A smart device for TCMS
By designing fault analysis and predictive analysis units for intelligent devices, real-time fault diagnosis and status prediction of the TCMS system were realized, solving the problems of untimely fault handling and lack of prediction in existing technologies, and improving the safety of train operation and maintenance efficiency.
Patent Information
- Application Number
- CN202611147373.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-17
AI Technical Summary
The existing TCMS system lacks real-time reliability in fault diagnosis and status prediction, and cannot achieve two-way linkage between faults and operating status, resulting in untimely fault handling and a lack of effective prediction, which affects train safety and maintenance efficiency.
Design an intelligent device for TCMS, including a main control communication unit, a backplane interconnection unit, a fault analysis unit, and a predictive analysis unit. Real-time interaction and mutual correction between fault analysis and status prediction are achieved through a bidirectional linkage data channel. Real-time data preprocessing and feature engineering pipelines are adopted, combined with fault detection and prediction models, to realize real-time acquisition and online analysis of multi-source asynchronous data from TCMS.
It significantly shortens fault response time, improves the real-time performance and reliability of fault analysis, enables online prediction of TCMS operating status, enhances train operation safety and maintenance initiative, solves the problem of the separation between fault analysis and status prediction in existing technologies, and forms a mutually reinforcing closed-loop collaborative mechanism.
Smart Images

Figure CN123104729A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban rail transit technology, and in particular to an intelligent device for TCMS. Background Technology
[0002] TCMS, or Train Control and Monitoring System, is the core intelligent control and monitoring hub for rail transit locomotives, EMUs, and urban rail trains. Often referred to as the "brain and nerve center" of the train, it typically adopts a distributed architecture based on WTB / Ethernet bus. With the Train Control Unit (CCU) at its core, it links local control units of various key subsystems such as traction, braking, and pantograph. It can coordinate and issue overall train operation control commands, collect various operating parameters in real time and display them on the DDU interface in the driver's cab, and has real-time fault diagnosis, hierarchical alarm and automatic protection functions. It can also realize vehicle-to-ground communication, vehicle-to-vehicle multiple-unit data interaction and human-machine command transmission. By integrating and controlling the coordinated operation of all electrical and mechanical subsystems of the train, it ensures the safe, stable and efficient operation of the train. It is the core hardware and software integration system for realizing automated and intelligent train operation.
[0003] In traditional TCMS, all data is recorded in the ERM via an Ethernet bus, and the core components within the TCMS interact with data and information via the Ethernet bus. Figure 1 The blue section (including the DDU) represents the core components of the TCMS (Train Control System). CS is the Ethernet switch, CCU is the train control unit, GW is the gateway, RIOM is the remote input / output unit, and ERM is the data recorder. If a TCMS or other equipment malfunctions, maintenance personnel can only manually download the fault data after the train enters the depot or stops, and then analyze the fault using fault analysis software. The fault analysis capability depends heavily on the technical personnel's analytical skills, resulting in low reliability, low efficiency, and a long processing time. Furthermore, it lacks effective fault and operational status prediction. In the event of a serious breakdown (the train stops on the main line due to a fault), the lack of real-time and reliable fault diagnosis and analysis often necessitates emergency rescue. Even train drivers with varying experience often struggle to make the correct decisions immediately, lacking guidance, wasting time and effort, and increasing uncertainty and risk. Simultaneously, the lack of real-time and accurate predictive analysis of the overall vehicle equipment status prevents two-way linkage between faults and status, making it impossible to predict faults in advance. Summary of the Invention
[0004] This invention provides an intelligent device for TCMS to overcome the technical problems that trains cannot reliably analyze TCMS faults in real time, cannot effectively predict TCMS operating status, and cannot link train faults and operating status in both directions.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A smart device for TCMS, connected to TCMS, includes: a main control communication unit, a backplane interconnection unit, a fault analysis unit, and a predictive analysis unit; The main control communication unit is used to access the train Ethernet bus, interact with the train Ethernet switch CS, central control unit CCU and event recorder ERM of the train control and monitoring system, and transmit multi-source asynchronous TCMS data to the fault analysis unit through the backplane interconnection unit. The backplane interconnection unit is used to construct an internal local area Ethernet and connect the main control communication unit, fault analysis unit, predictive analysis unit and power management module. The fault analysis unit is used to collect multi-source asynchronous TCMS data through the main control communication unit according to the set fault diagnosis trigger conditions, and receive the collected multi-source asynchronous TCMS data through the backplane interconnection unit. The multi-source asynchronous TCMS data is processed through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated first feature vector. The encapsulated first feature vector is then analyzed through the built-in train fault detection model to output the initial fault conclusion. The predictive analysis unit is used to determine the locomotive operation scenario. Based on the operation scenario determination result, it collects multi-source asynchronous TCMS data through the main control communication unit, processes the multi-source asynchronous TCMS data through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated second feature vector, and performs state prediction on the encapsulated second feature vector through the built-in train prediction model to output the initial prediction conclusion. The fault analysis unit and the predictive analysis unit establish a bidirectional linkage data channel through the backplane interconnection unit. When a predetermined trigger condition is met, the fault analysis unit sends the initial fault conclusion to the predictive analysis unit, and the predictive analysis unit sends the initial prediction conclusion to the fault analysis unit. The fault analysis unit updates its own model parameters based on the received initial prediction conclusion and re-executes the fault analysis, outputting the final fault conclusion. The predictive analysis unit updates its own model parameters based on the received initial fault conclusion and recalculates, outputting the final prediction conclusion.
[0006] Furthermore, the bidirectional linkage data channel includes a fault-to-prediction linkage channel, a prediction-to-fault linkage channel, and a bidirectional fusion reinforcement learning channel; The fault-to-prediction linkage channel is: When the fault analysis unit completes the fault analysis, it outputs the initial fault conclusion to the predictive analysis unit in real time through the backplane interconnection unit. After receiving the initial fault conclusion, the predictive analysis unit adjusts the predictive feature weights, health warning thresholds, or predictive analysis frequencies of the fault-related components or functional modules based on the fault conclusion. The prediction-to-fault linkage channel is: When the predictive analysis unit predicts that the health of any subsystem is lower than the health warning threshold or identifies that the monitoring data deviates from the preset normal operating condition baseline, it sends the high-risk prediction conclusion to the fault analysis unit. The fault analysis unit adds the monitoring features associated with the high-risk prediction conclusion to the real-time monitoring list, lowers the fault trigger threshold, and increases the priority of the diagnostic path for the fault mode indicated by the high-risk prediction conclusion, based on the high-risk prediction conclusion. The monitoring features associated with the high-risk prediction conclusion are determined through a preset fault mode and feature mapping table. The bidirectional fusion reinforcement learning channel is: After completing a fault-to-prediction linkage or prediction-to-fault linkage, a consistency check is performed to compare whether the subsystem where the fault actually occurred is consistent with the subsystem pointed to by the predicted anomaly. If they match, a positive reward signal is generated, and the reward signal is used to increase the weight of the monitoring feature corresponding to the current fault mode in the fault mode-feature mapping table. If there is a discrepancy, a negative penalty signal is generated, and a rule base optimization process is triggered when a predetermined number of negative signals are accumulated. The rule base optimization process is as follows: Call the predefined rule mining algorithm to search for new feature association patterns that have not yet been included in the rule base in the fault cases corresponding to negative signals from the historical ERM data, generate candidate rules, and add them to the candidate domain feature rule base after confirmation by maintenance personnel. After each consistency check is completed, the feature vectors, model outputs and actual maintenance feedback of this analysis are encapsulated into a training data package and written into the ERM dedicated model training partition, thus completing a data loop from fault triggering to training data archiving. When the first predetermined number of data loops is completed, an incremental model training is triggered. When the second predetermined number of data loops is completed and the rule base is updated, the confirmed valid candidate rules are formally written into the domain feature rule base.
[0007] Furthermore, the bidirectional linkage data channel also includes: Fault-to-prediction linkage extension operation: In the next prediction cycle, the prediction analysis unit uses the prediction model to perform prediction inference with parameter adjustment based on the adjusted feature weights, warning thresholds and sampling frequency to determine whether the fault has predictable fault precursor characteristics. If it finds that the characteristics of the predetermined period before the fault occurs show a monotonically deteriorating trend, a fault predictability analysis report is generated and sent to the prediction display screen. Prediction-to-fault linkage extended operation: In the next parking window after predicting an anomaly, the fault analysis unit sends a diagnostic request instruction to the CCU, requesting the CCU to perform an online self-check procedure on the predicted anomaly subsystem to obtain health status information in order to verify and correct the accuracy of the prediction model.
[0008] Furthermore, the fault analysis unit also includes a scenario-linked data retrieval strategy module and a model dynamic scheduling module; The operational scenario linkage data retrieval strategy module is used to read the train operation status code and speed signal to determine the current operational scenario when it receives fault information from the CCU through the main control communication unit. The fault information includes a fault code and a fault timestamp. If the current train speed is zero, a retrieval command is sent to the ERM via the Ethernet bus to request all recorded data within a predetermined time range before and after the time of the fault. If the current train speed is not zero, the fault information will be cached and the fault time will be written to the local queue. When the train stops again, the fault information will be retrieved according to the operation when the speed is zero. The model dynamic scheduling module is used to query the preset scheduling strategy table according to the current operating scenario before sending the encapsulated analysis request frame to the internal Ethernet switch, determine the computing resources and operating mode required for this analysis, and send a wake-up command and expected analysis load information to the corresponding computing board. If the shared data buffer between the fault analysis computing board and the predictive analysis computing board in the backplane interconnection unit is occupied by the predictive analysis computing board, the shared data buffer is allocated according to a preset priority arbitration rule. The allocation involves the fault analysis unit issuing a resource preemption instruction to the backplane interconnection unit to switch the access right of the shared data buffer to the fault analysis computing board. The priority arbitration rule is that the priority of the fault analysis task is higher than that of the predictive analysis task. The fault analysis computing board is used to perform fault analysis in the fault analysis unit, and the predictive analysis computing board is used to perform predictive analysis in the predictive analysis unit.
[0009] Furthermore, the predictive analysis unit also includes a scenario determination and sampling strategy selection module and a differentiated data reading module; The operation scenario determination and sampling strategy selection module is used to continuously monitor the train operation status code, perform scenario determination at preset intervals, query the preset scheduling strategy table according to the current operation scenario, and determine the sampling frequency, data range and model resource allocation for predictive analysis; the operation scenarios include, but are not limited to, high-speed cruise, braking, coasting and stopping; The determined sampling frequency, data range, and model resource allocation parameters are sent to the differentiated data reading module for processing. The differentiated data reading module is used to read differentiated data from the ERM backup Ethernet interface according to the received sampling frequency and data range parameters, specifically: When the sampling frequency corresponding to the running scenario is high frequency, the DMA burst read mode is used to read data; when the sampling frequency corresponding to the running scenario is medium frequency, the periodic polling mode is used to read data; when the sampling frequency corresponding to the running scenario is low frequency, the on-demand query mode is used to read only the data of the specified channel; the differentiated data reading module sends the read raw data to the real-time data preprocessing and feature engineering pipeline.
[0010] Furthermore, the fault analysis unit is also configured to perform the following operations after completing the fault analysis: The fault analysis conclusions and fault solutions are sent to the fault analysis display screen via the internal Ethernet for real-time display. The driver's driving suggestions are returned to the CCU via the original path. The CCU then performs a safety check based on the current vehicle status and sends the suggestions to the DDU for the driver's reference. The predictive analysis unit is further configured to perform the following operations after completing the state prediction: The predictive analysis results are sent to the predictive analysis display screen for real-time display. The system receives actual maintenance results from maintenance personnel via the predictive analysis display screen interface, and uses this feedback data as a reinforcement learning signal to optimize the confidence calibration of the predictive model.
[0011] Furthermore, the real-time data preprocessing and feature engineering pipeline includes: a multi-source asynchronous data alignment module, a data quality adaptive cleaning module, a missing data compensation module, a domain feature automatic extraction module, and a data encapsulation module; The multi-source asynchronous data alignment module is used to time-align data from different subsystems and with different sampling rates within a configured fixed time window using a hybrid algorithm of linear interpolation and spline interpolation, and output structured data frames with a unified time reference to the data quality adaptive cleaning module. The data quality adaptive cleaning module is used to receive the structured data frame, correct repairable anomalies using a forward and backward weighted smoothing algorithm, generate data quality labels for unrepairable anomalies, and send the cleaned data frame and data quality labels to the missing data compensation module. The missing data compensation module is used to receive cleaned data frames and data quality labels, use forward prediction compensation based on historical trends for data channels with a packet loss rate less than a preset ratio, generate warning labels for data channels with a packet loss rate greater than or equal to a preset ratio and skip the channel, and send the compensated data frames and data quality labels to the domain feature automatic extraction module. The automatic domain feature extraction module is used to receive the compensated data frame and data quality label, calculate the domain feature vector in real time based on the preset domain feature rule library, and the domain feature vector includes, but is not limited to, traction motor current fluctuation entropy value, brake cylinder pressure change gradient, pantograph contact force dynamic deviation, door travel curve integral deviation and air conditioning compressor start-stop frequency variation coefficient, and send the domain feature vector to the data encapsulation module. The data encapsulation module is used to encapsulate feature vectors, fault codes, data quality tags, and timestamps into standardized analysis request frames, and output the encapsulated first feature vector or second feature vector.
[0012] Beneficial Effects: This invention provides an intelligent device for TCMS (Train Control System). By setting up an independent fault analysis unit and incorporating a dedicated real-time data preprocessing and feature engineering pipeline and a train fault detection model, it achieves real-time acquisition, online preprocessing, and rapid inference analysis of multi-source asynchronous data from TCMS. This eliminates the need for reliance on ground servers or manual offline analysis, significantly shortening the response time from fault occurrence to fault conclusion output and improving the real-time performance and reliability of fault analysis. Furthermore, by setting up an independent predictive analysis unit and triggering data acquisition and predictive analysis based on dynamic judgment results of train operation scenarios, it achieves online, scenario-adaptive state prediction of TCMS operation status. This enables early identification of subsystem health decline trends before faults occur, solving the problem of insufficient effective prediction methods and inability to provide early warnings in existing technologies, thus improving train operation safety and proactive maintenance. Finally, by establishing a bidirectional linkage data channel between the fault analysis unit and the predictive analysis unit, real-time interaction and mutual correction of fault analysis results and state prediction results are achieved. On the one hand, fault analysis conclusions can guide the prediction model to adjust feature weights, warning thresholds, and prediction frequencies, making the prediction results more closely match the current fault situation. On the other hand, prediction conclusions can guide the fault analysis model to optimize diagnostic paths and monitoring priorities, making fault analysis more forward-looking. The two are no longer isolated modules, but rather form a mutually reinforcing closed-loop collaborative mechanism, solving the problem of the separation and inability to link fault analysis and state prediction in existing technologies, and comprehensively improving the device's integrated analysis and early warning capabilities. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1A topology diagram of a smart device for TCMS provided by the present invention; Figure 2 This is a schematic diagram illustrating the real-time analysis and working steps for train TCMS faults provided by the present invention. Figure 3 This is a schematic diagram illustrating the real-time predictive analysis and working steps for train TCMS faults provided by the present invention. Figure 4 This is a schematic diagram illustrating the real-time two-way linkage and working steps of fault detection and prediction in train TCMS. Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] This embodiment provides a smart device for TCMS, which is connected to TCMS, such as... Figure 1 As shown, it includes: a main control communication unit, a backplane interconnection unit, a fault analysis unit, and a predictive analysis unit; The main control communication unit is used to access the train Ethernet bus, interact with the train Ethernet switch CS, central control unit CCU and event recorder ERM of the train control and monitoring system, and transmit multi-source asynchronous TCMS data to the fault analysis unit through the backplane interconnection unit. The backplane interconnection unit is used to construct an internal local area Ethernet and connect the main control communication unit, fault analysis unit, predictive analysis unit and power management module. Specifically, the device of the present invention mainly consists of three parts, such as Figure 1 As shown, it includes a 3U40TE main unit chassis, a fault analysis display screen, and a predictive analysis display screen; The 3U40TE chassis consists of a power supply board, main control board, computing board, Ethernet switch, backplane, and 3U enclosure. The power supply board has a DC 110V input and DC 12V output, employing a redundant A / B power supply configuration. If one power supply fails, the other can still ensure normal system operation. The main control board uses a 12V backplane power supply and supports one independent Ethernet communication interface on both the front and back panels. The computing board also uses a 12V backplane power supply and supports one independent Ethernet communication interface on both the front and back panels. It supports 7B lightweight industry-specific models for trains, enabling local deployment on edge devices without relying on cloud computing power. It is suitable for real-time data processing at the field (such as local fault prediction and equipment health prediction analysis in TCMS), balancing model capabilities with practical engineering applications. The Ethernet switch is an 8-port industrial gigabit Ethernet switch, connecting the A / B computing boards and the A / B main control boards via the backplane to form an Ethernet local area network, enabling interconnection between any node devices within the network. Both the fault analysis display and the predictive analysis display are 15.6-inch touch screens, powered by DC110V, and support one Ethernet port, enabling real-time fault analysis and predictive analysis.
[0017] The fault analysis unit is used to collect multi-source asynchronous TCMS data through the main control communication unit according to the set fault diagnosis trigger conditions, and receive the collected multi-source asynchronous TCMS data through the backplane interconnection unit. The multi-source asynchronous TCMS data is processed through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated first feature vector. The encapsulated first feature vector is then analyzed through the built-in train fault detection model to output the initial fault conclusion. In this embodiment, the fault analysis unit includes a first main control board (main control board B) and a first computing board (computing board B). The first main control board also includes a running scenario linkage data retrieval strategy module, and the first computing board includes a model dynamic scheduling module. The operational scenario linkage data retrieval strategy module is used to read the train operation status code and speed signal to determine the current operational scenario when it receives fault information from the CCU through the main control communication unit. The fault information includes a fault code and a fault timestamp. If the current train speed is zero, a retrieval command is sent to the ERM via the Ethernet bus to request all recorded data within a predetermined time range before and after the time of the fault. If the current train speed is not zero, the fault information will be cached and the fault time will be written to the local queue. When the train stops again, the fault information will be retrieved according to the operation when the speed is zero. The operation scenario linkage data retrieval strategy module is also configured with an emergency fault exception channel: when the fault code in the fault information belongs to a preset emergency fault category, regardless of whether the current vehicle speed is zero, the operation scenario linkage data retrieval strategy module immediately sends a retrieval command to the ERM via the Ethernet bus to request all recorded data within a predetermined time range before and after the fault time; at the same time, the fault information and retrieval status are written to the local log.
[0018] The emergency fault categories are defined by a preset fault classification table, and include at least fault types that directly affect driving safety, such as loss of traction, brake failure, unexpected opening of doors, fire alarm, and ATP / ATO communication interruption. The model dynamic scheduling module is used to query the preset scheduling strategy table according to the current operating scenario before sending the encapsulated analysis request frame to the internal Ethernet switch, determine the computing resources and operating mode required for this analysis, and send a wake-up command and expected analysis load information to the corresponding computing board. The fault analysis unit is also configured to perform the following operations after completing the fault analysis: The fault analysis conclusions and fault solutions are sent to the fault analysis display screen via the internal Ethernet for real-time display. The driver's driving suggestions are returned to the CCU via the original path. The CCU then performs a safety check based on the current vehicle status and sends the suggestions to the DDU for the driver's reference. Specifically, such as Figure 2 As shown, the specific steps for each module and the steps for fault diagnosis are as follows: Step 1 (Fault Trigger): When the train CCU detects a TCMS fault, it sends the fault code and fault timestamp through the onboard Ethernet bus, which takes about 50ms.
[0019] Step 2 (Fault Forwarding): The train Ethernet switch CS forwards the fault information to the main control board B of the intelligent device, which takes microseconds.
[0020] Step 3 (Determining the Scenario and Selecting the Data Retrieval Strategy): Upon receiving the fault information, main control board B immediately executes its unique scene-linked data retrieval strategy: Sub-step 3.1: Read the train operation status code and speed signal broadcast by the CCU; Sub-step 3.2: Determine the current operating scenario (high-speed cruise / braking / stop / ramp, etc.) based on the status code; Sub-step 3.3: If the vehicle speed is 0 (stopped at the station or in the depot), immediately send a retrieval command to the ERM via the Ethernet bus to request all recorded data from 1 hour before and 10 minutes after the fault (total data size is approximately 200MB). Sub-step 3.4: If the train speed is not equal to 0 (during operation), data is not retrieved temporarily; only the fault information is cached, and the time of the fault occurrence is written to the local queue. After the train stops again (when the speed changes from non-zero to zero), a delayed retrieval is automatically triggered, and data is retrieved in the manner described in 3.3. This design ensures that ERM recording continues uninterrupted during operation, while also ensuring that fault data is not missed.
[0021] Compared to existing solutions that only perform simple "parking download", this step achieves adaptive retrieval of both running and parking states and is linked with the running scenario.
[0022] Step 4 (ERM Data Feedback): After receiving the instruction, ERM forwards the relevant record data to the main control board B via CS, which takes about 400ms.
[0023] Step 5: After receiving the ERM record data, the main control board B does not forward it directly to the computing board, but executes a dedicated, hard real-time data preprocessing and feature engineering pipeline to obtain the encapsulated analysis request frame. Step 6 (Model Dynamic Scheduling): If the shared data buffer of the fault analysis computing board and the predictive analysis computing board in the backplane interconnection unit is occupied by the predictive analysis computing board, the shared data buffer is allocated according to the preset priority arbitration rule. The allocation involves the fault analysis unit issuing a resource preemption instruction to the backplane interconnection unit and notifying the predictive analysis computing board to save its current processing state before switching, thus transferring the access right of the shared data buffer to the fault analysis computing board. The priority arbitration rule is that the priority of the fault analysis task is higher than that of the predictive analysis task. The fault analysis computing board is used to perform fault analysis in the fault analysis unit, and the predictive analysis computing board is used to perform predictive analysis in the predictive analysis unit. Specifically, before sending the encapsulated analysis request frame to the internal switch of the intelligent device, the main control board B first calls the scheduler to make resource decisions. Based on the determined vehicle speed, it consults the scheduling policy table to determine the computing board resources (full speed / high speed / medium speed / low speed mode) that need to be activated for this analysis, sends a wake-up command to the computing board B, and informs it of the expected analysis load. If the shared resources of the computing board B are currently occupied by other tasks, it decides whether to preempt the resources according to the priority arbitration rules (fault analysis priority > predictive analysis priority). In this invention, computing board A and computing board B are heterogeneous computing carriers dedicated to the predictive analysis unit and the fault analysis unit, respectively. Each has an independent AI inference core and local storage, and their core computing power is not shared. However, they share internal Ethernet bandwidth, ERM access channels, and system power consumption budget through a backplane interconnect module, which may lead to system-level resource competition. The computing power resources refer to the comprehensive abstraction of the computing power, storage bandwidth, communication channels, and power consumption quotas that the computing board can provide, and are divided by operating modes such as full speed, high speed, medium speed, and low speed. The analysis load refers to the resource requirements description of the model size, input feature dimension, preprocessing complexity, and expected inference time corresponding to a single fault analysis or state prediction task. The scheduler determines the computing board operating mode and sends a wake-up command based on the operating scenario and analysis load. When shared resources compete, coordination is carried out according to the arbitration rule that fault analysis has a higher priority than predictive analysis.
[0024] Step 7: The Ethernet switch inside the intelligent device forwards the analysis request frame to the computing board B, which takes microseconds. Then, step 8, fault diagnosis, is performed. After receiving the analysis request frame, computing board B extracts the domain feature vector (instead of the original data) and inputs it into the train fault detection model deployed on computing board B for inference, providing a fault root cause analysis conclusion (accurate to specific subsystems and fault modes). Fault classification (levels 1-4, level 1 for immediate stop, level 4 for negligible). Inference time is approximately 3000ms. The feature dimension of the model input is reduced from >200 in the original data to 18, the number of model parameters is compressed from 7B to 1.2B, and the inference speed is improved by 5 times. Specifically, the train fault detection model focuses on classification and root cause localization. The input is the preprocessed first feature vector, and the output is the fault mode and a level classification of 1-4, which can be selected as follows: Lightweight deep classification networks: such as convolutional networks based on MobileNet, ShuffleNet, and ResNet-18 (with channel pruning), which are suitable for identifying discrete fault modes of subsystems such as traction and braking from time slices or statistical features.
[0025] Lightweight temporal / recurrent networks, such as compressed LSTM, GRU, BiLSTM-Attention, or CNN-BiLSTM combination models, are used to capture short-term temporal dependencies before and after a fault occurs, and are suitable for fault mapping of multi-source asynchronous data on MVB / ETH buses.
[0026] Shallow machine learning models (lightweight deployment): such as pruned BP neural networks, SVM (Support Vector Machine), and compact versions of random forests, combined with 18-dimensional features input from a domain feature rule base, have extremely low latency and high interpretability at the edge (often used for mapping fault codes to root causes).
[0027] Lightweight vision-temporal fusion models, such as LAD-YOLO (a YOLO variant with knowledge distillation) or FastSAM variant with lightweight attention, can be used to locate visual faults such as abnormal pantograph-catenary contact if fault analysis involves vehicle-mounted images (such as pantograph video monitoring linkage). The above models all belong to existing technologies in the fields of artificial intelligence and rail transit fault diagnosis. Those skilled in the art can flexibly select or combine the above lightweight models according to the hardware specifications of the on-board computing board (such as memory and power consumption), the fault mechanisms and data sampling rates of each subsystem of TCMS (traction / braking / pantograph, etc.), and achieve an inference time of about 3000ms and a 5-fold speedup effect through pre-training fine-tuning, quantization (INT8 / FP16), etc. Therefore, we will not elaborate further on their internal neuron structure and training hyperparameters. After diagnosis, proceed to steps nine (distributing results to the fault display screen) and ten (driving suggestion feedback closed loop): Computing board B sends the initial fault analysis conclusions and fault solutions to the fault analysis display screen via internal Ethernet, which is displayed in real time, taking about 50ms. Simultaneously, computing board B returns the driver's driving suggestions via the original path (computing board B → switch → main control board B → train CS → CCU), taking approximately 200ms. After the CCU performs a safety check based on the current vehicle status, it sends the final suggestions to the DDU for the driver's reference.
[0028] Step 11 (Linkage Trigger): After the computing board B completes the fault analysis, it triggers the fault → prediction bidirectional linkage, and sends the fault analysis results to the computing board A via the backplane Ethernet in about 100ms, which is used for subsequent linkage. Step 12 (Data Archiving): Simultaneously, computing board B encapsulates the complete data of this analysis (original feature vectors, model output, driver's subsequent operation results, etc.) into a training data package and writes it into the ERM's dedicated model training data partition for subsequent vehicle-to-ground closed-loop iteration. From the occurrence of the fault to the delivery of driving advice to the driver, it takes approximately 4.5 seconds; from the completion of the fault analysis, it takes approximately 4.0 seconds.
[0029] The predictive analysis unit is used to determine the locomotive operation scenario. Based on the operation scenario determination result, it collects multi-source asynchronous TCMS data through the main control communication unit, processes the multi-source asynchronous TCMS data through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated second feature vector, and performs state prediction on the encapsulated second feature vector through the built-in train prediction model to output the initial prediction conclusion. The predictive analysis unit also includes a scenario determination and sampling strategy selection module and a differential data reading module; The operation scenario determination and sampling strategy selection module is used to continuously monitor the train operation status code, perform scenario determination at preset intervals, query the preset scheduling strategy table according to the current operation scenario, and determine the sampling frequency, data range and model resource allocation for predictive analysis; the operation scenarios include, but are not limited to, high-speed cruise, braking, coasting and stopping; The determined sampling frequency, data range, and model resource allocation parameters are sent to the differentiated data reading module for processing. The differentiated data reading module is used to read differentiated data from the ERM backup Ethernet interface according to the received sampling frequency and data range parameters, specifically: When the sampling frequency corresponding to the running scenario is high, the DMA burst read mode is used to read data; when the sampling frequency corresponding to the running scenario is medium, the periodic polling mode is used to read data; when the sampling frequency corresponding to the running scenario is low, the on-demand query mode is used to read only the data of the specified channel; the differentiated data reading module sends the read raw data to the real-time data preprocessing and feature engineering pipeline; The predictive analysis unit is further configured to perform the following operations after completing the state prediction: The predictive analysis results are sent to the predictive analysis display screen for real-time display. The system receives actual maintenance results from maintenance personnel via the predictive analysis display screen interface, and uses this feedback data as a reinforcement learning signal to optimize the confidence calibration of the predictive model.
[0030] Specifically, such as Figure 3 As shown, in order not to affect vehicle operation or occupy Ethernet bus bandwidth, the main control board A of the predictive analysis unit directly reads ERM data from the ERM's spare Ethernet interface. The modules and prediction steps are as follows: Step 1 (Operation Scenario Determination and Sampling Strategy Selection): Main control board A continuously monitors the train's operation status codes, performing a scenario determination every 100ms. Based on the current operation scenario (high-speed cruise / braking / stopping, etc.), it consults the scheduling strategy table to dynamically determine the sampling frequency, data range, and model resources for this round of predictive analysis. If scenario = high-speed cruise: sampling frequency = 100Hz, data range = traction + braking + pantograph, model operation mode = full speed. If scenario = coasting: sampling frequency = 20Hz, data range = running gear bearings + gearbox, model operation mode = low speed. If scenario = stopping: sampling frequency = 5Hz, data range = doors + air conditioning, model operation mode = hibernation (maintenance status self-check only).
[0031] Step 2 (Data Reading Strategy): Based on the strategy determined in Step 1, main control board A directly reads differentiated data from the ERM backup Ethernet interface.
[0032] High-frequency scenario (100Hz): DMA burst read mode is adopted, and the size of a single data block read is 1MB.
[0033] Mid-frequency scenario (20-50Hz): Uses periodic polling mode, with a single read of 256KB.
[0034] Low-frequency scenarios (≤10Hz): Use on-demand query mode to read only the data of the specified channel.
[0035] The total data reading time is dynamically adjusted between 200ms and 1000ms depending on the scenario.
[0036] Step 3 (Data Preprocessing and Feature Engineering): The main control board A performs real-time data preprocessing and feature engineering pipeline on the read data, taking 200ms~500ms to form predicted feature vectors. Typical features include: brake cylinder pressure change gradient (predicting brake system leakage trend), traction motor current fluctuation entropy value (predicting bearing wear trend), etc.
[0037] Step 4 (Model Scheduling and Resource Allocation): Main control board A calls the scheduler, which makes a decision based on the current operating scenario and the load status of computing board A (whether it is being occupied by fault-linked tasks) according to the scheduling policy table: If computing board A is idle, the analysis frequency and computing power allocation weight corresponding to the scenario will be executed.
[0038] If computing board A is occupied by a fault linkage task, the scheduler will arbitrate based on priority: if the priority of predictive analysis is lower than that of fault analysis, the predictive task will be automatically reduced in frequency or delayed in execution.
[0039] Specifically, configuring the analysis frequency and computing power allocation weight according to the operating scenario is a common scenario-adaptive resource management method in the prior art. In this embodiment, the initial weight value is allocated through a scheduling strategy table. Step 5 (Data Forwarding): The internal switch of the intelligent device forwards the predicted feature vector to computing board A, which takes microseconds.
[0040] Step 6 (Model Inference): After receiving the predicted feature vector, computing board A inputs it into the train prediction model for inference and outputs the health score of each subsystem (0-100 points), remaining service life prediction (unit: days / km / running hours), maintenance suggestions (next maintenance time, list of parts to be replaced), and abnormal trend warning (warning 3-5 operating cycles in advance). The inference takes about 3000ms.
[0041] Specifically, the train prediction model can be implemented using lightweight temporal prediction networks known in the art (such as pruned and quantized BiLSTM-Attention, lightweight Transformer, or CNN-BiLSTM, etc.). It is configured to receive the dimensionality-reduced prediction feature vector and output the subsystem health score (0-100 points), remaining service life (days / km / running hours), maintenance suggestions, and abnormal trend warnings through multiple output heads, and achieve an inference time of about 3000ms on the on-board edge computing board. The above model structure, multi-task output design, parameter compression, and edge deployment methods are all existing technologies in industrial prediction and health management systems. Those skilled in the art can select and implement them as needed, and there is no need to describe them further in this invention.
[0042] Step 7 (Result Distribution Prediction Display Screen): Computing board A sends the prediction analysis conclusions to the prediction analysis display screen, which displays the health dashboard and maintenance suggestions in real time, taking about 50ms.
[0043] Step 8 (Linkage Trigger): If a high-risk warning appears in the prediction results (health score < 60 points or remaining lifespan < 30 days), computing board A automatically triggers prediction → two-way fault linkage, sending the high-risk prediction conclusion to computing board B via the backplane Ethernet, and increasing the value of the attention weight of the fault analysis model.
[0044] Step 9 (Data Archiving): Computing board A encapsulates the complete data of this predictive analysis (feature vectors, model outputs, subsequent maintenance and verification results, etc.) into a training data package and writes it to the ERM dedicated partition.
[0045] Step 10 (Asynchronous Feedback Closed Loop): After maintenance personnel complete the maintenance operation based on the predicted suggestions, the actual maintenance result is fed back through the interactive interface of the predictive analysis display screen (e.g., "Health score improved to 95 after replacing brake pads"). This feedback data serves as a positive reinforcement learning signal and is used to optimize the confidence calibration of the predictive model in the next round of iteration training.
[0046] In this embodiment, the prediction and analysis unit includes a second main control board (main control board A) and a second computing board (computing board A). The first and second main control boards are equipped with dedicated, hard real-time data preprocessing and feature engineering pipelines. These pipelines execute before the raw data enters the small model on the computing board, serving as an original bridge connecting the vehicle-mounted raw data and the AI model. The pipeline includes a multi-source asynchronous data alignment module, a data quality adaptive cleaning module, a missing data compensation module, a domain feature automatic extraction module, and a data encapsulation module. The multi-source asynchronous data alignment module is used to time-align data from different subsystems and with different sampling rates within a configured fixed time window using a hybrid algorithm of linear interpolation and spline interpolation, and output structured data frames with a unified time reference to the data quality adaptive cleaning module. Specifically, the A and B main control boards have a built-in dynamic time window aligner that aligns data from different sources and sampling rates, such as CCU (sampling period 10ms), RIOM (sampling period 20ms), traction subsystem (sampling period 1ms), and braking subsystem (sampling period 5ms), within a configurable fixed time window (default 100ms). The aligner uses a hybrid algorithm of linear interpolation and spline interpolation to output a structured data frame with a unified time reference. Data timing unification is a common technical means in data processing. Those skilled in the art can write the corresponding program into the chip of the device according to actual needs, so it will not be described in detail. The data quality adaptive cleaning module is used to receive the structured data frame, correct repairable anomalies using a forward and backward weighted smoothing algorithm, generate data quality labels for unrepairable anomalies, and send the cleaned data frame and data quality labels to the missing data compensation module. Specifically, the A and B main control boards have built-in anomaly detectors that automatically identify and mark abnormal jumps, lost data, or invalid values caused by EMC interference, bus conflicts, poor connector contact, etc. For repairable anomalies (such as single-point jumps), a forward and backward weighted smoothing algorithm is used for correction; for unrepairable anomalies (such as continuous packet loss), data quality labels are generated and passed to the model for reference on data confidence during inference; modifiable anomalies refer to isolated, short-term abrupt dirty points on the timeline, such as single timestamp spikes, single-point range overflows, single-point sign errors, and single-point missing points caused by occasional communication packet loss. The data in the preceding and following neighborhoods are normal and the temporal continuity is not disrupted. The true value can be inferred and backfilled using methods such as forward and backward neighbor weighting, moving average, exponential smoothing, forward and backward Kalman smoothing, and forward and backward sequence prediction weighting. Unrepairable abnormal segments (continuous abnormal segments) refer to data from multiple consecutive periods that are distorted at the same time. Examples include long segments of zero values caused by sensor power failure, segments of constant values caused by bus interruption, temperature drift saturation segments, segments of a channel that are completely out of physical limits for N consecutive seconds, and continuous bad segments with no reliable neighbor references before and after. The above-mentioned anomaly detection and data quality processing methods are common techniques used in train control and monitoring systems and industrial data processing. Specifically: Anomaly detectors are used to automatically identify abnormal transition points, lost data, or invalid values caused by electromagnetic compatibility (EMC) interference, bus communication conflicts, poor connector contact, etc., and are a standard configuration in the data preprocessing stage in this field. Forward and backward weighted smoothing algorithm, as a classic time series anomaly correction method, is widely used in online cleaning and repair of industrial field data; The generation and transmission of data quality labels enable downstream models to refer to data confidence levels during inference, and are also a common practice in this field to ensure the credibility of analysis results under non-ideal data conditions.
[0047] The specific implementation methods of the above-mentioned technical means can be selected and configured by those skilled in the art according to the actual application scenario. They all belong to the prior art and therefore will not be elaborated in detail. The missing data compensation module is used to receive cleaned data frames and data quality labels, use forward prediction compensation based on historical trends for data channels with a packet loss rate less than a preset ratio, generate warning labels for data channels with a packet loss rate greater than or equal to a preset ratio and skip the channel, and send the compensated data frames and data quality labels to the domain feature automatic extraction module. Specifically, for data channels with minor missing data (packet loss rate <5%), forward prediction compensation based on historical trends is used; for data channels with severe missing data (packet loss rate ≥5%), a warning label is generated and the channel is skipped. Forward forecasting compensation based on historical trends, which uses the changing patterns of historical data to fill in short-term minor gaps, is a commonly used interpolation and forecasting method in time series data processing, and therefore will not be described in detail. The automatic domain feature extraction module is used to receive the compensated data frame and data quality label, calculate the domain feature vector in real time based on the preset domain feature rule library, and the domain feature vector includes, but is not limited to, traction motor current fluctuation entropy value, brake cylinder pressure change gradient, pantograph contact force dynamic deviation, door travel curve integral deviation and air conditioning compressor start-stop frequency variation coefficient, and send the domain feature vector to the data encapsulation module. Specifically, the A and B main control boards have a built-in configurable domain feature rule base. This rule base is built based on a knowledge graph of rail transit fault mechanisms and contains more than 50 feature extraction rules validated by industry experts. The main control boards calculate a series of domain feature vectors highly sensitive to fault and health status in real time from the aligned and cleaned raw data, rather than directly feeding the raw telemetry points into the model. Typical features include, but are not limited to: Traction motor current fluctuation entropy value: reflects early mechanical failures such as traction motor bearing wear and rotor bar breakage; Brake cylinder pressure change gradient: to identify braking system response delays, pipeline leaks, or valve sticking trends; Dynamic deviation of pantograph contact force: predicting the risk of poor pantograph-catenary contact and abnormal wear of carbon sliding plate; Door travel curve integral deviation: detects wear or obstruction in the door transmission mechanism; Air conditioner compressor start-stop frequency variation coefficient: indicates a decrease in refrigeration system efficiency or sensor drift; The above-mentioned feature extraction technology based on a domain feature rule base is a commonly used technique in the field of train condition monitoring and fault diagnosis. The construction method of the domain feature rule base, namely, based on the knowledge graph of rail transit fault mechanism and integrating feature extraction rules verified by industry experts, is the common practice of knowledge-driven feature engineering in this field. Calculating domain feature vectors in real time from the aligned and cleaned raw data, instead of directly feeding the raw telemetry points into the model, can effectively reduce the model input dimension, reduce noise interference, and improve inference efficiency and generalization ability. This is also a technical approach widely adopted in industrial AI applications in this field. The typical features listed include the entropy value of traction motor current fluctuation, the gradient of brake cylinder pressure change, the dynamic deviation of pantograph contact force, the integral deviation of door travel curve, and the coefficient of variation of air conditioning compressor start-stop frequency. These are all mature characteristic indicators commonly used in public literature and engineering practice in the field of rail transit to characterize the health status of equipment and early signs of failure. The selection of the above-mentioned feature types, calculation formulas, and specific entries of the rule base can be adapted and extended by those skilled in the art according to actual vehicle models, subsystems, and fault modes, without needing to be further elaborated in this invention; The data encapsulation module is used to encapsulate feature vectors, fault codes, data quality tags, and timestamps into standardized analysis request frames, and output the encapsulated first feature vector or second feature vector.
[0048] The fault analysis unit and the predictive analysis unit establish a bidirectional linkage data channel through the backplane interconnection unit. When a predetermined trigger condition is met, the fault analysis unit sends the initial fault conclusion to the predictive analysis unit, and the predictive analysis unit sends the initial prediction conclusion to the fault analysis unit. The fault analysis unit updates its own model parameters based on the received initial prediction conclusion and re-executes the fault analysis, outputting the final fault conclusion. The predictive analysis unit updates its own model parameters based on the received initial fault conclusion and recalculates, outputting the final prediction conclusion.
[0049] Specifically, such as Figure 4 As shown, due to the presence of the intelligent device backplane, the Ethernet switch, main control boards A and B, and computing boards A and B can be combined into an internal local area Ethernet, enabling interconnection and interoperability between any boards within the network, achieving a bidirectional, peer-to-peer, and mutually reinforcing closed-loop linkage. Specifically: The bidirectional linkage data channel includes a fault-to-prediction linkage channel, a prediction-to-fault linkage channel, and a bidirectional fusion reinforcement learning channel. The fault-to-prediction linkage channel is: When the fault analysis unit completes the fault analysis, it outputs the initial fault conclusion to the predictive analysis unit in real time through the backplane interconnection unit. Specifically, after completing the fault analysis, computing board B proactively broadcasts the fault data to computing board A via the backplane Ethernet without waiting for external requests. This includes fault type codes, fault occurrence times, fault location subsystems, key characteristic anomalies identified during the fault analysis process (such as the traction motor current fluctuation entropy value exceeding 2.5 times the threshold at a certain moment), and finally, a conclusion on the root cause of the fault. After receiving the initial fault conclusion, the predictive analysis unit adjusts the predictive feature weights, health warning thresholds, or predictive analysis frequencies of the fault-related components or functional modules based on the fault conclusion. Specifically, the weight of the predictive features related to the faulty subsystem is increased by 1.5 to 2 times (e.g., for traction motor faults, the weight of "current fluctuation entropy value" is increased from 0.3 to 0.6); the health warning threshold of the faulty subsystem is temporarily tightened from the default 60 points to 75 points to improve the sensitivity of subsequent predictions; and the predictive analysis frequency of the faulty subsystem is increased by one level from the frequency corresponding to the current scenario (e.g., from the mid-frequency 20Hz to the high-frequency 50Hz), and then restored after 24 hours or 100 samplings. The prediction-to-fault linkage channel is: When the predictive analysis unit predicts that the health of any subsystem is lower than the health warning threshold or identifies that the monitoring data deviates from the preset normal operating condition baseline, it sends the high-risk prediction conclusion to the fault analysis unit. Specifically, when computing board A predicts that the health score of any subsystem is less than 60 or an abnormal trend occurs, it actively broadcasts the high-risk prediction conclusion to computing board B via the backplane Ethernet. This includes, but is not limited to, the subsystem predicted to be abnormal, the type of abnormal characteristics, the prediction confidence level, and the prediction time window (the expected failure to occur within the next X days).
[0050] The fault analysis unit adds the monitoring features associated with the high-risk prediction conclusion to the real-time monitoring list, lowers the fault trigger threshold, and increases the priority of the diagnostic path for the fault mode indicated by the high-risk prediction conclusion, based on the high-risk prediction conclusion. The monitoring features associated with the high-risk prediction conclusion are determined through a preset fault mode and feature mapping table. Specifically, the fault analysis unit monitors the monitoring characteristics associated with the high-risk prediction conclusions in real time, even if the CCU fault is not currently triggered; it lowers the fault trigger threshold, for example, temporarily reducing the fault trigger threshold of the predicted abnormal subsystem by 20% (such as reducing the traction motor current fluctuation entropy threshold from 1.0 to 0.8) to achieve early warning; and it increases the priority of the diagnostic path for the fault mode indicated by the high-risk prediction conclusions, for example, within the fault analysis model, it raises the priority of the diagnostic path for the fault mode related to the predicted abnormality to the highest level, accelerating root cause localization when the fault actually occurs.
[0051] The bidirectional fusion reinforcement learning channel is: After completing a fault-to-prediction linkage or prediction-to-fault linkage, a consistency check is performed to compare whether the subsystem where the fault actually occurred is consistent with the subsystem pointed to by the predicted anomaly. If they match, a positive reward signal is generated, and the weight of the monitoring feature corresponding to the current fault mode in the fault mode-feature mapping table is increased using the reward signal; for example, the positive reward signal is multiplied or added to the weight to increase the value of the weight. If there is a discrepancy, a negative penalty signal is generated, and a rule base optimization process is triggered when a predetermined number of negative signals are accumulated. The rule base optimization process is as follows: Call the predefined rule mining algorithm to search for new feature association patterns that have not yet been included in the rule base in the fault cases corresponding to negative signals from the historical ERM data, generate candidate rules, and add them to the candidate domain feature rule base after confirmation by maintenance personnel. Specifically, in this embodiment, the rule mining algorithm uses well-known algorithms in the field of association rule mining (such as the pre-set Apriori algorithm and FP-Growth algorithm, whose minimum support, minimum confidence, and transaction construction method are determined by system configuration). It mines the association patterns between monitoring features from fault case samples marked with negative reward signals in ERM historical data, and generates candidate rules after comparison with the existing rule base to remove duplicates. Positive / negative reward signals are generated based on the consistency between the predicted conclusion and the actual measurement, and the weight of the corresponding monitoring feature in the fault mode-feature mapping table is increased or decreased according to the reward sign. The above algorithm calls, reward and punishment weight adjustment, incremental rule mining, and mapping table lookup methods are all existing technologies in the field of predictive maintenance and fault diagnosis. This invention does not improve them, but only organizes them in the fault-prediction linkage channel of the backplane interconnection. After each consistency verification is completed, the feature vector, model output, and actual maintenance feedback of this analysis are encapsulated into training data packets and written into the ERM dedicated model training partition, completing a data closed loop from fault triggering to training data archiving. When the first predetermined number of data loops is completed, an incremental model training is triggered. When the second predetermined number of data loops is completed and the rule base is updated, the confirmed valid candidate rules are formally written into the domain feature rule base.
[0052] The bidirectional linkage data channel also includes: Fault-to-prediction linkage extension operation: In the next prediction cycle, the prediction analysis unit uses the prediction model to perform prediction inference with parameter adjustment based on the adjusted feature weights, warning thresholds and sampling frequency to determine whether the fault has predictable fault precursor characteristics. If it finds that the characteristics of the predetermined period before the fault occurs show a monotonically deteriorating trend, a fault predictability analysis report is generated and sent to the prediction display screen. In the field of fault prediction and health management (PHM), the following methods are common knowledge: In the next cycle, the prediction unit uses adjusted feature weights, warning thresholds, and sampling frequency as model input parameters, and calls a pre-set prediction model (such as LSTM, VAE-AxLSTM, particle filter, Wiener process model, etc.) to perform predictive inference. The criterion for determining whether a fault is predictable is whether there are precursor features sensitively associated with the fault mode in the cycles preceding the fault. If the health indicators or original monitoring values of the precursor features show a continuous unidirectional deterioration trend within a predetermined cycle, a predictability analysis report containing precursor identifiers, deterioration slopes, and expected warning duration is generated and sent to the host computer / prediction display screen. The aforementioned methods of calling pre-set models, precursor judgment, monotonic degradation identification, and report display in this application are all existing technologies and are not improved; they are simply organized into the fault-prediction linkage channel.
[0053] Specifically, in the next prediction cycle, computing board A uses the adjusted model parameters to perform incremental predictions to determine whether the fault has predictable precursor characteristics. If the characteristics of the fault in the 3-5 cycles before the fault occurs show a monotonically worsening trend, a "Fault Predictability Analysis Report" will be generated and sent to the prediction display screen for maintenance personnel to optimize the prediction rule base. Prediction-to-fault linkage extended operation: In the next parking window after predicting an anomaly, the fault analysis unit sends a diagnostic request instruction to the CCU, requesting the CCU to perform an online self-check procedure on the predicted anomaly subsystem to obtain health status information in order to verify and correct the accuracy of the prediction model.
[0054] Specifically, at the next parking window after the predicted anomaly, the computing board B proactively sends a diagnostic request instruction to the CCU, requesting the CCU to perform an online self-check procedure on the predicted anomaly subsystem to obtain more detailed health status information for verifying and correcting the accuracy of the prediction model.
[0055] The intelligent device of this invention can achieve fault and predictive analysis at the second level, which is dozens of times more efficient than traditional manual fault analysis. At the same time, it can analyze abnormal data before the fault in a very short time, and achieve early fault warning and prediction at the second level.
[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A smart device for TCMS, connected to TCMS, characterized in that, include: Main control communication unit, backplane interconnection unit, fault analysis unit and predictive analysis unit; The main control communication unit is used to access the train Ethernet bus, interact with the train Ethernet switch CS, central control unit CCU and event recorder ERM of the train control and monitoring system, and transmit multi-source asynchronous TCMS data to the fault analysis unit through the backplane interconnection unit. The backplane interconnection unit is used to construct an internal local area Ethernet and connect the main control communication unit, fault analysis unit, predictive analysis unit and power management module. The fault analysis unit is used to collect multi-source asynchronous TCMS data through the main control communication unit according to the set fault diagnosis trigger conditions, and receive the collected multi-source asynchronous TCMS data through the backplane interconnection unit. The multi-source asynchronous TCMS data is processed through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated first feature vector. The encapsulated first feature vector is then analyzed through the built-in train fault detection model to output the initial fault conclusion. The predictive analysis unit is used to determine the locomotive operation scenario. Based on the operation scenario determination result, it collects multi-source asynchronous TCMS data through the main control communication unit, processes the multi-source asynchronous TCMS data through the built-in real-time data preprocessing and feature engineering pipeline to obtain the encapsulated second feature vector, and performs state prediction on the encapsulated second feature vector through the built-in train prediction model to output the initial prediction conclusion. The fault analysis unit and the prediction analysis unit establish a bidirectional linkage data channel through the backplane interconnection unit; the bidirectional linkage data channel is such that when a predetermined trigger condition is met, the fault analysis unit sends the initial fault conclusion to the prediction analysis unit, and the prediction analysis unit sends the initial prediction conclusion to the fault analysis unit. The fault analysis unit updates its own model parameters based on the received initial prediction conclusion and re-executes the fault analysis, outputting the final fault conclusion; the prediction analysis unit updates its own model parameters based on the received initial fault conclusion and recalculates, outputting the final prediction conclusion.
2. The intelligent device for TCMS according to claim 1, characterized in that, The bidirectional linkage data channel includes a fault-to-prediction linkage channel, a prediction-to-fault linkage channel, and a bidirectional fusion reinforcement learning channel. The fault-to-prediction linkage channel is: When the fault analysis unit completes the fault analysis, it outputs the initial fault conclusion to the predictive analysis unit in real time through the backplane interconnection unit. After receiving the initial fault conclusion, the predictive analysis unit adjusts the predictive feature weights, health warning thresholds, or predictive analysis frequencies of the fault-related components or functional modules based on the fault conclusion. The prediction-to-fault linkage channel is: When the predictive analysis unit predicts that the health of any subsystem is lower than the health warning threshold or identifies that the monitoring data deviates from the preset normal operating condition baseline, it sends the high-risk prediction conclusion to the fault analysis unit. The fault analysis unit adds the monitoring features associated with the high-risk prediction conclusion to the real-time monitoring list, lowers the fault trigger threshold, and increases the priority of the diagnostic path for the fault mode indicated by the high-risk prediction conclusion, based on the high-risk prediction conclusion. The monitoring features associated with the high-risk prediction conclusion are determined through a preset fault mode and feature mapping table. The bidirectional fusion reinforcement learning channel is: After completing a fault-to-prediction linkage or prediction-to-fault linkage, a consistency check is performed to compare whether the subsystem where the fault actually occurred is consistent with the subsystem pointed to by the predicted anomaly. If they match, a positive reward signal is generated, and the reward signal is used to increase the weight of the monitoring feature corresponding to the current fault mode in the fault mode-feature mapping table. If there is a discrepancy, a negative penalty signal is generated, and a rule base optimization process is triggered when a predetermined number of negative signals are accumulated. The rule base optimization process is as follows: Call the predefined rule mining algorithm to search for new feature association patterns that have not yet been included in the rule base in the fault cases corresponding to negative signals from the historical ERM data, generate candidate rules, and add them to the candidate domain feature rule base after confirmation by maintenance personnel. After each consistency check is completed, the feature vectors, model outputs and actual maintenance feedback of this analysis are encapsulated into a training data package and written into the ERM dedicated model training partition, thus completing a data loop from fault triggering to training data archiving. When the first predetermined number of data loops is completed, an incremental model training is triggered. When the second predetermined number of data loops is completed and the rule base is updated, the confirmed valid candidate rules are formally written into the domain feature rule base.
3. The intelligent device for TCMS according to claim 1, characterized in that, The bidirectional linkage data channel also includes: Fault-to-prediction linkage extension operation: In the next prediction cycle, the prediction analysis unit uses the prediction model to perform prediction inference with parameter adjustment based on the adjusted feature weights, warning thresholds and sampling frequency to determine whether the fault has predictable fault precursor characteristics. If it is found that the characteristics of the predetermined period before the fault occurs show a monotonically deteriorating trend, a fault predictability analysis report is generated and sent to the prediction display screen. Prediction-to-fault linkage extended operation: In the next parking window after predicting an anomaly, the fault analysis unit sends a diagnostic request instruction to the CCU, requesting the CCU to perform an online self-check procedure on the predicted anomaly subsystem to obtain health status information in order to verify and correct the accuracy of the prediction model.
4. The intelligent device for TCMS according to claim 1, characterized in that, The fault analysis unit also includes a runtime scenario linkage data retrieval strategy module and a model dynamic scheduling module; The operational scenario linkage data retrieval strategy module is used to read the train operation status code and speed signal to determine the current operational scenario when it receives fault information from the CCU through the main control communication unit. The fault information includes a fault code and a fault timestamp. If the current train speed is zero, a retrieval command is sent to the ERM via the Ethernet bus to request all recorded data within a predetermined time range before and after the time of the fault. If the current train speed is not zero, the fault information will be cached and the fault time will be written into the local queue. When the train stops again, the fault information will be retrieved according to the operation when the train speed is zero. The model dynamic scheduling module is used to query the preset scheduling strategy table according to the current operating scenario before sending the encapsulated analysis request frame to the internal Ethernet switch, determine the computing resources and operating mode required for this analysis, and send a wake-up command and expected analysis load information to the corresponding computing board. If the shared data buffer between the fault analysis computing board and the predictive analysis computing board in the backplane interconnection unit is occupied by the predictive analysis computing board, the shared data buffer is allocated according to a preset priority arbitration rule. The allocation involves the fault analysis unit issuing a resource preemption instruction to the backplane interconnection unit to switch the access right of the shared data buffer to the fault analysis computing board. The priority arbitration rule is that the priority of the fault analysis task is higher than that of the predictive analysis task. The fault analysis computing board is used to perform fault analysis in the fault analysis unit, and the predictive analysis computing board is used to perform predictive analysis in the predictive analysis unit.
5. A smart device for TCMS according to claim 1, characterized in that, include: The predictive analysis unit also includes a scenario determination and sampling strategy selection module and a differential data reading module; The operation scenario determination and sampling strategy selection module is used to continuously monitor the train operation status code, perform scenario determination at preset intervals, query the preset scheduling strategy table according to the current operation scenario, and determine the sampling frequency, data range and model resource allocation for predictive analysis; the operation scenarios include, but are not limited to, high-speed cruise, braking, coasting and stopping; The determined sampling frequency, data range, and model resource allocation parameters are sent to the differentiated data reading module for processing. The differentiated data reading module is used to read differentiated data from the ERM backup Ethernet interface according to the received sampling frequency and data range parameters, specifically: When the sampling frequency corresponding to the running scenario is high frequency, the DMA burst read mode is used to read data; when the sampling frequency corresponding to the running scenario is medium frequency, the periodic polling mode is used to read data; when the sampling frequency corresponding to the running scenario is low frequency, the on-demand query mode is used to read only the data of the specified channel; the differentiated data reading module sends the read raw data to the real-time data preprocessing and feature engineering pipeline.
6. A smart device for TCMS according to claim 1, characterized in that, The fault analysis unit is also configured to perform the following operations after completing the fault analysis: The fault analysis conclusions and fault solutions are sent to the fault analysis display screen via the internal Ethernet for real-time display. The driver's driving suggestions are returned to the CCU via the original path. The CCU then performs a safety check based on the current vehicle status and sends the suggestions to the DDU for the driver's reference. The predictive analysis unit is further configured to perform the following operations after completing the state prediction: The predictive analysis results are sent to the predictive analysis display screen for real-time display. The system receives actual maintenance results from maintenance personnel via the predictive analysis display screen interface, and uses this feedback data as a reinforcement learning signal to optimize the confidence calibration of the predictive model.
7. A smart device for TCMS according to claim 1, characterized in that, The real-time data preprocessing and feature engineering pipeline includes: a multi-source asynchronous data alignment module, a data quality adaptive cleaning module, a missing data compensation module, a domain feature automatic extraction module, and a data encapsulation module; The multi-source asynchronous data alignment module is used to time-align data from different subsystems and with different sampling rates within a configured fixed time window using a hybrid algorithm of linear interpolation and spline interpolation, and output structured data frames with a unified time reference to the data quality adaptive cleaning module. The data quality adaptive cleaning module is used to receive the structured data frame, correct repairable anomalies using a forward and backward weighted smoothing algorithm, generate data quality labels for unrepairable anomalies, and send the cleaned data frame and data quality labels to the missing data compensation module. The missing data compensation module is used to receive cleaned data frames and data quality labels, use forward prediction compensation based on historical trends for data channels with a packet loss rate less than a preset ratio, generate warning labels for data channels with a packet loss rate greater than or equal to a preset ratio and skip the channel, and send the compensated data frames and data quality labels to the domain feature automatic extraction module. The automatic domain feature extraction module is used to receive the compensated data frame and data quality label, calculate the domain feature vector in real time based on the preset domain feature rule library, and the domain feature vector includes, but is not limited to, traction motor current fluctuation entropy value, brake cylinder pressure change gradient, pantograph contact force dynamic deviation, door travel curve integral deviation and air conditioning compressor start-stop frequency variation coefficient, and send the domain feature vector to the data encapsulation module. The data encapsulation module is used to encapsulate feature vectors, fault codes, data quality tags, and timestamps into standardized analysis request frames, and output the encapsulated first feature vector or second feature vector.
