Semantic-based fault-tolerant scheduling and emergency operation methods and related devices for robots

By constructing a residual capability space matrix and an inspection intent knowledge graph, and generating downgrade replacement instructions, the emergency inspection problem when the robot experiences local hardware failure is solved, ensuring that the robot can still work effectively in a faulty state and improving operational fault tolerance and reliability.

CN123103765APending Publication Date: 2026-11-17NANJING TIANZONG EMBODIED ROBOT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611306263.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively utilize the remaining hardware capabilities to complete emergency inspection tasks when a robot experiences partial hardware failure, causing the robot to become paralyzed and affecting the smooth progress of inspection operations.

Method used

By collecting real-time data on the robot's underlying hardware status, a residual capability space matrix is ​​constructed. Combined with a pre-set capability quantification model and an inspection intent knowledge graph, downgraded replacement instructions are searched and generated, and emergency inspection operations are completed using the robot's residual capabilities.

Benefits of technology

This technology enables robots to effectively complete emergency inspection tasks even in the event of partial hardware failure, maximizing the effective working time after a malfunction, avoiding overall scrapping, and improving operational fault tolerance and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103765A_ABST
    Figure CN123103765A_ABST
Patent Text Reader

Abstract

A semantic-based method and apparatus for fault-tolerant scheduling and emergency operations of robots, relating to the field of intelligent scheduling, are disclosed. The method includes: when a partial hardware failure occurs, converting the robot's currently available hardware performance into a residual capability space matrix; extracting the target action from the user-input emergency voice command and mapping it to a task requirement space matrix; if the residual capability space matrix fails to fully encompass the corresponding dimension threshold in the task requirement space matrix, extracting the core task objective node corresponding to the emergency voice command, and searching for alternative action path combinations in the inspection intent knowledge graph; generating a downgraded replacement instruction from the alternative action path combination that passes secondary verification, and sending the downgraded replacement instruction to the robot's underlying actuator for execution. Implementing the above technical solution enables the robot to effectively complete emergency inspection operations even in the event of partial functional failure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent scheduling, and in particular to a semantic-based fault-tolerant scheduling and emergency operation method and related apparatus for robots. Background Technology

[0002] With the rapid development of technology, robots are being used more and more widely in various fields, such as industrial production, logistics and transportation, and inspection and monitoring. In inspection operations, robots can quickly and accurately acquire environmental information and promptly identify potential problems, playing a vital role in ensuring the stable operation of the system. However, as the functions and complexity of robots continue to increase, the risk of hardware failure also increases accordingly.

[0003] In traditional robot emergency response, the ability to handle hardware failures is limited. When a robot experiences partial hardware failure, the conventional approach is to either stop operation immediately or perform only some simple default actions. The main drawback of existing technology is that it cannot effectively utilize residual hardware capabilities to complete emergency inspection tasks when partial hardware failure occurs. Due to the lack of quantitative assessment of hardware performance and a reasonable task-capability matching mechanism, robots are prone to paralysis when faced with hardware failures, unable to respond to emergency commands in a timely manner, thus affecting the smooth progress of the entire inspection operation. Summary of the Invention

[0004] This application provides a semantic-based robot fault-tolerant scheduling and emergency operation method and related device, which can still effectively complete emergency inspection operations even when the robot's partial functions fail, thereby improving the fault tolerance and reliability of the operation.

[0005] Firstly, this application provides a semantic-based method for fault-tolerant scheduling and emergency operation of robots, the method comprising: Real-time acquisition of the robot's underlying hardware status data; when a local hardware failure occurs, based on a preset capability quantification model, the robot's current available hardware performance is converted into a residual capability space matrix, which represents the robot's current limit movement parameters, available perception dimensions, and remaining computing power. Receive emergency voice commands input by the user, extract the target actions from the emergency voice commands, and map the target actions into a task requirement space matrix; The task requirement space matrix and the residual capability space matrix are compared in a multi-dimensional envelope. If the residual capability space matrix fails to completely enclose the corresponding dimension threshold in the task requirement space matrix, the core operation objective node corresponding to the emergency voice command is extracted according to the pre-constructed inspection intention knowledge graph. Alternative action path combinations that can achieve the core operation objective node and have zero dependence on the failed hardware are searched in the inspection intention knowledge graph. The backup action path combination is compared and verified with the residual capability space matrix for the second time. The backup action path combination that passes the second comparison and verification is used to generate a downgrade replacement instruction, and the downgrade replacement instruction is sent to the robot's underlying actuator for execution, so as to complete the emergency inspection operation in the state of partial functional failure.

[0006] By employing the above technical solution, user intent (demand matrix) and equipment residual capacity (capability matrix) are compared in a mathematical space. When execution conditions are not met, the task is not simply terminated. Instead, the knowledge graph is used to search for alternative action paths that are intact and consume the least resources. This maximizes the utilization of the robot's remaining usable functions after a malfunction, significantly improving the success rate of tasks under extreme conditions and the efficiency of on-site emergency operations.

[0007] In some embodiments, converting the robot's currently available hardware performance into a residual capability space matrix includes: A set of capability dimensions is preset for the robot in its full-function state. The set of capability dimensions includes mobility capability dimension, perception capability dimension, computing power dimension, and overall stability dimension. In the robot's full-function state, rated capability baseline values ​​are set for each capability dimension to form an initial capability space matrix. In the initial capability space matrix, each sub-parameter of each capability dimension is configured with a corresponding rated value range. When any parameter in the underlying hardware status data is detected to exceed the corresponding preset threshold, each sub-parameter of each capability dimension is updated one by one according to the faulty hardware type and fault severity. The sub-parameters affected by the fault are lowered to the actual achievable value, while the sub-parameters not affected by the fault are kept unchanged or appropriately lowered according to preset rules to match the overall machine safety constraints, thereby generating the residual capability space matrix. The residual capability space matrix replaces the initial capability space matrix, and the underlying hardware status data is continuously monitored during the operation of the robot. In response to any secondary change in the hardware status, the capability space matrix recalculation process is repeatedly triggered.

[0008] By employing the above technical solution, a full-function benchmark coordinate system is constructed by pre-setting a four-dimensional capability set and refining it to the rated range of each sub-parameter. When a fault occurs, the system accurately distinguishes between sub-parameters affected by the fault and those unaffected, and adopts targeted adjustment or maintenance strategies. This real-time mapping of abstract hardware damage into a concrete, computer-solvable residual capability space matrix provides a solid data foundation for subsequent accurate intent-capability matching and alternative path search.

[0009] In some embodiments, extracting the target action from the emergency voice command and mapping the target action to a task requirement space matrix specifically includes: The voice signal of the emergency voice command is subjected to noise reduction and speech recognition processing to generate command text. The command text is then filled with semantic slots. The action type slot and the target object slot of the target action are extracted. The position attribute slot and detection type slot of the target object are also extracted. Based on the action type slot, query the preset inspection action-resource requirement mapping table to obtain the requirement level of the target action in each capability dimension; The requirement level of the target action in each capability dimension is converted into the corresponding requirement quantification value, and a task requirement space matrix with the same dimensional structure as the residual capability space matrix is ​​generated. The target slot, the location attribute slot, and the detection type slot are used as parameter labels and bound to the task requirement space matrix for caching.

[0010] By adopting the above technical solution, the extracted action type slots are transformed into quantitative requirement levels in each capability dimension of "mobility, perception, computing power, and stability" through a preset inspection action-resource requirement mapping table. Finally, a task requirement space matrix that is completely consistent with the dimensional structure of the residual capability space matrix is ​​generated. This enables the originally vague and heterogeneous natural language instructions and the precise and structured robot physical capability state to be aligned in the same mathematical coordinate system, laying the foundation for subsequent envelope comparison.

[0011] In some embodiments, the step of performing a secondary comparison and verification between the alternative action path combination and the residual capability space matrix, and generating a downgrade replacement instruction from the alternative action path combination that passes the secondary comparison and verification, specifically includes: Each action node in the backup action path combination is expanded into an atomic action sequence arranged in time sequence. The resource requirement value of each atomic action in each capability dimension is extracted. The resource requirement values ​​of each atomic action are superimposed in time sequence to generate a degradation requirement space matrix. The degradation demand space matrix and the residual capacity space matrix are compared dimension by dimension. When the demand value of each dimension in the degradation demand space matrix does not exceed the capacity value of the corresponding dimension in the residual capacity space matrix, the secondary comparison verification is deemed to be successful. If the secondary comparison and verification passes, the parameter label is filled into the variable parameter bit of each atomic action in the backup action path combination, and the parameterized atomic action sequence is compiled into a standardized control instruction sequence that conforms to the robot's underlying actuator instruction format, as the downgraded replacement instruction; Verify whether the instruction parameters of each atomic action in the downgrade replacement instruction are within the safe execution range of the underlying actuator under the constraint of the residual capability space matrix. If they exceed the safe execution range, reduce the instruction parameters of the corresponding atomic action to the upper limit of the safe execution range and generate corresponding voice feedback information to broadcast to the user.

[0012] By adopting the above technical solution, the abstract backup action path is expanded into a sequence of atomic actions arranged in time sequence, and the resource requirements of each atomic action in each capability dimension are extracted and superimposed in time sequence to generate an accurate degradation demand space matrix. This achieves accurate modeling of the dynamic resource demand curve of the alternative path, fundamentally eliminating the risk of secondary loss of control caused by the instantaneous resource peak exceeding the robot's actual residual capacity during the execution of the alternative solution.

[0013] In some embodiments, the step of extracting the core operational objective node corresponding to the emergency voice command based on a pre-constructed inspection intent knowledge graph, and searching in the inspection intent knowledge graph for alternative action path combinations that can achieve the core operational objective node and have zero dependence on failed hardware, specifically includes: Obtain the target slot of the emergency voice command, input the target action and the target slot into the inspection intent knowledge graph, start from the starting node of the inspection intent knowledge graph, trace along the directed edge to the first intent node corresponding to the target action, the inspection intent knowledge graph is composed of intent nodes, action nodes, device nodes and detection method nodes, and the directed edges between nodes represent the achievement, dependency or substitution relationship; Starting from the first intent node, the source is traced along the relationship in the inspection intent knowledge graph until the root intent node is reached. The root intent node is taken as the core operation objective node, and the core operation objective node represents the final operation objective corresponding to the emergency voice command. Taking the core task objective node as the endpoint, perform a reverse path search in the inspection intent knowledge graph to retrieve all action path combinations that can be derived from the core task objective node and formed through substitution and dependency relationships, as a set of candidate backup action paths. Traverse each action node in the candidate backup action path set, extract the resource requirements of each action node for the failed dimension in the residual capability space matrix, retain the backup action path combination with zero dependency on failed hardware, if there are multiple retained backup action path combinations, sort them according to the total resource consumption of each backup action path combination, and select the backup action path combination with the lowest total resource consumption as the final backup action path combination.

[0014] By employing the above technical solution, a knowledge graph of inspection intentions is constructed, comprising intention nodes, action nodes, device nodes, and detection method nodes, with edge relationships covering achievement, dependence, and substitution. This endows the robot with true semantic reasoning capabilities. Starting from a specific action node, it traces upwards along the achievement relationship to the core root intention node, and then performs a reverse path search with the root intention as the endpoint, retrieving all feasible action paths composed of substitution and dependence relationships. This upgrades fault-tolerant decision-making from passive table lookup to proactive cognition.

[0015] In some embodiments, the step of tracing from the starting node of the inspection intent knowledge graph along directed edges to the first intent node corresponding to the target action specifically includes: Starting from the action node corresponding to the target action, identify the precondition action chain required to complete the target action along the directed edge of the dependency relationship, and map each action node in the precondition action chain to the corresponding action intent node in the inspection intent knowledge graph to form a first intent node set. For each action intent node in the first intent node set, perform pairwise correlation determination. If the first action intent node and the second action intent node share the same core operation objective node and the shortest path distance in the inspection intent knowledge graph is less than a preset distance threshold, then determine that there is an intent coupling relationship between the first action intent node and the second action intent node, and merge multiple intent nodes with the intent coupling relationship into a composite intent node. The composite intent node replaces the multiple intent nodes to perform subsequent tracing operations. After tracing back to the core task destination node, the combined backup action path combination covering all sub-intents in the composite intent node is simultaneously retrieved in the reverse path search. If a target atomic action in the combined backup action path combination contributes to multiple sub-intents in the composite intent node, then a weighted bonus is assigned to the target atomic action to increase the selection priority of backup action path combinations that include atomic actions with the weighted bonus.

[0016] By employing the aforementioned technical solution, a series of preconditional action chains are used as clues to deduce a series of preconditions required to complete the target action, forming a first set of intent nodes. Then, by performing pairwise correlation determination on each intent node in the set, multiple intents with coupling relationships are intelligently merged into a single composite intent node. This mechanism enables the system to identify composite instructions commonly found in actual inspection operations and process them as a unified whole, fundamentally solving the problem of incomplete operations caused by traditional solutions that fragmentedly process composite instructions or simply discard some semantic information.

[0017] In some embodiments, the step of tracing back along the established relationships in the inspection intent knowledge graph, starting from the first intent node, until the root intent node is reached, and using the root intent node as the core task objective node, specifically includes: If multiple root intent nodes are detected in the inspection intent knowledge graph that can be derived to achieve the first intent node, then all root intent nodes that can achieve the first intent node are included in the core objective candidate set. The root intent node with the highest level or the highest confidence level is selected from the core objective candidate set as the core task objective node, and the remaining root intent nodes that are not selected in the core objective candidate set are recorded as second intent nodes. When performing the reverse path search, a first backup action path combination that can reach the core operation target node and has zero dependence on the failed hardware is searched. If there is no complete action path combination in the inspection intent knowledge graph that can reach the core operation target node and has zero dependence on the failed hardware, the second intent node is promoted to the current core operation target node, and the reverse path search is re-executed for the promoted current core operation target node.

[0018] By employing the aforementioned technical solution, the system detects whether multiple root intent nodes in the inspection intent knowledge graph can be deduced to reach the same first intent node, thus achieving a comprehensive perception of multiple high-level operational objectives served by the same action. Based on this, the system does not simply discard unselected root intents; instead, it selects the highest-level or most confident root intent from the core candidate set as the core operational objective node. Simultaneously, it systematically aggregates and marks other root intent nodes as second intent nodes, forming a complete hierarchical intent cognition structure of "intent graph + alternative intents." This mechanism ensures that in subsequent intent downgrading processing, the system always maintains a panoramic understanding of all possible operational objectives behind the user's command, rather than processing only a single surface-level intent.

[0019] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0020] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0021] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0022] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By using a capability quantification model, the robot's currently available limit movement parameters, available perception dimensions, and remaining computing power are precisely quantified into a residual capability space matrix. By mapping the target actions in the voice commands to a task requirement space matrix that is completely consistent with the dimensional structure of the capability matrix, this process transforms the originally ambiguous physical state and spoken commands into a standardized data structure that can be scientifically calculated and quantitatively arbitrated, laying a mathematical foundation for all subsequent intelligent decisions. 2. When the residual capability space matrix cannot fully encompass the task requirement space matrix, the pre-constructed inspection intent knowledge graph is used to trace back from the specific action to the core operation objective node behind the user instruction. Then, with the objective node as the endpoint, all alternative action path combinations that can achieve the objective and have zero dependence on the failed hardware are searched in reverse. This two-way cognitive reasoning mechanism realizes a paradigm shift from passive defense to active completion in fault tolerance logic. 3. By introducing a secondary comparison and verification mechanism, the searched alternative action path combinations are compared again dimension by dimension with the residual capability space matrix to ensure that the alternative solutions are absolutely safe and feasible within the actual physical boundaries. On this basis, the verified action path combinations are generated into standardized control commands that can be directly issued to the underlying actuators. This not only fundamentally eliminates the risk of secondary equipment damage caused by incorrect alternative solutions, but also ensures that as long as there is a trace of usable residual capability in any partial functional failure state of the robot, it can be automatically identified, called up, and put into effective alternative work by the system. This maximizes the effective working time after the failure and avoids the inspection task being completely scrapped due to partial failure. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the semantic-based robot fault-tolerant scheduling and emergency operation method in the embodiments of this application; Figure 2 This is a flowchart illustrating the process of searching for alternative action path combinations in an embodiment of this application; Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0025] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0026] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0027] The following is combined with Figure 1 The method of the embodiments of this application will be described below.

[0028] Please see Figure 1 This is a flowchart illustrating the semantic-based robot fault-tolerant scheduling and emergency operation method in this application embodiment. Figure 1 As shown, the method includes the following steps: S101. Real-time acquisition of the robot's underlying hardware status data. When a local hardware failure occurs, based on a preset capability quantification model, the robot's current available hardware performance is converted into a residual capability space matrix. The residual capability space matrix represents the robot's current limit movement parameters, available perception dimensions, and remaining computing power. S102. Receive emergency voice commands input by the user, extract the target action from the emergency voice commands, and map the target action into a task requirement space matrix; S103. Perform a multi-dimensional envelope comparison between the task requirement space matrix and the residual capability space matrix. If the residual capability space matrix fails to completely enclose the corresponding dimension threshold in the task requirement space matrix, then extract the core operation objective node corresponding to the emergency voice command according to the pre-constructed inspection intention knowledge graph, and search for a backup action path combination in the inspection intention knowledge graph that can achieve the core operation objective node and has zero dependence on the failed hardware. S104. Perform a second comparison and verification between the backup action path combination and the residual capability space matrix. Take the backup action path combination that passes the second comparison and verification to generate a downgrade replacement instruction. Send the downgrade replacement instruction to the robot's underlying actuator for execution, so as to complete the emergency inspection operation in the state of partial functional failure.

[0029] When the robot is in inspection mode, the system collects the robot's underlying hardware status data in real time. This underlying hardware status data includes, but is not limited to: torque feedback values, current values, temperature values, and encoder readings of each joint motor; attitude angle data, angular velocity data, and acceleration data of the inertial measurement unit; pressure data from the foot force sensors; online status and signal quality data of sensing devices such as LiDAR, depth cameras, and infrared thermal imagers; and computing resource data such as CPU utilization, GPU utilization, and available memory capacity. The system has a built-in preset capability quantification model, which is used to transform the robot's abstract hardware performance into quantifiable parameters that can be used by the computer for mathematical operations and logical judgments. Preferably, the capability quantification model includes multiple capability dimensions, at least covering the dimensions of mobility capability, perception capability, computing power, and overall stability. When any underlying hardware status data exceeds the corresponding preset failure threshold, the system determines that a partial hardware failure event has occurred. At this time, the system calls the capability quantification model and, based on the currently functioning hardware resources and the residual performance of some failed hardware, transforms the robot's current usable hardware performance into a multi-dimensional residual capability space matrix. In this application, the residual capability space matrix represents the robot's current limiting mobility parameters, available perception dimensions, and remaining computing power in a vector or tensor mathematical form. The limiting mobility parameters include at least one or more of the following sub-items: maximum translational speed, maximum obstacle-crossing height, maximum climbing angle, minimum turning radius, and executable gait types (e.g., diagonal trotting, crawling gait, standing posture). The available perception dimensions include at least one or more of the following sub-items: the type and number of available sensors, the effective detection distance of each available sensor, the detection accuracy of each available sensor, and the current signal-to-noise ratio of each available sensor. The remaining computing power includes at least one or more of the following sub-items: the percentage of available CPU computing power, the percentage of available GPU computing power, available memory capacity, and available storage space. Preferably, the residual capability space matrix is ​​stored in a standardized format, and the values ​​of each dimension adopt a unified dimension to facilitate mathematical comparison and calculation in subsequent steps. For example, the values ​​of each capability dimension are normalized to the interval [0, 1], where 0 indicates that the capability of that dimension is completely lost, and 1 indicates that the capability of that dimension is in a fully functional state. Preferably, after generating the residual capability space matrix, the system continuously monitors changes in the underlying hardware state data. When any hardware state undergoes a secondary change, the system automatically re-triggers the calculation process of the capability quantification model to dynamically update the residual capability space matrix, ensuring that the residual capability space matrix always remains synchronized with the robot's current actual physical capabilities.

[0030] When the robot experiences a partial hardware failure, on-site maintenance personnel can input emergency voice commands to the robot via voice. Specifically, the system collects the voice signals of the emergency voice commands through a microphone array mounted on the robot. Preferably, the microphone array employs multi-microphone beamforming technology to suppress environmental noise in the substation industrial setting and extract clear voice signals. The system performs the following processing on the collected voice signals sequentially: front-end noise reduction processing is performed on the voice signals. Preferably, front-end noise reduction processing can employ spectral subtraction, Wiener filtering, or a voice enhancement algorithm based on deep neural networks to filter out environmental noise and reverberation; speech recognition processing is performed on the noise-reduced voice signals. Preferably, speech recognition processing is based on an end-to-end deep speech recognition model to convert the voice signals into corresponding command text; semantic parsing is performed on the command text. Specifically, the system extracts the target action from the emergency voice command through a preset semantic parsing model. The target action represents the specific work behavior that the user wants the robot to perform, such as "reading instrument data," "capturing infrared images of equipment," or "inspecting a designated area." The system inputs the extracted target action into a preset task-resource mapping model, mapping the target action into a task requirement space matrix. The task requirement space matrix is ​​a mathematical vector or tensor with the same dimensional structure as the residual capability space matrix. Specifically, the task-resource mapping model pre-defines the correspondence between various standard actions and resource requirements in a power inspection scenario. The values ​​of each dimension in the task requirement space matrix represent the minimum capability level required to complete the target action in that dimension. Preferably, the values ​​of each dimension are also normalized to the [0, 1] interval to maintain consistency with the dimensions of the residual capability space matrix, ensuring comparability between the two. Preferably, during semantic parsing, the system also extracts semantic slot information from the emergency voice commands. The semantic slot information includes at least the target action's object (e.g., "high-voltage cabinet No. 1", "transformer No. 3"), the object's location attribute, and the detection type (e.g., "infrared temperature measurement", "visible light photography", "meter reading"). This semantic slot information is temporarily stored in the system cache for subsequent steps.

[0031] After obtaining the task requirement space matrix and the residual capability space matrix, the system performs a multi-dimensional envelope comparison. Specifically, envelope comparison refers to comparing the requirement value of each dimension in the task requirement space matrix with the capability value of the corresponding dimension in the residual capability space matrix one by one. If the capability value of each dimension in the residual capability space matrix is ​​greater than or equal to the requirement value of the corresponding dimension in the task requirement space matrix, it is determined that the residual capability space matrix completely encloses the task requirement space matrix, meaning the robot's current residual capability is sufficient to directly execute the emergency voice command. If the capability value of at least one dimension in the residual capability space matrix is ​​less than the requirement value of the corresponding dimension in the task requirement space matrix, it is determined that the residual capability space matrix fails to completely enclose the task requirement space matrix, meaning the robot's current residual capability is insufficient to complete the command according to the normal execution method of the target action. When a failure to completely enclose occurs, the system triggers an intent downgrading and reconstruction process. The system acquires a pre-built inspection intent knowledge graph. The inspection intent knowledge graph is a structured semantic network composed of multiple nodes and directed edges. The system comprises nodes including intent nodes, action nodes, device nodes, and detection method nodes. Directed edges represent the logical relationships between nodes, including at least achievement relationships, dependency relationships, and substitution relationships. An achievement relationship indicates that one node can achieve or accomplish another node; a dependency relationship indicates that one node requires another node as a prerequisite before execution; and a substitution relationship indicates that one node can be substituted by another node in an equivalent way. The system inputs the target action and target object corresponding to the emergency voice command into the inspection intent knowledge graph. Starting from the action node corresponding to the target action, the system traces upwards along the directed edges of the achievement relationships until it reaches one or more root intent nodes. The reached root intent nodes are taken as the core task objective node corresponding to the emergency voice command. The core task objective node represents the final task objective that the user expects to achieve through the emergency voice command. The system uses the core task objective node as the endpoint and performs a reverse path search in the inspection intent knowledge graph. The reverse path search refers to: starting from the core task objective node, searching downwards along the reverse direction of the achievement relationships, through substitution relationships and dependency relationships, for all action path combinations that can be derived from the core task objective node, as a set of candidate backup action paths. The system filters candidate backup action paths based on their dependency on failed hardware. Specifically, it iterates through each action node in the candidate backup action path combinations, extracts the type and quantity of hardware resources required for each action node during execution, and compares the extracted hardware resource requirements with the robot's failed hardware. Any backup action path combination containing action nodes that have a positive dependency on failed hardware is eliminated, retaining only those with zero dependency on failed hardware. Preferably, when multiple backup action path combinations are retained, the system sorts them according to their total resource consumption and selects the backup action path combination with the lowest total resource consumption as the preferred backup action path combination.The total resource consumption can be calculated using a weighted summation method, which assigns different weights to the consumption of each capability dimension and then sums them up.

[0032] After obtaining alternative action path combinations, the system performs a secondary comparison and verification. This secondary verification involves expanding each action node in the alternative action path combination into a sequence of atomic actions arranged in a time series. An atomic action sequence refers to the smallest indivisible execution unit sequence, such as "turn left 30 degrees," "move forward 2 meters," "tilt the gimbal 15 degrees," or "activate infrared shooting." The system extracts the resource requirement values ​​for each atomic action across each capability dimension and superimposes these values ​​in time series to generate a degraded requirement space matrix. This matrix represents the peak resource requirement for each dimension of the alternative action path combination over the entire time series. The system then performs a dimension-by-dimensional numerical comparison between the degraded requirement space matrix and the residual capability space matrix. The secondary verification is considered successful only if the requirement value for each dimension in the degraded requirement space matrix does not exceed the capability value for the corresponding dimension in the residual capability space matrix. If any dimension exceeds the requirement value, the secondary verification is considered unsuccessful. Preferably, if the secondary comparison verification fails, the system returns to step S103 to re-execute the reverse path search, excluding the current backup action path combination, selecting the second-best backup action path combination for secondary comparison verification, until a backup action path combination that passes the verification is found, or if no combination that passes the verification is found after traversing all candidate backup action path combinations, the system broadcasts an alarm message to the user that execution is not possible. If the secondary comparison verification passes, the system converts the backup action path combination into standardized control instructions that can be executed by the robot's underlying actuator. Specifically, the system first fills the variable parameter bits of each atomic action in the backup action path combination with semantic slot information to generate a parameterized atomic action instruction sequence. For example, if the atomic action is gimbal shooting, its variable parameter bit is "target device ID", then the system fills "No. 1 high-voltage cabinet" in the semantic slot into this parameter bit, generating "gimbal shooting (target = No. 1 high-voltage cabinet)". The system compiles the parameterized atomic action instruction sequence into a standardized control instruction sequence that conforms to the robot's underlying actuator instruction format, as a degraded replacement instruction. The system sends the degraded replacement instruction to the robot's underlying actuator for execution. The underlying actuators include, but are not limited to, joint motor drivers, gimbal controllers, image acquisition triggers, and mobile chassis controllers. After receiving and executing the degradation replacement command, the robot completes emergency inspection operations even with partial functional failure. Preferably, before issuing the degradation replacement command, the system also performs a safety boundary check on the command parameters of each atomic action in the degradation replacement command to verify whether the command parameters of each atomic action are within the safe execution range of the underlying actuator under the constraints of the residual capability space matrix. If they exceed the safe execution range, the system automatically reduces the command parameters of the corresponding atomic action to the upper limit of the safe execution range and provides feedback to the user via voice broadcast of the adjusted command parameters and the reason for the adjustment, achieving transparent human-machine interaction.

[0033] In some embodiments, converting the robot's currently available hardware performance into a residual capability space matrix includes: A set of capability dimensions is preset for the robot in its full-function state. The set of capability dimensions includes mobility capability dimension, perception capability dimension, computing power dimension, and overall stability dimension. In the robot's full-function state, rated capability baseline values ​​are set for each capability dimension to form an initial capability space matrix. In the initial capability space matrix, each sub-parameter of each capability dimension is configured with a corresponding rated value range. When any parameter in the underlying hardware status data is detected to exceed the corresponding preset threshold, each sub-parameter of each capability dimension is updated one by one according to the faulty hardware type and fault severity. The sub-parameters affected by the fault are lowered to the actual achievable value, while the sub-parameters not affected by the fault are kept unchanged or appropriately lowered according to preset rules to match the overall machine safety constraints, thereby generating the residual capability space matrix. The residual capability space matrix replaces the initial capability space matrix, and the underlying hardware status data is continuously monitored during the operation of the robot. In response to any secondary change in the hardware status, the capability space matrix recalculation process is repeatedly triggered.

[0034] The capability dimension set includes at least the following four main dimensions: mobility capability, perception capability, computing power, and overall stability. Each dimension is independent yet collectively constitutes a complete description of the robot's overall capabilities. The mobility capability dimension characterizes the robot's ability to perform spatial displacement-related operations. It includes one or more of the following sub-parameters: maximum translational speed, maximum obstacle-crossing height, maximum climbing angle, range of steerable angles, executable gait types (e.g., diagonal trotting, crawling gait, standing posture), and the upper limit of the duration of each gait type. The perception capability dimension characterizes the robot's ability to acquire environmental and equipment information through various sensors. It includes one or more of the following sub-parameters: available sensor types (e.g., LiDAR, depth camera, infrared thermal imager, visible light camera, microphone, etc.), the effective detection range of each available sensor, the detection accuracy or resolution of each available sensor, and the current signal validity status of each available sensor. The computing power dimension characterizes the computing resources provided by the computing platform onboard the robot. The computing power dimension includes one or more of the following sub-parameters: percentage or absolute value of available CPU computing power (e.g., the ratio of remaining computing power to rated computing power), percentage of available GPU computing power, available memory capacity (e.g., the ratio of remaining available memory to total memory), and available storage space (e.g., the ratio of remaining storage capacity to total storage capacity). The overall stability dimension characterizes the robot's ability to maintain overall balance and posture safety. It includes one or more of the following sub-parameters: adjustable center of gravity range (e.g., the offset margin of the center of gravity within the supporting polygon), posture maintenance margin (e.g., the margin of the current pitch and roll angles relative to the overturning boundary), and support phase width (e.g., the area of ​​the supporting triangle / polygon for a legged robot). When the robot is in a fully functional and healthy state, the system sets rated capability baseline values ​​for each sub-parameter of each capability dimension. The rated capability baseline values ​​characterize the robot's standard performance indicators when no hardware failures occur, such as a maximum translational speed of 1.2 m / s, a maximum obstacle crossing height of 0.2 m, 100% available CPU computing power, and an adjustable center of gravity range of ±0.05 m. The system organizes the rated capability baseline values ​​of all the above capability dimensions according to a predetermined format to form an initial capability space matrix. In the initial capability space matrix, each sub-parameter of each capability dimension is configured with a corresponding rated value range. When the real-time measured value of a sub-parameter is within its corresponding rated value range, the system determines that the sub-parameter is in a normal state; when the real-time measured value exceeds the rated value range, the system determines that the sub-parameter is abnormal.

[0035] Throughout the robot's inspection operation, the system continuously collects underlying hardware status data and compares the real-time values ​​of each parameter with the corresponding rated value ranges in the initial capability space matrix. Specifically, the underlying hardware status data comes from sensor feedback from various hardware subsystems of the robot, including but not limited to: joint angles and angular velocities read by the joint motor encoders, current and torque feedback values ​​from the joint motor drivers, attitude angles and angular accelerations output by the inertial measurement unit, pressure measurements from the foot force sensors, online status and signal quality indicators of each sensing device, real-time CPU and GPU utilization, and real-time memory and storage usage. The system presets corresponding failure thresholds for each sub-parameter. Failure thresholds can be the upper or lower boundary of the rated value range, or safety warning values ​​further extended from the rated value range. For example, if the rated operating current range of a certain joint motor is 0.5A to 2.0A, its failure threshold can be set to 2.5A (overcurrent protection threshold) and 0.3A (undercurrent / open circuit detection threshold). The system compares the real-time measured value of each sub-parameter with its corresponding failure threshold in real time. When any parameter is detected to exceed the corresponding preset threshold, the system determines that a local hardware failure event has occurred and triggers the subsequent capability matrix recalculation process. Preferably, to prevent false triggering due to transient noise or occasional interference, the system adopts a "continuous sampling window determination" strategy: only when the same parameter exceeds the threshold in multiple consecutive sampling periods is it determined to be a valid failure event.

[0036] When a localized hardware failure event is triggered, the system initiates a capability matrix recalculation process. The system identifies the type of faulty hardware in the failure event. Faulty hardware types include at least: joint motor failures, sensor failures, communication interface failures, power supply failures, and computing resource overload failures. The system determines the type of faulty hardware based on the specific parameter exceeding the threshold and the hardware subsystem to which that parameter belongs. The system quantifies the severity of the failure. Fault severity is quantified by the deviation of real-time measured values ​​from the corresponding failure threshold. Preferably, the deviation is divided into three levels: slight deviation (0% to 20% exceeding the threshold), moderate deviation (20% to 50% exceeding the threshold), and severe deviation (more than 50% exceeding the threshold), with different levels corresponding to different parameter reduction amounts. Based on the faulty hardware type and severity, the system updates the parameters of each sub-item in each capability dimension of the capability dimension set one by one. For sub-item parameters affected by the failure, the system reduces their value from the rated capability baseline to the currently achievable value. The actual achievable value is determined as follows: Based on the fault type and severity, a preset "fault-capability mapping table" is consulted to obtain the corresponding capability reduction coefficient. The rated capability baseline value is then multiplied by (1 - capability reduction coefficient) to obtain the actual achievable value. For example, when the left front leg joint motor torque is abnormal, the "maximum obstacle crossing height" sub-parameter under the mobility dimension is reduced from 0.2m to 0m, and the "maximum lateral speed" is reduced from 1.2m / s to 0.2m / s. For sub-parameters unaffected by the fault, the system determines its handling method according to preset rules: one method is to maintain them unchanged, i.e., keep their rated capability baseline value unchanged to preserve the usable functions of healthy hardware to the greatest extent; the other method is to appropriately reduce them according to preset rules to match the overall machine safety constraints. Overall robot safety constraints refer to the following: when a piece of robot hardware fails, even if other hardware functions perfectly, to ensure overall dynamic balance and system stability, the performance output of some healthy hardware must be appropriately limited. For example, if one leg fails, even if the joint motors of the other three legs are perfectly functioning, the overall movement speed limit must be appropriately lowered to ensure center of gravity stability. Preferably, sub-parameters unaffected by the failure adopt a "maintain unchanged" strategy, triggering "appropriate reduction" only when it is truly necessary to ensure overall robot safety. Through the above processing, the system generates a multi-dimensional residual capability space matrix. This matrix has the same dimensional structure and sub-parameter types as the initial capability space matrix; the difference lies in that the values ​​of each sub-parameter have been updated from the rated capability baseline value to the actual achievable value under the current failure state.

[0037] After generating the residual capability space matrix, the system replaces the initial capability space matrix with the residual capability space matrix, using it as the capability judgment benchmark for all subsequent fault-tolerant scheduling decisions (including but not limited to task requirement comparison, intent downgrading and reconstruction, secondary verification, etc.). Simultaneously, the real-time acquisition and monitoring process of underlying hardware status data continues uninterrupted, not stopping due to the recalculation of the capability matrix. The system continuously compares the real-time acquired parameters with their corresponding failure thresholds.

[0038] In some embodiments, extracting the target action from the emergency voice command and mapping the target action to a task requirement space matrix specifically includes: The voice signal of the emergency voice command is subjected to noise reduction and speech recognition processing to generate command text. The command text is then filled with semantic slots. The action type slot and the target object slot of the target action are extracted. The position attribute slot and detection type slot of the target object are also extracted. Based on the action type slot, query the preset inspection action-resource requirement mapping table to obtain the requirement level of the target action in each capability dimension; The requirement level of the target action in each capability dimension is converted into the corresponding requirement quantification value, and a task requirement space matrix with the same dimensional structure as the residual capability space matrix is ​​generated. The target slot, the location attribute slot, and the detection type slot are used as parameter labels and bound to the task requirement space matrix for caching.

[0039] The system collects the voice signals of emergency voice commands input by maintenance personnel through a microphone array mounted on a robot. Preferably, the microphone array employs multi-microphone beamforming technology to suppress industrial noise in the power plant environment. The system performs front-end noise reduction processing on the collected voice signals. Preferably, the front-end noise reduction uses an adaptive filtering algorithm or a speech enhancement model based on a deep neural network to filter out environmental noise and reverberation, extracting clear voice signals. The noise-reduced voice signals are then processed by speech recognition to convert them into corresponding command text. The system inputs the command text into a preset semantic parsing model and performs semantic slot filling. Semantic slot filling refers to extracting the text fragments corresponding to each semantic slot from the command text as slot values ​​based on a predefined semantic slot template. Specifically, the system extracts at least the following semantic slots: Action type slot, used to characterize the type of target action that the user expects the robot to perform, such as "reading a meter", "measuring temperature", "inspecting", "taking a picture", etc.; Target object slot, used to characterize the specific equipment or target to which the target action is directed, such as "No. 1 high-voltage switchgear", "No. 3 transformer", "circuit breaker phase A", etc.; Location attribute slot, used to characterize the spatial location information of the target object in the power station, such as "the third column of the GIS room", "outdoor area A", "cable layer B2", etc.; Detection type slot, used to characterize the detection method or sensing method used for the target action, such as "infrared temperature measurement", "visible light photography", "ultrasonic partial discharge detection", etc.

[0040] The system extracts the target action from the action type slot and uses it as a query index, inputting it into a pre-defined inspection action-resource requirement mapping table. This table pre-contains quantitative correspondences between various standard actions and resource requirements in power grid inspection scenarios. For example, the "meter reading" action corresponds to high mobility requirement, high perception requirement, and medium computing power requirement; the "full-area inspection" action corresponds to high mobility requirement, medium perception requirement, and low computing power requirement; and the "fixed-point photography" action corresponds to low mobility requirement, high perception requirement, and low computing power requirement. The system queries the inspection action-resource requirement mapping table based on the target action to obtain the requirement level of the target action in each capability dimension of the capability dimension set. The requirement levels include at least four levels: "high requirement," "medium requirement," "low requirement," and "zero requirement," used to characterize the degree of dependence on the corresponding capability dimension to complete the action.

[0041] The system converts the demand level of the target action in each capability dimension into corresponding quantitative demand values. Preferably, the quantitative demand values ​​and the residual capability space matrix use the same dimensions and normalization interval, for example, both are normalized to the interval [0, 1], where "zero demand" corresponds to a quantitative value of 0, "low demand" corresponds to a quantitative value of 0.3, "medium demand" corresponds to a quantitative value of 0.6, and "high demand" corresponds to a quantitative value of 1.0. Through a unified quantification standard, the system ensures that the quantitative demand values ​​are mathematically comparable to the capability values ​​in the subsequent residual capability space matrix. The system organizes the quantitative demand values ​​of each dimension according to the dimensional order of the capability dimension set to generate a task demand space matrix. The task demand space matrix and the residual capability space matrix have the same dimensional structure and dimensional order, thus allowing them to be compared and enveloped dimension by dimension in the same mathematical coordinate system.

[0042] The system associates and binds the slot values ​​of the target object slot, location attribute slot, and detection type slot as parameter tags with the task requirement space matrix, and stores the bound data structure in the cache. These parameter tags serve as the source of instruction parameters for filling the variable parameter bits of each atomic action in the subsequent degradation and replacement instruction generation stage. For example, when the backup action path combination includes the atomic action "pan-tilt-zoom (PTZ) shooting," the system can directly call the parameter values ​​of the "target object slot" and "detection type slot" from the cache to instantiate the atomic action as "perform infrared shooting of high-voltage cabinet No. 1," rather than the generalized "pan-tilt-zoom (PTZ) shooting." Through this parameter tag binding and caching mechanism, the system ensures that the precise semantic information in the user's original voice command is not lost or attenuated throughout the entire process of intent degradation, replacement search, and instruction generation, ensuring that the final degradation and replacement instruction accurately targets the user-specified device, rather than a generalized replacement operation.

[0043] In some embodiments, the step of performing a secondary comparison and verification between the alternative action path combination and the residual capability space matrix, and generating a downgrade replacement instruction from the alternative action path combination that passes the secondary comparison and verification, specifically includes: Each action node in the backup action path combination is expanded into an atomic action sequence arranged in time sequence. The resource requirement value of each atomic action in each capability dimension is extracted. The resource requirement values ​​of each atomic action are superimposed in time sequence to generate a degradation requirement space matrix. The degradation demand space matrix and the residual capacity space matrix are compared dimension by dimension. When the demand value of each dimension in the degradation demand space matrix does not exceed the capacity value of the corresponding dimension in the residual capacity space matrix, the secondary comparison verification is deemed to be successful. If the secondary comparison and verification passes, the parameter label is filled into the variable parameter bit of each atomic action in the backup action path combination, and the parameterized atomic action sequence is compiled into a standardized control instruction sequence that conforms to the robot's underlying actuator instruction format, as the downgraded replacement instruction; Verify whether the instruction parameters of each atomic action in the downgrade replacement instruction are within the safe execution range of the underlying actuator under the constraint of the residual capability space matrix. If they exceed the safe execution range, reduce the instruction parameters of the corresponding atomic action to the upper limit of the safe execution range and generate corresponding voice feedback information to broadcast to the user.

[0044] The system takes the alternative action path combinations obtained from the search as input. These alternative action path combinations are directed paths generated by reverse searching the inspection intent knowledge graph, consisting of multiple action nodes arranged in a temporal sequence. Examples include gimbal rotation → telephoto zoom → infrared capture → image storage. The system expands each action node in the alternative action path combination one by one, forming a sequence of atomic actions arranged in a time sequence. An atomic action is the smallest granularity action unit that the robot's underlying actuator can directly execute; it cannot be further subdivided, such as "turn left 30 degrees," "move forward 2 meters," "tilt the gimbal up 15 degrees," and "activate infrared capture." Any inspection action can be broken down into multiple atomic actions arranged in a specific temporal sequence. The system extracts the resource requirements for each atomic action across various capability dimensions. The resource requirements are extracted by querying a pre-set atomic action resource requirement library, which pre-defines the resource consumption of various atomic actions in terms of mobility, perception, computing power, and overall machine stability during execution. For example, the atomic action of "moving forward 2 meters" has a resource requirement of 0.3 for mobility, 0.1 for computing power, 0 for perception, and 0.2 for stability. The system superimposes the resource requirements of each atomic action in time sequence to generate a degraded demand space matrix. The superposition method is as follows: for each capability dimension, calculate the cumulative resource consumption of that dimension at any time during the entire execution of the complete atomic action sequence, and take the maximum value in the entire time sequence as the demand value of that dimension. Through the above processing, the degraded demand space matrix accurately represents the peak resource requirements of each dimension for executing the alternative action path combination in the entire time sequence.

[0045] The system performs a dimension-by-dimensional numerical comparison between the downgraded demand space matrix and the residual capability space matrix. Specifically, for each capability dimension, it compares whether the peak demand value of that dimension in the downgraded demand space matrix does not exceed the capability value of that dimension in the residual capability space matrix. When the demand value of each dimension in the downgraded demand space matrix does not exceed the capability value of the corresponding dimension in the residual capability space matrix, the system determines that the secondary comparison verification passes, indicating that the alternative action path combination can be safely executed under the physical constraints of the robot's current residual capability. If the demand value of any dimension exceeds the capability value of the corresponding dimension, the secondary comparison verification fails, indicating that the alternative action path combination exceeds the robot's current physical capability boundary, and the system needs to select another alternative action path combination for verification.

[0046] If the secondary comparison and verification pass, the system fills the variable parameter bits of each atomic action in the backup action path combination with the parameter tags stored in the cache. Variable parameter bits refer to parameters in the atomic action instruction that need to be dynamically filled according to the specific task objective. For example, the variable parameter bits for "pan-tilt rotation" are the spatial coordinates or horizontal / tilt rotation angle of the target device; the variable parameter bits for "start shooting" are the shooting mode and resolution settings. The specific values ​​of the parameter tags are: the target device number is provided in the target object slot (e.g., "High-voltage cabinet No. 1"), the location attribute slot provides the area where the target device is located (e.g., "third column of GIS room"), and the detection type slot provides the detection method (e.g., "infrared temperature measurement"). Based on the atomic action type, the system selects the corresponding slot value from the parameter tags and fills it into the variable parameter bits, instantiating the generalized atomic action into a specific operation for a specific target. For example, "pan-tilt rotation (variable parameter bit = target device)" is instantiated as "pan-tilt rotation (target = high-voltage cabinet No. 1, rotation angle = preset bit 3)". The system arranges the parameterized atomic motion sequences in chronological order and compiles them sequentially into standardized control instruction sequences that conform to the robot's underlying actuator instruction format, serving as the final downgraded replacement instructions. During compilation, the system encapsulates each atomic motion into a corresponding instruction frame format according to the communication protocol of the underlying actuator.

[0047] Before issuing the downgrade replacement command to the underlying actuator, the system performs a safety boundary check on the command parameters of each atomic action in the downgrade replacement command. This check verifies whether the command parameters of each atomic action are within the safe execution range of the underlying actuator under the constraints of the residual capability space matrix. The safe execution range of different atomic actions is determined by the capability values ​​of each dimension in the residual capability space matrix. For example, the safe upper limit of movement speed is determined by the value of the "limit movement parameter" in the residual capability space matrix, and the safe upper limit of gimbal rotation angle is determined by the attitude maintenance margin of the current overall stability dimension. If the command parameters of a certain atomic action exceed the corresponding safe execution range, the system does not directly reject or interrupt execution, but actively reduces the command parameters of the corresponding atomic action to the upper limit of the safe execution range. This pruning operation passively adapts the command parameters with the safety boundary as a hard constraint, while ensuring that the downgrade replacement command can be executed smoothly. At the same time, the system generates corresponding voice feedback information to broadcast to the user. The broadcast content includes at least: the name of the pruned atomic action, the original parameter value, the pruned parameter value, and the reason for pruning. For example, it might broadcast a message such as, "The gimbal rotation angle has been reduced from 30 degrees to 20 degrees, and the current attitude stability margin is insufficient." Through this feedback mechanism, transparent human-machine interaction is achieved, ensuring that maintenance personnel are aware in real time of the robot's instruction adjustments due to safety constraints. The system sends downgraded replacement instructions, after safety boundary verification and active trimming, to the robot's underlying actuators for execution, completing emergency inspection operations under partial functional failure conditions.

[0048] Figure 2 This is a flowchart illustrating the process of searching for alternative action path combinations in an embodiment of this application, as shown below. Figure 2 As shown, the step of extracting the core operational objective node corresponding to the emergency voice command based on the pre-constructed inspection intent knowledge graph, and searching in the inspection intent knowledge graph for alternative action path combinations that can achieve the core operational objective node and have zero dependence on failed hardware, specifically includes the following steps: S201. Obtain the target slot of the emergency voice command, input the target action and the target slot into the inspection intent knowledge graph, start from the starting node of the inspection intent knowledge graph, trace along the directed edge to the first intent node corresponding to the target action, the inspection intent knowledge graph is composed of intent nodes, action nodes, device nodes and detection method nodes, and the directed edges between nodes represent the achievement, dependency or substitution relationship; S202. Starting from the first intent node, continue tracing along the relationship in the inspection intent knowledge graph until the root intent node is reached. The root intent node is taken as the core operation objective node, and the core operation objective node represents the final operation objective corresponding to the emergency voice command. S203. Taking the core task objective node as the endpoint, perform a reverse path search in the inspection intent knowledge graph to retrieve all action path combinations that can be derived from the core task objective node and formed through substitution and dependency relationships, as a set of candidate backup action paths. S204. Traverse each action node in the candidate backup action path set, extract the resource requirements of each action node for the failed dimension in the residual capability space matrix, retain the backup action path combination with zero dependency on failed hardware, if there are multiple retained backup action path combinations, sort them according to the total resource consumption of each backup action path combination, and select the backup action path combination with the lowest total resource consumption as the final backup action path combination.

[0049] The system acquires the target slot of the emergency voice command (e.g., "High-voltage switchgear No. 1"), and inputs the target action (e.g., reading a meter) and the target slot into a pre-constructed inspection intent knowledge graph. The inspection intent knowledge graph is a structured semantic network composed of multiple nodes and directed edges. Specifically, the knowledge graph includes four node types: intent nodes (representing the purpose of the operation, such as obtaining equipment temperature or confirming the appearance of equipment), action nodes (representing specific operational behaviors, such as moving forward, taking a picture, or measuring temperature), equipment nodes (representing electrical equipment entities, such as high-voltage switchgear, transformers, and circuit breakers), and detection method nodes (representing detection methods, such as infrared thermometry, visible light photography, and ultrasonic detection). Directed edges between nodes represent logical relationships, including at least three types: achievement relationships (meaning one node can achieve or accomplish another node, such as the action node "shoot" achieving the intent node "obtain meter reading"), dependency relationships (meaning one node needs another node as a prerequisite before execution, such as the action node "shoot" depending on the action node "gimbal rotation into position"), and substitution relationships (meaning one node can be substituted by another node in an equivalent way, such as the action node "telephoto shooting" substituting for the action node "approach shooting"). The system starts with the action node corresponding to the target action and traces upwards along the directed edges. The tracing path follows these rules: tracing back along the achievement relationship to the direct intent served by the action, and tracing back along the dependency relationship to the indirect intent associated with the prerequisite. The system continues tracing upwards until it reaches one or more intent nodes directly associated with the target action, and designates that node as the first intent node. The first intent node represents the first-level operational purpose directly served by the target action.

[0050] Starting from the first intent node, the system traces upwards along the directed edges of the achievement relationships in the inspection intent knowledge graph. The direction of the achievement relationship is from the action node to the intent node (i.e., the action "achieves" the intent), so tracing upwards is traversing in the reverse direction of the directed edges of the achievement relationships. The system continues to trace in this direction until it reaches the root intent node, which is no longer achieved by any other intent node. The root intent node is the top-level task objective node in the knowledge graph, and typical examples include "obtain target device operating status data," "confirm the target device's appearance integrity," and "detect abnormal overheating of the target device." The root intent node represents the final task objective corresponding to the emergency voice command, that is, the ultimate goal that the user essentially wants to achieve through the voice command. The system uses the root intent node reached through tracing as the core task objective node and as the target endpoint for subsequent reverse path searches. Preferably, if a branch is encountered when tracing upwards along the relationship (i.e., an action can be achieved by multiple upper-level intentions), the system includes the root intention nodes on all branches into the core task objective candidate set, and selects the root intention node with the highest level as the core task objective node according to the preset priority rules, while the remaining root intention nodes are cached as alternative intention nodes for later use.

[0051] The system uses the core task objective node as the endpoint (target node) and performs a reverse path search in the inspection intent knowledge graph. Reverse path search refers to traversing the graph backwards along directed edges from the core task objective node, retrieving all action path combinations that can be derived downwards from the core task objective node. The search rules are as follows: starting from the core task objective node, proceeding downwards through the positive direction of achieving relationships (from intent node to action node), or through substitution relationships (from one action node to another equivalent action node), or through the positive direction of dependency relationships (from prerequisite action node to subsequent action node), expanding layer by layer until the end of the action node sequence is reached. Through this search, the system retrieves all action path combinations that can achieve the core task objective node and uses these action path combinations as a candidate set of backup action paths. Each action path combination in the candidate set is a complete action flow that is physically executable, arranged chronologically by action nodes. Preferably, the reverse path search uses a depth-first search algorithm or a breadth-first search algorithm to traverse the knowledge graph, setting a maximum search depth to prevent search space explosion.

[0052] The system iterates through each action node in the candidate backup action path set, extracting the type and quantity of hardware resources required for each action node during execution. This extraction is done by querying an atomic action resource library, which pre-defines the type and quantity of hardware resources required for each action node. The system compares the hardware resource requirements of each action node with the robot's current list of failed hardware. If any action node in a backup action path combination has a hardware resource type that matches an item in the list of failed hardware and a quantity greater than zero, then this backup action path combination has a positive dependency on failed hardware, and the system removes it from the candidate set. Through this filtering process, the system retains only backup action path combinations where each action node has a zero requirement for failed hardware resources—that is, backup action path combinations with zero dependency on failed hardware. If multiple backup action path combinations remain after this filtering, the system sorts them in ascending order based on their total resource consumption. The total resource consumption is calculated by weighted summing of the resource consumption values ​​of each action node in the backup action path combination across all capability dimensions. Preferably, the weights of each capability dimension are dynamically adjusted based on the remaining margin of each dimension in the current residual capability space matrix. Dimensions with lower remaining margins are assigned higher weights to prioritize the alternative path that consumes the least scarce resources. The system selects the backup action path combination with the lowest overall resource consumption as the final backup action path combination and outputs it to step S104 for secondary comparison and verification. If all candidate backup action path combinations depend on the failed hardware, i.e., the candidate set is empty after screening, the system broadcasts an alarm message "No executable alternative" to the user and outputs the specific dependencies of each candidate path on the failed hardware for maintenance personnel to refer to.

[0053] In some embodiments, the step of tracing from the starting node of the inspection intent knowledge graph along directed edges to the first intent node corresponding to the target action specifically includes: Starting from the action node corresponding to the target action, identify the precondition action chain required to complete the target action along the directed edge of the dependency relationship, and map each action node in the precondition action chain to the corresponding action intent node in the inspection intent knowledge graph to form a first intent node set. For each action intent node in the first intent node set, perform pairwise correlation determination. If the first action intent node and the second action intent node share the same core operation objective node and the shortest path distance in the inspection intent knowledge graph is less than a preset distance threshold, then determine that there is an intent coupling relationship between the first action intent node and the second action intent node, and merge multiple intent nodes with the intent coupling relationship into a composite intent node. The composite intent node replaces the multiple intent nodes to perform subsequent tracing operations. After tracing back to the core task destination node, the combined backup action path combination covering all sub-intents in the composite intent node is simultaneously retrieved in the reverse path search. If a target atomic action in the combined backup action path combination contributes to multiple sub-intents in the composite intent node, then a weighted bonus is assigned to the target atomic action to increase the selection priority of backup action path combinations that include atomic actions with the weighted bonus.

[0054] The system starts with the action node corresponding to the target action and performs a forward traversal along the directed edges of dependencies in the inspection intent knowledge graph. A dependency relationship indicates that one action requires another action as a precondition before execution. For example, the action node "take a picture" depends on the action node "pan-tilt unit rotates to position," and "pan-tilt unit rotates to position" depends on the action node "obtain target device coordinates." Tracing forward layer by layer along the dependencies, the system identifies all the precondition actions required to complete the target action and arranges these precondition actions in chronological order to form a precondition action chain. The end of the precondition action chain is an atomic action that requires no preconditions, and the beginning is the immediate preceding action that directly serves the target action. The entire action chain completely depicts all intermediate action steps required from the initial state to the execution of the target action. The system maps each action node in the precondition action chain to the corresponding action intent node in the inspection intent knowledge graph. An action intent node is an intent node that represents the operational purpose served by the action. The mapping method is: for each action node, trace upward one layer along the achievement relationship to obtain the intent node directly achieved by the action. Each action node and its corresponding action intent node together constitute the first intent node set. Each intent node in this set represents a sub-intent corresponding to a certain step required to complete the target action. For example, for the target action "reading the table", the precondition action chain may include actions such as "moving to cabinet number 1", "adjusting the pan-tilt angle", and "focusing clearly", whose corresponding intent nodes are "reaching the target position", "aligning with the target device", and "acquiring a clear image", which together constitute the first intent node set.

[0055] The system performs pairwise correlation determination on each action intent node in the first intent node set. The determination rules are as follows: Take the first action intent node and the second action intent node, check whether they share the same core task objective node, and calculate the shortest path distance between them in the inspection intent knowledge graph. The first and second action intent nodes are arbitrary action intent nodes, and the shortest path distance refers to the number of edges contained in the path from the first action intent node to the second action intent node with the fewest directed edges. If the first and second action intent nodes share the same core task objective node, and the shortest path distance in the inspection intent knowledge graph is less than a preset distance threshold, then an intent coupling relationship is determined to exist between the first and second action intent nodes. The typical value of the preset distance threshold is 3 to 5, indicating that the two intent nodes are semantically close enough, serve the same final purpose, and are closely related. Intent coupling indicates that the two sub-intents are intrinsically related in terms of semantics and purpose, and should not be treated as independent intents. When multiple intent nodes with intent coupling relationships are identified, the system merges these intent nodes into a composite intent node. A composite intent node is a special type of intent node structure that contains multiple sub-intents but presents a unified overall task objective to the outside world. In subsequent tracing and search operations, the composite intent node participates in processing as a single node, but internally retains the complete semantic information of each sub-intent.

[0056] The system replaces the original multiple independent intent nodes with a composite intent node and performs subsequent upward tracing operations. Specifically, it continues tracing upwards from the composite intent node along the achievement relationship until it reaches the root intent node, which is then designated as the core task objective node. After the core task objective node is determined, the system performs a reverse path search. However, unlike the reverse path search in a single intent scenario, the reverse path search in this application is constrained by covering all sub-intents in the composite intent node, simultaneously searching for combined backup action path combinations that can simultaneously achieve each sub-intent. A combined backup action path combination refers to a combination structure composed of multiple action paths, where each atomic action sequence can completely cover every sub-intent in the composite intent node. For example, if a composite intent node contains two sub-intents, "acquire temperature data" and "record device appearance," the combined backup action path combination must simultaneously include atomic action sequences that achieve both sub-intents, ensuring that all task objectives of the composite instruction are satisfied, rather than only achieving some sub-intents. Through the above processing, the system ensures that the alternative execution scheme completely covers all implicit task objectives in the user's composite instruction, avoiding task omissions due to semantic fragmentation.

[0057] The system iterates through each atomic action in the combined backup action path combination, identifying whether a single atomic action simultaneously contributes to multiple sub-intents within a composite intent node. Simultaneous contribution to multiple sub-intents within a composite intent node means that the execution result of the atomic action is reused by multiple sub-intents as a common basis or intermediate product for achieving each sub-intent. For example, the atomic action "capture a high-definition image of cabinet 1" simultaneously serves two sub-intents: "obtain meter readings" and "record equipment appearance," because the same high-definition image can be used for both OCR recognition to extract readings and for archiving appearance status. If such a shared atomic action is detected, the system assigns a weighted bonus to that atomic action. This weighted bonus is a positive incentive factor that plays a role in the subsequent overall resource consumption ranking. Specifically, when calculating the overall resource consumption of the backup action path combination, the system subtracts a preset bonus discount value from path combinations containing weighted atomic actions, ensuring that the final ranking score of this path combination is better than other path combinations without shared atomic actions but with similar resource consumption. Through the aforementioned weighting mechanism, the system prioritizes combinations of joint backup action paths that can simultaneously serve multiple sub-intents with a single atomic action. This mechanism fully leverages the synergistic effect of action reuse, minimizing the total execution time and overall resource consumption of alternative solutions while satisfying all sub-intents, thus achieving global optimization of alternative execution efficiency in multi-intent composite scenarios.

[0058] In some embodiments, the step of tracing back along the established relationships in the inspection intent knowledge graph, starting from the first intent node, until the root intent node is reached, and using the root intent node as the core task objective node, specifically includes: If multiple root intent nodes are detected in the inspection intent knowledge graph that can be derived to achieve the first intent node, then all root intent nodes that can achieve the first intent node are included in the core objective candidate set. The root intent node with the highest level or the highest confidence level is selected from the core objective candidate set as the core task objective node, and the remaining root intent nodes that are not selected in the core objective candidate set are recorded as second intent nodes. When performing the reverse path search, a first backup action path combination that can reach the core operation target node and has zero dependence on the failed hardware is searched. If there is no complete action path combination in the inspection intent knowledge graph that can reach the core operation target node and has zero dependence on the failed hardware, the second intent node is promoted to the current core operation target node, and the reverse path search is re-executed for the promoted current core operation target node.

[0059] The system starts from the first intent node and traces upwards along the achievement relationships within the inspection intent knowledge graph. When a branch appears during the tracing process, the system checks if multiple root intent nodes can be derived to achieve the first intent node. A root intent node is the top-level intent node that cannot be achieved by any other intent node. Typical examples include "acquire target device operating status data," "confirm the target device's appearance integrity," and "detect abnormal overheating of the target device." Different root intent nodes correspond to different final work objectives. If multiple root intent nodes are detected that can be derived to achieve the first intent node, the system includes all root intent nodes that can achieve the first intent node in the core objective candidate set. This candidate set fully records all possible high-level work objectives served by the first intent node, providing a complete alternative space for subsequent main intent selection and intent rollback.

[0060] The system selects a root intent node from the core objective candidate set as the core task objective node. The selection rule is based on at least one of the following two strategies: The first strategy is the highest-level priority strategy. The system calculates the level depth of each root intent node in the core objective candidate set within the inspection intent knowledge graph. Level depth refers to the number of edges contained in the longest path from the root intent node down to the leaf node (the terminal action node without further branches). The higher the level of the root intent node, the more abstract its representation and the closer it is to the final task objective, thus covering a wider range of task scenarios when used as the core task objective node. The second strategy is the highest-confidence priority strategy. The system pre-configures a confidence score for each root intent node, which reflects the prior probability that the root intent node is the user's actual task objective in the power inspection scenario. For example, the confidence score for "acquiring equipment operating status data" is usually higher than that for "recording equipment artistic images" because the essential purpose of inspection work is status detection rather than image recording. The system selects the root intent node from the core objective candidate set as the core task objective node according to the above strategy, and records the remaining unselected root intent nodes in the core objective candidate set as second intent nodes. The second intent node serves as an alternative job objective and is invoked in the subsequent intent rollback mechanism.

[0061] The system uses the selected core task objective node as the target endpoint and performs a reverse path search to find a first backup action path combination that can achieve the core task objective node and has zero dependence on failed hardware. If the search is successful, meaning at least one complete action path combination that can achieve the core task objective node without relying on any failed hardware is found, the system outputs this path combination to step S104 for secondary comparison and verification. If the search fails, meaning there is no complete action path combination in the inspection intent knowledge graph that can achieve the current core task objective node and has zero dependence on failed hardware, the system triggers the intent rollback mechanism. The execution process of the intent rollback mechanism is as follows: the system promotes the marked second intent node from the candidate state to the current core task objective node, replaces the original core task objective node with the promoted current core task objective node, and re-executes the reverse path search for this node. Preferably, when there are multiple second intent nodes, the system promotes each second intent node in order of priority. The priority order is based on the confidence score of each second intent node in the original core objective candidate set from high to low, or based on the shortest path distance between each second intent node and the first intent node in the graph from near to far. Preferably, if the current second intent node still cannot find a complete action path combination with zero dependency on the failed hardware after being promoted, the system continues to promote the next root intent node in the core objective candidate set in order of priority until a feasible alternative action path combination is found. If no feasible alternative action path combination is found after traversing all root intent nodes in the core objective candidate set, the system terminates the search, broadcasts an alarm message to the user that no effective work objective can be achieved under the current fault state, and outputs a summary of the specific reasons why each root intent node in the core objective candidate set cannot be executed for the reference of the operation and maintenance personnel. The essence of the above-mentioned progressive rollback mechanism at the intent layer is that when the robot is completely unable to achieve the optimal work objective corresponding to the user's instruction under the current hardware failure state, the system does not give up directly, but automatically reduces to the second-best work objective and tries again. Through step-by-step rollback, even under extreme fault states, it still outputs work results with actual operation and maintenance value as much as possible, and minimizes the possibility of the entire emergency dispatch process being idle or scrapped due to the unachievability of a single intent.

[0062] The semantic-based fault-tolerant scheduling and emergency operation method for robots in the embodiments of this application has been described above. The computer system in the embodiments of this application will be described in detail below in conjunction with the above semantic-based fault-tolerant scheduling and emergency operation method for robots.

[0063] Please see Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0064] In some embodiments, the computer system 300 includes a computer device, which may be a terminal device. The computer device includes a processor 301, a memory 302, a sensor module 303, a communication module 304, an input device 305, and an output device 306 connected via a system bus. The processor 301 of the computer device provides computing and control capabilities. The memory 302 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0065] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0066] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the computer system 300, cause the computer system 300 to execute the semantic-based robot fault-tolerant scheduling and emergency operation method of this application.

[0067] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 300, causes the computer system 300 to execute the semantic-based robot fault-tolerant scheduling and emergency operation method of this application.

[0068] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0069] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A semantic-based fault-tolerant scheduling and emergency operation method for robots, characterized in that, Applied to robots, the methods include: Real-time acquisition of the robot's underlying hardware status data; when a local hardware failure occurs, based on a preset capability quantification model, the robot's current available hardware performance is converted into a residual capability space matrix, which represents the robot's current limit movement parameters, available perception dimensions, and remaining computing power. Receive emergency voice commands input by the user, extract the target actions from the emergency voice commands, and map the target actions into a task requirement space matrix; The task requirement space matrix and the residual capability space matrix are compared in a multi-dimensional envelope. If the residual capability space matrix fails to completely enclose the corresponding dimension threshold in the task requirement space matrix, the core operation objective node corresponding to the emergency voice command is extracted according to the pre-constructed inspection intention knowledge graph. Alternative action path combinations that can achieve the core operation objective node and have zero dependence on the failed hardware are searched in the inspection intention knowledge graph. The backup action path combination is compared and verified with the residual capability space matrix for the second time. The backup action path combination that passes the second comparison and verification is used to generate a downgrade replacement instruction, and the downgrade replacement instruction is sent to the robot's underlying actuator for execution, so as to complete the emergency inspection operation in the state of partial functional failure.

2. The method according to claim 1, characterized in that, The step of converting the robot's current available hardware performance into a residual capability space matrix includes: A set of capability dimensions is preset for the robot in its full-function state. The set of capability dimensions includes mobility capability dimension, perception capability dimension, computing power dimension, and overall stability dimension. In the robot's full-function state, rated capability baseline values ​​are set for each capability dimension to form an initial capability space matrix. In the initial capability space matrix, each sub-parameter of each capability dimension is configured with a corresponding rated value range. When any parameter in the underlying hardware status data is detected to exceed the corresponding preset threshold, each sub-parameter of each capability dimension is updated one by one according to the faulty hardware type and fault severity. The sub-parameters affected by the fault are lowered to the actual achievable value, while the sub-parameters not affected by the fault are kept unchanged or appropriately lowered according to preset rules to match the overall machine safety constraints, thereby generating the residual capability space matrix. The residual capability space matrix replaces the initial capability space matrix, and the underlying hardware status data is continuously monitored during the operation of the robot. In response to any secondary change in the hardware status, the capability space matrix recalculation process is repeatedly triggered.

3. The method according to claim 1, characterized in that, The step of extracting the target action from the emergency voice command and mapping the target action into a task requirement space matrix specifically includes: The voice signal of the emergency voice command is subjected to noise reduction and speech recognition processing to generate command text. The command text is then filled with semantic slots. The action type slot and the target object slot of the target action are extracted. The position attribute slot and detection type slot of the target object are also extracted. Based on the action type slot, query the preset inspection action-resource requirement mapping table to obtain the requirement level of the target action in each capability dimension; The requirement level of the target action in each capability dimension is converted into the corresponding requirement quantification value, and a task requirement space matrix with the same dimensional structure as the residual capability space matrix is ​​generated. The target slot, the location attribute slot, and the detection type slot are used as parameter labels and bound to the task requirement space matrix for caching.

4. The method according to claim 3, characterized in that, The step of performing a secondary comparison and verification between the backup action path combination and the residual capability space matrix, and generating a downgrade replacement instruction from the backup action path combination that passes the secondary comparison and verification, specifically includes: Each action node in the backup action path combination is expanded into an atomic action sequence arranged in time sequence. The resource requirement value of each atomic action in each capability dimension is extracted. The resource requirement values ​​of each atomic action are superimposed in time sequence to generate a degradation requirement space matrix. The degradation demand space matrix and the residual capacity space matrix are compared dimension by dimension. When the demand value of each dimension in the degradation demand space matrix does not exceed the capacity value of the corresponding dimension in the residual capacity space matrix, the secondary comparison verification is deemed to be successful. If the secondary comparison and verification passes, the parameter label is filled into the variable parameter bit of each atomic action in the backup action path combination, and the parameterized atomic action sequence is compiled into a standardized control instruction sequence that conforms to the robot's underlying actuator instruction format, as the downgraded replacement instruction; Verify whether the instruction parameters of each atomic action in the downgrade replacement instruction are within the safe execution range of the underlying actuator under the constraint of the residual capability space matrix. If they exceed the safe execution range, reduce the instruction parameters of the corresponding atomic action to the upper limit of the safe execution range and generate corresponding voice feedback information to broadcast to the user.

5. The method according to claim 1, characterized in that, The process of extracting the core operational objective node corresponding to the emergency voice command based on the pre-constructed inspection intent knowledge graph, and searching the inspection intent knowledge graph for alternative action path combinations that can achieve the core operational objective node and have zero dependence on failed hardware, specifically includes: Obtain the target slot of the emergency voice command, input the target action and the target slot into the inspection intent knowledge graph, start from the starting node of the inspection intent knowledge graph, trace along the directed edge to the first intent node corresponding to the target action, the inspection intent knowledge graph is composed of intent nodes, action nodes, device nodes and detection method nodes, and the directed edges between nodes represent the achievement, dependency or substitution relationship; Starting from the first intent node, the source is traced along the relationship in the inspection intent knowledge graph until the root intent node is reached. The root intent node is taken as the core operation objective node, and the core operation objective node represents the final operation objective corresponding to the emergency voice command. Taking the core task objective node as the endpoint, perform a reverse path search in the inspection intent knowledge graph to retrieve all action path combinations that can be derived from the core task objective node and formed through substitution and dependency relationships, as a set of candidate backup action paths. Traverse each action node in the candidate backup action path set, extract the resource requirements of each action node for the failed dimension in the residual capability space matrix, retain the backup action path combination with zero dependency on failed hardware, if there are multiple retained backup action path combinations, sort them according to the total resource consumption of each backup action path combination, and select the backup action path combination with the lowest total resource consumption as the final backup action path combination.

6. The method according to claim 5, characterized in that, The step of tracing from the starting node of the inspection intent knowledge graph along directed edges to the first intent node corresponding to the target action specifically includes: Starting from the action node corresponding to the target action, identify the precondition action chain required to complete the target action along the directed edge of the dependency relationship, and map each action node in the precondition action chain to the corresponding action intent node in the inspection intent knowledge graph to form a first intent node set. For each action intent node in the first intent node set, perform pairwise correlation determination. If the first action intent node and the second action intent node share the same core operation objective node and the shortest path distance in the inspection intent knowledge graph is less than a preset distance threshold, then determine that there is an intent coupling relationship between the first action intent node and the second action intent node, and merge multiple intent nodes with the intent coupling relationship into a composite intent node. The composite intent node replaces the multiple intent nodes to perform subsequent tracing operations. After tracing back to the core task destination node, the combined backup action path combination covering all sub-intents in the composite intent node is simultaneously retrieved in the reverse path search. If a target atomic action in the combined backup action path combination contributes to multiple sub-intents in the composite intent node, then a weighted bonus is assigned to the target atomic action to increase the selection priority of backup action path combinations that include atomic actions with the weighted bonus.

7. The method according to claim 5, characterized in that, Starting from the first intent node, the process of tracing back along the established relationships in the inspection intent knowledge graph until the root intent node is reached, and then using the root intent node as the core task objective node, specifically includes: If multiple root intent nodes are detected in the inspection intent knowledge graph that can be derived to achieve the first intent node, then all root intent nodes that can achieve the first intent node are included in the core objective candidate set. The root intent node with the highest level or the highest confidence level is selected from the core objective candidate set as the core task objective node, and the remaining root intent nodes that are not selected in the core objective candidate set are recorded as second intent nodes. When performing the reverse path search, a first backup action path combination that can reach the core operation target node and has zero dependence on the failed hardware is searched. If there is no complete action path combination in the inspection intent knowledge graph that can reach the core operation target node and has zero dependence on the failed hardware, the second intent node is promoted to the current core operation target node, and the reverse path search is re-executed for the promoted current core operation target node.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.