A multi-stage alarm adaptive control method based on operation instruction analysis rules

Abstract

This invention relates to the field of instruction parsing technology, specifically a multi-level alarm adaptive control method based on operation instruction parsing rules. The method includes constructing a dependency syntax tree using a natural language processing model and extracting instruction semantic triples; mapping the triples to a semantic constraint graph to extract safety numerical boundaries and dynamically correcting them based on real-time environmental conditions; generating an instruction risk index by combining parameter limit exceedance and semantic incompleteness probability, and triggering hierarchical adaptive control based on this index and matching thresholds. This invention breaks the dependence of traditional industrial control on rigid protocols by constructing a multi-dimensional heterogeneous graph and a deep semantic parsing model, enhancing the inclusiveness of non-standard instructions, and using the risk vector magnitude to quantitatively assess instruction risk, thus realizing a multi-level control strategy from intelligent semantic completion to interactive correction guidance.

Description

Technical Field

[0001] This invention relates to the field of instruction parsing technology, specifically to a multi-level alarm adaptive control method based on operation instruction parsing rules. Background Technology

[0002] In the monitoring of switchgear drawer positions, comprehensive electrical protection, and coordinated control of equipment status in low-voltage power distribution systems, the issuance and execution of operation instructions are the core links to ensure the safe and reliable operation of the system. With the development of industrial intelligence, the sources and interaction forms of operation and maintenance instructions are becoming increasingly diversified, including text instructions issued by the upper-level dispatch system, manually filled electronic operation tickets, and human-machine interaction instructions from on-site personnel. Existing industrial control systems typically rely on rigid structured matching protocols and static threshold tables for compliance verification when processing these operational instructions. However, in complex industrial operation and maintenance scenarios, instruction texts often lack structured elements due to human abbreviations, colloquial expressions, or information flow. Faced with such non-standard instructions, existing mechanical parsing methods struggle to understand their deep semantics, easily leading to direct instruction failure or erroneous execution. Furthermore, this rigid processing mechanism often detaches from real-time on-site environmental conditions, relying solely on single static limits for safety verification, resulting in a high probability of misjudgment under extreme or complex conditions. More significantly, existing systems generally employ a one-size-fits-all approach of comprehensive interception and deadlock alarm strategies when faced with instruction parameter deviations or structural incompleteness. Lacking a refined dimensional assessment of potential risks, these systems cannot make safe and reasonable inferences to complete instructions in low-risk situations, nor can they provide precise guidance for operators to correct deviations. This not only leads to low efficiency in human-machine interaction but can also delay optimal processing opportunities during emergency repairs due to forced system shutdowns. To address this, a multi-level alarm adaptive control method based on operation instruction parsing rules is proposed. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-level alarm adaptive control method based on operation instruction parsing rules to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A multi-level alarm adaptive control method based on operation instruction parsing rules includes: Obtain the operation instruction text and construct a dependency syntax tree using a natural language processing model; extract instruction semantic triples containing the operation object, operation behavior, and optional operation parameters based on the dependency syntax tree; The instruction semantic triples are mapped to the corresponding ontology nodes of the semantic constraint graph to extract the safe numerical boundary. When the operation parameter is extracted, the absolute deviation between the operation parameter and the safe numerical boundary is calculated and converted into parameter limit violation through nonlinear mapping. The dependency syntax tree is topologically compared with the built-in standard instruction template to locate missing nodes and generate semantic incompleteness probability. An instruction risk index is generated based on the parameter limit violation and the semantic incompleteness probability. Based on the dynamic matching threshold of the instruction risk index, adaptive control is triggered: when the risk index is lower than the first risk threshold, the instruction is executed normally; when it is in the first interval, a primary warning is triggered and the semantic completion rule is invoked to reconstruct the instruction semantic triple; when it is in the second interval, an intermediate alarm is triggered to intercept the execution and provide syntactic clarification text; when it exceeds the highest threshold, an advanced alarm is triggered to block the operation and abnormal samples are recorded.

[0005] Preferably, the instruction semantic triple extraction process involves identifying the core root node of the dependency syntax tree, verifying the part-of-speech feature of the core root node as a verb, and extracting it as an operation behavior; traversing the path along the tree structure starting from the node corresponding to the operation behavior, and extracting nodes with verb-object dependency relations and object dependency relations as candidate operation objects; traversing the subtree branches of the operation behavior and the candidate operation objects, and extracting nodes with quantifier-modified dependency relations and unit identifiers as candidate operation parameters; calculating the dependency topology step length from each candidate operation object and each candidate operation parameter to the core root node, and selecting the node combination with the shortest dependency topology step length to form the instruction semantic triple.

[0006] Preferably, the safe numerical boundary extraction process involves: matching the names and aligning the attributes of the operation objects in the instruction semantic triples with the ontology nodes in the semantic constraint graph to determine the target ontology node; based on the operation behavior in the instruction semantic triples, retrieving action constraint edges associated with the target ontology node in the semantic constraint graph to determine the attribute dimensions under the current operation behavior; extracting the safe threshold interval corresponding to the attribute dimensions in the target ontology node as the initial boundary; retrieving real-time environment state nodes in the semantic constraint graph that have a topological association with the target ontology node, and dynamically correcting the initial boundary based on the values ​​of the real-time environment state nodes to obtain the safe numerical boundary.

[0007] Preferably, the semantic constraint graph is a multidimensional heterogeneous knowledge graph, including an entity node layer, a relation edge layer, and an attribute constraint layer; the entity node layer includes ontology nodes representing different types of operating devices and real-time environmental state nodes representing the perceived working conditions; the relation edge layer includes topological association edges connecting the influence relationships between the ontology nodes and the real-time environmental state nodes, and action constraint edges connecting the operating behavior and the ontology nodes; the attribute constraint layer includes a feature dictionary configured in each ontology node, the feature dictionary being used to map and store the attribute dimensions and corresponding safety threshold ranges under the action constraint edges.

[0008] Preferably, the parameter deviation is calculated as the absolute value of the difference between the operating parameter and the nearest safe boundary endpoint, and the absolute deviation is generated by nonlinearly mapping the absolute deviation through an exponential function. The semantic incompleteness probability is the ratio obtained by dividing the total number of nodes in the standard instruction template and the number of missing nodes located by topology comparison by the total number of nodes. The parameter exceedance is normalized, and the instruction risk index is constructed by using the parameter exceedance and the semantic incompleteness probability as independent risk dimensions to build a two-dimensional evaluation space. The sum of squares of the parameter exceedance and the semantic incompleteness probability is calculated, and the square root of the sum of squares is performed to obtain the magnitude of the risk vector in the two-dimensional evaluation space as the instruction risk index. When no operation parameters are extracted from the instruction semantic triple, the parameter exceedance is assigned a preset constant value, and the instruction risk index is generated solely based on the semantic incompleteness probability.

[0009] Preferably, the adaptive control includes: when the risk index is less than a first risk threshold, executing the command normally; when the command risk index is in a first interval, i.e., greater than or equal to the first risk threshold and less than a second risk threshold, triggering a primary warning and invoking semantic completion rules to reconstruct the command semantic triplet; when the command risk index is in a second interval, i.e., greater than or equal to the second risk threshold and less than a maximum threshold, triggering a mid-level alarm to intercept execution and providing syntactic clarification text; and when the command risk index is greater than or equal to the maximum threshold, triggering a high-level alarm to block execution and recording abnormal samples.

[0010] Preferably, the semantic completion rule involves: parsing the missing node type located by topological comparison to determine the missing element; obtaining a set of candidate nodes corresponding to the missing element; retrieving historical security execution logs and constructing a Markov state transition matrix based on the operation sequence; calculating the transition probability of the known node in the instruction semantic triplet to each candidate node in the candidate node set using the Markov state transition matrix, and generating the temporal co-occurrence of each candidate node; selecting the candidate node with the highest temporal co-occurrence to fill the missing node and complete the reconstruction; if no candidate node is matched and the missing element is an operation parameter, then the median of the security numerical boundary is extracted as a security benchmark parameter for reconstruction.

[0011] Preferably, the syntactic clarification text is generated by extracting the missing node labels after topological comparison of the dependency syntax tree and the standard instruction template, and mapping the semantic constraint graph to obtain the ontology attribute description corresponding to the missing node; retrieving the attribute dimensions associated with the current operation behavior in the semantic constraint graph, and generating a reference interval containing the safety numerical boundary corresponding to the attribute dimensions; and template-concatenating the ontology attribute description, the reference interval, and the top three candidate nodes with the highest temporal co-occurrence generated by the semantic completion rules to generate guiding text containing missing location hints, completion option guidance, and safety compliance range information as the syntactic clarification text.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a dependency syntax tree of a natural language processing model and deeply extracting instruction semantic triples containing the operation object, operation behavior, and operation parameters, breaks through the limitations of traditional industrial control systems that heavily rely on rigid structured matching protocols. Addressing the lack of non-standard instructions and structured elements in actual industrial operation and maintenance due to human abbreviations, colloquial expressions, or information flow, this method can accurately capture and understand the deep semantics of instructions, effectively avoiding the direct failure or erroneous execution of instructions that is easily caused by traditional mechanical parsing, thus endowing the system with stronger semantic inclusiveness and compatibility of interactive instructions.

[0013] 2. This invention introduces a multidimensional heterogeneous semantic constraint graph, maps instruction semantic triples to ontology nodes in the graph to extract initial safety boundaries, and innovatively combines real-time environmental state nodes with topological associations to dynamically correct these initial boundaries, thereby obtaining safety numerical boundaries that fit the current operating conditions. It completely abandons the rigid mechanism of existing technologies that rely solely on a single static threshold table for compliance verification, detached from the real-time field environment. This allows the safety verification process to be closely linked to the actual dynamic operating conditions on site, effectively eliminating verification errors caused by static limits under extreme or complex operating conditions, and effectively ensuring the safety of low-voltage power distribution systems and equipment collaborative control.

[0014] 3. This invention constructs an instruction risk index by combining the parameter limit deviation converted from the absolute deviation of comprehensive parameters with the semantic incompleteness probability generated by topological comparison. Based on this index, a multi-level alarm adaptive control strategy is implemented, encompassing automatic issuance, semantic completion, clarification guidance, and advanced blocking. This effectively addresses the technical pain points of existing methods that commonly employ a one-size-fits-all approach to interception and deadlock alarms when facing abnormal instructions. Thanks to this dimensional assessment and hierarchical processing of instruction risk, this invention can not only perform safe and reasonable inference and reconstruction based on historical state transition matrices under low-risk conditions, but also provide personnel with precise guiding correction text when deviations occur. This not only eliminates the problem of delaying the best handling time due to forced system deadlock during emergency repairs, but also greatly improves the smoothness of human-computer interaction and the continuity of daily operation and maintenance. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a multi-level alarm adaptive control method based on operation instruction parsing rules; Figure 2 This is a schematic diagram of the logical flow for assessing the instruction risk index of this invention; Figure 3 This is a diagram of the multi-level alarm adaptive control strategy of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a multi-level alarm adaptive control method based on operation instruction parsing rules, the technical solution of which is as follows: A multi-level alarm adaptive control method based on operation instruction parsing rules includes: Obtain the operation instruction text and construct a dependency syntax tree using a natural language processing model; extract instruction semantic triples containing the operation object, operation behavior, and optional operation parameters based on the dependency syntax tree; The instruction semantic triples are mapped to the corresponding ontology nodes of the semantic constraint graph to extract the safe numerical boundary. When the operation parameter is extracted, the absolute deviation between the operation parameter and the safe numerical boundary is calculated and converted into parameter limit violation through nonlinear mapping. The dependency syntax tree is topologically compared with the built-in standard instruction template to locate missing nodes and generate semantic incompleteness probability. An instruction risk index is generated based on the parameter limit violation and the semantic incompleteness probability. Based on the dynamic matching threshold of the instruction risk index, adaptive control is triggered: when the risk index is lower than the first risk threshold, the instruction is executed normally; when it is in the first interval, a primary warning is triggered and the semantic completion rule is invoked to reconstruct the instruction semantic triple; when it is in the second interval, an intermediate alarm is triggered to intercept the execution and provide syntactic clarification text; when it exceeds the highest threshold, an advanced alarm is triggered to block the operation and abnormal samples are recorded.

[0018] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0019] Example 1: Obtain the operation instruction text and construct a dependency syntax tree using a natural language processing model; extract instruction semantic triples containing the operation object, operation behavior, and optional operation parameters based on the dependency syntax tree; Specifically, the system receives initial instruction data input by the operator; performs text cleaning preprocessing on the initial instruction data, including removing invalid consecutive whitespace characters, removing newline control characters, and filtering special punctuation marks within non-standard control instruction systems. The cleaned and standardized pure Chinese string is extracted and saved to the cache key as the obtained operation instruction text. In the specific scenario of this embodiment, after receiving data through the text input box of the interactive interface and completing data cleaning, the specific content of the obtained operation instruction text is "Raise the temperature of reactor number one by thirty degrees Celsius". In this embodiment, the natural language processing model specifically employs a combined deep learning model constructed from a pre-trained bidirectional language encoder network based on a multi-head self-attention mechanism and a dual affine dependency attention network. The overall architecture includes a bottom-level lexical and character feature embedding layer, multiple stacked hidden layers in the middle responsible for extracting deep global context semantic feature vectors, and a top-level layer with two feedforward neural networks and a dual affine tensor multiplication module arranged in parallel. The two feedforward neural networks are used to generate dimensionality-reduced feature representations of individual word nodes as the central governing word and dependent word, respectively. The dual affine tensor multiplication module receives the aforementioned dimensionality-reduced feature representations and calculates and outputs the probability score of a direct dependency connection between any two word nodes and the corresponding dependency relationship category classification score. In this embodiment, the pre-trained bidirectional language encoder specifically adopts the BERT-base-Chinese version, whose architecture includes a 12-layer Transformer module, a hidden layer dimension of 768, and 12 parallel attention heads. Standard operating manuals, equipment maintenance logs, and historical scheduling instructions in the industrial control field were collected. Through professional human analysis, a completely correct standard reference answer was constructed for each text sequence, marking standard word segmentation node boundaries, standard part-of-speech attribute tags, and corresponding standard dependency syntax tree node dependency connection pairs and dependency type tags. This formed a labeled training corpus containing a total of 50,000 manually verified corpus samples. Before training, 10% of the data was randomly segmented and retained as a model validation set. The 50,000 samples used for training covered more than three typical industrial scenarios, including power, petrochemical, and discrete manufacturing. Approximately 25% of the instructions contained complex nested sentences, and 15% contained non-standard abbreviations or colloquial expressions, ensuring the model's generalization ability to heterogeneous instructions. The text data from the labeled training corpus was converted into feature vectors and input into the model using a pre-set batch size of 32 samples. After the hidden network layer performs forward computation, the top-level double affine tensor multiplication module outputs the predicted dependency connection probability and the predicted dependency relationship category probability. The cross-entropy calculation rule is invoked to calculate the first and second cross-entropy errors between the predicted probabilities and the true distribution of the standard reference answer, respectively. These are then summed with a fixed weight of 1:1 to obtain the overall global prediction loss. Based on this overall global prediction loss, the backpropagation algorithm is used to calculate gradient information, and the adaptive moment estimation optimization algorithm is invoked to fine-tune and update all feature weight nodes within the network with an initial learning rate parameter of 0.0001. The batch iteration process is executed cyclically, and the parsing accuracy is tested using validation set data after each cycle. When the accuracy no longer increases after 10 consecutive iterations, or when the overall global prediction loss is stably lower than the preset convergence threshold of 0.05, an early stopping mechanism is triggered to end training, and the final network weight matrix is ​​saved to the server's hard disk storage area, thus obtaining the natural language processing model. Call the natural language processing model to process the obtained operation instruction text; call the word segmentation algorithm to perform full segmentation on the obtained operation instruction text, and split it into a set of discrete word node sequences; in this embodiment, the word node sequence obtained after word segmentation is specifically preset as "will", "one", "number", "reactor", "of", "temperature", "increase", "thirty", "degrees Celsius"; input the word node sequence into the underlying vocabulary and character feature embedding layer of the natural language processing model to be transformed into an initial feature vector, and sequentially perform forward non-linear calculations through the intermediate self-attention stacked hidden layers to extract a semantic feature vector containing deep global context information; input the semantic feature vector into two feed-forward neural networks arranged in parallel at the top layer to respectively generate a reduced-dimensional feature representation of each word node in the sequence as the central governing word and the subordinate dependent word; the top-layer double-affine tensor multiplication operation module receives the foregoing reduced-dimensional feature representation, calculates the probability score of forming a parent-child dependency connection line between any two word nodes in the word node sequence, and the classification probability of the corresponding connection line's syntactic relationship type; finally, traverse each word node in the word node sequence as the current subordinate node, directly perform local optimal solution matching based on the calculated probability score, select another word node with the highest probability score of forming a parent-child dependency connection line with it as its uniquely associated central governing word node, and simultaneously extract the syntactic relationship type label with the highest classification probability on the corresponding connection line, establish the one-way domination and subordination hierarchical relationship between each word node in the sequence, and instantiate and generate a data structure model in memory with each word as a tree node and a directed edge containing a specific syntactic dependency relationship type label as the connection link, and construct the dependency syntax tree; The process of extracting the instruction semantic triple is to identify the core root node of the dependency syntax tree, extract the operation behavior after verifying the词性特征 of the core root node as a verb; start from the node corresponding to the operation behavior and traverse along the tree structure, and jointly extract the nodes with verb-object dependency relationships and object dependency relationships as candidate operation objects; traverse the sub-tree branches of the operation behavior and the candidate operation objects, and extract the nodes with a quantifier modification dependency relationship and a unit identifier as candidate operation parameters; calculate the dependency topological step lengths of each candidate operation object and each candidate operation parameter to the core root node respectively, and select the node combination with the shortest dependency topological step length to form the instruction semantic triple; It should be noted that the term "词性特征" in the original text is not clear. It might be a specific term in Chinese patent texts that needs to be further clarified according to the context. Here, a rough translation is provided.Specifically, the graph structure traversal algorithm is invoked to access all word nodes in the dependency syntax tree, and the in-degree value of each word node, i.e., the number of directed edges pointing to that node, is counted. The only node with an in-degree value of 0 is established as the core root node. In the specific parsing scenario of this embodiment, the node with an in-degree value of zero is "raise". A set of compliant verb part-of-speech features is preset in memory. This set is mapped from the core action words in the industrial control action ontology library. The specific verification process is as follows: the part-of-speech tag of the core root node is extracted. First, it is determined whether the tag belongs to the preset general verb category. Then, its text string is matched with the control instruction predicate subset in the action ontology library based on the edit distance algorithm. The control instruction predicate subset includes: raise, lower, close, disconnect, start, stop, etc. If the similarity of the text matching is greater than 0.85, it is determined that the verb has industrial control effect and is established as an operation behavior, thereby ensuring that the parsing logic can effectively match the mapping relationship between colloquial predicates and standard control actions. The "elevation" node, extracted as the operation behavior, is used as the starting center anchor point for traversal. The breadth-first search algorithm is invoked to traverse the path downwards along the directed edges of the dependency syntax tree, layer by layer, towards each leaf node. When visiting each lower-level node, the syntactic dependency relationship label between the current node and its direct parent node is read, and it is determined whether the dependency relationship label completely matches the preset verb-object dependency relationship label or whether it completely matches the preset object dependency relationship label. If the match is successful, the current node is extracted and added to the system's candidate operation object memory list. In this embodiment, during the downward traversal, the directed edge label between the "temperature" node and the "elevation" node is identified as a verb-object dependency relationship. At the same time, when continuing to traverse its child nodes downwards, the "reactor" node is identified as having a dependency relationship label representing the object attribute. Therefore, the text of the two nodes, "temperature" and "reactor," is extracted together as the candidate operation object. Using the operation behavior "increase" node, and the extracted candidate operation objects "temperature" node and "reactor" node as the root nodes of the local subtrees, independent threads are started to traverse all their respective subtree branches downwards. During the traversal, a strict two-step conditional screening is performed on each node: the first step is to read the directed edge labels connecting the currently visited node and its parent node and determine whether they are preset quantifier-modified dependency relationships; the second step, provided that the first step condition is met, is to compare the text strings of the currently visited node and its adjacent combined nodes with the system's preset physical measurement unit characters. The dictionary performs inclusion relationship detection to determine whether the text contains compliant unit identifiers. In this embodiment, the preset physical measurement unit dictionary includes characters such as "Celsius", "number", and "Pascal". When traversing the subtree, the "thirty" node is identified as having a quantifier dependency relationship, and its adjacent text links carry the unit identifier "Celsius". At the same time, the "one" node is identified as having a quantifier dependency relationship and its combined text carries the unit identifier "number". Thus, "thirty degrees Celsius" and "one number", which simultaneously meet these two extraction conditions, are extracted and stored in the candidate operation parameter list. Each physically directed edge in the dependency syntax tree is defined as a standard topological step unit. Using a tree-graph shortest path algorithm, the total number of directed edges required to backtrack upwards from each node in the candidate operation object list to the core root node "Raise" is calculated, and this number is used as the corresponding dependency topological step unit. Similarly, the total number of directed edges required to backtrack upwards from each node in the candidate operation parameter list to the core root node is calculated. In this embodiment, the dependency topological step unit from "Temperature" to "Raise" is calculated to be 1, while the dependency topological step unit from "Reaction Vessel" to "Raise" is 2. After comparing the values, the candidate operation object is selected... The node with the shortest step size is selected for "Temperature". The dependency topology step size from the candidate operation parameter "30 degrees Celsius" to "increase" is calculated to be 1, while the dependency topology step size from "number one" backtracking to "increase" is 3. After comparison, the node with the shortest step size, "30 degrees Celsius", is selected. Finally, the system encapsulates the extracted operation behavior "increase", the candidate operation object "temperature" after step size filtering, and the candidate operation parameter "30 degrees Celsius" after step size filtering into fields, and generates a data entity in memory with a structured expression of "operation behavior: increase, operation object: temperature, operation parameter: 30 degrees Celsius", which constitutes the complete instruction semantic triple. By locating the core root node and verifying part-of-speech tags, the system can reliably identify the actual command operation behavior and avoid recognition bias caused by complex sentence structures. By combining dependency relationships and unit identifiers for targeted traversal, non-critical interfering numerical words are effectively filtered out, ensuring that the extracted objects and parameters match the control logic. In addition, a dependency topology step-size calculation mechanism is introduced, which uses the feature of proximity dominance of syntactic structure to select the node combination with the shortest step size, effectively eliminating semantic ambiguity and modification misalignment that are prone to occur in long sentences or nested commands.

[0020] The instruction semantic triples are mapped to corresponding ontology nodes in the semantic constraint graph, and safe numerical boundaries are extracted. The safe numerical boundary extraction process involves matching the names and aligning the attributes of the operation objects in the instruction semantic triples with the ontology nodes in the semantic constraint graph to determine the target ontology node. Based on the operation behavior in the instruction semantic triples, action constraint edges associated with the target ontology node are retrieved in the semantic constraint graph to determine the attribute dimensions under the current operation behavior. The safe threshold interval corresponding to the attribute dimensions in the target ontology node is extracted as the initial boundary. Real-time environment state nodes that have a topological association with the target ontology node are retrieved in the semantic constraint graph, and the initial boundary is dynamically corrected according to the values ​​of the real-time environment state nodes to obtain the safe numerical boundary. Specifically, the operation object string is extracted from the semantic triple of the instruction, and candidate ontology nodes are retrieved in the semantic constraint graph. The name matching is implemented using a string similarity algorithm. The specific process of the string similarity algorithm is as follows: the operation object string is compared character by character with the name string of the ontology node in the semantic constraint graph; the minimum number of character editing steps required to convert one string into another is calculated, including the number of times characters are added, deleted, or replaced; the minimum number of character editing steps is divided by the total length of the longer of the two strings to obtain the difference rate, and 1 is subtracted from the difference rate to obtain the similarity score; when the similarity score exceeds 0.9 (which can be adjusted according to actual working conditions), the name matching is considered successful; simultaneously, the attributes include the device's physical number and... Loop identifier; retrieve the metadata of the source of the operation instruction, extract the physical number and loop identifier of the instruction, and perform a logical comparison with the physical number attribute and loop identifier attribute recorded by the matched nodes in the semantic constraint graph to determine whether the values ​​of the two are completely consistent. If the name matches successfully and the physical number and loop identifier are completely equivalent, the attribute alignment is determined to be successful, and the ontology node is determined to be the target ontology node. If a matching or alignment failure occurs when executing the name matching and attribute alignment logic, it is determined that the operation object node in the instruction semantic triple cannot find a mapping relationship in the semantic constraint graph, and the node is marked as missing target ontology node. After identifying the target ontology node, based on the operational behavior in the instruction semantic triple, the action constraint edges associated with the target ontology node are retrieved in the semantic constraint graph to determine the attribute dimensions under the current operational behavior. Taking the target ontology node as the starting point, the edges with action attributes directly connected to the target ontology node in the semantic constraint graph, i.e., action constraint edges, are traversed. The action constraint edges pre-store the mapping relationship between action names and physical quantity categories. By comparing the semantic content of the operational behavior, the corresponding action constraint edges are located, and the physical measurement standards defined in the action constraint edges are extracted as attribute dimensions. The attribute dimensions include milliamperes, degrees Celsius, volts, decibels, kilohertz, percentages, ampere-hours, charge / discharge rates, cycles, milliseconds, days, and pixels. The system enters the data storage space of the target ontology node. Based on the determined attribute dimension, it retrieves the threshold list encapsulated within the target ontology node, searches for a numerical range field whose name completely matches the attribute dimension, and reads the upper and lower limits of the recorded values. For example, for the megapascal dimension, if the upper limit is read as 15 and the lower limit as 2, then the numerical range from 2 to 15 is determined as the initial boundary. This initial boundary represents the basic safe operating range of the target ontology node under ideal operating conditions.

[0021] The process involves retrieving real-time environmental status nodes from the semantic constraint graph that have a topological association with the target ontology node. Specifically, using the target ontology node as the origin, nodes with first-order or second-order adjacency relationships are searched in the semantic constraint graph, and nodes marked with the "environmental monitoring" attribute are identified; these are the real-time environmental status nodes. These real-time environmental status nodes update external environmental monitoring values ​​in real time via an IoT interface, and their current values ​​are retrieved. The initial boundary is dynamically corrected based on the values ​​of the real-time environmental state nodes to obtain a safe numerical boundary. A range adjustment mapping table is used for conversion, recording the logical correspondence between environmental numerical deviations and operational parameter corrections. The values ​​of the real-time environmental state nodes are subtracted from the standard environmental reference values ​​recorded in the semantic constraint graph to obtain the environmental deviation value. The correction step size corresponding to this deviation value is retrieved from the range adjustment mapping table. The corrected upper limit is obtained by subtracting the correction step size from the upper limit of the initial boundary, and the corrected lower limit is obtained by adding the correction step size to the lower limit of the initial boundary. The numerical range formed by the corrected upper and lower limits is determined as the final safe numerical boundary. The target ontology node in the semantic constraint graph is located by using the string edit distance algorithm combined with loop identifier verification. The attribute dimensions are determined based on the associated action constraint edges and the static initial boundary is retrieved. At the same time, the real-time values ​​of the associated environmental nodes are obtained through the graph topology path. The environmental deviation is mapped to the physical correction step size through the interval adjustment mapping table. The upper and lower limits of the initial boundary are added and subtracted to output an adaptive safety numerical boundary that matches the current working condition. The semantic constraint graph is a multidimensional heterogeneous knowledge graph, including an entity node layer, a relation edge layer, and an attribute constraint layer. The entity node layer includes ontology nodes representing different types of operating devices and real-time environmental state nodes representing the perceived operating conditions. The relation edge layer includes topological association edges connecting the influence relationships between the ontology nodes and the real-time environmental state nodes, as well as action constraint edges connecting the operating behavior and the ontology nodes. The attribute constraint layer includes a feature dictionary configured in each ontology node, which is used to map and store the attribute dimensions and corresponding safety threshold ranges under the action constraint edges.

[0022] Specifically, the semantic constraint graph is a multi-dimensional heterogeneous knowledge graph stored in a server database. Logically, it is divided into three dimensions: entity node layer, relation edge layer, and attribute constraint layer, to support deep semantic verification of operation instructions. Through the graph database management engine, the physical entities, logical relationships, and control rules of the industrial site are digitally modeled to ensure that each operation object has a corresponding topological location and constraint logic in the graph. During the implementation of the entity node layer, two types of core nodes are pre-established: entity nodes and real-time environmental status nodes. Entity nodes are pre-defined as digital entities representing different types of operating equipment. For example, in this embodiment, "Centrifugal Pump No. 1" and "Heating Furnace No. 2" are set as entity nodes, storing the unique identification code and equipment type information of the equipment. Real-time environmental status nodes are virtual nodes representing perceived operating conditions, which are connected to the monitoring values ​​of field sensors in real time via IoT interfaces. For example, "Workshop Real-time Ambient Temperature" is set as a real-time environmental status node to reflect the external physical background of equipment operation. Entity nodes adopt a naming convention of "unique equipment serial number + physical location label + functional description." When equipment with the same name appears (such as "Pump No. 1"), the loop identifier from the instruction source is automatically extracted for spatial coordinate alignment, and node disambiguation is achieved through unique coding. In the implementation of the relational edge layer, logical interactions between nodes are defined by establishing topological association edges and action constraint edges. Topological association edges connect the entity node and the real-time environmental state node, representing the physical influence of the environment on the equipment. For example, the entity node "Heating Furnace No. 2" is connected to the real-time environmental state node "Workshop Real-time Ambient Temperature" via a topological association edge, meaning that fluctuations in the ambient temperature directly affect the safety control logic of the heating furnace. Action constraint edges connect operational behaviors to entity nodes, representing the specifications that the entity node should follow under the action. For example, the operational behavior "Heating Up" is pointed to "Heating Furnace No. 2" via an action constraint edge, thus determining the operating boundary of the equipment under this action. The establishment of these action constraint edges is based on the structured extraction of the equipment's official "Operating Procedure Manual" and "Safety Operation Standards," supporting manual rule definition and direct import from expert knowledge bases. During the implementation of the attribute constraint layer, a feature dictionary is configured within each ontology node. This feature dictionary is preset as a key-value pair mapping table in the underlying data structure, its function being to map and store the attribute dimensions and corresponding safety threshold ranges under action constraint edges. Specifically, when a "heating" action constraint edge is identified for the "Heating Furnace No. 2" ontology node, the feature dictionary within that node is automatically accessed, the corresponding attribute dimension is extracted as "degrees Celsius," and the preset safety threshold range under that dimension is retrieved. The initial values ​​of the safety threshold ranges in the feature dictionary are referenced from relevant industry safety standards in GB / T or IEC. The real-time environmental status node accesses field sensor data via the OPC-UA industrial communication protocol, with a data update frequency set to no less than 1Hz to ensure the real-time performance of dynamic boundaries. By constructing a multidimensional heterogeneous semantic constraint graph containing entity node layer, relation edge layer, and attribute constraint layer, a digital mapping system for industrial entities, real-time operating conditions, and action specifications was established. The logical association between equipment nodes and environmental influence factors was realized by using topological association edges and action constraint edges, and the mapping mechanism between attribute dimensions and safety thresholds under different operating behaviors was established through feature dictionary, thus providing a structured topological verification path for adaptive alarm control. See Figure 2 When the operation parameters are extracted, the absolute deviation of the operation parameters from the safety numerical boundary is calculated and converted into parameter limit violation through nonlinear mapping; the dependency syntax tree is topologically compared with the built-in standard instruction template to locate missing nodes and generate semantic incompleteness probability; an instruction risk index is generated based on the parameter limit violation and the semantic incompleteness probability. The parameter deviation is calculated by taking the absolute value of the difference between the operating parameter and the nearest safe boundary endpoint as the absolute deviation, and generating the value by nonlinearly mapping the absolute deviation through an exponential function. The semantic incompleteness probability is the ratio obtained by dividing the total number of nodes in the standard instruction template and the number of missing nodes located by topology comparison by the total number of nodes. The parameter exceedance is normalized, and the instruction risk index is constructed by using the parameter exceedance and the semantic incompleteness probability as independent risk dimensions to build a two-dimensional evaluation space. The sum of squares of the parameter exceedance and the semantic incompleteness probability is calculated, and the square root of the sum of squares is performed to obtain the magnitude of the risk vector in the two-dimensional evaluation space as the instruction risk index. When no operation parameters are extracted from the instruction semantic triple, the parameter exceedance is assigned a preset constant value, and the instruction risk index is generated solely based on the semantic incompleteness probability.

[0023] Specifically, the calculation process for parameter exceedance involves obtaining the operational parameter values ​​from the instruction semantic triplet and acquiring the corresponding safe numerical boundaries from the semantic constraint graph. The operational parameter is compared with the safe numerical boundaries; if the parameter exceeds the upper limit or falls below the lower limit, the absolute value of the difference between it and the nearest boundary endpoint is calculated to obtain the absolute deviation. For example, if the safe numerical boundary is set to 20 to 80, and the actual operational parameter is 100, the absolute deviation is 20. The absolute deviation is then nonlinearly mapped using an exponential function, specifically through a logistic regression S-curve function, introducing a preset deviation threshold and a deviation sensitivity coefficient. The deviation threshold is determined by statistically analyzing the fluctuation distribution of the operational parameter under normal execution conditions in historical operation and maintenance logs, taking its 95th percentile as the deviation threshold to filter out common industrial noise fluctuations. The sensitivity coefficient is determined through reverse derivation using numerical simulation: when the absolute deviation reaches one time the initial boundary range, the generated parameter exceedance value approaches 0.9; the sensitivity coefficient is then calculated and fixed accordingly. For core nodes with high criticality levels, the sensitivity of risk response is improved by increasing a preset weighting factor on top of the above. During calculation, the obtained absolute deviation is multiplied by the sensitivity coefficient, and the deviation threshold is subtracted to obtain an intermediate risk variable. The negative of this intermediate risk variable is used as the exponent of the natural constant for exponentiation, and the result is incremented by 1 to form the denominator. A division operation is then performed with 1 as the numerator. This fractional structure maps physical deviations of varying dimensions to a standardized numerical range between zero and one, generating the parameter limit exceedance.

[0024] The semantic incompleteness probability calculation process involves a pre-set standard instruction template. This template defines the node structure that an instruction must possess, including operation behavior nodes, operation object nodes, and environmental constraint nodes. The total number of nodes in the template is counted; in this embodiment, the total number of nodes is set to 5. When processing actual instructions, the generated dependency syntax tree is topologically compared with the standard instruction template to identify missing nodes not present in the instruction. The number of missing nodes is counted and divided by the total number of nodes. For example, if the comparison reveals that two necessary nodes are missing, the number of missing nodes (2) divided by the total number (5) yields a ratio of 0.4. This value represents the semantic incompleteness probability, quantifying the semantic logical incompleteness of the instruction.

[0025] The parameter exceeding the limit is normalized and scaled to the range of 0 to 1 using a linear scaling method to ensure that it is on the same dimension as the semantic incompleteness probability. The normalized parameter exceeding the limit and the semantic incompleteness probability are used as two independent risk dimension coordinates to construct a two-dimensional evaluation space in Cartesian coordinates. The square value of the parameter exceeding the limit and the square value of the semantic incompleteness probability are calculated. These two square values ​​are added together to obtain a sum of squares. The square root of the sum of squares is then performed to calculate the magnitude of the risk vector from the origin to the risk point in this two-dimensional space. This magnitude is determined as the instruction risk index. When no operation parameters are extracted from the instruction semantic triplet, it indicates that the current instruction only contains the operation action and the operation object, lacking specific values ​​as a basis for safety verification. At this time, the preset constant substitution logic will be activated. Since the absolute deviation cannot be calculated, the parameter limit is directly assigned to the preset constant value. In this embodiment, the preset constant value is set to 0.5 to represent the moderate level of basic uncertainty risk introduced due to missing parameter information. Under this calculation path, since the parameter limit is fixed to this constant, the coordinate coefficient value of this dimension in the two-dimensional evaluation space remains constant. Therefore, the numerical fluctuation of the instruction risk index will only depend on the semantic incompleteness probability. In specific implementation, the preset constant value of 0.5 and the semantic incompleteness probability obtained by actual statistical calculation are used as the horizontal and vertical axes. The sum of their squares is calculated and the square root is taken. The instruction risk index output thereby includes both the basic warning benchmark of missing parameters and realizes a single dynamic response to the integrity of the instruction structure. By introducing parameter limit-crossing, exponential nonlinear mapping was used to achieve sensitive capture and rapid amplification of risk deviation, significantly enhancing the early warning intensity for numerical violations. By introducing semantic incompleteness probability, the degree of missing instruction nodes was quantified through topological comparison, and the risk of incomplete semantic logic was identified. The two-dimensional evaluation model constructed by the two collaboratively achieved a deep fusion evaluation of numerical deviation and structural defects. See Figure 3Based on the risk index of the instruction, the threshold is dynamically matched to trigger adaptive control: when the risk index is lower than the first risk threshold, the instruction is executed normally; when it is in the first interval, a primary warning is triggered and the semantic completion rule is called to reconstruct the semantic triple of the instruction; when it is in the second interval, an intermediate alarm is triggered to intercept the execution and syntactic clarification text is fed back; when it exceeds the highest threshold, an advanced alarm is triggered to block the operation and abnormal samples are recorded. The adaptive control includes: when the risk index is less than a first risk threshold, the instruction is executed normally; when the instruction risk index is in a first interval (greater than or equal to the first risk threshold and less than the second risk threshold), a primary warning is triggered and semantic completion rules are invoked to reconstruct the instruction semantic triplet; when the instruction risk index is in a second interval (greater than or equal to the second risk threshold and less than the highest threshold), an intermediate alarm is triggered to intercept execution and provide syntactic clarification text; when the index is greater than or equal to the highest threshold, an advanced alarm is triggered to block the operation and abnormal samples are recorded.

[0026] Specifically, in this embodiment, the first risk threshold is set to 0.3, the second risk threshold is set to 0.6, and the highest threshold is set to 0.85. The reason for this setting is based on the statistical analysis of the risk index distribution of 10,000 historical abnormal instruction samples. When the calculated instruction risk index is less than the first risk threshold of 0.3, the instruction is determined to have a complete logical structure and extremely high execution security. Under this path, an execution trigger signal is directly generated. The complete instruction semantic triple is sent to the instruction scheduling gateway and executed normally according to the established business process. When the instruction risk index is in the first interval (i.e., the instruction risk index is greater than or equal to 0.3 and less than 0.6), the instruction is determined to have mild semantic incompleteness. At this point, a primary warning is triggered, indicated by a yellow indicator on the monitoring interface. Pre-defined semantic completion rules are automatically invoked to reconstruct the instruction's semantic triples. Specifically, the semantic completion rules retrieve historical high-frequency completion records from the knowledge graph and automatically fill in missing nodes. The semantic completion rule involves parsing the topology comparison to locate the type of missing node and determining the missing element; obtaining a set of candidate nodes corresponding to the missing element; retrieving historical security execution logs and constructing a Markov state transition matrix based on the operation sequence; using the Markov state transition matrix, calculating the transition probability of the known node in the instruction semantic triplet to each candidate node in the candidate node set, and generating the temporal co-occurrence of each candidate node; selecting the candidate node with the highest temporal co-occurrence to fill the missing node and complete the reconstruction; if no candidate node is matched and the missing element is an operation parameter, then the median of the security numerical boundary is extracted as the security benchmark parameter for reconstruction.

[0027] Specifically, the generated topology comparison results are retrieved, and the differences between the dependency syntax tree of the current instruction and the built-in standard instruction template are recorded. By identifying the positions of nodes that exist in the standard instruction template but are missing in the current instruction tree, the type of missing element is determined, and it is determined whether the missing element is an operation object, operation behavior, or operation parameter.

[0028] By accessing the semantic constraint graph, ontology nodes of the same type that are logically related to the currently identified operation behavior or operation object are retrieved to form a candidate node set; for example, if the missing element is identified as the target opening degree in the operation parameters, all compliant opening degree values ​​that have appeared in the history of the device are extracted from the semantic constraint graph and summarized into a candidate node set. Retrieve historical security execution logs stored in the database. These logs contain a preset number (10,000 in this embodiment) of consecutive successfully executed operation instruction sequences. Treat each instruction semantic triple as a state, count the number of times from state A to state B, and construct a Markov state transition matrix. The rows and columns of the Markov state transition matrix represent different instruction semantic nodes, and the values ​​in the matrix represent the transition frequency of the two nodes in time. The known nodes in the current instruction semantic triple are extracted. The state in the Markov state transition matrix is ​​defined as a discrete combination of "operation behavior + operation object". The probability of these known nodes evolving into each candidate node in the candidate node set is queried in the Markov state transition matrix. The specific calculation method is as follows: the number of transitions from a known node to a certain candidate node is divided by the total number of transitions from the known node to all nodes. The result is the transition probability of the candidate node. In this embodiment, the probability value is defined as the temporal co-occurrence degree. When the record of the operation sequence in the transition matrix is ​​empty, the system automatically switches to a global heat protection mechanism based on the single dimension of operation behavior, and extracts the attribute description with the highest frequency associated with the action in the historical security log as a candidate suggestion. To address the prediction failure problem caused by the sparsity of the state space, a conditional probability decomposition logic is used to reduce the dimensionality of Markov state transitions. Specifically, the frequency of two-dimensional temporal evolution between the known operation object and the missing element is independently counted under the premise of the operation behavior. When the record of the operation sequence in the transition matrix is ​​empty, it automatically switches to a global heat protection mechanism based on the single dimension of the operation behavior, that is, extracting the attribute description with the highest frequency associated with the action in the historical security log as a candidate suggestion. When calculating the transition probability, a smoothing constant is superimposed on each transition count term to avoid logical deadlock caused by zero probability. This hierarchical prediction mechanism ensures that the algorithm can still output a robust instruction completion scheme based on the operation context and historical experience when the sample is insufficient. The dimensionality reduction process specifically decomposes the original ternary joint probability of 'action-object-parameter' into the conditional probability of 'action-object' to 'parameter'. When constructing the Markov state transition matrix, the sliding window size is set to five consecutive operation instructions, and only the conditional probability of the operation object and a certain candidate node appearing sequentially in the same window under the operation behavior is counted.

[0029] The temporal co-occurrence of all nodes in the candidate node set is sorted, and the candidate node with the highest temporal co-occurrence value is selected as the best completion item. The best completion item is then filled into the corresponding missing node position in the dependency syntax tree, and the complete instruction semantic triple is reassembled and constructed.

[0030] If no valid node is found in the candidate node set, and the missing element is determined to be an operation parameter, a safety margin reconstruction mechanism is initiated. First, a safe numerical boundary is defined, consisting of an interval with a minimum and a maximum value. The arithmetic mean of this safe numerical boundary is calculated, i.e., the sum of the maximum and minimum values ​​divided by 2 is taken as the median. This median is extracted as the safe baseline parameter and forcibly filled into the parameter position of the instruction semantic triple to complete the reconstruction, thereby ensuring that the instruction can still be issued with the safest baseline value even when parameters are missing. By deeply mining the temporal patterns of historical security data through Markov state transition matrices, intelligent logical repair and reconstruction of incomplete instructions are achieved. On the basis of improving operational coherence, the safety boundary median fallback mechanism ensures that the system can still operate in the most robust state when critical information is missing. When the instruction risk index is in the second range (i.e., 0.6 or higher and less than 0.85), the instruction is determined to have moderate logical ambiguity and does not meet the conditions for direct completion and execution. In this case, a medium-level alarm is immediately triggered, forcibly interrupting the instruction delivery process via the instruction interceptor. Simultaneously, feedback logic is initiated, automatically generating syntactic clarification text based on the specific missing node name located in the dependency syntax tree. The syntactic clarification text is generated by extracting the missing node labels from the dependency syntax tree and the standard instruction template topological comparison, mapping the semantic constraint graph to obtain the ontology attribute descriptions corresponding to the missing nodes, retrieving the attribute dimensions associated with the current operation in the semantic constraint graph, generating a reference interval containing the safety numerical boundaries corresponding to the attribute dimensions, and then template-concatenating the ontology attribute descriptions, the reference intervals, and the top three candidate nodes with the highest temporal co-occurrence generated by the semantic completion rules to generate guiding text containing missing location hints, completion option guidance, and safety compliance range information as the syntactic clarification text. Specifically, the topology comparison results are retrieved, and the label information of all nodes located as missing is extracted. For example, if the standard instruction template contains an operation target value node but the current instruction is missing, the logical label of the node is extracted, and the logical label is mapped to the attribute constraint layer of the semantic constraint graph. By retrieving the feature dictionary associated with the current ontology node, the ontology attribute description corresponding to the label is obtained. In this embodiment, the description is set to an easy-to-understand natural language phrase. By accessing the relation edge layer of the semantic constraint graph, the action constraint edge connected to the currently identified operation behavior is retrieved. Based on the action constraint edge, the corresponding attribute dimension is located in the attribute constraint layer, the preset safety threshold range under the attribute dimension is extracted, and the value of the current real-time environment state node is dynamically corrected. The corrected value range is then encapsulated into a reference range containing the unit of measurement. The semantic completion rule logic is invoked, and the Markov state transition matrix is ​​used to predict the probability of missing positions. Instead of directly selecting the single node with the highest probability, the system sorts the temporal co-occurrence of all nodes in the candidate node set in descending order, selecting the top three candidate nodes by numerical ranking. These three nodes represent the three completion options that are most relevant to the current known operation context in the historical safe execution log. Retrieve the built-in guided dialogue template, which contains preset placeholders; fill in the missing position prompt placeholder with the ontology attribute description, fill in the top three candidate nodes of temporal co-occurrence with the completion option guide placeholder, and fill in the safety compliance range inform placeholder with the reference interval with dimensions. The generated final syntactic clarification text will look like this: "Command parsing blocked. Your operation is missing 'device operating target frequency'. Based on historical practice, we suggest filling in: '40', '45', or '35'. The current safety compliance range is: 30 Hz to 50 Hz." This text will be immediately fed back to the user interface, intercepting the current execution action and waiting for the operator to correct or confirm the command based on the prompts. This provides interactive completion suggestions while blocking high-risk operations. By transforming instruction syntax defects into a combined feedback that includes ontology attribute descriptions, temporal evolution suggestions, and dynamic safety intervals, an interactive upgrade from mechanical interception to intelligent error correction guidance is achieved. When there is logical ambiguity in the instruction, it can not only provide safety red line warnings through physical boundary information, but also use historical experience mined by Markov models to assist in rapid human decision-making. When the risk index of an instruction is greater than or equal to the highest threshold of 0.85, the current instruction is determined to be an extremely high-risk operation. At this time, the highest-level advanced alarm is triggered, and the execution mechanism immediately switches to a safety lockout state, blocking all operational links related to the instruction. Simultaneously, the abnormal sample recording process is executed, automatically collecting the operation time, original text, generated instruction semantic triplet, parameter limit violation value, and corresponding risk index coordinates of the instruction, and encapsulating them into a structured abnormal data file; this file is stored in the system security audit database as a negative example sample for identifying malicious instructions; This invention breaks away from the dependence of traditional control systems on rigid protocols by constructing a natural language processing model for deep analysis of instruction semantics, effectively solving the problem of non-standard instruction parsing failure. By using a multi-dimensional heterogeneous semantic constraint graph to link with real-time environmental conditions, it achieves adaptive dynamic correction of safety boundaries, fundamentally eliminating the risk of misjudgment caused by static threshold verification. Combined with a multi-level adaptive control strategy based on risk index, it provides guidance from intelligent completion to clarification based on quantified numerical deviations and structural defects.

[0031] Example 2: In this embodiment, taking the intelligent monitoring scenario of the drawer position status of a low-voltage switchgear as an example, the implementation process of a multi-level alarm adaptive control method based on operation command parsing rules is described in detail: During the instruction acquisition and parsing phase, the text instruction received from the scheduling system is: "Set the alarm frequency of the experimental position of drawer number one to 2,500 Hz". The natural language processing model constructs the dependency syntax tree of the instruction through a dual affine dependency attention network and identifies the core root node verb as "set". Through path traversal and step size calculation, the instruction semantic triple is extracted as: operation object: drawer number one, operation behavior: set, operation parameter: 2,500 Hz. The operation object "Drawer No. 1" is mapped to the corresponding ontology node in the semantic constraint graph. Attribute alignment is completed through physical number and loop identifier. Based on the "Set" operation behavior, the "Experimental Position" action constraint edge associated with the drawer is retrieved in the relation edge layer, and its attribute dimension is determined to be "Hertz". The initial safety boundary is retrieved as 1900 to 2100 Hertz (corresponding to the standard 2kHz low-frequency sound). Combined with the field ambient temperature value monitored by the real-time environmental status node, the boundary is dynamically corrected to 1850 to 2050 Hertz according to the interval adjustment mapping table. The absolute deviation of the operation parameter "2500 Hz" from the upper limit of the safety boundary (2050 Hz) is calculated to be 450 Hz, and is converted into parameter limit violation through an exponential nonlinear mapping. At the same time, the current dependency syntax tree is compared with the built-in standard instruction template (which contains 5 nodes by default). It is found that the instruction is missing the environmental constraint node "duration". The semantic incompleteness probability is calculated to be 0.2. Using the normalized parameter limit violation and semantic incompleteness probability as two-dimensional coordinates, the risk vector magnitude (instruction risk index) is calculated to be approximately 0.75. Since the instruction risk index of 0.75 falls within the second range (0.6 to 0.85), the instruction is determined to have moderate logical ambiguity and severely deviated parameters, immediately triggering a medium-level alarm. The instruction interceptor forcibly interrupts the issuance process to prevent excessively high output frequencies from damaging the audible and visual alarm module or causing misjudgment by maintenance personnel. Simultaneously, the semantic constraint graph is mapped to generate syntactic clarification text: "Instruction parsing is blocked. Your operation is missing the 'duration' parameter, and the frequency deviates from the compliance range. It is recommended to fill in: 2000, 2100, or 1900 Hz. The current safety compliance range is: 1850 Hz to 2050 Hz."

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-stage alarm adaptive control method based on operation instruction analysis rules, characterized in that, include: Obtain the operation instruction text and construct a dependency syntax tree using a natural language processing model; Based on the dependency syntax tree, extract instruction semantic triples containing the operation object, operation behavior, and optional operation parameters; The instruction semantic triples are mapped to the corresponding ontology nodes of the semantic constraint graph to extract the safe numerical boundary. When the operation parameter is extracted, the absolute deviation between the operation parameter and the safe numerical boundary is calculated and converted into parameter limit violation through nonlinear mapping. The dependency syntax tree is topologically compared with the built-in standard instruction template to locate missing nodes and generate semantic incompleteness probability. An instruction risk index is generated based on the parameter limit violation and the semantic incompleteness probability. Based on the dynamic matching threshold of the instruction risk index, adaptive control is triggered: when the risk index is lower than the first risk threshold, the instruction is executed normally; when it is in the first interval, a primary warning is triggered and the semantic completion rule is invoked to reconstruct the instruction semantic triple; when it is in the second interval, an intermediate alarm is triggered to intercept the execution and provide syntactic clarification text; when it exceeds the highest threshold, an advanced alarm is triggered to block the operation and abnormal samples are recorded.

2. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 1, characterized in that, The instruction semantic triple extraction process involves identifying the core root node of the dependency syntax tree, verifying the part-of-speech feature of the core root node as a verb, and extracting it as an operation behavior. Starting from the node corresponding to the operation behavior, a path traversal is performed along the tree structure, and nodes with verb-object dependency relations and object dependency relations are extracted together as candidate operation objects. The subtree branches of the operation behavior and the candidate operation objects are traversed, and nodes with quantifier-modified dependency relations and unit identifiers are extracted as candidate operation parameters. Calculate the dependency topology step length from each candidate operation object and each candidate operation parameter to the core root node, and select the node combination with the shortest dependency topology step length to form the instruction semantic triple.

3. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 1, characterized in that, The process of extracting the safety numerical boundary involves matching the names and aligning the attributes of the operation objects in the instruction semantic triples with the ontology nodes in the semantic constraint graph to determine the target ontology node; based on the operation behavior in the instruction semantic triples, retrieving the action constraint edges associated with the target ontology node in the semantic constraint graph to determine the attribute dimensions under the operation behavior; and extracting the safety threshold interval in the target ontology node corresponding to the attribute dimensions as the initial boundary. Retrieve real-time environment state nodes in the semantic constraint graph that have topological association with the target ontology node, and dynamically correct the initial boundary based on the values ​​of the real-time environment state nodes to obtain a safe numerical boundary.

4. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 3, characterized in that, The semantic constraint graph is a multidimensional heterogeneous knowledge graph, including an entity node layer, a relation edge layer, and an attribute constraint layer. The entity node layer includes ontology nodes representing different types of operating devices and real-time environmental state nodes representing the perceived operating conditions. The relation edge layer includes topological association edges connecting the influence relationships between the ontology nodes and the real-time environmental state nodes, as well as action constraint edges connecting the operating behavior and the ontology nodes. The attribute constraint layer includes a feature dictionary configured in each ontology node, which is used to map and store the attribute dimensions and corresponding safety threshold ranges under the action constraint edges.

5. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 1, characterized in that, The parameter deviation is calculated by taking the absolute value of the difference between the operating parameter and the nearest safe numerical boundary endpoint as the absolute deviation, and generating the value by nonlinearly mapping the absolute deviation through an exponential function. The semantic incompleteness probability is the ratio obtained by dividing the total number of nodes in the standard instruction template and the number of missing nodes located by topology comparison by the total number of nodes. The parameter exceedance is normalized, and the instruction risk index is constructed by using the parameter exceedance and the semantic incompleteness probability as independent risk dimensions to construct a two-dimensional evaluation space. The sum of squares of the parameter exceedance and the semantic incompleteness probability is calculated, and the square root operation is performed on the sum of squares to obtain the magnitude of the risk vector in the two-dimensional evaluation space as the instruction risk index. When no operation parameters are extracted from the instruction semantic triple, the parameter excess is assigned a preset constant value, and the instruction risk index is generated solely based on the semantic incompleteness probability.

6. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 1, characterized in that, The adaptive control includes: when the instruction risk index is less than a first risk threshold, the instruction is issued and executed normally; when the instruction risk index is in a first interval (greater than or equal to the first risk threshold and less than a second risk threshold), a primary warning is triggered and semantic completion rules are invoked to reconstruct the instruction semantic triplet; when the instruction risk index is in a second interval (greater than or equal to the second risk threshold and less than a maximum threshold), a mid-level alarm is triggered to intercept execution and provide syntactic clarification text; when the instruction risk index is greater than or equal to the maximum threshold, a high-level alarm is triggered to block execution and abnormal samples are recorded.

7. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 6, characterized in that, The semantic completion rule is to parse the missing node type located by topological comparison, determine the missing element; obtain the candidate node set corresponding to the missing element; retrieve the historical safe execution log, construct a Markov state transition matrix based on the operation sequence; and calculate the transition probability of the instruction semantic triple to each candidate node in the candidate node set through the Markov state transition matrix, and generate the temporal co-occurrence degree of each candidate node. The candidate node with the highest temporal co-occurrence is selected to fill the missing node and complete the reconstruction. If no candidate node is matched and the missing element is an operation parameter, the median of the safe numerical boundary is extracted as a safe baseline parameter for reconstruction.

8. The multi-stage alarm adaptive control method based on operation instruction analysis rules according to claim 6, characterized in that, The syntactic clarification text is generated by extracting the missing node labels from the dependency syntax tree and the standard instruction template topological comparison, and mapping the semantic constraint graph to obtain the ontology attribute descriptions corresponding to the missing nodes; retrieving the attribute dimensions associated with the current operation behavior in the semantic constraint graph, and generating a reference interval containing the safety numerical boundaries corresponding to the attribute dimensions; and template-concatenating the ontology attribute descriptions, the reference intervals, and the top three candidate nodes with the highest temporal co-occurrence generated by the semantic completion rules to generate guiding text containing missing location hints, completion option guidance, and safety compliance range information as the syntactic clarification text.

Images