Scheduling automation fault intelligent processing system and method based on knowledge reasoning navigation

Abstract

The invention discloses a dispatching automation fault intelligent processing system based on knowledge reasoning navigation. The dispatching automation fault intelligent processing system comprises a dispatching automation system environment sensing module which obtains machine room environment multi-source data in real time to serve as a data source; the intelligent analysis module performs preprocessing and labeling on the image type data and the text type data, then performs knowledge extraction to obtain entities and relationships among the entities, performs data fusion on the obtained entities, constructs an automatic duty knowledge graph, forms a scheduling automation system fault research and judgment mechanism reasoning model, and performs fault diagnosis on the scheduling automation system. A fault alarm event is generated, a knowledge reasoning environment is constructed, and a fault research and judgment conclusion is obtained through reasoning of a knowledge reasoning algorithm; an auxiliary decision-making module performs reasoning matching on the fault alarm event and the automatic duty knowledge graph through an RE2 model to obtain a fault handling auxiliary decision-making opinion; and constructing a fault exception handling flow navigation chart, and generating a fault handling coping strategy. The operation and maintenance management level of the dispatching automation machine room is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid control automation, and in particular to a dispatching automation fault intelligent handling system and method based on knowledge reasoning navigation. Background Art

[0002] At present, the operation and management of dispatching automation rooms are mostly decentralized management modes in various places, and the main problems are as follows: the patrol frequency of automation duty personnel is low, and faults in the room cannot be discovered and handled in time; the duty system functions are not perfect, and there is a lack of fault alarm event analysis, fault judgment and fault handling auxiliary decision-making means, and fault analysis and fault handling operations still mainly rely on the experience of operation and maintenance personnel.

[0003] Knowledge graph is the technical direction for realizing machine "cognition". At present, the static text knowledge graph for the field of power grid control has realized single-step reasoning based on a single triple, but it is not enough to support the fault event analysis and fault auxiliary decision-making business of the dispatching automation system. With the increase in the number of automated secondary equipment, the system will generate massive control operation data, and it is necessary to identify anomalies in multi-source data and merge them into fault alarm events. In addition, the intelligent handling technology of power grid faults based on knowledge graph reasoning has been preliminarily studied in the power grid dispatching system, but there is less research on intelligent handling of faults and auxiliary decision-making in the automated secondary system. Through the knowledge graph technology, multi-source data of automated duty are integrated to perform reasoning navigation, and the fault alarm event analysis and fault handling auxiliary decision-making opinions of the dispatching automation system are realized. Summary of the invention

[0004] The purpose of the present invention is to provide a dispatching automation fault intelligent handling system and method based on knowledge reasoning navigation, which improves the operation and maintenance management level of the dispatching automation room.

[0005] To achieve this purpose, the present invention designs a dispatching automation fault intelligent handling system based on knowledge reasoning navigation, which includes:

[0006] The environment perception module of the dispatching automation system is used to obtain multi-source data information of the computer room environment in real time through the computer room inspection robot, and send the multi-source data information of the computer room environment to the background server as the data source of the intelligent analysis module and the auxiliary decision-making module;

[0007] The intelligent analysis module is used to divide the data source into numerical data, image data and text data according to the data type, perform knowledge extraction after preprocessing and annotating the image data and text data to obtain entities and the relationship between each entity, perform data fusion on the obtained entities, and construct an automated duty knowledge graph based on the relationship between each entity and the entities after data fusion. Based on the multi-source data information of the computer room environment collected by the dispatching automation system environment perception module and combined with the automated duty knowledge graph, a reasoning model of the fault analysis mechanism of the dispatching automation system is formed, thereby generating a fault alarm event, and using the reinforcement learning path discovery algorithm to model the automated duty knowledge graph as a knowledge reasoning environment model, and using the knowledge reasoning environment model to draw a fault analysis conclusion through reasoning with a knowledge reasoning algorithm based on a combination of multi-level relationship paths;

[0008] The auxiliary decision-making module is used to input the node contents of fault alarm events and automated duty knowledge graphs into the RE2 model, and use the RE2 model to infer and match the node contents of fault alarm events and automated duty knowledge graphs, and combine the fault analysis conclusions to obtain auxiliary decision-making opinions for fault handling; the auxiliary decision-making opinions for fault handling are constructed into a multi-path selection fault abnormality handling process navigation map according to the time sequence of fault alarm events, and a fault handling response strategy is generated to realize the auxiliary execution of the response strategy based on knowledge guidance.

[0009] Preferably, the method for preprocessing and labeling image data includes: using a Gaussian filtering method to remove noise, using a histogram equalization method to adjust image contrast, and using an image labeling tool to manually label warning areas and fault objects in image recordings to obtain image label data;

[0010] The method for preprocessing and labeling text data includes: using data cleaning and standardization methods for preprocessing, removing special symbols and redundant spaces in the text, and marginal information irrelevant to fault alarm events and fault analysis, using text labeling tools to label plan entities, relationships and attributes in the text data, and obtaining text label data.

[0011] Preferably, the specific process of extracting knowledge from the preprocessed and annotated text data is as follows:

[0012] Input the fault alarm event sentence into the BERT model, perform pre-training on the input text through self-supervised learning method, capture the contextual semantic information of entity words, and generate word vector representation;

[0013] The word vector representation is passed to the bidirectional long short-term memory network model to continue capturing the previous and next dependencies between entity words;

[0014] The sequence composed of the hidden state vectors of each entity word is passed to the conditional random field model, and the Viterbi algorithm is used to screen out the automatic entity label sequence with the largest conditional probability as the prediction output of the conditional random field model to obtain the automatic entity recognition result. Based on entity recognition, the text convolutional neural network model is used to extract the relationship between entities to obtain the entities in the text data and the relationship between each entity.

[0015] Preferably, the specific process of extracting knowledge from the preprocessed and annotated image data is as follows:

[0016] The visual prefix of the image data is extracted through the encoder structure of the hierarchical visual transformer model and converted into a semantically embedded title. The semantically embedded title is then input into a Transformer-based text-to-text transformer model. The next frame title is predicted based on the output of the Transformer-based text-to-text transformer model training to generate a complete image description. Finally, the information in the image data is extracted and the relevant equipment failures or environmental anomalies in the image are output in the form of text description, realizing the transformation from image-type data to text description of the scheduling automation system.

[0017] Preferably, the specific method of performing data fusion on the obtained entity is:

[0018] Calculate the similarity between entities and determine whether different entities point to the same object based on the set threshold. The principle of similarity calculation is shown in the following formula:

[0019]

[0020]

[0021] In the formula, S e is the edit distance similarity, S o is the overlap coefficient similarity, S cos is the cosine similarity; S is the similarity between entities; a is the string a in the entity, b is the string b in the entity; E is the edit distance; length(a) is the length of string a, length(b) is the length of string b; I is the number of identical characters in string a and string b; set(a+b) is the set of strings a and b after removing identical characters; V a is the representation of string a in vector space, V b is the representation of string b in vector space; ||V a || represents the vector V a The length of ||V b || represents the vector V b Length;

[0022] Entities expressing the same meaning are merged according to the calculated similarity values.

[0023] Preferably, the automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm, and the specific process of drawing a fault analysis conclusion through reasoning using a knowledge reasoning algorithm based on a combination of multi-level relationship paths is as follows:

[0024] The automated duty knowledge graph is represented by entity-relationship triples, and these triples are embedded into a low-dimensional vector space using the translation embedding model in the vector space.

[0025] The path information between a pair of entities is defined as a segment, and the relationship between the source node and the target node is defined as the target relationship vector; the agent starts from the source node, samples action A according to the random strategy, and observes the reward value R and the next state s′. If the reward value is -1, it is defined as a failure, then the agent returns to the previous state, updates the parameters of the policy network, and repeats the previous steps;

[0026] The reward value is defined as follows:

[0027]

[0028] In the formula, R global is the reward value for each decision step. When the path reaches the target entity, the reward value is +1, otherwise the reward value is -1; R diversity is the diversity reward function, which is used to calculate the similarity between the currently discovered path and the obtained paths; |F| is the number of obtained paths; The obtained relationship chain r1→r2→···→r n The embedding path of P i is the currently discovered path;

[0029] If the target node is reached or the number of action selections exceeds the set maximum inference length, the loop ends, the cumulative reward value is calculated, and the parameters of the policy network are updated; after the above steps, a path from the source node to the target node is obtained;

[0030] The embedding vectors of entities and relations in the path are used as the input of the recurrent neural network model, and iterative learning is performed to obtain a result vector of the semantic combination of the relationship path, and the similarity between the result vector and the target relationship vector is calculated.

[0031] Preferably, the specific method of inferring and matching the node contents of the fault alarm event and the automated duty knowledge graph through the RE2 model to obtain auxiliary decision-making opinions on fault handling is as follows:

[0032] Collect the dispatch automation system fault alarm events and node contents in the automated duty knowledge graph identified in the intelligent analysis module, and annotate the entities and relationships in the fault alarm event text;

[0033] The fault alarm event, the node content in the automated duty knowledge graph, and the fault alarm event content obtained from the fault analysis conclusion are simultaneously input into the embedding layer of the RE2 model, and the text sequence of the fault alarm event is recorded as The text sequence of the node content in the automated duty knowledge graph is recorded as Sequence a and sequence b are simultaneously vector-encoded through the embedding layer and converted into the embedding vector and residual vector corresponding to sequence a and the embedding vector and residual vector corresponding to sequence b;

[0034] The text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph are semantically calculated through N blocks of enhanced residual connection. The dot product operation is used in the alignment layer to calculate the similarity value between the text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph, and the aligned text sequence is output. The similarity matrix between the two text sequences is defined as follows:

[0035] s ij =I(a i ) Τ I(b j )

[0036] Where I is a single-layer feedforward network, i.e., the dot product operation function; s ij is the similarity matrix, a i is the i-th word vector (1≤i≤l) in the fault alarm event text sequence a );b j is the jth word vector (1≤j≤l) in the text sequence of the node content of the automated duty knowledge graph b );

[0037] The aligned text sequence of fault alarm events is recorded as The text sequence of the node content in the aligned automated duty knowledge graph is recorded as Where a' and b' are defined as follows:

[0038]

[0039] Among them, s ik is the similarity value between the i-th word vector of sequence a and any word vector of sequence b; s kj is the similarity between the jth word vector of sequence b and any word vector of sequence a; l a is the length of sequence a; l bis the length of sequence b;

[0040] The fusion layer integrates the input and output sequences of the alignment layer, and aggregates the information of the two sequences through four operations: direct concatenation, combined subtraction, combined dot multiplication, and fusion concatenation. The specific definition is shown in the following formula:

[0041]

[0042] In the formula, F1, F2, F3, and F are single-layer feedforward networks with independent parameters, corresponding to the four operations of direct concatenation, combined subtraction, combined matrix dot multiplication, and fused concatenation; represents element-wise multiplication, Indicates the splicing operation between sequences;

[0043] The fusion layer output of the last block is used as the input of the pooling layer, the two sequences are reduced in dimension, classified through the prediction layer, and the text matching results are output; the specific definition is as follows:

[0044]

[0045] Where M is a multi-layer feedforward network; is the non-normalized prediction score of all classes, where C is the matching category and R is a real number set; u1 is the vector representation of sequence a after dimensionality reduction processing by the pooling layer, and u2 is the vector representation of sequence b after dimensionality reduction processing by the pooling layer; the category with the highest score in the non-normalized prediction score vector calculated by the above formula is selected as the final prediction category p, is the predicted category label, where argmax i Indicates taking the index i that maximizes the prediction score;

[0046] After the matching process between the above fault text sequences, the mapping of corresponding nodes in the automated duty knowledge graph is completed, and then further retrieval and reasoning are performed to obtain the fault emergency plan that best matches the fault alarm event, thereby realizing auxiliary decision-making opinion matching based on the automated duty knowledge graph.

[0047] Preferably, the fault handling auxiliary decision opinions are constructed into a multi-path selection fault abnormality handling process navigation map according to the time sequence of the fault alarm events, and the fault handling response strategy is generated. The specific process of implementing the knowledge-guided response strategy auxiliary execution is as follows:

[0048] For the triples (h, r, t) related to the sequence of fault abnormality handling events in the automated duty knowledge graph, the ALite BERT pre-trained model is used to create semantic representations for the entities and relations in the triples, and low-dimensional embedding vectors are obtained respectively, as shown below:

[0049] h0=Embbeding(h), t0=Embbeding(t), r0=Embbeding(r)

[0050] Where h is the head entity vector, t is the tail entity vector, and r is the relationship vector; h0 is the low-dimensional embedding vector corresponding to the head entity, t0 is the low-dimensional embedding vector corresponding to the tail entity, r0 is the low-dimensional embedding vector corresponding to the relationship, and Embbeding is the vector embedding operation;

[0051] Use the A Lite BERT model to encode the text features of the fault event information and the fault emergency plan information contained in the knowledge graph to obtain the encoding vector v1 and encoding vector v2; calculate the cosine similarity between the two vectors to measure the matching accuracy of the fault emergency plan, and search for the response strategy nodes and relationships most relevant to the fault event based on this;

[0052] The cosine similarity calculation is shown in the following formula:

[0053]

[0054] Where, ||v1|| is the modulus length of the encoding vector v1, ||v2|| is the modulus length of the encoding vector v2; sim(v1,v2) is the cosine similarity value between the encoding vector v1 and the encoding vector v2;

[0055] The graph neural network model is used for fault knowledge reasoning. The entities in the automated duty knowledge graph are used as nodes of the graph neural network model, and the relationships are used as edges. According to the calculation results of the previous step, corresponding weights are assigned to entities and relationships with high similarity as a priori conditions for subsequent node updates; the neighbor node information is integrated, and the embedded representation of each fault node is updated through the forward transmission of messages between nodes. The message transmission mechanism is shown in the following formula:

[0056]

[0057] In the formula, v1 is the fault event node, v2 is the response strategy node; r is the relationship between the two nodes; (l-1) is the (l-1)th layer of the graph neural network model; l is the lth layer of the graph neural network model after information aggregation; N(v1) is the set of all neighbor nodes of the fault event node; f is the activation function; m is the message transmitted from the response strategy node to the fault event node;

[0058] The embedded representation of the fault event node is gradually updated through message forward transmission. The specific definition is as follows:

[0059]

[0060] Where W is the weight matrix; b is the bias value; σ is the ReLU activation function;

[0061] After multiple layers of forward propagation of messages between nodes, the final representation of the faulty node is output by the last layer of the graph neural network model. The high-weight transmission path from the fault event node to all possible response strategy nodes is tracked using the breadth-first search algorithm, and each response strategy is returned in turn. At the same time, system operation instructions are generated, and the corresponding disposal instructions are automatically executed through authorization.

[0062] Through the NoSQL graph database and Cypher query operation, each step of the fault handling strategy is visualized in the Neo4jBrowser interface, and a navigation diagram of the fault handling process of the automation system is constructed according to the sequence of fault handling events;

[0063] Each process navigation entity node is mapped to the corresponding network device operating status, and the execution status of the entity nodes of the fault exception handling process navigation diagram is continuously judged in multiple rounds. The automation system fault exception handling process navigation diagram is driven and lit up according to the above handling steps until the fault exception event handling of the scheduling automation system is completed, realizing the knowledge-guided response strategy assisted execution.

[0064] A method for intelligently handling dispatching automation faults based on knowledge reasoning navigation comprises the following steps:

[0065] The computer room inspection robot obtains multi-source data information of the computer room environment in real time, and sends the multi-source data information of the computer room environment to the backend server as the data source of the intelligent analysis module and the auxiliary decision-making module;

[0066] The data source is divided into numerical data, image data and text data according to the data type, and knowledge is extracted after preprocessing and annotating the image data and text data to obtain entities and the relationship between each entity, and data fusion is performed on the obtained entities. An automated duty knowledge graph is constructed based on the relationship between each entity and the entities after data fusion. Based on the multi-source data information of the computer room environment collected by the environmental perception module of the dispatching automation system, combined with the automated duty knowledge graph, a reasoning model of the fault analysis mechanism of the dispatching automation system is formed, thereby generating a fault alarm event, and the automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm, and the knowledge reasoning environment model is used to reason out the fault analysis conclusion through a knowledge reasoning algorithm based on a combination of multi-level relationship paths;

[0067] The node contents of fault alarm events and automated duty knowledge graphs are input into the RE2 model. The node contents of fault alarm events and automated duty knowledge graphs are inferred and matched through the RE2 model, and combined with the fault analysis conclusions, auxiliary decision-making opinions for fault handling are obtained; the auxiliary decision-making opinions for fault handling are constructed into a multi-path selection fault abnormality handling process navigation map according to the time sequence of fault alarm events, and a fault handling response strategy is generated to realize the auxiliary execution of the response strategy based on knowledge guidance.

[0068] A computer program product comprises a computer program, wherein the computer program implements the steps of the above method when executed by a processor.

[0069] Beneficial effects of the present invention:

[0070] The present invention improves the work efficiency of on-duty personnel, can quickly and accurately identify and analyze faults, greatly reduces the time and errors of manual intervention, improves the efficiency and accuracy of operation and maintenance, greatly improves the operation and maintenance management level of the dispatching automation room, saves management costs, and achieves the purpose of unified management of the dispatching automation room equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a structural schematic diagram of the present invention;

[0072] Figure 2 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0074] Example 1

[0075] A dispatching automation fault intelligent handling system based on knowledge reasoning navigation, such as Figure 1 As shown, it includes:

[0076] The dispatching automation system environment perception module is used to obtain multi-source data information of the computer room environment in real time through the computer room inspection robot (the dispatching automation system will generate various types of data during operation, including environmental data, hardware data, service internal process data and unstructured text data), and send the multi-source data information of the computer room environment to the background server as the data source of the intelligent analysis module and the auxiliary decision-making module. This design can obtain multi-source data information of the computer room environment in real time for efficient use, providing strong support for the intelligent operation and maintenance of the computer room;

[0077] The intelligent analysis module is used to divide the data source into numerical data, image data and text data according to the data type (wherein disk space, running time, CPU utilization, memory utilization, packet loss rate, inverter voltage, output current, noise, temperature and humidity belong to numerical data; image recording belongs to image data; automation standard operation book, emergency plan, safety rules, duty procedures, log files belong to text data; numerical data does not need to be processed), pre-process and annotate the image data and text data, and then extract knowledge to obtain entities and the relationship between each entity, perform data fusion on the obtained entities, and build an automation based on the relationship between each entity and the entities after data fusion. The duty knowledge graph is based on the multi-source data information of the computer room environment collected by the dispatching automation system environment perception module, combined with the automated duty knowledge graph, to form a reasoning model of the dispatching automation system fault analysis mechanism, thereby generating fault alarm events. The automated duty knowledge graph is modeled as a knowledge reasoning environment model using the reinforcement learning path discovery algorithm. The knowledge reasoning environment model is used to reason through the knowledge reasoning algorithm based on the combination of multi-level relationship paths to draw the fault analysis conclusion. This design improves the efficiency and accuracy of data processing, and enhances the adaptive ability and operation and maintenance efficiency of the dispatching automation system. Through real-time monitoring and early warning, it can timely discover and deal with safety hazards in the computer room environment, thereby improving the safety of the computer room.

[0078] The auxiliary decision-making module is used to input the node content of the fault alarm event and the automated duty knowledge graph (the node content of the automated duty knowledge graph refers to the important information units or entities related to the automated duty operation in the knowledge graph, usually including: automated equipment information (such as equipment name, type, status, operating parameters), fault type and characteristics (such as fault analysis rules), fault handling steps, fault records and causes, duty personnel information, alarm levels and related equipment, expert decision rules) into the RE2 model (a model based on residual vectors, word embedding vectors, and encoding vectors), and use the RE2 model to infer and match the node content of the fault alarm event and the automated duty knowledge graph, and combine the fault analysis conclusion to obtain auxiliary decision opinions on fault handling; according to the time sequence of the fault alarm event (fault The alarm event sequence refers to: a causal chain formed by sorting a series of related events according to the time or logical relationship of the fault occurrence; this sequence is usually based on the alarm information recorded by the monitoring system, including the triggering time of the alarm, the event correlation or priority, which is used to determine the occurrence path and evolution process of the fault) The fault handling auxiliary decision-making opinions are constructed as a multi-path selection fault abnormality handling process navigation map, and a fault handling response strategy is generated to realize the knowledge-guided response strategy assisted execution. This design realizes the intelligent analysis and processing of fault alarm events, and provides operation and maintenance personnel with multi-path fault handling response strategies and knowledge-guided response strategy assisted execution, which improves the efficiency and accuracy of fault handling, enhances the intelligence level of the dispatching automation system and the work ability of operation and maintenance personnel.

[0079] In the above technical solution, the method for preprocessing and labeling image data includes: using Gaussian filtering method for denoising, adjusting image contrast by histogram equalization method (improving the inspection effect of alarm area and faulty equipment), and using image labeling tool (the image labeling tool is LabelImg tool) to manually label the alarm area and faulty object in the image recording to obtain image label data (the image label data is used as the visual prefix of the image, which contains the automatic semantic information in the image; the label category is determined by the information contained in the business scenario, such as the name of the faulty equipment, the state of the doors and windows of the computer room, and the working state of the air-conditioning unit);

[0080] The method for preprocessing and labeling text data includes: preprocessing by using data cleaning and standardization methods, removing special symbols and redundant spaces in the text, and marginal information irrelevant to fault alarm events and fault analysis (such as system prompts, status marks, and redundant comments in log files), and using a text labeling tool (the text labeling tool is an INCEpTION labeling tool) to label the plan entities, relationships, and attributes in the text data to obtain text label data (wherein the entity label categories include alarm information, warning level, fault level, and emergency disposal method);

[0081] The above technologies can improve the quality of image data and text data, and can also improve annotation efficiency.

[0082] In the above technical solution, the specific process of extracting knowledge from the preprocessed and annotated text data (including entity recognition and entity relationship extraction of key text information such as automated standard operating instructions, emergency plans, safety details, and duty procedures) is as follows:

[0083] Input the fault alarm event sentence into the BERT model, perform pre-training on the input text through self-supervised learning method, capture the contextual semantic information of entity words, and generate word vector representation (word vector is essentially a low-dimensional embedding vector of entity words);

[0084] The word vector representation is passed to the bidirectional long short-term memory network model (BiLSTM model) to continue capturing the front-end dependency relationship between entity words.

[0085] The sequence of hidden state vectors of each entity word is passed to the conditional random field model. The Viterbi algorithm selects the automatic entity label sequence with the largest conditional probability as the prediction output of the conditional random field model to obtain the automatic entity recognition result. Based on the entity recognition, the text convolutional neural network model is used to extract the relationship between entities to obtain the entities in the text data and the relationship between each entity.

[0086] The above techniques improve the accuracy of entity recognition.

[0087] In the above technical solution, the specific process of extracting knowledge from the preprocessed and annotated image data is as follows:

[0088] The visual prefix of the image data is extracted through the encoder structure of the hierarchical visual transformer model (Swin Transformer model) and converted into a semantically embedded title; the semantically embedded title is then input into the Transformer-based text-to-text transformer model (also known as the T5 model, full name: Text-to-Text Transfer Transformer), and the next frame title is predicted based on the output of the Transformer-based text-to-text transformer model training to generate a complete image description, and finally the information in the image data is extracted, and the relevant equipment failures or environmental anomalies in the image are output in the form of text descriptions, realizing the transformation from image-type data to text descriptions of the dispatching automation system (on this basis, a rule-based method is used to structure the text information obtained from the image-type data to obtain automation entities, and the specific categories include: server information, network equipment operating status, power supply status, security equipment, access control information, air conditioning working status, and environmental status);

[0089] The above technology improves the accuracy of image information extraction and realizes the automatic conversion from image to text.

[0090] In the above technical solution, the specific method for data fusion of the obtained entities (fusion of automated entities expressing the same meaning, elimination of entities with ambiguous content and redundant structures, and enhancement of entity fusion accuracy) is:

[0091] Calculate the similarity between entities and determine whether different entities point to the same object based on the set threshold. The principle of similarity calculation is shown in the following formula:

[0092]

[0093] In the formula, S e is the edit distance similarity, S o is the overlap coefficient similarity, S cos is the cosine similarity; S is the similarity between entities; a is the string a in the entity, b is the string b in the entity; E is the edit distance; length(a) is the length of string a, length(b) is the length of string b; I is the number of identical characters in string a and string b; set(a+b) is the set of strings a and b after removing identical characters; V a is the representation of string a in vector space, V b is the representation of string b in vector space; ||V a || represents the vector V a The length of ||V b || represents the vector V b Length;

[0094] Entities expressing the same meaning are merged according to the calculated similarity values. This can verify the extracted entities to a certain extent and optimize the erroneously extracted entities. Then, based on the standardized entities and the relationships between entities, the mapping of structured data to the knowledge graph is realized, where the entity component of the graph is composed of entities, and the relationship component of the graph is composed of the relationships between entities, thereby constructing an automated duty knowledge graph.

[0095] In the above technical solution, the automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm. The specific process of drawing a fault diagnosis conclusion through reasoning using a knowledge reasoning algorithm based on a combination of multi-level relationship paths is as follows:

[0096] The automated duty knowledge graph is represented by entity-relationship triples, and these triples are embedded into a low-dimensional vector space using a translation embedding model in a vector space (the translation embedding model is the TransE model);

[0097] The path information between a pair of entities is defined as a fragment, and the relationship between the source node (the head entity of the starting triplet of a fault diagnosis rule in the knowledge reasoning environment is called the source node) and the target node is defined as the target relationship vector; the agent (using the reinforcement learning path discovery algorithm, the automated duty knowledge graph is modeled as a knowledge reasoning environment model, and an agent Agent is set) starts from the source node, samples action A according to the random strategy, and observes the reward value R (the agent Agent continuously makes decisions until it reaches the target state, that is, reaches the tail entity of the termination triplet of the fault diagnosis rule, and obtains a decision Reasoning path, while accumulating rewards for each step to get the total expected reward value; training supervised policy network through reward function (a reward function based on goal-oriented rewards, giving positive rewards when the agent successfully reaches the goal, prompting the agent to focus on how to find an effective path to the goal as quickly as possible) to find the path between the starting state and the target state, that is, to select a series of actions to maximize the cumulative expected reward value) and the next state s′. If the reward value is -1, it is defined as failure, then the agent returns to the previous state, updates the parameters of the policy network, and repeats the previous steps;

[0098] The reward value is defined as follows:

[0099]

[0100] In the formula, R global is the reward value for each decision step. When the path reaches the target entity, the reward value is +1, otherwise the reward value is -1; R diversityis the diversity reward function, which is used to calculate the similarity between the currently discovered path and the obtained paths; |F| is the number of obtained paths; The obtained relationship chain r1→r2→···→r n The embedding path of P i is the currently discovered path;

[0101] If the target node is reached or the number of action selections exceeds the set maximum inference length, the loop ends, the cumulative reward value is calculated, and the parameters of the policy network are updated; after the above steps, a path from the source node to the target node is obtained;

[0102] The embedding vectors of entities and relations in the path are used as the input of the recurrent neural network model (RNN network model), and iterative learning is performed to obtain a result vector of the semantic combination of the relationship path, and the similarity between the result vector and the target relationship vector is calculated (the larger the similarity value, the higher the possibility of fault diagnosis through the path, thereby improving the accuracy of fault diagnosis);

[0103] The above technologies realize the integration and optimization of entity information, and improve the intelligence level and accuracy of fault analysis.

[0104] In the above technical solution, the RE2 model (based on residual vectors (the "residual vector" in the RE2 model refers to: the direct connection (i.e., residual connection) transmitted between layers of the RE2 model, which helps to retain the information of the previous layer, so that the RE2 model can better capture and transmit text features in the deep structure), word embedding vectors, and encoding vectors to train the automation operation related text to establish the RE2 model) is used to infer and match the node content of the fault alarm event and the automation duty knowledge graph, and the specific method for obtaining the auxiliary decision-making opinions on fault handling is as follows:

[0105] Collect the dispatch automation system fault alarm events and node contents in the automated duty knowledge graph identified in the intelligent analysis module, and annotate the entities and relationships in the fault alarm event text;

[0106] The fault alarm event, the node content in the automated duty knowledge graph, and the fault alarm event content obtained from the fault analysis conclusion are simultaneously input into the embedding layer of the RE2 model, and the text sequence of the fault alarm event is recorded as The text sequence of the node content in the automated duty knowledge graph is recorded as Sequence a and sequence b are simultaneously vector-encoded by the embedding layer and converted into the embedding vector and residual vector corresponding to sequence a and the embedding vector and residual vector corresponding to sequence b (the dimensions of the embedding vector and residual vector corresponding to sequence a are both 300, and the dimensions of the embedding vector and residual vector corresponding to sequence b are both 300);

[0107] The text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph are semantically calculated through N blocks (block is a network block composed of encoding layer, alignment layer and fusion layer) with enhanced residual connection. The dot product operation is used in the alignment layer to calculate the similarity value between the text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph, and the aligned text sequence is output. The similarity matrix between the two text sequences is defined as follows:

[0108] s ij =I(a i ) Τ I(b j )

[0109] Where I is a single-layer feedforward network, i.e., the dot product operation function; s ij is the similarity matrix, a i is the i-th word vector (1≤i≤l) in the fault alarm event text sequence a );b j is the jth word vector (1≤j≤l) in the text sequence of the node content of the automated duty knowledge graph b );

[0110] The aligned text sequence of fault alarm events is recorded as The text sequence of the node content in the aligned automated duty knowledge graph is recorded as Where a' and b' are defined as follows:

[0111]

[0112] Among them, s ik is the similarity value between the i-th word vector of sequence a and any word vector of sequence b; s kj is the similarity between the jth word vector of sequence b and any word vector of sequence a; l a is the length of sequence a; l b is the length of sequence b;

[0113] The fusion layer integrates the input and output sequences of the alignment layer, and aggregates the information of the two sequences through four operations: direct concatenation, combined subtraction, combined dot multiplication, and fusion concatenation. The specific definition is shown in the following formula:

[0114]

[0115] In the formula, F1, F2, F3, and F are single-layer feedforward networks with independent parameters, corresponding to the four operations of direct concatenation, combined subtraction, combined matrix dot multiplication, and fused concatenation; represents element-wise multiplication, Indicates the splicing operation between sequences;

[0116] The fusion layer output of the last block is used as the input of the pooling layer, the two sequences are reduced in dimension, classified through the prediction layer, and the text matching results are output; the specific definition is as follows:

[0117]

[0118] Where M is a multi-layer feedforward network; is the non-normalized prediction score of all classes, where C is the matching category (C takes the value of 0 or 1), R represents the set of real numbers; u1 represents the vector representation of sequence a after dimensionality reduction processing by the pooling layer, and u2 represents the vector representation of sequence b after dimensionality reduction processing by the pooling layer; the category with the highest score in the non-normalized prediction score vector calculated by the above formula is selected as the final prediction category p, is the predicted category label, where argmax i It means taking the index i that makes the prediction score the largest, that is, the category corresponding to the highest prediction score;

[0119] After the matching process between the above fault text sequences, the mapping of the corresponding nodes in the automated duty knowledge graph is completed, and then further retrieval and reasoning are performed to obtain the fault emergency plan that best matches the fault alarm event, thereby realizing the auxiliary decision-making opinion matching based on the automated duty knowledge graph;

[0120] The above technology realizes the precise matching of fault alarm events with the content of the nodes of the automated duty knowledge graph, improves the intelligence level, accuracy and efficiency of fault handling, reduces the cost of manual intervention, and promotes the continuous optimization of the automated duty knowledge graph.

[0121] In the above technical solution, the auxiliary decision opinions for fault handling are constructed into a multi-path selection fault abnormal handling process navigation map according to the time sequence of the fault alarm events, and the fault handling response strategy is generated. The specific process of implementing the knowledge-guided response strategy auxiliary execution is as follows:

[0122] For the triples (h, r, t) related to the sequence of fault abnormality handling events in the automated duty knowledge graph, the ALite BERT pre-trained model (A Lite BERT pre-trained model is also called the ALBERT model, i.e., a lightweight BERT model) is used to create semantic representations for the entities and relations in the triples, and low-dimensional embedding vectors are obtained respectively, as shown below:

[0123] h0=Embbeding(h), t0=Embbeding(t), r0=Embbeding(r)

[0124] Where h is the head entity vector, t is the tail entity vector, and r is the relationship vector; h0 is the low-dimensional embedding vector corresponding to the head entity, t0 is the low-dimensional embedding vector corresponding to the tail entity, r0 is the low-dimensional embedding vector corresponding to the relationship, and Embbeding is the vector embedding operation;

[0125] Use the A Lite BERT model to encode the text features of the fault event information and the fault emergency plan information contained in the knowledge graph to obtain the encoding vector v1 and encoding vector v2; calculate the cosine similarity between the two vectors to measure the matching accuracy of the fault emergency plan, and search for the response strategy nodes and relationships most relevant to the fault event based on this;

[0126] The cosine similarity calculation is shown in the following formula:

[0127]

[0128] Where, ||v1|| is the modulus length of the encoding vector v1, ||v2|| is the modulus length of the encoding vector v2; sim(v1,v2) is the cosine similarity value between the encoding vector v1 and the encoding vector v2;

[0129] The graph neural network model is used for fault knowledge reasoning. The entities in the automated duty knowledge graph are used as nodes of the graph neural network model, and the relationships are used as edges. According to the calculation results of the previous step, corresponding weights are assigned to entities and relationships with high similarity as a priori conditions for subsequent node updates; the neighbor node information is integrated, and the embedded representation of each fault node is updated through the forward transmission of messages between nodes. The message transmission mechanism is shown in the following formula:

[0130]

[0131] In the formula, v1 is the fault event node, v2 is the response strategy node; r is the relationship between the two nodes; (l-1) is the (l-1)th layer of the graph neural network model; l is the lth layer of the graph neural network model after information aggregation; N(v1) is the set of all neighbor nodes of the fault event node; f is the activation function; m is the message transmitted from the response strategy node to the fault event node;

[0132] The embedded representation of the fault event node is gradually updated through message forward transmission. The specific definition is as follows:

[0133]

[0134] Where W is the weight matrix; b is the bias value; σ is the ReLU activation function;

[0135] After multiple layers of forward propagation of messages between nodes, the final representation of the faulty node is output by the last layer of the graph neural network model. The high-weight transmission path from the fault event node to all possible response strategy nodes is tracked using the breadth-first search algorithm, and each response strategy is returned in turn. At the same time, system operation instructions are generated, and the corresponding disposal instructions are automatically executed through authorization.

[0136] Through the NoSQL graph database and Cypher query operation, each step of the fault handling strategy is visualized in the Neo4jBrowser interface. According to the sequence of fault exception handling events, a navigation map of the fault exception handling process of the automation system is constructed to realize the push of auxiliary decision-making suggestions for fault handling.

[0137] Each process navigation entity node is mapped to the corresponding network device operation status, and the execution status of the entity nodes of the fault abnormality handling process navigation diagram is continuously judged in multiple rounds. The automation system fault abnormality handling process navigation diagram is driven and lit up according to the above handling steps until the fault abnormality event handling of the dispatching automation system is completed, realizing the auxiliary execution of the response strategy based on knowledge guidance;

[0138] The above technology realizes the automation and visualization of the fault handling process by constructing a multi-path selection fault exception handling process navigation map, reducing the cost of manual intervention and improving the reliability and stability of the scheduling automation system.

[0139] Example 2

[0140] A method for intelligent handling of dispatching automation faults based on knowledge reasoning navigation, such as Figure 2 As shown, multi-source data of the computer room environment is obtained in real time as the data source; after preprocessing and annotating the image data and text data, knowledge extraction is performed to obtain entities and the relationship between each entity, and data fusion is performed on the obtained entities to build an automated duty knowledge graph, form a reasoning model of the fault analysis mechanism of the dispatching automation system, generate fault alarm events, build a knowledge reasoning environment, and draw fault analysis conclusions through knowledge reasoning algorithms; through the RE2 model, reasoning and matching of fault alarm events and automated duty knowledge graphs is performed to obtain auxiliary decision-making opinions for fault handling; a navigation map of the fault abnormality handling process is built to generate a fault handling response strategy.

[0141] The specific method of intelligent handling of dispatching automation faults includes the following steps:

[0142] The computer room inspection robot obtains multi-source data information of the computer room environment in real time, and sends the multi-source data information of the computer room environment to the backend server as the data source of the intelligent analysis module and the auxiliary decision-making module;

[0143] The data source is divided into numerical data, image data and text data according to the data type, and knowledge is extracted after preprocessing and annotating the image data and text data to obtain entities and the relationship between each entity, and data fusion is performed on the obtained entities. An automated duty knowledge graph is constructed based on the relationship between each entity and the entities after data fusion. Based on the multi-source data information of the computer room environment collected by the environmental perception module of the dispatching automation system, combined with the automated duty knowledge graph, a reasoning model of the fault analysis mechanism of the dispatching automation system is formed, thereby generating a fault alarm event, and the automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm, and the knowledge reasoning environment model is used to reason out the fault analysis conclusion through a knowledge reasoning algorithm based on a combination of multi-level relationship paths;

[0144] The node contents of fault alarm events and automated duty knowledge graphs are input into the RE2 model. The node contents of fault alarm events and automated duty knowledge graphs are inferred and matched through the RE2 model, and combined with the fault analysis conclusions, auxiliary decision-making opinions for fault handling are obtained; the auxiliary decision-making opinions for fault handling are constructed into a multi-path selection fault abnormality handling process navigation map according to the time sequence of fault alarm events, and a fault handling response strategy is generated to realize the auxiliary execution of the response strategy based on knowledge guidance.

[0145] Example 3

[0146] A computer program product includes a computer program, wherein the computer program implements the steps of the method described in Example 2 when executed by a processor.

[0147] The contents not described in detail in this specification belong to the prior art known to those skilled in the art. It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0148] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims of the invention.

[0152] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A dispatching automation fault intelligent handling system based on knowledge reasoning navigation, characterized in that: It includes: The environment perception module of the dispatching automation system is used to obtain multi-source data information of the computer room environment in real time through the computer room inspection robot, and send the multi-source data information of the computer room environment to the background server as the data source of the intelligent analysis module and the auxiliary decision-making module; The intelligent analysis module is used to divide the data source into numerical data, image data and text data according to the data type, perform knowledge extraction after preprocessing and annotating the image data and text data to obtain entities and the relationship between each entity, perform data fusion on the obtained entities, and construct an automated duty knowledge graph based on the relationship between each entity and the entities after data fusion. Based on the multi-source data information of the computer room environment collected by the dispatching automation system environment perception module and combined with the automated duty knowledge graph, a reasoning model of the fault analysis mechanism of the dispatching automation system is formed, thereby generating a fault alarm event, and using the reinforcement learning path discovery algorithm to model the automated duty knowledge graph as a knowledge reasoning environment model, and using the knowledge reasoning environment model to draw a fault analysis conclusion through reasoning with a knowledge reasoning algorithm based on a combination of multi-level relationship paths; The auxiliary decision-making module is used to input the node contents of the fault alarm events and the automated duty knowledge graph into the RE2 model, and use the RE2 model to infer and match the node contents of the fault alarm events and the automated duty knowledge graph, and combine the fault analysis conclusion to obtain auxiliary decision-making opinions on fault handling; According to the time sequence of fault alarm events, the fault handling auxiliary decision-making opinions are constructed into a multi-path selection fault abnormality handling process navigation map, and the fault handling response strategy is generated to realize the knowledge-guided response strategy auxiliary execution.

2. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The method for preprocessing and labeling image data includes: using a Gaussian filtering method to remove noise, using a histogram equalization method to adjust image contrast, and using an image labeling tool to manually label warning areas and fault objects in image recordings to obtain image label data; The method for preprocessing and labeling text data includes: using data cleaning and standardization methods for preprocessing, removing special symbols and redundant spaces in the text, and marginal information irrelevant to fault alarm events and fault analysis, using text labeling tools to label plan entities, relationships and attributes in the text data, and obtaining text label data.

3. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The specific process of extracting knowledge from the preprocessed and annotated text data is as follows: Input the fault alarm event sentence into the BERT model, perform pre-training on the input text through self-supervised learning method, capture the contextual semantic information of entity words, and generate word vector representation; The word vector representation is passed to the bidirectional long short-term memory network model to continue capturing the previous and next dependencies between entity words; The sequence composed of the hidden state vectors of each entity word is passed to the conditional random field model, and the Viterbi algorithm is used to screen out the automatic entity label sequence with the largest conditional probability as the prediction output of the conditional random field model to obtain the automatic entity recognition result. Based on entity recognition, the text convolutional neural network model is used to extract the relationship between entities to obtain the entities in the text data and the relationship between each entity.

4. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The specific process of extracting knowledge from the preprocessed and annotated image data is as follows: The visual prefix of the image data is extracted through the encoder structure of the hierarchical visual transformer model and converted into a semantically embedded title. The semantically embedded title is then input into a Transformer-based text-to-text transformer model. The next frame title is predicted based on the output of the Transformer-based text-to-text transformer model training to generate a complete image description. Finally, the information in the image data is extracted and the relevant equipment failures or environmental anomalies in the image are output in the form of text description, realizing the transformation from image-type data to text description of the scheduling automation system.

5. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The specific method for performing data fusion on the obtained entity is: Calculate the similarity between entities and determine whether different entities point to the same object based on the set threshold. The principle of similarity calculation is shown in the following formula: In the formula, S e is the edit distance similarity, S o is the overlap coefficient similarity, S cos is the cosine similarity; S is the similarity between entities; a is the string a in the entity, b is the string b in the entity; E is the edit distance; length(a) is the length of string a, length(b) is the length of string b; I is the number of identical characters in string a and string b; set(a+b) is the set of string a and string b after removing the identical characters; V a is the representation of string a in vector space, V b is the representation of string b in vector space; ||V a || represents the vector V a The length of ||V b || represents the vector V b Length; Entities expressing the same meaning are merged according to the calculated similarity values.

6. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm. The specific process of drawing a fault diagnosis conclusion through reasoning using a knowledge reasoning algorithm based on a combination of multi-level relationship paths is as follows: The automated duty knowledge graph is represented by entity-relationship triples, and these triples are embedded into a low-dimensional vector space using the translation embedding model in the vector space. The path information between a pair of entities is defined as a segment, and the relationship between the source node and the target node is defined as the target relationship vector; the agent starts from the source node, samples action A according to the random strategy, and observes the reward value R and the next state s′. If the reward value is -1, it is defined as a failure, then the agent returns to the previous state, updates the parameters of the policy network, and repeats the previous steps; The reward value is defined as follows: In the formula, R global is the reward value for each decision step. When the path reaches the target entity, the reward value is +1, otherwise the reward value is -1; R diversity is the diversity reward function, which is used to calculate the similarity between the currently discovered path and the obtained paths; |F| is the number of obtained paths; The obtained relationship chain r1→r2→···→r n The embedding path of P i is the currently discovered path; If the target node is reached or the number of action selections exceeds the set maximum inference length, the loop ends, the cumulative reward value is calculated, and the parameters of the policy network are updated; After the above steps, a path from the source node to the target node is obtained; The embedding vectors of entities and relations in the path are used as the input of the recurrent neural network model, and iterative learning is performed to obtain a result vector of the semantic combination of the relationship path, and the similarity between the result vector and the target relationship vector is calculated.

7. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: The specific method of inferring and matching the node contents of fault alarm events and automated duty knowledge graphs through the RE2 model and combining the fault analysis conclusions to obtain auxiliary decision-making opinions on fault handling is as follows: Collect the dispatch automation system fault alarm events and node contents in the automated duty knowledge graph identified in the intelligent analysis module, and annotate the entities and relationships in the fault alarm event text; The fault alarm event, the node content in the automated duty knowledge graph, and the fault alarm event content obtained from the fault analysis conclusion are simultaneously input into the embedding layer of the RE2 model, and the text sequence of the fault alarm event is recorded as The text sequence of the node content in the automated duty knowledge graph is recorded as Sequence a and sequence b are simultaneously vector-encoded through the embedding layer and converted into the embedding vector and residual vector corresponding to sequence a and the embedding vector and residual vector corresponding to sequence b; The text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph are semantically calculated through N blocks of enhanced residual connection. The dot product operation is used in the alignment layer to calculate the similarity value between the text sequence of the fault alarm event and the text sequence of the node content in the automated duty knowledge graph, and the aligned text sequence is output. The similarity matrix between the two text sequences is defined as follows: s ij =I(a i ) Τ I(b j ) Where I is a single-layer feedforward network, i.e., the dot product operation function; s ij is the similarity matrix, a i is the i-th word vector (1≤i≤l in the fault alarm event text sequence a );b j is the jth word vector (1≤j≤l) in the text sequence of the node content of the automated duty knowledge graph b ); The aligned text sequence of fault alarm events is recorded as The text sequence of the node content in the aligned automated duty knowledge graph is recorded as Where a' and b' are defined as follows: Among them, s ik is the similarity value between the i-th word vector of sequence a and any word vector of sequence b; s kj is the similarity between the jth word vector of sequence b and any word vector of sequence a; l a is the length of sequence a; l b is the length of sequence b; The fusion layer integrates the input and output sequences of the alignment layer, and aggregates the information of the two sequences through four operations: direct concatenation, combined subtraction, combined dot multiplication, and fusion concatenation. The specific definition is shown in the following formula: In the formula, F1, F2, F3, and F are single-layer feedforward networks with independent parameters, corresponding to the four operations of direct concatenation, combined subtraction, combined matrix dot multiplication, and fused concatenation; represents element-wise multiplication, Indicates the splicing operation between sequences; The fusion layer output of the last block is used as the input of the pooling layer, the two sequences are reduced in dimension, classified through the prediction layer, and the text matching results are output; the specific definition is as follows: Where M is a multi-layer feedforward network; is the non-normalized prediction score of all classes, where C is the matching category and R is a real number set; u1 is the vector representation of sequence a after dimensionality reduction processing by the pooling layer, and u2 is the vector representation of sequence b after dimensionality reduction processing by the pooling layer; the category with the highest score in the non-normalized prediction score vector calculated by the above formula is selected as the final prediction category p, is the predicted category label, where arg max i Indicates taking the index i that maximizes the prediction score; After the matching process between the above fault text sequences, the mapping of corresponding nodes in the automated duty knowledge graph is completed, and then further retrieval and reasoning are performed to obtain the fault emergency plan that best matches the fault alarm event, thereby realizing auxiliary decision-making opinion matching based on the automated duty knowledge graph.

8. The dispatching automation fault intelligent handling system based on knowledge reasoning navigation according to claim 1 is characterized by: According to the time sequence of fault alarm events, the fault handling auxiliary decision opinions are constructed into a multi-path selection fault abnormal handling process navigation map, and the fault handling response strategy is generated. The specific process of implementing the knowledge-guided response strategy auxiliary execution is as follows: For the triples (h, r, t) related to the sequence of fault abnormality handling events in the automated duty knowledge graph, the A LiteBERT pre-trained model is used to create semantic representations for the entities and relations in the triples, and low-dimensional embedding vectors are obtained respectively, as shown below: h0=Embbedding(h), t0=Embbeding(t), r0=Embbeding(r) Where h is the head entity vector, t is the tail entity vector, and r is the relationship vector; h0 is the low-dimensional embedding vector corresponding to the head entity, t0 is the low-dimensional embedding vector corresponding to the tail entity, r0 is the low-dimensional embedding vector corresponding to the relationship, and Embbeding is the vector embedding operation; Use the A Lite BERT model to encode the text features of the fault event information and the fault emergency plan information contained in the knowledge graph to obtain the encoding vector v1 and encoding vector v2; calculate the cosine similarity between the two vectors to measure the matching accuracy of the fault emergency plan, and search for the response strategy nodes and relationships most relevant to the fault event based on this; The cosine similarity calculation is shown in the following formula: Where, ||v1|| is the modulus length of the encoding vector v1, ||v2|| is the modulus length of the encoding vector v2; sim(v1,v2) is the cosine similarity value between the encoding vector v1 and the encoding vector v2; The graph neural network model is used for fault knowledge reasoning. The entities in the automated duty knowledge graph are used as nodes of the graph neural network model, and the relationships are used as edges. According to the calculation results of the previous step, corresponding weights are assigned to entities and relationships with high similarity as a priori conditions for subsequent node updates; the neighbor node information is integrated, and the embedded representation of each fault node is updated through the forward transmission of messages between nodes. The message transmission mechanism is shown in the following formula: In the formula, v1 is the fault event node, v2 is the response strategy node; r is the relationship between the two nodes; (l-1) is the (l-1)th layer of the graph neural network model; l is the lth layer of the graph neural network model after information aggregation; N(v1) is the set of all neighbor nodes of the fault event node; f is the activation function; m is the message transmitted from the response strategy node to the fault event node; The embedded representation of the fault event node is gradually updated through message forward transmission. The specific definition is as follows: Where W is the weight matrix; b is the bias value; σ is the ReLU activation function; After multiple layers of forward propagation of messages between nodes, the final representation of the faulty node is output by the last layer of the graph neural network model. The high-weight transmission path from the fault event node to all possible response strategy nodes is tracked using the breadth-first search algorithm, and each response strategy is returned in turn. At the same time, system operation instructions are generated, and the corresponding disposal instructions are automatically executed through authorization. Through the NoSQL graph database and Cypher query operation, each step of the fault handling strategy is visualized in the Neo4jBrowser interface, and a navigation diagram of the fault handling process of the automation system is constructed according to the sequence of fault handling events; Each process navigation entity node is mapped to the corresponding network device operating status, and the execution status of the entity nodes of the fault exception handling process navigation diagram is continuously judged in multiple rounds. The automation system fault exception handling process navigation diagram is driven and lit up according to the above handling steps until the fault exception event handling of the scheduling automation system is completed, realizing the knowledge-guided response strategy assisted execution.

9. A method for intelligent handling of dispatching automation faults based on knowledge reasoning navigation, characterized in that: It follows the steps: The computer room inspection robot obtains multi-source data information of the computer room environment in real time, and sends the multi-source data information of the computer room environment to the backend server as the data source of the intelligent analysis module and the auxiliary decision-making module; The data source is divided into numerical data, image data and text data according to the data type, and knowledge is extracted after preprocessing and annotating the image data and text data to obtain entities and the relationship between each entity, and data fusion is performed on the obtained entities. An automated duty knowledge graph is constructed based on the relationship between each entity and the entities after data fusion. Based on the multi-source data information of the computer room environment collected by the environmental perception module of the dispatching automation system, combined with the automated duty knowledge graph, a reasoning model of the fault analysis mechanism of the dispatching automation system is formed, thereby generating a fault alarm event, and the automated duty knowledge graph is modeled as a knowledge reasoning environment model using a reinforcement learning path discovery algorithm, and the knowledge reasoning environment model is used to reason out the fault analysis conclusion through a knowledge reasoning algorithm based on a combination of multi-level relationship paths; The node contents of the fault alarm events and the automated duty knowledge graph are input into the RE2 model. The RE2 model is used to infer and match the node contents of the fault alarm events and the automated duty knowledge graph, and combined with the fault analysis conclusion, auxiliary decision-making opinions on fault handling are obtained. According to the time sequence of fault alarm events, the fault handling auxiliary decision-making opinions are constructed into a multi-path selection fault abnormality handling process navigation map, and the fault handling response strategy is generated to realize the knowledge-guided response strategy auxiliary execution.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in claim 9 are implemented.

Images