CN103235206A - Transformer fault diagnosis method - Google Patents

Abstract

The invention provides a neural network based transformer fault diagnosis method. The method includes: taking parameters in a transformer as the object of study, and utilizing an artificial neutral network model for data acquisition, fault training, analysis and the like. Experimental results show that the transformer fault diagnosis method has the advantages of high convergence speed, high stability and strong sample adding capability.

Description

A kind of method for diagnosing faults of transformer

Technical field

The present invention relates to a kind of method for diagnosing faults of transformer, belong to power electronics and power technology field.

Technical background

Fault diagnosis starts from the mechanical equipment fault judgement, modern comfort technical merit and complexity improve constantly, equipment failure also significantly increases the influence of producing, therefore to guarantee that equipment is reliable, operation effectively, give full play to its benefit, must develop fault diagnosis technology, fault diagnosis technology was tested by means of the modern times, means such as monitoring and computing machine, research equipment is in operation or the relative status information under the quiescent conditions, the state of the art of analytical equipment is diagnosed character and the cause of its fault, and predicts fault and potential safety hazard, avoid unnecessary loss, thereby have very high economic and social benefit.

When in various degree fault takes place in operating transformer, can produce abnormal occurrence or information, fault analysis is exactly abnormal occurrence or the information of collecting transformer, analyzes according to these phenomenons or information, thus the type of failure judgement, the order of severity and trouble location, therefore, the purpose of transformer fault diagnosis at first is accurately to judge current normal condition or the abnormality of being in of operational outfit, if transformer is in abnormality fault is arranged, the then character of failure judgement, type and reason.For example be insulation fault, overheating fault or mechanical fault if insulation fault then is insulation ag(e)ing, are made moist, still discharge property fault.If discharge property fault is again the discharge of which kind of type etc., transformer fault diagnosis also will be according to failure message or according to the information processing result, may developing namely the order of severity of fault of prediction fault, development trend is made diagnosis, proposes the measure of control fault, prevent and eliminate fault, propose rational method and the corresponding anti-accident measures of maintenance of equipment, to the design of equipment, make, propositions such as assembling change 1 and advance suggestion, for the equipment modernization management provides scientific basis and suggestion.

Be the important means of power transformer interior fault diagnosis to the transformer oil dissolved gas, current extensive application be the improvement three-ratio method, but utilize three-ratio method to have the deficiency of two aspects as the criterion of transformer fault diagnosis, be that so-called encoding impairment and critical value criterion are damaged, the artificial neuron grid is with advantages such as its distributed parallel processing, self-adaptation, self study, associative memory and Nonlinear Mapping, for address this problem opened up newly by way of.

Summary of the invention

The invention provides a kind of transformer fault diagnosis method based on neural network, the parameter in the transformer as research object, is utilized the artificial neural network type, carry out data acquisition, the fault training is analyzed etc., and experimental result shows.This fault diagnosis algorithm fast convergence rate, stability is high, the advantage that the sample supplemental capabilities is strong.

For achieving the above object, the technical scheme that the present invention takes is: its neural network model as shown in Figure 1, comprise input layer, mode layer, summation layer, and output layer, its theoretical foundation is Bayes's minimum risk criterion, and input layer is received from the value of training sample, imitates proper vector and passes to network, its neuron number and sample vector dimension equate that each mode unit is output as

$f(X,W_i)=\exp \left[\frac{-{(X-W_i)}^T(X-W_i)}{2{\mathrm{\ξ}}^2}\right]---\left(1\right)$

In the formula, W _iBe the weights that input layer connects to mode layer, ζ is smoothing factor, and this method can be described as:

Suppose that two kinds of known fault modes are θ _A, θ _B, for the fault signature sample that will judge

X＝(x ₁，x ₂…，x _n)，

If h _Al _Af _A(X)＞h _Bl _Bf _B(X), X ∈ θ then _A

If h _Al _Af _A(X)＜h _Bl _Bf _B(X), X ∈ θ then _B

Its design cycle is:

Technique effect of the present invention: the invention provides a kind of transformer fault diagnosis method based on neural network, this algorithm the convergence speed is fast, stability is high, the sample supplemental capabilities is strong, in actual applications, need set up transformer fault sample storehouse, increase along with transformer fault, change and change, in sum, this new neural network transformer fault diagnosis method is in diagnosis speed, appends the sample ability and aspect performances such as accuracy rate of diagnosis in actual applications all are better than BP transformer fault diagnosis method.

Description of drawings

Fig. 1 is neural network model;

Fig. 2 is neural metwork training result among the embodiment;

Fig. 3 is neural metwork training error responses among the embodiment;

Fig. 4 predicts the outcome for transformer among the embodiment;

Embodiment

Embodiment:

For an operating transformer, gathering number is the matrix of 30*4, and preceding 3 classify improvement three-ratio method numerical value as, the 4th classifies fault category output as, use preceding 23 samples as train samples, back 7 conduct checking samples, its neural training result is respectively as Fig. 2, Fig. 3, shown in Figure 4.As can be seen from Figure 4, after training, as in the input substitution neural network, have only two samples to get the wrong sow by the ear training data.When predicting checking, have and have only two transformer faults, show the validity of this algorithm.

The invention provides a kind of transformer fault diagnosis method based on neural network, kind new neural network transformer fault diagnosis method is in diagnosis speed, appends the sample ability and aspect performances such as accuracy rate of diagnosis in actual applications all are better than other transformer fault diagnosis method.

Claims (3)

1. transformer fault diagnosis method, it is characterized in that: this method at first be just the parameter in the transformer as research object, utilize the artificial neural network type, carry out the training of data acquisition, fault, analyze experimental result and show, this method for diagnosing faults can accurately detect in the transformer faulty component.

2. method according to claim 1, it is characterized in that, its neural network model is as shown below, comprise input layer, mode layer, summation layer and output layer, its theoretical foundation is that value, the effect proper vector that Bayes's minimum risk criterion, input layer are received from training sample passes to network, and its neuron number and sample vector dimension equate.Each mode unit is output as

In the formula, W _iBe the weights that input layer connects to mode layer, ζ is smoothing factor.

3. method according to claim 1 is characterized in that, for the fault signature sample that will judge

X＝(x ₁，x ₂…，x _n)

If h _Al _Af _A(X)＞h _Bl _Bf _B(X), X ∈ θ then _A

If h _Al _Af _A(X)＜h _Bl _Bf _B(X), X ∈ θ then _B

Its design cycle is

。

Images