector and the value vector. The product of the query vector and the transpose of the key vector is calculated, and the attention score matrix is obtained after scaling and normalization. A gating generator is used to generate gating coefficients between 0 and 1 based on the query vector. The gating coefficients are then multiplied element-wise with the value vector after weighting the attention score matrix to output the fused multimodal feature vector. S4. Using the contrastive learning projection layer in the cross-modal spatiotemporal fusion network, the fused multimodal feature vector is projected onto the contrastive learning feature space to obtain a contrastive feature vector. The Euclidean distance between the contrastive feature vector and the center point of each fault prototype in the fault prototype library pre-installed in the cloud platform component is calculated. S5. Input the fused multimodal feature vector into the regression prediction layer of the cross-modal spatiotemporal fusion network, and output the remaining service life prediction value of the key components of the financial self-service terminal by the regression prediction layer; determine the confidence score based on the remaining service life prediction value and the Euclidean distance; when the remaining service life prediction value is lower than a preset threshold and the confidence score is higher than the confidence threshold, generate a maintenance work order.
2. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: In step S1, the key components of the financial self-service terminal include a cash dispensing module, a card reader module, and a receipt printing module. The process of intercepting business transaction log data through the middleware interface of the financial self-service terminal specifically includes: capturing the underlying hardware instruction response in real time through the XFS interface, parsing and obtaining the number of banknotes deposited or withdrawn in the current transaction, the card reader throughput action and module reset instruction, and recording the start and end times of instruction execution.
3. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: In step S2, the physical status data and the business transaction log data are aligned and sliced according to a unified timestamp, specifically including: Using the start and end times of each complete transaction in the business transaction log data as window boundaries, extract the physical status data stream within that time window; If the number of physical state data sampling points within the time window is less than the preset length, zero padding is performed; if it is more than the preset length, max pooling is performed to match its length with the time step of the business feature vector sequence.
4. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: In step S3, the calculation formula for the cross-modal gated attention unit is as follows: In the formula, Q is the query vector after linear transformation of the business feature vector sequence, K is the key vector after linear transformation of the physical feature vector sequence, V is the value vector after linear transformation of the physical feature vector sequence, σ is the Sigmoid activation function, and W... g This is the weight matrix of the gated generator. d represents the element-wise multiplication operation. k This is the scaling factor.
5. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: In step S4, the fault prototype library is constructed using the following method: The cloud platform component collects feature vectors from historical failure samples; The model is trained using a supervised contrastive loss function, which makes the features of samples belonging to the same fault type cluster in the feature space, while the features of samples belonging to different fault types exclude each other in the feature space. Calculate the centroid of each type of fault sample cluster, and use the centroid as the fault prototype center point of that type of fault.
6. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: In step S5, the confidence score is calculated using the Monte Carlo Dropout method: Keep the Dropout layer in the network enabled during the inference phase; Perform N forward propagations on the same sample to obtain N remaining lifetime prediction results; Calculate the variance of the N remaining useful life prediction results, and normalize the inverse of the variance to obtain the confidence score.
7. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: The method also includes the following steps: When the calculated confidence score is lower than the preset confidence threshold, the current data is determined to be a difficult sample; the original physical state data and business transaction log data of the difficult sample are uploaded to the cloud platform component, the difficult sample is re-labeled and trained, and the fault prototype library is updated.
8. The predictive maintenance method for a financial self-service terminal according to claim 1, characterized in that: The method also includes the following steps: Obtain the historical peak business time schedule of the branch where the financial self-service terminal is located; If the predicted remaining service life corresponds to a time point within the peak business period, the trigger threshold for generating maintenance work orders will be lowered to generate maintenance work orders in advance.