Tfds freight car fault image intelligent recognition application system

Abstract

The TFDS (Frequency Detection and Recognition System) intelligent image recognition application system for freight cars relates to a railway freight car fault image recognition system. To address the issue of existing TFDS systems being susceptible to small-scale failures, this invention's intelligent image recognition application system for freight cars is deployed on a centralized recognition service cluster to receive and process fault identification tasks from each detection station. This system, referred to as the TFDS cluster system, includes recognition modules corresponding to all fault identification types. Fault identification tasks are sent to fault identification nodes in the centralized recognition service cluster, which then assign the tasks to the corresponding recognition modules for processing according to the fault identification type. When any service node in the centralized recognition service cluster fails, redundant service nodes load the recognition module based on the failed node's configuration information and modify their own role to that of a service processing node. When the failed node restarts, the redundant service nodes enter a redundancy backup mode.

Description

Technical Field

[0001] This invention relates to a railway freight car fault image recognition system. Background Technology

[0002] By deploying line scan cameras along the trackside, photos of multiple parts of the train as it passes are captured and sent to a server. Advanced deep learning algorithms are then used to identify the faults in the original images, thus preventing train accidents, improving the safety level of railway operations, and reducing the workload of railway staff.

[0003] The TFDS truck fault image intelligent recognition system uses deep learning for fault identification. With the long-term evolution of the system versions, the increasing number of recognition modules, and the expansion of the recognition library, the system's requirements for the operating environment are becoming increasingly stringent. Compared to pursuing hardware upgrades with stronger processing capabilities, distributed, clustered, and horizontally scalable architectures are better directions for system upgrades and iterations. Single-point systems can also be affected by many uncontrollable factors, including physical machine crashes, power outages, and network outages. This not only affects the truck fault detection task but also directly impacts the safe, stable, and effective operation of the TFDS system (Truck Operation Fault Dynamic Image Detection System). Therefore, the current TFDS system needs to improve its availability, ensuring that it can still operate normally without human intervention in occasional small-scale fault scenarios, such as single-node power outages, crashes, or network outages.

[0004] Meanwhile, the current TFDS system also needs to be upgraded and modified in the following aspects:

[0005] 1. Simplify the development of the recognition module, unify the processing of common functions, improve the running performance, enrich the module interface, and support dynamic recognition library upgrades and updates.

[0006] 2. The system platform is designed with a centralized and clustered approach. A single platform handles freight car fault identification for multiple railway stations simultaneously, merging multiple systems into a unified platform. This consolidates business processing capabilities, reuses the hardware and software resources of all distributed stations, improves processing capacity while saving on hardware and software investment, reducing system costs, and simplifying system operation and maintenance. Because all software within the cluster runs in multiple replicas simultaneously, the overall availability of the system is enhanced. Summary of the Invention

[0007] This invention aims to address the problem that existing TFDS systems are susceptible to small-scale failures.

[0008] The TFDS truck fault image intelligent recognition system is applied as follows:

[0009] The central computer room is equipped with a centralized identification service cluster. The intelligent identification system for truck fault images is deployed on the centralized identification service cluster to receive and process fault identification tasks from each detection station. The intelligent identification system for truck fault images deployed on the centralized identification service cluster is referred to as the TFDS cluster system. The TFDS cluster system contains identification modules corresponding to all fault identification types.

[0010] The fault identification task is sent to the fault identification node in the centralized identification service cluster. The identification task is assigned to the corresponding identification module for processing according to the fault identification type. After all identification modules have completed their processing, the identification process collects the processing results and completes the vehicle passing identification task of the current detection station.

[0011] All detection station identification tasks are processed in parallel within a centralized identification service cluster, and all detection stations reuse all physical resources within the cluster.

[0012] The central computer room also has one or more redundant business nodes, which are in standby mode under normal circumstances. When any business node in the centralized identification service cluster fails, the redundant business node enters working mode. The redundant business node loads the identification module according to the configuration information of the failed node and changes its own role to a business processing node. When the failed node restarts, the redundant business node that has changed its own role to a business processing node automatically enters the redundancy backup mode.

[0013] Furthermore, in the centralized identification service cluster, the physical servers corresponding to all nodes are grouped and coded, and the machine codes cannot be duplicated. The physical servers are divided into multiple groups, and each group is an identification group. This group contains all the identification modules, and there are no duplicate identification modules. A group can independently complete the identification of vehicle information and is distributed across multiple physical machines within the group. Multiple groups are in a replication mode, and any two groups are completely identical, with the same number of physical machines. Physical servers with corresponding relationship numbers in different groups have the same identification module installed.

[0014] Furthermore, when the centralized identification service cluster expands, new group codes are set for newly added nodes on the deployment platform.

[0015] Furthermore, each physical machine deploys at least four recognition processes based on the number of GPUs.

[0016] Furthermore, the encoding rules for the physical servers are as follows:

[0017] Any physical machine uses two-part encoding:

[0018] The first segment is Group encoding, using Arabic numerals from 0 to n;

[0019] The second segment is encoded using the serial number of the physical machine within the group, and the characters A through Z are used for encoding.

[0020] Any identification process adds a third segment of encoding on top of the two-segment encoding used in the physical machine; the third segment is the identification process number within the physical machine, initially using encoding from 0 to 3, and continuing to expand as the number of GPU cores increases.

[0021] Furthermore, the fault identification task received and processed by the intelligent truck fault image recognition system is a finely granularly segmented task. When the fault identification task to be identified is finely segmented, it is segmented according to two dimensions.

[0022] The first dimension is the number of carriages in the entire train, that is, the entire train is divided into an identification group of m carriages, and the train is divided into M identification groups;

[0023] The second dimension is the number of fault identification types that the entire train needs to identify. If the number of fault identification types is N, then there are N corresponding identification modules.

[0024] If a train needs to be divided into M identification groups and corresponding to N identification modules, then the fault identification task to be identified needs to be divided into M*N identification subtasks and M*N identification subtasks need to be sent to the message queue.

[0025] Furthermore, after a task is submitted, only one identification module is run at a time, specifying the identification module ID to be used for the identification task. After running the identification ID of that module, the identification process no longer uses other modules within the process to identify the task to be detected.

[0026] Furthermore, the centralized identification service cluster provides a unified protocol interface to the outside world, and the centralized identification service cluster can simultaneously connect to multiple detection stations.

[0027] Furthermore, the process by which the intelligent truck fault image recognition system receives fault identification tasks from each detection station is as follows:

[0028] Each detection station packages the real-time vehicle passing image data collected by the trackside cameras into a single data packet. The image data packet is then sent to the TFDS cluster system access gateway through the network firewall. The access gateway generates multiple fault identification tasks from the vehicle passing image data and then sends them to the nodes in the centralized identification service cluster.

[0029] Furthermore, after all identification modules in the centralized identification service cluster have completed their processing, the identification process collects and returns the processing results to the cluster management and control platform, waiting for the cluster management and control platform to assign the next identification task. When all identification tasks have been processed, a final identification report is generated and reported to the data aggregation and display platform, completing the vehicle recognition task for that detection station.

[0030] Beneficial effects:

[0031] 1. Improve the speed of truck fault identification; multiple identification nodes in the cluster system process fault identification tasks concurrently, and the fault detection time for a single vehicle can be significantly reduced by horizontally scaling the system deployment.

[0032] 2. Improve system reliability; computer system hardware and software failures are generally difficult to avoid. When any failure occurs in the computer system, this invention can tolerate failures to the greatest extent, reduce or even avoid the impact of small-scale failures, and successfully complete train detection tasks. These are important indicators of a train fault identification system. This invention ensures that the identification task can be handled correctly by deploying multiple instances of the identification module, performing concurrent identification, and quickly retrying failed tasks. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the TFDS truck fault image intelligent recognition system.

[0034] Figure 2 A diagram showing the grouping and encoding of all physical servers in the cluster.

[0035] Figure 3 This is a diagram illustrating redundant backup nodes.

[0036] Figure 4 A schematic diagram illustrating the fine-grained breakdown of the identification task. Detailed Implementation

[0037] This invention proposes a deployment scheme for identification modules. All identification modules are deployed on multiple physical servers, forming a centralized identification service cluster. Multiple identification processes run simultaneously on each physical machine, with the number of identification processes corresponding to the number of GPU cores on each machine. Each identification process runs multiple identification modules, and within each process, a service management module also runs. This module communicates with the system platform service, which includes health monitoring, service governance, and heartbeat services for the identification processes. The service management module reports the running status, processing capacity, and throughput of the identification processes and receives platform management signals. All identification modules and processes have multiple replicas running simultaneously within the centralized identification service cluster. The number of replicas corresponds to the number of cluster groups. Each identification group runs a unique identification module. Two identification groups may have identical identification modules and processes. If an identification process in any identification group crashes abnormally, or if a physical machine experiences a network outage, power outage, or other abnormal situation, the same module within the cluster can still handle the identification work for the same type of failure, and subsequent identification services can continue to operate normally. In addition, one or more redundant service nodes are deployed, which are normally in standby mode. If any service server fails, the redundant service node enters working mode, loads the identification module based on the failed node's configuration information, and then changes its role to a service processing node. After the failed node restarts, it automatically enters the redundant backup mode. The following section provides a detailed explanation of the specific implementation method.

[0038] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0039] This embodiment describes a TFDS truck fault image intelligent recognition system, as follows:

[0040] The central computer room is equipped with a centralized identification service cluster. The intelligent identification system for truck fault images is deployed on the centralized identification service cluster to receive and process fault identification tasks from each detection station. The intelligent identification system for truck fault images deployed on the centralized identification service cluster is referred to as the TFDS cluster system. The TFDS cluster system contains identification modules corresponding to all fault identification types.

[0041] The fault identification task is sent to the fault identification node in the centralized identification service cluster. The identification task is assigned to the corresponding identification module for processing according to the fault identification type. After all identification modules have completed their processing, the identification process collects the processing results and completes the vehicle passing identification task of the current detection station.

[0042] All detection station identification tasks are processed in parallel within a centralized identification service cluster, and all detection stations reuse all physical resources within the cluster.

[0043] The central computer room also has one or more redundant business nodes, which are in standby mode under normal circumstances. When any business node in the centralized identification service cluster fails, the redundant business node enters working mode. The redundant business node loads the identification module according to the configuration information of the failed node and changes its own role to a business processing node. When the failed node restarts, the redundant business node that has changed its own role to a business processing node automatically enters the redundancy backup mode. Specific Implementation Method Two:

[0045] This embodiment is a TFDS truck fault image intelligent recognition application system. In the centralized recognition service cluster, the physical servers corresponding to all nodes are grouped and coded, and the machine codes cannot be repeated. The physical servers are divided into multiple groups, and each group is a recognition group. The group contains all the recognition modules, and there are no duplicate recognition modules. A group can independently complete the recognition of vehicle information and is distributed on multiple physical machines within the group. Multiple groups are in a replication mode. Any two groups are completely identical, with the same number of physical machines. Physical servers with corresponding relationship numbers in different groups are equipped with the same recognition module.

[0046] The other modules and processing procedures are the same as in Specific Implementation Method 1. Specific implementation method three:

[0048] This embodiment is a TFDS truck fault image intelligent recognition application system. When the centralized recognition service cluster is expanded, a new group code is set for the newly added nodes on the deployment platform.

[0049] The other modules and processing procedures are the same as in Specific Implementation Method Two. Specific implementation method four:

[0051] This embodiment is a TFDS truck fault image intelligent recognition application system, in which at least four recognition processes are deployed on each physical machine according to the number of GPUs.

[0052] The other modules and processing procedures are the same as in Specific Implementation Method 3. Specific implementation method five:

[0054] This embodiment is a TFDS truck fault image intelligent recognition application system. The encoding rules for the physical servers are as follows:

[0055] Any physical machine uses two-part encoding:

[0056] The first segment is Group encoding, using Arabic numerals from 0 to n;

[0057] The second segment is encoded using the serial number of the physical machine within the group, and the characters A through Z are used for encoding.

[0058] Any identification process adds a third segment of encoding on top of the two-segment encoding used in the physical machine; the third segment is the identification process number within the physical machine, initially using encoding from 0 to 3, and continuing to expand as the number of GPU cores increases.

[0059] The other modules and processing procedures are the same as in Specific Implementation Method Four. Specific implementation method six:

[0061] This embodiment is a TFDS truck fault image intelligent recognition application system. The fault recognition task received and processed by the truck fault image intelligent recognition application system is a task that has been finely divided. When the fault recognition task to be identified is finely divided, it is divided according to two dimensions.

[0062] The first dimension is the number of carriages in the entire train, that is, the entire train is divided into an identification group of m carriages, and the train is divided into M identification groups;

[0063] The second dimension is the number of fault identification types that the entire train needs to identify. If the number of fault identification types is N, then there are N corresponding identification modules.

[0064] If a train needs to be divided into M identification groups and corresponding to N identification modules, then the fault identification task to be identified needs to be divided into M*N identification subtasks and M*N identification subtasks need to be sent to the message queue.

[0065] The other modules and processing procedures are the same as those in one of the specific implementation methods one to five. Specific implementation method seven:

[0067] This embodiment is a TFDS truck fault image intelligent recognition application system. After the task is delivered, only one recognition module is run at a time. The recognition module ID to be recognized is specified. After the recognition ID of the module is run, the recognition process no longer uses other modules in the process to recognize the task to be detected.

[0068] The other modules and processing procedures are the same as those in one of the specific implementation methods one to six. Detailed implementation method eight:

[0070] This embodiment is a TFDS truck fault image intelligent recognition application system. The centralized recognition service cluster provides a unified protocol interface to the outside world, and the centralized recognition service cluster can simultaneously connect to multiple detection stations.

[0071] The other modules and processing procedures are the same as those in any of the specific implementation methods one through seven. Specific implementation method nine:

[0073] This embodiment is a TFDS truck fault image intelligent recognition application system. The process of the truck fault image intelligent recognition application system receiving fault recognition tasks from each detection station is as follows:

[0074] Each detection station packages the real-time vehicle passing image data collected by the trackside cameras into a single data packet. The image data packet is then sent to the TFDS cluster system access gateway through the network firewall. The access gateway generates multiple fault identification tasks from the vehicle passing image data and then sends them to the nodes in the centralized identification service cluster.

[0075] The other modules and processing procedures are the same as those in one of the specific implementation methods one to three. Specific Implementation Method Ten:

[0077] This embodiment is a TFDS truck fault image intelligent recognition application system. After all recognition modules in the centralized recognition service cluster have completed their processing, the recognition process collects the processing results and returns them to the cluster management and control platform, waiting for the cluster management and control platform to assign the next recognition task. When all recognition tasks have been processed, a final recognition report is generated and reported to the data aggregation and display platform, completing the vehicle recognition task of the current detection station.

[0078] The other modules and processing procedures are the same as those in any of the specific implementation methods one through nine.

[0079] Example

[0080] This embodiment of the TFDS truck fault image intelligent recognition system is as follows:

[0081] 1. Deployment Architecture

[0082] like Figure 1 As shown, the intelligent truck fault image recognition application system is deployed on a centralized recognition service cluster in the central computer room. It provides a unified protocol interface to the outside world and waits for the network access of the detection stations. The centralized recognition service cluster can simultaneously access multiple detection stations, receive and process the fault recognition tasks of each detection station. The intelligent truck fault image recognition application system deployed on the centralized recognition service cluster is referred to as the TFDS cluster system. The TFDS cluster system contains recognition modules corresponding to all fault recognition types.

[0083] Each detection station controls the trackside cameras to collect real-time images of passing vehicles. These images are then packaged into a single data packet, which is sent to the TFDS cluster system access gateway via a network firewall. The access gateway generates multiple fault identification tasks from the passing vehicle image data and sends these tasks to the fault identification nodes in the centralized identification service cluster. Through the cluster management and control platform, the fault identification tasks are evenly distributed to the fault identification processes within the nodes. Based on the fault identification type, the tasks are assigned to the identification modules for processing. After all identification modules have completed their processing, the identification process collects the results and returns them to the cluster management and control platform, awaiting the allocation of the next identification task. Once all identification tasks have been processed, a final identification report is generated and uploaded to the data aggregation and display platform, completing the vehicle passing identification task for that detection station.

[0084] All detection station identification tasks are processed in parallel within the centralized identification service cluster. All detection stations reuse all physical resources within the cluster, improving the resource utilization rate of the centralized identification service cluster and saving hardware cost investment.

[0085] 2. Group coding

[0086] like Figure 2 As shown, all physical servers in the cluster are grouped and coded. Machine codes need to be manually assigned by the system administrator during the system deployment phase, ensuring that no duplicates are assigned. When expanding the cluster size later, the system administrator sets new group codes for newly added nodes on the deployment platform. The system administrator first divides the physical servers into multiple groups. Each group is a recognition group, containing all recognition modules without duplicate modules. A group can independently complete the recognition of vehicle information and is distributed across multiple physical machines. Multiple groups operate in a replication mode, with any two groups being completely identical. The number of physical machines is the same, and the machines with corresponding numbers have the same recognition modules installed. Each physical machine deploys four recognition processes based on the number of GPUs, and each process carries multiple recognition modules and lead modules.

[0087] Encoding rules:

[0088] Any physical machine uses two-segment encoding, and any identification process uses three-segment encoding, with segments separated by "-".

[0089] The first segment is Group encoding, using Arabic numerals from 0 to n;

[0090] The second segment is encoded using the serial number of the physical machine within the group, and the characters A through Z are used for encoding.

[0091] The third segment is the identification process number within the physical machine, encoded using 0 to 3. As the number of GPU cores increases, it can be further expanded.

[0092] For example:

[0093] The identification process code for the first physical server in the first group is: 0-A-0.

[0094] The code for the second physical machine in the second group is 1-B.

[0095] 3. Redundancy backup

[0096] To improve system availability, in the event of node failure or severe hardware / software malfunction, redundant backup nodes are designed to ensure that the overall cluster system throughput does not decrease or only decreases slightly. For example... Figure 3 As shown, the cluster has 1 to n redundant service nodes. Under normal conditions, the redundant service nodes do not work, and only the node management service is enabled. When the health monitoring detects that a physical machine has crashed, lost network access, or has an extremely high failure rate in business processing, the health monitoring platform first attempts to issue a shutdown command to the failed node. Then, it assigns the group code information of the failed node to a redundant service node. After receiving the activation command, the redundant service node management service reads the local description file according to the machine code of the failed node to be replaced as indicated in the command, finds the identification module loaded by the failed node, and starts four identification processes. Each identification process loads the module list of the process corresponding to the failed node. Then, it registers its own information with the registry center and takes over subsequent business.

[0097] 4. Task Breakdown

[0098] Without affecting overall recognition performance, the task to be recognized is broken down into finer-grained steps, such as... Figure 4 As shown. When performing fine-grained task decomposition, the first dimension is: only a limited number of passing trains are identified each time. For example, the entire train is divided into identification groups of 8 cars each, and a train can be divided into multiple identification groups, with M identification groups. The second dimension is: the number of fault identification types that need to be identified for the entire train. Let the number of fault identification types be N, then there are N corresponding identification modules. Only one identification module is run each time. When submitting the task, the module ID to be identified is specified. After running the module to identify the ID, the identification process no longer uses other modules in the process to identify the task to be detected, and returns the identification result of the module.

[0099] Suppose a train needs to be divided into M identification groups, and has N identification modules. Then M*N identification subtasks need to be sent to the message queue.

[0100] This fine-grained task partitioning, both theoretically and practically, does not significantly impact the overall monitoring time. The benefits include refined control over the identification process, rapid detection of slow-running identification modules, and timely task transfer to other physical machines for rerunning. Only the modules that timed out are run, and retrying tasks does not impose significant additional business pressure on the cluster. Fine-grained partitioning facilitates load balancing and enhances the parallel processing capabilities of identification, maximizing cluster computing power. In scenarios with low overall load, it significantly improves identification speed, and even under full cluster load, fine-grained partitioning does not impose a noticeable burden on the cluster.

[0101] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. The TFDS truck fault image intelligent recognition application system, characterized in that: The central computer room is equipped with a centralized identification service cluster. The intelligent identification system for truck fault images is deployed on the centralized identification service cluster to receive and process fault identification tasks from each detection station. The intelligent truck fault image recognition application system deployed on the centralized recognition service cluster is referred to as the TFDS cluster system. The TFDS cluster system contains recognition modules corresponding to all fault recognition types. The fault identification task is sent to the fault identification node in the centralized identification service cluster. The identification task is assigned to the corresponding identification module for processing according to the fault identification type. After all identification modules have completed their processing, the identification process collects the processing results and completes the vehicle passing identification task of the current detection station. All detection station identification tasks are processed in parallel within a centralized identification service cluster, and all detection stations reuse all physical resources within the cluster. The central computer room is also equipped with one or more redundant business nodes, which are in standby mode under normal circumstances. When any business node in the centralized identification service cluster fails, the redundant business node enters working mode. The redundant business node loads the identification module according to the configuration information of the failed node and changes its own role to a business processing node. When a failed node restarts, the redundant business node that has changed its role to a business processing node will automatically enter the redundant backup mode. The fault identification task received and processed by the intelligent truck fault image recognition system is a fine-grained task. When the fault identification task to be identified is fine-grained, it is split according to two dimensions. The first dimension is the number of carriages in the entire train, that is, the entire train is divided into an identification group of m carriages, and the train is divided into M identification groups; The second dimension is the number of fault identification types that the entire train needs to identify. If the number of fault identification types is N, then there are N corresponding identification modules. Suppose a train needs to be divided into M identification groups and corresponding to N identification modules. Then the fault identification task to be identified needs to be divided into M*N identification subtasks and M*N identification subtasks need to be sent to the message queue. After a task is submitted, only one identification module is run at a time. The identification module ID to be identified is specified. After the identification ID of the module is run, the identification process will no longer use other modules in the process to identify the task to be detected.

2. The TFDS truck fault image intelligent recognition application system according to claim 1, characterized in that: In a centralized identification service cluster, all physical servers corresponding to nodes are grouped and coded, and machine codes cannot be duplicated. The physical servers are divided into multiple groups, each group being an identification group containing all identification modules, and no duplicate identification modules exist. A group can independently complete the identification of vehicle information and is distributed across multiple physical machines within the group. Multiple groups operate in a replication mode, with any two groups being completely identical, having the same number of physical machines. Physical servers with corresponding relationship numbers in different groups have the same identification module installed.

3. The TFDS truck fault image intelligent recognition application system according to claim 2, characterized in that: When expanding a centralized identification service cluster, new group codes are set for newly added nodes on the deployment platform.

4. The TFDS truck fault image intelligent recognition system according to claim 3, characterized in that: Each physical machine deploys at least four recognition processes based on the number of GPUs.

5. The TFDS truck fault image intelligent recognition system according to claim 4, characterized in that: The encoding rules for all physical servers are as follows: Any physical machine uses two-part encoding: The first segment is Group encoding, using Arabic numerals from 0 to n; The second segment is encoded using the serial number of the physical machine within the group, and uses the characters A to Z for encoding; Any identification process adds a third segment of encoding on top of the two-segment encoding used in the physical machine; the third segment is the identification process number within the physical machine, initially using encoding from 0 to 3, and continuing to expand as the number of GPU cores increases.

6. The TFDS truck fault image intelligent recognition application system according to any one of claims 1 to 5, characterized in that: The centralized identification service cluster provides a unified protocol interface to the outside world, and the centralized identification service cluster can simultaneously connect to multiple detection stations.

7. The TFDS truck fault image intelligent recognition application system according to claim 6, characterized in that: The process by which the intelligent truck fault image recognition system receives fault identification tasks from each detection station is as follows: Each detection station packages the real-time vehicle passing image data collected by the trackside cameras into a single data packet. The image data packet is then sent to the TFDS cluster system access gateway through the network firewall. The access gateway generates multiple fault identification tasks from the vehicle passing image data and then sends them to the nodes in the centralized identification service cluster.

8. The TFDS truck fault image intelligent recognition application system according to claim 7, characterized in that: After all identification modules in the centralized identification service cluster have completed their processing, the identification process collects the processing results and returns them to the cluster management and control platform, waiting for the cluster management and control platform to assign the next identification task. When all identification tasks have been processed, a final identification report is generated and reported to the data aggregation and display platform, completing the vehicle recognition task for that detection station.

Images