A ship industrial internet dynamic migration fault-tolerant control method and system

Abstract

The application discloses a kind of ship industrial internet dynamic migration fault-tolerant control method and system, it is related to ship industrial internet technical field, comprising: obtaining multiple nodes first state, multiple task second state time sequence and network topology data;Extract the first implicit feature representing node resource fluctuation and the second implicit feature representing task response degradation;Combined with the criticality of task, an initial matching degree mapping matrix is constructed;Based on state correlation, task load features are extracted and a deduction queue is generated in descending order;Determine the optimal node for the head task and perform virtual allocation;Based on the load and implicit feature attenuation target node matching degree, and based on the topology data, perform cohesion penalty to update the matrix;Iterate until the queue is empty to generate a task node mapping table and perform migration;The application effectively solves the control oscillation and topology tearing problem under complex sea conditions, ensuring the steady-state operation of the ship control loop.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology in the shipbuilding industry, specifically to a dynamic migration fault-tolerant control method and system for industrial internet in the shipbuilding industry. Background Technology

[0002] As ships evolve towards intelligence and high automation, the industrial internet of ships has become the infrastructure to ensure the stable operation of core businesses such as propulsion systems, power management, and damage control. In this distributed architecture, a large number of computing nodes collaboratively execute massive industrial control tasks. In order to ensure the high availability of the system, when some nodes experience performance bottlenecks or face the risk of failure, it is usually necessary to dynamically migrate control tasks to other healthy nodes to achieve fault tolerance and load balancing.

[0003] Most existing dynamic migration solutions typically monitor instantaneous resource indicators such as the CPU utilization or memory usage of nodes in real time. Once the system detects that the resource usage of a node exceeds the preset warning line or that there is a response delay, it will directly transfer the task to the node with the most abundant remaining resources.

[0004] However, ships often encounter complex external disturbances such as severe sea conditions and violent vibrations during actual navigation. These drastic changes in the physical environment can cause general and instantaneous performance fluctuations in computing clusters. At the same time, there are often extremely tight implicit data flow couplings between industrial control tasks. If the system still only uses instantaneous resource overruns as migration trigger conditions, when faced with environmental disturbances, tasks will be frequently moved back and forth between seemingly crowded nodes, causing serious control oscillations.

[0005] Even more fatally, this scheduling method, which only focuses on the resources of a single node, completely ignores the data links that are stretched in space after the tasks are forcibly broken up. This will inevitably tear apart the original closed-loop control loop and cause an uncontrollable surge in cross-node communication latency.

[0006] Therefore, accurately identifying the true degradation and evolution of the system, and on this basis completing task scheduling that takes into account the overall topological connectivity stability, so as to avoid the fault-tolerant migration process itself causing a larger systemic collapse, is the core technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a dynamic migration fault-tolerant control method and system for the shipbuilding industrial internet.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect, this invention discloses a dynamic migration fault-tolerant control method for the shipbuilding industrial internet, comprising the following steps:

[0010] Acquire first-state time-series data of multiple nodes, second-state time-series data of multiple tasks running on multiple nodes, and network topology data between nodes;

[0011] Based on a preset time window, feature extraction is performed on the time series data of the first state and the time series data of the second state respectively to generate a first latent feature representing the fluctuation of node resources and a second latent feature representing the deterioration of task response.

[0012] Based on the preset task criticality, first latent feature, second latent feature, and network topology data, a first mapping matrix is ​​constructed to represent the matching degree between the task to be migrated and the candidate nodes.

[0013] Based on the correlation between the first-state time series data and the second-state time series data, the load characteristics of the tasks to be migrated are extracted, and the tasks to be migrated are sorted in descending order according to the load characteristics to generate a simulation queue.

[0014] To deduce the current tasks to be migrated in the queue, the target candidate node with the highest matching degree is determined from the first mapping matrix;

[0015] The current task to be migrated is virtually assigned to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node is decayed. Based on the network topology data, the current task to be migrated and the load characteristics of the remaining unassigned tasks, the topological cohesion penalty is applied to the matching degree of the remaining unassigned tasks to other candidate nodes, and the second mapping matrix is ​​updated and generated.

[0016] The second mapping matrix is ​​used as the new first mapping matrix. The steps of determining the target candidate node and updating the matching degree are executed cyclically until the deduction queue is cleared and the task node mapping table is generated. The migration of the task to be migrated is then performed based on the task node mapping table.

[0017] Secondly, the present invention discloses a dynamic migration fault-tolerant control system for the shipbuilding industrial internet, comprising: a data acquisition module, used to acquire first state time-series data of multiple nodes, second state time-series data of multiple tasks running on multiple nodes, and network topology data between nodes;

[0018] The feature extraction module is used to extract features from the first state time series data and the second state time series data based on a preset time window, respectively, to generate a first latent feature representing the fluctuation of node resources and a second latent feature representing the deterioration of task response.

[0019] The matrix construction module is used to construct a first mapping matrix representing the matching degree between the task to be migrated and the candidate nodes based on the preset task criticality, first latent feature, second latent feature and network topology data.

[0020] The simulation queue generation module is used to extract the load characteristics of the tasks to be migrated based on the correlation between the first state time series data and the second state time series data, and to arrange the tasks to be migrated in descending order according to the load characteristics to generate a simulation queue.

[0021] The target node determination module is used to determine the target candidate node with the highest matching degree from the first mapping matrix for the current task to be migrated in the simulation queue.

[0022] The virtual allocation module is used to virtually allocate the current task to be migrated to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, it decays the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node. Based on the network topology data, the load characteristics of the current task to be migrated and the remaining unassigned tasks, it performs topological cohesion penalty on the matching degree of the remaining unassigned tasks to other candidate nodes and updates and generates the second mapping matrix.

[0023] The loop control module is used to repeatedly trigger the target node determination module and the virtual allocation module by taking the second mapping matrix as the new first mapping matrix until the deduction queue is cleared, generating a task node mapping table, and performing the migration of the task to be migrated based on the task node mapping table.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. A topological cohesion penalty is introduced in the virtual allocation process. By combining cross-node communication latency with task load characteristics for penalty, task clusters with strong correlations are forced to cluster in the same neighborhood in the physical topological space, thereby ensuring the cohesion of the closed-loop control chain and effectively avoiding large fluctuations in communication latency caused by task space splitting.

[0026] 2. By extracting the first latent feature representing the fluctuation of node resources and the second latent feature representing the deterioration of task response, the traditional scheme breaks through the limitation of relying only on instantaneous resource thresholds. It can accurately identify random noise caused by severe sea conditions or violent vibrations of ships, thereby avoiding the control oscillation phenomenon caused by the frequent back-and-forth of tasks between nodes and significantly improving the accuracy of fault-tolerant control.

[0027] 3. When generating the simulation queue, not only are the task load characteristics considered, but also a dynamic scheduling priority value is generated by constructing a nonlinear amplification mapping term. This mechanism can accurately identify those critical tasks that are not only large in scale but also whose response performance is accelerating and collapsing, and move them to the front of the queue for priority allocation, ensuring that the critical control loop can obtain the best computing resources before the system collapses globally.

[0028] 4. The system employs local potential field distortion attenuation to simulate the resource pressure disturbance on the target node after the task is positioned. Combined with iterative simulation, the system reshapes the global matching potential field multiple times in the virtual space, ensuring that the allocation of subsequent tasks always takes into account the new steady-state space after the previous tasks have occupied the positions. This sequential iterative process effectively prevents secondary overload caused by tasks blindly rushing to a single high-quality node, and achieves the optimal balance between global resource distribution and topology resilience. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0031] Figure 2 This is a flowchart illustrating the overall execution process of the method in Embodiment 1 of the present invention.

[0032] Figure 3 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In the evolution of modern ships towards intelligence and high automation, the industrial internet of ships is regarded as the digital foundation for ensuring the stable operation of core businesses such as propulsion control, power management, and damage control monitoring. This type of highly complex control system uses distributed computing nodes as the logical pivot point and real-time Ethernet or industrial buses as rigid information levers to transmit decision commands losslessly to end effectors (such as smart valve groups or propulsion motors), thereby driving the ship to complete navigation attitudes with specific dynamic characteristics at the physical level.

[0035] However, existing fault-tolerant migration techniques lack a verification mechanism for the consistency of energy transfer between the internal resource fluctuations of computing nodes and the task execution response performance. This leads to the system's inability to accurately identify pseudo-health states and topology misalignment issues that may exist in nodes under complex operating conditions. Pseudo-health states manifest as nodes having sufficient remaining resources in real time, but their actual resilience has deteriorated due to aging underlying physical equipment or environmental interference, resulting in a continuous increase in instruction response latency. Topology misalignment occurs when the system finds the optimal node at the computing level during migration, but ignores the tight data flow coupling between tasks, causing tasks to deviate from the original network topology after migration. The energy transfer efficiency is severely offset by cross-node communication losses. Consequently, a strict physical correspondence cannot be established between the computing power distribution of nodes and the response displacement of tasks, leading to misjudgments or ambiguous feedback of the system's risk situation, thus affecting the accuracy of fault-tolerant control and the targeted nature of instructional closed-loop adjustments.

[0036] For example, in real-world scenarios where ships encounter severe sea conditions and experience violent vibrations, conventional monitoring systems can only capture the instantaneous CPU utilization of computing nodes, but cannot distinguish whether the fluctuation is due to a normal increase in task load or hardware performance degradation caused by physical environmental stress. Furthermore, when the system detects that node resources are exceeding limits and initiates migration, if it fails to monitor the autocorrelation of task execution latency and the topological cohesion constraints between nodes, the system may only record the superficial data migration, ignoring abnormal fluctuations in the effective lever arm of the control loop and the decrease in kinetic energy transfer efficiency after migration. Specifically, the system may misclassify a faulty node as the correct target and provide allocation feedback, or incorrectly categorize latency spikes caused by topology tearing as insufficient computing resources. This leads to the control system continuously solidifying erroneous scheduling patterns, failing to form a steady-state operating trajectory consistent with industrial dynamics principles.

[0037] If the above problems are not resolved, the ship's fault-tolerant control system will continue to lose its ability to objectively judge the operating state. Specifically, the failure to identify the instability pressure of nodes will cause tasks to rely excessively on seemingly idle nodes, resulting in the kinetic energy transfer path at the limb level deviating from the system's steady-state principle, thereby weakening the rhythmic performance of the control logic. At the same time, the failure to correct the topology tearing problem between tasks will exacerbate communication friction, making command execution unable to exhibit inertial characteristics, ultimately causing the ship to lose the necessary control force.

[0038] Example 1:

[0039] like Figures 1-2 As shown, a dynamic migration fault-tolerant control method for the shipbuilding industrial internet includes the following steps:

[0040] Step S1: Obtain the first state time series data of multiple nodes, the second state time series data of multiple tasks running on multiple nodes, and the network topology data between nodes;

[0041] Specifically, before system operation, lightweight distributed monitoring agents were pre-deployed at the interfaces between each computing node of the cluster under control and the controlled subsystem. To ensure strict consistency of subsequent multi-dimensional data over time, a high-precision clock synchronization architecture based on the PTP (Precision Time Protocol) was configured at the cluster's underlying layer. This architecture was chosen because the ship propulsion control and power management system has extremely high real-time requirements for command issuance; any microsecond-level clock deviation will cause timing misalignment when analyzing task load characteristics, thus affecting the accuracy of migration decisions.

[0042] Against this backdrop, the system periodically extracts first-state timing data through a monitoring agent. In practical engineering scenarios, this data primarily reflects the CPU utilization sequence. The monitoring agent accesses the operating system kernel statistics interface of the computing nodes, continuously collecting and recording the instantaneous resource occupancy of the nodes at a preset sampling period of 100 milliseconds. Simultaneously, the system performs deep tracking of multiple control tasks running on each node to obtain second-state timing data. In this embodiment, this data specifically manifests as an instruction response delay time sequence. This process is achieved by adding a nanosecond-level timestamp at the control instruction issuance end and recording the acknowledgment timestamp on the execution feedback side of the controlled subsystem. The difference between the two constitutes the real-time performance feedback of the control task during execution on the current node.

[0043] It is worth noting that the system also constructs network topology data characterizing the communication latency between nodes by calculating the probe data packets periodically exchanged between nodes and measuring the round-trip time of the data packets between nodes. This network topology data is stored in matrix form, where each element represents the network round-trip latency time series between a pair of nodes. To cope with the complex electromagnetic environment of the shipbuilding industry and the possible random collisions on the network bus, the system introduces data cleaning and preprocessing logic based on statistical distribution after obtaining the above raw sequence. Specifically, for each sampling point within a preset time window, the system performs noise removal using the following outlier identification formula: ;

[0044] in, The sampled value represents the current moment. This is the average of all sampled values ​​within the window. The standard deviation is used. Based on this criterion, if the abnormal delay caused by instantaneous sensor pulses or occasional bus conflicts at a certain moment reaches several times the mean, that point will be judged as irregular noise and discarded. Subsequently, linear interpolation is used to complete the data. It is worth noting that to prevent issues arising from the standard deviation under system steady-state conditions... To address false rejections caused by noise levels approaching zero, a preset noise threshold is introduced into the formula. .when At that time, the system stops performing outlier removal in order to preserve the true physical fluctuation signals.

[0045] For example, the system is monitoring a task responsible for ship steering control. At this moment, the first state time-series data acquired by the system shows a slight fluctuation in CPU utilization between 30% and 45%, while the simultaneously acquired second state time-series data, i.e., control command response latency, remains around 5 milliseconds. However, if the system detects a sudden jump in the network round-trip latency time series in the network topology data from 2 milliseconds to 50 milliseconds, and satisfies the outlier identification formula mentioned above, the system will trigger a cleaning process. Through this deep-aware acquisition method, the system not only obtains accurate CPU utilization sequences, command response latency time series, and network round-trip latency time series, but also ensures that all data has undergone high-precision spatiotemporal alignment and noise filtering before entering the subsequent feature extraction stage.

[0046] Step S2: Based on a preset time window, extract features from the time series data of the first state and the time series data of the second state respectively to generate a first latent feature representing the fluctuation of node resources and a second latent feature representing the deterioration of task response.

[0047] After initial data acquisition and cleaning, this embodiment proceeds to feature extraction of the first-state and second-state time-series data. The aim is to transform the original physical quantity observations into implicit feature indicators that characterize the system's intrinsic state. Specifically, considering the periodicity of control tasks and the stability of node load fluctuations in the shipbuilding industrial internet, the preset time window is set to 300 seconds to ensure coverage of the complete business cycle and statistical significance. Simultaneously, the system pre-constructs a set of equidistant numerical intervals for the CPU utilization sequence in the first-state time-series data, dividing the 0% to 100% utilization range into 10 preset numerical intervals, each spanning 10%. This interval division method is used to discretize the continuous fluctuation data, thereby enabling the use of information entropy theory to capture the degree of disorder in node resource consumption, rather than simply focusing on absolute values.

[0048] Based on this configuration, in actual operation, this embodiment first performs probability statistical processing on the CPU usage rate sequence within a preset time window. The system counts the frequency of sampling points falling into each numerical interval (e.g., 0-10%, 10-20%) within this window, and divides it by the total number of sampling points to obtain the probability of data occurrence corresponding to each numerical interval. Furthermore, the system calculates the resource load fluctuation entropy based on these probability values ​​and uses it as the first latent feature characterizing node resource fluctuations. The specific calculation formula for this first latent feature is as follows:

[0049] ;

[0050] in, This represents the resource load fluctuation entropy of the node being evaluated at the current moment, i.e., the first latent feature. If the node's resource usage remains concentrated in a certain range (e.g., 30-40%), its entropy value is close to 0, indicating that the node is operating extremely robustly. If the resource usage fluctuates drastically across multiple ranges, the entropy value increases, indicating that the node is under significant instability pressure. Assuming that within a certain detection period, the probability distribution of a node's CPU utilization is uniform across 10 ranges, the calculated first latent feature will reach its maximum value, and the system will determine that the node has an extremely high risk of failure.

[0051] Simultaneously, this embodiment extracts the instruction response delay time series of tasks running on the nodes from the second state time series data to generate a second latent feature characterizing task response degradation. This process does not directly use the average delay, but rather uses time series autocorrelation analysis to uncover the trend characteristics of delay fluctuations. Specifically, the system extracts the instruction response delay sequence within the current preset time window and calculates its first-order autocorrelation coefficient as the second latent feature. The calculation process for this second latent feature first requires calculating the mean of the sequence. With variance The first-order autocorrelation coefficient is then extracted using the following formula:

[0052] ;

[0053] in, Representative moment The instruction response delay value, This is the second implicit feature characterizing the degradation of task response. The physical meaning of this feature is to determine whether the evolution of latency has memory: if the coefficient is significantly greater than 0 and close to 1, it indicates that the current increase in latency is not random noise, but shows a continuous inertial trend of degradation, that is, if the current latency is high, it is likely to be even higher in the next moment.

[0054] Through the above calculations, this embodiment successfully transforms the original time-series data into a set of latent feature pairs with deep logical relationships. This allows for the prior identification of potential risk states of nodes and tasks without waiting for actual failures to occur. It is worth noting that these generated first and second latent features are directly output to the subsequent matrix construction stage as key input parameters for constructing the first mapping matrix, thereby ensuring that migration decisions are based on an accurate grasp of the system's dynamic degradation trend.

[0055] After extracting the first and second latent features, this embodiment further performs the fault-tolerant migration trigger determination and object establishment steps.

[0056] Specifically, in order to overcome the problem of false triggering caused by existing technologies relying solely on instantaneous resource overruns, this embodiment constructs a comprehensive instability pressure index to characterize the operational risk of nodes. This index is not a simple instantaneous value, but a second-order fusion of node resource disorder and task performance degradation trend. Its calculation logic is as follows:

[0057] ;

[0058] in, It is the first latent feature. This is the second implicit characteristic of each task running on this node. and These are the preset weighting coefficients.

[0059] The migration trigger condition is defined as: the system continuously monitors the comprehensive instability stress index of each computing node. When a certain node's Exceeding the preset risk threshold If this state is met continuously within a preset confirmation window (such as three consecutive sampling periods), the system determines that the node has entered a state of substantial performance degradation and then triggers the fault-tolerant migration process.

[0060] After the process is triggered, the system further identifies the objects of the migration operation:

[0061] Tasks awaiting migration: These refer to all activity control tasks carried on nodes currently deemed to be in a state of substantial performance degradation. These tasks face the physical risk of being interrupted or exceeding latency limits at any time due to resource disorder or excessive autocorrelation in the responses of their respective nodes.

[0062] Candidate nodes: These refer to healthy computing nodes in the cluster, excluding those experiencing substantial performance degradation, whose overall instability stress index is below a preset safety threshold. These nodes possess sufficient stability potential to accept new tasks, forming the search space for subsequently constructing the mapping matrix.

[0063] After identifying the aforementioned objects, the system, carrying the identification information of the task to be migrated and the state parameters of the candidate nodes, formally enters the mapping matrix construction stage described in step S3. Through this triggering logic based on the evolutionary trend of implicit features, the system achieves a proactive fault-tolerance shift from reporting faults only upon discovery to acting immediately upon degradation, ensuring that migration decisions are based on an accurate assessment of the node's resilience.

[0064] Step S3: Based on the preset task criticality, first latent feature, second latent feature and network topology data, construct a first mapping matrix representing the matching degree between the task to be migrated and the candidate nodes;

[0065] After feature extraction, this embodiment further constructs a first mapping matrix representing the matching degree between the task to be migrated and the candidate nodes. This aims to quantify the optimal placement probability for each task in the candidate cluster through nonlinear fusion of multi-dimensional features. Before formally executing matrix construction, this embodiment pre-constructs and configures a predictive evolution model, i.e., a time series prediction model for the first-state time series data. Specifically, this model employs the Autoregressive Differential Integrated Moving Average (ARIMA) algorithm. This algorithm was chosen because resource consumption in shipbuilding control tasks typically exhibits significant short-term linear trends and local stationarity. The ARIMA model can effectively eliminate non-stationarity and capture the slope characteristics of load evolution by performing third-order differencing on historical CPU usage sequences. The ARIMA model performs parameter optimization using the Akaike Information Content Criterion (AIC), dynamically determining the autoregressive order p and the moving average order q. Model training uses node load data from the past 24 hours as the training set, and online parameter recalibration is performed every hour to adapt to the periodic changes in ship operating conditions.

[0066] The system inputs the CPU utilization sequence within a preset time window into the model, outputs the load prediction value of the nodes within a preset future time period (e.g., the next 10 seconds), and converts it into a parameter of future available resource margin. The physical meaning of this parameter is the expected idle computing capacity of the node, and its calculation logic is as follows: ,in The predicted usage is normalized to the [0,1] interval.

[0067] Based on this, this embodiment generates a first mapping matrix using a weighted and penalized calculation logic, based on a preset task criticality, the first implicit feature, the second implicit feature, and the network topology data. Specifically, the system first retrieves the static task criticality attribute stored in a preset database. This attribute is pre-defined based on the ship's navigation safety level; for example, steering control tasks are assigned a higher weight. Furthermore, the system combines this with the resource load fluctuation entropy generated in the preceding steps. With first-order autocorrelation coefficient And network round-trip delay time series characterizing the physical connectivity between nodes. Numerical fusion is performed according to the matching degree calculation formula defined below:

[0068] ;

[0069] in, Indicates the task to be migrated With candidate nodes The matching degree value between them; These represent the preset first, second, and third weighting coefficients, which are set to 0.5, 0.3, and 0.2 respectively in this embodiment to balance the priorities of computing resources, communication overhead, and operational resilience. The system performs normalization mapping on the calculated matching degree. If the calculation result... If the value is zero, it will be forcibly set to 0, indicating that the node is in an unreachable / unacceptable state in terms of physical topology or performance metrics.

[0070] The physical meaning of this formula is that the matching degree is proportional to the future available resource margin of the candidate node and the stability of its resource fluctuations (characterized by the reciprocal of the first latent feature), while being negatively penalized by the network latency overhead caused by cross-node migration and the historical deterioration trend of the target node's response to the task.

[0071] Assuming there is a criticality For a migration control task with a value of 0.9, the first latent feature of the target candidate node A. The value is 0.1 (extremely stable), representing the predicted future available resource balance. The first latent feature of candidate node A is 0.7 (abundant resources), while the first latent feature of candidate node B is 0.6 (higher volatility), and the predicted resource reserve is 0.8. Even though node B has more remaining resources, after substituting into the formula, node A's final matching degree will be significantly higher than node B's because node A has a higher stability gain and a smaller penalty term for its first latent feature. The system iterates through all tasks to be migrated and all healthy candidate nodes, filling each calculated matching degree value into the corresponding matrix coordinates. The resulting two-dimensional matrix is ​​the first mapping matrix. It is worth noting that this first mapping matrix is ​​not the final migration instruction, but rather serves as the basic situational field for subsequent dynamic updates and topology reshaping of the deduction queue, ensuring that the subsequent iterative determination steps are performed within a globally optimal space that considers the future trends and historical resilience of nodes.

[0072] Step S4: Based on the correlation between the first state time series data and the second state time series data, extract the load characteristics of the task to be migrated, and arrange the tasks to be migrated in descending order according to the load characteristics to generate a simulation queue.

[0073] After completing the initial construction of the first mapping matrix, this embodiment proceeds to extract the load characteristics of the tasks to be migrated and generate a simulation queue. This aims to quantify the implicit resource consumption of each task through correlation analysis between data. Specifically, before executing step S4, this embodiment pre-constructs a load characteristic evaluation model based on statistical correlation. The model's construction logic is based on the following mechanism: In an industrial control node, if the instruction execution performance of a certain task is highly synchronized with the global resource fluctuations of the node, it indicates that the task is the main contributor to the node's resource consumption or is extremely sensitive to resource status. Since multiple tasks share computing resources in actual operating environments, it is difficult to isolate the absolute CPU utilization of a single task through direct observation. Therefore, this embodiment uses the Pearson correlation coefficient as a proxy indicator to measure the degree of coupling between tasks and node resources, effectively revealing the true load intensity of tasks.

[0074] In actual operation step S4, the system first extracts the instruction response delay time series of the task to be migrated on the original faulty node from the second-state timing data, and simultaneously extracts the CPU utilization sequence corresponding to the faulty node from the first-state timing data. The system calculates the Pearson correlation coefficient for sampling points of these two sequences within a preset time window. This coefficient characterizes the degree of linear correlation between changes in task execution delay and fluctuations in node resource usage. Furthermore, to differentiate the priority of different tasks, the system introduces a nonlinear mapping logic based on the natural constant, combined with a preset task criticality, to calculate the load characteristics of each task to be migrated. The specific formula for calculating these load characteristics is as follows:

[0075] ;

[0076] in, Indicates the task to be migrated The load characteristics; Indicates the criticality of the task; This represents the Pearson correlation coefficient between the instruction response latency time series and the CPU utilization time series. Taking the absolute value of the correlation coefficient is to capture any form of strong coupling between tasks and resources.

[0077] Suppose there are two tasks to be migrated on the current faulty node. Task A is the core propulsion control task, and its task criticality is... The criticality is 0.95, and its execution latency fluctuates drastically with CPU usage. The calculated Pearson correlation coefficient has an absolute value of 0.85. Task B is an environmental monitoring data forwarding task, with a criticality of... The correlation coefficient is 0.4, indicating relatively stable execution latency with minimal CPU influence; the absolute value of the correlation coefficient is 0.2. Substituting these values ​​into the formula, due to the amplifying effect of the exponential term, the load characteristics of task A... The load characteristics will be much greater than those of task B. .

[0078] After acquiring the load characteristics of all tasks to be migrated, the system sorts them in descending order based on the specific values ​​of the load characteristics, thereby generating a simulation queue. The descending order is used because tasks with larger load characteristics have the most significant impact on the resource allocation of target candidate nodes and often have higher criticality. Therefore, they must be prioritized in the subsequent potential field simulation to ensure that the system locks in the optimal topology placement for large tasks in the early stages of the simulation, avoiding global resource imbalance caused by smaller tasks occupying high-quality nodes. The final generated simulation queue, as an ordered set of tasks, will be output to the subsequent cyclical simulation process, serving as the input sequence for successively determining target candidate nodes and updating matching degrees. Thus, this embodiment achieves a closed loop from task volume perception to simulation order decision-making, providing a priority-weighted execution benchmark for subsequent adaptive topology reconstruction.

[0079] Step S5: For the current task to be migrated in the deduction queue, determine the target candidate node with the highest matching degree from the first mapping matrix;

[0080] To ensure that in multi-task concurrent migration scenarios, the system can prioritize rescuing control tasks that are not only heavily loaded but also have the most severe degradation trend, this embodiment first performs dynamic reordering before formally establishing target candidate nodes.

[0081] Specifically, at the initial moment of executing step S5, the system first initiates the calculation process of dynamic scheduling priority values ​​for each task to be migrated in the simulation queue. This process is not a simple linear weighting, but is achieved by constructing a nonlinear amplification mapping term with the second latent feature as the independent variable. The reason for introducing the nonlinear amplification mechanism is to capture the nonlinear coupling relationship between the task response degradation rate and the task size, and to prevent some medium-sized but already out-of-control critical control loops from being suppressed at the back of the simulation queue for a long time by large but stable tasks. This nonlinear amplification mapping term is defined as a positive correlation mapping function with respect to the second latent feature (i.e., the first-order autocorrelation coefficient). In this embodiment, an exponential function is specifically used to enhance the sensitivity to high autocorrelation values. Based on this, the system calculates the dynamic scheduling priority value that characterizes the urgency of task rescue, and its calculation formula is as follows:

[0082] ;

[0083] in, Indicates the task to be migrated Dynamic scheduling priority value; This describes the load characteristics of the task. The second latent feature generated by the preceding steps; This is a preset bias amplification factor used to adjust the weight of degradation trend in priority determination.

[0084] After obtaining the dynamic scheduling priority values ​​of all tasks to be migrated, the system reorders the tasks in the simulation queue according to the numerical values ​​from largest to smallest, and generates an updated simulation queue.

[0085] Suppose that Task A has a large load characteristic but its second latent characteristic is close to 0 (i.e., latency fluctuations are uncorrelated over time), while Task B has a medium load characteristic but its second latent characteristic is close to 1 (i.e., latency exhibits extremely strong degradation inertia). After nonlinear amplification mapping, Task B's dynamic scheduling priority will surpass that of Task A, thus being moved to the head of the updated inference queue. This dynamic adjustment mechanism ensures that the system not only considers the absolute resource occupancy of tasks but also proactively responds to the risk of control loop failure.

[0086] Subsequently, the system extracts the task at the head of the updated deduction queue as the current task to be migrated. For this task, the system retrieves the first mapping matrix constructed in the previous steps and iterates through this matrix to search for the matching degree values ​​corresponding to all candidate nodes. At this point, the system executes the maximum value retrieval algorithm to determine the target candidate node using the following mathematical definition:

[0087] ;

[0088] in, That is, for the current task to be migrated The target candidate node with the highest matching degree is identified; This represents the matching degree value in the first mapping matrix; This is the set of all healthy candidate nodes in the current cluster.

[0089] Through the above process, this embodiment not only completes the mapping from matrix to specific execution path, but also achieves real-time protection of critical control loops through in-depth deconstruction of load volume and degradation acceleration. This determination process ensures that the task most in need of a stable environment can prioritize selecting the optimal computing resources in the current global situation. It is worth noting that the determined target candidate nodes... The data will be used as input parameters in subsequent virtual allocation and update processes, thereby triggering local distortion and topological reshaping of the potential field matrix. This ensures that the allocation of remaining tasks is carried out within a new steady-state space that takes into account the current task's occupancy. Thus, this embodiment completes the evolution from static priority to dynamic rescue-style decision-making.

[0090] Step S6: Virtually assign the current task to be migrated to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, reduce the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node. Based on the network topology data, the load characteristics of the current task to be migrated and the remaining unassigned tasks, perform topological cohesion penalty on the matching degree of the remaining unassigned tasks to other candidate nodes, and update and generate the second mapping matrix.

[0091] After identifying target candidate nodes, this embodiment proceeds to the stage of virtually assigning the currently migrated task to the target candidate node and reshaping the global matching degree distribution accordingly. This process is not an actual task migration execution, but rather a simulation of the disturbance to the system resource potential field after the task is assigned through serialized virtual probing. Its core lies in deconstructing the two types of physical impacts generated after the task is assigned: first, the shrinkage of available resources and increased instability of the target node due to the new task; and second, the strain on the original control chain caused by the cross-node communication overhead resulting from the task space splitting.

[0092] Specifically, in step S6 of this embodiment, the system first performs a matching degree attenuation calculation on the target candidate nodes. The system extracts the load characteristics of the current task to be migrated and the first latent characteristics of the target candidate nodes, and combines them with a preset attenuation factor. The system performs fusion and calculates the local potential field distortion attenuation. The physical meaning of this attenuation is to quantify the decrease in the target node's ability to continue carrying subsequent tasks after it is expected to accept the current heavily loaded task. The system subtracts this attenuation from the matching degree before the update to obtain the updated matching degree for the target candidate node.

[0093] ;

[0094] in, Indicates the first In this iteration of the simulation, there are still unassigned tasks in the simulation queue. For target candidate nodes Updated matching accuracy; The load characteristics of the task currently to be migrated; This is the first latent feature of the target candidate node. Through this non-linear deduction, the system can effectively prevent overload crashes caused by a large number of tasks rushing to the same high-quality node due to static optimization.

[0095] Simultaneously, this embodiment applies a topology cohesion penalty to all candidate nodes in the cluster except the target candidate node. The system extracts network topology data between the target candidate node and other candidate nodes, and calculates the topology tearing penalty by combining the load characteristics of the current task to be migrated and the remaining unassigned tasks. This penalty aims to simulate the performance loss caused by the physical splitting of the control loop, and its calculation logic is as follows:

[0096] ;

[0097] in, The preset cohesion penalty weight parameter; The network round-trip delay time series between the target candidate node and other candidate nodes; To infer the load characteristics of the remaining unassigned tasks in the queue, the system subtracts the penalty from the original matching degree to obtain the updated matching degree for other candidate nodes.

[0098] ;

[0099] For example, assuming that task A to be migrated has been virtually assigned to node A, when the system evaluates the matching degree of the remaining task B to migrate to node B, if the network round-trip latency between A and B is significantly higher than a preset benchmark, the calculated topology tear penalty will increase dramatically, thereby significantly reducing the matching degree value of task B on node B. This mechanism forces task clusters with strong implicit associations to tend to cluster in the same physical neighborhood in the network topology space, thus ensuring the cohesion of the ship control chain.

[0100] Finally, the system summarizes all the matching degree values ​​after attenuation and penalty processing, and updates and generates the second mapping matrix. It is worth noting that this second mapping matrix reflects the dynamic feedback of the current task allocation decision on the global load distribution and communication topology in real time. It will serve as the input benchmark for the next round of inference, ensuring that the allocation of subsequent tasks is always in an optimal space of continuous self-reshaping and dynamic game.

[0101] Step S7: Use the second mapping matrix as the new first mapping matrix, and repeatedly execute the steps of determining the target candidate node and updating the matching degree until the deduction queue is cleared and the task node mapping table is generated, and execute the migration of the task to be migrated based on the task node mapping table.

[0102] After completing a single calculation of virtual allocation and potential field reshaping, this embodiment enters the cyclic execution steps of determining target candidate nodes and updating matching degrees until the final task node mapping table is generated and physical migration is performed. This stage searches for the globally optimal topological steady state in a dynamically distorted potential field through serialized iterative game theory. Specifically, before formally entering the cyclic control logic, this embodiment pre-constructs a deduction state management operator. The construction logic of this operator is to maintain a dynamically updated task allocation state vector to record the virtual allocation records generated in each iteration. The reason for adopting this recording method is that in the multi-task scheduling of the Industrial Internet, the placement of each task will instantly change the perception space of subsequent tasks. The system must have the ability to track the entire deduction path to ensure that the generated mapping relationship has spatiotemporal consistency.

[0103] In the actual execution of step S7, the system first determines the state of the deduction queue. Specifically, after a single virtual allocation is completed and the second mapping matrix is ​​generated, the system immediately removes the currently pending migration task that has completed virtual allocation from the deduction queue and replaces the current first mapping matrix with the newly generated second mapping matrix. Subsequently, the system determines whether the deduction queue is empty; this determination process constitutes the termination trigger condition for the loop execution. If the deduction queue is determined not to be empty, the system extracts the task currently at the head of the deduction queue as the new current pending migration task and returns to steps S5 and S6 with the replaced first mapping matrix, that is, under the new potential field situation, it re-searches for the target candidate node with the highest matching degree and performs local distortion and topology penalty update again.

[0104] It is worth noting that if the deduction queue is determined to be empty, it means that all tasks to be migrated have found their optimal placement in the simulation space after multiple potential field reshaping processes. At this point, the system extracts the virtual allocation records of each task to be migrated and its corresponding target candidate node from each iteration, and performs aggregation mapping processing based on all virtual allocation records to generate the final task node mapping table. Logically, this task node mapping table is defined as a set of key-value pairs consisting of task identifiers and node identifiers:

[0105] ;

[0106] in, This represents the generated task node mapping table. The total number of tasks to be migrated. Indicates the first The mapping relationship between each task and its corresponding target candidate node.

[0107] For example, assuming there are three tasks to be migrated in the current simulation queue, the system first assigns node A to task 1 with the highest load characteristics, and then reduces the matching degree of task 2 and task 3 to node A; in the next round of the loop, due to the decrease in the matching degree of node A, task 2 is assigned to node B with the second best matching degree, which in turn triggers the topology penalty for task 3 again; finally, when task 3 is also assigned, the system aggregates and generates a mapping table containing the precise locations of the three tasks.

[0108] Based on the generated task node mapping table, the system enters the physical phase of migration execution. Specifically, the system converts the task node mapping table into a target routing control policy and distributes it to the software-defined network control gateway in the cluster to be controlled. This gateway seamlessly and quickly redirects the data flow of the control task to the mapped target candidate node for execution by updating the switch flow table or modifying the packet encapsulation header of the task execution entity. At this point, the task to be migrated has completed the migration process from the faulty node to the highly stable topology node at the physical level. Subsequently, the system synchronously triggers the collection of new first-state timing data, second-state timing data, and network topology data of the reconstructed cluster, and returns these data streams to step S1 to participate in the next round of perception and analysis, thereby building a continuously adaptive and self-healing closed-loop control in the shipbuilding industrial internet.

[0109] In summary, this embodiment achieves a leap from observing instantaneous physical indicators to evaluating the intrinsic stability of the system by deconstructing the node resource fluctuation entropy and task response autocorrelation. This dynamic mapping matrix, constructed based on implicit features, can not only identify potential degradation trends of computing resources a priori, but also simulate and quantify the dynamic disturbances to the global resource potential field and communication links after the task load is deployed in the extrapolation space through the synergistic effect of local potential field distortion and topological cohesion penalty. It fundamentally avoids frequent control oscillations caused by disturbances in the ship's external environment and effectively solves the problems of topological tearing and data link latency spikes during task migration. Thus, while ensuring the efficiency of control command transmission, it significantly enhances the survivability and closed-loop self-healing capability of the ship's industrial internet under extremely complex working conditions.

[0110] Example 2:

[0111] like Figure 3 As shown, a dynamic migration fault-tolerant control system for the shipbuilding industrial internet includes:

[0112] The data acquisition module is used to acquire the first state time-series data of multiple nodes, the second state time-series data of multiple tasks running on multiple nodes, and the network topology data between nodes;

[0113] The feature extraction module is used to extract features from the first state time series data and the second state time series data based on a preset time window, respectively, to generate a first latent feature representing the fluctuation of node resources and a second latent feature representing the deterioration of task response.

[0114] The matrix construction module is used to construct a first mapping matrix representing the matching degree between the task to be migrated and the candidate nodes based on the preset task criticality, first latent feature, second latent feature and network topology data.

[0115] The simulation queue generation module is used to extract the load characteristics of the tasks to be migrated based on the correlation between the first state time series data and the second state time series data, and to arrange the tasks to be migrated in descending order according to the load characteristics to generate a simulation queue.

[0116] The target node determination module is used to determine the target candidate node with the highest matching degree from the first mapping matrix for the current task to be migrated in the simulation queue.

[0117] The virtual allocation module is used to virtually allocate the current task to be migrated to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, it decays the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node. Based on the network topology data, the load characteristics of the current task to be migrated and the remaining unassigned tasks, it performs topological cohesion penalty on the matching degree of the remaining unassigned tasks to other candidate nodes and updates and generates the second mapping matrix.

[0118] The loop control module is used to repeatedly trigger the target node determination module and the virtual allocation module by taking the second mapping matrix as the new first mapping matrix until the deduction queue is cleared, generating a task node mapping table, and performing the migration of the task to be migrated based on the task node mapping table.

[0119] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0120] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0121] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A dynamic migration fault-tolerant control method for the shipbuilding industrial internet, characterized in that, Includes the following steps: Acquire first-state time-series data of multiple nodes, second-state time-series data of multiple tasks running on multiple nodes, and network topology data between nodes; Based on a preset time window, feature extraction is performed on the first state time series data and the second state time series data respectively to generate a first latent feature representing node resource fluctuations and a second latent feature representing task response degradation. Based on the preset task criticality, the first latent feature, the second latent feature, and the network topology data, a first mapping matrix is ​​constructed to represent the matching degree between the task to be migrated and the candidate nodes; Based on the correlation between the first state time series data and the second state time series data, the load characteristics of the task to be migrated are extracted, and the task to be migrated is sorted in descending order according to the load characteristics to generate a simulation queue. For the current task to be migrated in the simulation queue, determine the target candidate node with the highest matching degree from the first mapping matrix; The current task to be migrated is virtually assigned to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node is reduced. Based on the network topology data, the current task to be migrated and the load characteristics of the remaining unassigned tasks, a topological cohesion penalty is applied to the matching degree of the remaining unassigned tasks to other candidate nodes, and the second mapping matrix is ​​updated and generated. Using the second mapping matrix as the new first mapping matrix, the steps of determining the target candidate node and updating the matching degree are executed cyclically until the deduction queue is cleared to generate a task node mapping table, and the migration of the task to be migrated is performed based on the task node mapping table.

2. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: The first state timing data includes at least a CPU utilization sequence, the second state timing data includes at least an instruction response delay time sequence, and the network topology data includes a network round-trip delay time sequence between nodes. The extraction process of the first latent feature includes: dividing the CPU usage rate sequence within a preset time window into a preset number of numerical intervals, calculating the probability of data occurrence corresponding to each numerical interval, calculating the resource load fluctuation entropy based on the probability of each data occurrence, and using the resource load fluctuation entropy as the first latent feature. The extraction process of the second latent feature includes: extracting the instruction response delay time series of the task running on the node from the second state time series data, performing time series autocorrelation analysis on the instruction response delay time series, extracting the first-order autocorrelation coefficient, and using the first-order autocorrelation coefficient as the second latent feature.

3. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: The process of constructing the first mapping matrix includes: Based on the first state time series data, the percentage of available resources of the node in the future within a preset time period is calculated by the time series prediction model to obtain the future available resource reserve parameter; Based on preset first, second, and third weighting coefficients, the task criticality, the first implicit feature, the future available resource margin parameter, the network topology data, and the second implicit feature are weighted and penalized to obtain the matching degree of each candidate node corresponding to the task to be migrated. The first mapping matrix is ​​constructed by the matching degrees of all tasks to be migrated and all candidate nodes. The formula for calculating the matching degree is: ； in, This indicates the matching degree of the candidate nodes corresponding to the task to be migrated. These represent the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. Indicates the criticality of the task. This represents the first latent feature of the candidate node. This parameter represents the future available resource balance of a candidate node. This represents the network topology data between the original node of the task to be migrated and the candidate node. This represents the second implicit feature of the candidate node in handling the task to be migrated.

4. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 2, characterized in that: The process of extracting the load characteristics of the task to be migrated includes: Extract the instruction response delay time series of the task to be migrated from the second state time series data, and extract the CPU utilization sequence of the node from the first state time series data; calculate the Pearson correlation coefficient between the instruction response delay time series and the CPU utilization sequence; The absolute value of the Pearson correlation coefficient is used as the exponent of the natural constant, and multiplied by the task criticality of the task to be transferred to obtain the load characteristics of the task to be transferred.

5. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: Attenuating the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node includes: multiplying the first latent feature of the target candidate node, the load feature of the current task to be migrated, and a preset attenuation factor to obtain the local potential field distortion attenuation amount; Subtracting the local potential field distortion attenuation from the original matching degree yields the updated matching degree for the target candidate node; the formula for calculating the updated matching degree is: ； in, This indicates the matching degree of the remaining unassigned tasks to the target candidate node after the update. This indicates the matching degree of the remaining unassigned tasks to the target candidate node before the update. This represents the preset attenuation factor. This indicates the load characteristics of the task currently to be migrated. This represents the first latent feature of the target candidate node.

6. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: Perform topology cohesion penalty on the matching degree of the remaining unassigned tasks to other candidate nodes, including: multiplying the network topology data between the target candidate node and other candidate nodes, the load characteristics of the current task to be migrated, the load characteristics of the remaining unassigned tasks, and the preset cohesion penalty weight parameter to obtain the topology tear penalty amount. Subtract the topology tear penalty from the original matching degree to obtain the updated matching degree for other candidate nodes; the formula for calculating the updated matching degree is: ； in, This indicates the matching degree of the remaining unassigned tasks with other candidate nodes after the update. This indicates the matching degree of the remaining unassigned tasks with other candidate nodes before the update. This represents the preset cohesion penalty weight parameter. This represents the network topology data between the target candidate node and other candidate nodes. This indicates the load characteristics of the task currently to be migrated. This indicates the load characteristics of the remaining unassigned tasks.

7. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: The process of generating the task node mapping table includes: Remove the currently pending migration task that has completed virtual allocation from the deduction queue, replace the current first mapping matrix with the updated second mapping matrix, and determine whether the deduction queue is empty; If the deduction queue is not empty, the task to be migrated at the head of the queue is extracted as the new current task to be migrated. Based on the replaced first mapping matrix, the steps of determining the target candidate node with the highest matching degree and updating the matching degree are returned. If the deduction queue is empty, the virtual allocation record of each task to be migrated and the corresponding target candidate node in each iteration is extracted. Aggregate and map all virtual allocation records to generate a task node mapping table.

8. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 1, characterized in that: After generating the simulation queue, and before determining the target candidate node with the highest matching degree from the first mapping matrix for the current migration task in the simulation queue, the process further includes: Extract the second latent features of each task to be transferred in the inference queue; The second implicit features of each task to be migrated are fused with the corresponding load features to generate a dynamic scheduling priority value that represents the urgency of task rescue. Based on the dynamic scheduling priority value of all tasks to be migrated, the tasks to be migrated in the simulation queue are reordered to generate an updated simulation queue. The task at the head of the updated simulation queue is taken as the current task to be migrated, and the step of determining the target candidate node with the highest matching degree is executed.

9. The method for dynamic migration and fault-tolerant control of the shipbuilding industrial internet according to claim 8, characterized in that: The process of generating the dynamic scheduling priority value includes: constructing a nonlinear amplification mapping term with the second latent feature as the independent variable, wherein the nonlinear amplification mapping term is positively correlated with the second latent feature; multiplying the nonlinear amplification mapping term with the load feature corresponding to the task to be migrated to obtain the dynamic scheduling priority value; The reordering process includes: reordering all tasks to be migrated in the simulation queue in descending order according to the dynamic scheduling priority value of each task to be migrated, so as to move the task to be migrated with the largest dynamic scheduling priority value to the head of the updated simulation queue.

10. A dynamic migration fault-tolerant control system for the shipbuilding industrial internet, characterized in that: Using a dynamic migration fault-tolerant control method for the shipbuilding industrial internet as described in any one of claims 1-9, comprising: The data acquisition module is used to acquire the first state time-series data of multiple nodes, the second state time-series data of multiple tasks running on multiple nodes, and the network topology data between nodes; The feature extraction module is used to extract features from the first state time series data and the second state time series data based on a preset time window, respectively, to generate a first latent feature representing node resource fluctuations and a second latent feature representing task response degradation. The matrix construction module is used to construct a first mapping matrix representing the matching degree between the task to be migrated and the candidate nodes based on the preset task criticality, the first latent feature, the second latent feature and the network topology data. The deduction queue generation module is used to extract the load characteristics of the task to be migrated based on the correlation between the first state time series data and the second state time series data, and to arrange the task to be migrated in descending order according to the load characteristics to generate a deduction queue. The target node determination module is used to determine the target candidate node with the highest matching degree from the first mapping matrix for the current task to be migrated in the simulation queue. The virtual allocation module is used to virtually allocate the current task to be migrated to the target candidate node. Based on the load characteristics of the current task to be migrated and the first implicit characteristics of the target candidate node, the matching degree of the remaining unassigned tasks in the inference queue to the target candidate node is reduced. Based on the network topology data, the current task to be migrated and the load characteristics of the remaining unassigned tasks, the module performs topological cohesion penalty on the matching degree of the remaining unassigned tasks to other candidate nodes and updates and generates the second mapping matrix. The loop control module is used to repeatedly trigger the target node determination module and the virtual allocation module by taking the second mapping matrix as the new first mapping matrix until the deduction queue is cleared, generate a task node mapping table, and perform the migration of the task to be migrated based on the task node mapping table.

Images