Method, device, equipment, medium and system for power coordination scheduling and task fragmentation of heterogeneous vehicle platoon

Abstract

This application provides a method, apparatus, device, medium, and system for collaborative scheduling and task sharding of computing power in heterogeneous vehicle platoons. The method includes: acquiring parking trigger signals sent by vehicle nodes with mobile energy storage and computing platforms; establishing a temporary computing federation group based on the parking trigger signals and extracting hardware specifications containing concurrent throughput dimension information; parsing the operator execution dependencies of the target processing tasks for hierarchical decoupling, generating first and second instruction processing sets; configuring the first instruction processing set as a concurrent processing sub-stream according to throughput requirements; dividing the second instruction processing set into serially associated task shards based on timing constraints; distributing the concurrent sub-streams and task shards to different vehicle nodes according to the specifications; and performing time-series concatenation of the returned local results to generate a global result. This application achieves efficient matching of idle onboard computing power and fine-grained task sharding collaboration.

Description

Technical Field

[0001] This application relates to the fields of distributed computing and vehicle network control technology, and more specifically, to a method, apparatus, device, computer-readable storage medium, and system for collaborative scheduling and task sharding of heterogeneous vehicle formations. Background Technology

[0002] With the evolution of new energy vehicles and intelligent connected vehicle technologies, vehicles are gradually being equipped with mobile energy storage and computing platforms with high computing power (such as onboard neural network acceleration units). In scenarios where multiple vehicles are parked together or parked for extended periods, a large amount of onboard computing power is idle. How to interconnect these distributed, heterogeneous vehicle nodes to cope with large-scale processing tasks with high throughput or complex timing constraints has created an urgent need for a distributed computing power collaborative scheduling mechanism.

[0003] In existing multi-device computing power collaboration solutions, a coarse-grained task distribution architecture based on a fixed central node is typically adopted. This solution first relies on a cloud-based control center to poll the network online status of each node in real time to maintain a static list of available nodes. Then, the control center obtains the tasks to be processed and forcibly divides the tasks into several homogeneous sub-tasks according to a pre-set single task distribution rule. Finally, the control center uniformly distributes these homogeneous sub-tasks to the various hardware nodes in the list and waits for all nodes to complete their independent processing before returning the results for simple plaintext concatenation.

[0004] However, this coarse-grained collaborative scheme based on a fixed central node has significant technical drawbacks. Because it employs static polling and a single equal-division mechanism, it completely ignores the inherent operator dependencies within complex computational tasks and fails to consider the heterogeneous differences in concurrency throughput among the underlying hardware of each node. This leads to data backflow congestion between nodes caused by coupled instructions with timing constraints, while independent instructions requiring high concurrency cannot find suitable throughput resources. Furthermore, this homogeneous equal-division mode is difficult to adapt to the discrete computing power of vehicle nodes during dynamic parking, resulting in low flexibility in the control strategy and an inability to meet the fine-grained task sharding and computing power collaborative scheduling requirements of heterogeneous vehicle platoons in complex task scenarios. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, computer-readable storage medium, and system for computational collaborative scheduling and task sharding of heterogeneous vehicle formations, so as to at least alleviate the above-mentioned technical problems.

[0006] A method for collaborative scheduling and task partitioning of computing power in heterogeneous vehicle platooning includes the following steps: Acquire parking trigger signals sent by multiple vehicle nodes with mobile energy storage and computing platforms; Based on the parking trigger signal, a temporary computing federation group is formed, and the specifications of the hardware acceleration units of each vehicle node in the temporary computing federation group are extracted. The specifications include the concurrent throughput dimension information of the vehicle neural network acceleration unit. Receive the target processing task, parse the operator execution dependency of the target processing task, perform hierarchical decoupling of the target processing task according to the operator execution dependency, and generate a first instruction processing set and a second instruction processing set; Extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream if the concurrent throughput dimension information meets the throughput requirement index. Extract the timing delivery constraints of the second instruction processing set, construct a cascaded execution topology sequence based on the timing delivery constraints, and divide the second instruction processing set into multiple serially associated computing task fragments along the cascaded execution topology sequence; According to the specifications, the concurrent processing subflow is sent to the first vehicle node in the temporary computing federation group, and the multiple serially associated computing task fragments are sent to the second and third vehicle nodes in the temporary computing federation group in sequence. The system receives local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, performs a time-series splicing operation on the local processing results, and generates the global execution result of the target processing task.

[0007] Optionally, the step of establishing a temporary computing federation group based on the parking trigger signal includes: Extract the location coordinate data and network connectivity identifier carried by the parking trigger signal; Spatial clustering is performed on the location coordinate data to extract candidate vehicle nodes that are within the same physical range and whose network connectivity identifier indicates that they are in a connectable state. Obtain a preset security domain identity credential, and use the security domain identity credential to authenticate the candidate vehicle node. Obtain the local area network address of the candidate vehicle node that has passed the identity authentication, and form a network group based on the local area network address to generate the temporary computing federation group.

[0008] Optionally, the step of extracting the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, wherein the specifications include concurrent throughput dimension information of the onboard neural network acceleration unit, includes: Send probe query commands to each vehicle node within the temporary computing federation group; Receive the underlying hardware status logs for the hardware acceleration unit returned by each vehicle node in response to the probe query command. Extract the number of tensor operation cores and available video memory capacity of the vehicle-mounted neural network acceleration unit from the underlying hardware status log; The number of tensor operation cores and the available video memory capacity are formatted and combined to form the concurrent throughput dimension information, and the formatted combination data containing the concurrent throughput dimension information is encapsulated into the specification of the hardware acceleration unit of each vehicle node.

[0009] Optionally, the step of parsing the operator execution dependencies of the target processing task, performing hierarchical decoupling of the target processing task based on the operator execution dependencies, and generating a first instruction processing set and a second instruction processing set includes: The target processing task is parsed using a directed acyclic graph to obtain the operator execution dependencies consisting of nodes and edges; The hierarchical decoupling is performed based on the edge connection state in the operator execution dependency relationship, and independent operator nodes without data dependency edges and coupled operator nodes with data dependency edges are extracted. The independent operator nodes are aggregated to generate the first instruction processing set; Extract the data flow direction indicated by the data dependency edge between the coupled operator nodes, and combine the coupled operator nodes with the same data flow direction in a time sequence to generate the second instruction processing set.

[0010] Optionally, the steps of extracting the timing transit constraints of the second instruction processing set, constructing a cascaded execution topology sequence based on the timing transit constraints, and dividing the second instruction processing set into multiple serially associated computational task fragments along the cascaded execution topology sequence include: Extract the input tensor size of the first instruction data and the output tensor size of the last instruction data from the second instruction processing set, and use the input tensor size and the output tensor size as the timing transmission constraint. Based on the input tensor size and the output tensor size in the timing transit constraints, calculate the activation value overhead parameter when the second instruction processing set is internally transited; Based on the activation value overhead parameter, multiple data communication nodes are inserted into the second instruction processing set, and the second instruction processing set is connected in series using the data communication nodes to construct and generate the cascaded execution topology sequence. Using the data communication node as the dividing point, the second instruction processing set is divided into multiple serially associated computing task fragments along the cascaded execution topology sequence.

[0011] Optionally, the steps of distributing the concurrent processing substream to the first vehicle node in the temporary computing federation group according to the specification, and distributing the multiple serially associated computing task fragments sequentially to the second and third vehicle nodes in the temporary computing federation group, include: Under the condition that the concurrent throughput dimension information in the specification meets the first preset value, candidate nodes with matching computing power are selected from the temporary computing federation group as the first vehicle node, and the concurrent processing subflow is sent to the first vehicle node. Extract the network bandwidth transmission index contained in the specification. Under the condition that the network bandwidth transmission index meets the second preset value, select candidate nodes that meet the transmission requirements from the temporary computing federation group as the second vehicle node and the third vehicle node. According to the serial association order of the cascaded execution topology sequence, the preceding computing task fragments in the multiple serially associated computing task fragments are sent to the second vehicle node, and the remaining computing task fragments are sent to the third vehicle node.

[0012] Optionally, the step of receiving the partial processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node includes: Real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node; If the parking maintenance state is a normal parking state, then the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node are received. If a parking release signal is detected from the first vehicle node, it is determined that the parking maintenance state of the first vehicle node is abnormal, and the data receiving channel of the first vehicle node is blocked. Extract the concurrent processing sub-streams that have not yet been completed in the first vehicle node, redirect and distribute the concurrent processing sub-streams to other vehicle nodes in the temporary computing federation group that are in the normal parking state for continued execution, and receive the local processing results returned by the other vehicle nodes.

[0013] Optionally, the step of real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node includes: Obtain the battery management system interface protocol, and periodically extract the remaining power battery capacity of the first vehicle node, the second vehicle node, and the third vehicle node through the battery management system interface protocol; The remaining power of the power battery is compared and analyzed with the preset power protection lower limit value; If the remaining power of the power battery is not lower than the power protection lower limit value, then the corresponding vehicle node is determined to meet the computing power maintenance condition of the parking maintenance state. If the remaining power of the power battery is lower than the power protection lower limit, it is determined that the corresponding vehicle node does not meet the computing power maintenance condition, a computing power cancellation command is generated and issued, and the corresponding vehicle node is forcibly triggered to exit the parking maintenance state.

[0014] Optionally, the step of performing a time-series concatenation operation on the local processing results to generate the global execution result of the target processing task includes: The output timestamps of the local processing results are aligned, and the time-series splicing operation is performed on the local processing results based on the aligned output timestamps to obtain the initial splicing result. A memory-safe overwrite instruction is issued to the first vehicle node, the second vehicle node, and the third vehicle node to erase the local processing results on their respective localities. Receive the underlying storage erase credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node in response to the memory-secure overwrite instruction; Under the condition that the underlying storage erasure certificate is valid, the initial splicing result is confirmed as the global execution result of the target processing task and output encapsulated.

[0015] The technical advantages of the technical solution provided in this application are: This application's method for collaborative scheduling and task sharding of heterogeneous vehicle platoons addresses the technical shortcomings of traditional coarse-grained task distribution architectures, which ignore operator dependencies and heterogeneous differences in vehicle hardware, leading to data congestion and computing power mismatch. By acquiring parking trigger signals sent by vehicle nodes with mobile energy storage and computing platforms, temporary computing federation groups are formed. Furthermore, specifications containing information on the concurrent throughput of onboard neural network acceleration units are extracted, solving the problems of lagging static polling strategies and the inability to perceive underlying heterogeneous computing power details in traditional solutions. Compared to the traditional method of relying on a static node list maintained in the cloud, this application uses dynamic parking signals for on-demand grouping and delves into the specifications of hardware acceleration units, resulting in a more granular resource profile of idle computing power and laying a solid data-aware foundation for subsequent heterogeneous scheduling.

[0016] Based on the established temporary computing federation group, this application parses the operator execution dependencies of the target processing task for hierarchical decoupling, generating first and second instruction processing sets. Under the condition that the concurrency throughput dimension information meets the throughput requirements, the first instruction processing set is configured as a concurrent processing sub-stream, solving the problem of severe computing power mismatch caused by the traditional single equal-division rule that forcibly splits tasks. Traditional solutions ignore task characteristics and perform homogeneous equal division, while this application extracts independent operators based on the actual underlying dependency logic, and then provides targeted computing power supply according to the actual concurrency throughput dimension of the vehicle nodes. Compared with homogeneous distribution, its computing power adaptability and the processing efficiency of independent concurrent streams are significantly improved.

[0017] Furthermore, by extracting the timing delivery constraints of the second instruction processing set to construct a cascaded execution topology sequence, this sequence is divided into multiple serially associated computational task slices. According to the specifications, concurrent processing sub-streams and serial slices are distributed to different vehicle nodes for processing. Finally, the local processing results are sequentially concatenated, solving the problem of large-scale data backflow waiting that easily occurs when processing coupled instructions in traditional collaborative schemes. Traditional schemes lack a cascaded slicing mechanism, leading to the accumulation and lag of timing tasks. This application, however, specifically constructs a topology sequence for coupled operators with constraints and implements a pipelined serial distribution, allowing complex processing tasks to be smoothly relayed among heterogeneous vehicles. The timing concatenation operation can quickly close the loop and produce global results. Compared to the traditional independent operation and coarse aggregation mode, this greatly avoids data flow conflicts, reliably ensuring the collaborative efficiency of heterogeneous vehicle formations in processing complex tasks. Attached Figure Description

[0018] Figure 1 This application provides an embodiment of a resource federation scheduling and execution verification scenario for cross-domain heterogeneous terminals. Figure 2 This application provides a method for resource federation scheduling and execution verification of cross-domain heterogeneous terminals. Figure 3 This application provides a resource federation scheduling and execution verification device for cross-domain heterogeneous terminals. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This is a computer-readable storage medium according to an embodiment of the present application. Figure 6 This application presents a computational collaborative scheduling and task sharding system for heterogeneous vehicle formations. Detailed Implementation

[0019] like Figure 1 The image shows a scenario of resource federation scheduling and execution verification for cross-domain heterogeneous terminals according to an embodiment of this application; as shown... Figure 2The image shows a resource federation scheduling and execution verification method for cross-domain heterogeneous terminals according to an embodiment of this application. The resource federation scheduling and execution verification method for cross-domain heterogeneous terminals includes the following steps: Acquire parking trigger signals sent by multiple vehicle nodes with mobile energy storage and computing platforms; Based on the parking trigger signal, a temporary computing federation group is formed, and the specifications of the hardware acceleration units of each vehicle node in the temporary computing federation group are extracted. The specifications include the concurrent throughput dimension information of the vehicle neural network acceleration unit. Receive the target processing task, parse the operator execution dependency of the target processing task, perform hierarchical decoupling of the target processing task according to the operator execution dependency, and generate a first instruction processing set and a second instruction processing set; Extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream if the concurrent throughput dimension information meets the throughput requirement index. Extract the timing delivery constraints of the second instruction processing set, construct a cascaded execution topology sequence based on the timing delivery constraints, and divide the second instruction processing set into multiple serially associated computing task fragments along the cascaded execution topology sequence; According to the specifications, the concurrent processing subflow is sent to the first vehicle node in the temporary computing federation group, and the multiple serially associated computing task fragments are sent to the second and third vehicle nodes in the temporary computing federation group in sequence. The system receives local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, performs a time-series splicing operation on the local processing results, and generates the global execution result of the target processing task.

[0020] Optionally, the step of establishing a temporary computing federation group based on the parking trigger signal includes: Extract the location coordinate data and network connectivity identifier carried by the parking trigger signal; Spatial clustering is performed on the location coordinate data to extract candidate vehicle nodes that are within the same physical range and whose network connectivity identifier indicates that they are in a connectable state. Obtain a preset security domain identity credential, and use the security domain identity credential to authenticate the candidate vehicle node. Obtain the local area network address of the candidate vehicle node that has passed the identity authentication, and form a network group based on the local area network address to generate the temporary computing federation group.

[0021] Preferably, the specific implementation process of forming a temporary computing federation group based on the parking trigger signal is as follows: After receiving the parking trigger signal sent by multiple vehicle nodes, the parking trigger signal is first parsed to obtain a parking trigger parsing record; the parking trigger parsing record includes the location coordinate data, the network connectivity identifier, the vehicle node identity digest, and the parking trigger timestamp. The parking trigger timestamp is used to determine whether multiple parking trigger signals belong to the same parking formation window, and the same parking formation window is used to limit the parking trigger time range participating in this network formation; the vehicle node identity digest is used to deduplicate and merge multiple parking trigger signals continuously sent by the same vehicle node to obtain a deduplicated parking trigger parsing record; the location coordinate data and the network connectivity identifier continue to enter the subsequent spatial clustering processing and connectivity status judgment along with the deduplicated parking trigger parsing record. By first creating the parking trigger parsing record, instead of directly adding all online vehicle nodes to the fixed node list, the parking status, spatial location, and connectivity status can be jointly judged within the same data record, thus preventing vehicle nodes that are visible to the network at a distance but cannot form a data transmission channel from entering the temporary computing federation group.

[0022] Preferably, in the specific technical implementation of extracting the location coordinate data carried by the parking trigger signal, the location coordinate data is analyzed to obtain the satellite positioning coordinates reported by the vehicle node, the local coordinates of the parking area, and the location sampling timestamp. Then, according to a pre-configured parking area coordinate reference, the satellite positioning coordinates and the local coordinates of the parking area are normalized to obtain a unified parking spatial coordinate record. The location sampling timestamp is used to verify time consistency with the parking trigger timestamp, ensuring that the unified parking spatial coordinate record corresponds to the parking position within the same parking platoon window. The parking area coordinate reference describes the correspondence between the parking space layout direction, the lane extension direction, and the local area network coverage boundary within the same parking area; the parking space layout direction, the lane extension direction, and the local area network coverage boundary together define the spatial reference direction and communication coverage range in the parking area coordinate reference. The unified parking space coordinate record is used to convert the location coordinate data of different vehicle nodes under different positioning sources into the same spatial reference. When performing spatial clustering processing on the location coordinate data, the unified parking space coordinate record is actually used as the clustering basis. This makes it possible to determine whether vehicle nodes are in the same physical range not only based on rough distance, but also by combining the parking space layout direction, lane extension direction, local area network coverage boundary, and the same parking platoon window in the parking area coordinate reference.

[0023] Preferably, in the specific technical implementation of extracting the network connectivity identifier, the network connectivity identifier is jointly generated by the vehicle node after parking via a local area network (LAN) discovery message, short-range wireless beacon reception status, and LAN address reachability detection result. The LAN discovery message indicates whether the vehicle node is within the same LAN discovery range, the short-range wireless beacon reception status indicates whether the vehicle node is within a short-range wireless coverage range, and the LAN address reachability detection result indicates whether a data transmission channel can be established between vehicle nodes. After parsing the network connectivity identifier, a connectivity confirmation record is obtained; the connectivity confirmation record includes a discovery response status, a beacon continuity status, and an address reachability status. The discovery response status originates from the local area network discovery message, the beacon continuity status originates from the short-range wireless beacon reception status, and the address reachability status originates from the local area network address reachability detection result. When extracting the candidate vehicle nodes, the discovery response status, the beacon continuity status, and the address reachability status in the connectivity confirmation record, together with the unified parking space coordinate record, participate in the judgment to ensure that vehicle nodes located in the same physical range but unable to establish a data transmission channel are not mistakenly selected as candidate vehicle nodes.

[0024] Preferably, when performing spatial clustering processing on the location coordinate data, based on the unified parking spatial coordinate record, multiple vehicle nodes are first mapped to the same parking area partition according to the parking area coordinate reference, and then spatial clustering labels are generated according to the spatial adjacency relationship between the vehicle nodes. The spatial clustering labels are used to indicate whether multiple vehicle nodes are located in the same parking area partition, and further indicate whether multiple vehicle nodes are located in the same parking space layout area, the same local area network coverage area, or the same near-field wireless coverage area. After generating the spatial clustering labels, vehicle nodes with the same spatial clustering labels are jointly filtered with the connectivity confirmation record. Under the condition that the discovery response status, beacon continuity status, and address reachability status in the connectivity confirmation record all indicate a connectable status, the corresponding vehicle node is written into the candidate vehicle node list. The candidate vehicle node list continues to serve as input for identity authentication, so that only vehicle nodes whose spatial clustering labels and connectivity confirmation records both meet the filtering conditions enter the verification process of the security domain identity credential.

[0025] Preferably, the security domain identity credential is configured when vehicle nodes are pre-added to the same open computing power security domain. The security domain identity credential includes a security domain identifier, a vehicle node identity digest, a credential validity period record, a computing power access permission record, and signature verification material. The security domain identifier is used to limit the open computing power security domain to which the candidate vehicle node belongs. The vehicle node identity digest is used for consistency comparison with the vehicle node identity digest in the parking trigger parsing record. The credential validity period record is used to exclude expired security domain identity credentials. The computing power access permission record is used to limit whether the corresponding vehicle node is allowed to participate in the temporary computing federation group. The signature verification material is used to verify the signature response data in subsequent identity authentication. By using the security domain identity credential in correspondence with the candidate vehicle node list, it is possible to avoid adding unknown vehicle nodes to the temporary computing federation group solely based on the connectivity confirmation record.

[0026] Preferably, when authenticating the candidate vehicle nodes using the security domain identity credentials, a security domain authentication request is first sent to each candidate vehicle node in the candidate vehicle node list. This security domain authentication request carries one-time random challenge data and the security domain identifier. Upon receiving the security domain authentication request, the candidate vehicle node signs the one-time random challenge data based on the locally stored security domain identity credentials, forming signature response data, and returns the vehicle node identity digest, the credentials validity time record, and the signature response data. After verifying the vehicle node identity digest, the credentials validity time record, and the signature response data, a security domain authentication record is formed. This security domain authentication record is used to mark whether each candidate vehicle node belongs to the same open computing power security domain, whether it is in a valid credential state, and whether it has open computing power permissions. Only when the security domain authentication record simultaneously satisfies the conditions of consistent vehicle node identity digest, valid credentials validity time record, and successful signature response data verification, is the corresponding candidate vehicle node confirmed as having passed the identity authentication.

[0027] Preferably, when obtaining the local area network (LAN) address of a candidate vehicle node that has passed the identity authentication, the LAN address returned by the candidate vehicle node is subjected to address format verification, address conflict detection, and port reachability probe to obtain a LAN address confirmation record. The address format verification is used to determine whether the LAN address belongs to a pre-configured range of available LAN addresses, which is determined based on the LAN coverage boundary in the parking area coordinate reference. The address conflict detection is used to eliminate intra-group communication routing ambiguities caused by multiple vehicle nodes reporting the same LAN address. The port reachability probe is used to determine whether the LAN address can carry subsequent probe query commands and task data transmission. The LAN address confirmation record includes the verification results of the address format verification, the detection results of the address conflict detection, and the detection results of the port reachability probe. The LAN address confirmation record is further associated with the security domain authentication record, ensuring that the candidate vehicle node that has passed the identity authentication not only meets the verification conditions corresponding to the security domain authentication record but is also in a connectable state on the data transmission channel corresponding to the LAN address confirmation record.

[0028] Preferably, when platooning candidate vehicle nodes authenticated through the identity verification based on the local area network address, a member platooning record is first generated according to the security domain authentication record, the local area network address confirmation record, and the spatial clustering label. The member platooning record includes the vehicle node identity digest, the local area network address, the spatial clustering label, the connectivity confirmation record, and the parking trigger timestamp. After generating the member platooning record, intra-group communication routes between vehicle nodes are established according to the local area network address, and the member platooning record is written into the member registration area of ​​the temporary computing federation group to form and generate the temporary computing federation group. The intra-group communication route is used to carry the issuance of probe query commands and the transmission of task data between vehicle nodes within the temporary computing federation group. The member registration area of ​​the temporary computing federation group is not a fixed list of nodes, but a dynamic range of members jointly defined by the parking trigger signal, the security domain authentication record, and the local area network address confirmation record. The dynamic range of members continues to participate in the issuance of subsequent probe query commands and the transmission of task data through the member registration area of ​​the temporary computing federation group. When extracting the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, probe query commands are only issued to the vehicle nodes in the member registration area of ​​the temporary computing federation group.

[0029] Preferably, after the temporary computing federation group is generated, a federation group lifecycle status record is also generated based on the member queuing record. The federation group lifecycle status record is used to characterize the temporary computing federation group's transition from a pending formation state to a computing power registration pending execution state. The pending formation state corresponds to the state where the member queuing record has not yet been written into the member registration area of ​​the temporary computing federation group, and the computing power registration pending execution state corresponds to the state where the member queuing record has been written into the member registration area of ​​the temporary computing federation group and is ready to issue the probe query command. The computing power registration pending execution state is used to trigger the subsequent extraction of specifications for the hardware acceleration units of each vehicle node, enabling the generation result of the temporary computing federation group to be directly connected to the concurrent throughput dimension information collection of the onboard neural network acceleration unit. If a vehicle node stops returning a LAN reachability response while the computing power registration is pending, the member platoon record is updated based on the LAN address confirmation record, and the vehicle node that stops returning a LAN reachability response is removed from the member registration area of ​​the temporary computing federation group. The LAN reachability response is generated by the vehicle node responding to the port reachability probe based on the intra-group communication route. The LAN reachability response is used to determine whether the vehicle nodes in the member registration area of ​​the temporary computing federation group are still in a connectable state, so as to reduce the impact of the vehicle node that stops returning a LAN reachability response on the subsequent probe query command issuance and task data transmission.

[0030] Preferably, in a multi-vehicle centralized parking application scenario, multiple vehicle nodes in adjacent parking spaces send parking trigger signals after entering the parking state. If the unified parking space coordinate records of some of these vehicle nodes fall into the same parking area partition, and the discovery response status, beacon continuity status, and address reachability status in the connectivity confirmation record all indicate a connectable state, then these vehicle nodes are first written into the candidate vehicle node list. Subsequently, the vehicle nodes in the candidate vehicle node list are authenticated using the security domain identity credentials, and a local area network address confirmation record is generated for the local area network addresses of the candidate vehicle nodes that have passed the authentication. If the verification results of the address format check, the detection results of the address conflict detection, and the detection results of the port reachability probe in the local area network address confirmation record meet the conditions for establishing intra-group communication routes, network grouping is performed based on the local area network address to form the temporary computing federation group. The conditions for establishing intra-group communication routes are used to limit the local area network (LAN) address to fall within the pre-configured range of available LAN addresses, ensure that the LAN address does not cause ambiguity in intra-group communication routes, and that the LAN address can carry subsequent probe query commands and task data transmission. This process differs from the coarse-grained approach in the prior art that relies on a fixed central node polling online status. It establishes a hierarchical filtering relationship between the parking trigger signal, the spatial clustering label, the connectivity confirmation record, the security domain authentication record, and the LAN address confirmation record, ensuring that the vehicle nodes entering the temporary computing federation group are consistent with the parking scenarios, intra-group communication routes, and open security domains required for subsequent heterogeneous computing power scheduling.

[0031] Optionally, the step of extracting the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, wherein the specifications include concurrent throughput dimension information of the onboard neural network acceleration unit, includes: Send probe query commands to each vehicle node within the temporary computing federation group; Receive the underlying hardware status logs for the hardware acceleration unit returned by each vehicle node in response to the probe query command. Extract the number of tensor operation cores and available video memory capacity of the vehicle-mounted neural network acceleration unit from the underlying hardware status log; The number of tensor operation cores and the available video memory capacity are formatted and combined to form the concurrent throughput dimension information, and the formatted combination data containing the concurrent throughput dimension information is encapsulated into the specification of the hardware acceleration unit of each vehicle node.

[0032] Preferably, the specific implementation process of extracting the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group is as follows: After the temporary computing federation group is generated, the member registration area and the federation group lifecycle status record of the temporary computing federation group are read first. The member registration area of ​​the temporary computing federation group is used to provide the vehicle node identity digest, local area network address, and intra-group communication route of each vehicle node. The federation group lifecycle status record is used to determine whether the temporary computing federation group has entered the computing power registration pending execution state. Under the condition that the federation group lifecycle status record indicates that the temporary computing federation group is in the computing power registration pending execution state, a probe deployment plan is generated according to the member registration area of ​​the temporary computing federation group. The probe deployment plan includes the vehicle node to be queried, the local area network address corresponding to the vehicle node to be queried, the probe session identifier, and the hardware acceleration unit query field. The vehicle node to be queried comes from the vehicle node in the member registration area of ​​the temporary computing federation group that is still in a connected state. The probe deployment plan is used to generate the probe query instruction, so that the probe query instruction is only issued to the vehicle node to be queried, rather than performing a generalized query on any online vehicle node.

[0033] Preferably, when issuing probe query commands to each vehicle node within the temporary computing federation group, the probe query commands are encapsulated according to the probe issuance plan. The probe query command includes the probe session identifier, the vehicle node identity digest, the local area network address, the hardware acceleration unit query field, the log return address, and the status sampling time base. The probe session identifier is used to map the subsequently returned underlying hardware status logs to this probe query command. The vehicle node identity digest is used to verify whether the vehicle node returning the underlying hardware status logs belongs to the member registration area of ​​the temporary computing federation group. The local area network address is used to issue the probe query command along the intra-group communication route. The hardware acceleration unit query field is used to limit the range of underlying hardware data queried in this instance. The log return address is used to limit the return path of the underlying hardware status logs. The status sampling time base is used to constrain each vehicle node to read its local hardware operating status within the sampling time window corresponding to the status sampling time base. Through this encapsulation method, the probe query command, from its generation, establishes a correspondence with the member registration area of ​​the temporary computing federation group, the intra-group communication route, the status sampling time base, and the subsequent underlying hardware status logs.

[0034] Preferably, the hardware acceleration unit query field includes at least an onboard neural network acceleration unit identification field, a tensor operation core count query field, and an available video memory capacity query field. The onboard neural network acceleration unit identification field confirms whether the hardware acceleration units actually participating in computing power collaboration within the vehicle node include an onboard neural network acceleration unit. The tensor operation core count query field reads the number of tensor operation cores in the onboard neural network acceleration unit that are in a schedulable state. The available video memory capacity query field reads the available video memory capacity that the onboard neural network acceleration unit can allocate to task execution within the sampling time window corresponding to the state sampling time benchmark. The hardware acceleration unit query field may also carry a network bandwidth transmission index query field and an estimated contribution duration query field. The network bandwidth transmission index query field is used to determine the task data transmission capability during subsequent vehicle node screening, and the estimated contribution duration query field is used to determine the valid time range of the specification during subsequent vehicle node state management. The network bandwidth transmission index query field and the estimated contribution duration query field do not change the technical relationship formed by the format combination of the tensor operation core count and the available video memory capacity for the concurrent throughput dimension information.

[0035] Preferably, after receiving the probe query instruction, each vehicle node first performs local probe verification on the probe session identifier, vehicle node identity digest, and state sampling time base in the probe query instruction, forming a local probe verification record. If the local probe verification record indicates that the probe session identifier, vehicle node identity digest, and state sampling time base have all passed verification, the vehicle node reads the local hardware operating status according to the hardware acceleration unit query field, forming a low-level hardware status log. The low-level hardware status log includes the probe session identifier, vehicle node identity digest, hardware acceleration unit identifier, vehicle neural network acceleration unit identification result, hardware operating status field, network bandwidth transmission index, estimated contribution duration, and state sampling timestamp. The hardware operating status field includes the number of tensor operation cores, available video memory capacity, video memory capacity field, video memory usage field, and video memory reservation field. The state sampling timestamp is used to correspond with the state sampling time base. The identification result of the vehicle-mounted neural network acceleration unit is used to indicate that the number of tensor operation cores and the available video memory capacity come from the vehicle-mounted neural network acceleration unit in the vehicle node. The underlying hardware status log is then returned along the log back address so that log reception and log verification can be performed subsequently based on the same probe session identifier.

[0036] Preferably, when receiving the underlying hardware status logs for the hardware acceleration unit returned by each vehicle node in response to the probe query command, the probe session identifier in the underlying hardware status log is first matched with the probe session identifier in the probe distribution plan to obtain a probe session matching result; then, the vehicle node identity digest in the underlying hardware status log is compared with the vehicle node identity digest in the member registration area of ​​the temporary computing federation group to obtain a member identity matching result; subsequently, a time validity check is performed based on the state sampling timestamp and the state sampling time base to obtain a state sampling validity result. The probe session matching result, the member identity matching result, and the state sampling validity result together form an underlying hardware status log verification record; the underlying hardware status log verification record is used to filter out underlying hardware status logs that do not match probe sessions, do not match vehicle node identity digests, or whose state sampling timestamps exceed the sampling time window corresponding to the state sampling time base, so that the underlying hardware status logs that have passed the verification by the underlying hardware status log verification record are used when extracting the number of tensor operation cores and the available video memory capacity.

[0037] Preferably, when extracting the number of tensor operation cores of the vehicle-mounted neural network acceleration unit from the underlying hardware status log, the corresponding hardware acceleration unit identifier is first located based on the vehicle-mounted neural network acceleration unit identification result in the underlying hardware status log, and then the number of tensor operation cores is read from the hardware running status field corresponding to the hardware acceleration unit identifier. The number of tensor operation cores is used to characterize the number of cores of the vehicle-mounted neural network acceleration unit that can participate in tensor operation scheduling simultaneously within the sampling time window corresponding to the state sampling time benchmark. The number of tensor operation cores is not the number of ordinary central processing unit threads of the vehicle node, nor the number of rendering units of the graphics display unit; the number of ordinary central processing unit threads and the number of rendering units of the graphics display unit are not written into the concurrency throughput dimension information. The number of tensor operation cores is related to the parallel data processing processes such as tensor multiplication and addition, tensor convolution, and batch tensor mapping performed by the vehicle-mounted neural network acceleration unit. When forming the concurrency throughput dimension information later, the number of tensor operation cores is used to describe the scale of parallel execution resources when the vehicle node undertakes high-concurrency instruction processing.

[0038] Preferably, when extracting the available video memory capacity of the vehicle-mounted neural network acceleration unit from the underlying hardware status log, the video memory capacity field, the video memory occupied field, and the video memory reserved field are first read from the hardware running status field corresponding to the hardware acceleration unit identifier. Then, the available video memory capacity is generated based on the video memory capacity field, the video memory occupied field, and the video memory reserved field. The video memory capacity field is used to characterize the video memory space that the vehicle-mounted neural network acceleration unit can manage, the video memory occupied field is used to characterize the video memory space that has been occupied by resident tasks, and the video memory reserved field is used to characterize the video memory space that the vehicle node reserves for local safe operation and basic services. The resident task is a local task that the vehicle node has been running and occupying video memory space within the sampling time window corresponding to the status sampling time reference. The available video memory capacity is used to characterize the video memory space that can be provided for subsequent target processing tasks without encroaching on the corresponding space of the video memory occupied field and the video memory reserved field; when the concurrent throughput dimension information is subsequently formed, the available video memory capacity is used to describe the storage capacity of the vehicle node to accommodate input data, intermediate tensor data and local processing results when it undertakes concurrent processing sub-streams.

[0039] Preferably, before formatting the number of tensor operation cores and the available video memory capacity, the number of tensor operation cores and the available video memory capacity are first normalized to obtain a normalized field for the number of tensor operation cores and a normalized field for the available video memory capacity. The normalized field for the number of tensor operation cores retains the integer resource meaning of the number of tensor operation cores, and the normalized field for the available video memory capacity retains the storage capacity meaning of the available video memory capacity. The normalized field for the number of tensor operation cores and the normalized field for the available video memory capacity are not mixed in terms of dimensions, nor are they added or multiplied. Through the field normalization process, the number of tensor operation cores and the available video memory capacity respectively carry the field name, field value, field unit, and state sampling timestamp in the same concurrency throughput dimension field structure. The concurrency throughput dimension field structure is then used to form the concurrency throughput dimension information, so that the subsequent concurrency throughput dimension information can express both parallel computing resources and video memory carrying resources, avoiding the miscombination of hardware state data with different dimensions into a single value, which would cause unclear scheduling meaning.

[0040] Preferably, when combining the number of tensor operation cores and the available video memory capacity in a formatted manner, the tensor operation core quantity specification field and the available video memory capacity specification field are bound according to the vehicle node identity digest, the same probe session identifier, and the same hardware acceleration unit identifier of the same vehicle node, forming formatted combined data. The formatted combined data includes a vehicle node identity digest field, a hardware acceleration unit identifier field, a tensor operation core quantity field, an available video memory capacity field, and a state sampling timestamp field. The vehicle node identity digest field is used to associate vehicle nodes in the member registration area of ​​the temporary computing federation group; the hardware acceleration unit identifier field is used to associate the queried hardware acceleration unit; the tensor operation core quantity field and the available video memory capacity field together serve as the concurrent throughput dimension information; and the state sampling timestamp field is used to determine whether the concurrent throughput dimension information is still within a valid time range. The formatted combined data then undergoes specification encapsulation processing, enabling the concurrent throughput dimension information to participate in subsequent throughput requirement indicator matching in the form of structured fields.

[0041] Preferably, when encapsulating the formatted combination data containing the concurrent throughput dimension information into the specification of the hardware acceleration unit of each vehicle node, the formatted combination data, the network bandwidth transmission index, the expected contribution duration, and the underlying hardware status log verification record are jointly written into the vehicle node specification registration record; the vehicle node specification registration record is used to register the hardware status of each vehicle node that can participate in computing power collaboration within the temporary computing federation group. The vehicle node specification registration record is further indexed according to the vehicle node identity digest to generate the specification; the specification includes the vehicle node identity digest, the local area network address, the hardware acceleration unit identifier, the concurrent throughput dimension information, the network bandwidth transmission index, the expected contribution duration, and the underlying hardware status log verification record. The concurrent throughput dimension information is used in the specification to determine whether the vehicle node meets the throughput requirements of the first instruction processing set. The network bandwidth transmission index is used in the specification to determine whether the vehicle node is suitable for undertaking data transmission between serially associated computing task fragments. The estimated contribution time is used in the specification to determine whether the vehicle node is still suitable to remain in the temporary computing federation group to continue participating in task execution.

[0042] Preferably, in a multi-vehicle clustered parking application scenario, multiple vehicle nodes within the temporary computing federation group respond to the probe query command and return the underlying hardware status log. If one of the vehicle nodes has a high number of tensor operation cores and a large available video memory capacity, then the concurrent throughput dimension information corresponding to that vehicle node can reflect that it is more suitable for undertaking high-concurrency tensor processing tasks. If another vehicle node has a low number of tensor operation cores but a relatively stable network bandwidth transmission index, then the specification description corresponding to that vehicle node can reflect that it is more suitable for participating in the fragmented transmission of serially associated computing tasks. The above judgment is not achieved by pre-maintaining a static list of available nodes through a fixed central node, but rather through hierarchical data processing between the probe query command, the underlying hardware status log, the number of tensor operation cores, the available video memory capacity, the format combination data, the concurrent throughput dimension information, and the specification description. This enables the temporary computing federation group to establish a vehicle node-level computing power status description record based on the specification description after parking is formed. The vehicle node-level computing power status description record continues to be used for subsequent concurrent processing sub-stream distribution and serially associated computing task fragment distribution.

[0043] Preferably, after the specification is formed, a specification validity record is generated based on the state sampling timestamp and the expected contribution duration. This specification validity record describes whether the specification can still serve as a basis for the vehicle node's hardware status before subsequent task sharding is distributed. If the specification validity record indicates that the specification is still within its valid time range, the specification is written into the specification registration area of ​​the temporary computing federation group. The specification registration area of ​​the temporary computing federation group stores specifications still within their valid time range and provides them to subsequent concurrent processing sub-stream distribution and serially associated computing task sharding distribution. If the specification validity record indicates that the specification has exceeded its valid time range, a new probe query command is reissued to the corresponding vehicle node to obtain new underlying hardware status logs and re-form the specification. Through the specification validity record, the concurrent throughput dimension information can be updated following changes in the hardware occupancy of the parked vehicle node, reducing the computing power matching deviation caused by subsequent distribution of concurrent processing sub-streams or serially associated computing task shards based on specifications that have exceeded their valid time range.

[0044] Optionally, the step of parsing the operator execution dependencies of the target processing task, performing hierarchical decoupling of the target processing task based on the operator execution dependencies, and generating a first instruction processing set and a second instruction processing set includes: The target processing task is parsed using a directed acyclic graph to obtain the operator execution dependencies consisting of nodes and edges; The hierarchical decoupling is performed based on the edge connection state in the operator execution dependency relationship, and independent operator nodes without data dependency edges and coupled operator nodes with data dependency edges are extracted. The independent operator nodes are aggregated to generate the first instruction processing set; Extract the data flow direction indicated by the data dependency edge between the coupled operator nodes, and combine the coupled operator nodes with the same data flow direction in a time sequence to generate the second instruction processing set.

[0045] Preferably, the specific implementation process of parsing the operator execution dependencies of the target processing task is as follows: After receiving the target processing task, the task structure is first parsed to obtain a task structure parsing record; the task structure parsing record includes an input data description, an instruction execution description, an instruction input port, an instruction output port, and instruction execution order constraints. The input data description is used to identify the data source required by each instruction in the target processing task; the instruction execution description is used to identify the tensor processing, embedding lookup, attention calculation, result aggregation, or other data processing actions corresponding to each instruction; the instruction input port and the instruction output port are used to identify the data connection relationship between each instruction; and the instruction execution order constraints are used to identify the pre-processing relationship that cannot be executed in parallel. By first forming the task structure parsing record, the subsequent directed acyclic graph parsing is not directly based on a rough split of the task text, but rather identifies the data connection relationship and pre-processing relationship within the target processing task based on the input data description, the instruction execution description, the instruction input port, the instruction output port, and the instruction execution order constraints.

[0046] Preferably, when performing directed acyclic graph parsing on the target processing task, operator nodes are generated based on the instruction execution description in the task structure parsing record, and data dependency edges are generated based on the instruction input port and instruction output port in the task structure parsing record, thus forming the operator execution dependency relationship composed of operator nodes and data dependency edges. The operator node represents an instruction processing unit in the target processing task that can be scheduled independently, and the data dependency edge represents the output data of one operator node that needs to be used as input data for another operator node to continue processing. The starting point of the data dependency edge corresponds to the operator node from which the output data originates, and the ending point of the data dependency edge corresponds to the operator node that receives the output data. The operator execution dependency relationship thus formed can express the instruction processing units in the target processing task, the data connection relationships between instruction processing units, and the pre- and post-processing relationships between instruction processing units as a computable data structure, facilitating subsequent hierarchical decoupling based on edge connection states.

[0047] Preferably, after forming the operator execution dependency relationship consisting of operator nodes and data dependency edges, the operator execution dependency relationship is checked for acyclicity to obtain an acyclicity check record. The acyclicity check record is used to determine whether there is a cyclic connection in the operator execution dependency relationship where subsequent operator nodes depend on preceding operator nodes in reverse. If the acyclicity check record indicates the existence of a cyclic connection, the cyclic connection is broken down according to the instruction execution order constraints in the task structure parsing record to obtain a cyclic connection location record. Based on the cyclic connection location record, the data dependency edges are corrected or split to ensure that the operator execution dependency relationship after correction or splitting satisfies the directed acyclicity condition. The operator execution dependency relationship confirmed by the acyclicity check record continues to be used for hierarchical decoupling, preventing task structures with cyclic connections from being incorrectly classified into instruction processing ranges that can be directly executed serially or in parallel. The instruction processing ranges that can be directly executed serially or in parallel are continued to be taken over by the subsequent first instruction processing set and the second instruction processing set, respectively.

[0048] Preferably, when performing hierarchical decoupling based on the edge connection states in the operator execution dependency relationship, the incoming and outgoing data dependency edges of each operator node are first traversed to obtain edge connection state records. The incoming data dependency edges indicate whether the operator node depends on the output data of other operator nodes, and the outgoing data dependency edges indicate whether the output data of the operator node continues to be supplied to other operator nodes. The edge connection state records include the number of incoming data dependency edges, the number of outgoing data dependency edges, adjacent operator node identity records, and the data flow direction between adjacent operator nodes. The adjacent operator node identity records record other operator nodes directly connected to the current operator node through the data dependency edges, and the data flow direction between adjacent operator nodes records the direction in which the output data is transmitted from one operator node to another along the data dependency edges. The edge connection state records continue to serve as the basis for determining hierarchical decoupling, ensuring that the extraction process of independent and coupled operator nodes has a clear data source, rather than relying solely on isomorphic distribution rules based on instruction names or task sizes.

[0049] Preferably, when extracting the independent operator nodes, operator nodes without data-dependent edges are identified based on the edge connection state record. These operator nodes include those without inbound and outbound data-dependent edges, or those directly associated with the data source corresponding to the input data description and not dependent on the output data of other operator nodes. Independent execution markers are applied to the operator nodes without data-dependent edges to obtain independent operator node marker records. The independent execution marker indicates that the corresponding operator node can execute without waiting for intermediate output data from other operator nodes. The independent operator node marker record stores the operator node written with the independent execution marker and the judgment criteria corresponding to the independent execution marker. The independent operator node marker record continues to participate in generating the first instruction processing set, enabling the independent instruction processing units in the first instruction processing set to have a technical basis for parallel issuance and to be matched with the concurrent throughput dimension information subsequently.

[0050] Preferably, when extracting the coupled operator nodes, operator nodes with data-dependent edges are identified based on the edge connection state record. These operator nodes include those with inbound data-dependent edges, those with outbound data-dependent edges, and those with both inbound and outbound data-dependent edges. The operator nodes with data-dependent edges are then coupled and marked with a coupling execution flag to obtain a coupled operator node mark record. This flag indicates that the execution result of the corresponding operator node needs to be passed along the data-dependent edge, or that the corresponding operator node needs to wait for the output data of other operator nodes before execution. The coupled operator node mark record stores the operator node written with the coupled execution flag and the corresponding judgment criteria. The coupled operator node mark record continues to participate in generating the second instruction processing set, enabling the second instruction processing set to retain the temporal transmission constraints within the target processing task, rather than forcibly breaking down instructions with sequential dependencies into independent isomorphic subtasks.

[0051] Preferably, when aggregating the independent operator nodes to generate the first instruction processing set, the independent operator node tagging record is first read, and the corresponding operator node, instruction execution description, input data description, and output data encapsulation method are extracted according to the independent operator node tagging record to form an independent instruction aggregation record. The independent instruction aggregation record is used to organize multiple independent operator nodes according to a parallel execution granularity, rather than changing the instruction processing content of the independent operator node itself. Subsequently, multiple independent operator nodes are aggregated into the first instruction processing set according to the independent instruction aggregation record; the first instruction processing set includes multiple independent instruction processing units capable of concurrent execution, each independent instruction processing unit originating from the independent operator node and retaining its respective input data description and output data encapsulation method. The first instruction processing set is subsequently used to extract throughput requirement metrics, thereby matching them with the concurrent throughput dimension information in the specification.

[0052] Preferably, after generating the first instruction processing set, a concurrency boundary check is performed on the first instruction processing set to obtain a concurrency boundary check record. The concurrency boundary check record is used to determine whether there are any missing data dependency edges between the independent instruction processing units in the first instruction processing set. During the concurrency boundary check, each independent instruction processing unit in the first instruction processing set is traced back to the operator execution dependency relationship, and the inbound and outbound data dependency edges of the corresponding operator node in the operator execution dependency relationship are checked. If the corresponding operator node has missing data dependency edges in the operator execution dependency relationship, the corresponding independent instruction processing unit is removed from the first instruction processing set, and the corresponding operator node is rewritten into the coupled operator node marker record according to the missing data dependency edges. Through the concurrency boundary check record, the situation where operator nodes with missing data dependency edges are mistakenly included in the first instruction processing set can be reduced, ensuring that the technical meaning of the first instruction processing set remains consistent with that of subsequent concurrent processing sub-streams.

[0053] Preferably, when extracting the data flow direction indicated by the data dependency edge between the coupled operator nodes, the coupled operator node marker record and the edge connection status record are read first, and the direction of each data dependency edge connecting two coupled operator nodes is read to obtain a data flow direction record. The data flow direction record includes the output data source operator node, the input data receiving operator node, the data object name, the data object form, and the transmission order position. The output data source operator node indicates from which coupled operator node the data is output; the input data receiving operator node indicates to which coupled operator node the data enters for further processing; the data object name and the data object form indicate the data content transmitted along the data dependency edge; and the transmission order position describes the sequential order of the data dependency edge during coupled execution. The data flow direction record is further used for subsequent timing combinations, so that the generation of the second instruction processing set is based on the data flow direction in the operator execution dependency relationship.

[0054] Preferably, when combining coupled operator nodes with the same data flow direction in a time sequence, the coupled operator nodes with consistent continuous transmission directions are first identified according to the data flow direction record to obtain a co-directional coupled operator node record. The consistent continuous transmission direction means that the output data of the preceding coupled operator node enters the following coupled operator node along the data dependency edge, and the output data of the following coupled operator node continues to enter the next coupled operator node along the same data flow direction. The co-directional coupled operator node record is used to store coupled operator nodes that satisfy the consistent continuous transmission direction and their corresponding data dependency edges. Subsequently, the coupled operator nodes are sorted according to the co-directional coupled operator node record to obtain a coupled operator time sequence record. The coupled operator time sequence record retains the data succession order between the coupled operator nodes, so that the preceding coupled operator node, the following coupled operator node, and the data dependency edge between them can all be described in the same time sequence structure. The coupled operator time sequence record is further used to generate the second instruction processing set, so that the second instruction processing set can express the serial transmission relationship between coupled instructions.

[0055] Preferably, when generating the second instruction processing set, the coupling operator nodes, the data dependency edges, and the data flow direction records are combined according to the coupling operator timing records to form the second instruction processing set. The second instruction processing set includes coupled instruction processing units, data dependency edges between coupled instruction processing units, data flow direction records corresponding to each data dependency edge, and the coupling operator timing records. The coupled instruction processing units originate from the coupling operator nodes. The data dependency edges between coupled instruction processing units are used to maintain the data transmission relationship between the coupled instruction processing units. The data flow direction records are used to describe the transmission direction of data between the coupled instruction processing units, and the coupling operator timing records are used to describe the execution order of the coupled instruction processing units. The second instruction processing set is subsequently used to extract timing transmission constraints, and then used to construct a cascaded execution topology sequence and split multiple serially associated computational task fragments.

[0056] Preferably, in application scenarios where multiple vehicles are parked together and the temporary computing federation group is used to process large-scale target processing tasks, the target processing tasks may include batch image processing, video clip rendering, edge-side large model inference, or multi-frame perceptual data processing; the batch image processing, the video clip rendering, the edge-side large model inference, or the multi-frame perceptual data processing can all be expressed as the operator execution dependencies through the task structure parsing record. In the above application scenarios, embedding lookup, independent preprocessing of several frames of images, or parallel feature extraction of multiple sub-batches can usually form the independent operator nodes, which are then aggregated to form the first instruction processing set; the intermediate tensor transfer between attention computing layers, the transfer of inference results from the previous layer to the next layer, and the transfer of pre-rendered segments to the aggregation of post-results can usually form the coupled operator nodes, which are then combined with the data flow direction record and the coupled operator timing record to form the second instruction processing set. This processing method differs from the approach of dividing the target processing task into multiple isomorphic subtasks on an average basis. Instead, it first identifies the parts that can be executed concurrently and the parts that need to be passed in sequence from the operator execution dependencies, and then forms the first instruction processing set and the second instruction processing set respectively.

[0057] Preferably, after the first instruction processing set and the second instruction processing set are generated, a hierarchical decoupling verification record is also generated. This hierarchical decoupling verification record is used to record whether each operator node in the target processing task has been assigned to either the first or second instruction processing set, and to record whether each data dependency edge in the target processing task has been retained in the second instruction processing set. If the hierarchical decoupling verification record indicates the existence of operator nodes not assigned to either the first or second instruction processing set, then the corresponding operator node should be re-determined according to the edge connection status record to be assigned to either the independent operator node marker record or the coupled operator node marker record. If the hierarchical decoupling verification record indicates the existence of unretained data dependency edges, then the corresponding data dependency edges are added to the second instruction processing set according to the data flow direction record. Through the hierarchical decoupling verification record, the target processing task retains complete operator execution dependencies after hierarchical decoupling, and ensures that subsequent concurrent processing sub-stream configurations and cascaded execution topology sequence construction have a consistent data foundation.

[0058] Optionally, the steps of extracting the timing transit constraints of the second instruction processing set, constructing a cascaded execution topology sequence based on the timing transit constraints, and dividing the second instruction processing set into multiple serially associated computational task fragments along the cascaded execution topology sequence include: Extract the input tensor size of the first instruction data and the output tensor size of the last instruction data from the second instruction processing set, and use the input tensor size and the output tensor size as the timing transmission constraint. Based on the input tensor size and the output tensor size in the timing transit constraints, calculate the activation value overhead parameter when the second instruction processing set is internally transited; Based on the activation value overhead parameter, multiple data communication nodes are inserted into the second instruction processing set, and the second instruction processing set is connected in series using the data communication nodes to construct and generate the cascaded execution topology sequence. Using the data communication node as the dividing point, the second instruction processing set is divided into multiple serially associated computing task fragments along the cascaded execution topology sequence.

[0059] Preferably, the specific implementation process of extracting the timing transit constraints of the second instruction processing set is as follows: After generating the second instruction processing set, the coupled instruction processing units, the data dependency edges between the coupled instruction processing units, the data flow direction records corresponding to each data dependency edge, and the timing records of the coupling operators in the second instruction processing set are read first. The timing records of the coupling operators are used to determine the execution order of the coupled instruction processing units in the second instruction processing set; according to the timing records of the coupling operators, the data corresponding to the coupled instruction processing unit located at the execution start position is determined as the head instruction data, and the data corresponding to the coupled instruction processing unit located at the execution end position is determined as the tail instruction data. The head instruction data and the tail instruction data both originate from the second instruction processing set and continue to participate in the subsequent extraction of input tensor size and output tensor size, so that the timing transit constraints can directly reflect the serial acceptance boundary inside the second instruction processing set; the serial acceptance boundary includes the starting input boundary defined by the head instruction data and the ending output boundary defined by the tail instruction data.

[0060] Preferably, when extracting the input tensor size of the header instruction data in the second instruction processing set, an input tensor description record is read according to the instruction input port corresponding to the header instruction data. The input tensor description record includes the number of dimensions of the input tensor, the data length corresponding to each dimension, the data type of the input tensor, and the batch organization method of the input tensor. The number of dimensions of the input tensor is used to characterize how many data expansion directions the input tensor has, the data length corresponding to each dimension is used to characterize the number of elements in each data expansion direction, the data type of the input tensor is used to characterize the storage bit width occupied by each element, and the batch organization method of the input tensor is used to characterize whether multiple input samples are merged and processed in the header instruction data. The input tensor size is formed according to the input tensor description record, and the input tensor size is further written with the timing transit constraints to limit the data scale that the second instruction processing set needs to handle when it starts execution. The data scale continues to participate in the formation of the subsequent input tensor element scale description.

[0061] Preferably, when extracting the output tensor size of the tail instruction data in the second instruction processing set, the output tensor description record is read according to the instruction output port corresponding to the tail instruction data. The output tensor description record includes the number of dimensions of the output tensor, the data length corresponding to each dimension, the data type of the output tensor, and the result organization method of the output tensor. The number of dimensions of the output tensor is used to characterize the data expansion direction of the final output result of the tail instruction data, the data length corresponding to each dimension is used to characterize the number of elements in each data expansion direction of the final output result, the data type of the output tensor is used to characterize the storage bit width occupied by each element in the final output result, and the result organization method of the output tensor is used to characterize the encapsulation relationship between the final output result and the subsequent local processing results. The output tensor size is formed according to the output tensor description record. The output tensor size and the input tensor size are jointly written into the timing transit constraint, so that the timing transit constraint simultaneously limits the starting input boundary and the ending output boundary of the second instruction processing set. The ending output boundary continues to participate in the formation of the subsequent output tensor element size description.

[0062] Preferably, when the input tensor size and the output tensor size are used as the timing transitivity constraints, the input tensor size and the output tensor size are not combined into a single value. Instead, the input tensor size, the output tensor size, the instruction input port corresponding to the header instruction data, the instruction output port corresponding to the tail instruction data, and the timing record of the coupling operator are all written into the timing transitivity constraint record. The timing transitivity constraint record is used to describe the timing transitivity constraint relationship of the second instruction processing set from the header instruction data, through the data dependency edges between the coupled instruction processing units, and through the tail instruction data. The timing transitivity constraint record continues to serve as the carrier of the timing transitivity constraints, so that when calculating the activation value overhead parameter, the input tensor size, the output tensor size, and the timing record of the coupling operator can be read simultaneously, instead of simply dividing it equally according to the number of tasks.

[0063] Preferably, when calculating the activation value overhead parameter based on the input tensor size and the output tensor size in the timing transit constraints, an input tensor element size description is first formed based on the input tensor size, and an output tensor element size description is formed based on the output tensor size. The input tensor element size description characterizes the number of tensor elements and the element storage width that need to be carried when the header instruction data enters the second instruction processing set, and the output tensor element size description characterizes the number of tensor elements and the element storage width that need to be carried when the tail instruction data is output from the second instruction processing set. Subsequently, combining the data dependency edges between adjacent coupled instruction processing units in the coupling operator timing record, the data object form corresponding to each data dependency edge is read to form an intermediate tensor transit description. The intermediate tensor transit description continues to participate in the formation of the activation value overhead parameter together with the input tensor element size description and the output tensor element size description, so that the activation value overhead parameter can reflect the intermediate tensor transmission burden generated by the second instruction processing set during internal transit.

[0064] Preferably, the activation value overhead parameters include an input-side activation value overhead description, an intermediate-side activation value overhead description, and an output-side activation value overhead description. The input-side activation value overhead description is derived from the input tensor element size description and is used to characterize the activation value size that needs to be written or transmitted when the header instruction data enters the second instruction processing set. The intermediate-side activation value overhead description is derived from the intermediate tensor transfer description and is used to characterize the intermediate tensor size transmitted between adjacent coupled instruction processing units along the data dependency edge. The output-side activation value overhead description is derived from the output tensor element size description and is used to characterize the activation value size that needs to be returned or temporarily stored when the tail instruction data forms the output result. The input-side activation value overhead description, the intermediate-side activation value overhead description, and the output-side activation value overhead description all retain their respective tensor size sources and do not perform mixed operations of different dimensions, so that the insertion position of the data communication node can be determined subsequently based on the activation value overhead parameters.

[0065] Preferably, when inserting multiple data communication nodes within the second instruction processing set according to the activation value overhead parameter, the data dependency edges between adjacent coupled instruction processing units are first read sequentially along the coupling operator timing record. The intermediate side activation value overhead description corresponding to each data dependency edge is then analyzed in correspondence with the network bandwidth transmission index in the specification of the temporary computing federation group to obtain a data communication node candidate position record. This record is used to mark which intermediate tensor sizes corresponding to data dependency edges are suitable as fragmentation transmission boundaries between vehicle nodes, and which are not. Subsequently, the data communication nodes are inserted within the second instruction processing set according to the data communication node candidate position record, such that the data communication nodes are located on the data dependency edges between adjacent coupled instruction processing units, and are used to receive the output data of the previous coupled instruction processing unit and the input data of the next coupled instruction processing unit.

[0066] Preferably, the data communication node does not replace the computational content of the coupled instruction processing unit, but is a transmission boundary node set on the data dependency edge within the second instruction processing set. The data communication node includes a pre-coupled instruction processing unit identifier, a post-coupled instruction processing unit identifier, an intermediate tensor transmission description, a data flow direction record, a network bandwidth transmission index corresponding record, and a fragmentation boundary identifier. The pre-coupled instruction processing unit identifier indicates which coupled instruction processing unit's output intermediate tensor the data communication node receives; the post-coupled instruction processing unit identifier indicates which coupled instruction processing unit the data communication node transfers the intermediate tensor to; the intermediate tensor transmission description characterizes the scale of the intermediate tensor received by the data communication node; the data flow direction record maintains the intermediate tensor transmission direction unchanged; the network bandwidth transmission index corresponding record explains the correspondence between the data transmission capabilities of the data communication node and the vehicle node; and the fragmentation boundary identifier is used for subsequent fragmentation of the multiple serially associated computational tasks.

[0067] Preferably, when the second instruction processing set is chained together using the data communication node, the execution order between each coupled instruction processing unit is preserved according to the coupling operator timing record, and the data communication node is embedded in the data dependency edge between adjacent coupled instruction processing units to form a cascade node connection record. The cascade node connection record includes the coupled instruction processing unit, the data communication node, the data dependency edge between the coupled instruction processing units, and the data flow direction record corresponding to the data communication node. The cascade execution topology sequence is generated based on the cascade node connection record; the cascade execution topology sequence is used to express the serial arrangement relationship between each coupled instruction processing unit and each data communication node in the second instruction processing set. Through the cascade execution topology sequence, the second instruction processing set is organized into a task chain structure with task issuance boundaries, fragmentation transmission boundaries, and execution order; the task issuance boundaries, the fragmentation transmission boundaries, and the execution order continue to participate in the subsequent partitioning of multiple serially associated computation task fragments.

[0068] Preferably, after constructing and generating the cascaded execution topology sequence, a topology continuity check is performed on the cascaded execution topology sequence to obtain a topology continuity check record. The topology continuity check record is used to determine whether the corresponding data dependency edges are retained between adjacent nodes in the cascaded execution topology sequence, and whether the data communication nodes are all located between adjacent coupled instruction processing units. If the topology continuity check record indicates that a certain data communication node does not correspond to any data dependency edge, then the data communication node is removed and the candidate position record of the data communication node is reread; if the topology continuity check record indicates that a certain data dependency edge is not retained by the cascaded execution topology sequence, then the data dependency edge is added to the cascaded execution topology sequence according to the coupling operator timing record. The topology continuity check record enables the cascaded execution topology sequence to retain the original data flow direction within the second instruction processing set, reducing the situation where the calculation results before and after cannot be connected after the split.

[0069] Preferably, when dividing the second instruction processing set into multiple serially associated computational task fragments along the cascaded execution topology sequence using the data communication node as the demarcation point, the fragment boundary identifier in the cascaded execution topology sequence is first read, and a fragment instruction segment is formed based on the coupled instruction processing units between two adjacent fragment boundary identifiers. The fragment instruction segment includes one or more consecutively arranged coupled instruction processing units, and retains all data dependency edges within the fragment instruction segment. Subsequently, each fragment instruction segment is encapsulated as a computational task fragment, and the data communication nodes between adjacent computational task fragments are written into an inter-fragment transfer record. The inter-fragment transfer record is used to explain how the output data of the previous computational task fragment enters the next computational task fragment along the data communication node, thereby ensuring that the multiple serially associated computational task fragments maintain a consistent data transfer relationship with the second instruction processing set after division.

[0070] Preferably, the plurality of serially associated computation task fragments each include a fragmentation instruction segment, a fragmentation input boundary, a fragmentation output boundary, intra-fragment data dependency edges, and inter-fragment transfer records. The fragmentation input boundary originates from the instruction input port corresponding to the beginning of the fragmentation instruction segment or the preceding data communication node; the fragmentation output boundary originates from the instruction output port corresponding to the end of the fragmentation instruction segment or the following data communication node; the intra-fragment data dependency edges are used to preserve the data transfer relationship between coupled instruction processing units within the same computation task fragment; and the inter-fragment transfer records are used to preserve the data transfer relationship between adjacent computation task fragments. Through the above encapsulation method, the plurality of serially associated computation task fragments can be respectively dispatched to the second vehicle node and the third vehicle node for execution, while maintaining the execution order and intermediate tensor transfer paths along the cascaded execution topology sequence.

[0071] Preferably, in a multi-vehicle centralized parking application scenario, if the second instruction processing set originates from an edge-side large model inference task, then the input tensor size of the first instruction data corresponds to the size of the first-layer inference input tensor, and the output tensor size of the tail instruction data corresponds to the size of the last-layer inference output tensor. The activation value overhead parameter is used to describe the overhead of transferring intermediate tensors between adjacent inference layers between vehicle nodes. If the second instruction processing set originates from a video segment rendering task, then the input tensor size of the first instruction data corresponds to the input frame data size of the preceding rendered segment, and the output tensor size of the tail instruction data corresponds to the output frame data size of the subsequent result aggregation. The activation value overhead parameter is used to describe the overhead of transferring intermediate frame data between adjacent rendering stages between vehicle nodes. The above examples do not change the technical relationship formed by the input tensor size and the output tensor size of the temporal transmission constraints, nor do they change the technical relationship of the multiple serially associated computation task slices being divided along the cascaded execution topology sequence.

[0072] Preferably, after the multiple serially associated computing task fragments are formed, a fragment association verification record is also generated. This fragment association verification record records the position of each computing task fragment in the cascaded execution topology sequence, the fragment input boundary corresponding to each computing task fragment, the fragment output boundary corresponding to each computing task fragment, and the data communication nodes between adjacent computing task fragments. If the fragment association verification record indicates that a computing task fragment lacks a fragment input boundary, then the data communication node or header instruction data preceding that computing task fragment is traced back according to the cascaded execution topology sequence; if the fragment association verification record indicates that a computing task fragment lacks a fragment output boundary, then the data communication node or tail instruction data following that computing task fragment is traced back according to the cascaded execution topology sequence. Through the fragment association verification record, the multiple serially associated computing task fragments can retain complete input sources, output destinations, and serial association order before being distributed, providing a consistent data foundation for subsequently distributing the preceding computing task fragments to the second vehicle node and the remaining computing task fragments to the third vehicle node.

[0073] Optionally, the steps of distributing the concurrent processing substream to the first vehicle node in the temporary computing federation group according to the specification, and distributing the multiple serially associated computing task fragments sequentially to the second and third vehicle nodes in the temporary computing federation group, include: Under the condition that the concurrent throughput dimension information in the specification meets the first preset value, candidate nodes with matching computing power are selected from the temporary computing federation group as the first vehicle node, and the concurrent processing subflow is sent to the first vehicle node. Extract the network bandwidth transmission index contained in the specification. Under the condition that the network bandwidth transmission index meets the second preset value, select candidate nodes that meet the transmission requirements from the temporary computing federation group as the second vehicle node and the third vehicle node. According to the serial association order of the cascaded execution topology sequence, the preceding computing task fragments in the multiple serially associated computing task fragments are sent to the second vehicle node, and the remaining computing task fragments are sent to the third vehicle node.

[0074] Preferably, the specific implementation process of distributing the concurrent processing sub-stream to the first vehicle node in the temporary computing federation group according to the specification is as follows: After forming the concurrent processing sub-stream and the multiple serially associated computing task fragments, the specification registration area of ​​the temporary computing federation group, the throughput requirement index corresponding to the first instruction processing set, the concurrent processing sub-stream, and the specification validity record are read first. The specification registration area of ​​the temporary computing federation group is used to provide the specifications of each vehicle node, the specification validity record is used to filter out specifications that have exceeded the valid time range, and the throughput requirement index is used to limit the requirements of the concurrent processing sub-stream for parallel computing resources and video memory resources when it is executed. After reading the specification registration area of ​​the temporary computing federation group, the specifications that are still within the valid time range are retained first, and the concurrent throughput dimension information of each vehicle node is extracted according to the specifications that are still within the valid time range, so that the subsequent filtering process is based on the hardware status of the vehicle nodes that are still within the valid time range, rather than based on the specifications that have exceeded the valid time range for task distribution.

[0075] Preferably, before filtering based on the concurrent throughput dimension information in the specification meeting the first preset value, the first preset value is configured according to the throughput requirement index corresponding to the concurrent processing sub-stream. The first preset value is not a fixed single threshold applicable to all tasks, but is configured based on the number of concurrent instructions, the input data size of the concurrent processing sub-stream, the intermediate tensor occupancy size of the concurrent processing sub-stream, and the local processing result temporary storage size of the concurrent processing sub-stream, all included in the throughput requirement index. The number of concurrent instructions characterizes the number of independent instruction processing units that need to be scheduled simultaneously in the concurrent processing sub-stream; the input data size of the concurrent processing sub-stream characterizes the amount of data that needs to be read when the concurrent processing sub-stream starts execution; the intermediate tensor occupancy size of the concurrent processing sub-stream characterizes the amount of intermediate tensor data that needs to be temporarily stored during the execution of the concurrent processing sub-stream; and the local processing result temporary storage size of the concurrent processing sub-stream characterizes the amount of local processing result data that needs to be temporarily stored and returned after the concurrent processing sub-stream completes. The first preset value is used to limit the minimum parallel computing resources and minimum video memory capacity required for a vehicle node to handle the concurrent processing sub-stream. Subsequently, the concurrent throughput dimension information of each vehicle node is compared with the first preset value using a field correspondence determination. Specifically, the number of tensor operation cores in the concurrent throughput dimension information is used to determine the parallel computing resources, and the available video memory capacity in the concurrent throughput dimension information is used to determine the video memory capacity. These two values ​​are determined separately without any mixed-dimensional calculations.

[0076] Preferably, when selecting candidate nodes with matching computing power from the temporary computing federation group, concurrent throughput matching processing is performed on the concurrent throughput dimension information of each vehicle node to obtain a concurrent throughput matching record. The concurrent throughput matching record includes a vehicle node identity digest, tensor operation core quantity matching result, available memory capacity matching result, specification validity record, and concurrent processing sub-stream matching status. The tensor operation core quantity matching result indicates whether the number of tensor operation cores of the vehicle node has reached the lower limit of parallel computing resources corresponding to the first preset value. The available memory capacity matching result indicates whether the available memory capacity of the vehicle node has reached the lower limit of memory carrying resources corresponding to the first preset value. The specification validity record indicates whether the specification of the corresponding vehicle node is still within the valid time range. The concurrent processing sub-stream matching status summarizes the tensor operation core quantity matching result, the available memory capacity matching result, and the specification validity record. If the concurrent processing substream matching status characterization vehicle node simultaneously satisfies the judgment conditions corresponding to the tensor operation core quantity matching result, the available video memory capacity matching result, and the specification validity record, then the corresponding vehicle node is written into the computing power matching candidate node record; the computing power matching candidate node record is then used to determine the first vehicle node.

[0077] Preferably, when determining the first vehicle node, the vehicle node identity digest, local area network address, intra-group communication route, concurrent throughput dimension information, and estimated contribution time of each computing power matching candidate node are first read according to the computing power matching candidate node record. Then, a sustainable execution judgment is performed on the estimated contribution time based on the estimated execution time of the concurrent processing sub-flow to obtain the first vehicle node determination record. The estimated execution time of the concurrent processing sub-flow is derived from the throughput requirement index and is used to characterize the time range required for the concurrent processing sub-flow to return a partial processing result from the start of execution. The sustainable execution judgment is used to determine whether the estimated contribution time can cover the estimated execution time of the concurrent processing sub-flow. The first vehicle node determination record is used to indicate that the selected computing power matching candidate node not only meets the first preset value in the concurrent throughput dimension information but also covers the execution requirements of the concurrent processing sub-flow in terms of the estimated contribution time. Based on the first vehicle node determination record, candidate nodes that meet the computing power matching conditions are determined as the first vehicle node; the first vehicle node continues to participate in the subsequent concurrent processing sub-stream distribution, so that the concurrent processing sub-stream can be distributed to vehicle nodes with parallel computing resources, video memory carrying resources and sustainable execution time.

[0078] Preferably, when the concurrent processing sub-stream is sent to the first vehicle node, a concurrent processing sub-stream sending record is first generated based on the first vehicle node determination record. The concurrent processing sub-stream sending record includes the vehicle node identity digest of the first vehicle node, the local area network address of the first vehicle node, the concurrent processing sub-stream, the input data encapsulation method corresponding to the concurrent processing sub-stream, the output data encapsulation method corresponding to the concurrent processing sub-stream, and the concurrent processing sub-stream sending session identifier. The input data encapsulation method corresponding to the concurrent processing sub-stream originates from an independent instruction processing unit in the first instruction processing set and is used to define the input data organization form of the concurrent processing sub-stream. The output data encapsulation method corresponding to the concurrent processing sub-stream originates from an independent instruction processing unit in the first instruction processing set and is used to define the organization form of the local processing result formed after the concurrent processing sub-stream is completed. Subsequently, according to the local area network address of the first vehicle node and the intra-group communication route, the concurrent processing sub-stream and the concurrent processing sub-stream sending record are sent to the first vehicle node. The concurrent processing substream session identifier is used for corresponding processing when receiving the partial processing result returned by the first vehicle node. The input data encapsulation method corresponding to the concurrent processing substream is used to enable the first vehicle node to read the input data of the concurrent processing substream. The output data encapsulation method corresponding to the concurrent processing substream is used to enable the first vehicle node to return the partial processing result in the format required by the subsequent timing splicing operation.

[0079] Preferably, in the specific technical implementation of extracting the network bandwidth transmission indicators contained in the specification, firstly, specifications still within the valid time range are read from the specification registration area of ​​the temporary computing federation group. Then, network bandwidth transmission indicators, local area network addresses, intra-group communication routes, and underlying hardware status log verification records are extracted from each specification still within the valid time range to form a network bandwidth transmission indicator extraction record. The network bandwidth transmission indicators originate from the underlying hardware status logs and have been written into the specification registration area of ​​the temporary computing federation group along with the specifications. The underlying hardware status log verification record is used to indicate that the underlying hardware status logs from which the network bandwidth transmission indicators originate have undergone probe session matching, vehicle node identity digest comparison, and status sampling validity verification. The network bandwidth transmission indicator extraction record is further used to determine whether a vehicle node is suitable for undertaking data transmission between the multiple serially associated computing task fragments, avoiding the distribution of the multiple serially associated computing task fragments solely based on the online status of the vehicle nodes.

[0080] Preferably, before filtering based on the network bandwidth transmission index meeting the second preset value, the second preset value is configured according to the cascaded execution topology sequence and the multiple serially associated computing task fragments. The second preset value is configured based on inter-fragment transfer records between adjacent computing task fragments, intermediate tensor transfer descriptions, data flow direction records corresponding to data communication nodes, and the serial association order of the multiple serially associated computing task fragments. The second preset value is used to limit the lower limit of data transmission capacity that a vehicle node needs to achieve when undertaking data transmission between the multiple serially associated computing task fragments. Subsequently, the network bandwidth transmission index and the second preset value are compared using the same dimensions to obtain a network bandwidth matching record. The network bandwidth matching record is used to indicate whether a vehicle node can undertake intermediate tensor transfers between adjacent computing task fragments when transmitting task data along the intra-group communication route.

[0081] Preferably, when selecting candidate nodes that meet the transmission requirements from the temporary computing federation group, a transmission matching candidate node record is generated based on the network bandwidth matching record, the specification validity record, and the expected contribution duration. The transmission matching candidate node record includes a vehicle node identity digest, local area network address, intra-group communication route, network bandwidth transmission index, network bandwidth matching record, expected contribution duration, and specification validity record. If the network bandwidth matching record of a vehicle node indicates that the network bandwidth transmission index meets the second preset value, and the specification validity record indicates that the vehicle node's specification is still within the valid time range, then the vehicle node is written into the transmission matching candidate node record. The transmission matching candidate node record continues to be used to determine the second vehicle node and the third vehicle node, such that the determination of the second vehicle node and the third vehicle node is based on the data transmission requirements between the multiple serially associated computing task fragments, rather than from the homogeneous distribution rule.

[0082] Preferably, when determining the second vehicle node and the third vehicle node, the serial association order of the multiple serially associated computing task fragments is first read according to the cascaded execution topology sequence, and the output data inheritance relationship between the preceding computing task fragment and the remaining computing task fragments is determined according to the inter-fragment transmission record, thus obtaining a serial fragment assignment record. The serial fragment assignment record includes the preceding computing task fragment, the remaining computing task fragments, the fragment output boundary corresponding to the preceding computing task fragment, the fragment input boundary corresponding to the remaining computing task fragments, and the data communication nodes between adjacent computing task fragments. The preceding computing task fragment originates from the computing task fragment located before the cascaded execution topology sequence among the multiple serially associated computing task fragments, and the remaining computing task fragments originate from the computing task fragment located after the preceding computing task fragment among the multiple serially associated computing task fragments. Subsequently, the second vehicle node is selected for the preceding computation task fragment based on the transmission matching candidate node record, and the third vehicle node is selected for the remaining computation task fragments; the second vehicle node and the third vehicle node are respectively written into the serial fragmentation assignment record, so that the subsequent distribution process can be executed in the serial association order of the cascaded execution topology sequence.

[0083] Preferably, after selecting the second vehicle node and the third vehicle node, a fragmentation succession relationship verification is performed on the second vehicle node and the third vehicle node to obtain a fragmentation succession relationship verification record. This record is used to determine whether the intra-group communication route of the second vehicle node can transmit the output data of the preceding computation task fragment to the third vehicle node, and whether the intra-group communication route of the third vehicle node can receive the output data of the preceding computation task fragment. The fragmentation succession relationship verification record reads the inter-fragment transmission record, the data communication node, the network bandwidth transmission index, and the intra-group communication route for corresponding judgment. If the fragmentation succession relationship verification record indicates that the intra-group communication route between the second vehicle node and the third vehicle node does not meet the second preset value, the transmission matching candidate node record is reread, and the second and third vehicle nodes are re-determined. Through the fragmentation succession relationship verification record, the multiple serially associated computation task fragments maintain a sequential data succession relationship during assignment.

[0084] Preferably, when distributing tasks according to the serial association order of the cascaded execution topology sequence, the preceding computation task fragment, the remaining computation task fragment, the second vehicle node, the third vehicle node, and the data communication node are first read from the serial fragment distribution record, and a serial task fragment distribution record is generated. The serial task fragment distribution record includes a preceding computation task fragment distribution session identifier, a remaining computation task fragment distribution session identifier, an inter-fragment transfer record, the LAN address of the second vehicle node, the LAN address of the third vehicle node, and the serial association order in the cascaded execution topology sequence. Based on the serial task fragment distribution record, the preceding computation task fragment is distributed to the second vehicle node, and the remaining computation task fragment is distributed to the third vehicle node, so that the second vehicle node executes the preceding computation task fragment first, and then provides the input data required for subsequent execution to the third vehicle node through the inter-fragment transfer record corresponding to the data communication node.

[0085] Preferably, when the pre-computation task fragment is sent to the second vehicle node, the pre-computation task fragment, the fragment input boundary corresponding to the pre-computation task fragment, the fragment output boundary corresponding to the pre-computation task fragment, the data dependency edges within the fragment, and the pre-computation task fragment sending session identifier are all written into the pre-computation task fragment sending data packet. The pre-computation task fragment sending data packet is used to enable the second vehicle node to read input data according to the fragment input boundary corresponding to the pre-computation task fragment, execute the pre-computation task fragment according to the data dependency edges within the fragment, and generate pre-computation fragment output data that can be further processed by the third vehicle node according to the fragment output boundary corresponding to the pre-computation task fragment. The pre-computation fragment output data continues to be transmitted along the inter-fragment transmission record and the data communication node, so that the output of the pre-computation task fragment can become the input of the remaining computation task fragments.

[0086] Preferably, when the remaining computation task fragments are sent to the third vehicle node, the remaining computation task fragments, the fragment input boundaries corresponding to the remaining computation task fragments, the fragment output boundaries corresponding to the remaining computation task fragments, the data dependency edges within the fragments, the inter-fragment transmission records, and the remaining computation task fragment sending session identifier are all written into the remaining computation task fragment sending data packet. The remaining computation task fragment sending data packet is used to enable the third vehicle node to receive the preceding fragment output data according to the inter-fragment transmission records, and to continue execution using the preceding fragment output data as the input data for the remaining computation task fragments. After the third vehicle node completes the remaining computation task fragments, it forms serial fragment output data according to the fragment output boundaries corresponding to the remaining computation task fragments; the serial fragment output data continues to be part of the subsequent local processing results, and participates in the timing splicing operation together with the local processing results returned by the first vehicle node.

[0087] Preferably, after the concurrent processing sub-stream, the pre-computation task shard, and the remaining computation task shards are distributed, a task distribution consistency record is generated. This record records the correspondence between the concurrent processing sub-stream and the first vehicle node, the correspondence between the pre-computation task shard and the second vehicle node, the correspondence between the remaining computation task shards and the third vehicle node, and whether the serial association order in the cascaded execution topology sequence has been written into the serial task shard distribution record. If the task distribution consistency record indicates that the concurrent processing sub-stream, the pre-computation task shard, or the remaining computation task shard lacks a corresponding vehicle node, then the vehicle node selection is re-executed based on the computing power matching candidate node record or the transmission matching candidate node record. If the task distribution consistency record indicates that a certain computation task shard lacks an inter-shard transmission record, then the corresponding data communication node and inter-shard transmission record are added according to the cascaded execution topology sequence. The task distribution consistency record continues to be used for task source verification when subsequently receiving local processing results.

[0088] Preferably, in a multi-vehicle centralized parking application scenario, if the concurrent processing substream formed by the first instruction processing set contains multiple independent image preprocessing instructions or multiple independent sub-batch feature extraction instructions, then vehicle nodes whose tensor operation core count and available video memory capacity both meet the first preset value are selected as the first vehicle node based on the concurrent throughput dimension information; if the multiple serially associated computation task slices originate from the continuous inference layer in the edge-side large model inference task, or from the pre- and post-rendering stages in the video segment rendering task, then vehicle nodes that meet the second preset value are selected as the second vehicle node and the third vehicle node based on the network bandwidth transmission index. The independent image preprocessing instructions and the independent sub-batch feature extraction instructions both originate from the independent instruction processing units in the first instruction processing set, and the continuous inference layer and the pre- and post-rendering stages both originate from the slice instruction segments in the multiple serially associated computation task slices. The above distribution process differs from the method of distributing the target processing task equally to all online vehicle nodes. Instead, it first determines the first vehicle node based on the concurrent throughput requirements of the concurrent processing sub-stream, and then determines the second and third vehicle nodes based on the data transmission requirements between the multiple serially associated computing task fragments, so that a correspondence is formed between vehicle node selection and task fragment type.

[0089] Preferably, after the task distribution consistency record is generated, a distribution status tracking record is also generated based on the specification validity record, the first vehicle node determination record, the fragmentation acceptance relationship verification record, and the serial task fragmentation distribution record. The distribution status tracking record continuously records whether the specifications corresponding to the first vehicle node, the second vehicle node, and the third vehicle node are still within the valid time range during task execution, and whether the intra-group communication route can still carry concurrent processing sub-stream distribution and serial fragmentation transmission. If the distribution status tracking record indicates that the first vehicle node, the second vehicle node, or the third vehicle node experiences specification failure or intra-group communication route unreachable during task execution, the distribution status tracking record continues to provide data for subsequent parking maintenance status monitoring, concurrent processing sub-stream redirection distribution, and local processing result reception. Through the distribution status tracking record, the task fragmentation distribution result can maintain continuity with subsequent vehicle node dynamic exit processing.

[0090] Optionally, the step of receiving the partial processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node includes: Real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node; If the parking maintenance state is a normal parking state, then the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node are received. If a parking release signal is detected from the first vehicle node, it is determined that the parking maintenance state of the first vehicle node is abnormal, and the data receiving channel of the first vehicle node is blocked. Extract the concurrent processing sub-streams that have not yet been completed in the first vehicle node, redirect and distribute the concurrent processing sub-streams to other vehicle nodes in the temporary computing federation group that are in the normal parking state for continued execution, and receive the local processing results returned by the other vehicle nodes.

[0091] Preferably, the specific implementation process of real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node is as follows: After the concurrent processing sub-stream, the pre-computation task sharding, and the remaining computation task sharding are distributed, the task distribution consistency record, the distribution status tracking record, the first vehicle node determination record, the serial sharding assignment record, and the member registration area of ​​the temporary computation federation group are read first. The task distribution consistency record is used to provide the correspondence between the concurrent processing sub-stream and the first vehicle node, the correspondence between the pre-computation task sharding and the second vehicle node, and the correspondence between the remaining computation task sharding and the third vehicle node; the distribution status tracking record is used to provide the specification validity record and intra-group communication routing status of the first vehicle node, the second vehicle node, and the third vehicle node during task execution; the member registration area of ​​the temporary computation federation group is used to provide the vehicle node identity digest, local area network address, and intra-group communication routing of the first vehicle node, the second vehicle node, and the third vehicle node. Based on the task distribution consistency record, the distribution status tracking record, the first vehicle node determination record, the serial sharding and dispatch record, and the member registration area of ​​the temporary computing federation group, a parking maintenance monitoring record is generated. The parking maintenance monitoring record is used to periodically read the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node, so that the subsequent local processing result reception process can maintain correspondence with the current parking status of the first vehicle node, the second vehicle node, and the third vehicle node.

[0092] Preferably, the parking maintenance monitoring record includes a vehicle node identity summary, local area network address, intra-group communication route, task issuance session identifier, parking status sampling timestamp, parking status feedback field, local area network reachability feedback field, and task execution progress field. The task issuance session identifier includes a concurrent processing sub-stream issuance session identifier, a pre-computation task fragment issuance session identifier, and a remaining computation task fragment issuance session identifier. The parking status feedback field records whether the vehicle node is still in a parking maintenance state; the local area network reachability feedback field records whether the vehicle node can still interact with data through the corresponding intra-group communication route; and the task execution progress field records the execution progress of the corresponding vehicle node for the concurrent processing sub-stream, the pre-computation task fragment, or the remaining computation task fragment. The parking maintenance monitoring records are indexed according to the vehicle node identity digest, so that the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node can correspond to the concurrent processing sub-stream, the pre-computation task segment, or the remaining computation task segment they undertake, respectively, avoiding receiving partial processing results only based on the online status of the vehicle node.

[0093] Preferably, when monitoring the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node in real time, parking status query messages are sent to the first vehicle node, the second vehicle node, and the third vehicle node respectively according to the parking maintenance monitoring record, and parking status feedback messages returned by the first vehicle node, the second vehicle node, and the third vehicle node are received. The parking status query message carries a vehicle node identity digest, a task distribution session identifier, and a parking status sampling timestamp; the parking status feedback message carries a parking status feedback field, a local area network reachability feedback field, and a task execution progress field. After parsing the fields of the parking status feedback message, a parking maintenance status judgment record is formed; the parking maintenance status judgment record is used to determine whether the first vehicle node, the second vehicle node, and the third vehicle node are still in a normal parking state, and continues to serve as the basis for receiving the partial processing result, blocking the data receiving channel, or redirecting the distribution of the concurrent processing sub-stream.

[0094] Preferably, if the parking maintenance status judgment record indicates that the parking status feedback field indicates parking hold, the local area network reachability feedback field indicates reachability, and the task execution progress field matches the corresponding task distribution session identifier, then the parking maintenance status of the corresponding vehicle node is determined to be a normal parking status. For the first vehicle node, the task execution progress field corresponds to the concurrent processing sub-stream distribution session identifier; for the second vehicle node, the task execution progress field corresponds to the pre-computation task fragment distribution session identifier; for the third vehicle node, the task execution progress field corresponds to the remaining computation task fragment distribution session identifier. Under the condition that the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node is a normal parking status, the partial processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node are received according to the local area network address and the intra-group communication route, so that the partial processing results originate from the vehicle node that is still in a normal parking status and whose task distribution session identifier matches.

[0095] Preferably, when receiving the partial processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, the source of the partial processing results is verified according to the task distribution consistency record to obtain a partial processing result source verification record. The partial processing result source verification record includes a vehicle node identity digest matching result, a task distribution session identifier matching result, a partial processing result type matching result, and an output timestamp record. The vehicle node identity digest matching result is used to determine whether the vehicle node returning the partial processing result is the first vehicle node, the second vehicle node, or the third vehicle node. The task distribution session identifier matching result is used to determine whether the partial processing result corresponds to the distributed concurrent processing sub-stream, the pre-computation task fragment, or the remaining computation task fragment. The partial processing result type matching result is used to distinguish the partial processing results corresponding to the concurrent processing sub-stream, the partial processing results corresponding to the pre-computation task fragment, and the partial processing results corresponding to the remaining computation task fragment. The partial processing result source verification record is further used in subsequent timing splicing operations, enabling the partial processing results to be organized according to the task source and output timestamp.

[0096] Preferably, if a parking release signal sent by the first vehicle node is detected, the parking release signal is first parsed to obtain a parking release parsing record. The parking release parsing record includes the vehicle node identity digest of the first vehicle node, a parking release timestamp, a parking release status field, a concurrent processing sub-flow distribution session identifier, and a task execution progress field. The parking release status field indicates that the first vehicle node has exited the parking maintenance state or is in the preparation process for exiting the parking maintenance state. The parking release timestamp determines the time correspondence between the parking release signal and the concurrent processing sub-flow execution process. The task execution progress field determines the execution progress of the first vehicle node on the concurrent processing sub-flow. The parking maintenance status judgment record is updated according to the parking release parsing record, and the parking maintenance status of the first vehicle node is determined to be abnormal.

[0097] Preferably, after determining that the parking maintenance state of the first vehicle node is abnormal, a data receiving channel blocking record is generated based on the local area network address of the first vehicle node, the intra-group communication route, and the concurrent processing sub-stream distribution session identifier. The data receiving channel blocking record then blocks the data receiving channel of the first vehicle node. The data receiving channel blocking record is used to mark that the local processing result of the first vehicle node will no longer directly enter the subsequent timing splicing operation, and to mark that the concurrent processing sub-stream distribution session identifier corresponding to the first vehicle node has entered the redirection distribution preparation state. The redirection distribution preparation state is used to trigger the formation of the remaining execution records of the subsequent concurrent processing sub-streams. After blocking the data receiving channel of the first vehicle node, the local processing result that the first vehicle node has returned and has passed the verification of the local processing result source verification record is still retained. Simultaneously, the reception of local processing results returned by the first vehicle node after the parking maintenance state is abnormal and that have not passed the verification of the local processing result source verification record is stopped, in order to reduce the entry of incomplete data into the subsequent timing splicing operation during the parking release process.

[0098] Preferably, when extracting the concurrent processing sub-streams that have not yet been completed in the first vehicle node, the following steps are taken: first, read the concurrent processing sub-stream distribution record, the concurrent processing sub-stream distribution session identifier, the task execution progress field in the parking release parsing record, and the partial processing result that has been verified by the partial processing result source verification record, to form the remaining execution record of the concurrent processing sub-stream. The remaining execution record of the concurrent processing sub-stream is used to distinguish between independent instruction processing units that have been completed and returned partial processing results in the concurrent processing sub-stream, and independent instruction processing units that have not yet been completed. Based on the remaining execution record of the concurrent processing sub-stream, the concurrent processing sub-streams that have not yet been completed in the first vehicle node are extracted; the concurrent processing sub-streams that have not yet been completed retain their original input data encapsulation method, output data encapsulation method, and concurrent processing sub-stream distribution session identifier, and continue to be used for subsequent redirection and distribution, so that the task object of the redirection and distribution corresponds to the concurrent processing sub-stream originally distributed to the first vehicle node.

[0099] Preferably, when determining other vehicle nodes in the normal parking state within the temporary computing federation group, the member registration area, specification registration area, specification validity record, and parking maintenance status judgment record of the temporary computing federation group are read first. The first vehicle node and vehicle nodes whose parking maintenance status is not normal parking state are then excluded from the temporary computing federation group. Subsequently, the concurrent throughput dimension information, intra-group communication routes, and expected contribution duration of the remaining vehicle nodes are filtered to obtain redirection candidate vehicle node records. These redirection candidate vehicle node records include a vehicle node identity digest, local area network address, concurrent throughput dimension information, intra-group communication routes, expected contribution duration, and parking maintenance status judgment record. These redirection candidate vehicle node records are then used to determine other vehicle nodes in the normal parking state, enabling the concurrent processing subflow that has not yet been completed to be transferred to vehicle nodes that still have concurrent execution capabilities and data return paths.

[0100] Preferably, when the concurrent processing subflow is redirected and distributed to other vehicle nodes in the temporary computing federation group that are in the normal parking state for continued execution, a concurrent processing subflow redirection distribution record is generated based on the remaining execution record of the concurrent processing subflow and the redirection candidate vehicle node record. The concurrent processing subflow redirection distribution record includes the concurrent processing subflow that has not yet been completed, the vehicle node identity digest of the redirecting vehicle node, the local area network address of the redirecting vehicle node, the intra-group communication route of the redirecting vehicle node, the original concurrent processing subflow distribution session identifier, and the redirection distribution session identifier. The original concurrent processing subflow distribution session identifier is used to associate with the concurrent processing subflow previously undertaken by the first vehicle node, and the redirection distribution session identifier is used to mark the continued execution process of the concurrent processing subflow that has not yet been completed on the other vehicle nodes. The concurrent processing subflow that has not yet been completed and the concurrent processing subflow redirection distribution record are sent to the other vehicle nodes according to the local area network address and intra-group communication route of the redirecting vehicle node.

[0101] Preferably, after receiving the concurrent processing sub-stream that has not yet been completed, the other vehicle nodes read the input data encapsulation method, output data encapsulation method, and redirection session identifier of the concurrent processing sub-stream according to the concurrent processing sub-stream redirection distribution record, and read the input data according to the input data encapsulation method to continue executing the concurrent processing sub-stream that has not yet been completed. After completing the concurrent processing sub-stream that has not yet been completed, the other vehicle nodes form a redirection partial processing result according to the output data encapsulation method, and return the redirection partial processing result and the redirection session identifier together. The redirection partial processing result continues to enter the partial processing result source verification record, and corresponds to the concurrent processing sub-stream through the original concurrent processing sub-stream distribution session identifier, so that subsequent timing splicing operations can identify that the redirection partial processing result belongs to the concurrent processing sub-stream originally distributed to the first vehicle node.

[0102] Preferably, when receiving the partial processing result returned by the other vehicle nodes, the partial processing result returned by the other vehicle nodes, the redirection session identifier, the original concurrent processing sub-flow distribution session identifier, and the remaining execution record of the concurrent processing sub-flow are verified to obtain a redirection partial processing result verification record. The redirection partial processing result verification record is used to determine whether the partial processing result returned by the other vehicle nodes corresponds to the concurrent processing sub-flow that has not yet been completed, and to determine whether the partial processing result returned by the other vehicle nodes is duplicated with the partial processing result already returned by the first vehicle node. If the redirection partial processing result verification record indicates that there is a duplicate partial processing result, the partial processing result corresponding to the remaining execution record of the concurrent processing sub-flow is retained, and the duplicate partial processing result is marked for deduplication; the deduplication mark is continued to be written into the partial processing result source verification record to prevent duplicate partial processing results from entering subsequent timing splicing operations. If the redirection partial processing result verification record indicates that there is no duplicate partial processing result, the partial processing result returned by the other vehicle nodes is written into the input record of subsequent timing splicing operations.

[0103] Preferably, for the second vehicle node and the third vehicle node, when monitoring the parking maintenance status in real time, it is also determined whether they are in a normal parking state based on the parking maintenance monitoring record, the serial task sharding distribution record, and the inter-sharding transmission record. If the parking maintenance status of the second vehicle node or the third vehicle node is in a normal parking state, the local processing result returned by the second vehicle node or the third vehicle node continues to be received, and the source is verified based on the pre-computation task sharding distribution session identifier or the remaining computation task sharding distribution session identifier. If the parking maintenance status of the second vehicle node or the third vehicle node is abnormal, a local processing result receiving status record is generated based on the serial task sharding distribution record and the inter-sharding transmission record; the local processing result receiving status record is used to mark whether the local processing result of the corresponding computation task shard has been received, whether it has passed the local processing result source verification record, and whether it still has the conditions to continue receiving. The local processing result receiving status record continues to be used for subsequent local processing result receiving and timing splicing operations.

[0104] Preferably, in a multi-vehicle centralized parking application scenario, the first vehicle node undertakes the concurrent processing sub-stream, the second vehicle node undertakes the pre-computation task segmentation, and the third vehicle node undertakes the remaining computation task segmentation. When the first vehicle node enters an abnormal parking maintenance state due to a parking release signal, the data receiving channel blocking record first blocks the data receiving channel of the first vehicle node, the remaining execution record of the concurrent processing sub-stream then separates the concurrent processing sub-streams that have not yet been completed from the concurrent processing sub-streams, the redirection candidate vehicle node record then selects other vehicle nodes in the normal parking state from the temporary computation federation group, and the concurrent processing sub-stream redirection distribution record then distributes the concurrent processing sub-streams that have not yet been completed to the other vehicle nodes for continued execution. The above processing differs from the method of waiting for all vehicle nodes to return results before uniformly determining failure. Instead, it continuously processes the data receiving channel blocking, the extraction of the concurrent processing sub-streams that have not yet been completed, and the redirection distribution when the parking release signal appears, thereby reducing the impact of the first vehicle node exiting the parking state on the global execution result generation process of the target processing task.

[0105] Optionally, the step of real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node includes: Obtain the battery management system interface protocol, and periodically extract the remaining power battery capacity of the first vehicle node, the second vehicle node, and the third vehicle node through the battery management system interface protocol; The remaining power of the power battery is compared and analyzed with the preset power protection lower limit value; If the remaining power of the power battery is not lower than the power protection lower limit value, then the corresponding vehicle node is determined to meet the computing power maintenance condition of the parking maintenance state. If the remaining power of the power battery is lower than the power protection lower limit, it is determined that the corresponding vehicle node does not meet the computing power maintenance condition, a computing power cancellation command is generated and issued, and the corresponding vehicle node is forcibly triggered to exit the parking maintenance state.

[0106] Preferably, in the specific technical implementation of real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node, after generating the parking maintenance monitoring record, the vehicle node identity digest, local area network address, intra-group communication route, task issuance session identifier, and specification validity record corresponding to the first vehicle node, the second vehicle node, and the third vehicle node are read first. A power sampling plan is then generated based on the vehicle node identity digest, the local area network address, the intra-group communication route, the task issuance session identifier, and the specification validity record. The power sampling plan includes the vehicle node identity digest, the local area network address, the intra-group communication route, the battery management system interface protocol, the power battery remaining power sampling field, the power sampling period, and the task issuance session identifier. The power sampling plan is used to limit from which vehicle nodes the power battery remaining power is read, according to which power battery remaining power sampling field the power battery remaining power is read, and which task issuance session identifier corresponds to the read power battery remaining power. By associating the power sampling plan with the parking maintenance monitoring record, the remaining power of the power battery is no longer treated as an isolated vehicle state value, but can be directly mapped to the concurrent processing sub-stream undertaken by the first vehicle node, the pre-computation task segment undertaken by the second vehicle node, and the remaining computation task segment undertaken by the third vehicle node.

[0107] Preferably, when obtaining the battery management system interface protocol, the vehicle node identity digests and local area network addresses of the first vehicle node, the second vehicle node, and the third vehicle node are first read from the member registration area of ​​the temporary computing federation group. Then, the vehicle node identity digests are matched with the pre-registered battery management system interface protocol version records of the vehicle nodes to form a battery management system interface protocol confirmation record. The battery management system interface protocol version record is used to describe the name of the remaining battery power field, the unit of the remaining battery power field, the update timestamp of the remaining battery power field, the verification method of the remaining battery power field, and the return method of the abnormal field of the remaining battery power that the vehicle node can provide. The battery management system interface protocol confirmation record is used to indicate which battery management system interface protocol the corresponding vehicle node uses to read the remaining battery power. The battery management system interface protocol confirmation record is further written into the power sampling plan to ensure that the remaining battery power sampling field in the power sampling plan is consistent with the name, unit, and verification method of the remaining battery power field that the corresponding vehicle node can actually return.

[0108] Preferably, when periodically extracting the remaining battery power of the first vehicle node, the second vehicle node, and the third vehicle node through the battery management system interface protocol, a battery status query message is sent to the first vehicle node, the second vehicle node, and the third vehicle node respectively according to the battery sampling plan, and a battery status feedback message returned by the first vehicle node, the second vehicle node, and the third vehicle node based on the battery management system interface protocol is received. The battery status query message carries a vehicle node identity digest, a task issuance session identifier, a battery sampling timestamp, and a battery remaining power sampling field; the battery status feedback message carries a vehicle node identity digest, a task issuance session identifier, the remaining battery power, the unit of the remaining battery power field, and a battery remaining power update timestamp. After parsing the fields of the battery status feedback message, a battery remaining power sampling record is formed; the battery remaining power sampling record continues to participate in subsequent comparison and analysis, so that the remaining battery power can enter the parking maintenance status judgment process in the form of data with a vehicle node identity digest and a task issuance session identifier.

[0109] Preferably, after forming the remaining battery power sampling record, the sampling validity of the remaining battery power sampling record is first verified to obtain a remaining battery power sampling validity record. The remaining battery power sampling validity record includes vehicle node identity digest matching results, task distribution session identifier matching results, remaining battery power field unit matching results, and remaining battery power update timestamp matching results. The vehicle node identity digest matching results are used to determine whether the remaining battery power originates from the first vehicle node, the second vehicle node, or the third vehicle node. The task distribution session identifier matching results are used to determine whether the remaining battery power corresponds to the currently executing concurrent processing sub-stream, the pre-computation task segment, or the remaining computation task segment. The remaining battery power field unit matching results are used to determine whether the remaining battery power uses a unit consistent with the power protection lower limit value. The remaining battery power update timestamp matching results are used to determine whether the remaining battery power belongs to the current power sampling period. Only the remaining battery power after verification by the remaining battery power sampling validity record will continue to be compared and analyzed with the power protection lower limit value.

[0110] Preferably, the lower limit value for power protection is preset when the temporary computing federation group is established, and is written into the power protection lower limit configuration record of the corresponding vehicle node along with the parking maintenance monitoring record. When setting the lower limit value for power protection, the power battery capacity level of the vehicle node, the basic operating power reserve of the vehicle node, the control power reserve required for the vehicle node to exit the parking maintenance state, and the estimated duration of the target processing task are read, and a power protection lower limit configuration record is formed based on the power battery capacity level, the basic operating power reserve of the vehicle node, the control power reserve required for the vehicle node to exit the parking maintenance state, and the estimated duration of the target processing task. The power battery capacity level is used to define the capacity reference range of the lower limit value of the power protection. The vehicle node basic operation reserved power is used to define the remaining power battery power required for the vehicle node to maintain local basic services. The control reserved power required for the vehicle node to exit the parking maintenance state is used to define the remaining power battery power required for the vehicle node to perform the exit parking maintenance state processing. The expected duration of the target processing task is used to define the execution time range that the vehicle node needs to cover when continuing to participate in the concurrent processing sub-stream, the pre-computation task sharding, or the remaining computation task sharding. The power protection lower limit configuration record is used to explain the source of the power protection lower limit value. The power protection lower limit value is used to define the boundary of the remaining power battery power that the vehicle node still needs to retain during the participation of computing power collaboration. The power protection lower limit configuration record continues to be associated with the power battery remaining power sampling record, so that subsequent comparison analysis can be based on the power protection lower limit value of the corresponding vehicle node, rather than using a fixed power judgment value that is exactly the same for all vehicle nodes.

[0111] Preferably, when comparing the remaining battery power with a pre-set lower limit value for battery power protection, the remaining battery power, verified by the sampling validity record of the remaining battery power, is first read. Then, the lower limit configuration record for battery power protection of the corresponding vehicle node is read, and the lower limit value for battery power protection is extracted from the configuration record. Subsequently, the remaining battery power and the lower limit value for battery power protection are compared in the same unit to form a battery power comparison record. The battery power comparison record includes a vehicle node identity digest, a task distribution session identifier, the remaining battery power, the lower limit value for battery power protection, and the battery power comparison result. The battery power comparison result is used to characterize whether the remaining battery power is not lower than the lower limit value for battery power protection. The battery power comparison record is further used to generate a computing power maintenance condition judgment record, so that the source of the judgment of the computing power maintenance condition has a traceable data field.

[0112] Preferably, if the remaining power battery charge is not lower than the power protection lower limit, a computing power maintenance condition judgment record is generated based on the power comparison record, and the corresponding vehicle node is marked in the computing power maintenance condition judgment record as meeting the computing power maintenance condition of the parking maintenance state. For the first vehicle node, the computing power maintenance condition judgment record continues to be associated with the concurrent processing sub-stream distribution session identifier, indicating that the first vehicle node still has the power basis to continue executing the concurrent processing sub-stream; for the second vehicle node, the computing power maintenance condition judgment record continues to be associated with the pre-computation task shard distribution session identifier, indicating that the second vehicle node still has the power basis to continue executing the pre-computation task shard; for the third vehicle node, the computing power maintenance condition judgment record continues to be associated with the remaining computation task shard distribution session identifier, indicating that the third vehicle node still has the power basis to continue executing the remaining computation task shard. The computing power maintenance condition judgment record is then written into the parking maintenance state judgment record, so that the judgment of the normal parking state simultaneously includes the parking state feedback field, the local area network reachability feedback field, and the computing power maintenance condition corresponding to the remaining power battery charge.

[0113] Preferably, if the remaining power battery charge is lower than the power protection lower limit, a computing power maintenance condition judgment record is generated based on the power comparison record, and the corresponding vehicle node is marked as not meeting the computing power maintenance condition in the computing power maintenance condition judgment record. After marking the corresponding vehicle node as not meeting the computing power maintenance condition, the task distribution session identifier, intra-group communication route, local area network address, and task execution progress field of the corresponding vehicle node are read, and a computing power deregistration preparation record is generated based on the task distribution session identifier, intra-group communication route, local area network address, and task execution progress field. The computing power deregistration preparation record is used to indicate that the corresponding vehicle node needs to exit the parking maintenance state because the remaining power battery charge is lower than the power protection lower limit, and is used to limit the vehicle node identity digest, task distribution session identifier, and deregistration reason field that the subsequent computing power deregistration command needs to carry. The computing power deregistration preparation record is further used to generate the computing power deregistration command, so that the computing power deregistration command is directly linked to the judgment result that the remaining power battery charge is lower than the power protection lower limit.

[0114] Preferably, when generating the computing power deregistration instruction, a computing power deregistration instruction issuance record is formed by encapsulating the vehicle node identity digest, local area network address, intra-group communication route, task issuance session identifier, power comparison record, computing power maintenance condition judgment record, and deregistration reason field in the computing power deregistration preparation record. The computing power deregistration instruction issuance record is used to determine the recipient, issuance path, and corresponding task object of the computing power deregistration instruction; the deregistration reason field records that the remaining power battery power of the corresponding vehicle node is lower than the power protection lower limit value; and the task issuance session identifier indicates which task object among the concurrent processing sub-stream, the pre-computation task slice, or the remaining computation task slice the computing power deregistration instruction corresponds to. According to the computing power deregistration instruction issuance record, the computing power deregistration instruction is issued to the corresponding vehicle node through the intra-group communication route, causing the corresponding vehicle node to perform data processing before exiting the parking maintenance state based on the computing power deregistration instruction.

[0115] Preferably, after the computing power deregistration command is issued to the corresponding vehicle node, the corresponding vehicle node reads the locally executing task distribution session identifier, task execution progress field, and locally stored partial processing results according to the computing power deregistration command, and generates a computing power deregistration response record. The computing power deregistration response record includes a vehicle node identity digest, a task distribution session identifier, a task execution progress field, a partial processing result temporary storage status, and an exit parking maintenance status feedback field; the task execution progress field is used to determine whether the corresponding task object has been completed, the partial processing result temporary storage status is used to determine whether there are any partial processing results that have not yet been returned locally, and the exit parking maintenance status feedback field is used to indicate that the corresponding vehicle node has started to execute the process of exiting the parking maintenance status. After receiving the computing power deregistration response record, the parking maintenance status judgment record is updated according to the computing power deregistration response record, and the parking maintenance status of the corresponding vehicle node is marked as parking maintenance status abnormal; the parking maintenance status abnormality is further used to trigger the local processing result receiving status update or data receiving channel blocking processing of the corresponding vehicle node.

[0116] Preferably, for the first vehicle node, when the computing power maintenance condition judgment record indicates that the first vehicle node does not meet the computing power maintenance condition, the computing power deregistration response record continues to be used to trigger the data receiving channel blocking record and the remaining execution record of the concurrent processing sub-stream. If the computing power deregistration response record indicates that the concurrent processing sub-stream has not yet been completed, the concurrent processing sub-streams that have not yet been completed in the first vehicle node are extracted according to the task execution progress field, and the concurrent processing sub-streams that have not yet been completed are redirected and distributed to other vehicle nodes in the temporary computing federation group that are in the normal parking state for continued execution. Thus, the judgment result that the remaining power battery charge is lower than the power protection lower limit value can be directly connected to the redirection and distribution process of the concurrent processing sub-stream, reducing the impact of the first vehicle node exiting the parking maintenance state due to insufficient power on the execution process of the concurrent processing sub-stream.

[0117] Preferably, for the second vehicle node and the third vehicle node, when the computing power maintenance condition judgment record indicates that the second vehicle node or the third vehicle node does not meet the computing power maintenance condition, a local processing result reception status record is generated based on the computing power cancellation response record, the serial task sharding distribution record, and the inter-sharding transmission record. The local processing result reception status record is used to mark whether the preceding computing task shard or the remaining computing task shard has been completed, whether a local processing result has been returned, and whether there is inter-sharding data that has not yet been transmitted through the inter-sharding transmission record; the inter-sharding data comes from the output data of the preceding computing task shard or the input data of the remaining computing task shard, and is further used to determine whether the data connection relationship between the preceding computing task shard and the remaining computing task shard is complete. If the local processing result reception status record indicates that the corresponding computing task segment has not yet been completed, then the local processing result reception status of the corresponding computing task segment is marked according to the serial task segment issuance record, so that subsequent local processing result reception and timing splicing operations can identify that the second vehicle node or the third vehicle node exits the parking maintenance state because it does not meet the computing power maintenance conditions.

[0118] Preferably, in a multi-vehicle centralized parking application scenario, the first vehicle node, the second vehicle node, and the third vehicle node return the remaining power battery charge according to the power sampling plan during task execution. If the remaining power battery charge of the first vehicle node is not lower than the power protection lower limit value, the computing power maintenance condition judgment record marks the first vehicle node as meeting the computing power maintenance condition of the parking maintenance state, and continues to receive the partial processing results returned by the first vehicle node; if the remaining power battery charge of the first vehicle node is lower than the power protection lower limit value, the computing power deregistration instruction issuance record issues the computing power deregistration instruction to the first vehicle node, and triggers subsequent redirection distribution through the computing power deregistration response record, the data receiving channel blocking record, and the concurrent processing substream remaining execution record. This process differs from the method of judging availability solely based on whether the vehicle node is online. Instead, while the vehicle node is still in a network reachable state, it further uses the comparison result between the remaining power battery charge and the power protection lower limit value to determine whether the vehicle node is suitable to continue providing computing power, thereby connecting the parking computing power coordination process with the power status changes of the vehicle node.

[0119] Optionally, the step of performing a time-series concatenation operation on the local processing results to generate the global execution result of the target processing task includes: The output timestamps of the local processing results are aligned, and the time-series splicing operation is performed on the local processing results based on the aligned output timestamps to obtain the initial splicing result. A memory-safe overwrite instruction is issued to the first vehicle node, the second vehicle node, and the third vehicle node to erase the local processing results on their respective localities. Receive the underlying storage erase credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node in response to the memory-secure overwrite instruction; Under the condition that the underlying storage erasure certificate is valid, the initial splicing result is confirmed as the global execution result of the target processing task and output encapsulated.

[0120] Preferably, the specific implementation process of performing time-series splicing operation on the local processing results is as follows: After receiving the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, the task distribution consistency record, the local processing result source verification record, the redirected local processing result verification record, the serial task fragment distribution record, the cascaded execution topology sequence, and the inter-fragment transmission record are read first. Then, a local processing result reception summary record is generated based on the task distribution consistency record, the local processing result source verification record, the redirected local processing result verification record, the serial task fragment distribution record, the cascaded execution topology sequence, and the inter-fragment transmission record. The local processing result reception summary record includes a vehicle node identity digest, a task distribution session identifier, a local processing result type, an output timestamp of the local processing result, the task object corresponding to the local processing result, and the local processing result reception status. The local processing result reception summary record is used to indicate that each local processing result originates from the concurrent processing sub-stream, the pre-computation task fragment, the remaining computation task fragment, or the concurrent processing sub-stream after redirection. By first forming a summary record of the local processing results, subsequent time-series splicing operations can organize the local processing results according to the task source, task distribution session identifier, and output timestamp of the local processing results, instead of simply merging the data returned by each vehicle node in sequence.

[0121] Preferably, before aligning the output timestamps of the local processing results, the output timestamps of the local processing results in the local processing result reception summary record are first validated to obtain a local processing result output timestamp validity record. The local processing result output timestamp validity record includes a task distribution session identifier matching result, a local processing result output timestamp field integrity result, a local processing result output timestamp to task object matching result, and a local processing result output timestamp sampling source record. The task distribution session identifier matching result is used to determine whether the output timestamp of the local processing result corresponds to the already distributed concurrent processing sub-stream, the pre-computation task fragment, or the remaining computation task fragment. The local processing result output timestamp field integrity result is used to determine whether the local processing result output timestamp has a sortable data format. The local processing result output timestamp to task object matching result is used to determine whether the local processing result output timestamp is consistent with the corresponding local processing result task object. The local processing result output timestamp sampling source record indicates that the local processing result output timestamp originates from the local processing result returned by the corresponding vehicle node. The validity record of the output timestamp of the local processing result continues to participate in the alignment process of the output timestamp of the local processing result, so that the local processing result that fails the timestamp validity check will not directly enter the timing splicing operation.

[0122] Preferably, when aligning the output timestamps of the local processing results, the local processing result reception summary record and the local processing result output timestamp validity record are read first, and the local processing results that pass the verification of the local processing result output timestamp validity record are written into the splicing input queue. The splicing input queue includes a concurrent processing sub-stream result queue, a pre-computation task fragment result queue, a remaining computation task fragment result queue, and a redirected local processing result queue. The concurrent processing sub-stream result queue is used to store the local processing results corresponding to the concurrent processing sub-streams originating from the first vehicle node or the other vehicle nodes. The pre-computation task fragment result queue is used to store the local processing results corresponding to the pre-computation task fragments originating from the second vehicle node. The remaining computation task fragment result queue is used to store the local processing results corresponding to the remaining computation task fragments originating from the third vehicle node. The redirected local processing result queue is used to store the local processing results corresponding to the concurrent processing sub-streams after redirection and distribution. Subsequently, using the cascaded execution topology sequence and the task distribution consistency record as references, the output timestamps of the local processing results in the splicing input queue are mapped to the same task time base to form a timestamp alignment record. The same task time base is used to describe the time reference relationship of the concurrent processing sub-stream, the pre-computation task fragment, the remaining computation task fragment, and the concurrent processing sub-stream after redirection and distribution in the same target processing task. The timestamp alignment record includes the arrangement position of the same task time base and the output timestamps of the local processing results under the same task time base. The timestamp alignment record is further used to determine the arrangement position of the local processing results in the timing splicing operation.

[0123] Preferably, after forming the timestamp alignment record, the timestamp alignment record is subjected to result integrity verification to obtain a splicing integrity verification record. The splicing integrity verification record is used to determine whether the local processing result corresponding to the concurrent processing sub-stream has covered the independent instruction processing unit in the first instruction processing set, and to determine whether the local processing result corresponding to the pre-computation task slice and the local processing result corresponding to the remaining computation task slice maintain a sequential relationship along the cascaded execution topology sequence. During the result integrity verification, the task distribution consistency record is used to check whether the concurrent processing sub-stream, the pre-computation task slice, and the remaining computation task slice all have corresponding local processing results, and the inter-slice transfer record is used to check whether the output data of the pre-computation task slice has been used as input data for the remaining computation task slices. The splicing integrity verification record continues to serve as the input constraint for the timing splicing operation, ensuring that the initial splicing result retains the operator execution dependency and task slice continuity relationship corresponding to the target processing task.

[0124] Preferably, when performing the time-series splicing operation on the local processing results based on the output timestamps of the aligned local processing results, the arrangement position of the local processing results corresponding to the concurrent processing sub-stream in the splicing input queue is first determined according to the timestamp alignment record. Then, the serial arrangement position between the local processing results corresponding to the pre-computation task slice and the local processing results corresponding to the remaining computation task slice is determined according to the cascaded execution topology sequence. Subsequently, according to the splicing integrity verification record, the local processing results corresponding to the concurrent processing sub-stream, the local processing results corresponding to the pre-computation task slice, and the local processing results corresponding to the remaining computation task slice are written into the initial splicing result. The initial splicing result includes concurrent result fragments, serial result fragments, task distribution session identifiers, output timestamp alignment information, and local processing result source information; the concurrent result fragments originate from the local processing results corresponding to the concurrent processing sub-stream, the serial result fragments originate from the local processing results corresponding to the pre-computation task slice and the remaining computation task slice, the output timestamp alignment information originates from the timestamp alignment record, and the local processing result source information originates from the local processing result reception summary record. The initial splicing result then enters the underlying storage erasure credential validity verification process, so that the initial splicing result can establish a correspondence with the local processing result cleanup status of each vehicle node before being confirmed as the global execution result.

[0125] Preferably, after obtaining the initial splicing result, a memory-safe overwrite plan is generated based on the received summary record of the local processing result, the task distribution consistency record, and the initial splicing result. The memory-safe overwrite plan includes the vehicle node identity digest of the first vehicle node, the vehicle node identity digest of the second vehicle node, the vehicle node identity digest of the third vehicle node, the task distribution session identifier, a description of the local local processing result storage location, a description of the local intermediate tensor cache location, a memory-safe overwrite method, and a memory-safe overwrite completion feedback path. The description of the local local processing result storage location is used to define the storage location of the local processing result to be erased on the vehicle node. The description of the local intermediate tensor cache location is used to define the location of the local intermediate tensor cache formed during the execution of the concurrent processing sub-stream, the pre-computation task shard, or the remaining computation task shard. The memory-safe overwrite method is used to define a local storage cleanup method combining overwrite processing and release flag update. The memory-safe overwrite completion feedback path is used to define the intra-group communication route used when the underlying storage erase certificate is returned. The memory-safe overwrite plan continues to be used to generate memory-safe overwrite instructions, enabling the memory-safe overwrite instructions to correspond to the received partial processing results and the formed initial splicing results.

[0126] Preferably, when issuing memory-safe overwrite instructions to the first vehicle node, the second vehicle node, and the third vehicle node, the memory-safe overwrite instructions for the corresponding vehicle node are encapsulated according to the memory-safe overwrite plan. The memory-safe overwrite instruction includes a vehicle node identity digest, a task issuance session identifier, a description of the local partial processing result storage location, a description of the local intermediate tensor cache location, a memory-safe overwrite method, an overwrite execution timestamp, and a memory-safe overwrite completion feedback path. The vehicle node identity digest is used to limit the recipient of the memory-safe overwrite instruction; the task issuance session identifier is used to limit the task object corresponding to the partial processing result to be erased; the description of the local partial processing result storage location and the description of the local intermediate tensor cache location are used to limit the storage location where the memory-safe overwrite instruction is executed locally on the vehicle node; and the overwrite execution timestamp is used to record the time and position at which the vehicle node executes the memory-safe overwrite instruction. After issuing the memory-secure overwrite instruction to the first vehicle node, the second vehicle node, and the third vehicle node respectively according to the intra-group communication route, the first vehicle node, the second vehicle node, and the third vehicle node perform local storage cleanup according to the local local processing result storage location description and the local intermediate tensor cache location description, so as to erase their local local local processing results and the local intermediate tensor cache corresponding to the local processing results.

[0127] Preferably, when the first vehicle node, the second vehicle node, and the third vehicle node respond to the memory-safe overwrite instruction, they first locate the local partial processing result and the local intermediate tensor cache according to the task-issuing session identifier, then perform overwrite processing on the local partial processing result and the local intermediate tensor cache according to the memory-safe overwrite method, and update the release flag of the corresponding storage location after the overwrite processing, forming a local overwrite execution record. The local overwrite execution record includes the task-issuing session identifier, a description of the storage location of the local partial processing result, a description of the location of the local intermediate tensor cache, an overwrite processing status, a release flag update status, and an overwrite execution timestamp. The local overwrite execution record is further used to generate an underlying storage erasure certificate, so that the underlying storage erasure certificate can indicate that the vehicle node has performed local storage cleanup for the storage location specified by the memory-safe overwrite instruction, rather than simply returning a completion feedback for an unbound storage location.

[0128] Preferably, when receiving the underlying storage erasure credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node in response to the memory-secure overwrite instruction, the underlying storage erasure credentials are first parsed to obtain an underlying storage erasure credential parsing record. The underlying storage erasure credential includes a vehicle node identity digest, a task issuance session identifier, a local overwrite execution record, an overwrite execution timestamp, a release flag update status, and credential verification material. The underlying storage erasure credential parsing record is used to describe which vehicle node, which task issuance session identifier, and which local processing result storage location the underlying storage erasure credential corresponds to. The credential verification material is used to perform integrity verification on the underlying storage erasure credential. The underlying storage erasure credential parsing record continues to participate in subsequent underlying storage erasure credential validity verification, ensuring that the local storage cleanup status of the vehicle node can be obtained before the initial splicing result is confirmed as the global execution result.

[0129] Preferably, when verifying the validity of the underlying storage erasure credential, the vehicle node identity digest in the underlying storage erasure credential parsing record is first compared with the vehicle node identity digest in the task distribution consistency record to obtain a credential vehicle node matching result; then, the task distribution session identifier in the underlying storage erasure credential parsing record is compared with the task distribution session identifier in the local processing result reception summary record to obtain a credential task session matching result; subsequently, the overwrite processing status and release flag update status in the local overwrite execution record are jointly verified to obtain a local overwrite status verification result. The credential vehicle node matching result, the credential task session matching result, and the local overwrite status verification result together form an underlying storage erasure credential validity record; the underlying storage erasure credential validity record is used to determine whether the underlying storage erasure credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node correspond to vehicle nodes that have participated in task execution and returned local processing results, and is used to determine whether the first vehicle node, the second vehicle node, and the third vehicle node have completed the local storage cleanup corresponding to the local processing result.

[0130] Preferably, if the underlying storage erasure credential validity record indicates that the underlying storage erasure credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node are all valid, then the initial concatenation result is confirmed as the global execution result of the target processing task. If the concurrent processing substream is redirected and distributed to other vehicle nodes for continued execution, the redirection local processing result verification record and the underlying storage erasure credentials returned by the other vehicle nodes are also read, and the underlying storage erasure credentials returned by the other vehicle nodes are written into the underlying storage erasure credential validity record. Only after the underlying storage erasure credential validity record indicates that all vehicle nodes that have returned local processing results have completed local storage cleanup, is the initial concatenation result confirmed as the global execution result of the target processing task, so that the confirmation process of the global execution result is consistent with the receiving process of the local processing result and the execution process of the memory-safe overwrite instruction.

[0131] Preferably, when encapsulating the global execution result of the target processing task, a global execution result encapsulation record is generated based on the initial splicing result, the timestamp alignment record, the splicing integrity verification record, the underlying storage erasure credential validity record, and the task distribution consistency record. The global execution result encapsulation record includes the global execution result, output timestamp alignment information, local processing result source information, task distribution session identifier, underlying storage erasure credential validity record, and a result encapsulation verification field. The global execution result originates from the concurrent result fragment and the serial result fragment in the initial splicing result; the output timestamp alignment information originates from the timestamp alignment record; the local processing result source information originates from the local processing result reception summary record; the underlying storage erasure credential validity record indicates that the local local processing result of the vehicle node participating in the task execution has been stored and cleaned up; and the result encapsulation verification field records the verification status of the splicing integrity verification record and the underlying storage erasure credential validity record. The encapsulated record of the global execution result is used to output the global execution result, so that the global execution result includes not only the execution content of the target processing task, but also the data verification basis related to the time-series splicing operation and local storage cleanup.

[0132] Preferably, in a multi-vehicle centralized parking application scenario, the first vehicle node returns the local processing result corresponding to the concurrent processing sub-stream, the second vehicle node returns the local processing result corresponding to the pre-computation task slice, and the third vehicle node returns the local processing result corresponding to the remaining computation task slice. If the first vehicle node is redirected due to an abnormal parking maintenance state, the other vehicle nodes return the local processing result corresponding to the concurrent processing sub-stream after the redirection. The aforementioned local processing results are first entered into the timing splicing operation through the local processing result receiving summary record and the timestamp alignment record to form the initial splicing result. Subsequently, the memory-safe overwrite instruction is generated through the memory-safe overwrite plan, and the local storage cleanup status of each vehicle node is confirmed through the underlying storage erase credential validity record. Under the condition that the underlying storage erase credential validity record is valid, the initial splicing result is confirmed as the global execution result of the target processing task. This processing method differs from the method of immediately splicing plaintext output after the vehicle node returns the result. Instead, it establishes a sequential processing relationship between timestamp alignment, result integrity verification, local storage cleanup, and credential validity verification, so that the formation process of the global execution result can be connected with the task fragmentation execution process of heterogeneous vehicle formation.

[0133] like Figure 3As shown, this application discloses a computing power collaborative scheduling and task sharding device for heterogeneous vehicle formations. The computing power collaborative scheduling and task sharding device for heterogeneous vehicle formations includes: The parking trigger signal acquisition module is used to acquire parking trigger signals sent by multiple vehicle nodes with mobile energy storage and computing platforms; The specification extraction module is used to form a temporary computing federation group based on the parking trigger signal and extract the specification of the hardware acceleration unit of each vehicle node in the temporary computing federation group. The specification includes the concurrent throughput dimension information of the vehicle neural network acceleration unit. The hierarchical decoupling module is used to receive the target processing task, parse the operator execution dependency of the target processing task, perform hierarchical decoupling on the target processing task according to the operator execution dependency, and generate a first instruction processing set and a second instruction processing set. The concurrent processing sub-stream configuration module is used to extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream under the condition that the concurrent throughput dimension information meets the throughput requirement index. The computation task sharding module is used to extract the timing delivery constraints of the second instruction processing set, construct a cascaded execution topology sequence based on the timing delivery constraints, and shard the second instruction processing set into multiple serially associated computation task shards along the cascaded execution topology sequence. The task distribution module is used to distribute the concurrent processing subflow to the first vehicle node in the temporary computing federation group according to the specification, and to distribute the multiple serially associated computing task fragments to the second and third vehicle nodes in the temporary computing federation group in sequence. The global execution result generation module is used to receive the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, perform a time-series splicing operation on the local processing results, and generate the global execution result of the target processing task.

[0134] like Figure 4 As shown, an electronic device according to this application includes a processor and a memory. The memory stores a computer program, and when the processor runs the computer program, it performs the steps of the heterogeneous vehicle formation computing power collaborative scheduling and task sharding method as described in any one of this application.

[0135] like Figure 5As shown, this application discloses a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the heterogeneous vehicle formation computing power collaborative scheduling and task sharding method as described in any one of this application.

[0136] like Figure 6 As shown, this application discloses a computing power collaborative scheduling and task sharding system for heterogeneous vehicle formations. The system is characterized in that it includes multiple vehicle nodes with mobile energy storage and computing platforms, as well as a computing power collaborative scheduling and task sharding device for heterogeneous vehicle formations as described in this application. The vehicle node is used to send a parking trigger signal to the computing power collaborative scheduling and task sharding device of the heterogeneous vehicle formation, and to receive concurrent processing sub-streams or computing task shards, and return the local processing result after processing. The heterogeneous vehicle platoon's computational power collaborative scheduling and task sharding device is used to acquire the parking trigger signal sent by the vehicle nodes, establish a temporary computing federation group based on the parking trigger signal, extract the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, wherein the specifications include the concurrent throughput dimension information of the onboard neural network acceleration unit; receive the target processing task, parse the operator execution dependency relationship of the target processing task, perform hierarchical decoupling of the target processing task according to the operator execution dependency relationship, generate a first instruction processing set and a second instruction processing set; extract the throughput requirement index of the first instruction processing set, and, under the condition that the concurrent throughput dimension information meets the throughput requirement index, process the first instruction... The second instruction processing set is configured as a concurrent processing sub-stream; the timing delivery constraints of the second instruction processing set are extracted, and a cascaded execution topology sequence is constructed based on the timing delivery constraints. The second instruction processing set is then divided into multiple serially associated computation task fragments along the cascaded execution topology sequence. According to the specification, the concurrent processing sub-stream is sent to the first vehicle node in the temporary computing federation group, and the multiple serially associated computation task fragments are sequentially sent to the second and third vehicle nodes in the temporary computing federation group. The local processing results returned by the first, second, and third vehicle nodes are received, and a timing splicing operation is performed on the local processing results to generate the global execution result of the target processing task. Zero-knowledge execution proof data is obtained, and if the zero-knowledge execution proof data passes verification, the credential locking mechanism is unfrozen, and the task processing results returned by the target execution node are then aggregated and merged.

[0137] The above Figures 3-6 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for collaborative scheduling and task fragmentation of computing power in heterogeneous vehicle platooning, characterized in that, Includes the following steps: Acquire parking trigger signals sent by multiple vehicle nodes with mobile energy storage and computing platforms; Based on the parking trigger signal, a temporary computing federation group is formed, and the specifications of the hardware acceleration units of each vehicle node in the temporary computing federation group are extracted. The specifications include the concurrent throughput dimension information of the vehicle neural network acceleration unit. Receive the target processing task, parse the operator execution dependency of the target processing task, perform hierarchical decoupling of the target processing task according to the operator execution dependency, and generate a first instruction processing set and a second instruction processing set; Extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream if the concurrent throughput dimension information meets the throughput requirement index. Extract the timing delivery constraints of the second instruction processing set, construct a cascaded execution topology sequence based on the timing delivery constraints, and divide the second instruction processing set into multiple serially associated computing task fragments along the cascaded execution topology sequence; According to the specifications, the concurrent processing subflow is sent to the first vehicle node in the temporary computing federation group, and the multiple serially associated computing task fragments are sent to the second and third vehicle nodes in the temporary computing federation group in sequence. The system receives local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, performs a time-series splicing operation on the local processing results, and generates the global execution result of the target processing task.

2. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The steps for establishing a temporary computing federation group based on the parking trigger signal include: Extract the location coordinate data and network connectivity identifier carried by the parking trigger signal; Spatial clustering is performed on the location coordinate data to extract candidate vehicle nodes that are within the same physical range and whose network connectivity identifier indicates that they are in a connectable state. Obtain a preset security domain identity credential, and use the security domain identity credential to authenticate the candidate vehicle node. Obtain the local area network address of the candidate vehicle node that has passed the identity authentication, and form a network group based on the local area network address to generate the temporary computing federation group.

3. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The step of extracting the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, wherein the specifications include concurrent throughput dimension information of the onboard neural network acceleration unit, includes: Send probe query commands to each vehicle node within the temporary computing federation group; Receive the underlying hardware status logs for the hardware acceleration unit returned by each vehicle node in response to the probe query command. Extract the number of tensor operation cores and available video memory capacity of the vehicle-mounted neural network acceleration unit from the underlying hardware status log; The number of tensor operation cores and the available video memory capacity are formatted and combined to form the concurrent throughput dimension information, and the formatted combination data containing the concurrent throughput dimension information is encapsulated into the specification of the hardware acceleration unit of each vehicle node.

4. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The steps of parsing the operator execution dependencies of the target processing task, performing hierarchical decoupling of the target processing task based on the operator execution dependencies, and generating a first instruction processing set and a second instruction processing set include: The target processing task is parsed using a directed acyclic graph to obtain the operator execution dependencies consisting of nodes and edges; The hierarchical decoupling is performed based on the edge connection state in the operator execution dependency relationship, and independent operator nodes without data dependency edges and coupled operator nodes with data dependency edges are extracted. The independent operator nodes are aggregated to generate the first instruction processing set; Extract the data flow direction indicated by the data dependency edge between the coupled operator nodes, and combine the coupled operator nodes with the same data flow direction in a time sequence to generate the second instruction processing set.

5. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The steps of extracting the timing transit constraints of the second instruction processing set, constructing a cascaded execution topology sequence based on the timing transit constraints, and dividing the second instruction processing set into multiple serially associated computational task fragments along the cascaded execution topology sequence include: Extract the input tensor size of the first instruction data and the output tensor size of the last instruction data from the second instruction processing set, and use the input tensor size and the output tensor size as the timing transmission constraint. Based on the input tensor size and the output tensor size in the timing transit constraints, calculate the activation value overhead parameter when the second instruction processing set is internally transited; Based on the activation value overhead parameter, multiple data communication nodes are inserted into the second instruction processing set, and the second instruction processing set is connected in series using the data communication nodes to construct and generate the cascaded execution topology sequence. Using the data communication node as the dividing point, the second instruction processing set is divided into multiple serially associated computing task fragments along the cascaded execution topology sequence.

6. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The steps of distributing the concurrent processing substream to the first vehicle node in the temporary computing federation group according to the specification, and distributing the multiple serially associated computing task fragments sequentially to the second and third vehicle nodes in the temporary computing federation group, include: Under the condition that the concurrent throughput dimension information in the specification meets the first preset value, candidate nodes with matching computing power are selected from the temporary computing federation group as the first vehicle node, and the concurrent processing subflow is sent to the first vehicle node. Extract the network bandwidth transmission index contained in the specification. Under the condition that the network bandwidth transmission index meets the second preset value, select candidate nodes that meet the transmission requirements from the temporary computing federation group as the second vehicle node and the third vehicle node. According to the serial association order of the cascaded execution topology sequence, the preceding computing task fragments in the multiple serially associated computing task fragments are sent to the second vehicle node, and the remaining computing task fragments are sent to the third vehicle node.

7. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The step of receiving the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node includes: Real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node; If the parking maintenance state is a normal parking state, then the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node are received. If a parking release signal is detected from the first vehicle node, it is determined that the parking maintenance state of the first vehicle node is abnormal, and the data receiving channel of the first vehicle node is blocked. Extract the concurrent processing sub-streams that have not yet been completed in the first vehicle node, redirect and distribute the concurrent processing sub-streams to other vehicle nodes in the temporary computing federation group that are in the normal parking state for continued execution, and receive the local processing results returned by the other vehicle nodes.

8. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 7, characterized in that, The step of real-time monitoring of the parking maintenance status of the first vehicle node, the second vehicle node, and the third vehicle node includes: Obtain the battery management system interface protocol, and periodically extract the remaining power battery capacity of the first vehicle node, the second vehicle node, and the third vehicle node through the battery management system interface protocol; The remaining power of the power battery is compared and analyzed with the preset power protection lower limit value; If the remaining power of the power battery is not lower than the power protection lower limit value, then the corresponding vehicle node is determined to meet the computing power maintenance condition of the parking maintenance state. If the remaining power of the power battery is lower than the power protection lower limit, it is determined that the corresponding vehicle node does not meet the computing power maintenance condition, a computing power cancellation command is generated and issued, and the corresponding vehicle node is forcibly triggered to exit the parking maintenance state.

9. The computational power collaborative scheduling and task fragmentation method for heterogeneous vehicle formation as described in claim 1, characterized in that, The step of performing a time-series concatenation operation on the local processing results to generate the global execution result of the target processing task includes: The output timestamps of the local processing results are aligned, and the time-series splicing operation is performed on the local processing results based on the aligned output timestamps to obtain the initial splicing result. A memory-safe overwrite instruction is issued to the first vehicle node, the second vehicle node, and the third vehicle node to erase the local processing results on their respective localities. Receive the underlying storage erase credentials returned by the first vehicle node, the second vehicle node, and the third vehicle node in response to the memory-secure overwrite instruction; Under the condition that the underlying storage erasure certificate is valid, the initial splicing result is confirmed as the global execution result of the target processing task and output encapsulated.

10. A computational power collaborative scheduling and task sharding device for heterogeneous vehicle platooning, characterized in that, The computational power collaborative scheduling and task sharding device for heterogeneous vehicle formation includes: The parking trigger signal acquisition module is used to acquire parking trigger signals sent by multiple vehicle nodes with mobile energy storage and computing platforms; The specification extraction module is used to form a temporary computing federation group based on the parking trigger signal and extract the specification of the hardware acceleration unit of each vehicle node in the temporary computing federation group. The specification includes the concurrent throughput dimension information of the vehicle neural network acceleration unit. The hierarchical decoupling module is used to receive the target processing task, parse the operator execution dependency of the target processing task, perform hierarchical decoupling on the target processing task according to the operator execution dependency, and generate a first instruction processing set and a second instruction processing set. The concurrent processing sub-stream configuration module is used to extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream under the condition that the concurrent throughput dimension information meets the throughput requirement index. The computation task sharding module is used to extract the timing delivery constraints of the second instruction processing set, construct a cascaded execution topology sequence based on the timing delivery constraints, and shard the second instruction processing set into multiple serially associated computation task shards along the cascaded execution topology sequence. The task distribution module is used to distribute the concurrent processing subflow to the first vehicle node in the temporary computing federation group according to the specifications, and to distribute the multiple serially associated computing task fragments to the second and third vehicle nodes in the temporary computing federation group in sequence. The global execution result generation module is used to receive the local processing results returned by the first vehicle node, the second vehicle node, and the third vehicle node, perform a time-series splicing operation on the local processing results, and generate the global execution result of the target processing task.

11. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program, and when the processor runs the computer program, it performs the steps of the computational power collaborative scheduling and task sharding method for heterogeneous vehicle formation as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the heterogeneous vehicle formation computing power collaborative scheduling and task sharding method as described in any one of claims 1 to 9.

13. A computational power collaborative scheduling and task sharding system for heterogeneous vehicle platooning, characterized in that, The heterogeneous vehicle formation computing power collaborative scheduling and task sharding system includes multiple vehicle nodes with mobile energy storage and computing platforms, as well as the heterogeneous vehicle formation computing power collaborative scheduling and task sharding device as described in claim 10. The vehicle node is used to send a parking trigger signal to the computing power collaborative scheduling and task sharding device of the heterogeneous vehicle formation, and to receive concurrent processing sub-streams or computing task shards, and return the local processing result after processing. The heterogeneous vehicle platoon's computational power collaborative scheduling and task sharding device is used to acquire the parking trigger signal sent by the vehicle nodes, establish a temporary computing federation group based on the parking trigger signal, extract the specifications of the hardware acceleration units of each vehicle node within the temporary computing federation group, wherein the specifications include the concurrent throughput dimension information of the on-board neural network acceleration unit; receive a target processing task, parse the operator execution dependency relationship of the target processing task, perform hierarchical decoupling of the target processing task according to the operator execution dependency relationship, and generate a first instruction processing set and a second instruction processing set; extract the throughput requirement index of the first instruction processing set, and configure the first instruction processing set as a concurrent processing sub-stream under the condition that the concurrent throughput dimension information meets the throughput requirement index; extract the timing delivery constraint condition of the second instruction processing set, construct a cascaded execution topology sequence according to the timing delivery constraint condition, and divide the second instruction processing set into multiple serially associated computational task shards along the cascaded execution topology sequence; According to the specifications, the concurrent processing subflow is distributed to the first vehicle node in the temporary computing federation group, and the multiple serially associated computing task fragments are sequentially distributed to the second and third vehicle nodes in the temporary computing federation group. The local processing results returned by the first, second, and third vehicle nodes are received, and a time-series splicing operation is performed on the local processing results to generate the global execution result of the target processing task. Zero-knowledge execution proof data is obtained, and if the zero-knowledge execution proof data passes verification, the credential locking mechanism is unfrozen, and then the task processing results returned by the target execution node are summarized and merged.

Images