CN115827888A - Fault prediction method for complex equipment - Google Patents

Abstract

The invention discloses a fault prediction method of complex equipment, which comprises the following steps: preprocessing the acquired information data of the complex equipment to be detected; wherein the information data includes timing information of the fault information; performing entity extraction, attribute extraction and relationship extraction on the preprocessed information data; performing knowledge fusion on the entities, the attributes and the relationship among the entities; constructing an equipment fault time sequence knowledge graph based on the entity information, the entity attributes and the incidence relation among the entities after knowledge fusion; constructing an equipment fault prediction model based on the equipment fault time sequence knowledge graph; and performing fault prediction on the complex equipment through the equipment fault prediction model. The invention can obtain more accurate predicted fault information, thereby better assisting operation and maintenance personnel to accurately maintain the equipment, effectively reducing the occurrence of fault shutdown events caused by untimely maintenance of the equipment and improving the production efficiency of the equipment.

Description

Fault prediction method for complex equipment

Technical Field

The invention relates to the technical field of fault prediction, in particular to a fault prediction method of complex equipment.

Background

At present, in the operation and maintenance work of complex equipment, a large amount of manpower is required to be invested to find out fault equipment and equipment fault points, and meanwhile, compiling of technical reports and regular maintenance work are required. The shutdown phenomenon caused by equipment maintenance often occurs, which not only affects the production efficiency, but also often causes huge economic loss due to service interruption. Therefore, the possibility of reducing the problem of equipment failure and finding out the equipment failure in time is one of the necessary means for effectively improving the production efficiency of enterprises and saving manpower and time cost.

There are many failure prediction methods for complex devices in the field of operation and maintenance services, and they can be classified into two categories, i.e. a method based on statistical analysis and a method based on artificial intelligence:

the equipment failure prediction method based on the statistical analysis is to determine the correlation information between two or more variables through the statistical analysis and judge the possibility of equipment failure occurrence based on the correlation information. For stable data, the statistical analysis method has higher operation efficiency, and the prediction model can be realized in a shorter time. However, these statistical models often require more stringent constraints on equipment health. On the other hand, the device failure prediction method based on statistical analysis often uses one or more signals such as vibration, sound, current, etc. as input signals, and performs failure prediction by methods such as signal processing, statistical analysis, etc., however, most failure prediction methods so far require a large amount of historical data to perform training and learning processes of prediction models, but when many signals are mixed, feature extraction of device failure information is often difficult. Although such a fault prediction method based on statistical analysis often has a good effect in a laboratory test environment, it often focuses mainly on intermediate diagnosis of a single fault, and cannot perform comprehensive fault analysis on the entire device, such as short circuit of a motor rotor, and the like, and it cannot effectively perform perfect fault diagnosis and prediction on complex devices, and under the actual complex device operation condition, it is often difficult to satisfy corresponding condition assumptions, and there is a great limitation.

The equipment fault prediction method based on artificial intelligence is to learn to dynamically extract characteristics and optimize parameters from a large amount of data by an artificial intelligence method without human intervention so as to fit corresponding input and output data. Support Vector Regression (SVR) is a common machine learning method, and for small sample data, it often obtains better experimental results, but SVR computational complexity may show m as the number of samples m increases ² Or m ³ Due to the increase of the vector quantity, the vector quantity is not suitable for large sample data, and no proper method is provided for selecting a kernel function to support the sensitivity of the vector quantity to an error boundary at present, and meanwhile, the vector quantity is sensitive to a penalty factor and has weak self-adaption capability; the neural network does not need to establish a mathematical model of the system, has self-learning property and stronger adaptability, robustness and nonlinear mapping, so the neural network is widely applied to equipment failure prediction, such as a BP (back propagation) neural network, a convolutional neural network, a Recurrent Neural Network (RNN), a long short-term memory (LSTM) neural network and the like, but the neural network cannot explain the reasoning process and reasoning basis of the neural network, cannot provide necessary inquiry for a user, cannot work when the data is insufficient, simultaneously changes all the characteristics of the problems into numbers, changes all the reasoning into numerical calculation, and has the effect of losing information; the neural network expert system is a new knowledge expression system, it is a low layer number model, the information processing is carried on through the interaction among a large number of simple processing elements (nodes), it can combine logical reasoning with numerical operation, utilize learning function, associative memory function, distributed parallel information processing function of the neural network, solve the uncertain knowledge in the diagnostic system to represent, obtain and carry on the reasoning question in parallel, through studying experience sample, store expert's knowledge in the network in the form of weight and threshold value, and utilize the information retention of the network to finish the imprecise diagnosis and reasoning, has simulated the expert by experience, intuition but not the reasoning process of the complicated calculation well, but its limit is limited by brain scienceThe existing research results do not establish a complete and mature theoretical system, have thick strategy colors and are immature in interface with the traditional computing technology.

Disclosure of Invention

In view of this, the present invention provides a method for predicting a failure of a complex device, so as to solve the above technical problems.

The invention discloses a fault prediction method of complex equipment, which comprises the following steps:

preprocessing the acquired information data of the complex equipment to be detected; wherein the information data includes timing information of the fault information;

performing entity extraction, attribute extraction and relationship extraction on the preprocessed information data;

performing knowledge fusion on the entities, the attributes and the relationship among the entities;

constructing an equipment fault time sequence knowledge graph based on the entity information, the entity attributes and the incidence relation among the entities after knowledge fusion;

constructing an equipment fault prediction model based on the equipment fault time sequence knowledge graph;

and performing fault prediction on the complex equipment through the equipment fault prediction model.

Further, the knowledge fusion of the entities, the attributes, and the relationships between the entities includes:

removing repeated data in the entity information, the entity attributes and the association relationship among the entities;

unifying entity information, integrating entity attributes, fusing entities and complementing entity missing attributes of the data subjected to duplicate removal; the entity fusion comprises the steps that entity information of entities with the same attribute, the same associated entity and different entity information is fused, and the entity information needing to be fused is separated by '|'; and the entity missing attribute completion comprises the steps of carrying out verification reasoning on the entity current attribute and the associated entity, searching the most similar entity set in the similar entity set, and deducing the missing attribute value of the same entity set based on the entity current attribute and the associated entity.

Further, the constructing an equipment fault time sequence knowledge graph based on the entity information, the entity attributes and the association relationship among the entities after knowledge fusion includes:

selecting fault components, fault phenomena, fault reasons and effective operation time to construct a time sequence knowledge graph, importing all time sequence quadruples into a graph database, and drawing an equipment fault time sequence knowledge graph; the time sequence quadruplet comprises a fault component, a fault phenomenon, a fault reason and effective running time;

converting each time sequence quadruple into two triples, namely a first triplet and a second triplet; the first triple is a fault component, effective operation time and fault phenomenon; the second triple is a fault phenomenon, effective operation time and a fault reason; the first triad represents that the fault component has a fault phenomenon in the effective operation time; the second triplet indicates that a fault phenomenon occurs at the effective operating time due to a fault reason;

and importing the first triple and the second triple into a graph database eo4j to obtain a corresponding equipment fault time sequence knowledge graph.

Further, the constructing an equipment fault prediction model based on the equipment fault timing knowledge graph comprises:

preprocessing the equipment fault time sequence knowledge graph to obtain a training set, a verification set and a test set;

acquiring a predicted fault event and the predicted probability thereof at the moment t;

training an equipment fault prediction model based on a training set based on a predicted fault event and the predicted probability of the predicted fault event at the moment t;

and performing single-step fault prediction and multi-step fault prediction based on the trained equipment fault prediction model.

Further, the preprocessing the equipment fault timing knowledge graph to obtain a training set, a verification set and a test set includes:

defining a fault event as (s, r, o, t), wherein s is a fault component, t is time, r is a fault phenomenon, and o is a fault reason;

numbering equipment failure event quadruple sequences corresponding to an equipment failure time sequence knowledge graph, respectively acquiring entity sets after duplication of ' failure components ', ' failure phenomena ' and ' failure causes ', respectively numbering and assigning values from 0,1 and 1 in sequence, and respectively storing corresponding ' numbers: the system comprises an entity dictionary, a component selection unit, a component analysis unit and a component analysis unit, wherein the entity dictionary is used for setting unit time interval delta t at the same time, converting effective operation time of the component, sequencing based on event time sequence, interpolating the sequenced data sequence, carrying out interpolation processing from time 0, and fitting the quadruple time if the quadruple time with a fault deviates from the planned sample time; after the processing is finished, the data set is divided from time 0, and the data set is divided into a training set, a verification set and a test set according to the proportion.

Further, the obtaining of the predicted fault event at the time t and the predicted probability thereof includes:

define all failure events G = { G = { ₀ ,...,G _t-1 Where k is more than or equal to 0 and less than or equal to t-1, G _k Representing a fault event set at the moment k, constructing an equipment fault prediction model based on a time sequence knowledge graph and RE-NET, and predicting a fault event at the moment t through fault event information at the moment t-m, t-m +1, \ 8230;

the conditional probabilities of the failed component s, the failure phenomenon r and the failure cause o in each event at time t are calculated and are respectively represented as p(s) _t |G _t-m:t-1 )、p(r _t |s _t ,G _t-m:t-1 ) And p (o) _t |s _t ,r _t ,G _t-m:t-1 ) Based on the conditional probability corresponding to each element in the event, G is obtained _t-m:t-1 Conditional fault event(s) _t ,r _t ,o _t ) Probability of (c): p(s) _t ,r _t ,o _t |G _t-m:t-1 )＝p(s _t |G _t-m:t-1 )p(r _t |s _t ,G _t-m:t-1 )p(o _t |s _t ,r _t ,G _t-m:t-1 )

After the calculation of the prediction probability of all the events is finished, the fault event set G at the time t is obtained _t All the following reasonsAnd (4) the set of fault events completes the screening of the fault events at the time t, and the predicted fault events and the predicted probability thereof at the time t are obtained.

Further, cross entropy of multiple classes is adopted as a loss function of the equipment failure prediction model:

wherein G is _t ' is a set of predicted failure events at time t, G _t Is the set of actual fault events at time t, E _i ＝(s _i ,r _i ,o _i T) is a fault event belonging to the set G of predicted fault events _t ' with actual failure event set G _t Alpha and beta are respectively loss term weight coefficients, p(s) _i |G _t-m:t-1 )、p(r _i |s _i ,G _t-m:t-1 )、p(o _i |s _i ,r _i ,G _t-m:t-1 ) Respectively, predicted events(s) _i ,r _i ,o _i T) failed component s _i Failure phenomenon r _i Cause of failure o _i The predicted conditional probability of (2).

Further, the training process of the equipment fault prediction model is as follows:

input device failure time-series knowledge map sequences, i.e. training sets: { G ₀ ,...,G _T }；

Setting a time to be predicted t = m;

if T is less than or equal to T, m samples { G }are obtained _t-m ,...,G _t-1 Predicting p (s | G) corresponding to each fault event _t-m:t-1 )、p(r|s,G _t-m:t-1 ) And p (o | s, r, G) _t-m:t-1 ) Let the set of predicted fault events be G' _t And combining the actual event set G _t Calculating the loss value L at the time t _t ；

Returning a loss value and correcting a model;

training times reach verification conditions, and model evaluation and model parameter correction are carried out based on the verification set;

adding 1 to the time t to be measured; until T is greater than T; and finally evaluating the equipment fault prediction model based on the test set.

Further, the single-step fault prediction is performed based on the trained equipment fault prediction model, and includes:

firstly, acquiring a fault event time sequence G from an equipment fault time sequence knowledge graph _t-m ,...,G _t-1 Predicting the fault event at the future time T by taking the predicted fault event as input, and calculating the residual running time delta T of the equipment for generating the single-step fault prediction event based on the latest historical fault event time T-1, the actual running time of the current equipment and the time conversion set in the data preprocessing ₁ Predicting the probability p (s, r, o | G) of each fault event of the equipment at the moment t-1+ i based on the obtained equipment fault prediction model _t-m:t-1 ) The fault event is predicted according to its corresponding probability p (s, r, o | G) _t-m:t-1 ) Sorting from big to small, and selecting the first k fault events {(s) ₁ ,r ₁ ,o ₁ ,t),...,(s _k ,r _k ,o _k T), rejecting events in which the fault information and the fault reason are 0, and taking the events as a candidate prediction event set;

aiming at a fault event corresponding to a key concern component, converting a fault component number, a fault information number and a fault reason number in the event into specific information, and feeding back the specific information to a user by combining the obtained prediction event; and screening the fault events corresponding to the non-key attention components based on a set probability threshold, eliminating the fault events with the event probability smaller than the probability threshold, converting the serial numbers of the fault components, the serial numbers of the fault information and the serial numbers of the fault reasons in the events into specific information, and feeding the specific information back to a user.

Further, the multi-step fault prediction is performed based on the trained equipment fault prediction model, and includes:

continuously taking the single-step fault prediction result as a fault event set corresponding to the prediction time, performing multiple single-step fault predictions, thereby predicting the fault events at the future (t-1) + n moments, and feeding back the obtained multiple-step fault prediction events to a specific user;

the calculation formula of the residual running time of the single-step and multi-step fault prediction event occurrence equipment is as follows:

ΔT _n ＝(t-1+n)*Δt-T ₁

wherein, delta T _n Predicting the residual operation time of the equipment for n steps of fault prediction events, wherein n is the number of prediction steps and is more than or equal to 1, namely the predicted fault event can occur after the equipment operates again by delta T, T is single step fault prediction time, delta T is set unit time interval, and T is the set unit time interval ₁ The current actual running time of the equipment.

Due to the adoption of the technical scheme, the invention has the following advantages:

1. the equipment fault time sequence knowledge graph is simple to construct and can be fused with the traditional knowledge graph. The equipment failure time sequence knowledge graph body comprises four parts, namely a failure part, failure information, a failure reason and effective occurrence time (actual equipment operation time), and the effective operation time is simultaneously merged into the knowledge graph in a relational form, so that the time sequence knowledge graph is stored in a graph database in a traditional knowledge graph form.

2. The equipment fault prediction is fast and accurate, the monitoring of the heavy point parts and the multi-step fault prediction are supported, and operation and maintenance personnel are assisted to carry out accurate operation and maintenance on the equipment, so that the occurrence of fault shutdown events caused by untimely maintenance of the equipment is effectively reduced, and the production efficiency of the equipment is improved. The device fault prediction model takes device fault Event time sequence data as input, consists of a Multi-relation Aggregator (Multi-Relational Aggregator), a self-circulation Event Encoder (current Event Encoder) and a self-circulation Event decoder (current Event Encoder), and fully considers the time sequence and the relevance among fault events. Based on the obtained equipment fault prediction model and the historical fault event time sequence data, single-step and multi-step fault prediction results of the equipment, including fault components, fault phenomena, fault reasons and effective operation time, can be quickly obtained, and meanwhile, based on the customized key concerned components and the prediction probability threshold, more accurate and concerned prediction fault information can be obtained, so that operation and maintenance personnel can be better assisted to accurately maintain the equipment, the occurrence of fault shutdown events caused by untimely maintenance of the equipment is effectively reduced, a series of economic losses caused by equipment shutdown are avoided, and the production efficiency of the equipment is improved.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments described in the embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings.

FIG. 1 is a schematic diagram of a fault timing knowledge graph construction process according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating a local structure of an apparatus failure timing knowledge graph according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a local structure of an operation and maintenance knowledge graph of a device according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an equipment failure prediction model according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a training process of an equipment failure prediction model according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of a failure prediction process according to an embodiment of the present invention;

fig. 7 is a flowchart illustrating a failure prediction method for a complex device according to an embodiment of the present invention.

Detailed Description

The present invention will be further described with reference to the accompanying drawings and examples, it being understood that the examples described are only some of the examples and are not intended to limit the invention to the embodiments described herein. All other embodiments available to those of ordinary skill in the art are intended to be within the scope of the embodiments of the present invention.

Referring to fig. 7, the present invention provides an embodiment of a failure prediction method for a complex device, which includes:

s1, preprocessing acquired information data of complex equipment to be detected; the information data comprises time sequence information of fault information;

s2, performing entity extraction, attribute extraction and relationship extraction on the preprocessed information data;

s3, carrying out knowledge fusion on each entity, attributes and the relation among the entities;

s4, constructing an equipment fault time sequence knowledge graph based on the information of each entity, the attributes of the entities and the incidence relation among the entities after knowledge fusion;

s5, constructing an equipment fault prediction model based on the equipment fault time sequence knowledge graph;

and S6, carrying out fault prediction on the complex equipment through an equipment fault prediction model.

In this embodiment:

1. fault timing knowledge graph construction

The construction of the complex equipment fault time sequence knowledge graph can be divided into five core processes of data acquisition, data preprocessing, knowledge extraction, knowledge fusion and time sequence knowledge graph drawing, and the construction process of the complex equipment fault time sequence knowledge graph is shown in figure 1.

1.1 data acquisition

The data acquisition comprises various semi-structured forms, web pages, unstructured text data and partially structured data, and meanwhile, the data acquisition also needs to pay attention to the time sequence information of the acquired fault information.

1.2 data preprocessing

The data preprocessing refers to preprocessing the acquired text data based on the NLP technology, and particularly processing unstructured data. For Data collected from each Data source, an ETL tool (e.g. a button) is used to sequentially perform Data extraction (for different Data source characteristics, different extraction methods are selected to be extracted from each Data source into an Operational Data Store (ODS)), data cleaning (dirty Data and incomplete Data are filtered, conversion of inconsistent Data and Data granularity is performed, and time sequence unification is performed at the same time) and loading (Data writing into a Data Warehouse (DW) after Data cleaning is completed) operations, so as to obtain a required Data format, thereby performing subsequent knowledge extraction and fusion work.

1.3 knowledge extraction

The knowledge extraction is to extract corresponding knowledge units from unstructured and semi-structured data through automation and manual technologies, and comprises entity extraction, attribute extraction and relationship extraction. The entity extraction identifies and extracts entity information from the text data set, wherein the entity information comprises equipment (device), "equipment model (device _ type)," fault phenomenon (fault _ php), "fault component (fault _ part)," component type (part _ type), "fault cause (fault _ recovery)," and the like; relation extraction extracts association relations among entities, but the relation extraction is different from the relation extraction of a common knowledge graph, and a time sequence is required to be extracted as the association relations among the entities, so that the association relations of dependencies, actions, time and the like among the entities are obtained based on the relation extraction, the relationships of the dependencies, the actions and the like are defined as a fixed relation, and the time is defined as a time relation; the attribute extraction is to construct an attribute list of an entity, and acquire attribute information of entities such as 'equipment', 'fault component', 'fault reason', and the like from a plurality of data sources, wherein the attribute information comprises 'equipment number' (equipment attribute), 'component installation position' (fault component attribute), 'fault consequence' (fault reason attribute), 'error correction suggestion' (fault reason attribute), so that the entity is completely explained, and the time is effective running time and refers to the actual running time of the equipment or the component (the downtime does not belong to the effective running time).

1.4 knowledge fusion

The knowledge fusion is to complete data integration on data (main entities, relations and guest entities) of knowledge information from different sources, perform a series of operations such as duplicate removal, disambiguation, verification inference, updating and the like under the same standard frame, remove rough data, take accurate data of the rough data for utilization, and enhance the expression capability and the logicality of a fault information internal knowledge base of the equipment. The method specifically comprises the steps of duplicate removal (removal of duplicate data), unification of entity information (entity name synonym conversion, full name conversion for short and the like), entity attribute integration (for the same entity, duplicate attribute information deduplication and fusion of different attribute information are performed on the attribute of the entity), entity fusion (for the entities with the same attribute, the same associated entity and different entity information, the entity information of the entities is fused, the entity information needing to be fused is separated by '|', if two different 'fault phenomena' entity association 'fault reasons' are the same as the 'fault component' entity and the 'fault phenomena' are different, the two 'fault phenomena' can be fused), entity missing attribute completion (based on the verification reasoning on the current attribute of the entity and the associated entity, the entity set which is most similar to the entity set is searched in the same entity set, and the missing attribute value of the entity set is deduced based on the verification reasoning), and the like.

1.5 time series knowledge mapping

A quadruple (s, r, o, t), where s is the subject entity, r is the relationship, o is the object entity, and t is the time. The time sequence knowledge graph drawing is to construct a time sequence knowledge graph by selecting fault components, fault phenomena, fault reasons and effective operation time based on single equipment, other entities (such as component types, equipment and the like) are constructed in the form of a traditional knowledge graph and serve as the self knowledge graph of the equipment, and the self knowledge graph can be integrated into time sequence quadruples (fault components, fault reasons, fault phenomena and effective operation time), namely the fault components generate the fault phenomena due to the fault reasons in the effective operation time, and the equipment fault time sequence quadruples are shown in a table 1. When the equipment failure time sequence four-tuple data set is imported into a graph database Neo4j to draw an equipment failure time sequence knowledge graph, a four-tuple (a failure component, a failure reason, a failure phenomenon and effective operation time) is converted into two triplets (a failure component, effective operation time, a failure phenomenon) and a (failure phenomenon, effective operation time and a failure reason) which respectively represent that the failure component has the failure phenomenon in the effective operation time and the failure phenomenon has the failure reason in the effective operation time, and the obtained triplets are imported into the Neo4j to obtain the corresponding equipment failure time sequence knowledge graph, wherein the local structure of the equipment failure time sequence knowledge graph is shown in fig. 2, and the equipment failure time sequence knowledge graph can be further fused with the traditional equipment operation and maintenance knowledge graph (a main entity, a fixed relation and a guest entity) conveniently, so that the equipment operation and maintenance knowledge graph containing time sequence information and general information is obtained, and the local structure of the equipment operation and maintenance knowledge graph is shown in fig. 3.

TABLE 1 Equipment Fault timing Quaternary set of data

2. Model construction and fault prediction: the method mainly comprises four steps of data preprocessing, model construction, model training and fault prediction.

2.1 data preprocessing

One fault event is defined as (s, r, o, t), where s is the faulty component, t is the time, r is the fault phenomenon, and o is the fault cause. Numbering the equipment fault event quadruple sequence corresponding to the equipment fault time sequence knowledge graph, respectively acquiring an entity set after duplicate removal of a ' fault component ', ' fault phenomenon ' and ' fault reason ', respectively numbering and assigning values from 0,1 and 1 in sequence (for the ' fault phenomenon ' and the ' fault reason ', a code 0 is reserved to represent no fault phenomenon and no fault reason), and respectively storing the corresponding ' numbers: the method includes the steps of a physical dictionary, setting unit time intervals delta T (such as 1200, 2400, 3600, unit seconds), converting the effective operation time of the component, if the converted time series is (0, 1, \8230; k, \8230; m), sorting the converted time series based on the event time series, interpolating the sorted data series, performing interpolation processing from time 0 (if the fault component is numbered 101, no fault event occurs at k time, inserting (101, 0, k), and if the fault component 101 is not, no, k, \\ delta T), namely, the equipment component 101 operates normally at k T time of the effective operation time of the component, no fault information and no fault reason exist, and if the fault four time and the planning sample time are deviated, performing correction on the four tuple time (the planning sample time is a time tuple of the four tuple of the equipment component is a latest time group). After the processing is finished, the data set can be divided from time 0, and divided into a training set, a verification set and a test set according to the proportion of 80%, 10% and 10%. See table 2 for an example of a training set sequence for equipment failure.

TABLE 2 Equipment Fault training set sequence example

2.2 model construction

Define all failure events G = { G = { ₀ ,...,G _t-1 Where k is more than or equal to 0 and less than or equal to t-1, G _k The method comprises the steps of representing a fault Event set at the moment k, constructing an equipment fault prediction model based on a time sequence knowledge graph and RE-NET (Recurrent Event Network), predicting fault events (fault components s, fault phenomena r and fault reasons o) at the moment t through fault Event information at the moment t-m, t-m +1, \8230;, and t-1, and enabling the structure of the equipment fault model to be shown in FIG. 4.

Wherein, the local information refers to the set of fault events at the corresponding moment (the specific input of the model, G) _t-m To G _t-1 ) The global information refers to the global representation of event information from the starting state to the current state (along with the continuous input of the fault event set at each moment, the model is continuously accumulated and updated); the Aggregator is a Multi-Relational Aggregator (Multi-Relational Aggregator), which is composed of a Multi-layer Relational Graph Convolutional Network (RGCN), supports Multi-layer aggregation, and can aggregate information from Multi-Relational (Multi-Relational) neighbors and Multi-hop (Multi-hop) neighbors; the Encoder is a self-circulation Event Encoder (a current Event Encoder) which comprises 3 RNN (Recurrent Neural Network) modules and is used for respectively encoding fault Event global information, fault Event (fault component s) local information and fault Event (fault component s, fault phenomenon r) local information at corresponding time; the decoder is a self-circulation Event decoder (current Event Encoder) which further calculates the failure component s, failure phenomenon r and failure reason o in each Event at time t by transmitting the code obtained by the Encoder to a Multi-Layer Perceptron (MLP) decoderConditional probabilities, respectively denoted as p(s) _t |G _t-m:t-1 )、p(r _t |s _t ,G _t-m:t-1 ) And p (o) _t |s _t ,r _t ,G _t-m:t-1 ) Based on the conditional probability corresponding to each element in the event, G is obtained _t-m:t-1 Conditional failure event(s) _t ,r _t ,o _t ) The probability of (c) is shown in the following formula, and after the calculation of the prediction probabilities of all the events is completed, the fault event set G at the time t is obtained _t And (4) the screening of the fault events at the time t can be completed based on the corresponding selection rule after the collection of all the fault events is performed, and the predicted fault events at the time t and the predicted probability thereof are obtained.

p(s _t ,r _t ,o _t |G _t-m:t-1 )＝p(s _t |G _t-m:t-1 )p(r _t |s _t ,G _t-m:t-1 )p(o _t |s _t ,r _t ,G _t-m:t-1 )

2.3 model training

The prediction of the faulty component s can be regarded as a multi-classification task (where each class corresponds to each faulty component, i.e. obtaining the probability of each component failing), and similarly, the prediction of the fault phenomenon r given (faulty component s) and the prediction of the fault cause o given (faulty component s, fault phenomenon r) can be regarded as a multi-classification task, so that the cross entropy of the multi-classification is used as a model loss function, as shown in the following formula.

Wherein G is _t ' is a set of predicted failure events at time t, G _t Is the set of actual fault events at time t, E _i ＝(s _i ,r _i ,o _i T) is a fault event belonging to the set G of predicted fault events _t ' with actual failure event set G _t Alpha and beta are respectively loss term weight coefficients, and can be selected based on actual conditions, and p(s) _i |G _t-m:t-1 )、p(r _i |s _i ,G _t-m:t-1 )、p(o _i |s _i ,r _i ,G _t-m:t-1 ) Respectively, predicted events(s) _i ,r _i ,o _i T) failed component s _i Failure phenomenon r _i Cause of failure o _i The predicted conditional probability of (2). Defining a set of input training samples as a sequence of event sets G ₀ ,...,G _T And there is an equipment failure prediction model training process as shown in fig. 5.

2.4 Fault prediction

The equipment fault prediction aims at predicting fault events (fault components, fault information and fault reasons) which may occur in the equipment at a certain time in the future, and can be divided into a single-step fault prediction mode and a multi-step fault prediction mode, wherein a fault prediction process is shown in fig. 6. When the single-step fault prediction of the equipment is carried out, firstly, a fault event time sequence { G ] is obtained from an equipment fault time sequence knowledge map _t-m ,...,G _t-1 Predicting the fault event at the future time T by taking the predicted fault event as input, and calculating the residual running time delta T of the equipment for generating the single-step fault predicted event based on the latest historical fault event time T-1, the actual running time of the current equipment and the time conversion rule set in the data preprocessing ₁ Predicting probability p (s, r, o | G) of each fault event of the equipment at the moment of t-1+ i based on the obtained equipment fault prediction model _t-m:t-1 ) The fault event is predicted according to its corresponding probability p (s, r, o | G) _t-m:t-1 ) Sorting from big to small, and selecting the first k fault events {(s) ₁ ,r ₁ ,o ₁ ,t),...,(s _k ,r _k ,o _k T), rejecting events in which the fault information and the fault reason are both 0 (if the fault information and the fault reason are both 0, no fault occurs, that is, non-fault events are rejected) as a set of alternative prediction events; for a fault event corresponding to a key attention component (an event in which a component in an alternative prediction event set is a key attention component, wherein the key attention component set is set by a user), converting a fault component number, a fault information number and a fault reason number in the event into specific information, and feeding back the specific information to the specific user by combining the obtained prediction event; for the fault event corresponding to the non-important part (the part in the alternative prediction event set is not the event of the important part), further screening is carried out based on a set probability threshold (such as 0.1)And selecting and eliminating fault events with the event probability smaller than the probability threshold, converting the fault component number, the fault information number and the fault reason number in the event into specific information, and feeding the specific information back to the user.

When the multi-step fault prediction of the equipment is carried out, the single-step fault prediction result is continuously used as a fault event set corresponding to the prediction time in the input, the multi-step fault prediction is carried out for many times, so that the fault events at the future (t-1) + n moments are predicted, and the obtained multi-step fault prediction events are fed back to a specific user.

The calculation formula of the residual running time of the single-step and multi-step fault prediction event occurrence equipment is shown as the following formula:

ΔT _n ＝(t-1+n)*Δt-T ₁

wherein, delta T _n Predicting the residual operation time of the equipment for n steps of fault prediction events, wherein n is the number of prediction steps and is more than or equal to 1, namely the predicted fault event can occur after the equipment operates again by delta T, T is single step fault prediction time (the operation timestamp of the predicted equipment is divided by the unit time interval), delta T is set unit time interval, and T is the unit time interval ₁ The current actual running time of the equipment.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims (10)

1. A failure prediction method for a complex device, comprising:

preprocessing the acquired information data of the complex equipment to be detected; wherein the information data includes timing information of the fault information;

performing entity extraction, attribute extraction and relationship extraction on the preprocessed information data;

performing knowledge fusion on the entities, the attributes and the relationship among the entities;

constructing an equipment fault time sequence knowledge graph based on the entity information, the entity attributes and the incidence relation among the entities after knowledge fusion;

constructing an equipment fault prediction model based on the equipment fault time sequence knowledge graph;

and performing fault prediction on the complex equipment through the equipment fault prediction model.

2. The method of claim 1, wherein the knowledge fusion of each of the entities, attributes, and relationships between entities comprises:

removing repeated data in the entity information, the entity attributes and the association relationship among the entities;

unifying entity information, integrating entity attributes, fusing entities and complementing entity missing attributes of the data subjected to duplicate removal; the entity fusion comprises the steps that entity information of entities with the same attribute, the same associated entity and different entity information is fused, and the entity information needing to be fused is separated by '|'; and the entity missing attribute completion comprises the steps of carrying out verification reasoning on the entity current attribute and the associated entity, searching the most similar entity set in the similar entity set, and deducing the missing attribute value of the same entity set based on the entity current attribute and the associated entity.

3. The method according to claim 1, wherein constructing an equipment fault timing knowledge graph based on the entity information, the entity attributes and the association relationship among the entities after knowledge fusion comprises:

selecting fault components, fault phenomena, fault reasons and effective operation time to construct a time sequence knowledge graph, importing all time sequence quadruples into a graph database, and drawing an equipment fault time sequence knowledge graph; the time sequence quadruplet comprises a fault component, a fault phenomenon, a fault reason and effective running time;

converting each time sequence quadruplet into two triplets, namely a first triplet and a second triplet; the first triple is a fault component, effective operation time and fault phenomenon; the second triple is a fault phenomenon, effective operation time and a fault reason; the first triad represents that the fault component has a fault phenomenon in the effective operation time; the second triplet indicates that a fault phenomenon occurs at the effective operating time due to a fault reason;

and importing the first triple and the second triple into a graph database eo4j to obtain a corresponding equipment fault time sequence knowledge graph.

4. The method of claim 1, wherein constructing an equipment fault prediction model based on the equipment fault timing knowledgegraph comprises:

preprocessing the equipment fault time sequence knowledge graph to obtain a training set, a verification set and a test set;

acquiring a predicted fault event and the predicted probability thereof at the moment t;

training an equipment fault prediction model based on a training set based on the predicted fault event and the predicted probability of the predicted fault event at the time t;

and performing single-step fault prediction and multi-step fault prediction based on the trained equipment fault prediction model.

5. The method of claim 4, wherein the preprocessing the equipment failure timing knowledge graph to obtain a training set, a validation set, and a test set comprises:

defining a fault event as (s, r, o, t), wherein s is a fault component, t is time, r is a fault phenomenon, and o is a fault reason;

numbering equipment failure event quadruple sequences corresponding to an equipment failure time sequence knowledge graph, respectively acquiring entity sets after duplication of ' failure components ', ' failure phenomena ' and ' failure causes ', respectively numbering and assigning values from 0,1 and 1 in sequence, and respectively storing corresponding ' numbers: the system comprises an entity dictionary, a component selection unit, a component analysis unit and a component analysis unit, wherein the entity dictionary is used for setting unit time interval delta t at the same time, converting effective operation time of the component, sequencing based on event time sequence, interpolating the sequenced data sequence, carrying out interpolation processing from time 0, and fitting the quadruple time if the quadruple time with a fault deviates from the planned sample time; after the processing is finished, the data set is divided from time 0, and the data set is divided into a training set, a verification set and a test set according to the proportion.

6. The method of claim 5, wherein the obtaining the predicted fault event and the predicted probability thereof at time t comprises:

define all failure events G = { G = { ₀ ,...,G _t-1 Where k is more than or equal to 0 and less than or equal to t-1, G _k Representing a fault event set at the moment k, constructing an equipment fault prediction model based on a time sequence knowledge graph and RE-NET, and predicting a fault event at the moment t through fault event information at the moment t-m, t-m +1, \ 8230;

the conditional probabilities of the failed component s, the failure phenomenon r and the failure cause o in each event at time t are calculated and are respectively represented as p(s) _t |G _t-m:t-1 )、p(r _t |s _t ,G _t-m:t-1 ) And p (o) _t |s _t ,r _t ,G _t-m:t-1 ) Based on the conditional probability corresponding to each element in the event, G is obtained _t-m:t-1 Conditional fault event(s) _t ,r _t ,o _t ) Probability of (c):

p(s _t ,r _t ,o _t |G _t-m:t-1 )＝p(s _t |G _t-m:t-1 )p(r _t |s _t ,G _t-m:t-1 )p(o _t |s _t ,r _t ,G _t-m:t-1 )

after the calculation of the prediction probability of all the events is finished, the fault event set G at the time t is obtained _t And (5) collecting all the fault events, and finishing screening of the fault events at the time t to obtain the predicted fault events at the time t and the predicted probability thereof.

7. The method of claim 6, wherein cross-entropy for multiple classes is employed as a loss function for the equipment failure prediction model:

wherein, G' _t Is a set of predicted fault events at time t, G _t Is the set of actual fault events at time t, E _i ＝(s _i ,r _i ,o _i T) is a fault event belonging to the set G 'of predicted fault events' _t And actual failure event set G _t Alpha and beta are respectively the weight coefficient of the loss term, p(s) _i |G _t-m:t-1 )、p(r _i |s _i ,G _t-m:t-1 )、p(o _i |s _i ,r _i ,G _t-m:t-1 ) Respectively, predicted events(s) _i ,r _i ,o _i T) failed component s _i Failure phenomenon r _i And cause of failure o _i The predicted conditional probability of (2).

8. The method of claim 7, wherein the training process of the equipment failure prediction model is as follows:

input device failure time-series knowledge map sequences, i.e. training sets: { G ₀ ,...,G _T }；

Setting a time to be predicted t = m;

if T is less than or equal to T, m samples { G }are obtained _t-m ,...,G _t-1 Predicting p (s | G) corresponding to each fault event _t-m:t-1 )、p(r|s,G _t-m:t-1 ) And p (o | s, r, G) _t-m:t-1 ) Let the set of predicted fault events be G' _t And combining the actual event set G _t Calculating the loss value L at the time t _t ；

Returning a loss value and correcting a model;

training times reach verification conditions, and model evaluation and model parameter correction are carried out based on the verification set;

adding 1 to the time t to be measured; until T is greater than T; and finally evaluating the equipment fault prediction model based on the test set.

9. The method of claim 8, wherein the conducting single-step fault prediction based on the trained equipment fault prediction model comprises:

firstly, acquiring a fault event time sequence G from an equipment fault time sequence knowledge graph _t-m ,...,G _t-1 Predicting the fault event at the future time T by taking the predicted fault event as input, and calculating the residual running time delta T of the equipment for generating the single-step fault prediction event based on the latest historical fault event time T-1, the actual running time of the current equipment and the time conversion set in the data preprocessing ₁ Predicting the probability p (s, r, o | G) of each fault event of the equipment at the moment t-1+ i based on the obtained equipment fault prediction model _t-m:t-1 ) The fault event is predicted according to its corresponding probability p (s, r, o | G) _t-m:t-1 ) Sorting from big to small, and selecting the first k fault events {(s) ₁ ,r ₁ ,o ₁ ,t),...,(s _k ,r _k ,o _k T), rejecting events in which the fault information and the fault reason are both 0, and taking the events as a candidate prediction event set;

aiming at a fault event corresponding to a key concern component, converting a fault component number, a fault information number and a fault reason number in the event into specific information, and feeding back the specific information to a user by combining the obtained prediction event; and screening the fault events corresponding to the non-key attention components based on a set probability threshold, eliminating the fault events with the event probability smaller than the probability threshold, converting the serial numbers of the fault components, the serial numbers of the fault information and the serial numbers of the fault reasons in the events into specific information, and feeding the specific information back to a user.

10. The method of claim 8, wherein the conducting a multi-step fault prediction based on the trained equipment fault prediction model comprises:

continuously taking the single-step fault prediction result as a fault event set corresponding to the prediction time, performing multiple single-step fault predictions, thereby predicting the fault events at the future (t-1) + n moments, and feeding back the obtained multiple-step fault prediction events to a user;

the calculation formula of the residual running time of the single-step and multi-step fault prediction event generation equipment is as follows:

ΔT _n ＝(t-1+n)*Δt-T ₁

wherein, delta T _n Predicting the residual operation time of the equipment for n steps of fault prediction events, wherein n is the number of prediction steps and is more than or equal to 1, namely the predicted fault event can occur after the equipment operates again by delta T, T is single step fault prediction time, delta T is set unit time interval, and T is the set unit time interval ₁ The current actual running time of the equipment.

Images