CN-121976918-A - Fan abnormality monitoring and operation detection optimizing method based on knowledge graph and deep learning

CN121976918ACN 121976918 ACN121976918 ACN 121976918ACN-121976918-A

Abstract

The invention relates to the technical field of operation detection of wind turbine equipment, and discloses a fan abnormality monitoring and operation detection optimizing method based on knowledge graph and deep learning, wherein wind turbine operation data are collected through a SCADA system, and wind turbine state data are collected through a CMS system; the method comprises the steps of obtaining operation and maintenance data of a wind turbine generator, constructing a TCN-BiGRU-attribute network model to analyze the multi-dimensional data, outputting abnormal types and confidence degrees of model prediction, constructing a double-layer knowledge graph, taking the abnormal types as input, extracting a feasible scheme according to a graph reasoning algorithm, outputting operation and detection strategy triples, inputting the operation and detection strategy triples and the multi-dimensional data into a large language model, carrying out multi-round semantic reasoning and strategy optimization, and outputting an optimal operation and detection strategy. The method has the advantages of realizing high accuracy of anomaly identification and high reliability of operation and detection decision, having good expandability and practical value, and being suitable for various scenes such as intelligent wind power plant operation and maintenance, remote state monitoring and intelligent scheduling.

Inventors

LIU YOUBO
Gong Haochen
LIU ZIHAO
He Shuaijia
ZHENG WEI

Assignees

四川大学
广州新源禾信息科技有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251204

Claims (5)

1. The fan abnormality monitoring and operation optimizing method based on knowledge graph and deep learning is characterized by comprising the steps of, S1, acquiring running data of a wind turbine through an SCADA system and acquiring state data of the wind turbine through a CMS system, acquiring running and maintenance data of the wind turbine, and performing data preprocessing on the running data, the state data and the running and maintenance data to obtain multi-dimensional data with aligned data; S2, analyzing the multidimensional data by constructing a TCN-BiGRU-attribute network model, so as to monitor the abnormality of the fan and output the abnormality type and confidence of model prediction; S3, constructing a double-layer knowledge graph, wherein the double-layer knowledge graph comprises an anomaly-fault knowledge graph and a fault-operation detection strategy knowledge graph; s4, based on a double-layer knowledge graph, taking the abnormal type as input, extracting a feasible scheme according to a graph reasoning algorithm, and outputting an operation and detection strategy triplet; S5, inputting the operation and detection strategy triples and the multidimensional data into a large language model, carrying out multi-round semantic reasoning and strategy optimization, and outputting the optimal operation and detection strategy under the multi-constraint condition.
2. The fan anomaly monitoring and operation detection optimizing method based on knowledge graph and deep learning of claim 1, wherein the sequential convolution network TCN of the deep neural network in S2 firstly carries out convolution transformation on an original time sequence to extract local dynamic characteristics, and the input time sequence is set as follows: ; wherein, T is the length of the input sequence, d is the characteristic dimension of each moment; The output of the time-sequential convolutional network TCN is: ; In the formula, The ith convolution kernel parameter is the first layer; For the layer 1 TCN convolutional layer at time step A feature vector on the first and second images; K is the size of a convolution kernel, d is an expansion factor, and sequential characteristics of layer-by-layer progression are extracted through stacking a plurality of convolution layers; The extracted time sequence features are input into a bidirectional gating cycle unit BiGRU to acquire deeper forward and backward time dependence, and the output of the bidirectional gating cycle unit BiGRU is as follows: ; Calculating similarity scores of the features and the global context at each moment through an attention mechanism, and distributing attention weights: ; In the formula, Importance score for the t-th time step on the final decision; For the weight matrix of the attention layer, Query vectors for learning attentiveness; Is a bias vector for the attention layer; an attention weight indicating the time t; A context representation obtained for the weighting; The prediction probability is: ; wherein y is a classification target; For the weight matrix of the output layer, Is the bias vector of the output layer.
3. The fan anomaly monitoring and operation detection optimizing method based on knowledge graph and deep learning of claim 1 is characterized in that the hiding state updating process of a single GRU unit is as follows: ; ; ; ; In the formula, To update the door; is a reset gate; is a candidate new state; is the current state; for the hidden state vector at the previous moment, Input weight matrices for update gate, reset gate and candidate states respectively, Hidden layer weight matrices for update gate, reset gate and candidate states respectively, The bias terms of the update gate, reset gate and candidate state, respectively, and the Hadamard product.
4. The fan anomaly monitoring and operation and detection optimizing method based on knowledge graph and deep learning of claim 1, wherein the specific step of S3 comprises, Constructing a double-layer knowledge graph through a relation extraction model RERE, and outputting each relation to any fan operation detection data sequence c by the relation extraction model RERE Probability of (2): ; In the formula, A coded representation of sequence head-end [ CLS ] characters for BERT (Bidirectional Encoder Representations from Transformers, BERT); A classification weight matrix for the relationship r; a bias term for relationship r; Activating a function for sigmod; The relationship extraction model re model then extracts, for each identified relationship r, its corresponding subject and object (s, o), the inputs as: and encoded into a matrix by BERT The relation extraction model re locates the sequence boundaries by four pointers: ; Wherein, the To predict probability distributions belonging to boundary type k at various positions in the sequence, The function is activated for the purpose of sigmod, As a weight matrix for the pointer k, As a biasing term for the pointer k, Four pointers respectively representing a subject starting point, a subject ending point, an object starting point and an object ending point; the total objective function of the relation extraction model re is a likelihood function that maximizes all training samples: ; Wherein, the C i is a fan operation detection data sequence of the ith training sample, and T i is a target relation and entity boundary triplet set corresponding to c i ; The conditional probability chain rule is utilized to decompose into: 。
5. The fan anomaly monitoring and operation detection optimizing method based on the knowledge graph and the deep learning of claim 1, wherein the step S4 specifically comprises, Setting the knowledge graph as a set with entities Sum edge set Is a directed graph of (1): ; deducing the current possible fault types to form an initial input set according to the knowledge graph: ; Wherein, the F i is the i candidate fault type node, and m is the number of candidate fault types; Is provided with As a candidate policy node, And For its initial embedding, the semantic matching between the fault and the policy is scored normalized by: ; In the formula, For sharing a linear transformation matrix; Is a trainable attention weight vector; Is a fault node Is a candidate policy node Is the feature concatenation operation; For a policy in a multi-policy candidate path Relative to The semantic importance of the graph of the fault; defining all paths from fault f to policy s as: ; In the formula, For the currently considered failed node, s is a candidate operation and detection strategy node corresponding to the failure f, In the knowledge graph, from the fault node The set of all possible paths from the departure to the policy node s, p i is the set The i-th specific path in (a), n is the number of paths; each path P is represented as a sequence of nodes The path scoring function is defined as: ; In the formula, Is the ith edge on the path Semantic weights of (2); Scoring semantic credibility or context consistency of the path intermediate entity nodes; each policy node ultimately gets the maximum path score from the path correlation score for the current fault: ; Calculating comprehensive candidate strategy scores through a double fusion mechanism: ; In the formula, To adjust parameters; Semantic matching weights obtained through a graph attention mechanism; is a strategy Total correlation score relative to fault f; Introducing policy resource matching degree index Modeling resource cost and feasibility constraints are performed in combination with policies: ; In the formula, The weight regulation and control coefficient between the semantic drive and the resource drive; A comprehensive evaluation score for the policy s after considering both semantic relevance and resource constraints.

Description

Fan abnormality monitoring and operation detection optimizing method based on knowledge graph and deep learning Technical Field The invention relates to the technical field of operation detection of wind turbine equipment, in particular to a fan abnormality monitoring and operation detection optimizing method based on knowledge graph and deep learning. Background With the rapid development of wind power technology and the centralized grid-connected operation of a large-scale wind power plant, the operation stability and reliability of the wind turbine generator have become keys for guaranteeing the safe and efficient operation of the wind power plant. The wind turbine generator system has a complex structure and comprises a plurality of highly coupled components such as blades, a main shaft, a gear box and the like, and the components are easy to generate abnormal conditions such as abrasion, cracks and the like when operating in complex environments such as high altitude, high wind speed, strong corrosion and the like for a long time. Failure to discover and handle in time may lead to equipment damage, economic loss, and even grid disturbances. In order to improve the running state sensing and fault early warning capability of the wind turbine, a state monitoring system (Condition Monitoring System, CMS) and a monitoring and data acquisition system (Supervisory Control and Data Acquisition, SCADA) are widely deployed in a wind turbine. The system mainly aims at signal acquisition and spectrum analysis of vibration, temperature and the like of key rotating components and is used for identifying early mechanical faults, and the SCADA system is used for realizing real-time acquisition and remote monitoring of the running parameters (such as wind speed, rotating speed, current, voltage, power and the like) of the whole machine. However, the data of the current CMS and SCADA systems are generally distributed on different platforms, the sampling frequency is inconsistent, the dimension is complex, the data semantics are heterogeneous, and an effective data fusion and association interpretation mechanism is lacked. The traditional abnormality detection and fault diagnosis mainly depend on threshold setting, expert rules or experience models, and have the problems of high false alarm rate, poor expandability, difficulty in adapting to nonlinear complex faults and the like. In recent years, the deep learning method has strong modeling capability in the aspect of processing multi-source time sequence data, in particular to a time sequence convolution network, a two-way gating circulation unit and an Attention mechanism combined structure (TCN-BiGRU-Attention), and can automatically extract key features from complex multi-source time sequence data and realize high-precision abnormality identification. However, depth models often lack good interpretability, and are difficult to support for subsequent engineering decisions and inspection procedures. The knowledge graph technology provides a structural expression means for equipment knowledge modeling, and can represent the association among the wind power equipment composition structure, the fault mode, the operation experience and the operation detection strategy through the form of the entity-relation-entity triplet, so as to support the logic reasoning based on semantic paths and embedding. However, the current knowledge graph construction often depends on manual extraction rules, has limited coverage, and cannot be fully linked with real-time monitoring data. The large language model (Large Language Model, LLM) is a deep learning-based natural language processing technique, and has achieved significant results in a variety of general-purpose natural language processing tasks. In the face of the field of wind turbine operation and detection, technical manuals and technical terms and concepts in the field of fan operation and detection are more, the application range is limited, and complex requirements are difficult to adapt. The traditional large language model can have the conditions of braiding answers, answer errors, machine illusions and the like, so that fan operation and inspection staff cannot timely process maintenance problems, and operation and inspection efficiency is reduced. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides a fan abnormity monitoring and operation detection optimizing method based on knowledge graph and deep learning. The invention aims at realizing the fan abnormity monitoring and operation detection optimizing method based on knowledge graph and deep learning, which comprises the following steps, S1, acquiring running data of a wind turbine through an SCADA system and acquiring state data of the wind turbine through a CMS system, acquiring running and maintenance data of the wind turbine, and performing data preprocessing on the running data, the state data and the running and maintenanc