CN105809196A - Priori topic model-based train control system on-board equipment intelligent fault diagnosis method - Google Patents

Abstract

The invention relates to a priori topic model-based train control system on-board equipment intelligent fault diagnosis method. The method mainly comprises two steps of feature extraction and intelligent diagnosis.

Description

Train Operation Control System Onboard device intelligence method for diagnosing faults based on priori topic model

Technical field

The invention belongs to technical field of rail transportation operation control, relate to the Train Operation Control System Onboard device intelligence method for diagnosing faults based on priori topic model.

Background technology

In the prior art, the troubles diagnosis and fixing to mobile unit, the Main Diagnosis measure of onsite application has:

(1) desk checking, this is a kind of traditional diagnosis and maintenance mode, because its labor intensity is big, the low inferior reason of diagnosis efficiency, increasingly it is not suitable with modern demand for development.

(2) data monitoring system, this system is by the operational factor of specific data acquisition unit Real-time Collection mobile unit, and is pooled to data monitoring center, and plant maintenance personnel comprehensively analyze the data collected to carry out fault diagnosis and maintenance.Data monitoring system achieves the Real-time Collection of mobile unit and possesses certain trouble diagnosibility, is the important means of current telecommunication and signaling branch regular maintenance and accident analysis.But, this system still suffers from misdiagnosis rate height, the coarse problem of fault location in actual applications, and its data analysis needs a large amount of artificial participation simultaneously, and this brings heavy burden to attendant undoubtedly.

(3) standardization of rds data: some scholars solve railway territory data and the skimble-scamble problem of information based on the method for body.XML, XSL is utilized to set up the standardization document of Europe data.But still cannot analyze and process fault data existing at present.

Summary of the invention

The invention aims to solve the problems referred to above, it is provided that based on the Train Operation Control System Onboard device intelligence method for diagnosing faults of priori topic model, solve problems of the prior art.

Solve the desk checking inefficiency of on-the-spot mobile unit fault diagnosis and the present situation that data monitoring system misdiagnosis rate is higher.

A kind of Train Operation Control System Onboard device intelligence method for diagnosing faults based on priori topic model, described method mainly includes feature extraction and two key steps of intelligent diagnostics.

Described feature extraction refers to that the mobile unit fault signature based on priori topic model extracts, and describes document from lexical item spatial alternation to theme feature space by priori LDA by the phenomenon of the failure of mobile unit.

Priori in fault case library text is extracted, and the theme update probability that priori topic model is LDA model can be expressed as formula (2):

$P(z_i=j|z_{-i},w,\mathrm{\α},\mathrm{\β})=\left(\frac{n_{-i,j}^{w_i}+\mathrm{\β}}{{\mathrm{\Σ}}_{w^{\′}}^W{n}_{-i,j}^{w^{\′}}+W\mathrm{\β}}\right)\left(\frac{n_{-i,j}^{d_i}+\mathrm{\α}}{{\mathrm{\Σ}}_j^T{n}_{-i,j}^{d_i}+T\mathrm{\α}}\right)---\left(2\right)$

Wherein, z_i=j represents i-th lexical item in document and is assigned under theme j, z_-iThe theme representing out the every other word outside i-th word distributes；It it is the word w ' number of times being assigned under theme j；It is document d_iLower word i is assigned to the number of theme j；The Dirichlet (Di Li Cray) that α and β is document subject matter and theme lexical item respectively is distributed hyper parameter.

The main thought of priori LDA is by integrating priori DM_ij, revise theme update probability, in theme renewal process, formula (2) will be multiplied by an extra saturation by us, such as formula (3):

For priori LDA model, firstly the need of setting theme number x, the priori number t of the phenomenon that breaks down can be obtained according to fault data collection, assuming that x meets original lexical item space lexical item number > x >=min (t, T), secondly, lexical item document matrix is inputted priori LDA and carries out theme feature extraction by us, and then obtains theme feature space and subject document matrix.

Based on Stratified Strategy method for diagnosing faults, the grader based on SVM is selected to carry out intelligent diagnostics, based on the grader of SVM, there is general generalization ability preferably on the one hand, on the other hand, it is possible to the complexity that reply mobile unit failure diagnostic process phenomenon of the failure reveals as mark sheet.

For mobile unit two-stage fault mode, take substep diagnosis policy, first, spatially build SVM classifier A at the theme feature extracted, carry out level fault pattern (FFP) fault diagnosis；Then, the diagnostic result of level fault pattern is constituted new feature space as new feature and theme feature Space integration by us, builds SVM classifier B, carry out secondary failure pattern (SFP) fault diagnosis on this feature space.

Beneficial effect

This patent adopts the high speed railway vehicle mounted ATP failure logging of collection in worksite, and totally 1046 data are verified test.The itemized record of the train control on board equipment fault that data are accumulated by scene, including fault time, train number, running section, phenomenon of the failure describes, if impact is stopped, accident analysis, fault category, troubleshooting situation.Diagnostic data is that the phenomenon of the failure in failure logging describes, and diagnostic result is fault category.

(1) train control on board equipment fault signature extracts: traditional feature extracting method does not consider priori, poor for high-speed railway vehicular equipment fault diagnosis field adaptability.The present invention adopts the basic thought of Topics Crawling, is incorporated in theme feature mining process by field priori, to extract the feature of applicable mobile unit fault diagnosis.

(2) train control on board equipment fault diagnosis: diagnosing for mobile unit two-stage fault mode, complexity is higher, has had a strong impact on diagnosis effect.The present invention proposes classification Fault Diagnosis Strategy, level fault diagnostic result is fused to lexical item feature space, and then secondary failure pattern is diagnosed.

Accompanying drawing explanation

Fig. 1 is based on the feature extraction flow chart figure of priori topic model；

Fig. 2 is based on the Troubleshooting Flowchart figure of SVM.

Detailed description of the invention

Embodiment 1

The Intelligent fault diagnosis scheme that the present invention proposes mainly includes feature extraction and two key steps of intelligent diagnostics.

1 extracts based on the mobile unit fault signature of priori topic model

LDA model is proposed in 2003 by D.M.Blei et al., is current most widely used a kind of probability topic model, and it has assumes than the more fully text generation of other models.But, it is a kind of unsupervised model, there is certain blindness in Topics Crawling process.Therefore, this patent, by mobile unit field priori is incorporated into LDA model Topics Crawling process, improves accuracy rate and the efficiency of the excavation of fault theme feature.Core concept is, by priori LDA, the phenomenon of the failure of mobile unit is described document from lexical item spatial alternation to theme feature space.Specific features extraction step, refers to accompanying drawing 1.

1.1 mobile unit field prioris are extracted

It is firstly introduced into three concepts:

" theme related term ": under certain corpus, in certain theme, the frequency of occurrences is higher, and the word that the frequency of occurrences is not high in other themes.

" public lexical item ": under certain corpus, the lexical item that occurrence number is all higher under multiple themes (>=2).

" weak relevant lexical item ": under certain corpus, the lexical item higher with theme co-occurrence rate.

On the one hand, based on the topic model of Term co-occurrence, only only in accordance with word frequency (TF) to term clustering, it is impossible to avoid the impact that public word brings.Such as, in " BTM fault " and " DMI fault " subject categories, the number of times that " fault " word occurs all compares many, is public word.And theme related term " BTM ", the number of times that " DMI " occurs respectively in two themes wants many, but according to the topic model of Term co-occurrence, it is impossible to differentiates this word and is the theme related term or public word.Therefore, word " BTM " and " DMI " are higher with " fault " Term co-occurrence rate, thus by mistake by " BTM ", " fault ", " DMI " three words are assigned under a theme, and cause the confusion of Topics Crawling.But, we can adopt the weight reducing public word, and solves to share the adverse effect that word brings, and improves LDA Topics Crawling accuracy.

On the other hand, the word higher with certain theme correlation degree obviously should give higher weight, is assigned to the probability of this theme promoting lexical item.For weak relevant lexical item, should not be assigned under this theme, it should give a very low weight.In natural language processing, PMI algorithm is often used in the dependency calculating two lexical items, and it is defined formula (1):

$\mathrm{PMI}(t,z)={\log }_2\frac{P_{t\&z}}{P_t{P}_z}---\left(1\right)$

Wherein, P_t&zIt is descriptor z and the lexical item w Joint Distribution probability of co-occurrence, P in text collection_tIt is the probability that occurs in text collection of lexical item, P_zIt it is the descriptor z probability occurred in text collection.

The dependency obtained through above-mentioned steps is successive value, also can be integrated directly into the excavation carrying out theme feature in topic model in theory as priori.But in order to improve the Topics Crawling efficiency of topic model, dependency PMI will be carried out sliding-model control by this patent.According to dependency size, relevant degree is divided into Three Estate: strong correlation SR, generally relevant GR, weak relevant WR.

This patent adopts the discretization method based on cluster (K-means), and setting cluster number K (i.e. discrete segment number) is 5.Concrete operating procedure is shown in algorithm 1:

Expert carries out the process of fault diagnosis and often carries out fault location and fault category judgement according to the combination of phenomenon or several phenomenon.For mobile unit fault diagnosis field, " phenomenon " occurred after theme fault just described in phenomenon of the failure description, and one phenomenon of the failure describes and is often bound to several interrelated vocabulary occur, what we vocabulary can necessarily occur with these represents the theme occurred in description, and we claim this inevitable relation to be Strong-Link.Such that it is able to realize PMI algorithm to excavate theme related term and public word and to calculate their weight.

By the summary of experience of mobile unit domain-specialist knowledge and field engineer, the phenomenon of the failure that high speed railway vehicle mounted ATP often occurs, the most of theme namely occurred, summing up to get up to have T=44 class, phenomenon list 1 is as follows:

The theme of each phenomenon of the failure of table 1. (theme) is correlated with lexical item table

1.2 priori topic models

The theme update probability of LDA model can be expressed as formula (2):

$P(z_i=j|z_{-i},w,\mathrm{\α},\mathrm{\β})=\left(\frac{n_{-i,j}^{w_i}+\mathrm{\β}}{{\mathrm{\Σ}}_{w^{\′}}^W,n_{-i,j}^{w^{\′}}+W\mathrm{\β}}\right)\left(\frac{n_{-i,j}^{d_i}+\mathrm{\α}}{{\mathrm{\Σ}}_j^T{n}_{-i,j}^{d_i}+T\mathrm{\α}}\right)---\left(2\right)$

Wherein, z_i=j represents i-th lexical item in document and is assigned under theme j, z_-iThe theme representing out the every other word outside i-th word distributes；It it is the word w ' number of times being assigned under theme j；It is document d_iLower word i is assigned to the number of theme j；The Dirichlet that α and β is document subject matter and theme lexical item respectively is distributed hyper parameter.

But, LDA model is unsupervised Topics Crawling model.There is certain blindness in Topics Crawling process, Topics Crawling accuracy rate is relatively low.Extracted by above-mentioned priori and obtain Prior knowledge matrix DM.By by these priori DM_ij∈ DM is incorporated into the Topics Crawling process of LDA model, it is clear that improving the Topics Crawling process of model, result of the test also demonstrates this point.The main thought of priori LDA is by integrating priori DM_ij, revise theme update probability.It means that in theme renewal process, formula (2) will be multiplied by an extra saturation by us, such as formula (3):

$P(z_i=j|z_{-i},w,\mathrm{\α},\mathrm{\β})={\mathrm{DM}}_{ij}*\left(\frac{n_{-i,j}^{w_i}+\mathrm{\β}}{{\mathrm{\Σ}}_{w^{\′}}^W{n}_{-i,j}^{w^{\′}}+W\mathrm{\β}}\right)\left(\frac{n_{-i,j}^{d_i}+\mathrm{\α}}{{\mathrm{\Σ}}_j^T{n}_{-i,j}^{d_i}+T\mathrm{\α}}\right)---\left(3\right)$

Shown in formula (2) and formula (3), DM_ijTopics Crawling process plays critically important effect, represent priori to theme feature extract process have crucial effect.For priori LDA model, it is necessary first to set theme number x, the priori number t of the phenomenon that breaks down can be obtained according to fault data collection.Assuming that x meets original lexical item space lexical item number > x >=min (t, T).Secondly, lexical item document matrix is inputted priori LDA and carries out theme feature extraction by us, and then obtains theme feature space and subject document matrix.

2 based on Stratified Strategy method for diagnosing faults

The grader based on SVM is selected to carry out intelligent diagnostics.Based on the grader of SVM, there is general generalization ability preferably on the one hand, on the other hand, it is possible to the complexity that reply mobile unit failure diagnostic process phenomenon of the failure reveals as mark sheet.

For mobile unit two-stage fault mode, take substep diagnosis policy, refer to accompanying drawing 2.First, spatially build SVM classifier A at the theme feature extracted, carry out level fault pattern (FFP) fault diagnosis；Then, the diagnostic result of level fault pattern is constituted new feature space as new feature and theme feature Space integration by us, builds SVM classifier B, carry out secondary failure pattern (SFP) fault diagnosis on this feature space.

Result of the test

For RBFSVM grader, by its comparison with the method for other feature extractions three kinds traditional, observe the classification performance of priori topic model.Wherein, feature space includes VSM primary features space, i.e. original lexical item space；TFIDF feature space；It is not added with the LDA feature space of priori, i.e. BLDA feature space.

For the LDA feature space of priori, we are by the comparison of it and other three kinds of classical taxonomy devices, carry out comprehensive verification it is proposed that method for diagnosing faults feasibility.Including the grader based on BP neutral net, based on KNN grader (KNN), based on BAYESIAN NETWORK CLASSIFIER (BN), observe the classification performance of the grader based on RBFSVM；With this, in our test, carried out the comprehensive assessment of diagnosis effect by F-measures.

Based on SVM classifier at different characteristic classifying quality spatially

The classification results of grader requires over setting confidence interval and is divided into positive class and negative class.We check diagnosis effect to be that to adopt the F1 of overall classification be leading indicator, and precision and two auxiliary characteristicss of recall rate are verified.When table 2 and table 3 respectively describe using the grader of SVM structure as diagnostic tool, diagnosis effect to level fault classification and secondary failure classification on different feature spaces.

Table 2: based on SVM classifier in different characteristic level fault pattern classification effect spatially

Table 3: based on SVM classifier in different characteristic secondary failure pattern classification effect spatially

From upper table we it follows that

(1), under PLDA feature space, the classifying quality of grader is best, and level fault model F 1-measure value can reach 81.22%, and secondary failure model F 1-measure value can reach 69.43%.It, when ensureing stable accuracy, improves recall rate.

(2) by comparing BLDA feature space and PLDA feature space, it can be seen that after adding priori, classifying quality is had greatly improved by the feature space of priori LDA model extraction.

The diagnosis effect of each grader under 7.3PLDA theme space

Table 4 shows, by comparing other three main flow graders, i.e. K-NearestNeighbor (KNN) grader, grader based on BP neutral net, based on the grader of Bayesian network, assess our the grader effect to the diagnosis under PLDA theme feature space of the level fault pattern.Table 5 shows, by comparing other four main flow graders, i.e. K-NearestNeighbor (KNN) grader, grader based on BP neutral net, grader based on Bayesian network, and do not consider the SVM classifier (SVMOFF) of convergence strategy, assess our the grader effect to the diagnosis under PLDA theme feature space of the secondary failure pattern.

Table 4: level fault pattern is the diagnosis effect of each grader under PLDA theme space

Table 5: secondary failure pattern is the diagnosis effect of each grader under PLDA theme space

Table 4 shows: for level fault pattern, the grader performance based on SVM is better than other graders, and reaching F1-measure value is 81.22%, and precision is the good classification effect of 82.02% recall rate 74.18%.

Secondly, table 5 shows:

(1) although the precision of SVM classifier is not best, but it has the ability to weigh precision and recall rate and reach whole structure optimum

(2) owing to we consider the feature of VOBE self, and the strategy of Feature Fusion is taked, thus improve the diagnosis effect of secondary failure pattern.

Therefore, we can show that the grader based on SVM has higher reliability and efficiency for VOBE fault diagnosis.

Last it is noted that obvious, above-described embodiment is only for clearly demonstrating example of the present invention, and is not the restriction to embodiment.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here without also cannot all of embodiment be given exhaustive.And the apparent change thus extended out or variation are still among protection scope of the present invention.

Claims (6)

1. the Train Operation Control System Onboard device intelligence method for diagnosing faults based on priori topic model, it is characterised in that described method mainly includes feature extraction and two key steps of intelligent diagnostics.

2. diagnostic method according to claim 1, it is characterized in that, described feature extraction refers to that the mobile unit fault signature based on priori topic model extracts, and describes document from lexical item spatial alternation to theme feature space by priori LDA by the phenomenon of the failure of mobile unit.

3. diagnostic method according to claim 2, it is characterised in that the priori in fault case library text is extracted, and the theme update probability that priori topic model is LDA model can be expressed as formula (2):

$P(z_i=j|z_{-i},w,\mathrm{\α},\mathrm{\β})=\left(\frac{n_{-i,j}^{w_i}+\mathrm{\β}}{{\Σ}_{w^{\′}}^W{n}_{-i,j}^{w^{\′}}+W\mathrm{\β}}\right)\left(\frac{n_{-i,j}^{d_i}+\mathrm{\α}}{{\Σ}_j^T{n}_{-i,j}^{d_i}+T\mathrm{\α}}\right)---\left(2\right)$

Wherein, z_i=j represents i-th lexical item in document and is assigned under theme j, z_-iThe theme representing out the every other word outside i-th word distributes；It it is the word w ' number of times being assigned under theme j；It is document d_iLower word i is assigned to the number of theme j；The Dirichlet (Di Li Cray) that α and β is document subject matter and theme lexical item respectively is distributed hyper parameter.

4. diagnostic method according to claim 3, it is characterised in that the main thought of priori LDA is by integrating priori DM_ij, revise theme update probability, in theme renewal process, formula (2) will be multiplied by an extra saturation by us, such as formula (3):For priori LDA model, firstly the need of setting theme number x, the priori number t of the phenomenon that breaks down can be obtained according to fault data collection, assuming that x meets original lexical item space lexical item number > x >=min (t, T), secondly, lexical item document matrix is inputted priori LDA and carries out theme feature extraction by us, and then obtains theme feature space and subject document matrix.

5. diagnostic method according to claim 4, it is characterized in that, based on Stratified Strategy method for diagnosing faults, the grader based on SVM is selected to carry out intelligent diagnostics, based on the grader of SVM, there is general generalization ability preferably on the one hand, on the other hand, it is possible to the complexity that reply mobile unit failure diagnostic process phenomenon of the failure reveals as mark sheet.

6. diagnostic method according to claim 5, it is characterised in that for mobile unit two-stage fault mode, take substep diagnosis policy, first, spatially build SVM classifier A at the theme feature extracted, carry out level fault pattern (FFP) fault diagnosis；Then, the diagnostic result of level fault pattern is constituted new feature space as new feature and theme feature Space integration by us, builds SVM classifier B, carry out secondary failure pattern (SFP) fault diagnosis on this feature space.