Adaptive inference model updating method for multi-core-board scene change and rack-mounted AI edge inference terminal

Abstract

The embodiment of the invention provides an inference model adaptive updating method for multi-core board scene change and a rack-mounted AI edge inference terminal, and relates to the technical field of edge calculation and artificial intelligence model optimizing.The inference model adaptive updating method comprises the steps that firstly, an inference model adjustment factor library is constructed, and a corresponding set of scene feature types and model adjustment factors is included; the scene features cover core cooperation states, task load distribution and data transmission path features, the model adjustment factors cover input node association, intermediate layer parameter interaction and output dimension adaptation factors, the real-time scene features of the multi-core board are extracted, a target model adjustment factor is matched and determined, a structure parameter adjustment scheme is generated, and the real-time scene features of the multi-core board are matched and determined. Structural reconstruction and parameter updating are carried out on the reasoning model, an updated reasoning model matched with the current scene is obtained, deployment and operation are carried out, meanwhile, new scene features are collected to update and adjust the corresponding relation of the factor library, the self-adaptive updating capacity of the reasoning model is maintained, and the adaptability and performance of the multi-core-board system in different scenes are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge computing and artificial intelligence model optimization, in particular to an inference model adaptive updating method for multi-core board scene changes and a rack-mounted AI edge inference terminal. BACKGROUND

[0002] With the wide application of artificial intelligence technology in edge computing scenarios, inference models deployed on edge devices face severe resource constraint challenges. Existing model compression techniques are mostly based on static resource assumptions, using fixed compression ratio pruning or quantization methods, and do not fully consider the dynamic changes in resource status of edge devices in actual operation. In particular, in high-performance edge devices such as rack-mounted AI inference terminals, due to the simultaneous running of multiple AI modules and the processing of complex inference tasks, the computing resources, storage resources and inference delay requirements will change in real time with factors such as workload, temperature changes and power fluctuations.

[0003] There are three significant defects in current model compression methods: first, the traditional compression strategy is disconnected from the actual resource status of the device, and the compressed model may not adapt to resource fluctuations, leading to inference interruption or performance degradation; second, existing methods ignore the dependency relationship of inter-layer feature transmission within the model, and rough pruning often destroys the integrity of feature transmission, causing serious accuracy loss; finally, static compression schemes lack dynamic adjustment mechanisms and cannot optimize online according to resource changes during inference. These problems are particularly prominent in military, industrial and other applications that require high reliability.

[0004] Although existing technologies have proposed some adaptive compression ideas, most of them only focus on a single compression dimension, lack of collaborative optimization of neuron pruning, channel pruning and connection sparsification and other multi-dimensional compression methods, and have not established a dynamic association mechanism between compression operations and inter-layer feature transmission. In addition, existing methods have a lag in responding to resource fluctuations on edge devices, making it difficult to maintain the stability of inference performance when resource constraints change suddenly. SUMMARY

[0005] Therefore, the purpose of the present application is to provide an inference model adaptive updating method for multi-core board scene changes and a rack-mounted AI edge inference terminal.

[0006] According to a first aspect of the present application, an inference model adaptive updating method for multi-core board scene changes is provided, the method comprising: Construct a library of inference model adjustment factors. The library contains a set of corresponding scene feature types and model adjustment factors. The scene feature types cover core collaboration state features, task load distribution features and data transmission path features. The model adjustment factors cover input node association factors, intermediate layer parameter interaction factors and output dimension adaptation factors. Extract real-time scene features during the operation of the multi-core board, match the real-time scene features with the scene feature types in the inference model adjustment factor library, and determine the target model adjustment factor corresponding to the current real-time scene feature; Based on the degree of change of the current real-time scene features and the target model adjustment factor, a structural parameter adjustment scheme for the inference model is generated. The structural parameter adjustment scheme includes input node association adjustment rules, intermediate layer parameter interaction adjustment rules, and output dimension adaptation adjustment rules. Based on the structural parameter adjustment scheme, the inference model is restructured and its parameters are updated to obtain an updated inference model that is adapted to the current multi-core board scenario. The updated inference model is then deployed to the multi-core board for operation. Collect new scene features of the multi-core board during the operation of the updated inference model, and use the new scene features to update the correspondence between scene feature types and model adjustment factors in the inference model adjustment factor library to maintain the adaptive update capability of the inference model.

[0007] According to a second aspect of this application, a rack-mounted AI edge inference terminal is provided, the rack-mounted AI edge inference terminal including a processor and a readable storage medium, the readable storage medium storing a program that, when executed by the processor, implements the aforementioned adaptive update method for inference models in response to changes in multi-core board scenarios.

[0008] Based on any of the above aspects, an inference model adjustment factor library is constructed, containing a set of corresponding scene feature types and model adjustment factors. This library covers key scene features such as core collaborative state features, task load distribution features, and data transmission path features, as well as model adjustment factors such as input node association factors, intermediate layer parameter interaction factors, and output dimension adaptation factors. During multi-core board operation, scene features can be extracted in real time and matched with the adjustment factor library to accurately determine the target model adjustment factors. This generates a structural parameter adjustment scheme that includes input node association adjustment rules, intermediate layer parameter interaction adjustment rules, and output dimension adaptation adjustment rules. Based on this structural parameter adjustment scheme, the inference model is restructured and its parameters are updated. The resulting updated inference model is highly adaptable to the current multi-core board scenario, effectively improving the model's adaptability and inference accuracy in different scenarios. Simultaneously, new scene features are collected during the operation of the updated inference model, and the corresponding relationships in the adjustment factor library are updated, maintaining the inference model's adaptive update capability. This allows the model to continuously adapt to the dynamic changes in the multi-core board scenario, ensuring stable operation and high performance of the system in different scenarios. In other words, by establishing a dual correlation between compression dimensions and inter-layer feature transfer, dynamic protection of feature transfer integrity is achieved during model compression, effectively avoiding the accuracy loss caused by traditional compression methods. By collecting edge device resource data in real time and constructing dynamic resource constraint information, the compression strategy can accurately adapt to the actual operating state of the devices, significantly improving the adaptability of the compressed model in resource fluctuation environments, and ensuring that the final compressed model maintains optimal inference performance while meeting strict resource constraints. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the adaptive update method for inference models in multi-core board scenarios provided in this application embodiment is shown. Figure 2 A schematic diagram of the component structure of the rack-mounted AI edge inference terminal provided in an embodiment of this application is shown. Detailed Implementation

[0010] Figure 1 The diagram illustrates a flowchart of an adaptive update method for inference models in multi-core board scenarios provided in this application. It should be understood that in other embodiments, the order of some steps in the adaptive update method for inference models in multi-core board scenarios can be interchanged as needed, or some steps can be omitted or deleted. The detailed steps of this adaptive update method for inference models in multi-core board scenarios are described below.

[0011] Step S110: Construct an inference model adjustment factor library. The inference model adjustment factor library contains a set of corresponding scene feature types and model adjustment factors. The scene feature types cover core collaborative state features, task load distribution features and data transmission path features. The model adjustment factors cover input node association factors, intermediate layer parameter interaction factors and output dimension adaptation factors.

[0012] In this embodiment, constructing the inference model adjustment factor library requires systematically integrating the mapping relationship between multi-core board operating scenarios and inference model adjustment parameters. The core function of this inference model adjustment factor library is to accurately match the corresponding model adjustment factor value range through quantitative description of scenario feature types.

[0013] Step S111: Collect historical operating scenario data of multi-core boards. The historical operating scenario data of multi-core boards includes core collaboration data, task allocation data and data transmission data of different operating stages. Each operating stage data is labeled with a corresponding scenario feature description.

[0014] During the data collection phase, it is necessary to cover the operational data of the multi-core board throughout its entire lifecycle, from startup initialization and normal operation to peak load and anomaly recovery. Core collaboration data should include fine-grained information such as inter-core interrupt request signals, shared memory access records, and bus arbitration results; task allocation data should cover task creation timestamps, core binding status, priority change logs, and resource preemption records; data transmission data should include transmission protocol type, packet verification results, cache hit status, and error retry calculations. Data for each operational phase should be labeled with a scenario characteristic description, which should include technical attributes such as the core cluster working mode (e.g., symmetric multiprocessing, asymmetric multiprocessing), task scheduling strategy (e.g., preemptive scheduling, cooperative scheduling), and data transmission topology (e.g., star connection, ring connection).

[0015] Step S112: Extract features from the core collaboration data. Based on the synchronization frequency between cores, the number of cores initiating collaboration, and the response time of collaboration signals, divide different core collaboration state features. Each core collaboration state feature corresponds to a specific range of core collaboration data.

[0016] The core synchronization frequency is calculated by comprehensively considering the number of times global synchronization signals (such as barriers signals) are triggered per unit time, the acquisition and release cycle of mutexes, and the waiting and wake-up frequency of condition variables. The number of cores initiating collaboration needs to distinguish between cores that actively initiate collaboration requests and cores that passively respond to collaboration requests. This is determined by parsing the source address identifier in the bus transaction and then deduplicating the count. The collaboration signal response time needs to record the system clock values ​​at the time the signal is initiated and at the time the response is completed, calculating the difference between the two, while excluding abnormal delay samples caused by bus congestion. Based on these three characteristics, a clustering algorithm is used to divide the core collaboration data into feature categories. For example, when the core synchronization frequency is in the first preset range, the proportion of collaboration-initiating cores to the total number of cores is in the second preset range, and the collaboration signal response time is in the third preset range, it corresponds to the first type of core collaboration state feature; when each feature falls into other preset range combinations, it corresponds to other core collaboration state features. The specific numerical range boundaries of the three feature parameters must be clearly defined for each feature category.

[0017] Step S113: Extract features from the task allocation data. Based on the task's dwell time in each core, task transfer frequency, and core load ratio, divide different task load distribution features. Each task load distribution feature corresponds to a specific task allocation data range.

[0018] The duration of tasks on each core needs to be statistically analyzed separately according to task type (e.g., compute-intensive, I / O-intensive). The total time from entering the core's ready queue to completion for each task should be calculated, deducting the blocking time spent waiting for I / O. Task transfer frequency is statistically analyzed by counting the number of times tasks migrate between different cores per unit time, combined with the reasons for task migration (e.g., load balancing, core failure). Core load percentage is determined by calculating the weighted average of indicators such as CPU utilization, memory bandwidth utilization, and cache hit rate for each core per unit time. Based on the above characteristics, factors such as task type distribution and core performance heterogeneity need to be considered when classifying task load distribution characteristics. For example, combinations of parameters such as the percentage of time compute-intensive tasks spend on high-performance cores, the transfer frequency of tasks between cores of the same performance level, and the standard deviation of load percentage for each core constitute different task load distribution characteristics, each corresponding to a specific numerical range of the above parameters.

[0019] Step S114: Extract features from the data transmission data. Based on the number of intermediate cores through which the data transmission passes, the frequency of transmission link switching, and the transmission duration of each link segment, different data transmission path features are defined. Each data transmission path feature corresponds to a specific range of data transmission data.

[0020] The number of intermediate cores traversed by data transmission requires parsing the routing records of data packets and counting the number of forwarding cores traversed along the transmission path from the source core to the destination core. If multi-path transmission exists, the average value is taken. The transmission link switching frequency is measured by monitoring the connection status change logs of physical links (such as PCIe channels and Ethernet ports) and counting the number of link switches per unit time, including active switches (such as those triggered by link aggregation policies) and passive switches (such as those recovering from link failures). The transmission duration of each link segment requires measuring the transmission time of data packets between every two adjacent intermediate cores (or between the source core and the first intermediate core, or between the last intermediate core and the destination core), including sending delay, propagation delay, processing delay, and queuing delay. When defining data transmission path characteristics, factors such as the transmission protocol (such as TCP, UDP) and data packet size distribution are considered, and different combinations of the above three characteristics are defined as different data transmission path characteristics.

[0021] Step S115: Determine the input node association factor, intermediate layer parameter interaction factor, and output dimension adaptation factor of the inference model as model adjustment factors. The input node association factor includes the node connection quantity coefficient, connection priority coefficient, and connection activation condition coefficient. The intermediate layer parameter interaction factor includes the parameter coupling strength coefficient, parameter transmission order coefficient, and parameter feedback coefficient. The output dimension adaptation factor includes the output node quantity coefficient, output data format coefficient, and output correspondence coefficient with core requirements.

[0022] In the input node association factors, the node connection quantity coefficient controls the connection density between the input layer and the intermediate layer of the inference model, and its value is positively correlated with the number of collaborative initiating cores in the core collaborative state features; the connection priority coefficient determines the signal weights of different input nodes and is positively correlated with the quantified value of the core's computing power; the connection activation condition coefficient sets the signal threshold conditions required for establishing connections between input nodes and intermediate layer nodes and is negatively correlated with the collaborative signal response time. In the intermediate layer parameter interaction factors, the parameter coupling strength coefficient controls the connection weights between intermediate layer neurons and is positively correlated with the task dwell time; the parameter transmission order coefficient determines the forward propagation path of intermediate layer parameters and is adjusted in conjunction with the task transfer frequency; the parameter feedback coefficient adjusts the gradient update amplitude during backpropagation and is negatively correlated with the core load ratio. In the output dimension adaptation factors, the output node quantity coefficient is positively correlated with the number of intermediate cores in the data transmission path; the output data format coefficient dynamically adjusts the data encoding method according to the transmission link switching frequency; and the output-core requirement correspondence coefficient matches the interface protocols between output nodes and cores and is negatively correlated with the link transmission time.

[0023] Step S116: Match the corresponding input node association factor for each core collaboration state feature, and determine the specific numerical range of the input node association factor under different core collaboration state features; match the corresponding intermediate layer parameter interaction factor for each task load distribution feature, and determine the specific numerical range of the intermediate layer parameter interaction factor under different task load distribution features; match the corresponding output dimension adaptation factor for each data transmission path feature, and determine the specific numerical range of the output dimension adaptation factor under different data transmission path features.

[0024] For matching core collaboration state characteristics with input node association factors, a mapping model needs to be established through training on historical data. For example, when core collaboration state characteristics are characterized by high synchronization frequency, multiple initiating cores, and short response time, the node connection quantity coefficient takes a high value range, the connection priority coefficient is assigned according to core performance ranking, and the connection activation condition coefficient takes a low value range. Matching task load distribution characteristics with intermediate layer parameter interaction factors needs to be determined through simulation experiments. For example, balanced load distribution characteristics correspond to medium parameter coupling strength coefficients, sequential parameter transmission order coefficients, and small-amplitude parameter feedback coefficients. Matching data transmission path characteristics with output dimension adaptation factors needs to consider hardware interface characteristics. For example, long link transmission path characteristics correspond to multiple output node quantity coefficients, fault-tolerant coding output data format coefficients, and high-priority output corresponding to core requirements coefficients. Each matching relationship needs to be fitted with a large amount of sample data to determine the upper and lower limits and boundary conditions of the factor numerical range.

[0025] Step S117: Integrate the scene feature types and the numerical range and matching logic of the corresponding model adjustment factors into a data set, store it in a dedicated database to form an inference model adjustment factor library, and set an update interface for the inference model adjustment factor library for subsequent updates of content when receiving new scene feature data.

[0026] The integration of the dataset must adopt a structured data format, with each scene feature type serving as the primary key, and including a matrix of numerical ranges for input node association factors, intermediate layer parameter interaction factors, output dimension adaptation factors, and matching logic description text. The matching logic must include details such as feature parameter weight allocation, threshold judgment rules, and factor value interpolation algorithms. The dedicated database must support high-concurrency read / write and transaction processing, employing index optimization techniques to improve the efficiency of scene feature matching queries. Update interfaces must define data transmission protocols, data verification rules, and version control mechanisms to ensure the integrity and compatibility of new scene feature data. Interface calls must be logged, including call time, data source, and update content, for traceability and auditing purposes.

[0027] Step S120: Extract real-time scene features during the operation of the multi-core board, match the real-time scene features with the scene feature types in the inference model adjustment factor library, and determine the target model adjustment factor corresponding to the current real-time scene features.

[0028] Step S121: During the operation of the multi-core board, the operation data of the core processing unit is acquired in real time through the data acquisition interface. The operation data includes the synchronization data signal between cores, the real-time record of task scheduling, and the real-time log of data transmission.

[0029] The data acquisition interface needs to be integrated into the system monitoring module of the multi-core board, employing a Hardware Abstraction Layer (HAL) design to achieve cross-platform compatibility. Inter-core synchronization data signals are acquired by monitoring transaction signals on the system bus (such as the AMBA bus or AXI bus), including fields such as the synchronization instruction's opcode, source core ID, destination core ID, and operation result. Real-time task scheduling records are output by the operating system kernel's scheduler module, containing information such as task ID, currently running core ID, state transition timestamp, and remaining execution time. Real-time data transmission logs are generated by the Network Interface Controller (NIC) or the monitoring unit of the Network on-Chip (NoC), recording detailed information such as the source address, destination address, length, checksum, transmission path, and timestamp of data packets. The data acquisition frequency needs to be dynamically adjusted according to the multi-core board's operating clock frequency to ensure that real-time requirements are met without affecting system performance; it is typically set to the range of one-thousandth to one-hundredth of the core's highest clock frequency.

[0030] Step S122: Process the core synchronization data signal, count the number of core synchronization interactions per unit time, the number of core identifiers that initiate synchronization interactions, and the signal response time of each synchronization interaction, and integrate the statistical results of the number of core synchronization interactions per unit time, the number of core identifiers that initiate synchronization interactions, and the signal response time of each synchronization interaction to form core collaboration status characteristics.

[0031] The processing of inter-core synchronization data signals involves three stages: signal filtering, protocol parsing, and feature extraction. The signal filtering stage uses a sliding window algorithm to remove high-frequency noise and transient interference signals, retaining valid signals whose duration exceeds the minimum synchronization period. The protocol parsing stage analyzes the synchronization interaction type (e.g., barrier, lock, unlock) and participating core sets based on the multi-core board's synchronization protocol specifications (e.g., OpenMP synchronization commands, MPI communication protocol). The feature extraction stage performs statistical analysis on the valid synchronization signals.

[0032] Step S1221: Set a fixed time statistics interval, and divide the core synchronization data signal into multiple continuous data signal segments according to the time statistics interval, with each data signal segment corresponding to a time statistics interval.

[0033] The length of the time statistics interval needs to comprehensively consider the task cycle and sensitivity to scene changes of the multi-core board, and is usually set to 10 milliseconds to 100 milliseconds. The interval division adopts a timer interrupt triggering mechanism, and the start time of each time statistics interval is synchronized with the system clock to ensure time alignment of data signal segments. For synchronization signals that cross intervals, they need to be assigned to the corresponding time statistics interval according to the signal initiation time to avoid data fragmentation. Each data signal segment needs to be appended with a timestamp identifier, accurate to the microsecond level, for subsequent time series analysis.

[0034] Step S1222: In each data signal segment, traverse all synchronization interaction records, count the total number of synchronization interactions between cores, and use the total number as the number of synchronization interactions between cores per unit time.

[0035] The traversal process employs multi-threaded parallel processing. Each core is assigned an independent processing thread responsible for parsing local synchronization interaction records, and then the total global count is aggregated through atomic operations. Synchronization interaction records need to distinguish between different types of synchronization operations (such as spinlocks, semaphores, and message passing). When counting, all types or specific types of synchronization interactions can be selected based on business needs. For nested synchronizations (such as a synchronization operation containing another synchronization operation), the outermost synchronization operation should be counted only once to avoid duplicate counting. The counting result needs to be stored as a 32-bit unsigned integer; if it exceeds the maximum value, it will automatically overflow and reset to zero, and the overflow count will be recorded.

[0036] Step S1223: In each data signal segment, extract the initiating core identifier from each synchronization interaction record, remove duplicate core identifiers, count the number of remaining core identifiers, and use this number as the number of core identifiers for initiating synchronization interaction.

[0037] The initiating core identifier is extracted from the source core ID field of the synchronization interaction record. A hash table data structure is used to store the core identifiers that have appeared, achieving a deduplication operation with an O(1) time complexity. The key of the hash table is the core identifier (usually an integer from 0 to N-1, where N is the total number of cores), and the value is a boolean indicating whether it has been recorded. After traversing all synchronization interaction records in the data signal segment, the number of keys in the hash table is the number of core identifiers that initiated the synchronization interaction. For broadcast synchronization signals (such as global barriers), the initiating core identifier is the ID of the broadcast controller, which needs to be processed separately to avoid duplicate counting.

[0038] Step S1224: In each data signal segment, for each synchronization interaction record, read the synchronization signal initiation time and the synchronization response signal reception time, calculate the time difference between the two times, and obtain the signal response duration of each synchronization interaction.

[0039] The synchronization signal initiation time is obtained from the timestamp field of the synchronization command transmission, and the synchronization response signal reception time is obtained from the timestamp field of the synchronization completion interruption. Both times are timed using the system monotonic clock to avoid time rollback issues caused by system time adjustments. The time difference is calculated using 64-bit integer subtraction, with the unit being nanoseconds. For synchronization interaction records that do not respond within a timeout period, the response duration is recorded as a preset maximum threshold (e.g., 1.5 times the length of the time statistics interval) and marked as an abnormal synchronization event. The signal response duration of each synchronization interaction needs to be associated and stored with the corresponding synchronization type and core identifier to form multi-dimensional statistical data.

[0040] Step S1225: Record the number of core synchronization interactions per unit time, the number of core identifiers that initiate synchronization interactions, and the signal response duration of each synchronization interaction for each time statistical interval. Add a corresponding time statistical interval identifier to each record to form core collaboration status characteristics.

[0041] The associated records use a structured array data format, with each array element corresponding to a time statistics interval. This interval includes four fields: interval identifier, number of synchronization interactions, number of initiating cores, and a response duration array (containing the response duration of each synchronization interaction). The time statistics interval identifier is encoded in a "date-time-sequence number" format to ensure global uniqueness. The response duration array is stored in chronological order of synchronization interactions to facilitate subsequent calculations of response duration distribution characteristics (such as mean, median, and variance). Core collaboration status characteristic data undergoes verification and calculation to ensure data integrity during transmission and storage.

[0042] Step S123: Process the real-time task scheduling records, extract the dwell time of tasks in each core processing unit, the number of times tasks are transferred between different core processing units, and the load percentage of each core processing unit. Integrate the extracted results of the dwell time of tasks in each core processing unit, the number of times tasks are transferred between different core processing units, and the load percentage of each core processing unit to form task load distribution characteristics.

[0043] The processing of real-time task scheduling records needs to be combined with the operating system's process management mechanism, and feature extraction is achieved by parsing the state change records of the Task Control Block (TCB). The task dwell time is calculated by the difference between the timestamps of the task entering the core ready queue and leaving the core execution queue; the number of task transfers is counted by tracking the switching count between different core IDs of the task ID; the core load ratio is determined by statistically analyzing the percentage of time the core is in each unit of time for its running state (running, ready, blocked).

[0044] Step S124: Process the real-time log of the data transmission, identify the number of intermediate core processing unit identifiers, the number of transmission link switching times, and the transmission time length of each transmission link during the data transmission process, and integrate the identification results of the number of intermediate core processing unit identifiers, the number of transmission link switching times, and the transmission time length of each transmission link during the data transmission process to form data transmission path features.

[0045] The processing of real-time data transmission logs requires parsing the header information of each layer of the network protocol stack. The physical layer log is used to extract the number of link switches, the network layer log is used to identify intermediate core identifiers in the transmission path, and the transport layer log is used to calculate the link transmission time. The intermediate core identifier is determined through routing table lookups and packet fragmentation and reassembly. The number of link switches is counted by physical port status changes, and the link transmission time is calculated by the difference between the sending and receiving timestamps of the data packets.

[0046] Step S125: Combine the core collaboration status features, the task load distribution features, and the data transmission path features into real-time scene features.

[0047] Real-time scene features are represented using a three-dimensional vector structure, with each dimension corresponding to a scene feature type. The core collaboration status feature dimension includes the number of synchronous interactions, the number of initiating cores, and the response time distribution feature vector; the task load distribution feature dimension includes the dwell time length matrix, the task transfer count matrix, and the core load percentage vector; the data transmission path feature dimension includes the number of intermediate cores, the number of link switching times, and the link transmission time length distribution vector. Each element of the three-dimensional vector needs to be normalized to convert feature parameters of different dimensions to the [0, 1] interval for subsequent matching calculations.

[0048] Step S126: Access the inference model adjustment factor library, compare the real-time scene features with the scene feature types in the inference model adjustment factor library, and find the input node association factor corresponding to the core collaborative state feature type with the highest matching degree; compare the task load distribution features in the real-time scene features with the task load distribution feature types in the library, and find the intermediate layer parameter interaction factor corresponding to the task load distribution feature type with the highest matching degree; compare the data transmission path features in the real-time scene features with the data transmission path feature types in the library, and find the output dimension adaptation factor corresponding to the data transmission path feature type with the highest matching degree.

[0049] Scene feature matching employs a weighted Euclidean distance algorithm. Weight coefficients (obtained through training based on historical matching accuracy) are assigned to each element of the real-time scene feature vector and the scene feature type vector in the database. After calculating the weighted distance, the scene feature type corresponding to the minimum value is taken as the matching result. The extraction of input node association factors, intermediate layer parameter interaction factors, and output dimension adaptation factors requires reading the corresponding numerical range matrix from the adjustment factor database based on the matched scene feature type, and then performing interpolation calculations within this range using the real-time scene feature parameters to obtain the specific factor values.

[0050] Step S127: Integrate the found input node association factors, intermediate layer parameter interaction factors, and output dimension adaptation factors into the target model adjustment factors corresponding to the current real-time scene features.

[0051] The target model adjustment factors are stored using a dictionary data structure, with the key being the factor name (such as "node connection quantity coefficient", "parameter coupling strength coefficient", etc.) and the value being the specific numerical value obtained through interpolation. During the integration process, the reasonableness of each factor value needs to be verified to ensure its physical validity (e.g., the connection quantity coefficient is a positive integer, and the coupling strength coefficient is in the range [0, 1]). If the verification fails, the default safe value for that factor is used, and an exception log is recorded.

[0052] Step S130: Based on the degree of change of the current real-time scene features and the target model adjustment factor, generate a structural parameter adjustment scheme for the inference model. The structural parameter adjustment scheme includes input node association adjustment rules, intermediate layer parameter interaction adjustment rules, and output dimension adaptation adjustment rules.

[0053] This step involves quantitatively analyzing the magnitude of changes in real-time scene features, combining this with the dynamic values ​​of the target model adjustment factors, and formulating specific adjustment strategies for each level of the inference model to ensure the relevance and effectiveness of the adjustment plan.

[0054] Step S131: Calculate the difference between the current real-time scene features and the previously extracted real-time scene features. The difference includes differences in core collaboration status features, task load distribution features, and data transmission path features. The differences in core collaboration status features are represented by the change in the number of synchronization interactions per unit time, the change in the number of core identifiers that initiate synchronization, and the change in signal response time. The differences in task load distribution features are represented by the change in task dwell time, the change in task transfer frequency, and the change in core load percentage. The differences in data transmission path features are represented by the change in the number of intermediate cores, the change in link switching frequency, and the change in link transmission time.

[0055] The difference calculation employs an incremental comparison method, calculating the difference (relative rate of change) between the current value and the previous value for each element of the real-time scene feature vector. In the core collaboration state feature differences, the change in the number of synchronous interactions is (current number - previous number) / previous number; the change in the number of initiating core identifiers is the absolute difference between the current number and the previous number; and the change in signal response duration is the difference between the current average response duration and the previous average response duration. The calculation methods for task load distribution feature differences and data transmission path feature differences are similar, both requiring a distinction between absolute changes and relative rates of change, and selecting an appropriate difference representation method based on the properties of the feature parameters.

[0056] Step S132: Adjust the input node association factor values ​​in the target model adjustment factors according to the differences in core collaboration status characteristics. If the number of synchronization interactions per unit time increases, increase the node connection quantity coefficient; if the number of core identifiers initiating synchronization increases, increase the connection priority coefficient corresponding to the core whose computing power meets the high-load task processing requirements of multi-core boards; if the signal response time increases, tighten the connection activation condition coefficient; based on the adjusted input node association factor values, generate input node association adjustment rules to determine the increase / decrease ratio of the number of input node connections, the adjustment level of connection priority, and the specific requirements of connection activation conditions.

[0057] The adjustment of input node association factors adopts a feedback control mechanism, using the differences in core collaborative state characteristics as control input, and calculating the adjustment amount of input node association factors through a proportional-integral-derivative (PID) controller. The adjustment ratio of the node connection quantity coefficient is positively correlated with the change in the number of synchronous interactions, and the proportional coefficient is determined through stability analysis; the adjustment level of the connection priority coefficient is dynamically allocated according to the change in the number of initiating core identifiers and the core performance ranking, with the priority increase of high-performance cores being greater than that of low-performance cores; the tightening degree of the connection activation condition coefficient is positively correlated with the change in signal response time, specifically manifested as an increase in the activation threshold and a decrease in the activation window.

[0058] Step S1321: Obtain the number of core synchronization interactions per unit time in the current real-time scene features, denoted as the current synchronization count; and obtain the number of core synchronization interactions per unit time in the previous real-time scene features, denoted as the historical synchronization count.

[0059] The current synchronization count is extracted from the core collaborative state feature dimension of the real-time scene characteristics, while the historical synchronization count is read from the previous real-time scene feature data cached locally. Data reading requires locking protection to avoid concurrent access conflicts. If the historical synchronization count is found to be zero during the reading process (such as the first run), the initial synchronization count preset by the system is used (determined according to the hardware configuration of the multi-core board).

[0060] Step S1322: Calculate the first difference between the current synchronization count and the historical synchronization count. When the first difference is positive, calculate the first ratio between the first difference and the historical synchronization count, and increase the node connection quantity coefficient by the value corresponding to the first ratio. When the first difference is negative, calculate the absolute value of the first difference and the second ratio between the historical synchronization count, and decrease the node connection quantity coefficient by the value corresponding to the second ratio.

[0061] The first difference is calculated by subtracting the historical synchronization count from the current synchronization count. The first ratio is the first difference divided by the historical synchronization count (positive value). The second ratio is the absolute value of the first difference divided by the historical synchronization count. The adjustment amount for the node connection quantity coefficient is the current coefficient value multiplied by the ratio, and the adjustment direction is determined by the sign of the first difference. For example, if the current coefficient value is 0.5 and the first ratio is 0.2, then the adjusted coefficient value is 0.5 × (1 + 0.2) = 0.6.

[0062] Step S1323: Obtain the set of core identifiers that initiate synchronous interaction in the current real-time scene features, denoted as the current core set; and obtain the set of core identifiers that initiate synchronous interaction in the previous real-time scene features, denoted as the historical core set. Compare the current core set with the historical core set to identify newly added and reduced core identifiers.

[0063] The core identifier set is stored using an unordered set data structure. Comparison operations are implemented through set difference operations: adding a new core identifier is the result of subtracting the historical core identifier from the current core identifier set, and reducing a core identifier is the result of subtracting the current core identifier from the historical core identifier set. Set operations are optimized using bitwise operations (core identifiers are integers). For a 32-core system, a 32-bit integer bitmask can be used to represent the set, and set difference operations are implemented through bitwise AND and bitwise NOT operations.

[0064] Step S1324: Obtain the predefined core performance configuration table of the multi-core board, which records the quantized value of the benchmark computing power corresponding to each core identifier.

[0065] The core performance configuration table is stored in the non-volatile memory (such as EEPROM, Flash) of the multi-core board and loaded into memory during system startup. The benchmark computing power is quantified through benchmark tests (such as Dhrystone, CoreMark), including individual scores for integer arithmetic, floating-point arithmetic, memory bandwidth, and a comprehensive score. In this step, the comprehensive score is used as the basis for priority adjustment.

[0066] Step S1325: For each newly added core identifier, query the core performance configuration table to obtain its baseline computing power quantification value, determine the connection priority coefficient boost value corresponding to the newly added core identifier according to the preset priority boost coefficient mapping relationship, and add this boost value to the connection priority coefficient of the target model adjustment factor.

[0067] The priority boosting coefficient mapping relationship is represented by a piecewise linear function, dividing the baseline computing power quantization value into multiple intervals, each interval corresponding to a fixed boost value. For example, the boost value is 0.1 in the interval [0, Q1), 0.2 in the interval [Q1, Q2), and so on, where Q1 and Q2 are preset performance thresholds. The accumulation operation is implemented using atomic addition to avoid multi-threaded conflicts.

[0068] Step S1326: For each reduced core identifier, query the core performance configuration table to obtain its baseline computing power quantification value, determine the connection priority coefficient reduction value corresponding to the reduced core identifier according to the preset priority reduction coefficient mapping relationship, and subtract this reduction value from the connection priority coefficient of the target model adjustment factor.

[0069] The priority reduction coefficient mapping relationship is similar to that of the promotion coefficient, but the absolute value of the reduction is positively correlated with the quantified value of the baseline computing power. For example, the reduction value when reducing high-performance cores is greater than that when reducing low-performance cores. Subtraction operations also use atomic operations to ensure data consistency.

[0070] Step S1327: Obtain the signal response duration data in the current real-time scene features, calculate the average signal response duration, and record it as the current average duration; obtain the average signal response duration in the previous real-time scene features, and record it as the historical average duration.

[0071] The current average duration is obtained by taking the arithmetic mean of the response duration distribution feature vectors of the core collaborative state features in the real-time scene features. The historical average duration is read from the cached previous scene feature data. If the response duration data is empty (no synchronous interaction), the average duration is taken as the preset minimum value.

[0072] Step S1328: Calculate the second difference between the current average duration and the historical average duration; when the second difference is positive and exceeds the preset response threshold, adjust the connection activation condition coefficient to a mode where at least two cores need to initiate activation signals simultaneously to establish a connection; when the second difference is negative and exceeds the preset response threshold, adjust the connection activation condition coefficient to a mode where only a single core needs to initiate an activation signal to establish a connection; when the absolute value of the second difference does not exceed the preset response threshold, keep the connection activation condition coefficient unchanged.

[0073] The preset response threshold is set according to system stability requirements, typically 20% of the historical average duration. The mode switching of the connection activation condition coefficient is implemented by modifying the state machine. In dual-core activation mode, the activation signal must satisfy "AND" logic, while in single-core activation mode, it must satisfy "OR" logic. During mode switching, the input node connection needs to be briefly frozen to avoid state inconsistencies.

[0074] Step S1329: Record the final adjustment ratio of the node connection quantity coefficient as the increase or decrease ratio of the input node connection quantity; record the final adjustment amount of the connection priority coefficient and convert it into the corresponding level change in the preset priority level system as the adjustment level of the connection priority; record the final value of the connection activation condition coefficient, clarify its corresponding core quantity requirements and signal type as the specific requirements of the connection activation condition.

[0075] The adjustment ratio is (adjusted coefficient value - original coefficient value) / original coefficient value, expressed as a percentage; the priority level system adopts a 5-level system (levels 1-5), and the mapping relationship between the adjustment amount and the level change is realized by looking up a table; the specific requirements for connection activation conditions include parameters such as the number of cores (1 or 2), signal type (level signal or pulse signal), and signal duration threshold.

[0076] Step S13210: Integrate the increase / decrease ratio of the number of input node connections, the adjustment level of the connection priority, and the specific requirements of the connection activation conditions into structured data records to generate input node association adjustment rules.

[0077] Structured data records are formatted using Extensible Markup Language (XML) and include elements such as rule ID, generation timestamp, adjustment ratio, priority level, and activation conditions. XML documents must undergo syntax validation and signing to ensure integrity and immutability.

[0078] Step S133: Adjust the values ​​of intermediate layer parameter interaction factors in the target model adjustment factors according to the differences in task load distribution characteristics. If the task dwell time increases, increase the parameter coupling strength coefficient; if the task transfer frequency increases, adjust the parameter transfer order coefficient to adapt to the task transfer rhythm; if the core load proportion increases, optimize the parameter feedback coefficient to balance the load; based on the adjusted values ​​of intermediate layer parameter interaction factors, generate intermediate layer parameter interaction adjustment rules, and determine the adjustment range of parameter coupling strength, the arrangement of parameter transfer order, and the triggering conditions of parameter feedback.

[0079] The adjustment of the intermediate layer parameter interaction factors employs an adaptive filtering algorithm, using the differences in task load distribution characteristics as noise input and dynamically adjusting the factor values ​​through the minimum mean square error (LMS) algorithm. The adjustment magnitude of the parameter coupling strength coefficient is positively correlated with the change in task dwell time; the longer the dwell time, the higher the coupling strength to enhance feature extraction capabilities. The parameter transfer order coefficient undergoes path reordering based on a Markov chain model of task transfer frequency, ensuring the transfer order matches the task transfer path. The optimization objective of the parameter feedback coefficient is to minimize the variance of the core load percentage, and the adjustment amount of the feedback coefficient is calculated using a gradient descent algorithm.

[0080] Step S134: Adjust the output dimension adaptation factor value in the target model adjustment factor according to the differences in data transmission path characteristics. If the number of intermediate cores increases, increase the output node number coefficient; if the link switching frequency increases, adjust the output data format coefficient to adapt to link switching; if the link transmission time increases, optimize the output-core requirement correspondence coefficient to reduce transmission delay; based on the adjusted output dimension adaptation factor value, generate output dimension adaptation adjustment rules, determine the adjustment scale of the number of output nodes, the conversion standard of output data format, and the matching method between output and core requirements.

[0081] The adjustment of the output dimension adaptation factor adopts a predictive control method, predicting future transmission needs based on differences in data transmission path characteristics and adjusting factor values ​​in advance. The adjustment scale of the output node quantity coefficient is positively correlated with the change in the number of intermediate cores, and a greedy algorithm is used to determine the order of node addition and subtraction. The output data format coefficient selects the encoding scheme according to the link switching frequency. Low-redundancy compression encoding is used for high-frequency switching, and high-error-correction encoding is used for low-frequency switching. The optimization of the coefficient corresponding to the output and core needs is achieved through reinforcement learning, with the minimization of transmission delay as the reward function, and the weight of the correspondence is dynamically adjusted.

[0082] Step S135: Integrate the input node association adjustment rules, the intermediate layer parameter interaction adjustment rules, and the output dimension adaptation adjustment rules, and sort them according to the execution order of the inference model adjustment to form an inference model structure parameter adjustment scheme.

[0083] The integration of adjustment rules employs a Directed Acyclic Graph (DAG) scheduling model, determining the execution order of rules through topological sorting. Input node-related adjustment rules are designated as preceding nodes, intermediate layer parameter interaction adjustment rules as intermediate nodes, and output dimension adaptation adjustment rules as following nodes. The DAG model must be verified for acyclicity; if circular dependencies are found, they are eliminated through rule decomposition. The integrated structural parameter adjustment scheme includes an execution order table, a rule dependency graph, and a parameter adjustment matrix, stored in binary format to improve loading speed.

[0084] For example, step S1351: Analyze the dependencies of each rule during the inference model adjustment process, and determine that the input node association adjustment rule should be executed first; the intermediate layer parameter interaction adjustment rule should be executed after the input node association adjustment rule; and the output dimension adaptation adjustment rule should be executed after the intermediate layer parameter interaction adjustment rule.

[0085] Dependency analysis employs static code analysis techniques to scan the operation objects of each adjustment rule (such as input nodes, intermediate layer parameters, and output nodes). If the operation object of rule B depends on the operation result of rule A, then rule A is a prerequisite dependency of rule B. Input node association adjustment rules modify the input layer structure to provide input signals to the intermediate layer, and therefore must be executed first. Intermediate layer parameter interaction adjustment rules adjust intermediate layer parameters based on the input layer output and must be executed after the input layer adjustment. Output dimension adaptation adjustment rules adjust the output layer structure based on the intermediate layer output and must be executed after the intermediate layer adjustment.

[0086] Step S1352: Determine the execution order based on the dependency relationship as follows: input node association adjustment rules, intermediate layer parameter interaction adjustment rules, and output dimension adaptation adjustment rules.

[0087] The execution order is stored using a linear list, where each element is a rule ID and an execution entry address. The list must support dynamic insertion and deletion to adjust the order when rules are updated. Once the execution order is determined, simulation verification is required to confirm that there are no resource conflicts or data races when the rules are executed sequentially.

[0088] Step S1353: Add an execution step description for each rule. The execution step description for the input node association adjustment rule includes the specific operation process for adding / deleting input nodes, the operation order for adjusting connection priority, and the operation method for configuring activation conditions. The execution step description for the intermediate layer parameter interaction adjustment rule includes the operation process for adjusting parameter coupling strength, the operation order for reorganizing parameter passing order, and the operation method for configuring parameter feedback. The execution step description for the output dimension adaptation adjustment rule includes the operation process for adding or removing output nodes, the operation order for converting output data format, and the operation method for configuring output correspondence.

[0089] The execution steps are described using a mix of natural language and pseudocode. Each step includes elements such as the operation object, operation type, parameter list, and expected result. The process for adding an input node must describe the order of node memory allocation, weight initialization, and connection establishment; the process for reordering parameter passing must describe the steps of path search, node sorting, and connection update; and the process for output data format conversion must describe the process of encoding method selection, format conversion algorithm, and checksum calculation.

[0090] Step S1354: Add parameter descriptions for each rule execution step, determine the source of parameters and the basis for parameter values ​​involved in each operation step, and the source of parameters all point to the adjusted target model adjustment factor value.

[0091] The parameter descriptions are presented in tabular form (text description), listing the parameter name, data type, value range, source rule ID, calculation method, and other information. For example, the parameter for the number of new input nodes comes from the node connection quantity coefficient in the input node association adjustment rule, and its value is the coefficient value multiplied by the number of base nodes; the parameter coupling strength adjustment range comes from the parameter coupling strength coefficient in the intermediate layer parameter interaction adjustment rule, and its value is the product of the coefficient value and the baseline strength.

[0092] Step S1355: Integrate the input node association adjustment rules, intermediate layer parameter interaction adjustment rules, output dimension adaptation adjustment rules arranged in execution order, as well as the corresponding rule execution step descriptions and parameter descriptions into a target document.

[0093] The target document is generated in Portable Document Format (PDF) and includes a cover, table of contents, main text, and appendices. The main text describes each rule in chapters according to the execution order, while the appendices contain auxiliary information such as parameter description tables and dependency diagrams. The PDF document must be watermarked and have a version number added for easy traceability and management.

[0094] Step S140: Based on the structural parameter adjustment scheme, the inference model is restructured and its parameters are updated to obtain an updated inference model adapted to the current multi-core board scenario. The updated inference model is then deployed to the multi-core board for operation.

[0095] This step is the core execution stage of adaptive updating of the inference model. By dynamically adjusting the model structure and optimizing parameters, the inference model can be matched with the current multi-core board operating scenario, thereby improving inference efficiency and resource utilization.

[0096] Step S141: Call the inference model reconstruction tool to convert the input node association adjustment rules in the structural parameter adjustment scheme into node operation instructions. The node operation instructions include input node addition instructions, input node deletion instructions, connection priority adjustment instructions, and activation condition configuration instructions.

[0097] The inference model reconstruction tool adopts a modular design, comprising three core modules: a rule parser, an instruction generator, and a model compiler. The rule parser is responsible for parsing the input node association adjustment rules (XML format) in the structural parameter adjustment scheme into an Abstract Syntax Tree (AST). The instruction generator converts the AST into a model reconstruction instruction set (custom binary instruction format), with each instruction containing an opcode and operands (node ​​ID, parameter values, etc.). The model compiler optimizes the instruction set (e.g., instruction rearrangement, redundancy elimination) to generate an executable node operation instruction flow. The opcode for adding an input node is 0x01, and the operands include the node type, initial weight, and a list of target nodes to connect to. The opcode for deleting an input node is 0x02, and the operand is a list of node IDs. The opcode for adjusting connection priority is 0x03, and the operands are the node ID and priority coefficient. The opcode for configuring activation conditions is 0x04, and the operands are the node ID, activation threshold, and activation window size.

[0098] Step S142: Execute the node operation instructions, add a corresponding number of input nodes to the input layer of the inference model according to the input node addition instruction, and each new input node corresponds to a newly added collaborative core in the multi-core board; remove the input nodes in the input layer corresponding to the cores in the multi-core board that no longer participate in collaboration according to the input node deletion instruction; rearrange the connection priority order of the input nodes and intermediate layer nodes according to the connection priority adjustment instruction; set new activation conditions for the connection between the input nodes and intermediate layer nodes according to the activation condition configuration instruction.

[0099] The execution of node operation instructions adopts a pipeline architecture, comprising four stages: instruction fetch, decoding, execution, and write-back. The instruction fetch stage reads instructions from the instruction stream; the decoding stage parses the opcode and operands; the execution stage calls the corresponding hardware acceleration unit (such as GPU or FPGA) to execute the node operation; and the write-back stage updates the model structure parameter storage area. When an input node is added, memory space needs to be allocated for the node (requested from the model memory pool), weight parameters are initialized (using the Xavier initialization method), and connection relationships with intermediate layer nodes are established (stored in the connection matrix). When an input node is deleted, the node's memory space is released to the memory pool, all connection records for that node are cleared from the connection matrix, and the input degree count of the intermediate layer node is updated. When the connection priority is adjusted, the priority field in the connection matrix is ​​modified, and subsequent data transmission will be sorted according to the new priority. When configuring activation conditions, the threshold and window parameters in the connection control register are updated, and the hardware acceleration unit will perform activation judgment based on these parameters during data transmission.

[0100] Step S143: Convert the intermediate layer parameter interaction adjustment rules in the structural parameter adjustment scheme into parameter operation instructions.

[0101] The conversion process of the intermediate layer parameter interaction adjustment rules is similar to that of the input node association adjustment rules. The rule parser parses the XML format rule document and generates an AST containing information such as parameter coupling strength, transmission order, and feedback configuration. The instruction generator converts the AST into a parameter operation instruction set, and the opcodes include parameter coupling strength adjustment (0x11), parameter transmission order reordering (0x12), and parameter feedback configuration (0x13). The model compiler performs dependency analysis on the instructions to ensure the consistency of the parameter operation order.

[0102] Step S144: Execute the parameter operation instructions, modify the parameter coupling coefficients between intermediate layer nodes according to the parameter coupling strength adjustment instructions, and the coefficient modification range is consistent with the adjustment range in the intermediate layer parameter interaction adjustment rules; replan the parameter transmission paths between intermediate layer nodes according to the parameter transmission order reorganization instructions, and the path planning method is consistent with the arrangement method in the intermediate layer parameter interaction adjustment rules; add parameter feedback channels to intermediate layer nodes according to the parameter feedback configuration instructions, and the triggering conditions of the feedback channels are consistent with the triggering conditions in the intermediate layer parameter interaction adjustment rules.

[0103] The execution of parameter operation instructions is completed by a dedicated parameter processing unit, which works in parallel with the model calculation unit. When adjusting the parameter coupling strength, the current coupling coefficient is read from the parameter storage area, multiplied by the adjustment range to obtain the new coefficient, and written back. When reorganizing the parameter transmission order, the transmission path table (using a linked list or array structure) is updated according to the arrangement. Each entry in the transmission path table contains the source node ID, the destination node ID, and the transmission delay. When configuring parameter feedback, the trigger condition register of the feedback channel is set (such as the core load ratio threshold and the parameter deviation threshold). When the conditions are met, the feedback channel is automatically enabled, and the intermediate layer output is transmitted back to the front layer node.

[0104] Step S145: Convert the output dimension adaptation adjustment rules in the structural parameter adjustment scheme into output operation instructions.

[0105] The conversion process for output dimension adaptation and adjustment rules is also implemented through a rule parser, instruction generator, and model compiler. The opcodes of the output operation instruction set include output node addition / reduction (0x21), output data format conversion (0x22), output correspondence configuration (0x23), etc., and the operands include the number of nodes, format type, core ID mapping table, etc.

[0106] Step S146: Execute the output operation instructions to increase or decrease the number of output nodes in the inference model output layer according to the output node increase / decrease instructions. The adjustment scale of the number of nodes is consistent with the adjustment scale in the output dimension adaptation adjustment rules. Convert the data format of the output nodes to a format that matches the core requirements of the multi-core board according to the output data format conversion instructions. The format conversion standard is consistent with the conversion standard in the output dimension adaptation adjustment rules. Set the corresponding interaction relationship between the output nodes and each core of the multi-core board according to the output correspondence configuration instructions. The correspondence relationship is consistent with the matching method in the output dimension adaptation adjustment rules.

[0107] The execution of output operation commands is handled by the output interface controller, which is directly connected to the on-chip network interface of the multi-core board. When output nodes are added or removed, node resources are allocated / released from the model memory pool according to the adjusted scale, and the output buffer is initialized / cleared. When output data format is converted, the format conversion library is called according to the conversion standard (supporting JSON, Protobuf, custom binary formats, etc.). The conversion process includes data serialization, compression, verification, and calculation. When configuring output correspondence, the core interaction mapping table (a many-to-many mapping from output node IDs to core IDs) is updated, and each core is notified to update its interface configuration.

[0108] Step S147: After adjusting the input layer, intermediate layer and output layer, generate a new structure parameter file for the inference model. The new structure parameter file contains the adjusted node connection relationship, parameter values ​​and output configuration information. Load the new structure parameter file into the inference model running framework to form an updated inference model adapted to the current multi-core board scenario.

[0109] The new structure parameter file adopts a hierarchical storage format, including a file header (version number, checksum, creation time), input layer parameter section (number of nodes, connection matrix, activation conditions), intermediate layer parameter section (number of nodes, coupling coefficient, propagation path table, feedback configuration), and output layer parameter section (number of nodes, format information, core mapping table). After the file is generated, it needs to undergo integrity verification (such as CRC32 check). After the verification passes, it is stored in the model storage area. The inference model runtime framework loads the new structure parameter file through dynamic linking, parses the parameters of each layer, and initializes the model instance. During the initialization process, parameter range checks and format conversions are performed to ensure the correctness of the model instance.

[0110] Step S1471: Collect the adjusted node information of the input layer, including the identifier, quantity, connection relationship with intermediate layer nodes and connection activation conditions of the input nodes. Record the identifier, quantity, connection relationship with intermediate layer nodes and connection activation conditions of the input nodes in a preset format to form an input layer structure parameter table.

[0111] Input node identifiers are generated using UUID format to ensure global uniqueness; the number of nodes is the adjusted total number of input layer nodes (32-bit integer); connection relationships are represented using a sparse matrix, with each row corresponding to one input node and each column corresponding to one intermediate layer node, and matrix elements representing connection weights (floating-point numbers); connection activation conditions are recorded as a list of tuples (node ​​ID, threshold, window size). The default format uses a binary stream, with each field stored in a fixed number of bytes (e.g., UUID 16 bytes, quantity 4 bytes, weight 4-byte floating-point number, etc.).

[0112] Step S1472: Collect the adjusted parameter information of the intermediate layer, including the identifier and quantity of intermediate layer nodes, the parameter coupling coefficient between each node, the parameter transmission path and the parameter feedback channel configuration. Record the identifier and quantity of intermediate layer nodes, the parameter coupling coefficient between each node, the parameter transmission path and the parameter feedback channel configuration in a preset format to form an intermediate layer structure parameter table.

[0113] The intermediate layer node identifiers also use UUIDs, and the number is the adjusted total number of intermediate layer nodes; the parameter coupling coefficients are stored in a two-dimensional array (node ​​pairs × coefficient values); the parameter transmission path is recorded as a linked list of (source UUID, destination UUID, delay); the parameter feedback channel configuration includes an array of (trigger threshold, feedback weight, channel status) structures. The preset format is similar to that of the input layer, but variable-length fields for path and feedback configurations are added.

[0114] Step S1473: Collect the adjusted configuration information of the output layer, including the identifier, quantity, data format of the output nodes and their corresponding interaction with each core of the multi-core board. Record the identifier, quantity, data format of the output nodes and their corresponding interaction with each core of the multi-core board according to the preset format to form the output layer structure parameter table.

[0115] Output nodes are identified by UUIDs, and their number corresponds to the adjusted total number of output layer nodes. Data format records are format type enumeration values ​​(e.g., 0x01 represents JSON, 0x02 represents Protobuf). Corresponding interaction relationships are stored using a dictionary structure (output node UUID → core ID list). In the default format, the dictionary structure is implemented using a key-value pair array.

[0116] Step S1474: Integrate the input layer structure parameter table, the intermediate layer structure parameter table, and the output layer structure parameter table into a new structure parameter file for the inference model. The format of the new structure parameter file is consistent with the format supported by the inference model running framework. At the same time, add a generation time identifier and a corresponding multi-core board scenario description to the new structure parameter file.

[0117] The integration process uses a file concatenation method, sequentially writing the input layer, intermediate layer, and output layer structure parameter tables into a file. The file header contains the offset and length of each segment. The generation time is identified by a Unix timestamp (64-bit integer), and the multi-core board scenario description is a UTF-8 encoded string (containing a summary of core collaboration, task load, and data transmission characteristics). The file format must be compatible with the model format of the inference model runtime framework (such as TensorFlow Lite or PyTorchMobile). If incompatible, it must be converted using a format conversion tool.

[0118] Step S1475: Start the inference model running framework, read the new structure parameter file through the parameter loading interface of the inference model running framework, and automatically parse the input layer, intermediate layer and output layer parameter information in the new structure parameter file, and construct the new network structure of the inference model according to the parsing results.

[0119] The parameter loading interface of the inference model's runtime framework is a standard C function (e.g., `model_load(const char *filename)`). Internally, it uses memory mapping (mmap) to read files, avoiding memory overflow when loading large files. The parsing process uses recursive descent analysis to process the parameter tables of each layer. The input layer parses to generate a list of input node objects, the intermediate layers parse to generate a computation graph (node ​​objects + connection edge objects), and the output layer parses to generate output interface objects. The new network structure is represented using a directed graph data structure. Graph nodes contain operation types, parameter pointers, and input / output ports, while graph edges contain connection weights and activation conditions.

[0120] Step S1476: Initialize the parameters of the new network structure in the inference model running framework, assign parameters such as parameter coupling coefficients and parameter passing order from the intermediate layer structure parameter table to the corresponding nodes, configure the connection relationships and activation conditions from the input layer structure parameter table to the connection between the input nodes and the intermediate layer nodes, and apply the output configuration information from the output layer structure parameter table to the output nodes.

[0121] Parameter initialization employs a multi-threaded parallel approach, with each core responsible for initializing a portion of the node parameters. When assigning values ​​to intermediate-layer node parameters, the corresponding values ​​are read from the parameter coupling coefficient array and written to the node's weight register. When configuring the connection between input nodes and intermediate-layer nodes, the non-zero elements of the connection relationship matrix are converted into connection objects (containing source node pointers, destination node pointers, and weight values), and activation conditions (threshold comparator, window timer) are set. When configuring output nodes, the data format and core mapping table are written to the configuration register of the output interface controller.

[0122] Step S1477: After initialization, the inference model running framework generates a new model instance. This model instance is used to perform inference calculations based on the input multi-core board running data, and the calculation process is adapted to the current multi-core board scenario. This model instance is determined as the updated inference model adapted to the current multi-core board scenario.

[0123] Model instances are generated by calling the `create_instance()` method of the inference model runtime framework. Each instance object contains members such as a network structure pointer, parameter cache, input / output buffers, and a state machine. During generation, a self-check is performed, including parameter range checks, connection integrity checks, and resource allocation checks. Once the self-check is passed, the model instance enters a ready state, ready to receive input data and perform inference calculations.

[0124] Step S148: Transfer the updated inference model's runtime file to the model storage area of ​​the multi-core board, replace the original inference model runtime file in the multi-core board, start the model loading program of the multi-core board, and enable the updated inference model to participate in the multi-core board's task processing, data interaction, and core collaboration processes.

[0125] The transmission of model runtime files uses a reliable transmission protocol (such as TCP). Before transmission, the MD5 hash value of the file is calculated and sent to the receiving end. After receiving the file, the receiving end verifies the hash value to ensure file integrity. The model storage area of ​​the multi-core board is on-chip Flash or external eMMC. Replacement operations require erasing the sector where the old file is located before writing the new file. The erasing and writing processes must be protected by locks to prevent interruption and file corruption. The model loader is the multi-core board's bootloader or operating system service. It starts the updated inference model through system calls (such as execve()). After the model starts, it registers with the system service list, receives inference tasks assigned by the task scheduler, and participates in the real-time operation of the multi-core board.

[0126] Step S150: Collect new scene features of the multi-core board during the operation of the updated inference model, and use the new scene features to update the correspondence between scene feature types and model adjustment factors in the inference model adjustment factor library to maintain the adaptive update capability of the inference model.

[0127] This step is a closed-loop feedback loop for the adaptive update of the inference model. By continuously learning new scenario features and constantly optimizing and adjusting the factor library, the inference model can maintain good adaptability when the multi-core board scenario changes.

[0128] Step S151: During the operation of the updated inference model, a fixed data acquisition time interval is set, and the new operating data of the core processing unit is obtained through the data acquisition module of the multi-core board according to the acquisition time interval. The new operating data includes new inter-core synchronization data signals, new real-time records of task scheduling, and new real-time logs of data transmission.

[0129] The data acquisition interval is dynamically adjusted based on the scene change rate of the multi-core board. Initially set to 1 minute, if the scene feature change rate from multiple consecutive acquisitions is less than a threshold (e.g., 5%), the interval is increased (maximum 5 minutes); conversely, the interval is decreased (minimum 10 seconds). The data acquisition module uses DMA (Direct Memory Access) to read data from core registers and peripheral buffers, avoiding CPU intervention. During the acquisition process, a timestamp (accurate to milliseconds) and a version number (incrementing with each model update) are added to the new running data. The format of new inter-core synchronization data signals, real-time task scheduling records, and real-time data transmission logs is consistent with the running data format in step S121, ensuring compatibility for subsequent processing.

[0130] Step S152: Process the new core-to-core synchronization data signal, and count the number of core-to-core synchronization interactions per unit time, the number of core identifiers that initiate synchronization interactions, and the signal response time of each synchronization interaction to form new core collaboration state characteristics.

[0131] The processing procedure is the same as step S122, including time interval division, counting the number of synchronous interactions, deduplication of the core identifier, and response time calculation, ultimately generating a new core collaboration state feature record containing a timestamp identifier. Unlike historical processing, the new core collaboration state feature needs to be stored in association with the version number of the updated inference model in order to analyze changes in scene features under different model versions.

[0132] Step S153: Process the new task scheduling real-time record, extract the dwell time of the new task in each core processing unit, the number of times the new task is transferred between different core processing units, and the load ratio of each core processing unit to form a new task load distribution characteristic.

[0133] The processing method is the same as in step S123, using the same feature extraction algorithm and data structure to ensure that the new task load distribution characteristics are comparable to historical characteristics. When calculating the new load percentage, the computational resource consumption of the updated inference model (such as CPU time and memory bandwidth used by the model) must be included in the statistics.

[0134] Step S154: Process the real-time log of the new data transmission to identify the number of intermediate core processing unit identifiers, the number of new transmission link switching times, and the transmission time length of each new transmission link during the new data transmission process, and form new data transmission path characteristics.

[0135] The processing procedure is the same as step S124. By parsing the new data transmission real-time logs, feature parameters such as intermediate core identifiers, link switching counts, and link transmission time lengths are extracted to form a new data transmission path feature vector. The new feature vector must include output data transmission-related indicators of the updated inference model to distinguish it from the feature vector before the model update.

[0136] Step S155: Combine the new core collaboration state feature, the new task load distribution feature, and the new transmission path feature into a new multi-core board scenario feature.

[0137] The combination method is the same as in step S125, using a three-dimensional normalized vector to represent the new scene features. Each element of the vector corresponds to the normalized value of the new scene feature. During the combination process, a model version number field needs to be added to record the version of the inference model used to generate the new scene feature.

[0138] Step S156: Access the inference model adjustment factor library, compare the new scene feature with the existing scene feature types in the inference model adjustment factor library, if there is a scene feature type that perfectly matches the new scene feature, extract the model adjustment factor corresponding to the scene feature type, and adjust the numerical range of the model adjustment factor according to the specific value of the new scene feature; if there is no perfectly matching scene feature type, define the new scene feature as a new scene feature type, analyze the impact of the new scene feature type on the operation of the inference model, determine the model adjustment factor and numerical range that are suitable for the new scene feature type, and add the new scene feature type and the corresponding model adjustment factor to the inference model adjustment factor library.

[0139] The comparison algorithm uses cosine similarity calculation. A complete match is determined when the cosine similarity between the new scene feature vector and the feature type vectors of each scene in the library is greater than a threshold (e.g., 0.95). For a complete match, the numerical range of the model adjustment factor is adjusted using a sliding window method. The specific values ​​of the new scene features are included in the statistical window, the mean and variance within the window are recalculated, and the upper and lower limits of the numerical range are updated. For incomplete matches, the definition of the new scene feature type must include a description of the value range of the feature parameters. Impact analysis is conducted through simulation experiments (running the inference model under the new scene features in a test environment and measuring performance indicators such as inference latency, accuracy, and resource utilization). The model adjustment factor is determined using a grid search method, traversing the factor value space to find the optimal numerical range.

[0140] Step S157: Regularly organize the updated inference model adjustment factor library, delete scene feature types and corresponding model adjustment factors that have not appeared for a long time, optimize the storage structure of data in the library, and improve the efficiency of subsequent scene feature matching.

[0141] The cleanup cycle is set to 24 hours and triggered by a scheduled task. The criterion for a prolonged absence is that the last matching time of the scene feature type exceeds a preset period (e.g., 7 days). Before deletion, the data must be backed up to the archive database. Storage structure optimization includes creating feature parameter indexes (B+ tree indexes), compressing redundant data (e.g., duplicate numerical range matrices), and defragmentation (for database files). After optimization, performance testing (e.g., query response time) verifies whether matching efficiency has improved (target improvement of over 20%).

[0142] Step S158: Store the updated inference model adjustment factor library into the database of the multi-core board.

[0143] The stored procedure is executed using a transactional approach. The updated content is first written to a temporary table, and the transaction is committed (atomic operation) after successful verification, ensuring database consistency. The multi-core board uses an embedded database (such as SQLite), stored in a system-specified non-volatile storage partition. After storage, the checksum of the database file is calculated and written to the file header. Verification is performed upon the next startup to detect file corruption.

[0144] Furthermore, Figure 2 A schematic diagram of a hardware structure for implementing the adaptive update method for inference models in multi-core board scenarios provided in this application embodiment is shown. This hardware structure specifically corresponds to a rack-mounted AI edge inference terminal 100, and the system 100 can be implemented through this rack-mounted AI inference terminal—a fully high-performance AI inference computer whose hardware configuration and functional design are fully adapted to the full-process execution requirements of the method in this application. Figure 2 As shown, the rack-mounted AI edge inference terminal 100 may include at least one processor 102, a memory 104 for storing data, a transmission device 106 for implementing communication functions, and a controller 108. Specifically, the processor 102 adopts a combined architecture of an RK3588 high-performance CPU and multiple high-performance AI inference modules. The RK3588 features a multi-core design (including multiple high-performance cores and high-efficiency cores), capable of handling global system scheduling, data statistical analysis, and rule generation tasks. Each of the multiple AI inference modules possesses independent inference computing power, resulting in a high-performance output for the entire machine. This supports high-density computing tasks such as inference model structure analysis, inter-layer feature transfer dependency calculation, and dynamic adjustment factor matching, ensuring the efficiency of adaptive updates of the inference model in multi-core board scenarios. Those skilled in the art should understand that… Figure 2The structure shown is only illustrative and does not limit the actual structure of the rack-mounted AI edge inference terminal 100. In practical applications, the system 100 can rely on the hardware expansion capabilities of the rack-mounted AI inference terminal to add a dedicated signal conditioning module to adapt to different types of core operation data acquisition, or add a backup power supply unit to improve the operational stability in industrial scenarios. It can also adjust the configuration number of AI inference modules according to the number of cores on a multi-core board to match the needs of inference tasks of different scales.

[0145] The memory 104 adopts a hierarchical storage architecture design, specifically including memory for RK3588, dedicated memory for each AI inference module, and multiple hard disks. The memory of RK3588 is used to load the program instructions corresponding to the method of this application (such as the execution code of the scene feature extraction module, the adjustment factor library matching module, and the model structure reconstruction module) and real-time computing data (such as the core collaboration status data of the multi-core board, task load distribution data, and data transmission path data) in real time. The dedicated memory of the AI ​​inference module is used to store intermediate parameters during the inference model adjustment process (such as the real-time calculated value of the input node association factor and the adjustment range of the intermediate layer parameter interaction factor). The multiple hard disks are used for long-term archiving of the inference model adjustment factor library, historical scene feature data, the original structure file of the inference model to be updated, and the running file of the updated inference model. The hard disks adopt a partition allocation strategy, which allows RK3588 and each AI inference module to use independently to avoid data read and write conflicts. The program instructions and data stored in memory 104 can be directly called by processor 102. By running these software programs and functional modules, processor 102 can perform various functional applications and data processing operations related to the adaptive update of the inference model. For example, it can extract features from the collected real-time running data of the multi-core board to generate core collaboration status features, task load distribution features, and data transmission path features; match the real-time scene features with the scene feature types in the adjustment factor library to determine the target model adjustment factor; generate a structural parameter adjustment scheme based on the degree of change of scene features and the target adjustment factor; and reconstruct and update the inference model according to the adjustment scheme, ultimately realizing an adaptive update method for the inference model for changes in the multi-core board scene.

[0146] The transmission device 106, as the core component for data interaction between the system 100 and the multi-core board, external devices, and external networks, integrates multiple types of industrial-grade aviation connectors, specifically including an Ethernet connector, a USB connector, multiple sets of serial communication connectors (RS232, RS422, RS485), and a CAN bus connector. The Ethernet connector supports multi-rate adaptive communication and can be connected to the network interface of the multi-core board via a network cable to realize the transmission of inference model files (such as transmitting the updated inference model's runtime file to the multi-core board) and the acquisition of multi-core board runtime data (such as real-time acquisition of inter-core synchronization data signals and real-time recording of task scheduling). It also supports TCP / UDP protocols to ensure the reliability and real-time performance of data transmission. The USB connector can connect to input devices such as keyboards and mice. The device is used to configure the operating parameters of system 100 (such as setting the scene feature acquisition time interval and adjusting the factor matching threshold), and can also be connected to an external storage device to import historical scene feature data or export the updated adjustment factor library; RS232 / RS422 / RS485 serial communication interface and CAN bus interface are adapted to the heterogeneous communication needs of multi-core boards in industrial scenarios, and can collect fine-grained information such as core load data and data transmission link status data of multi-core boards in real time, providing comprehensive data support for scene feature extraction; in addition, the transmission device 106 also includes an HDMI connector interface, which can be connected to a display to visualize key data in the inference model adjustment process (such as scene feature matching results and parameter comparison before and after model structure adjustment), which is convenient for operators to monitor the method execution status in real time.

[0147] The controller 108 coordinates the runtime sequence and operating status of various hardware components within the system 100. Its functionality relies on the power management board, filter, lithium battery, and indicator lights and switches within the rack-mounted AI inference terminal's enclosure. Specifically, the controller 108 controls the power management board to receive external DC power input, which, after filtering, provides stable power to the processor 102, memory 104, transmission device 106, and AI inference modules. Simultaneously, it controls the power management board to charge the lithium battery and triggers battery discharge when the external power is cut off, ensuring continuous operation of the system 100 without interrupting method execution. During system startup, the controller 108 triggers switch signals to sequentially start the motherboard and each AI inference module, providing startup progress feedback through indicator light status (e.g., a corresponding indicator light stays on when the motherboard starts, and sequentially stays on when an AI inference module starts). It also completes register initialization, memory self-test, hard disk partition mounting, and external interface parameter configuration to ensure all hardware components are ready. During method execution... During the process, the controller 108 monitors the operating status of each component in real time (such as the core temperature of the processor 102, the read / write rate of the memory 104, and the interface connection status of the transmission device 106). If a hardware abnormality is detected (such as interface communication interruption or memory read / write error), a fault handling mechanism is triggered (such as switching to a backup communication interface or restarting the abnormal component), and the fault type is displayed through indicator lights to provide a basis for fault diagnosis. In addition, the controller 108 can also dynamically adjust the allocation of computing resources of the processor 102 (such as allocating more core computing power to the feature extraction module when scene features change frequently) and the read / write priority of the memory 104 (such as increasing the read / write priority of model files during the model update stage) according to the frequency of changes in the multi-core board scenario, to ensure that each step of the method is executed efficiently and collaboratively.

[0148] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. A method for updating a reasoning model for multi-core board scene changes, characterized in that, The method comprises: constructing a reasoning model adjustment factor library, the reasoning model adjustment factor library comprising a corresponding set of scene feature types and model adjustment factors, the scene feature types covering core cooperation state features, task load distribution features, and data transmission path features, and the model adjustment factors covering input node association factors, intermediate layer parameter interaction factors, and output dimension adaptation factors; extracting real-time scene features during the operation of the multi-core board, matching the real-time scene features with the scene feature types in the reasoning model adjustment factor library, and determining target model adjustment factors corresponding to the current real-time scene features; generating a structure parameter adjustment scheme for the reasoning model based on the degree of change of the current real-time scene features and the target model adjustment factors, the structure parameter adjustment scheme comprising input node association adjustment rules, intermediate layer parameter interaction adjustment rules, and output dimension adaptation adjustment rules; reconstructing and updating the structure of the reasoning model according to the structure parameter adjustment scheme to obtain an updated reasoning model adapted to the current multi-core board scene, and deploying the updated reasoning model to the multi-core board for operation; collecting new scene features of the multi-core board during the operation of the updated reasoning model, and updating the corresponding relationship between the scene feature types and the model adjustment factors in the reasoning model adjustment factor library using the new scene features to maintain the self-adaptive updating capability of the reasoning model.

2. The method of claim 1, wherein, The construction of the reasoning model adjustment factor library comprises: collecting multi-core board historical operation scene data, the multi-core board historical operation scene data comprising core cooperation data, task allocation data, and data transmission data of different operation stages, and each operation stage data being labeled with a corresponding scene feature description; extracting features from the core cooperation data, dividing different core cooperation state features according to the core synchronization frequency, the number of cores initiating cooperation, and the cooperation signal response time, each core cooperation state feature corresponding to a specific core cooperation data range; extracting features from the task allocation data, dividing different task load distribution features according to the task residence time on each core, the task transfer frequency, and the core load proportion, each task load distribution feature corresponding to a specific task allocation data range; extracting features from the data transmission data, dividing different data transmission path features according to the number of intermediate cores through which data transmission passes, the transmission link switching frequency, and the transmission time of each link, each data transmission path feature corresponding to a specific data transmission data range; determining the input node association factors, the intermediate layer parameter interaction factors, and the output dimension adaptation factors of the reasoning model as model adjustment factors, the input node association factors comprising a node connection number coefficient, a connection priority coefficient, and a connection activation condition coefficient, the intermediate layer parameter interaction factors comprising a parameter coupling strength coefficient, a parameter transmission order coefficient, and a parameter feedback coefficient, and the output dimension adaptation factors comprising an output node number coefficient, an output data format coefficient, and an output and core demand correspondence coefficient; matching a corresponding input node association factor to each core cooperation state feature, and determining a specific numerical range of the input node association factor under different core cooperation state features; matching a corresponding intermediate layer parameter interaction factor to each task load distribution feature, and determining a specific numerical range of the intermediate layer parameter interaction factor under different task load distribution features; matching a corresponding output dimension adaptation factor to each data transmission path feature, and determining a specific numerical range of the output dimension adaptation factor under different data transmission path features; The scene feature types, the numerical ranges of the corresponding model adjustment factors, and the matching logics are integrated into a data set and stored in a special database to form a reasoning model adjustment factor library. An update interface is set for the reasoning model adjustment factor library to receive new scene feature data for content update in the future. 3.The method of claim 1, wherein, The real-time scene features during the operation of the multi-core board are extracted, and the real-time scene features are matched with the scene feature types in the reasoning model adjustment factor library to determine the target model adjustment factor corresponding to the current real-time scene features, including: During the operation of the multi-core board, the running data of the core processing units are acquired in real time through a data acquisition interface. The running data includes inter-core synchronization data signals, task scheduling real-time records, and data transmission real-time logs. The inter-core synchronization data signals are processed to count the number of inter-core synchronization interactions per unit time, the number of core identifiers initiating synchronization interactions, and the signal response time of each synchronization interaction. The counting results of the number of inter-core synchronization interactions per unit time, the number of core identifiers initiating synchronization interactions, and the signal response time of each synchronization interaction are integrated to form a core cooperation state feature. The task scheduling real-time records are processed to extract the residence time of tasks in each core processing unit, the number of task transfers between different core processing units, and the load proportion values of each core processing unit. The extraction results of the residence time of tasks in each core processing unit, the number of task transfers between different core processing units, and the load proportion values of each core processing unit are integrated to form a task load distribution feature. The data transmission real-time logs are processed to identify the number of intermediate core processing unit identifiers passed during data transmission, the number of transmission link switches, and the transmission time length of each transmission link. The identification results of the number of intermediate core processing unit identifiers passed during data transmission, the number of transmission link switches, and the transmission time length of each transmission link are integrated to form a data transmission path feature. The core cooperation state feature, the task load distribution feature, and the data transmission path feature are combined to form a real-time scene feature. Accessing the inference model adjustment factor library, comparing the core cooperation state feature in the real-time scene feature with the core cooperation state feature type in the library, finding the input node association factor corresponding to the core cooperation state feature type with the highest matching degree; comparing the task load distribution feature in the real-time scene feature with the task load distribution feature type in the library, finding the intermediate layer parameter interaction factor corresponding to the task load distribution feature type with the highest matching degree; comparing the data transmission path feature in the real-time scene feature with the data transmission path feature type in the library, finding the output dimension adaptation factor corresponding to the data transmission path feature type with the highest matching degree; Integrating the found input node association factor, intermediate layer parameter interaction factor and output dimension adaptation factor into the target model adjustment factor corresponding to the current real-time scene feature.

4. The method of claim 3, wherein, The core cooperation state feature is formed by integrating the number of inter-core synchronization interactions per unit time, the number of core identifiers initiating synchronization interaction and the signal response time of each synchronization interaction in a unit time, including: Setting a fixed time statistical interval, dividing the inter-core synchronization data signal into a plurality of continuous data signal segments according to the time statistical interval, each data signal segment corresponding to a time statistical interval; In each data signal segment, traverse all synchronization interaction records, count the total number of inter-core synchronization interactions, and take the total number as the number of inter-core synchronization interactions per unit time; In each data signal segment, extract the initiating core identifier in each synchronization interaction record, count the number of remaining core identifiers after removing duplicate core identifiers, and take the number as the number of core identifiers initiating synchronization interaction; In each data signal segment, for each synchronization interaction record, read the synchronization signal initiation time and the synchronization response signal reception time, calculate the time difference between the two time points, and obtain the signal response time of each synchronization interaction; Correlate the number of inter-core synchronization interactions per unit time, the number of core identifiers initiating synchronization interaction and the signal response time of each synchronization interaction corresponding to each time statistical interval, add the corresponding time statistical interval identifier to each record, and form the core cooperation state feature.

5. The method of claim 1, wherein, The structure parameter adjustment scheme of the inference model is generated based on the change degree of the current real-time scene feature and the target model adjustment factor, including: Differences between the current real-time scene features and the last extracted real-time scene features are calculated, including core cooperation state feature differences, task load distribution feature differences, and data transmission path feature differences. The core cooperation state feature differences are represented by the number of synchronous interactions per unit time, the number of synchronous core initiations, and the signal response time. The task load distribution feature differences are represented by the task residence time, the task transfer frequency, and the core load proportion. The data transmission path feature differences are represented by the number of intermediate cores, the link switching frequency, and the link transmission time. The input node association factor values in the target model adjustment factor are adjusted based on the core cooperation state feature differences. If the number of synchronous interactions per unit time increases, the node connection number coefficient is increased. If the number of synchronous core initiations increases, the connection priority coefficient corresponding to the core that meets the multi-core board high load task processing demand is increased. If the signal response time increases, the connection activation condition coefficient is tightened. Based on the adjusted input node association factor values, input node association adjustment rules are generated to determine the increase / decrease ratio of input node connection number, the adjustment level of connection priority, and the specific requirements of connection activation conditions. The intermediate layer parameter interaction factor values in the target model adjustment factor are adjusted based on the task load distribution feature differences. If the task residence time increases, the parameter coupling strength coefficient is increased. If the task transfer frequency increases, the parameter transfer order coefficient is adjusted to adapt to the task transfer rhythm. If the core load proportion increases, the parameter feedback coefficient is optimized to balance the load. Based on the adjusted intermediate layer parameter interaction factor values, intermediate layer parameter interaction adjustment rules are generated to determine the adjustment amplitude of parameter coupling strength, the arrangement method of parameter transfer order, and the triggering conditions of parameter feedback. The output dimension adaptation factor values in the target model adjustment factor are adjusted based on the data transmission path feature differences. If the number of intermediate cores increases, the output node number coefficient is increased. If the link switching frequency increases, the output data format coefficient is adjusted to adapt to the link switching. If the link transmission time increases, the output and core demand corresponding coefficient is optimized to reduce transmission delay. Based on the adjusted output dimension adaptation factor values, output dimension adaptation adjustment rules are generated to determine the adjustment scale of output node number, the conversion standard of output data format, and the matching method of output and core demand correspondence. The input node association adjustment rules, the intermediate layer parameter interaction adjustment rules, and the output dimension adaptation adjustment rules are integrated and sorted according to the execution order of the inference model adjustment to form an inference model structure parameter adjustment scheme.

6. The method of claim 5, wherein, The input node association factor value in the target model adjustment factor is adjusted according to the core cooperation state feature difference. If the number of synchronous interactions per unit time increases, the node connection quantity coefficient is increased. If the number of synchronous core identifiers increases, the connection priority coefficient corresponding to the core that meets the multi-core board high-load task processing requirement is increased. If the signal response time increases, the connection activation condition coefficient is tightened. Based on the adjusted input node association factor value, an input node association adjustment rule is generated to determine the increase / decrease ratio of the input node connection quantity, the adjustment level of the connection priority, and the specific requirements of the connection activation condition, including: Obtain the number of synchronous interactions between cores per unit time in the current real-time scene feature, denoted as the current synchronization number, and obtain the number of synchronous interactions between cores per unit time in the last real-time scene feature, denoted as the historical synchronization number. Calculate the first difference value of the current synchronization number and the historical synchronization number. When the first difference value is positive, calculate the first ratio of the first difference value and the historical synchronization number, and increase the node connection quantity coefficient by the value corresponding to the first ratio. When the first difference value is negative, calculate the second ratio of the absolute value of the first difference value and the historical synchronization number, and decrease the node connection quantity coefficient by the value corresponding to the second ratio. Obtain the core identifier set that initiates synchronous interaction in the current real-time scene feature, denoted as the current core set, and obtain the core identifier set that initiates synchronous interaction in the last real-time scene feature, denoted as the historical core set. Compare the current core set with the historical core set to identify the added core identifiers and the reduced core identifiers. Obtain the core performance configuration table predefined by the multi-core board. The core performance configuration table records the quantitative value of the baseline operation capacity corresponding to each core identifier. For each added core identifier, query the core performance configuration table to obtain its baseline operation capacity quantitative value, determine the connection priority coefficient increase value corresponding to the added core identifier according to the preset priority increase coefficient mapping relationship, and add this increase value to the connection priority coefficient of the target model adjustment factor. For each reduced core identifier, query the core performance configuration table to obtain its baseline operation capacity quantitative value, determine the connection priority coefficient decrease value corresponding to the reduced core identifier according to the preset priority decrease coefficient mapping relationship, and subtract this decrease value from the connection priority coefficient of the target model adjustment factor. Obtain the signal response time data in the current real-time scene feature, calculate the average signal response time, denoted as the current average time length. Obtain the average signal response time in the last real-time scene feature, denoted as the historical average time length. Calculate the second difference value of the current average time length and the historical average time length. When the second difference value is positive and exceeds the preset response threshold, adjust the connection activation condition coefficient to the mode that requires at least two cores to simultaneously initiate an activation signal to establish a connection. When the second difference value is negative and exceeds the preset response threshold, adjust the connection activation condition coefficient to the mode that only requires a single core to initiate an activation signal to establish a connection. When the absolute value of the second difference value does not exceed the preset response threshold, keep the connection activation condition coefficient unchanged. record the final adjustment proportion of the node connection quantity coefficient as the increase or decrease proportion of the input node connection quantity; record the final adjustment amount of the connection priority coefficient and convert it into the corresponding level change in the preset priority level system as the adjustment level of the connection priority; record the final value of the connection activation condition coefficient and specify its corresponding core quantity requirement and signal type as the specific requirement of the connection activation condition; integrate the increase or decrease proportion of the input node connection quantity, the adjustment level of the connection priority, and the specific requirement of the connection activation condition into a structured data record to generate an input node association adjustment rule.

7. The method of claim 1, wherein, reconstruct the structure and update the parameters of the inference model according to the structure parameter adjustment scheme to obtain an updated inference model adapted to the current multi-core board scene, and deploy the updated inference model to the multi-core board for operation, which comprises: calling an inference model reconstruction tool to convert the input node association adjustment rule in the structure parameter adjustment scheme into node operation instructions, wherein the node operation instructions include input node addition instructions, input node deletion instructions, connection priority adjustment instructions, and activation condition configuration instructions; executing the node operation instructions, adding a corresponding number of input nodes to the input layer of the inference model according to the input node addition instructions, each new input node corresponding to a newly added cooperative core in the multi-core board; removing the input nodes in the input layer corresponding to the cores that no longer participate in cooperation in the multi-core board according to the input node deletion instructions; rearranging the connection priority order of the input nodes and the intermediate layer nodes according to the connection priority adjustment instructions; and setting new activation conditions for the connections between the input nodes and the intermediate layer nodes according to the activation condition configuration instructions; converting the intermediate layer parameter interaction adjustment rule in the structure parameter adjustment scheme into parameter operation instructions, wherein the parameter operation instructions include parameter coupling strength adjustment instructions, parameter transfer order reorganization instructions, and parameter feedback configuration instructions; executing the parameter operation instructions, modifying the parameter coupling coefficients between the nodes in the intermediate layer according to the parameter coupling strength adjustment instructions, the coefficient modification amplitude being consistent with the adjustment amplitude in the intermediate layer parameter interaction adjustment rule; re-planning the parameter transfer path between the nodes in the intermediate layer according to the parameter transfer order reorganization instructions, the path planning mode being consistent with the arrangement mode in the intermediate layer parameter interaction adjustment rule; and adding a parameter feedback channel to the nodes in the intermediate layer according to the parameter feedback configuration instructions, the trigger condition of the feedback channel being consistent with the trigger condition in the intermediate layer parameter interaction adjustment rule; converting the output dimension adaptation adjustment rule in the structure parameter adjustment scheme into output operation instructions, wherein the output operation instructions include output node addition or deletion instructions, output data format conversion instructions, and output correspondence configuration instructions; The output operation instruction is executed, the number of output nodes in the output layer of the inference model is increased or decreased according to the output node increase / decrease instruction, the node number adjustment scale is consistent with the adjustment scale in the output dimension adaptation adjustment rule, the data format of the output node is converted to a format matched with the core demand of the multi-core board according to the output data format conversion instruction, the format conversion standard is consistent with the conversion standard in the output dimension adaptation adjustment rule, and the corresponding interaction relationship between the output node and each core of the multi-core board is set according to the output corresponding relationship configuration instruction, and the corresponding relationship is consistent with the matching mode in the output dimension adaptation adjustment rule; After the adjustment of the input layer, the intermediate layer and the output layer is completed, a new structure parameter file of the inference model is generated, the new structure parameter file includes the adjusted node connection relationship, the parameter value and the output configuration information, the new structure parameter file is loaded into the inference model running framework to form an updated inference model adapted to the current multi-core board scene; The running file of the updated inference model is transmitted to the model storage area of the multi-core board to replace the original inference model running file in the multi-core board, and the model loading program of the multi-core board is started to enable the updated inference model to participate in the task processing, data interaction and core cooperation process of the multi-core board.

8. The method of claim 1, wherein, The new scene characteristics of the multi-core board during the running of the updated inference model are collected, and the corresponding relationship between the scene characteristic type and the model adjustment factor in the inference model adjustment factor library is updated using the new scene characteristics, so that the self-adaptive updating capability of the inference model is maintained, including: During the running of the updated inference model, a fixed collection time interval is set, and the new running data of the core processing unit is obtained by the data collection module of the multi-core board at the collection time interval, the new running data includes new inter-core synchronization data signals, new task scheduling real-time records and new data transmission real-time logs; The new inter-core synchronization data signals are processed to count the number of inter-core synchronization interactions in a unit of time, the number of core identifiers initiating synchronization interaction and the signal response time of each synchronization interaction, and a new core cooperation state feature is formed; The new task scheduling real-time records are processed to extract the residence time length of the new task in each core processing unit, the number of times of transferring the new task between different core processing units and the load proportion value of each core processing unit, and a new task load distribution feature is formed; The new data transmission real-time logs are processed to identify the number of intermediate core processing unit identifiers passed in the new data transmission process, the number of switching times of the new transmission link and the transmission time length of each transmission link, and a new data transmission path feature is formed; The new core cooperation state feature, the new task load distribution feature and the new data transmission path feature are combined into a new scene feature of the multi-core board; The new scene feature is compared with the existing scene feature types in the inference model adjustment factor library. If there is a scene feature type that completely matches the new scene feature, the model adjustment factor corresponding to the scene feature type is extracted, and the numerical range of the model adjustment factor is adjusted according to the specific numerical value of the new scene feature. If there is no completely matched scene feature type, the new scene feature is defined as a new scene feature type, the influence of the new scene feature type on the operation of the inference model is analyzed, the model adjustment factor and the numerical range suitable for the new scene feature type are determined, and the new scene feature type and the corresponding model adjustment factor are added to the inference model adjustment factor library. The updated inference model adjustment factor library is periodically sorted, and scene feature types and corresponding model adjustment factors that have not appeared for a long time are deleted to optimize the storage structure of the data in the library and improve the subsequent scene feature matching efficiency. The updated inference model adjustment factor library is stored in the database of the multi-core board.

9. The method of claim 7, wherein the method further comprises: After the adjustment of the input layer, the middle layer and the output layer is completed, a new structure parameter file of the inference model is generated, the new structure parameter file includes the adjusted node connection relationship, parameter value and output configuration information, the new structure parameter file is loaded into the inference model running framework to form an updated inference model adapted to the current multi-core board scene, including: Collecting the adjusted node information of the input layer, including the identification, number, connection relationship with the middle layer node and connection activation condition of the input node, recording the identification, number, connection relationship with the middle layer node and connection activation condition of the input node in a preset format to form an input layer structure parameter table; Collecting the adjusted parameter information of the middle layer, including the identification, number, parameter coupling coefficient between nodes, parameter transmission path and parameter feedback channel configuration of the middle layer node, recording the identification, number, parameter coupling coefficient between nodes, parameter transmission path and parameter feedback channel configuration of the middle layer node in a preset format to form a middle layer structure parameter table; Collecting the adjusted configuration information of the output layer, including the identification, number, data format and corresponding interaction relationship with each core of the multi-core board of the output node, recording the identification, number, data format and corresponding interaction relationship with each core of the multi-core board of the output node in a preset format to form an output layer structure parameter table; The input layer structure parameter table, the middle layer structure parameter table and the output layer structure parameter table are integrated into a new structure parameter file of the inference model, the format of the new structure parameter file is consistent with the format supported by the inference model running framework, and a generation time identifier and a corresponding multi-core board scene description are added for the new structure parameter file; The inference model running framework is started, the new structure parameter file is read through the parameter loading interface of the inference model running framework, the input layer, the middle layer and the output layer parameter information in the new structure parameter file are automatically parsed by the inference model running framework, and a new network structure of the inference model is constructed according to the parsing result. In the inference model running framework, parameters of a new network structure are initialized, parameter coupling coefficients, parameter transmission sequences and other parameters in the intermediate layer structure parameter table are assigned to corresponding nodes, connection relationships, activation condition configurations in the input layer structure parameter table are configured to the connection between the input nodes and the intermediate layer nodes, and output configuration information in the output layer structure parameter table is applied to the output nodes. After the initialization is completed, the inference model running framework generates a new model instance, the model instance is used for inference calculation according to input multi-core board running data, and the calculation process is adapted to the current multi-core board scene, and the model instance is determined as an updated inference model adapted to the current multi-core board scene.

10. A rack-mounted AI edge inference terminal, characterized by, The method comprises a processor and a readable storage medium, the readable storage medium stores a program, and the program is executed by the processor to realize the inference model adaptive updating method for the multi-core board scene change in any one of claims 1-9.

Images