CN116991681A

CN116991681A - NLP-combined fly ash fusion processing system abnormality report identification method and server

Info

Publication number: CN116991681A
Application number: CN202311256647.6A
Authority: CN
Inventors: 张淼; 刘向金; 刘玉坤
Original assignee: Beijing Zhongke Runyu Environmental Protection Technology Co ltd
Current assignee: Beijing Zhongke Runyu Environmental Protection Technology Co ltd
Priority date: 2023-09-27
Filing date: 2023-09-27
Publication date: 2023-11-03
Anticipated expiration: 2043-09-27
Also published as: CN116991681B

Abstract

The embodiment of the application provides a fly ash fusion processing system abnormal report identification method and a server combined with NLP, which are used for carrying out abnormal trigger node positioning to generate a reference system abnormal event sequence when a system abnormal report text is detected, carrying out redundant optimization on the reference system abnormal event sequence according to the abnormal report text to generate an optimized system abnormal event sequence and a corresponding context description paragraph sequence, limiting the optimized system abnormal event corresponding to each context relation characteristic in the optimized system abnormal event sequence, and generating an abnormal event limiting area for characterizing the trigger abnormal node in the system abnormal report text. Therefore, through abnormal trigger node positioning, redundant optimization and extraction of the contextual characteristics, the abnormal event limiting area in the system abnormal report text can be more accurately identified and understood, so that the fly ash fusion processing system can be conveniently and timely inspected and examined by combining the abnormal event limiting area, and the running efficiency and stability of the fly ash fusion processing system are improved.

Description

NLP-combined fly ash fusion processing system abnormality report identification method and server

Technical Field

The application relates to the technical field of report mining, in particular to a report identification method and a server for an abnormality of a fly ash fusion processing system combined with NLP.

Background

In industrial production, particularly for complex systems such as fly ash fusion processing systems, the generation and processing of system anomaly report text is very important. They provide for the recording and delivery of critical information such as equipment operating conditions, fault causes, repair measures, etc. However, since these report texts generally contain a large amount of expertise and complex semantic information, efficient parsing and understanding by Natural Language Processing (NLP) techniques is required.

Currently, there is no efficient way to locate and extract critical anomaly event defined regions in system anomaly report text. For example, the prior art generally only performs simple keyword extraction and entity recognition, but cannot accurately locate specific nodes and events triggering anomalies, and often ignores complex relationships between context features and entities in the report text, resulting in insufficiently comprehensive understanding and analysis of the anomalies. And since the system abnormality report text generally contains a large amount of redundant information, the related art lacks a corresponding optimization process.

Disclosure of Invention

In view of the above, an object of the present application is to provide a method and a server for identifying abnormal reports of a fly ash fusion processing system combined with NLP.

According to a first aspect of the present application, there is provided a fly ash fusion processing system anomaly report identification method in combination with NLP, applied to a server, the method comprising:

when detecting a system abnormality report text for a fly ash fusion processing system, positioning a system abnormality trigger node of the system abnormality report text to generate a reference system abnormality event sequence;

performing redundancy optimization on the reference system abnormal event sequence according to the system abnormal report text, and generating an optimized system abnormal event sequence corresponding to the reference system abnormal event sequence and a context description paragraph sequence corresponding to the optimized system abnormal event sequence; the context description paragraph sequence comprises W context description paragraphs; a context description paragraph is obtained by extracting a text description paragraph from the system exception report text according to an optimization system exception event in the optimization system exception event sequence; w is an integer greater than 0;

extracting the context relation of each of the W context description paragraphs respectively, and generating context relation data corresponding to the system abnormal trigger node; the contextual data includes G contextual features; g is a positive integer not greater than W, and the contextual characteristics are used for representing the relationship characteristics among the identified entities in the contextual description paragraphs, wherein the relationship characteristics comprise anomaly description keywords and relationship vectors among the anomaly description keywords;

And respectively limiting the optimization system abnormal events corresponding to the G context relation features in the optimization system abnormal event sequence to generate G abnormal event limiting areas representing the system abnormal trigger nodes in the system abnormal report text.

In a possible implementation manner of the first aspect, before the step of locating a system anomaly trigger node of the system anomaly report text and generating a sequence of reference system anomaly events, the method further includes:

acquiring a first example exception report text for updating parameters of a first natural language processing network and first priori annotation data of a real exception triggering event representing the first example exception report text; the first example exception report text is generated after initializing a system exception trigger node of the initial example exception report text, and the natural language processing network adopts a converter network converter;

positioning the first example exception report text according to the first natural language processing network, and generating an estimated system exception event of the first example exception report text aiming at the system exception trigger node;

Determining training error values of an estimated system abnormal event of the first example abnormal report text and a real abnormal trigger event of the first example abnormal report text, and generating training error values of the first natural language processing network;

according to the training error value of the first natural language processing network, carrying out parameter updating on the first natural language processing network to generate a first network layer weight information updating result;

if the first natural language processing network after the first network layer weight information updating result representation parameter updating meets the first network deployment requirement, taking the first natural language processing network meeting the first network deployment requirement as a second natural language processing network;

analyzing the functional parameter layer of the second natural language processing network to generate functional parameter layer analysis data;

if the function parameter layer analysis data represents that a decoder is arranged in the function parameter layer of the second natural language processing network, replacing the decoder with a first encoder carrying a wandering window parameter;

performing parameter migration configuration for the first encoder according to the network layer weight information of the decoder to generate a second encoder;

Outputting a second natural language processing network comprising the second encoder as an abnormal trigger node positioning network for positioning the system abnormal trigger node of the system abnormal report text.

In a possible implementation manner of the first aspect, the performing redundancy optimization on the reference system abnormal event sequence according to the system abnormal report text, generating an optimized system abnormal event sequence corresponding to the reference system abnormal event sequence and a context description paragraph sequence corresponding to the optimized system abnormal event sequence, includes:

performing redundancy optimization on the abnormal event sequence of the reference system according to a redundancy optimization strategy of event filtering and classifying to generate a candidate system abnormal event sequence; the candidate system abnormal event sequence comprises a reference system abnormal event Lx; x is a positive integer not greater than D; d, representing the total number of abnormal trigger nodes in the candidate system abnormal event sequence;

performing correction optimization on the reference system abnormal event Lx based on the system abnormal report text according to a paragraph context correction optimization network to generate an optimization system abnormal event Rx and a context description paragraph corresponding to the optimization system abnormal event Rx;

When D optimizing system abnormal events are obtained, carrying out redundancy optimization on the D optimizing system abnormal events according to the redundancy optimization strategy;

outputting the optimized system abnormal event after the redundancy optimization as an optimized system abnormal event sequence corresponding to the reference system abnormal event sequence, and outputting the context description paragraph corresponding to the optimized system abnormal event after the redundancy optimization as a context description paragraph sequence corresponding to the optimized system abnormal event sequence.

In a possible implementation manner of the first aspect, the reference system anomaly event sequence is obtained when positioning a system anomaly trigger node of the system anomaly report text according to an anomaly trigger node positioning network;

the reference system abnormal event sequence comprises F reference system abnormal events; f is an integer greater than 0; the abnormal trigger node positioning network is further used for determining estimated system abnormal events corresponding to each reference system abnormal event in the F reference system abnormal events respectively; the redundant optimization strategy according to the event filtering and classifying carries out redundant optimization on the abnormal event sequence of the reference system to generate a candidate abnormal event sequence of the system, which comprises the following steps:

According to keyword rules mapped by the redundant optimization strategies of event filtering and classifying, respectively filtering text content of each reference system abnormal event in the F reference system abnormal events to generate F filtered system abnormal events;

according to the probability values of F estimated system abnormal events, carrying out order arrangement on the F filtered system abnormal events to generate an abnormal event order sequence;

outputting the filtered system abnormal events with the maximum probability value in the abnormal event sequence as first filtered system abnormal events, and outputting (F-1) filtered system abnormal events except the first filtered system abnormal events in the abnormal event sequence as a first temporary system abnormal event sequence;

determining text description similarity between the first filtered system exception event and each filtered system exception event in the first temporary system exception event sequence;

if the first temporary system abnormal event sequence has a common system abnormal event with text description similarity larger than preset similarity, the first filtered system abnormal event is reserved, and the common system abnormal event is cleaned in the first temporary system abnormal event sequence;

In the cleaned first temporary system abnormal event sequence, taking the filtered system abnormal event with the largest estimated system abnormal event as a second filtered system abnormal event, and taking the filtered system abnormal events except the second filtered system abnormal event as a second temporary system abnormal event sequence;

the second filtered system abnormal event is reserved, redundancy optimization is carried out on the common system abnormal event in the second temporary system abnormal event sequence according to the text description similarity between the second filtered system abnormal event and each filtered system abnormal event in the second temporary system abnormal event sequence until the second temporary system abnormal event sequence after redundancy optimization is an empty set, and the reserved first filtered system abnormal event and the reserved second filtered system abnormal event are output as an iterative system abnormal event sequence;

and expanding each filtered system abnormal event in the iterative system abnormal event sequence according to the keyword rule to generate a candidate system abnormal event sequence.

In a possible implementation manner of the first aspect, the correcting and optimizing network according to the context of the paragraph corrects and optimizes the reference system abnormal event Lx based on the system abnormal report text, and generates an optimized system abnormal event Rx and a context description paragraph corresponding to the optimized system abnormal event Rx, which includes:

According to the paragraph context deviation correcting and optimizing network, based on the system abnormality report text, carrying out paragraph context deviation parameter estimation on the reference system abnormality event Lx to generate a first paragraph context deviation parameter;

correcting the deviation of the abnormal event Lx of the reference system according to the context deviation parameter of the first paragraph, and generating a first abnormal event of the optimization system corresponding to the abnormal event Lx of the reference system;

if the first paragraph context deviation parameter belongs to a threshold parameter interval, outputting a first optimization system abnormal event corresponding to the reference system abnormal event Lx as an optimization system abnormal event Rx corresponding to the reference system abnormal event Lx;

and filtering text content of the system abnormal report text according to the positioning information of the optimizing system abnormal event Rx, and generating a context description paragraph corresponding to the optimizing system abnormal event Rx.

In a possible implementation manner of the first aspect, after obtaining a first optimization system abnormal event corresponding to the reference system abnormal event Lx, the method further includes:

if the first paragraph context deviation parameter does not belong to the threshold parameter interval, according to the paragraph context deviation correcting optimization network, based on the system abnormality report text, carrying out paragraph context deviation parameter estimation on a first optimization system abnormality event corresponding to the reference system abnormality event Lx, and generating a second paragraph context deviation parameter;

Correcting the first optimization system abnormal event corresponding to the reference system abnormal event Lx according to the second paragraph context deviation parameter, generating a second optimization system abnormal event corresponding to the reference system abnormal event Lx until the second paragraph context deviation parameter belongs to the threshold parameter interval, and outputting the second optimization system abnormal event corresponding to the reference system abnormal event Lx as an optimization system abnormal event Rx corresponding to the reference system abnormal event Lx.

In a possible implementation manner of the first aspect, before optimizing the network according to the paragraph context deskew, the method further includes:

acquiring a second example exception report text for updating parameters of a basic deviation correcting optimization network and second priori label data of prior positioning information of a system exception trigger node representing the second example exception report text;

according to the basic deviation correction optimization network, performing paragraph context deviation parameter learning on the second example exception report text, and generating training paragraph context deviation parameters of the second example exception report text;

correcting the deviation of the positioning information of the second example exception report text according to the context deviation parameter of the training paragraph, and generating training positioning information of the second example exception report text;

According to the training positioning information and the priori positioning information, parameter updating is carried out on the basic deviation correcting optimization network, and a second network layer weight information updating result is generated;

and if the basic deviation rectifying optimization network with the updated characteristic parameters of the second network layer weight information updating result meets the second network deployment requirement, outputting the basic deviation rectifying optimization network meeting the second network deployment requirement as the paragraph context deviation rectifying optimization network.

In a possible implementation manner of the first aspect, the extracting a context relation from each of the W context description paragraphs, to generate context relation data corresponding to the system exception triggering node includes:

determining a target context description paragraph from the W context description paragraphs;

based on a prediction network in a target context relation extraction model, carrying out abnormal category prediction on the target context description paragraph, and generating an abnormal category probability value corresponding to the target context description paragraph;

based on the self-coding network in the target context extraction model, carrying out context feature prediction on the target context description paragraph, and generating context features corresponding to the target context description paragraph;

And outputting the context relation characteristic corresponding to the target context description section as context relation data corresponding to the system abnormal trigger node if the abnormal category probability value corresponding to the target context description section is larger than the set probability value.

In a possible implementation manner of the first aspect, before extracting the model according to the target context, the method further includes:

acquiring training text data for updating parameters of a basic natural language processing network and priori labeling data corresponding to the training text data;

according to a third example exception report text in the training text data and third priori label data representing priori context relation characteristics of the third example exception report text, parameter updating is carried out on a shared network function layer and a self-coding network in the basic natural language processing network, and a first context relation extraction model is generated;

according to a fourth example exception report text in the training text data and fourth priori annotation data representing a real exception trigger event of the fourth example exception report text, carrying out parameter updating on a shared network function layer and a prediction network in the first context extraction model to generate a second context extraction model;

Performing context feature prediction on the third example exception report text according to the second context extraction model, and generating training context features corresponding to the third example exception report text;

the text entity of the third example exception report text is walked, and the walked text entity is output as a target text entity;

outputting the probability value of the target text entity in the priori context feature indicated by the third priori label data as a first probability value, and outputting the probability value of the target text entity in the training context feature in the target text entity as a second probability value;

determining a training error value of the first probability value and the second probability value, and generating a training error value of the second context extraction model;

according to the training error value of the second context relation extraction model, carrying out parameter updating on the shared network function layer and the self-coding network in the second context relation extraction model, and generating a third network layer weight information updating result;

outputting the second context relation extraction model meeting the first training termination requirement as a third context relation extraction model if the second context relation extraction model with the updated characterization parameter of the third network layer weight information updating result meets the first training termination requirement in the third network deployment requirement;

Fixing the shared network function layer and the self-coding network in the third context extraction model, and outputting the fixed third context extraction model as a fourth context extraction model; performing abnormal category prediction on the fourth example abnormal report text according to the fourth context extraction model, and generating an abnormal category training probability value corresponding to the fourth example abnormal report text;

determining a training error value of the real abnormal class indicated by the fourth priori annotation data and the training probability value of the abnormal class, and generating a training error value of the fourth context extraction model;

according to the training error value of the fourth context relation extraction model, parameter updating is carried out on a prediction network in the fourth context relation extraction model, and a fourth network layer weight information updating result is generated;

and if the fourth context relation extraction model with the fourth network layer weight information updating result representing parameter updated meets the second training termination requirement in the third network deployment requirement, outputting the fourth context relation extraction model meeting the second training termination requirement as a target context relation extraction model.

According to a second aspect of the present application, there is provided a server comprising a processor and a readable storage medium storing a program which, when executed by the processor, implements the aforementioned NLP-combined fly ash fusion processing system anomaly report identification method.

According to a third aspect of the present application, there is provided a computer-readable storage medium having stored therein computer-executable instructions for implementing the aforementioned NLP-combined fly ash fusion processing system anomaly report identification method when it is monitored that the computer-executable instructions are executed.

According to any one of the aspects, in the application, when the system exception report text is detected, the exception triggering node is positioned, so that a reference system exception event sequence is generated, then the reference system exception event sequence is subjected to redundancy optimization according to the exception report text, an optimized system exception event sequence and a corresponding context description paragraph sequence are generated, in the optimized system exception event sequence, the optimized system exception event corresponding to each context relation feature is limited, and an exception event limiting area for characterizing the triggering exception node in the system exception report text is generated. Therefore, through abnormal trigger node positioning, redundant optimization and extraction of the contextual characteristics, the abnormal event limiting area in the system abnormal report text can be more accurately identified and understood, so that the fly ash fusion processing system can be conveniently and timely inspected and examined by combining the abnormal event limiting area, and the running efficiency and stability of the fly ash fusion processing system are improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of a method for identifying an abnormality report of a NLP-combined fly ash fusion treatment system according to an embodiment of the present application;

fig. 2 is a schematic diagram of a component structure of a server for implementing the method for identifying an abnormality report of a fly ash fusion processing system with NLP according to an embodiment of the present application.

Detailed Description

In order that those skilled in the art will better understand the present application, a technical solution in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in which it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by those skilled in the art in light of the embodiments of the present application without undue burden, are intended to be within the scope of the present application.

It should be noted that the terms "first," "second," and the like in the description and the claims of the present application and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the application described herein may be implemented in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Fig. 1 is a flow chart illustrating an abnormality report identification method of a fly ash fusion processing system with NLP according to an embodiment of the present application, and it should be understood that, in other embodiments, the order of part of the steps in the abnormality report identification method of a fly ash fusion processing system with NLP according to the present application may be interchanged according to labeling requirements, or part of the steps may be omitted or deleted. The detailed steps of the NLP-combined fly ash fusion processing system abnormality report identification method are described below.

Step S110, when detecting a system abnormality report text for the fly ash fusion processing system, positioning a system abnormality trigger node of the system abnormality report text to generate a reference system abnormality event sequence.

Assuming that the fly ash fusion treatment system is operating in a large power plant, the system automatically generates a detailed system anomaly report text on a day:

"14:32, the second host triggers an alarm. Preliminary monitoring results showed that the melter temperature was abnormally increased, and furthermore, the pressure value of the cooling system was lower than the normal range. After detailed examination, it is presumed that this may be caused by the failure of the primary water pump in the cooling system. At the same time, the hydrogen sulfide emissions also exceed safety standards, which may be related to cooling system failure. Because, from historical data, the amount of hydrogen sulfide emissions tends to increase once the cooling system fails. At present, emergency maintenance has been performed on the cooling system, and the temperature of the melter gradually returns to normal. But close attention must be paid to its operation, especially during the next 72 hours, to prevent similar problems from occurring again. In addition, further investigation and treatment of the hydrogen sulfide emissions problem is also required. "

Thus, the system abnormality report text can be targeted to the following system abnormality trigger nodes: "abnormal rise in melter temperature", "cooling system pressure below normal range", "failure of water pump No. one" and "hydrogen sulfide discharge exceeding safety standards". Then, a sequence of reference system anomaly events is generated.

The reference system abnormal event sequence is to arrange and combine the system abnormal trigger nodes according to the occurrence sequence or the mutual influence logic relationship to form an event sequence. This sequence of events helps to understand the general view of system anomalies and helps to find clues to solve the problem. For example, patterns may be found, such as that a particular system exception trigger node always results in another system exception trigger node, or that certain system exception trigger nodes often occur simultaneously, etc. For example, the reference system anomaly event sequence may include the following:

the melter begins to heat.

The melter temperature abnormally increases.

The cooling system attempts to adjust the temperature, but the pressure is below the normal range.

The water pump number one starts, trying to increase the pressure of the cooling system, but fails due to a malfunction.

Hydrogen sulfide begins to accumulate and eventually is discharged, the discharge exceeds safety standards, and so on.

It should be noted that, in the step S110, the system exception triggering node positioning, the implementation manner of generating the reference system exception event sequence may be implemented by a natural language processing algorithm, and particularly, reference may be made to further description of the subsequent embodiments.

Step S120, performing redundancy optimization on the reference system abnormal event sequence according to the system abnormal report text, and generating an optimized system abnormal event sequence corresponding to the reference system abnormal event sequence and a context description paragraph sequence corresponding to the optimized system abnormal event sequence.

In this embodiment, the context description paragraph sequence includes W context description paragraphs. And a context description paragraph is obtained by extracting a text description paragraph from the system exception report text according to an optimization system exception event in the optimization system exception event sequence. W is an integer greater than 0.

For example, the sequence of reference system anomalies may be redundantly optimized, duplicated or insignificant events deleted, forming an optimized sequence of system anomalies.

For example, assume the original reference system anomaly event sequence described above is as follows:

1. The melter begins to heat.

2. The melter temperature abnormally increases.

3. The cooling system attempts to adjust the temperature, but the pressure is below the normal range.

4. The water pump number one is started in an attempt to increase the pressure of the cooling system.

5. The first water pump fails.

6. The cooling system pressure cannot be raised normally.

7. The hydrogen sulfide begins to accumulate and eventually is discharged.

8. The hydrogen sulfide emissions exceed safety standards.

Wherein the two events of "the cooling system tries to adjust the temperature but the pressure is lower than the normal range" and "the cooling system pressure cannot be raised normally" actually describe the same problem, so they can be combined into one event. At the same time, there is a repetition of "hydrogen sulfide starts to accumulate and finally discharges" and "hydrogen sulfide discharge amount exceeds the safety standard", and the combination can be performed.

After redundant optimization, the obtained optimized system abnormal event sequence is as follows:

1. the melter begins to heat.

2. The melter temperature abnormally increases.

3. The cooling system pressure is below the normal range.

4. The water pump number one starts and fails to increase the pressure of the cooling system.

5. The hydrogen sulfide emissions exceed safety standards.

Therefore, the development context of the abnormal event can be seen more clearly, and the subsequent problem analysis and treatment can be conveniently carried out.

At the same time, relevant text description paragraphs are extracted from the original system anomaly report text according to the optimized system anomaly event sequences, for example, for the event of 'No. one water pump failure', the corresponding context description paragraphs may be 'preliminary monitoring result display that the temperature of the melting device is abnormally increased', and moreover, the pressure value of the cooling system is lower than the normal range. After detailed examination, it is presumed that this may be caused by the failure of the primary water pump in the cooling system. "

It should be noted that, the implementation manner of the redundancy optimization in the step S120 may be implemented by a natural language processing algorithm, which may be specifically described in the following embodiments.

And step S130, extracting the context relation of each context description paragraph in the W context description paragraphs respectively, and generating context relation data corresponding to the system abnormal trigger node.

In this embodiment, the context data includes G context features. G is a positive integer not greater than W, and the contextual characteristics are used for representing the relationship characteristics among the identified entities in the contextual description paragraphs, wherein the relationship characteristics comprise anomaly description keywords and relationship vectors among the anomaly description keywords.

For example, context features may be extracted from each of the W context description paragraphs described above, including entities (e.g., "melter," "cooling system," "water pump No. one," etc.) and their relationship vectors (e.g., "cause and effect relationships between" abnormal rise in melter temperature "and" water pump No. one "failure"). In more detail, the following contextual characteristics may be found:

there may be a causal relationship between "abnormal rise in the melter temperature" and "cooling system pressure below the normal range". That is, the overheating of the melter may be caused by the cooling system being too low in pressure to cool effectively.

The failure of the "water pump No. one to start and fail to increase the pressure of the cooling system" may be a direct consequence of the "cooling system pressure being below the normal range" and may also be a direct cause of the "hydrogen sulfide emissions exceeding the safety standard". That is, the failure of the water pump number one may further exacerbate the pressure problem of the cooling system, resulting in emissions of hydrogen sulfide exceeding the standard.

Through the analysis, the system abnormality can be understood more deeply, and the nature of the system abnormality can be better grasped, so that support is provided for subsequent problem solving.

It should be noted that, the implementation manner of extracting the contextual characteristics in the present step S130 may be implemented by a natural language processing algorithm, which may be specifically described in the following embodiments.

And step S140, in the optimizing system abnormal event sequence, limiting optimizing system abnormal events corresponding to the G contextual characteristics respectively, and generating G abnormal event limiting areas representing the system abnormal trigger nodes in the system abnormal report text.

For example, through step S130, the following contextual characteristics are obtained:

there may be a causal relationship between "abnormal rise in the melter temperature" and "cooling system pressure below the normal range".

The failure of the "water pump No. one to start and fail to increase the pressure of the cooling system" may be a direct consequence of the "cooling system pressure being below the normal range" and may also be a direct cause of the "hydrogen sulfide emissions exceeding the safety standard".

Applying these contextual characteristics to an optimized system anomaly event sequence may form the following anomaly event definition regions:

the melter begins to heat.

[ abnormal melter temperature ] elevation ] - > initiate- [ Cooling System pressure ] lower than the normal range ]

[ Cooling System pressure is lower than normal Range ] -) leading to the starting and failure of the first water pump, failing to increase the pressure of the cooling system

The No. one water pump is started and fails, the pressure of the cooling system is not increased, and the result is that the discharge amount of hydrogen sulfide exceeds the safety standard

Therefore, an event chain with causal relation is constructed, the development process of the problem can be better understood, and the links which need to be repaired or optimized can be more accurately determined, so that the problem can be more effectively solved.

Based on the steps, when the system abnormal report text is detected, abnormal trigger node positioning is carried out, a reference system abnormal event sequence is generated, redundancy optimization is carried out on the reference system abnormal event sequence according to the abnormal report text, an optimized system abnormal event sequence and a corresponding context description paragraph sequence are generated, in the optimized system abnormal event sequence, optimization system abnormal events corresponding to each context relation characteristic are limited, and an abnormal event limiting area for triggering abnormal nodes in the system abnormal report text is generated. Therefore, through abnormal trigger node positioning, redundant optimization and extraction of the contextual characteristics, the abnormal event limiting area in the system abnormal report text can be more accurately identified and understood, so that the fly ash fusion processing system can be conveniently and timely inspected and examined by combining the abnormal event limiting area, and the running efficiency and stability of the fly ash fusion processing system are improved.

In a possible implementation manner, before step S110, the embodiment of the present application may further include the following steps:

step S101, acquiring a first case exception report text for parameter updating of a first natural language processing network and first priori labeling data of a real exception triggering event characterizing the first case exception report text.

In this embodiment, the first example exception report text is generated after initializing (e.g., preprocessing steps such as data cleaning, noise removal, normalization, etc.) the system exception trigger node of the initial example exception report text, and the natural language processing network adopts a converter network converter.

For example, the first example anomaly report text may include first a priori labeled data such as "a rise in the melt temperature anomaly," "a cooling system pressure below a normal range," "a water pump failure," and "a hydrogen sulfide emission exceeding safety standards" specific actual anomaly triggering events. For example, the abnormal trigger node can be marked by special staff according to the professional knowledge of the special staff, for example, the abnormal rise of the temperature of the melting device, the abnormal pressure of the cooling system being lower than the normal range, the failure of the first water pump and the discharge of hydrogen sulfide exceeding the safety standard can be marked, and the corresponding real abnormal trigger event sequence can be marked.

Step S102, locating the first example exception report text according to the first natural language processing network, and generating an estimated system exception event of the first example exception report text for the system exception trigger node.

Step S103, determining a training error value for the estimated system anomaly event of the first example anomaly report text and the real anomaly triggering event of the first example anomaly report text, and generating a training error value for the first natural language processing network.

Step S104, according to the training error value of the first natural language processing network, parameter updating is carried out on the first natural language processing network, and a first network layer weight information updating result is generated.

Step S105, if the first natural language processing network after the updating of the first network layer weight information representation parameter meets the first network deployment requirement, using the first natural language processing network meeting the first network deployment requirement as a second natural language processing network.

In this embodiment, the first example exception report text may be located according to the first natural language processing network to generate an estimated system exception event, and then the estimated system exception event and the actual exception triggering event are compared to determine a training error value, so as to generate a training error value of the first natural language processing network.

For example, assume that the estimated system anomaly event is "the cooling system is attempting to adjust temperature, but the pressure is below the normal range", and the actual anomaly trigger event is "the melter temperature is abnormally elevated". In determining the training error value, the two results need to be compared.

Assume that the error is calculated using a cross entropy loss function. This is a penalty function commonly used for classification problems and allows a comparison of the gap between the predicted probability distribution of the first natural language processing network and the true tag probability distribution. In this scenario, all possible system anomalies (e.g. "cooling system tries to adjust temperature, but pressure is below normal range", "melter temperature is abnormally elevated", etc.) can be seen as different categories.

When the first natural language processing network predicts that the "cooling system tries to adjust temperature but the pressure is below the normal range", its corresponding probability distribution may be [0.1,0.9] (assuming that the first element represents the probability of "the melter temperature abnormally increases", the second element represents the probability of "the cooling system tries to adjust temperature but the pressure is below the normal range"). The true tag probability distribution is [1,0], because the true anomaly is "the cooling system tries to adjust temperature, but the pressure is below the normal range".

The two probability distributions are put into the cross entropy loss function, and a positive number can be obtained as a training error value. The larger the value, the larger the difference between the predicted result and the real result of the model is; conversely, if this value is smaller, it is stated that the estimated result of the first natural language processing network is closer to the true result.

Then, the parameters of the first natural language processing network can be updated according to the error value, so that the first natural language processing network can predict the abnormal event of the system more accurately next time. This process is typically accomplished by a back-propagation algorithm and a gradient descent algorithm.

And S106, analyzing the functional parameter layer of the second natural language processing network to generate functional parameter layer analysis data.

Step S107, if the function parameter layer parsing data characterizes that the function parameter layer of the second natural language processing network has a decoder, replacing the decoder with a first encoder carrying a wandering window parameter.

Step S108, parameter migration configuration is carried out for the first encoder according to the network layer weight information of the decoder, and a second encoder is generated.

In the second natural language processing network, the decoder and encoder are two important components. The encoder is responsible for converting the input data (e.g., text) into an internal representation (typically a vector) and the decoder converts the internal representation back into the original data or other form of output.

The first encoder carrying the wandering window parameter may refer to an encoder with a specific function. The wander window is a common technique for processing time series data, and it slides on a time series, and a piece of data with a fixed length is taken each time to process, so that a local mode in the data can be captured, and the wander window is very useful for processing time series data such as an abnormality report of a fly ash melting system of a power plant.

When replacing the decoder with the first encoder carrying the runaway window parameters, it is in fact the structure of the network that is changed to improve its performance. Encoders using such a walk-through window are more efficient than the original decoder, for example, in handling power plant fly ash fusion system anomaly reports. Thus, the original decoder is replaced and the new encoder is used.

On the basis, parameter migration configuration is carried out for the first encoder according to the network layer weight information of the decoder, and a second encoder, also called migration learning, is generated. For example, parameters (weights) that have been learned by the decoder are used as the migration network parameters of the first encoder.

Step S109, outputting a second natural language processing network including the second encoder as an anomaly trigger node positioning network for positioning a system anomaly trigger node of the system anomaly report text.

Based on the steps, automatic prediction and estimation of the abnormal events of the system are realized, and the requirements of people and manual processing are greatly reduced, so that the efficiency is improved. Second, by constantly training and updating network parameters, the accuracy of natural language processing networks is gradually improved. This means that the error rate of the predictions will decrease over time, further enhancing the reliability of the system. The model can then be tuned and optimized according to the specific needs and application scenario by parsing the functional parameter layer and replacing the encoder. This increases the flexibility and adaptability of the natural language processing network. Finally, the encoder carrying the parameters of the wandering window is used for parameter migration configuration, so that a natural language processing network can better understand and process time sequence data, and the method is better in predicting abnormal events of a dynamic system. Overall, this process can increase the efficiency, reliability, flexibility and adaptability of the system, as well as the processing power of the dynamic data.

In one possible implementation, the aforementioned step S120 may include:

step S121, performing redundancy optimization on the reference system abnormal event sequence according to the redundancy optimization strategy of event filtering and classifying, so as to generate a candidate system abnormal event sequence.

In this embodiment, the candidate system abnormal event sequence includes a reference system abnormal event Lx. x is a positive integer not greater than D. And D, representing the total number of abnormal trigger nodes in the candidate system abnormal event sequence.

For example, in one possible implementation, the reference system anomaly event sequence is obtained when locating a system anomaly trigger node of the system anomaly report text in accordance with an anomaly trigger node locating network.

The reference system anomaly event sequence includes F reference system anomalies. F is an integer greater than 0. The abnormal trigger node positioning network is further used for determining estimated system abnormal events corresponding to each reference system abnormal event in the F reference system abnormal events respectively.

Step S121 may include the steps of:

1. and respectively filtering text contents of each reference system abnormal event in the F reference system abnormal events according to keyword rules mapped by the redundant optimization strategies for event filtering and classifying, and generating F filtered system abnormal events.

For example, depending on the redundancy optimization strategy of event filtering and categorization, a set of keyword rules may be established that are important information for identifying and categorizing system anomalies. Through this set of rules, the text content of each reference system anomaly event can be filtered.

For example, a set of keywords is defined for each reference system anomaly event. For example, for a "melter temperature anomaly rise", then the keywords may be "melter", "temperature", "anomaly" and "rise" matching the text content in each reference system anomaly with its corresponding keyword. In this process, only the part containing the keywords is retained, and the other non-keyword parts are filtered out.

After filtering, F filtered system abnormal events are obtained, and all the F filtered system abnormal events remove insignificant information and only remain keyword parts.

For example, among the original "F reference system anomalies", one reference system anomaly is "cooling system pressure is below the normal range due to a failure of the water pump No. one". Assuming that the keyword rules are "water pump", "failure", "cooling system" and "pressure", the reference system anomaly will become "water pump failure number one" after filtering, cooling system pressure is below the normal range. In this way, the original reference system abnormal event can be simplified and optimized, so that the subsequent problem analysis and processing are more convenient.

2. And according to the probability values of the F estimated system abnormal events, carrying out order arrangement on the F filtered system abnormal events to generate an abnormal event order sequence.

3. Outputting the filtered system abnormal events with the maximum probability value in the abnormal event sequence as a first filtered system abnormal event, and outputting (F-1) filtered system abnormal events except the first filtered system abnormal event in the abnormal event sequence as a first temporary system abnormal event sequence.

4. And respectively determining the text description similarity between the first filtered system abnormal event and each filtered system abnormal event in the first temporary system abnormal event sequence.

For example, the text description similarity may employ cosine similarity, jaccard similarity, euclidean distance, or semantic similarity calculation based on deep learning, or the like.

5. If the first temporary system abnormal event sequence has the common system abnormal event with the text description similarity larger than the preset similarity, the first filtered system abnormal event is reserved, and the common system abnormal event is cleaned in the first temporary system abnormal event sequence.

6. And in the cleaned first temporary system abnormal event sequence, taking the filtered system abnormal event with the largest estimated system abnormal event as a second filtered system abnormal event, and taking the filtered system abnormal events except the second filtered system abnormal event as a second temporary system abnormal event sequence.

7. And reserving the second filtered system abnormal event, performing redundancy optimization on the common system abnormal event in the second temporary system abnormal event sequence according to the text description similarity between the second filtered system abnormal event and each filtered system abnormal event in the second temporary system abnormal event sequence until the second temporary system abnormal event sequence after redundancy optimization is an empty set, and outputting the reserved first filtered system abnormal event and the reserved second filtered system abnormal event as an iterative system abnormal event sequence.

8. And expanding each filtered system abnormal event in the iterative system abnormal event sequence according to the keyword rule to generate a candidate system abnormal event sequence.

For example, each filtered system exception in the existing iterative system exception sequence may be further refined and expanded based on certain rules (e.g., keywords, phrases, patterns, etc.).

Illustratively, in the system exception sequence of the foregoing example, each system exception may be extended according to a keyword rule:

1. "melter starts heating": since this event itself does not contain a specific problem, it may not be extended. However, if there are relevant rules, such as "when the melter starts to heat, it is necessary to check the energy supply and the equipment status", then "energy supply check" and "equipment status check" may be added to the candidate sequence.

2. "abnormal rise in melter temperature": "energy supply overdose" and "heat control failure" may be added to the candidate sequence.

3. "Cooling systems attempt to adjust temperature, but pressure is below the normal range": "coolant pump pressure low" and "cooling system leak" may be added to the candidate sequence.

4. "Water pump one started, attempting to increase the pressure of the cooling system, but failed due to failure": "No. one pump failure" and "cooling system pressure adjustment failure" may be added to the candidate sequence.

5. "Hydrogen sulfide starts to accumulate and eventually is discharged, and the discharge amount exceeds the safety standard": "harmful gas treatment system failure" and "environmental standard violation" may be added to the candidate sequence.

Thus, according to these rules, the original sequence of abnormal events may be expanded into the following candidate sequence of system abnormal events:

"melting vessel start heating" is extended to [ "melting vessel start heating", "energy supply inspection", "equipment status inspection" ]

"abnormal rise in melter temperature" extends to [ "abnormal rise in melter temperature", "excessive energy supply", "heat control failure" ]

"cooling system tries to adjust the temperature, but the pressure is lower than the normal range" is extended to [ "cooling system tries to adjust the temperature, but the pressure is lower than the normal range", "Low pressure of Cooling liquid Pump", "Cooling System leakage

"Water Pump started, attempts to increase the pressure of the Cooling System, but fails because of failure" spread to [ "Water Pump started, attempts to increase the pressure of the Cooling System, but fails because of failure", "Water Pump failure", "Cooling System pressure adjustment failure")

"hydrogen sulfide starts to accumulate and finally discharges", the discharge amount is expanded beyond the safety standard "to [" hydrogen sulfide starts to accumulate and finally discharges, the discharge amount exceeds the safety standard "," the harmful gas treatment system is invalid "," the environmental protection standard is violation "]

Such extensions may help to more fully understand each system anomaly event, and their possible related factors and consequences, for more efficient problem analysis and processing.

Step S122, according to the context correction optimization network, correcting and optimizing the reference system abnormal event Lx based on the system abnormal report text, and generating an optimization system abnormal event Rx and a context description paragraph corresponding to the optimization system abnormal event Rx.

For example, step S122 may include:

1. and according to the paragraph context deviation correcting and optimizing network, based on the system abnormality report text, estimating deviation parameters of the paragraph context for the reference system abnormality event Lx, and generating a first paragraph context deviation parameter.

2. Correcting the deviation of the abnormal event Lx of the reference system according to the context deviation parameter of the first paragraph, and generating a first abnormal event of the optimization system corresponding to the abnormal event Lx of the reference system.

The context deviation parameter is used to measure the difference between the description of the reference system anomaly event in the original system anomaly report text and the expected or standard description. Such parameters may include, but are not limited to, the following:

1. semantic deviation: this refers to whether the event description is semantically in-coming or-going, e.g. "the melter to start heating" may semantically deviate from a true abnormal situation, such as "the melter temperature is abnormally elevated".

2. Position deviation: this is concerned with whether the location of the event description in the text deviates from the conventional or expected location. For example, if an important exception event is placed at the end of the text, then there is a positional deviation.

3. Frequency deviation: deviations may also be indicated if an event is referred to in the text too much or too little. Such as a certain fault event actually occurs with a high frequency but is mentioned a small number of times in the report.

4. Correlation bias: deviations may also be considered to exist if the event description is ambiguous or erroneous in relation to other related events. For example, if the two events "water pump No. one failed" and "cooling system pressure is below the normal range" should be causally related, but not explicitly stated in the text, then there is a correlation bias.

The above are some possible context deviation parameters, and in practical application, specific parameters may be defined and designed according to practical requirements and situations.

For example, if the reference system anomaly event Lx is "melter start heating", the network analysis system anomaly report text can be optimized by paragraph context deskewing to obtain a context deviation parameter associated with this event. Taking the foregoing example as an example, the expression of "the melting device starts to heat" may be found to be too broad according to the obtained context deviation parameter, and thus may be corrected to "the abnormal rise of the melting device temperature" as the first optimization system abnormal event.

In a more specific scenario, after context deviation parameters such as semantic deviation, position deviation, frequency deviation, relevance deviation and the like are determined, correction is performed on the abnormal event Lx of the reference system according to the determined context deviation parameters. This may involve modifying the text content to eliminate semantic bias, adjusting the location of the text to eliminate location bias, adjusting the frequency of occurrence of the text to eliminate frequency bias, and adjusting the degree of relevance of the text to other text to eliminate relevance bias. Finally, the corrected reference system abnormal event Lx becomes a first optimization system abnormal event. At this point it should have been expected, whether in terms of meaning, location, frequency of content, or degree of association with other text.

Note that the above steps may need to be iterated until all deviation parameters are within acceptable threshold limits. Meanwhile, in order to ensure the effectiveness of correction, filtering of text content may be required to exclude information that may cause misunderstanding or confusion.

3. And if the first paragraph context deviation parameter belongs to a threshold parameter interval, outputting a first optimization system abnormal event corresponding to the reference system abnormal event Lx as an optimization system abnormal event Rx corresponding to the reference system abnormal event Lx.

If the first paragraph context deviation parameter belongs to the threshold parameter interval, the paragraph context deviation parameter indicating the system abnormal event Lx (now called the first optimization system abnormal event) after optimization and deviation correction processing is already within the preset acceptable range. The "threshold parameter interval" may be set according to actual needs, for example, the maximum and minimum values of various bias parameters such as semantic bias, position bias, frequency bias, and relevance bias may be set.

If the context deviation parameters of the paragraphs are all within the threshold parameter interval, the first optimization system abnormal event is considered to be satisfactory, and can be used as the optimization system abnormal event Rx output corresponding to the reference system abnormal event Lx.

In practice, this process may require iterative iterations, as after one optimization and correction process, it may also be necessary to further check whether the newly generated anomaly event meets all threshold parameters. If not, the optimization and correction are continued until all deviation parameters are within the threshold parameter interval.

4. And filtering text content of the system abnormal report text according to the positioning information of the optimizing system abnormal event Rx, and generating a context description paragraph corresponding to the optimizing system abnormal event Rx.

For example, positioning information may be obtained that optimizes the system for an anomaly event Rx, which may include information about the time, location, device, etc. that the anomaly occurred. And searching the content related to the obtained positioning information in the system exception report text, filtering out irrelevant information, and only keeping the part related to the exception event. This step may require the use of Natural Language Processing (NLP) techniques such as keyword extraction, text classification, emotion analysis, etc. Finally, the filtered text is organized into one or more paragraphs, forming context description paragraphs that optimize the system anomaly event Rx. This context description section should contain all important information related to the abnormal event while excluding irrelevant content.

Through the steps, a context description paragraph corresponding to the abnormal event Rx of the optimizing system can be generated, so that the specific situation of the abnormal event can be understood, and references can be provided for further fault diagnosis and problem solving.

In a possible implementation manner, after the first optimization system abnormal event corresponding to the reference system abnormal event Lx is obtained, if the first paragraph context deviation parameter does not belong to the threshold parameter interval, according to the paragraph context deviation rectifying optimization network, based on the system abnormality report text, performing paragraph context deviation parameter estimation on the first optimization system abnormal event corresponding to the reference system abnormal event Lx, generating a second paragraph context deviation parameter, according to the second paragraph context deviation parameter, rectifying the first optimization system abnormal event corresponding to the reference system abnormal event Lx, generating a second optimization system abnormal event corresponding to the reference system abnormal event Lx until the second paragraph context deviation parameter belongs to the threshold parameter interval, and outputting the second optimization system abnormal event corresponding to the reference system abnormal event Lx as the optimization system abnormal event Rx corresponding to the reference system abnormal event Lx.

And step S123, when D optimization system abnormal events are obtained, performing redundancy optimization on the D optimization system abnormal events according to the redundancy optimization strategy.

Step S123 is to further optimize the D optimization system abnormal events using the redundant optimization strategy after the events are obtained in the previous step.

Redundancy optimization is a data processing method aimed at eliminating or reducing duplicates or similar items in a data set, thereby improving data quality and processing efficiency. In this process, various techniques may be used, such as deduplication, clustering, pattern matching, and the like.

Specifically, assuming that D optimization system anomalies have been obtained, some of those events may be found to be very similar or identical. These redundant events, if not removed, may interfere with subsequent analysis and decision making. Thus, it is necessary to merge these similar or identical events into one, or to retain only the most important one of them, by a redundancy optimization strategy.

For example, assume that there are two optimization system anomalies: "System A fails at 13:00" and "System A fails at 13:01". Both events, although slightly different in time, essentially describe the same problem. Thus, the two events can be combined into one: "System A fails around 13:00".

In general, the step S123 improves the data quality and the processing efficiency through the redundancy optimization strategy, so that the subsequent analysis and decision making are more accurate and effective.

Step S124, outputting the optimized system abnormal event after the redundancy optimization as an optimized system abnormal event sequence corresponding to the reference system abnormal event sequence, and outputting the context description paragraph corresponding to the optimized system abnormal event after the redundancy optimization as a context description paragraph sequence corresponding to the optimized system abnormal event sequence.

In one possible implementation manner, the training step of the above paragraph context deskewing optimization network is described below in conjunction with an embodiment, and a specific implementation manner of the relevant step may be referred to the description of the foregoing embodiment, where the training step specifically includes:

step A101, obtaining a second example anomaly report text for updating parameters of the basic deviation rectifying optimization network and second priori label data of prior positioning information of system anomaly triggering nodes representing the second example anomaly report text.

In this embodiment, the priori positioning information may refer to a priori text content positioning portion of a system abnormal event corresponding to a system abnormal trigger node of the second example abnormal report text.

Step a102, according to the basic deviation correction optimization network, performing paragraph context deviation parameter learning on the second example anomaly report text, and generating a training paragraph context deviation parameter of the second example anomaly report text.

In this embodiment, according to the basic correction optimization network, the deviation parameter learning of the paragraph context may be performed on the system exception event corresponding to the system exception trigger node of the second example exception report text, so as to generate the training paragraph context deviation parameter.

And step A103, rectifying the positioning information of the second example exception report text according to the context deviation parameter of the training paragraph, and generating the training positioning information of the second example exception report text.

And step A104, carrying out parameter updating on the basic deviation correcting optimization network according to the training positioning information and the priori positioning information, and generating a second network layer weight information updating result.

In this embodiment, training positioning information and priori positioning information may be used as input data, and sent to a basic correction optimization network for forward propagation, and the output value of each layer is calculated until the last layer. In the process, the current network weight is used for calculation, and then the value of the loss function is calculated according to the difference between the output result of the basic deviation correction optimization network and the actual result (namely the target output). Common loss functions are mean square error, cross entropy, etc. The back propagation is then performed, calculating the gradient of the loss function with respect to the weights of each layer. This procedure is to find out which weight changes can drop the value of the loss function. Finally, the weights of the neural network are updated according to a gradient descent method or other optimization algorithms. This is the so-called parameter update. Repeating the steps for a plurality of times until the performance of the model meets the requirement or the preset training times are reached. The obtained network weight is the updating result of the weight information of the second network layer.

Therefore, the parameter updating of the basic deviation correcting optimizing network can be carried out through training the positioning information and the priori positioning information, and a second network layer weight information updating result is generated.

And step A108, outputting the basic deviation rectifying optimization network meeting the second network deployment requirement as the paragraph context deviation rectifying optimization network if the basic deviation rectifying optimization network with the updated second network layer weight information updating result representing parameter meets the second network deployment requirement.

In one possible implementation, the step S130 includes:

step S131, determining a target context description paragraph from the W context description paragraphs.

For example, in this process, a description paragraph containing critical information or directly related to a system anomaly may be selected. For example, a paragraph of "water pump No. one failure" or "abnormal rise in the temperature of the melter" is described.

Step S132, based on the prediction network in the target context extraction model, carrying out abnormal category prediction on the target context description paragraph, and generating an abnormal category probability value corresponding to the target context description paragraph.

For example, in this process, the target context description paragraphs are input into the prediction network in the target context relation extraction model, and the probability values of the anomaly categories corresponding to the target context description paragraphs are calculated. For example, the abnormality category that "water pump No. one failure" may correspond to is "equipment failure", and the abnormality category that "melter temperature is abnormally increased" may correspond to is "overheat".

And step S133, carrying out context feature prediction on the target context description paragraph based on the self-coding network in the target context extraction model, and generating the context feature corresponding to the target context description paragraph.

For example, in this process, the object context description paragraphs are input into the self-encoding network, and the relationship features between the entities in these object context description paragraphs are calculated. For example, there may be a causal relationship between "pump number one failure" and "cooling system pressure below the normal range".

And step S134, outputting the context relation characteristic corresponding to the target context description section as the context relation data corresponding to the system abnormal trigger node if the abnormal category probability value corresponding to the target context description section is larger than the set probability value.

For example, in this process, if the probability value of the abnormal category corresponding to a certain context description paragraph is greater than a set threshold (e.g., 0.5), then it is considered that the paragraph does describe an abnormal event, and the corresponding context relation feature is taken as the output result.

In one possible embodiment, the training procedure of the above-described target context extraction model is described below.

And step B101, acquiring training text data for updating parameters of a basic natural language processing network and priori labeling data corresponding to the training text data.

And step B102, carrying out parameter updating on the shared network function layer and the self-coding network in the basic natural language processing network according to a third example exception report text in the training text data and third priori label data representing the priori context characteristics of the third example exception report text, and generating a first context extraction model.

For example, known exception report text may be gathered as a third example exception report text. In addition, corresponding prior labeling data is needed, which includes information that labels the contextual characteristics of the third example exception report text. On this basis, the underlying natural language processing network is updated with exception report text and its corresponding a priori contextual feature annotation data of the third paradigm, e.g., by adjusting the shared functional layers and the self-encoding network in the underlying natural language processing network. In this way, a first context extraction model is generated that is capable of extracting context features from any anomaly report text.

Wherein the shared functional layer refers to a group of neural network layers commonly used by a plurality of modules or tasks in the basic natural language processing network. Its function is to perform feature extraction on the input text and capture semantic and contextual information. The shared functional layer is typically composed of multiple convolutional layers, a recurrent neural network (e.g., LSTM or GRU), or a transducer, etc.

Specifically, the shared function layer may perform the following operations:

feature extraction: the shared functional layer extracts useful feature representations from the input text by convolution, recurrent neural network, or transfomer operations, etc., to capture critical semantic information.

Context modeling: the shared functionality layer may learn context information to help understand the meaning of text by considering the relationships between the front and back words or sentences. For example, in a recurrent neural network, the hidden state may store information of historical text to facilitate processing of the current word.

Parameter sharing: parameters of the shared functional layer may be shared between different tasks to reduce the number of training parameters of the model. This can improve the efficiency and generalization ability of the model.

Migration learning: by pre-training the shared functional layer on one task and then migrating it to other related tasks for fine-tuning, it can help improve the performance of the model on the new task.

In summary, the shared functional layer plays an important role in natural language processing networks, and it is capable of extracting meaningful feature representations from text and capturing contextual information, thereby helping to better understand and process report text data.

And step B103, carrying out parameter updating on the shared network function layer and the prediction network in the first context extraction model according to a fourth example exception report text in the training text data and fourth priori label data representing a real exception triggering event of the fourth example exception report text, and generating a second context extraction model.

In this embodiment, parameter updating is performed on the shared network function layer and the prediction network in the first context extraction model again according to a fourth example exception report text in the training text data and fourth priori label data representing a real exception triggering event of the fourth example exception report text, so as to generate a second context extraction model.

And step B104, according to the second context extraction model, carrying out context feature prediction on the third example exception report text, and generating training context features corresponding to the third example exception report text.

Step B105, the text entity of the third example exception report text is walked, and the walked text entity is output as a target text entity.

And step B106, outputting the probability value of the target text entity in the prior context feature indicated by the third prior labeling data as a first probability value, and outputting the probability value of the target text entity in the training context feature in the target text entity as a second probability value.

In this embodiment, the first probability value may be understood as a priori context feature probability value, which refers to information that labels the context of the exception report in the prior labeling data, and may be calculated by analyzing the context of the labeling entity in the prior labeling data. These probability values may be calculated based on statistical methods, rule-based methods, or machine learning models.

The second probability value may be understood as a training context feature probability value, where a training context feature is a feature modeled and extracted according to a context in the training data set, and during the training process, a machine learning model (e.g. a neural network) may be used to predict probability values of the target text entity in the training context feature, where the probability values may represent importance or relevance of the target text entity in the training data.

And step B107, determining a training error value for the first probability value and the second probability value, and generating a training error value of the second context extraction model.

For the training error of the first probability value and the second probability value, a mean square error loss function may be used. Assuming that the first probability value is y and the second probability value is p, the mean square error loss is defined as follows: loss= (p-y)/(2).

And step B108, carrying out parameter updating on the shared network function layer and the self-coding network in the second context relation extraction model according to the training error value of the second context relation extraction model, and generating a third network layer weight information updating result.

And step B109, if the second context relation extraction model with the updated characterization parameter of the third network layer weight information updating result meets the first training termination requirement in the third network deployment requirement, outputting the second context relation extraction model meeting the first training termination requirement as a third context relation extraction model.

And B110, fixing the shared network function layer and the self-coding network in the third context extraction model, outputting the fixed third context extraction model as a fourth context extraction model, and carrying out abnormal category prediction on the fourth example abnormal report text according to the fourth context extraction model to generate an abnormal category training probability value corresponding to the fourth example abnormal report text.

And step B112, determining a training error value for the real abnormal class indicated by the fourth priori label data and the training probability value of the abnormal class, and generating the training error value of the fourth context extraction model.

And step B113, carrying out parameter updating on the prediction network in the fourth context relation extraction model according to the training error value of the fourth context relation extraction model, and generating a fourth network layer weight information updating result.

And step B114, if the fourth context relation extraction model with the updated fourth network layer weight information updating result representing parameters meets the second training termination requirement in the third network deployment requirement, outputting the fourth context relation extraction model meeting the second training termination requirement as a target context relation extraction model.

Based on the steps, the first context relation extraction model is generated by acquiring training text data and priori labeling data and updating parameters based on a shared network function layer of a basic natural language processing network and a self-coding network. And extracting a model and training data by using the first context relation, generating training context relation characteristics, performing entity migration, and outputting a target text entity. And calculating probability values of the target text entity in the prior context feature and the training context feature, determining a training error value, and updating parameters of the second context extraction model. And outputting a third context relation extraction model according to the second network deployment requirement if the training termination requirement is met. And fixing the shared network function layer and the self-coding network of the third context relation extraction model, and outputting a fourth context relation extraction model. And carrying out abnormal category prediction according to the fourth context relation extraction model, calculating a training error value, and carrying out parameter updating on the fourth context relation extraction model. And if the second training termination requirement is met, outputting a target context relation extraction model, so that the model performance is gradually optimized, and the extraction accuracy of the context relation features and the abnormal category prediction capability are improved.

Further, fig. 2 shows a schematic hardware structure of a server 100 for implementing the method provided by the embodiment of the present application. As shown in fig. 2, the server 100 may include one or more processors 102 (the processor 102 may include, but is not limited to, a microprocessor MCU or a processing device such as a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions, and a controller 108. It will be appreciated by those of ordinary skill in the art that the structure shown in fig. 2 is merely illustrative and is not intended to limit the structure of the server 100 described above. For example, the server 100 may also include more or fewer components than shown in FIG. 2, or have a different configuration than shown in FIG. 2.

The memory 104 may be used to store software programs and modules of application software, such as program instructions corresponding to the above-described method embodiments in the embodiments of the present application, and the processor 102 executes the software programs and modules stored in the memory 104, thereby performing various functional applications and data processing, that is, implementing a method for identifying abnormal reports of a fly ash fusion processing system in combination with NLP. Memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include memory remotely located with respect to the processor 102, which may be connected to the server 100 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission means 106 is arranged to receive or transmit data via a network. The specific example of the network described above may include a wireless network provided by a communication provider of the server 100. In one example, the transmission device 106 includes a network adapter that can connect to other network equipment through a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency module for communicating wirelessly with the internet.

It should be noted that: the sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the foregoing description has been directed to specific embodiments of this application. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

The embodiments of the present application are described in a progressive manner, and identical and similar parts of the embodiments are all referred to each other, and each embodiment is mainly described as a difference from other embodiments. In particular, for the different embodiments above, since they are substantially similar to the method embodiments, the description is relatively simple, and reference is made to the description of the method embodiments in part.

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

Claims

1. A method for identifying abnormal reports of a fly ash fusion treatment system in combination with NLP, the method comprising:

2. The method of claim 1, wherein prior to the step of locating a system anomaly trigger node of the system anomaly report text to generate a sequence of reference system anomaly events, the method further comprises:

3. The method for identifying abnormal report of the fly ash fusion processing system in combination with NLP according to claim 1, wherein the redundant optimization of the abnormal event sequence of the reference system according to the abnormal report text of the system, the generation of the optimized abnormal event sequence of the reference system corresponding to the abnormal event sequence of the reference system and the context description paragraph sequence corresponding to the abnormal event sequence of the optimized system, comprises:

4. The method for identifying abnormal reports of a fly ash fusion processing system in combination with NLP according to claim 3, wherein the reference system abnormal event sequence is obtained when locating a system abnormal trigger node of the system abnormal report text according to an abnormal trigger node locating network;

5. The method for identifying abnormal report of NLP-combined fly ash fusion processing system according to claim 3, wherein the optimizing network for rectifying deviation according to context of paragraphs, based on the system abnormal report text, optimizes rectifying deviation of the reference system abnormal event Lx, and generates an optimizing system abnormal event Rx and a context description paragraph corresponding to the optimizing system abnormal event Rx, comprises:

6. The method for identifying an NLP-combined fly ash fusion processing system exception report according to claim 5, wherein after obtaining a first optimization system exception event corresponding to the reference system exception event Lx, the method further comprises:

7. A fly ash fusion processing system anomaly report identification method in combination with NLP according to claim 3, wherein prior to optimizing the network according to paragraph context deskewing, the method further comprises:

8. The method for identifying abnormal reports of NLP-combined fly ash fusion processing system according to any one of claims 1 to 7, wherein the extracting the context relation from each of the W context description paragraphs to generate the context relation data corresponding to the abnormal trigger node of the system comprises:

9. The NLP-combined fly ash fusion treatment system anomaly report identification method of claim 8, wherein prior to extracting the model from the target context, the method further comprises:

10. A server comprising a processor and a readable storage medium storing a program which when executed by the processor implements the NLP-combined fly ash fusion processing system anomaly report identification method of any one of claims 1-9.