A Metro BOM Ticketing Processing Method and System Based on Artificial Intelligence

Abstract

This application relates to an AI-based method and system for subway BOM ticketing processing, addressing the problems of decreased verification accuracy, low processing efficiency, and insufficient overall adaptability in existing subway BOM systems due to over-reliance on a single electronic data dimension and fixed processes. The method includes: collecting ticket images and electronic data and adaptively switching collection modes to extract and verify visual and electronic features. Upon successful verification, a second verification is performed using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement to associate with discount rules. If verification fails or features are not up to standard, user sentiment is analyzed, the root cause of the anomaly is diagnosed, and system parameters are dynamically optimized to generate guidance. Finally, intelligent ticketing routing is performed based on reinforcement learning, or an adaptation scheme is executed based on risk assessment, achieving intelligent processing throughout the entire process. This application has the following effects: achieving dynamic adaptation in subway ticketing processing, improving efficiency and recognition accuracy.

Description

Technical Field

[0001] This application relates to the technical field of intelligent ticketing processing for urban rail transit, and in particular to a subway BOM ticketing processing method and system based on artificial intelligence. Background Technology

[0002] With the continuous expansion of urban rail transit networks, the role of subway BOMs (semi-automatic ticket vending / ticket replenishment machines) in station ticketing processing, abnormal ticket collection, and passenger affairs handling is becoming increasingly prominent. Faced with increasing passenger flow and diverse ticket types, balancing recognition accuracy and processing efficiency in high-concurrency scenarios has become a critical issue that subway operators urgently need to address. Traditional ticketing processing methods relying on manual visual inspection and input are no longer sufficient to meet modern operational needs. Therefore, developing an intelligent subway BOM ticketing processing method is of great significance for improving traffic efficiency and reducing operation and maintenance costs.

[0003] Currently, subway BOM systems primarily rely on electronic reading technologies, such as RFID and NFC, to read electronic data from the chip in the ticket. These technologies can quickly acquire key ticket data, meeting basic ticketing processing needs.

[0004] The above-mentioned processing logic is relatively fixed. When encountering scenarios such as damaged electronic chips on tickets, abnormal data, peak passenger flow, or diversified ticket types, the over-reliance on a single electronic data dimension and fixed process often leads to a decrease in the accuracy of information verification or a limitation in processing efficiency. There is still room for improvement in overall adaptability. Summary of the Invention

[0005] To achieve dynamic adaptation in subway ticketing processing and improve efficiency and recognition accuracy, this application provides a subway BOM ticketing processing method and system based on artificial intelligence.

[0006] Firstly, this application provides a subway BOM ticketing processing method based on artificial intelligence, employing the following technical solution:

[0007] An AI-based subway BOM ticketing processing method includes:

[0008] Collect ticket images and electronic recognition data, and switch the collection mode according to the passenger flow density threshold;

[0009] Visual features are extracted from ticket images using object detection algorithms, and electronic features are obtained by parsing electronic identification data.

[0010] Simultaneously monitor the quality of visual features and electronic features, and filter features whose quality is below a preset basic threshold.

[0011] Perform cross-validation on the quality-compliant features to confirm their consistency.

[0012] If the verification passes, a second verification is performed using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement technology to improve recognition accuracy, and local discount rules are associated.

[0013] If the verification fails or the feature quality is substandard, user facial images are collected to establish an emotional baseline, and abnormal indicators of substandard visual and electronic features are extracted. The abnormal indicators and the difference records of the aforementioned cross-validation are input into the cross-modal abnormal causal association model to calculate the root cause confidence, thereby generating an abnormal feature vector with root cause labels. Specific abnormal types are identified through clustering. Based on the abnormal type and collection mode, a Bayesian optimization parameter tuning process is triggered with the comprehensive performance of the emotional baseline and the feature extraction and verification process as the dual optimization objectives. This process constructs an emotion-performance joint objective function and iteratively solves it to obtain the Pareto optimal solution, thereby determining the optimized system parameters. The system parameters include processing parameters associated with the feature extraction and verification process. Based on the optimized system parameters and the micro-expression recognition results corresponding to the emotional baseline, adaptation guidance information is output.

[0014] If the retest after adjustment still fails to meet the standards, a diagnostic report will be generated based on the aforementioned multimodal knowledge graph and abnormal feature vectors with root cause labels, and manual verification or maintenance instructions will be pushed to the system.

[0015] If the identification is successful, intelligent ticketing routing based on reinforcement learning will be executed and the operation result will be displayed; if it fails, the level will be determined by a preset risk assessment model and an appropriate handling plan will be executed.

[0016] By adopting the above technical solutions, accurate, efficient, and intelligent processing of subway ticketing has been achieved. High recognition accuracy is ensured through multimodal feature fusion and cross-validation; the system's processing efficiency and robustness are improved by combining a passenger flow density-adaptive data collection mode and parameter optimization; user emotion perception and root cause analysis are introduced to achieve personalized and intelligent anomaly handling; and finally, a complete closed loop of decision-making-execution-optimization is formed through reinforcement learning-based intelligent routing and risk assessment, significantly improving operational efficiency and passenger experience.

[0017] Secondly, this application provides a subway BOM ticketing processing system based on artificial intelligence, which adopts the following technical solution:

[0018] An AI-based subway BOM ticketing processing system includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is loaded and executed by the processor, it implements the AI-based subway BOM ticketing processing method as described in the first aspect. Attached Figure Description

[0019] Figure 1This is a flowchart of a subway BOM ticketing processing method based on artificial intelligence, according to an embodiment of this application.

[0020] Figure 2 This is a graph of the F1-score training of the verification set in the embodiments of this application. Detailed Implementation

[0021] Reference Figure 1 This application discloses an artificial intelligence-based subway BOM ticketing processing method, including the following steps:

[0022] Step S1: Collect ticket images and electronic identification data, switching the collection mode according to the passenger flow density threshold. The ticket image refers to a visible light image captured by the industrial CMOS sensor built into the BOM machine, including the ticket's QR code, barcode, and physical damage features. The electronic identification data refers to the information embedded in the ticket chip read by the NFC radio frequency module, including UID, balance, validity period, and key status. Dual-modal acquisition is achieved through hardware triggering and polling in parallel, synchronized by protocol encapsulation and timestamp addition, and written to a buffer queue for processing.

[0023] This application constructs a ticket information acquisition system covering both visual and electronic dimensions through the aforementioned dual-modal synchronous acquisition mechanism. More importantly, through the adaptive switching of the acquisition mode driven by the passenger flow density prediction model in subsequent steps S11 to S15, dynamic collaborative configuration of algorithm decision-making (passenger flow prediction) and hardware resources (image sensor, RF module operating parameters) is achieved, laying the foundation for a dynamic balance between overall system efficiency and accuracy from the data acquisition source.

[0024] Step S2: Visual features are extracted from the ticket image using an object detection algorithm, and electronic features are obtained by parsing the electronic identification data.

[0025] Visual features refer to the multi-dimensional representations extracted from the ticket image, including the location of the QR code area, ticket type identification, and the location and morphological information of the physical damage area. These are extracted from the ticket image cached in step S1 using an object detection algorithm. Electronic features refer to the structured fields parsed from the electronic identification data, including UID, balance, validity period, key authentication status, and transaction records. These are decoded from the electronic identification data cached in step S1 using a protocol parsing engine.

[0026] The necessary process is as follows: Visual feature extraction employs an improved YOLO object detection model, introducing an attention mechanism into the backbone network to enhance the perception of damaged QR codes and defective regions. The model is trained on a self-built ticket dataset through transfer learning, combining localization loss and classification loss for optimization. During the inference phase, it outputs detection boxes, class confidence scores, and damage region masks, extracts the QR code ROI, and labels the physical damage location and severity. The structural adjustments (introduction of the attention mechanism) and training strategy (transfer learning combined with a self-built subway ticket dataset) of this improved YOLO model are targeted optimizations for specific technical challenges in subway ticketing scenarios, such as image blurring caused by rapid ticket passage, physical wear caused by frequent use, and variable station lighting conditions. Key training parameters of the model include: using a stochastic gradient descent optimizer (initial learning rate 0.001, momentum 0.9), using a composite loss function combining GIoU localization loss and Focal classification loss (weight ratio 3:1), and training for 200 epochs on a dedicated dataset containing 100,000 labeled images. These designs ensure the model's high accuracy and robustness in the complex environment of the subway BOM.

[0027] Electronic signature analysis is achieved through a hierarchical state machine: the physical layer performs encoding and decoding and frame synchronization, the link layer verifies the checksum and frame integrity, and the application layer parses the fields in TLV format and verifies the card's legitimacy through a key algorithm.

[0028] The bimodal features are written into shared memory in a structured form, with timestamps and quality scores added, providing input for quality monitoring in step S3 and cross-validation in step S4.

[0029] Step S3: Simultaneously monitor the quality of visual features and electronic features, and filter features whose quality is below a preset basic threshold. For features that meet the quality standard, proceed to step S4 to perform cross-validation; for features whose quality is below the preset basic threshold, directly proceed to step S6 to initiate the anomaly adaptive processing flow.

[0030] Visual feature quality refers to the credibility measure of visual features, including QR code clarity score, detection box confidence, image exposure uniformity, and degree of physical damage interference. The visual features extracted in step S2 are quantified using a quality assessment algorithm. Electronic feature quality refers to the credibility measure of electronic features, including signal strength, verification pass rate, data integrity, and key authentication status. The electronic features parsed in step S2 are quantified using a quality assessment algorithm.

[0031] Preset basic threshold: refers to the minimum acceptable quality standard set by the system. The visual quality threshold and the electronic quality threshold can be configured independently. Features that do not meet the standard will be filtered and trigger the abnormal handling process in step S6.

[0032] The necessary process is described as follows: A four-dimensional evaluation system is constructed for visual feature quality monitoring. QR code clarity is calculated using the Laplacian operator to determine the variance; a variance below 100 indicates blurriness and requires resampling. Detection box confidence is directly taken from the YOLO model output, with a threshold set to 0.85. Exposure uniformity is evaluated using the standard deviation of the grayscale histogram; a standard deviation greater than 50 indicates uneven illumination. Physical damage interference is determined by the proportion of damaged mask area output in step S2; a proportion exceeding 30% indicates severe interference. The four dimensions are weighted and fused according to a 3:3:2:2 weighting to form a comprehensive quality score; a score below 60 is marked as substandard.

[0033] Electronic signature quality monitoring employs a two-layer evaluation system. The signal layer assesses communication stability using RSSI (Receiving Signal Strength Indicator) and SNR (Signal-to-Noise Ratio). An RSSI below -70dBm or an SNR below 10dB is considered a weak signal. The data layer verifies the CRC32 checksum, TLV field integrity, DES key authentication status, and validity period. Failure to perform any of these checks results in a non-compliance assessment.

[0034] Dual-modal quality monitoring is performed simultaneously. Features meeting both criteria are sent to step S4 for cross-validation; features failing to meet criteria are recorded as specific failure dimensions, such as "visually blurred - electronically normal" or "visually normal - weak electronic signal," triggering step S6 for adaptive optimization. For example, in high-density scenarios during morning rush hour, QR code blurring occurs frequently due to exposure time compression. The system records this abnormal pattern and initiates Bayesian optimization parameter tuning.

[0035] Step S4: Perform cross-validation on the features that meet the quality standards to confirm their consistency. Refer to steps S41 to S45 for details. If the validation passes, proceed to step S5 for secondary validation of the knowledge graph; if the validation fails, proceed to step S6 to initiate the anomaly adaptive handling process.

[0036] Step S5: If the verification passes, a second verification is performed using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement technology to improve recognition accuracy, and local discount rules are associated. For details, please refer to steps S51 to S56.

[0037] Step S6: If the verification fails or the feature quality is substandard, collect user facial images to establish an emotional baseline and extract abnormal indicators of substandard visual and electronic features; input the abnormal indicators and the aforementioned cross-validation difference records into the cross-modal abnormal causal association model, calculate the root cause confidence, and then generate an abnormal feature vector with root cause labels; identify specific abnormal types through clustering, and based on the abnormal type and collection mode, trigger a Bayesian optimization parameter tuning process with the comprehensive performance of the emotional baseline and the feature extraction and verification process as the dual optimization objectives. This process constructs an emotion-performance joint objective function and iteratively solves it to obtain the Pareto optimal solution, thereby determining the optimized system parameters, which include processing parameters associated with the feature extraction and verification process; based on the optimized system parameters and combined with the micro-expression recognition results corresponding to the emotional baseline, output adaptation guidance information.

[0038] The emotional baseline refers to the reference value of the current emotional state established through user facial image analysis, including dimensions such as anxiety, confusion, and calmness. It is extracted from user facial images captured by a camera using a micro-expression recognition model. Anomaly indicators refer to the specific quantitative representations of visual and electronic features failing to meet standards, including types such as insufficient clarity, signal attenuation, and verification failure. They are extracted from the quality monitoring results of step S3. The cross-modal anomaly causal association model refers to a deep learning model that integrates visual and electronic anomaly features to mine root cause relationships. It is constructed through training on multimodal data, inputting anomaly indicators and cross-validation differences, and outputting root cause confidence. Anomaly feature vectors refer to anomaly representations with root cause labels, used for clustering to identify specific anomaly types. They are generated by the causal association model. Bayesian optimization parameter tuning refers to a method that iteratively solves for Pareto optimal parameter configuration with emotion and performance as dual objectives. By constructing a joint objective function, system parameters are adaptively adjusted.

[0039] The necessary process is as follows: If the verification in step S4 fails or the feature quality in step S3 is substandard, the system activates the camera to capture the user's facial image and establishes an emotion benchmark through an improved and efficient network micro-expression recognition model. This model is pre-trained on a publicly available expression dataset and fine-tuned on a self-built subway scene dataset (containing 5000 images of crowded, backlit, and multi-angle scenes). It outputs the probability distribution of seven emotion categories, including anxiety, confusion, and calmness. The sum of the anxiety and confusion dimensions is taken as the emotional stress index (range 0-2), and an index exceeding 1.2 is considered a high-stress state. Simultaneously, abnormal indicators of substandard features are extracted, including the Laplacian variance, detection confidence, and damage area ratio in the visual dimension, and the signal strength, check code, and authentication status code in the electronic dimension. The cross-validation difference records from step S4 are integrated to construct a visual-electronic difference matrix. The matrix elements include the matching markers of the QR code parsing UID and the electronic UID, the timestamp deviation in milliseconds, and the balance logical consistency marker.

[0040] Anomaly indicators and a difference matrix are input into a cross-modal anomaly causal association model. The encoder of this model employs a 6-layer Transformer structure, with each layer containing an 8-head self-attention mechanism, an embedding dimension of 256, and a feedforward network dimension of 1024. It learns the cross-modal association between visual and electronic anomalies through masked multi-head attention. The model is trained on 100,000 historical anomaly samples (including labeled root causes), using InfoNCE contrastive loss to optimize the distance between positive and negative samples. It outputs a confidence distribution of five root causes (abnormal lighting, radio frequency interference, equipment aging, operational errors, and ticket damage). The root cause with the highest confidence is selected as the label, generating a labeled anomaly feature vector (128 dimensions).

[0041] The model's 6-layer Transformer encoder structure (8 heads per layer, 256 embedding dimensions) provides powerful cross-modal feature fusion capabilities. Its training framework employs a semi-supervised learning strategy, optimizing on 100,000 historical anomaly samples using the AdamW optimizer (learning rate 1e-4) and InfoNCE contrastive loss. Notably, the causal graph structure connected in the model's decoding part (containing 15 nodes across three categories: visual, electronic, and environmental) was defined by subway equipment maintenance experts based on historical fault reports and maintenance manuals, ensuring the domain rationality and interpretability of root cause reasoning.

[0042] The density peak clustering algorithm is used to cluster abnormal feature vectors, automatically determining the number of cluster centers and identifying specific anomaly types (such as "visual failure caused by strong light" or "electronic failure caused by multi-card collision"). Based on the anomaly type and the acquisition mode of step S1 (low density, medium density, high density), the Bayesian optimization parameter tuning process is triggered with the negative value of the emotional stress index and the feature extraction accuracy as dual optimization objectives.

[0043] A Gaussian process surrogate model is constructed as the joint objective function for emotion and performance. The kernel function adopts a combination of radial basis functions and polynomial kernels. The acquisition mode and anomaly type are used as context variables (categorical variables are encoded as one-hot vectors). The acquisition function is defined as a weighted combination of expected improvement and upper confidence bound. The Pareto front of system parameters such as exposure time (0.5ms-5ms), polling frequency (50ms-1000ms), infrared illumination power (0W-3W), and region of interest cropping ratio (50%-100%) is iteratively solved. The solution with the largest reduction in emotion stress index and the largest improvement in accuracy is selected as the optimal configuration.

[0044] Based on the optimized system parameters, the acquisition strategy is updated, and combined with the micro-expression recognition results corresponding to the emotion benchmark, adaptive guidance information is output. For example, when a scene of "high density + strong light + user anxiety" is identified, the optimization results increase the exposure time from 0.8ms to 2ms, enable full-power infrared illumination, switch to full-frame acquisition mode, and simplify the interface prompt to "Please place the ticket flat" with voice guidance to reduce the user's operational burden.

[0045] Step S7: If the retest after adjustment still fails to meet the standard, a diagnostic report is generated based on the aforementioned multimodal knowledge graph and abnormal feature vectors with root cause labels, and manual verification or maintenance instructions are pushed to the system.

[0046] The necessary process is as follows: If the retest after optimization in step S6 still fails to meet the standards, the system initiates a deep diagnostic process. Based on the multimodal knowledge graph constructed in step S5, the abnormal feature vectors with root cause labels generated in step S6 are integrated, and a diagnostic report is generated through graph neural networks for reasoning.

[0047] The multimodal knowledge graph adopts an entity-relation-attribute triple structure. The entity layer includes ticket entities (single-journey tickets, stored-value tickets, etc.), equipment entities (card readers, cameras, etc.), fault entities (QR code wear, RF module aging, etc.), and handling entities (manual verification, module replacement, etc.). The relation layer defines semantic associations such as "cause," "applicable to," and "located in." The attribute layer stores entity feature parameters and historical occurrence frequencies. The graph is initialized using domain expert rules and trained and optimized using the TransE algorithm with 100,000 sets of historical fault cases, achieving an entity embedding dimension of 256.

[0048] The graph neural network employs a two-layer GraphSAGE architecture, using mean pooling as the aggregation function. Each layer samples 25 neighbor nodes and fuses anomaly feature vectors with knowledge graph neighbor node information through a message passing mechanism. The input is a graph subgraph corresponding to the root cause label (containing faulty entities and associated equipment, tickets, and environmental entities). The output node embedding is mapped through a fully connected layer to generate fault location (equipment level / ticket level / environment level), impact assessment (individual / batch / system), and handling priority score.

[0049] The core of this diagnostic system lies in the deep integration of graph neural networks and domain knowledge graphs. The knowledge graph is trained on 100,000 historical cases using the TransE algorithm, learning 256-dimensional embeddings of entities and relationships. The GraphSAGE model (two layers, mean pooling, 25 neighbors) utilizes these embeddings for message passing and aggregation, enabling graph-based reasoning from abnormal symptoms to the root cause of the fault. This approach, combining data-driven (GNN) and knowledge-driven (graph) methods, is particularly suitable for complex scenarios like subway equipment fault diagnosis, which require combining real-time data with historical experience.

[0050] The diagnostic report is generated based on the reasoning results: if the problem is determined to be individual ticket damage (such as severe creases causing QR code failure), manual verification instructions are sent to guide the user to the service desk to exchange tickets; if the problem is determined to be a device module failure (such as aging RF antenna causing card reading failure), a repair work order is automatically generated and sent to the maintenance system, including fault location, recommended replacement parts, and historical records of similar fault handling. For example, when three consecutive tickets show "electronic feature quality not up to standard" on the same BOM machine and the root cause is "RF interference," the system infers that the device's RF module is abnormal, triggers repair instructions, and marks the device as pending repair.

[0051] In step S8, if the final identification is successful, intelligent ticketing routing based on reinforcement learning will be executed and the operation result will be displayed; if it fails, the level will be determined by a preset risk assessment model and an adaptation processing plan will be executed.

[0052] Among them, intelligent ticketing routing refers to a decision-making mechanism that dynamically selects the optimal business processing path based on real-time status, including strategies such as window allocation, channel selection, and resource scheduling. It learns the optimal strategy from historical interaction data through a reinforcement learning model. Risk assessment model refers to an assessment framework that comprehensively quantifies and identifies the level of failure risk based on multiple dimensions, including indicators such as failure frequency, anomaly type, user sentiment, and device status. It is built through a rule engine and statistical models. Adaptive handling solutions refer to differentiated handling strategies matched according to the risk level, including measures such as service degradation, manual intervention, and security control.

[0053] The necessary process is as follows: If the identification is successful after optimization in step S5 or S6, the system executes intelligent ticketing routing based on reinforcement learning. This model employs a near-end strategy optimization algorithm. The state space includes eight continuous variables such as current passenger flow density, window queue length, equipment load rate, and user emotional stress index. The action space is defined as five types of routing decisions (fast track, regular window, self-service terminal, manual assistance, and reservation service). The reward function is designed by integrating waiting time, user satisfaction, and resource utilization. The model is trained for 1 million steps on simulation environment and real historical data to learn the optimal state-action mapping. For example, for anxious users successfully identified during the morning rush hour, the system prioritizes assigning them to the fast track and preloads frequently used destinations to shorten the interaction steps.

[0054] If the identification ultimately fails, the system activates a risk assessment model. This model constructs a three-tiered assessment system: the primary indicator is failure frequency (number of consecutive failures, historical failure rate); the secondary indicator is the severity of the anomaly (root cause type, scope of impact); and the tertiary indicator is contextual factors (user emotional stress, device health, time-sensitive factors). Each indicator is weighted using the entropy weight method to output a risk level (low / medium / high / urgent). Low risk triggers a downgrade plan (e.g., manual ticket number input); medium risk triggers manual assistance (remote video customer service); high risk initiates security control (temporarily locking the transaction and recording audit logs); and emergency risk prompts on-site security intervention. For example, if identification fails and the emotional stress index is 1.8, there are 3 consecutive failures, and the root cause is "unknown ticket type," it is judged as medium to high risk, triggering manual window guidance and simultaneously reporting to the duty station manager.

[0055] Switching the data collection mode based on passenger flow density threshold includes:

[0056] Step S11: Based on real-time passenger flow density and combined with historical data reflecting periodic statistical patterns and external events reflecting sudden impacts, a corrected density threshold is dynamically generated through the passenger flow prediction model.

[0057] Historical data refers to periodic passenger flow statistics stored in a time-series database, including hourly, daily, weekly, and yearly passenger flow distribution patterns, which are periodically extracted from the data warehouse. External events refer to sudden factors affecting passenger flow, including weather conditions, holiday arrangements, large-scale events in the surrounding area, and subway line malfunctions, which are obtained in real time through meteorological interfaces, publicly available government data, and social media monitoring. Passenger flow prediction model refers to a deep learning model that integrates time-series patterns and external factors to generate passenger flow predictions for future periods. Corrected density thresholds refer to dynamically adjusted three-level density boundary values ​​based on prediction results, replacing fixed thresholds to adapt to complex scenarios.

[0058] The necessary procedures are as follows:

[0059] Based on the passenger flow prediction model, a revised density threshold is dynamically generated. The model input includes three types of features: real-time features (current density, queue length, and inbound flow rate), time-series features (historical passenger flow during the same period, weekly cycle index, and holiday markers), and external features (rainfall level, temperature range, distance to large-scale events, and line delay status).

[0060] The model employs an encoder-decoder architecture. The encoder layer includes a one-dimensional convolutional module to extract local temporal patterns, a bidirectional long short-term memory network to capture long-range dependencies, and external features are mapped and concatenated with temporal features through an embedding layer. The decoder layer uses a self-attention mechanism to weightedly fuse multi-source information, outputting passenger flow predictions at three scales: 15 minutes, 30 minutes, and 60 minutes. The model is trained on historical data (2 years of station-level passenger flow records), with a loss function that is a weighted combination of mean squared error and quantile loss, an optimizer named Adam, a learning rate of 1e-3, and 200 training epochs.

[0061] The corrected density thresholds are dynamically calculated using predicted values ​​and confidence intervals. The base thresholds (5 people / minute for low density, 15 people / minute for medium density) are corrected for predicted passenger flow deviations: if a passenger flow surge of over 30% is predicted in the next 30 minutes, the threshold is lowered by 20% to trigger the high-efficiency mode earlier; if a sudden drop in passenger flow is predicted or external events cause obstruction at the station, the threshold is raised by 15% to extend the high-precision mode. For example, during a rainy morning rush hour, if the model predicts a 40% increase in passenger flow from 8:15 to 8:45 compared to the historical average, the corrected medium-density threshold is lowered from 15 people / minute to 12 people / minute, allowing the high-efficiency data collection mode to be activated earlier to avoid congestion.

[0062] Step S12: Compare the real-time monitored current passenger flow density with the corrected density threshold.

[0063] The necessary procedures are as follows:

[0064] Step S12 performs a comparison and determination between real-time density and dynamic threshold. The system reads the current passenger flow density at a 1-second interval and performs a double-boundary comparison with the corrected threshold output in step S11.

[0065] The comparison rule adopts a dual-threshold structure: let the corrected low-to-medium density threshold be T1, and the medium-to-high density threshold be T2. If the current density is less than T1, it is determined to be a low density level; if the current density is greater than T2, it is determined to be a high density level; if the current density is between T1 and T2, it is determined to be a medium density level. To avoid threshold boundary jitter, a hysteresis buffer mechanism is introduced: when the density crosses T1 from low to high, it must remain above T1 for 5 seconds to trigger a level change; when the density falls back to T1 from high to low, it must remain below T1 for 10 seconds to confirm a level change, and the same applies to T2.

[0066] For example, after step S11, T1=12 people / minute and T2=28 people / minute. The current real-time density is 30 people / minute and lasts for 6 seconds, which is judged as a high-density level, triggering the efficient collection mode in step S13. If the real-time density drops to 11 people / minute but only lasts for 3 seconds, the original level is maintained to avoid frequent switching.

[0067] Step S13: If the current passenger flow density is greater than the corrected density threshold, then the preset high-efficiency collection mode is activated.

[0068] The high-efficiency data collection mode refers to a rapid response data collection strategy for high-density passenger flow, which improves processing efficiency by optimizing data collection parameters and simplifying verification procedures. It is preset in the system configuration table and is activated by density comparison results.

[0069] The necessary process is as follows: Step S13: When the current passenger flow density is greater than the corrected density threshold, the efficient collection mode is enabled.

[0070] The visual acquisition layer employs a resolution compression and region cropping strategy, compressing the image acquisition from full frame to 1280×960 and shrinking the region of interest to 800×600 in the center; the exposure time is fixed at 0.5 milliseconds, and adaptive exposure calculation is turned off; the object detection model is switched to a lightweight variant, and the confidence threshold is lowered from 0.85 to 0.70.

[0071] The electronic acquisition layer adopts a high-frequency polling and fast interaction strategy, increases the radio frequency power to 1.5 watts, compresses the polling period to 50 milliseconds, reduces the number of handshakes from 3 to 2, and changes the data verification from dual CRC to single CRC.

[0072] The dual-modal fusion layer uses key fields for rapid comparison, only verifying the first 4 bytes of the UID and the ticket type, skipping full-field cross-validation and the knowledge graph secondary validation in step S5, and directly outputting the recognition result.

[0073] For example, in high-density scenarios during the morning rush hour, passengers continuously submit tickets at a rate of 1.2 seconds per person. After the system activates the high-efficiency mode, the processing time for a single ticket is reduced from 1200 milliseconds to 650 milliseconds, and the queue backlog length is reduced from 15 people to 5 people. While meeting the priority requirement of passage efficiency, the recognition accuracy is adjusted from 99.2% to 97.5%, which is still within an acceptable range.

[0074] Step S14: If the current passenger flow density is less than the corrected density threshold, then the preset high-precision acquisition mode is activated. The high-precision acquisition mode refers to a refined identification and acquisition strategy for low-density passenger flow, ensuring recognition accuracy by improving image quality and enhancing verification depth. It is preset in the system configuration table and is activated by the density comparison result.

[0075] The necessary process is as follows: Step S14: When the current passenger flow density is less than the corrected density threshold, the high-precision acquisition mode is enabled.

[0076] The visual acquisition layer adopts a full-frame high-resolution strategy, maintaining a full-frame output of 2592×1944 for image acquisition, with the region of interest covering the entire ticket surface; the exposure time uses an adaptive algorithm, dynamically adjusting to the optimal value according to the ambient light, and automatically turning on infrared supplementary light in low-light scenes; the target detection model uses the full version, with the confidence threshold set to 0.90 to improve detection accuracy, and simultaneously outputs a mask of the physical damage area for quality assessment.

[0077] The electronic acquisition layer adopts a robust communication strategy, maintains the RF power at a standard of 0.1 watts, sets the polling period to 200 milliseconds to reduce power consumption, performs a complete three-way handshake protocol, and enables dual CRC check and frame sequence number continuity check to ensure data integrity and reliability.

[0078] The dual-modal fusion layer employs full-field deep verification, performing complete cross-validation of visual and electronic features, including full-byte UID comparison, balance logic verification, and validity period consistency check. Upon passing these checks, it triggers a second verification using the multimodal knowledge graph in step S5, and then performs refined processing in conjunction with local discount rules.

[0079] For example, in low-density scenarios during off-peak hours, when passengers hand over their tickets at a rate of 5 seconds per person, the system takes about 1500 milliseconds to process a single ticket after enabling high-precision mode, with a recognition accuracy rate of 99.8%. It can accurately identify minor damage to tickets and fully verify eligibility for discounts, providing passengers with highly reliable ticketing services.

[0080] Step S15: If the current passenger flow density is equal to the corrected density threshold, enable the preset mixed transition mode.

[0081] The hybrid transition mode refers to a balanced acquisition strategy oriented towards density critical states. It achieves a smooth transition between timeliness and accuracy by dynamically fusing parameters from both high-efficiency and high-precision modes. It is preset in the system configuration table and is activated by density comparison results.

[0082] The necessary procedures are as follows:

[0083] Step S15 activates the hybrid transition mode when the current passenger flow density equals the corrected density threshold. This mode uses a parameter interpolation mechanism to take an intermediate configuration between the high-efficiency mode and the high-precision mode, avoiding system oscillations caused by mode jumps.

[0084] The visual acquisition layer uses a medium resolution of 1600×1200 and fast adaptive exposure. The confidence threshold of the target detection model is set to 0.80, which is between the efficient 0.70 and the high-precision 0.90.

[0085] The electronic acquisition layer adopts a variable frequency polling strategy with a radio frequency power of 0.8 watts and a polling period of 120 milliseconds, both of which are the midpoint between the two modes, and are finely adjusted in real time based on the success rate of the previous 3 transactions.

[0086] The dual-modal fusion layer adopts a hierarchical verification strategy, prioritizing the rapid comparison of key fields. If the confidence level is higher than 0.85, the data is output directly; otherwise, a full-field deep verification is triggered.

[0087] For example, in the critical scenario of transitioning from off-peak to peak hours, the passenger flow gradually increases from 10 people / minute to 15 people / minute. Directly switching modes can easily cause fluctuations in the collected parameters. After enabling the hybrid mode, the system smoothly transitions with a time of 1000 milliseconds and an accuracy of 98.5%. This not only responds promptly to the trend of passenger flow growth but also avoids recognition fluctuations caused by frequent mode switching, achieving a dynamic balance between timeliness and accuracy.

[0088] Perform cross-validation on the quality-compliant features to confirm consistency between the two, including:

[0089] Step S41: Using a preset feature association mapping library, preset key identifiers in visual features are associated with preset core fields in electronic features.

[0090] The system includes a pre-defined feature association mapping library, which stores the correspondence between key visual feature identifiers and core fields of electronic features. This library includes field name mappings, format conversion rules, and verification logic definitions. It is built using domain expert definitions and historical data training and is embedded in the system configuration. Pre-defined key identifiers are the core information extracted from visual features for identity recognition, including the ticket number, ticket type identifier, and printed validity period characters from QR code parsing. Pre-defined core fields are the basic information parsed from electronic features for identity verification, including the UID stored on the chip, ticket type code, and validity period timestamp.

[0091] The necessary procedures are as follows:

[0092] First, key identifiers are extracted from visual features: ticket number (e.g., hexadecimal string) is obtained through QR code decoding, printing validity period (e.g., year, month, and day characters) is obtained through optical character recognition, and type identifier (e.g., printed pattern code) is obtained through ticket layout analysis. Simultaneously, core fields are extracted from electronic features: UID binary stream, type code enumeration value, and timestamp integer value are obtained through protocol parsing.

[0093] Subsequently, the mapping library is invoked to perform format standardization and semantic association: the basic identification layer converts the visual number and electronic UID into a standardized hexadecimal format and performs hash verification to ensure unique matching; the type determination layer maps the visual pattern code and electronic type code to a unified ticket type semantic through a lookup table (e.g., mapping "01" to "one-way ticket"); the timeliness verification layer converts the printed characters and electronic timestamp into a standard calendar format and calculates the time deviation.

[0094] Finally, the output is a structured association pair, which includes a matching score for the identifier pair, type pair, and timeliness pair.

[0095] Step S42: Adjust the verification dimensions according to the current acquisition mode. In the high-efficiency acquisition mode, enable the preset core verification dimensions. In the high-precision acquisition mode, enable the preset full verification dimensions. In the hybrid transition mode, allocate the preset verification dimensions according to the resource ratio.

[0096] The verification dimensions refer to the set of checks specifically executed in cross-validation, including dimensions such as identifier consistency, type consistency, timeliness consistency, and balance logic. These are preset by the system configuration and selected by the current collection mode. The preset core verification dimensions refer to the minimum necessary checks enabled in the high-efficiency collection mode, including only rapid comparisons of key fields (such as unique identifier prefixes and ticket type codes). These are defined through the high-efficiency mode configuration table. The preset full verification dimensions refer to the complete set of checks enabled in the high-precision collection mode, including deep verification of all fields such as the full-byte unique identifier, type, balance, validity period, and transaction records. These are defined through the high-precision mode configuration table. Resource allocation refers to the scheduling strategy in the hybrid transition mode that dynamically determines the number of verification items based on the system's real-time load. This is calculated using the hybrid mode algorithm.

[0097] The necessary procedures are as follows:

[0098] If the current data collection mode is high-efficiency, the system enables preset core verification dimensions. Only lightweight comparisons of key fields are performed, such as verifying whether the first 4 bytes of the unique identifier match and whether the ticket type code is consistent, skipping balance verification, validity period detail comparison, and transaction record checks. This strategy minimizes verification time, meeting the rapid passage requirements in high-density scenarios. For example, during morning rush hour, the system only needs to confirm that the ticket number prefix matches the type to determine passage, with a single ticket verification taking approximately 50 milliseconds.

[0099] If the current data acquisition mode is high-precision, the system enables a preset full-scale verification dimension. It performs deep validation across all fields, including exact matching of unique identifiers, semantic consistency of ticket type, logical verification of balance values, validity period validity checks, and hash comparison of recent transaction records. This strategy ensures maximum recognition accuracy and is suitable for refined services in low-density scenarios. For example, during off-peak periods, the system fully verifies all ticket information and triggers a correction prompt if the visual balance display deviates from the electronic record by 0.5 yuan.

[0100] If the current system is in a hybrid transition mode, it will implement a resource allocation strategy. Based on the average processing time of the first five transactions and the current processing unit utilization, the system will dynamically calculate the number of verification items: when resources are plentiful, verify 80% of the dimensions (such as the full-byte unique identifier, type, and balance); when resources are limited, verify 50% of the dimensions (such as the full-byte unique identifier and type). Alternatively, a stratified sampling strategy can be used, alternating between full and core verification. For example, in a critical state, the system will prioritize ensuring accurate unique identifier matching, mandatory verification of the type field, and verification of the balance field every three transactions, achieving a smooth transition between timeliness and accuracy.

[0101] Step S43: Align the acquisition timing of the two types of features using a preset timestamp mechanism, and complete feature matching within a preset time window. The preset timestamp mechanism refers to a synchronization mechanism that adds acquisition time markers to visual and electronic features, employing a combination of hardware-triggered timestamps and network time protocol synchronization to ensure consistency of the dual-modal time base. This is achieved collaboratively by the acquisition layer hardware module and the system clock module in step S1. The acquisition timing of the two types of features refers to the time sequence relationship between the ticket image acquisition time and the electronic recognition data reading time, reflecting the order and time interval of the capture of physical ticket and chip data. This is extracted from the timestamp field of the data packet acquired in step S1. The preset time window refers to the allowable time deviation range for determining whether visual and electronic features belong to the same ticketing event, dynamically set according to the acquisition mode (high-efficiency mode ±300 milliseconds, high-precision mode ±800 milliseconds, mixed mode ±500 milliseconds). This is preset through the system timing configuration table.

[0102] The necessary procedures are as follows:

[0103] Step S43 aligns the timing of the two types of feature acquisitions using a preset timestamp mechanism. The system first extracts the visual feature timestamp (the moment the image sensor hardware triggers) and the electronic feature timestamp (the moment the RF module completes decoding), and calculates the absolute time deviation between the two. A precise time protocol is used to ensure that the dual-modal clock synchronization error is less than 1 millisecond, eliminating the impact of clock drift between devices.

[0104] Feature matching is performed within a preset time window: if the time deviation is within the allowable range, it is confirmed that the two features originated from the synchronous acquisition of the same ticket, and subsequent consistency verification is performed; if the deviation exceeds the window, it is judged as an asynchronous anomaly (such as the confusion between the previous passenger's ticket image and the subsequent passenger's electronic data), the feature pair is discarded and re-acquisition is triggered. For example, if the visual feature timestamp is 10:00:00:100 milliseconds and the electronic feature timestamp is 10:00:00:350 milliseconds, the deviation of 250 milliseconds is within the ±500 millisecond window of the mixed mode, and the matching process is allowed; if the electronic feature timestamp is 10:00:00:950 milliseconds and the deviation of 850 milliseconds exceeds the window, the system judges it as a time sequence misalignment, rejects the match, and records the anomaly.

[0105] Step S44: Determine consistency based on the preset matching threshold. If the matching degree meets the standard, it passes. If it is within the preset fault tolerance range, it is marked as weakly consistent and a supplementary verification is triggered. If it exceeds the range, it is determined to be inconsistent and the difference features are recorded.

[0106] The system includes several key features: **Preset Matching Thresholds:** These refer to the system's set matching degree grading standards, including a complete pass threshold (e.g., 95%), a weak consistency tolerance range (e.g., 85%-94%), and an inconsistency judgment line (e.g., below 85%). These are preset through the system configuration table and can be dynamically adjusted according to the acquisition mode. **Consistency:** This refers to the degree of matching between visual and electronic features in the associated fields, quantified by a matching degree score. It is obtained by calculating the matching ratio of the associated pairs in step S41 using a field comparison algorithm. **Weak Consistency:** This refers to a matching degree within the tolerance range, indicating that the features are basically consistent but have slight deviations (e.g., second-level differences in timestamps, inconsistencies in character capitalization). It is identified through threshold range judgment. **Supplementary Verification:** This refers to a secondary confirmation process triggered for weak consistency states, including calling a multimodal knowledge graph for verification or switching to a high-precision acquisition mode for re-acquisition. This is triggered by the rule engine. **Difference Features:** These refer to the specific field deviation information recorded when inconsistency is determined, including field name, visual value, electronic value, and deviation type. This is stored in a structured difference log.

[0107] The necessary procedures are as follows:

[0108] Step S44 classifies the verification results of step S42 according to a preset matching threshold. The system first calculates the comprehensive matching degree: the comparison results of each verification dimension in step S42 are weighted and averaged, with the identifier matching weight accounting for 60%, type matching accounting for 20%, and timeliness matching accounting for 20%, generating a matching degree score of 0-100%.

[0109] Then, a three-level threshold judgment is performed: if the matching degree is higher than the complete pass threshold (e.g., 95%), it is judged as consistent and directly proceeds to step S5 or step S8; if the matching degree is in the fault tolerance range (e.g., 85%-94%), it is marked as weakly consistent and triggers the supplementary verification process, such as calling the multimodal knowledge graph in step S5 to confirm whether the timeliness deviation is within the allowable range, or prompting the user to re-display a specific area of ​​the ticket; if the matching degree is lower than the fault tolerance lower limit (e.g., 85%), it is judged as inconsistent, the difference features are recorded (e.g., "UID mismatch: visual 0A3F9B / electronic 0B4E8C"), and the abnormal indicators are input into the cross-modal abnormal causal association model in step S6.

[0110] For example, in high-precision mode, the ticket identifier and type match perfectly (100%), but there is a 1-day discrepancy between the printed validity period "20251231" and the electronic timestamp "2025-12-30 23:59:59" (matching degree drops to 90%). The system marks this as weakly consistent and triggers the knowledge graph to verify whether the ticket has a record of validity period change. If the identifier does not match at all (matching degree 0%), then the system directly determines that there is no match and records the difference characteristics.

[0111] Step S45: Output the verification result.

[0112] The recognition accuracy is enhanced through secondary verification using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement technology, and local discount rules are associated with:

[0113] Step S51: Based on the features that have passed cross-validation, call the appropriate graph resources according to the current acquisition mode. In the high-efficiency acquisition mode, load the preset core sub-graph and generate a fast verification result through the rule engine. In the high-precision acquisition mode, enable the full graph and generate a deep verification result based on the graph neural network. The two together constitute the basic verification result.

[0114] The multimodal knowledge graph comprises the following components: Multimodal Knowledge Graph: This refers to a semantic network integrating ticketing domain knowledge, equipment status, historical cases, and preferential policies, containing ticket entities, transaction entities, equipment entities, rule entities, and their relationships. It is constructed through domain expert annotation and historical data extraction. Adaptive Knowledge Graph Resources: This refers to dynamically loaded subsets of the knowledge graph based on the collection mode, including a core sub-graph (simplified key entities and relationships) and a full graph (complete entity-relationship network). It is obtained through hierarchical storage of the knowledge graph and pattern configuration mapping. Pre-defined Core Sub-graph: This refers to a lightweight knowledge base built for efficient collection modes, containing only core nodes such as ticket types, basic transaction rules, and common anomaly patterns. It is pre-generated using a graph pruning algorithm. Rule Engine: This refers to a decision-making component that performs rapid reasoning based on predefined logical rules, supporting conditional judgment and result generation. It is implemented through a rule base and an inference engine. Graph Neural Network: This refers to a deep learning model that performs message passing and node embedding on the full knowledge graph for complex relationship reasoning. It is constructed through training on historical data. Basic verification results: These refer to the preliminary conclusions generated after verification through the core sub-graph or the full graph, including determinations of ticket legality and transaction compliance. These results are generated and then merged through the two paths described above.

[0115] The necessary procedures are as follows:

[0116] If the current acquisition mode is high-efficiency, the system loads a preset core subgraph. This subgraph is extracted from the full graph using a pruning algorithm, retaining only ticket type nodes (single-journey tickets, stored-value tickets, etc.), basic transaction rule nodes (balance greater than zero, valid validity period), and direct relationships, with the node size controlled within 20% of the full graph. The verification process relies on a rule engine: the association pairs in step S41 are mapped to fact assertions (e.g., "ticket type = stored-value ticket", "balance = 50 yuan"), and pattern matching is performed using predefined production rules (e.g., "if the type is stored-value ticket and the balance is greater than zero and the validity period is valid, then it is considered valid") to quickly generate verification results. The processing time for a single ticket is approximately 30 milliseconds.

[0117] If the current data acquisition mode is high-precision, the system activates a full graph, including heterogeneous entities and multi-hop relationships such as tickets, users, devices, transaction history, and preferential policies (e.g., "user-holder-ticket-applicable-preferential rules"). The verification process relies on a graph neural network: a two-layer graph attention network structure is adopted, with an 8-head attention mechanism in each layer and a node embedding dimension of 256. Neighbor features are aggregated through message passing to infer potential associations (e.g., "this ticket is a senior citizen discount card, and the remaining number of discounts for today is determined based on historical transaction records"). The processing time for a single ticket is approximately 200 milliseconds.

[0118] The verification results generated by the two paths are fused using a confidence-weighted method (0.6 for the high-efficiency mode and 0.4 for the high-precision mode, dynamically adjusted) to form the basic verification result. This outputs a ticket legality determination, risk score, and discount eligibility marker, providing input for the spatiotemporal optimization in step S52. For example, the high-efficiency mode quickly determines "ticket legal, confidence 0.85," while the high-precision mode uses deep reasoning to determine "senior discount eligibility is valid, used twice today, 1 discount remaining." The fusion of these results forms the complete basic verification result.

[0119] Step S52: Based on the basic verification results, integrate real-time passenger flow, station and time segment features, dynamically adjust the graph node weights through an attention mechanism, and output the optimized verification results.

[0120] The system includes the following features: **Site Characteristics:** Station attribute information, including station type (hub station, transfer station, regular station), historical equipment failure rate, and special policy identifiers (e.g., airport line, tourist line). This information is obtained through the station configuration database. **Time Period Characteristics:** Time-dimensional attributes, including operating hours (morning peak, off-peak, evening peak) and date type (weekday, weekend, holiday). This information is obtained through the system clock and operating calendar. **Attention Mechanism:** A dynamic weight calculation component based on a graph attention network, which calculates the importance coefficient of each node in the graph based on contextual features. It is built using historical transaction data and employs a multi-head attention structure. **Graph Node Weights:** The influence weight of each entity in the knowledge graph (e.g., transaction rules, abnormal cases, preferential policies) on the current verification decision, dynamically adjusted through the attention mechanism.

[0121] The necessary process is as follows: Step S52 is based on the basic verification results of step S51, integrates real-time context features, dynamically adjusts the weights of graph nodes through an attention mechanism, and outputs the optimized verification results.

[0122] First, feature fusion is performed. The basic verification results (ticket validity score, discount eligibility mark, initial confidence level) output in step S51 are concatenated with the real-time context (current passenger flow density level, station type code, time period mark) and mapped into a joint feature vector of a unified dimension through the embedding layer.

[0123] Secondly, dynamic attention weights are calculated. A graph attention layer (using an 8-head attention mechanism) is constructed, using the joint feature vector as the query and the features of relevant nodes in the knowledge graph (such as "senior discount rules", "morning peak fast passage strategy", and "device fault tolerance mode") as the key values ​​to calculate the attention score. For example, in a morning peak high-traffic scenario, the attention score of the "fast passage rules" node increases from the default 0.3 to 0.8, while the score of the "detailed audit rules" node decreases from 0.7 to 0.2; if the current site has a high historical failure rate, the weight of the "fault tolerance verification rules" node is increased accordingly.

[0124] Then, the verification conclusions are weighted and revised. The graph node information is weighted and aggregated according to attention weights to adjust the confidence level and risk assessment of the basic verification results. For example, if the basic verification confidence level is 0.85, it is increased to 0.92 after weighted optimization at regular off-peak sites, passing directly; at sites with high failure rates, it is reduced to 0.75 after weighted revision, triggering the strict review in step S53; at tourist route sites, the high weight of the "special discount rules" node may revise the discount eligibility assessment.

[0125] Finally, the optimized verification results are output, including the adjusted confidence score, risk level label, and applicable rule version, as input for step S53, spatiotemporal anomaly enhancement.

[0126] Step S53: Based on the optimized verification results, introduce spatiotemporal anomaly enhancement technology: use the isolated forest algorithm to screen for abnormal features that deviate from the historical transaction baseline, and simultaneously use the LSTM model to learn the same site and time period within the previous preset number of days to generate a dynamic spatiotemporal normal interval.

[0127] The spatiotemporal anomaly enhancement technology combines spatial (site attributes) and temporal (time period patterns) anomaly detection methods to identify features deviating from normal patterns through historical baseline comparison and time-series prediction. This is achieved through a collaborative approach using the Isolation Forest algorithm and Long Short-Term Memory (LSTM) networks. Historical transaction baselines refer to the distribution of normal transaction characteristics at a specific site during a specific time period, including statistics such as transaction frequency, amount range, and ticket type proportion. This is obtained through historical data aggregation and is updated daily. LTM networks are recurrent neural networks with gating mechanisms used to learn long-term dependencies in time-series data. They are constructed by training on transaction sequences from the same site and time period within a preset number of days (e.g., 30 days). The network structure contains two hidden layers, each with 128 memory units. The same site and time period within the preset number of days refers to historical reference data from the same period. For example, if the current time is 8:00 AM on a weekday, transaction records from the same site between 7:00 AM and 9:00 AM within the previous 30 weekdays are extracted as training samples. Dynamic spatiotemporal normal range refers to the range of normal transaction characteristics generated through real-time learning, with boundary thresholds dynamically adjusted over time. It is formed by predicting outputs from the Long Short-Term Memory Network and then correcting them using an isolated forest.

[0128] The necessary process is as follows: First, anomalies deviating from historical transaction baselines are screened using the Isolation Forest algorithm. The algorithm constructs 128 isolation trees based on historical transaction data, with 256 samples randomly selected from each tree for recursive spatial partitioning. The transaction characteristics (type, amount, time period, frequency) of the current ticket are input into the forest, and the average isolation path length is calculated. The shorter the path, the higher the anomaly score. For example, if a ticket exhibits the characteristic of "large-amount abnormal recharge" during the morning rush hour, its path length is significantly shorter than the historical baseline, resulting in an anomaly score of 0.85, triggering a suspicious flag.

[0129] Simultaneously, a Long Short-Term Memory (LSTM) network is used to learn normal transaction characteristics from the same station and time period over the previous 30 days. The network input is a time-series vector (hourly transaction mean, variance, and ticket type distribution). Input gates, forget gates, and output gates are used to capture periodic patterns (such as weekday morning rush hour commuting patterns), and the output is a predicted feature for the next hour and its confidence interval. The LSTM model used here has a two-layer structure with 128 memory units per layer, specifically designed to learn the spatiotemporal patterns of subway transactions. Its training data consists of historical transaction sequences from the same station and time period over the previous 30 days. This allows the model to capture the unique commuting tidal patterns, weekend leisure travel patterns, and passenger flow characteristics of specific stations (such as airport lines and tourist lines) in the subway scenario. The dynamic normal interval output by the model provides a domain-adaptive judgment benchmark for identifying abnormal transactions that deviate from the norm.

[0130] The results of two algorithms are combined to generate a dynamic spatiotemporal normal range: the Long Short-Term Memory network provides a baseline range for time series prediction (e.g., normal transaction amount during morning rush hour is 5 to 50 yuan), and the Isolation Forest provides a corrected boundary based on spatial distribution (e.g., the tightened range after removing historical outliers). If the current feature falls within the range, it is judged as normal; if it exceeds the range, it is marked as spatiotemporal anomaly, and the process proceeds to step S54 to calculate the deviation. For example, at 8:00 AM on a weekday, the normal range for a certain station is dynamically set as "60% to 80% of one-way tickets, with an average consumption of 4 to 6 yuan per person." Transaction amounts exceeding this range, such as "full-price tickets" or "ultra-low-price transactions," will be captured.

[0131] Step S54: Calculate the deviation between the abnormal feature and the normal range. If the deviation exceeds the threshold, call the preceding quality compliance feature to construct a double verification chain to complete the abnormality review.

[0132] Among them, the preceding quality compliance features refer to the historical feature data that passed the quality monitoring in step S3 and is cached, including recently successfully identified ticket records and transaction logs. This is obtained through a circular buffer or a short-term historical database. The dual verification chain refers to a composite verification mechanism that combines current abnormal features with historical compliance features for cross-validation, confirming data consistency by constructing a feature association chain. This is implemented through the verification chain construction module. Anomaly review refers to the process of secondary confirmation of features marked as abnormal; if determined to be normal, it is allowed to proceed; otherwise, it proceeds to step S55. This is executed through the review decision engine.

[0133] The necessary process is as follows: First, calculate the multidimensional deviation. The system uses Mahalanobis distance to calculate the distance between the current feature vector and the center of the normal interval, comprehensively considering the variance and correlation of each dimension to generate a dimensionless deviation score. For example, a deviation threshold of 2.0 is set. A score below 2.0 is considered mild deviation, 2.0 to 3.0 is considered moderate deviation, and a score above 3.0 is considered severe deviation.

[0134] For abnormal features exceeding the threshold (e.g., deviation > 2.0), the system calls the preceding quality-compliant features to construct a double verification chain. It extracts the three most recent successfully identified records of the ticket from the cache queue (e.g., preceding entry station, balance, timestamp) to construct a temporal chain. It verifies the logical consistency between the current abnormal feature and the preceding records: checking UID continuity to confirm ticket identity, verifying whether balance changes conform to transaction logic (e.g., entry fee rules), and checking spatiotemporal rationality (e.g., physical feasibility across stations within a short period). For example, if the current feature shows "balance suddenly decreased by 50 yuan" (deviation 2.5, exceeding the threshold), and the preceding record shows "entered this station 5 minutes ago, balance sufficient," double verification confirms it's a normal fee deduction (50 yuan deduction upon exit), the deviation weight is reduced, and the verification passes. If the preceding record shows "balance is zero" or "previous transaction at a distant station," the verification chain breaks, confirming an anomaly, and proceeding to step S55 to trigger facial recognition and anomaly handling.

[0135] If the review fails, skip the associated local discount rules and trigger the exception handling process to collect the user's facial image.

[0136] Step S56: If the review is successful, the local discount rules are associated according to the data collection mode. Local discount rules refer to the ticketing discount policies implemented by the subway operator at specific stations, including the definition and usage restrictions (such as maximum number of uses and validity period) for rules such as senior citizen card discounts, student card discounts, transfer fee reductions, and time-based discounts. These rules are stored in a local rule database or knowledge graph and dynamically loaded according to station codes and time-based markers.

[0137] The necessary process is as follows: In high-efficiency data collection mode, the system calls the preset core discount rule library. Only high-frequency and simple rules (such as fixed discount rates) are loaded, and a lightweight rule engine (based on the Rete algorithm) quickly matches the ticket type and discount strategy, skipping historical cumulative verification and complex tiered calculations, ensuring processing is completed within 800 milliseconds. For example, a "50% off" discount is directly applied to senior citizen cards without verifying the remaining credit limit for the month.

[0138] In high-precision acquisition mode or hybrid transition mode, the system uses the full discount rule base. Deep association is performed based on knowledge graphs or relational databases: verifying ticket discount eligibility (e.g., student card registration status), querying historical usage records (e.g., cumulative discount times in the current month), and calculating multi-trip transfer discount amounts. A transaction mechanism ensures the atomicity of discount deduction and ticket transactions, with processing time approximately 200 to 500 milliseconds. For example, for a senior citizen discount card that passes review in step S54, 2 yuan (50% off) is deducted directly in high-efficiency mode; in high-precision mode, the system queries the local database to confirm that the card has been used 58 times in the current month (limit 60 times), determines compliance with the rules, deducts the remaining amount, records the transaction, and updates the local discount rule usage record to 59 times.

[0139] The anomaly indicators and the discrepancies recorded in the aforementioned cross-validation are input into the cross-modal anomaly causal association model to calculate the root cause confidence score, thereby generating an anomaly feature vector with root cause labels, including:

[0140] Step S61: The extracted visual feature anomaly indicators and electronic feature anomaly indicators, along with the specific difference features recorded during the cross-validation process, are time-stamped and spatially normalized according to a preset spatiotemporal mapping relationship. The aligned multi-source data are then weighted and fused using a preset weighted fusion coefficient to generate a related fusion dataset with preset dimensions and format.

[0141] The cross-validation discrepancy record refers to the specific deviation information generated by the comparison of visual and electronic features in step S44, including field mismatches (such as inconsistent UIDs or conflicting type codes), timestamp deviations, and logical conflict types. It is extracted from the discrepancy log of step S44. The spatiotemporal mapping relationship refers to the mapping rules that unify data from different sources and at different collection times to the same time axis and spatial coordinate system. This is defined through the system's preset clock synchronization protocol (such as a precise time protocol) and site coding system.

[0142] The necessary process is as follows: First, perform spatiotemporal alignment. Extract the timestamps of visual anomaly indicators (image acquisition time), electronic anomaly indicators (RF decoding time), and event timestamps from the difference records. Use a precise time protocol to calibrate the time deviations of the three sources to within ±1 millisecond, establishing a unified time reference. Simultaneously, normalize the spatial identifiers of each data source (such as device number, site code) into standard geographic coordinates or logical site IDs to ensure spatial consistency.

[0143] Secondly, data alignment and fusion are performed. Visual anomaly indicators (e.g., sharpness value 85, confidence level 0.6), electronic anomaly indicators (e.g., signal strength -75dBm, checksum 0xFF indicating failure), and difference records (e.g., UID deviation 3 bytes, time deviation 200ms) are formatted according to a preset dimension mapping table. A weighted fusion coefficient is used to weight the multi-source data: visual indicators 0.4, electronic indicators 0.4, and difference records 0.2 (visual weight can be increased to 0.5 in high-precision mode), generating a fusion feature vector with preset dimensions (e.g., 128 or 256 dimensions).

[0144] Finally, the output is a fused dataset. This dataset contains aligned timestamps, spatial coordinates, weighted fused anomaly feature vectors, and data quality labels, providing standardized input for step S62 to the cross-modal anomaly causal association model. For example, in a verification failure event, visual indicators show "blurry QR code (clarity 60)," electronic indicators show "weak signal (-72dBm)," and difference records show "UID first 4 bytes mismatch." After spatiotemporal alignment and weighted fusion, a standardized dataset containing the above information is generated and labeled "spatiotemporal alignment completed, data quality good, awaiting root cause inference."

[0145] Step S62: Input the associated fusion dataset into the pre-trained cross-modal anomaly causal association model; the model, based on the predefined causal graph structure containing visual anomaly nodes, electronic anomaly nodes and external environment nodes, uses the preset node transition probability and causal strength matrix, and through probabilistic graphical model inference or graph attention network forward computation, quantifies at least one potential root cause node that caused the ticket verification failure and its corresponding confidence probability value.

[0146] The necessary process is as follows: Step S62 inputs the associated fusion dataset generated in step S61 into the pre-trained cross-modal anomaly causal association model, and quantifies the potential root cause nodes and their confidence levels through probabilistic graphical model inference or graph attention network forward computation.

[0147] The model first loads a predefined cause-effect graph structure. This structure contains three types of nodes: visual anomaly nodes (blur, overexposure, smudges, angle deviation), electronic anomaly nodes (weak signal, decoding failure, authentication timeout, data corruption), and external environment nodes (strong light, shadow, electromagnetic interference, multi-card conflict, equipment aging, abnormal temperature). Causal relationships are defined between nodes, such as "strong light → overexposure," "electromagnetic interference → weak signal," and "equipment aging → decoding failure." Edge weights are initialized using the co-occurrence frequency of 100,000 historical anomaly cases and then solidified after expert verification.

[0148] Subsequently, root cause inference is performed. If a probabilistic graphical model (Bayesian network) is used, the fused features from step S61 are used as evidence variables (observed abnormal symptoms). The posterior probability of each root cause node is calculated using a belief propagation algorithm, and the node with the highest conditional probability is output as the potential root cause. If a graph attention network (using a 4-layer Graph Transformer, 256 hidden dimensions, and 8-head attention) is used, the fused features are mapped to query vectors, and the causal graph nodes are embedded as key-value vectors. Multiple neighbor information is aggregated through message passing, and the root cause node confidence distribution is output after softmax normalization. For example, if the input features show "low clarity + normal signal strength + mismatched UID", the model infers: the visual abnormality node "blurred" is activated, the electronic abnormality node "normal", and the difference record points to "visual misreading". Through causal graph backtracking, the probability of strong light causing overexposure blur is 0.75, the probability of device camera offset causing blur is 0.15, and the probability of the ticket itself being damaged is 0.10. The model outputs the potential root cause node "strong light irradiation" with a confidence probability value of 0.75, providing a core label for generating the abnormal feature vector with root cause label in step S63.

[0149] The cross-modal anomaly causal association model was trained using a semi-supervised approach. The causal graph structure was predefined by domain experts, and the node transition probabilities and causal strength matrices were iteratively optimized using the EM algorithm: the E-step calculated the posterior of the latent root cause variables based on the current parameters, and the M-step maximized the full data likelihood to update the parameters. The graph attention network was trained on a labeled dataset (50,000 anomaly-root cause pairs), using cross-entropy loss to optimize node classification accuracy. The learning rate was 1e-4, and the training lasted 100 epochs. The validation set F1-score reached 0.94 (the validation set consisted of 30,000 pairs, accounting for 30% of the total 100,000 pairs, covering five root cause categories; the remaining 50,000 pairs were used for graph attention network training, and 20,000 pairs were used for testing; details of the validation set F1-score variation with training epochs can be found in [link to relevant documentation]). Figure 2 ).

[0150] Step S63: Based on the calculated at least one potential root cause node and its confidence probability value, according to the preset vector encoding template, the visual and electronic feature anomaly indicators, difference records, root cause node identifiers and confidence probability values ​​are dimensionally concatenated and standardized to generate a structured anomaly feature vector with a preset data structure, accompanied by at least one root cause type label and corresponding quantitative confidence value.

[0151] The necessary process is as follows: First, multi-source data encoding is performed. The vector encoding template adopts a segmented structure: the first segment is filled with visual and electronic anomaly indicators (such as sharpness, signal strength, etc., continuous values ​​are normalized to the 0 to 1 range by min-max normalization), the middle segment is filled with one-hot encoding of difference records (such as UID mismatch encoding as [1,0,0,0], timestamp deviation encoding as [0,1,0,0]), and the last segment is filled with root cause labels and confidence levels (root cause types are identified by 8-dimensional one-hot vectors, and the 9th dimension stores the confidence probability value).

[0152] Subsequently, dimensional concatenation and standardization are performed. The encoded segmented vectors are concatenated sequentially to form an initial vector with a preset dimension (e.g., 128 or 256 dimensions). Z-score standardization or Min-Max normalization is used to unify the scale of the entire vector, ensuring that the values ​​of each dimension are at similar magnitudes, thus avoiding the dominance of dimensional differences in distance calculation during subsequent clustering. For example, in step S62, the root cause is determined to be "strong light irradiation" (assuming it corresponds to the 3rd class, one-hot encoded [0,0,1,0,0,0,0,0]), with a confidence level of 0.75; the visual anomaly index is normalized to [0.6,0.4,0.2] (sharpness, damage, exposure deviation); the electronic anomaly index is [0.3,0.1] (signal strength, bit error rate); and the difference record is encoded as [1,0] (identifier mismatch). After splicing and standardization, a 128-dimensional structured abnormal feature vector is generated, which includes both low-dimensional representations of the original abnormal symptoms and high-level root cause semantic labels. This allows the density peak clustering algorithm in subsequent steps to identify specific abnormal types (such as "visual failure caused by strong light") and provides labeled node features for graph neural network inference in step S7.

[0153] Specific anomaly types are identified through clustering. Based on these anomaly types and the data collection pattern, a Bayesian optimization and parameter tuning process is triggered, with the combined performance of the sentiment benchmark and the feature extraction and verification process as dual optimization objectives. This process constructs a joint sentiment-performance objective function and iteratively solves it to obtain the Pareto optimal solution, thereby determining the optimized system parameters, including:

[0154] Step S6A: Based on the structured anomaly feature vectors and combined with the current acquisition mode, adaptive clustering analysis is performed according to the preset clustering parameter configuration to identify clusters representing different fault modes. Each cluster is then assigned a specific anomaly type label according to a preset mapping relationship. Here, the clustering parameter configuration refers to the control parameters of the adaptive clustering algorithm, including neighborhood radius, density threshold, minimum sample size, etc., which are dynamically set according to the acquisition mode and preset through the system configuration table. Fault mode clusters refer to data sets with similar anomaly features, representing specific types of faults (such as strong light blurring, multi-card conflict, equipment aging). These are obtained by grouping anomaly feature vectors using the clustering algorithm. Anomaly type labels refer to semantic classification identifiers assigned to each cluster (such as "lighting anomaly type A", "RF interference type B"). These are obtained by matching the cluster center with the fault mode through a preset mapping relationship.

[0155] The necessary process is as follows: The system first loads preset clustering parameters according to the current acquisition mode: in the high-efficiency acquisition mode, lenient parameters (larger neighborhood radius, lower density threshold) are used to quickly identify coarse-grained fault categories; in the high-precision acquisition mode, strict parameters (smaller neighborhood radius, higher density threshold) are used to finely classify fault subcategories; in the hybrid transition mode, dynamic parameters are used (adaptively adjusted based on the previous clustering results). Then, the density peak clustering algorithm is executed: the Euclidean or Mahalanobis distance between each abnormal feature vector is calculated, local density peak points are identified as cluster centers, and the remaining samples are assigned to the nearest cluster center to form clusters. For example, in the high-efficiency mode during the morning peak, the system quickly identifies two major clusters: "weak signal - cluster 1" and "blurred image - cluster 2"; in the high-precision mode during the off-peak period, "blurred image" is further subdivided into subclusters such as "overexposed strong light," "lens damage," and "ticket folding."

[0156] Finally, based on the preset mapping relationship, the clusters are assigned anomaly type labels: a similarity comparison is established between the cluster center features and the standard fault template (e.g., calculating the cosine similarity with the "strong light illumination" template vector). If the similarity exceeds the threshold, a corresponding label is assigned (e.g., "lighting anomaly - strong light type"). Unmatched clusters are marked as "unknown type" and await manual labeling. For example, a cluster center feature showing "low clarity + high exposure + root cause strong light" is mapped to "lighting anomaly type A"; another cluster shows "sudden drop in signal strength + multiple UID conflicts" and is mapped to "radio frequency interference type B", providing a classification basis for loading the corresponding optimization strategy in step S6B.

[0157] Step S6B: Based on the identified anomaly type label and the current acquisition mode, match and load the corresponding initial system parameter configuration from the preset optimization strategy library.

[0158] The necessary process is as follows: Step S6B: Based on the identified anomaly type label and the current acquisition mode, match and load the corresponding initial system parameter configuration from the preset optimization strategy library.

[0159] The optimization strategy library adopts a three-dimensional index structure of "anomaly type - acquisition mode - parameter configuration". The construction process integrates expert rules and data-driven approaches: the expert rule layer defines the causal relationship between faults and parameters based on domain knowledge (e.g., "overexposure due to strong light → shorten exposure time", "multi-card conflict → increase RF power"); the data-driven layer extracts the optimal parameter snapshots when resolving specific anomalies by mining historical successful parameter tuning logs, and then stores them in the library after statistical filtering. During matching execution, the system searches using the anomaly type as the primary key and the acquisition mode as the filtering condition. For example, if it is identified as "lighting anomaly type A" and the current mode is "high precision", conservative initial parameters (exposure 1.0ms, turn off supplementary lighting, full ROI) are loaded; if it is "high efficiency mode", aggressive parameters (exposure 0.5ms, resolution compression, skip adaptive calculation) are loaded. The loaded parameters serve as the initial starting point (warmstart) for step S6D Bayesian optimization, ensuring that the search starts from the neighborhood of empirically effective parameters, avoiding blind random exploration, and significantly shortening the convergence time.

[0160] Step S6C: The emotion benchmark is transformed into a first optimization indicator using a preset emotion quantification model, and the feature extraction success rate and verification process time are selected as second optimization indicators. Based on preset weights, the first and second indicators are combined to construct a joint emotion-performance objective function. The feature extraction success rate refers to the ratio of successfully extracted effective visual and electronic features in steps S2 to S3 (number of successful attempts / total number of attempts). This is obtained through real-time system statistics. The verification process time refers to the processing time (in milliseconds) for cross-validation in step S4. This is obtained through system clock measurement.

[0161] The necessary process is as follows: First, perform emotion quantification and transformation. Call the emotion baseline established in step S6 (the sum of anxiety and confusion dimensions) and map it to an emotional stress index (range 0 to 2) through a quantification model. For example, if micro-expression recognition shows a user frowning and pursing their lips, the probability of anxiety is 0.6 and the probability of confusion is 0.5. After weighted summation, the emotional stress index is 1.1, which is the first optimization indicator to minimize (a higher index indicates a higher priority for appeasing the user). Second, calculate the performance index. Statistically calculate the feature extraction success rate of the last 10 transactions (e.g., 8 successful attempts result in a success rate of 0.8) and the average verification time (e.g., 650 milliseconds). Normalize the time (e.g., set a baseline of 1000 milliseconds, normalization value 0.65), and construct the second optimization index using a weighted combination: Performance Score = Success Rate × 0.6 + (1 - Normalized Time) × 0.4, with the goal of maximizing this score. Finally, construct a joint objective function. The weights are dynamically set based on the current data collection mode: In the high-efficiency data collection mode, the performance weight is 0.7 and the emotion weight is 0.3, with the objective function being to minimize (0.3 × emotion stress index - 0.7 × performance score); in the high-precision data collection mode, the emotion weight is 0.6 and the performance weight is 0.4, with the objective function being to minimize (0.6 × emotion stress index - 0.4 × performance score). For example, in the high-efficiency mode during peak hours, the emotion stress is 1.1 and the performance score is 0.75, with a joint objective function value of 0.3 × 1.1 - 0.7 × 0.75 = -0.195. Bayesian optimization will iteratively solve for the system parameters that minimize this value (i.e., reducing emotion stress while improving performance).

[0162] Step S6D: Starting from the initial system parameter configuration, Bayesian optimization iteration is performed within the preset parameter space with the joint objective function as the objective. The iteration process terminates according to the preset convergence condition, and the final system parameter combination is selected as the output from the obtained non-dominated solution set based on the preset Pareto solution evaluation strategy.

[0163] The initial system parameter configuration refers to the starting search point loaded from the optimization strategy library in step S6B, including specific values ​​for processing parameters such as exposure duration and polling frequency. This is obtained through strategy library matching queries. The preset parameter space refers to the feasible value range of each system parameter (e.g., exposure time 0.5ms to 5ms, polling frequency 50ms to 1000ms), forming a multi-dimensional search space. This is preset based on system hardware specifications and business requirements. The joint objective function refers to the sentiment-performance weighted optimization objective constructed in step S6C, which needs to be minimized or maximized. This is obtained through sentiment quantification models and performance statistical calculations. Bayesian optimization iteration refers to an efficient global optimization method based on a Gaussian process surrogate model, which guides the next set of sampling points through a sampling function to find the optimal solution by minimizing the number of observations. This is implemented through a Bayesian optimization framework. The convergence condition refers to the criteria for terminating the iteration, including the maximum number of iterations (e.g., 50 times) or the improvement in the objective function being less than a threshold (e.g., improvement <1% after 5 consecutive iterations). This is preset through system configuration. Non-dominated solution set: refers to the set of solutions on the Pareto front in multi-objective optimization, where no objective can be further optimized without compromising other objectives. It is obtained through Pareto dominance filtering.

[0164] Pareto solution evaluation strategy: refers to the decision rule for selecting the final solution from the non-dominated solution set (such as prioritizing minimizing emotional stress, or minimizing the weighted Euclidean distance). It is dynamically determined through a preset strategy or data acquisition mode. Final system parameter combination: refers to the optimal values ​​of processing parameters such as exposure duration and polling frequency, determined through optimization. It is selected from the non-dominated solution set using the Pareto solution evaluation strategy.

[0165] The necessary process is as follows: First, construct a Gaussian process surrogate model. Using the initial parameter configuration and its corresponding objective function value as the initial observation points, establish a black-box approximation model of the objective function. Characterize the correlation of the parameter space through radial basis function kernels to predict the function values ​​and uncertainties in the unexplored region.

[0166] Subsequently, iterative optimization is performed. Each iteration calculates the acquisition function (as desired) based on the current surrogate model and selects the next most promising sampling point within the parameter space. The system injects this parameter combination into the actual operating environment, collects sentiment and performance data, calculates the joint objective function value, and updates the model. Iteration continues until a preset convergence condition is met (e.g., reaching 50 iterations or 5 consecutive improvements <1%). After iteration terminates, a non-dominated solution set is obtained. A Pareto solution evaluation strategy is executed based on the current acquisition mode: in high-precision mode, the solution with the lowest sentiment pressure is prioritized; in high-efficiency mode, the solution with the highest performance score is prioritized; in mixed mode, the compromise solution with the shortest weighted Euclidean distance is selected. For example, after 30 iterations, three non-dominated solutions are obtained (sentiment 1.0 / performance 0.8, sentiment 0.8 / performance 0.7, sentiment 1.2 / performance 0.9). Since the current mode is high-precision, the solution with sentiment 0.8 is selected, and the final system parameter combination (e.g., exposure 2.0ms, polling 100ms, power 1.0W) is output to update the acquisition strategy in step S1.

[0167] Based on the optimized system parameters and the micro-expression recognition results corresponding to the emotion benchmark, the following adaptation guidance information is output:

[0168] Step S6a: Based on the facial image data corresponding to the emotion benchmark, facial micro-expression features are extracted using a preset feature extraction algorithm to generate a quantitative vector of emotional state. The facial image data corresponding to the emotion benchmark refers to the user's facial images collected in step S6, including micro-expression information such as frowning and pursing lips. These images are acquired using the front-facing camera of a BOM (640×480 resolution, 30fps). The preset feature extraction algorithm is a deep learning model that extracts micro-expression features from facial images, employing either an improved EfficientNet-B0 network or a facial landmark detection algorithm. It is pre-trained on a public expression dataset and a self-built subway scene dataset. This micro-expression recognition model is based on the EfficientNet-B0 architecture, with optimizations for subway scenes: after pre-training on the public expression dataset, it is fine-tuned using a self-built subway scene face dataset (5000 images, covering crowded, backlit, and multi-angle scenarios). During fine-tuning, a cross-entropy loss function and an Adam optimizer (learning rate 5e-5) are used, freezing some lower-level networks and only fine-tuning the higher-level networks and the classification head. This training strategy ensures that the model maintains its general recognition capabilities while exhibiting better robustness to typical lighting, angle, and facial expression changes in subway environments.

[0169] Facial micro-expression features: These represent subtle changes in facial muscle movement, including facial action unit (AU) activation intensity, keypoint displacement vectors, and texture variation features. They are extracted from aligned facial images using feature extraction algorithms. Emotional state quantification vector: This refers to a numerical vector mapping micro-expression features to multi-dimensional emotional probabilities, covering emotional dimensions such as anxiety, confusion, and calmness. It is generated through forward computation in the classification layer.

[0170] The necessary process is as follows: First, facial image preprocessing is performed. Face detection (using MTCNN or RetinaFace algorithm) and alignment are performed on the facial images acquired in step S6. Standardized facial regions (224×224 pixels) are cropped, and histogram equalization and denoising are performed to eliminate ambient light interference. Next, micro-expression feature extraction is performed. An improved efficient network (EfficientNet-B0, width coefficient 1.0, depth coefficient 1.0) is used as the backbone network to extract deep convolutional features; or a facial landmark detection algorithm (Dlib68-point model) is used to extract facial action unit (AU) features. For example, the activation intensity of micro-expression action units such as eyebrow drooping (AU4) and lip tightening (AU10) is detected, outputting 128-dimensional or 256-dimensional feature vectors. Finally, an emotion state quantification vector is generated. Micro-expression features are input into a fully connected classification layer and mapped to a probability distribution of 7 types of emotions (anxiety, confusion, anger, calmness, satisfaction, surprise, and neutrality). The distribution is output as a vector (e.g., anxiety 0.6, confusion 0.5, calmness 0.1, and the rest below 0.05), which serves as the basic input for matching the emotion-parameter fusion guide library in step S6b.

[0171] Step S6b involves matching the quantified emotional state vector with the anomaly types associated with the optimized system parameters, and then calling a pre-defined emotion-parameter fusion guidance library to generate preliminary guidance content. The emotion-parameter fusion guidance library is a knowledge base that stores the mapping relationship between "emotional state - anomaly type - guidance content," defining interaction strategies for different combinations of emotions and anomalies. It is constructed through domain expert annotation and historical interaction data mining. The preliminary guidance content refers to the original interaction instructions generated based on the matching results, including prompt text, voice scripts, and interface element configurations. It is generated through guidance library queries and template filling.

[0172] The necessary procedures are as follows:

[0173] The system first establishes a joint query key: using the anomaly type (e.g., "lighting anomaly") as the primary key and the dominant emotion of the emotional state quantification vector (e.g., "anxiety") as the secondary key, it retrieves corresponding entries from the guidance library. The guidance library uses a two-dimensional index structure, with the horizontal axis representing the anomaly type (lighting, radio frequency, equipment failure, etc.) and the vertical axis representing the emotional level (high anxiety, moderate anxiety, calm, etc.). Each cell stores a differentiated guidance template. For example, when matching the combination of "lighting anomaly + high anxiety," a gentle and reassuring template is loaded: "The light is dim, please wait, adjusting..."; when matching the combination of "lighting anomaly + calm," a concise and operational template is loaded: "Please lay the ticket flat." Simultaneously, the specific parameters optimized in step S6D (e.g., "exposure increased to 2ms") are transformed into a user-understandable system status description (e.g., "enhancing recognition"), and embedded in the template to generate preliminary guidance content. The preliminary guidance content includes multimodal information such as text, voice tone markings (soothing / rapid), and interface simplification level (displaying the complete steps / displaying only icons), providing raw materials for step S6c to adjust the presentation format according to the acquisition mode.

[0174] Step S6c: Based on the current acquisition mode and the preset processing priority for the anomaly type, dynamically adjust the level of detail and interaction of the guidance content. The preset processing priority for the anomaly type refers to the urgency and timeliness requirements of different fault categories (e.g., "multi-card conflict" has high priority and requires immediate handling, while "equipment aging" has low priority and can be delayed for repair). This is preset through an anomaly type and priority mapping table. The level of detail refers to the richness of the guidance information, including detailed step-by-step instructions, simplified prompts, or pure icon guidance. This is determined jointly by the acquisition mode and priority. The interaction format refers to the output method of the guidance information, including voice broadcast, on-screen text, light prompts, or a combination of vibration feedback. This is defined through multimodal output configuration.

[0175] The necessary process is as follows: In high-efficiency acquisition mode, a minimalist strategy is adopted regardless of priority: the guidance is compressed to within 8 characters through keyword extraction (such as "Adjusting, please wait"), the voice is at a high rate (320 words per minute), and the interface only displays the central progress animation to ensure that the interaction time is less than 3 seconds.

[0176] In high-precision acquisition mode, rendering is differentiated according to priority: high-priority anomalies are displayed with step-by-step graphic guidance and soothing voice (220 words per minute), and key areas are highlighted in the full screen; low-priority anomalies are displayed with brief status prompts.

[0177] In the hybrid transition mode, a layered lazy loading strategy is adopted: the first screen displays the core prompts, and after the user stays for more than 3 seconds, the detailed explanation automatically expands, achieving a smooth transition between detailed and concise information. For example, in the "abnormal lighting" scenario: the high-efficiency mode during the morning rush hour only announces "Please wait" and displays a rotating progress bar; while the high-precision mode during off-peak hours displays a graphic guide stating "The light is dim, automatic lighting has been added, please place the ticket flat in the recognition area," accompanied by a soothing voice to ensure that the user understands the system's adjustment intentions.

[0178] In step S6d, if the quantification vector of the emotional state deviates from the preset acceptability range, then based on the optimized system parameters and emotional benchmark, supplementary contextual information is matched from the preset emotion-scenario response library. The preset acceptability range refers to the threshold range for determining whether a user's emotion requires intervention (e.g., an anxiety index of 0 to 0.7 is acceptable). This is preset based on user satisfaction surveys and stress test data. The emotion-scenario response library is a knowledge base storing the mapping relationship between "high emotional state - abnormal type - coping strategy," including reassurance scripts, technical explanation templates, etc. It is constructed through customer service expert experience annotation and historical success case mining. Supplementary contextual information refers to additional reassurance content or explanations generated for users with high emotional states, used to alleviate anxiety and explain the system's adjusted behavior.

[0179] The necessary process is as follows: Step S6d compares the emotional state quantification vector with the preset acceptance range. If there is a deviation, supplementary information is matched from the emotion-scene response library.

[0180] The system performs threshold determination: it compares each emotion dimension (anxiety, confusion, etc.) output by S6a with the upper limit of the preset range. For example, if the anxiety threshold is set to 0.7, and an anxiety index of 0.85 (outside the range) is detected, it is determined to be an emotion deviation, triggering the supplementary information generation mechanism; if it is within the range (e.g., anxiety 0.4), this step is skipped.

[0181] When querying the emotion-scene response library, a three-dimensional index key is constructed: anomaly type (e.g., abnormal lighting), optimized system parameter status (e.g., "exposure has been increased to 2ms"), and dominant emotion (high anxiety). The response library stores differentiated response templates: high-anxiety scenarios are matched with technical explanation-based reassurance (e.g., "We are increasing your brightness, please rest assured"), and high-confusion scenarios are matched with operational guidance prompts (e.g., "Please keep your ticket still, the system is recognizing it"). The templates are populated with specific parameter values ​​through a rule engine to generate contextualized supplementary content. This supplementary information is integrated with the basic guidance in step S6c to ensure that the final output not only conveys the system's adjustment intent (e.g., exposure has been optimized) but also alleviates the user's negative emotions through empathetic language, avoiding operational errors caused by heightened anxiety due to misunderstanding of technical adjustments.

[0182] Step S6e outputs the final guidance, which integrates the initial guidance content, the dynamically adjusted presentation format, and supplementary contextual information.

[0183] The necessary process is as follows: The system first converts the optimization parameters determined in step S6D (such as increasing exposure to 2ms and adjusting power to 1.5W) into a technical status description that the user can understand (such as "brightness has been increased" and "sensing range has been expanded"), and then splices them with the presentation detail determined in step S6c and the supplementary information in step S6d to form complete content. Subsequently, it outputs through multi-channel rendering: the voice channel adopts a speech rate and tone that matches the emotion adaptation results (soothing when there is high anxiety and clear when there is high confusion); the visual channel draws status icons (such as a sun icon to indicate supplemental lighting), progress indicators, and operation prompts; the auxiliary channel controls the color of the LED light strip and the buzzer prompts.

[0184] For example, in a scenario of "morning rush hour + abnormal lighting + high anxiety," the system outputs: "Brightness is being increased, please wait" (soothing voice), and the interface displays a yellow progress ring and a sun icon; in a scenario of "off-peak hours + radio frequency interference + calm," the system outputs: "Senses have been enhanced, please swipe your card again" (normal voice), and the interface displays a green checkmark and a placement diagram. This fusion ensures that the passenger perception system proactively optimizes while providing an interactive experience that is appropriate to the passenger's emotions.

[0185] If the quantification vector of the emotional state deviates from the preset acceptability range, then based on the optimized system parameters and emotional benchmark, contextual information is matched and supplemented from the preset emotion-scenario response library, including:

[0186] Step S6d1 involves performing a difference operation between the quantified emotional state vector and the emotional baseline to calculate the emotional deviation index. Based on a preset threshold range, the emotional deviation index is mapped to the corresponding emotional intervention level. The emotional deviation index (range 0-1) is a numerical indicator quantifying the degree of deviation between the current emotion and the expected baseline. It is obtained through vector difference operations and weighted summation. The emotional intervention level refers to the response level categorized according to the degree of deviation (e.g., no intervention required, mild reassurance, moderate intervention, deep intervention). It is preset using threshold range mapping rules.

[0187] The necessary procedures are as follows:

[0188] The system first performs vector differencing: the current emotion vector (dimension 7, covering anxiety, confusion, calmness, etc.) from step S6a is subtracted element-wise from the preset baseline vector. For example, if the current anxiety is 0.8 and confusion is 0.3, and the baseline is 0.3 and confusion is 0.1, the difference result is anxiety +0.5 and confusion +0.2. Next, the emotion deviation index is calculated: the difference results are weighted and summed (anxiety weight 0.4, confusion weight 0.3, other negative emotion weight 0.3), generating a scalar index of 0.5 × 0.4 + 0.2 × 0.3 = 0.26, which is normalized to the 0-1 range. Finally, intervention levels are mapped: threshold ranges are set (0-0.2 for normal, 0.2-0.5 for mild reassurance, 0.5-0.8 for moderate intervention, and 0.8-1.0 for deep intervention). A score of 0.26 falls into the mild range, corresponding to a "mild reassurance" level, triggering step S6d2 to load a mild response strategy. A score of 0.85 corresponds to "deep intervention," initiating a mandatory reassurance and preparatory mechanism for manual intervention. This grading system ensures that resources are precisely allocated to high-deviation scenarios, avoiding over-response.

[0189] Step S6d2 involves encoding the emotion intervention level, the abnormality type associated with the optimized system parameters, and the current acquisition mode into features, and then concatenating and combining them according to preset dimensions to generate a multi-dimensional retrieval vector. Based on this multi-dimensional retrieval vector, a preset graph retrieval algorithm is used to calculate its correlation score with each coping strategy node in the emotion-scenario response library, and the correlation scores are used to sort the nodes to determine the matching contextualized coping templates. The emotion-scenario response library uses a preset graph structure to store the correlation between emotional states, scene features, and coping strategies.

[0190] The necessary procedures are as follows:

[0191] The system first performs feature encoding: the emotion intervention level uses 4-dimensional one-hot encoding (e.g., mild soothing is [0,1,0,0]), the abnormality type uses a 16-dimensional embedding vector (generated through Word2Vec or autoencoder pre-training), and the acquisition mode uses 3-dimensional one-hot encoding (efficient [1,0,0], high precision [0,1,0], mixed [0,0,1]). These are concatenated according to a preset dimensional order to generate a 23-dimensional retrieval vector (example: mild soothing + abnormal lighting + efficient mode = [0,1,0,0,0.2,-0.5,0.8,...,1,0,0]).

[0192] Subsequently, a graph retrieval algorithm is used to calculate the relevance. The emotion-scene response library adopts a heterogeneous graph structure: emotion nodes, abnormal nodes, and pattern nodes are condition nodes, and policy nodes are result nodes. Edge weights are initialized using the co-occurrence frequency of historical successful cases. The algorithm uses a graph attention network (2 layers, 8 attention heads, 64 hidden dimensions), using the retrieval vector as the query and calculating the attention score with each policy node; or using cosine similarity to calculate the similarity between the retrieval vector and the policy node embedding vector. For example, the query vector has a similarity of 0.85 with the "gentle lighting prompt" policy node and 0.12 with the "forced manual intervention" policy node. Finally, based on the relevance score, the top-1 or top-K policy nodes are selected, and the corresponding contextualized response templates (such as "Please wait, we are enhancing your brightness") are extracted. For example, mild soothing + abnormal lighting + efficient mode matches the "concise technical explanation" template, while deep intervention + radio frequency interference + high precision mode matches the "detailed soothing + operation guidance" template, providing basic materials for the personalized generation of step S6d3.

[0193] In step S6d3, if the emotion deviation index exceeds the preset dynamic generation threshold, a preset generative supplementation mechanism is activated. Based on the preset dimensional features of the emotion state quantification vector, the technical status of the system parameters, and the aforementioned contextualized response template, personalized supplementary information is generated. The dynamic generation threshold (e.g., an emotion deviation index of 0.8) is higher than the basic intervention threshold in step S6d1, targeting deep emotion deviation scenarios requiring dynamic synthesis rather than template matching. It is preset through system risk configuration. The generative supplementation mechanism is a technical mechanism for dynamically synthesizing reassuring text based on a deep generative model, employing a lightweight sequence generation network (e.g., a text generation model based on a transformer architecture, a 6-layer encoding / decoding structure, and 256 hidden dimensions), distinct from the fixed template retrieval in step S6d2. It is constructed through fine-tuning of subway customer service dialogue data and emotion annotation data. The technical status of the system parameters refers to the specific parameter values ​​optimized in step S6d and their natural language semantic mappings (e.g., an exposure time of 2 milliseconds maps to "increasing brightness," and a radio frequency power of 1.5 watts maps to "expanding the sensing range"). It is obtained through a parameter-semantic mapping table.

[0194] The necessary procedures are as follows:

[0195] Step S6d3 first performs a trigger judgment: the emotion deviation index calculated in step S6d1 is compared with a dynamic generation threshold (e.g., 0.8). If the index exceeds the threshold (e.g., anxiety 0.85), a generative supplementation mechanism is activated; otherwise, the fixed template matched in step S6d2 is used. After activation, the system constructs multi-source conditional inputs: key dimensions of the emotion state quantification vector (e.g., anxiety 0.85, confusion 0.4) are encoded into emotional conditional vectors, and the technical status of system parameters (e.g., "exposure compensation to 2 milliseconds") are encoded into technical description vectors. These are then fused with the scenario-based response template provided in step S6d2 (as generation constraints or style references) to form a composite prompt for the generative model. The generative model adopts an encoder-decoder architecture: the encoder (based on a bidirectional transformer) extracts the semantic features of the conditional inputs, and the decoder (based on an autoregressive transformer) generates personalized text word by word. The model is fine-tuned on 100,000 sets of subway scene customer service dialogues (including emotion tags and system status annotations). The loss function is a weighted sum of cross-entropy and emotion consistency loss, ensuring that the generated content is both consistent with technical facts and reassuring. The generation process constrains the output through a length control mechanism (limited to 15 characters in efficient mode and allowed to 30 characters in high-precision mode) and emotional polarity control (forced use of soothing words).

[0196] For example, when an anxiety index of 0.85 (exceeding the threshold) is detected and the system parameters display "exposure time 2 milliseconds (lighting up)," the model generates a personalized supplement: "The light is a little dim, which is making you anxious. We are brightening it for you, and it will be ready soon." If the anxiety index is 0.9 and the parameters display "RF power 1.5 watts (enhancing)," then the model generates: "The sensor is a little weak. We have enhanced it for you. Please try again." This dynamically generated content is integrated with the template in step S6d2 to form the final supplementary contextual information, ensuring that the technical adjustment status is accurately conveyed while also specifically alleviating user anxiety.

[0197] Step S6d4: Integrate the contextualized response template and personalized supplementary information according to the preset fusion strategy to generate the final supplementary contextualized information.

[0198] The necessary process is as follows: Step S6d4 integrates the contextualized response template and personalized supplementary information according to the preset fusion strategy based on the triggering state of step S6d3.

[0199] If step S6d3 is not triggered (emotional deviation index does not exceed the dynamic generation threshold), the template matched in step S6d2 is directly output as the final supplementary information to ensure response efficiency. If step S6d3 has been triggered, a layered fusion strategy is executed: a slot identification and filling mechanism is used to parse the markers in the template (such as "[emotional reassurance] Please wait, [technical action]") and fill the corresponding slots with personalized supplementary information (such as "The dim lighting has made you anxious"); or a prefix concatenation strategy is used to use personalized content as an emotional reassurance prefix and technical descriptions in the template as suffixes to form a complete sentence.

[0200] For example, the template for step S6d2 is "Adjusting, please wait," and step S6d3 generates "Sorry to keep you waiting, the lighting is a bit dim." After being integrated using a primary and secondary fusion strategy, it becomes: "Sorry to keep you waiting, the lighting is a bit dim, adjusting, please wait." The fusion process ensures natural splicing and avoids semantic breaks through text coherence checks (based on language model perplexity scoring). Based on the same inventive concept, this invention provides an AI-based subway BOM ticketing processing system, including a memory and a processor. The memory stores information that can run on the processor to implement... Figure 1 The procedure for the method shown.

[0201] It should be noted that all data collection, processing, and decision-making processes involved in this application strictly adhere to the following compliance principles: 1. Facial image collection: Triggered only when ticket verification is abnormal and user emotion analysis is required. Before collection, users are clearly informed through the interface and their consent is obtained. The collection scope is limited to the facial areas required for emotion recognition. Image data is processed in real time, not persistently stored, and destroyed immediately after processing. 2. Fairness of algorithmic decisions: All intelligent routing, risk assessment, and other decisions are based on objective technical indicators (such as passenger flow density, equipment status, anomaly type, processing time, etc.) and do not involve any discriminatory rules based on sensitive attributes such as gender, age, race, or region, ensuring a fair and transparent processing flow. 3. Legitimacy of technical purpose: This application aims to improve the efficiency, accuracy, and passenger experience of subway ticketing processing. It is a legitimate technical improvement for urban rail transit operation and management, conforms to the public interest, and does not involve any abuse or violation.

[0202] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A subway BOM ticketing processing method based on artificial intelligence, characterized in that, include: Collect ticket images and electronic recognition data, and switch the collection mode according to the passenger flow density threshold; Visual features are extracted from ticket images using object detection algorithms, and electronic features are obtained by parsing electronic identification data. Simultaneously monitor the quality of visual features and electronic features, and filter features whose quality is below a preset basic threshold. Perform cross-validation on the quality-compliant features to confirm their consistency. If the verification passes, a second verification is performed using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement technology to improve recognition accuracy, and local discount rules are associated. If the verification fails or the feature quality is substandard, user facial images are collected to establish an emotional baseline, and abnormal indicators of substandard visual and electronic features are extracted. The abnormal indicators and the difference records of the aforementioned cross-validation are input into the cross-modal abnormal causal association model to calculate the root cause confidence and generate an abnormal feature vector with root cause label. By clustering to identify specific anomaly types, and based on the anomaly type and collection mode, a Bayesian optimization parameter tuning process is triggered with the combined performance of the emotion benchmark and the feature extraction and verification process as the dual optimization objectives. This parameter tuning process obtains the Pareto optimal solution by constructing a joint emotion-performance objective function and iteratively solving it, thereby determining the optimized system parameters. The system parameters include the processing parameters associated with the feature extraction and verification process. Based on the optimized system parameters and the micro-expression recognition results corresponding to the emotion benchmark, adaptation guidance information is output. If the retest after adjustment still fails to meet the standards, a diagnostic report will be generated based on the aforementioned multimodal knowledge graph and abnormal feature vectors with root cause labels, and manual verification or maintenance instructions will be pushed to the system. If the final identification is successful, intelligent ticketing routing based on reinforcement learning will be executed, and the operation result will be displayed; If a failure occurs, the risk level will be determined by a pre-set risk assessment model, and an appropriate handling plan will be implemented. The anomaly indicators and the discrepancies recorded in the aforementioned cross-validation are input into the cross-modal anomaly causal association model to calculate the root cause confidence score, thereby generating an anomaly feature vector with root cause labels, including: The extracted visual feature anomaly indicators and electronic feature anomaly indicators, along with the specific difference features recorded during the cross-validation process, are time-stamped and spatially normalized according to a preset spatiotemporal mapping relationship. The aligned multi-source data are then weighted and fused using a preset weighted fusion coefficient to generate an associated fusion dataset with preset dimensions and format. The associated fusion dataset is input into a pre-trained cross-modal anomaly causal association model. Based on a predefined causal graph structure containing visual anomaly nodes, electronic anomaly nodes, and external environment nodes, the model uses a preset node transition probability and causal strength matrix to quantify at least one potential root cause node that caused the ticket verification failure and its corresponding confidence probability value through probabilistic graphical model inference or graph attention network forward computation. Based on at least one potential root cause node and its confidence probability value, the visual and electronic feature anomaly indicators, difference records, root cause node identifiers and confidence probability values ​​are dimensionally concatenated and standardized according to a preset vector encoding template to generate a structured anomaly feature vector with a preset data structure, at least one root cause type label and corresponding quantitative confidence value.

2. The subway BOM ticketing processing method based on artificial intelligence according to claim 1, characterized in that, Switching the data collection mode based on passenger flow density threshold includes: By combining real-time passenger flow density with historical data reflecting periodic statistical patterns and external events reflecting sudden impacts, a modified density threshold is dynamically generated through a passenger flow prediction model. The current passenger flow density is compared with the corrected density threshold in real time; If the current passenger flow density is greater than the corrected density threshold, the preset high-efficiency data collection mode will be activated. If the current passenger flow density is less than the corrected density threshold, the preset high-precision data collection mode will be activated. If the current passenger flow density is equal to the corrected density threshold, the preset mixed transition mode is enabled.

3. The subway BOM ticketing processing method based on artificial intelligence according to claim 2, characterized in that, Perform cross-validation on the quality-compliant features to confirm consistency between the two, including: A preset feature association mapping library is used to associate preset key identifiers in visual features with preset core fields in electronic features; Adjust the verification dimensions according to the current acquisition mode. The high-efficiency acquisition mode enables the preset core verification dimensions, the high-precision acquisition mode enables the preset full-quantity verification dimensions, and the hybrid transition mode allocates the preset verification dimensions according to the resource ratio. The two types of feature acquisition time sequences are aligned by a preset timestamp mechanism, and feature matching is completed within a preset time window; Consistency is determined based on a preset matching threshold. If the matching degree meets the standard, it passes; if it is within the preset fault tolerance range, it is marked as weakly consistent and supplementary verification is triggered; if it is outside the range, it is determined to be inconsistent and the difference characteristics are recorded. Output the verification results.

4. The subway BOM ticketing processing method based on artificial intelligence according to claim 3, characterized in that, The recognition accuracy is enhanced through secondary verification using a multimodal knowledge graph, combined with spatiotemporal anomaly enhancement technology, and local discount rules are associated with: Based on the features that pass cross-validation, the appropriate graph resources are called according to the current acquisition mode. In the high-efficiency acquisition mode, the preset core sub-graph is loaded and the rule engine is used to generate fast verification results. In the high-precision acquisition mode, the full graph is enabled and the graph neural network is used to generate deep verification results. The two together constitute the basic verification results. Based on the basic verification results, real-time passenger flow, station and time segment features are integrated, and the weight of graph nodes is dynamically adjusted through an attention mechanism to output the optimized verification results. Based on the optimized verification results, a spatiotemporal anomaly enhancement technique is introduced: the isolated forest algorithm is used to screen for abnormal features that deviate from the historical transaction baseline, and the LSTM model is used to learn the same site and time period within the previous preset number of days to generate a dynamic spatiotemporal normal interval. Calculate the deviation between abnormal features and the normal range. If the deviation exceeds the threshold, call the preceding quality compliance features to construct a double verification chain to complete the abnormality review. If the review fails, the associated local discount rules will be skipped, and the exception handling process that starts by collecting the user's facial image will be triggered. If the review is successful, the local discount rules will be associated according to the collection mode.

5. The subway BOM ticketing processing method based on artificial intelligence according to claim 1, characterized in that, Specific anomaly types are identified through clustering. Based on these anomaly types and the data collection pattern, a Bayesian optimization and parameter tuning process is triggered, with the combined performance of the sentiment benchmark and the feature extraction and verification process as dual optimization objectives. This process constructs a joint sentiment-performance objective function and iteratively solves it to obtain the Pareto optimal solution, thereby determining the optimized system parameters, including: Based on structured anomaly feature vectors, combined with the current acquisition mode, adaptive clustering analysis is performed according to the preset clustering parameter configuration to identify clusters that represent different fault modes, and each cluster is assigned a specific anomaly type label according to the preset mapping relationship. Based on the identified anomaly type label and the current acquisition mode, the corresponding initial system parameter configuration is matched and loaded from the preset optimization strategy library; The emotion benchmark is transformed into the first optimization index through a pre-set emotion quantification model, and the feature extraction success rate and the verification process time are selected as the second optimization indexes. Based on the pre-set weights, the first optimization index and the second optimization index are combined to construct an emotion-performance joint objective function. Starting with the initial system parameter configuration, Bayesian optimization iteration is performed within the preset parameter space with the joint objective function as the objective. The iteration process terminates according to the preset convergence condition, and the final combination of system parameters is selected as the output from the obtained non-dominated solution set based on the preset Pareto solution evaluation strategy.

6. The subway BOM ticketing processing method based on artificial intelligence according to claim 5, characterized in that, Based on the optimized system parameters and the micro-expression recognition results corresponding to the emotion benchmark, the following adaptation guidance information is output: Based on facial image data corresponding to the emotion benchmark, facial micro-expression features are extracted through a preset feature extraction algorithm to generate a quantitative vector of emotion state. The emotional state quantification vector is matched with the anomaly type associated with the optimized system parameters, and a preset emotion-parameter fusion guidance library is invoked to generate preliminary guidance content. Based on the current data collection mode and the pre-defined processing priority for the anomaly type, the level of detail and interactive format of the guidance content are dynamically adjusted. If the emotional state quantification vector deviates from the preset acceptability range, then contextual information is matched and supplemented from the preset emotion-scenario response library based on the optimized system parameters and emotional benchmark. The output integrates the initial guidance content, the dynamically adjusted presentation format, and the final guidance with supplementary contextual information.

7. The subway BOM ticketing processing method based on artificial intelligence according to claim 6, characterized in that, If the quantification vector of the emotional state deviates from the preset acceptability range, then based on the optimized system parameters and emotional benchmark, contextual information is matched and supplemented from the preset emotion-scenario response library, including: The emotional state quantification vector is differentially calculated with the emotional benchmark to obtain the emotional deviation index; the emotional deviation index is then mapped to the corresponding emotional intervention level based on a preset threshold range. The emotional intervention level, the abnormal type associated with the optimized system parameters, and the current collection mode are respectively feature-encoded and concatenated according to preset dimensions to generate a multi-dimensional retrieval vector. Based on this multi-dimensional retrieval vector, a preset graph retrieval algorithm is used to calculate its correlation score with each coping strategy node in the emotion-scene response library. The correlation scores are then sorted to determine the matching contextualized coping templates. The emotion-scene response library uses a preset graph structure to store the correlation between emotional state, scene features, and coping strategies. If the emotion deviation index exceeds the preset dynamic generation threshold, the preset generative supplementation mechanism is activated; based on the preset dimensional features of the emotion state quantification vector, the technical status of the system parameters, and the aforementioned contextualized response template, personalized supplementary information is generated. The contextualized response templates and personalized supplementary information are integrated according to a preset fusion strategy to generate the final supplementary contextualized information.

8. A subway BOM ticketing processing system based on artificial intelligence, characterized in that, It includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements an AI-based subway BOM ticketing processing method as described in any one of claims 1 to 7.

Images