CN102564496A - Micro-analysis method for transformer device based on BP nerve network and manual shoal - Google Patents

Abstract

A micro-analysis method for a transformer device based on a BP nerve network and a manual shoal comprises the following steps: obtaining the data such as furfural, brackish water, disruptive voltage and acid value; optimizing the weight and the threshold of the BP nerve network by utilizing a manual shoal optimizing AFSO algorithm, so as to define the weight and the threshold; extracting the state micro-analysis result of a breakdown case base of the transformer device to train the BP nerve network; and inputting the newly obtained furfural, brackish water, disruptive voltage, acid value, hydrogen content, acetylene content, methane content, ethylene content, commissioning time, breakdown frequency in the recent three years, and breakdown grade in the recent three years into the BP nerve network for micro-analysis of the state of the transformer device, so as to define the state of the transformer device. The method has the advantages of scientificity, efficiency and accuracy.

Description

Converting equipment micro-analysis method based on BP neural network and artificial fish-swarm

Technical field

The present invention relates to converting equipment state micro-analysis method, relate in particular to a kind of converting equipment state micro-analysis method based on BP neural network and artificial fish school optimization.

Background technology

Grid company is as capital-intensive enterprise, and its core competitiveness is assets efficiency maximization and minimization of cost.From early stage fault repair afterwards, to stressing the preventive maintenance of maintenance in advance, the consciousness of grid equipment assets fine-grained management is progressively set up.The micro-analysis of converting equipment state is the important foundation of realization company converting equipment life-cycle cycle management and fine-grained management.

The converting equipment lean management is a systems engineering, and it is a little many to involve a wide range of knowledge, and is technical strong, is the basis of carrying out the converting equipment scientific management, is the foundation stone of realizing the converting equipment reliability service, is to realize converting equipment security maximization, the optimized approach of economy.Converting equipment state microcosmic situation is the important evidence of equipment lean management, and the accuracy of state micro-analysis directly influences the science of equipment control.On the basis that fully obtains status information of equipment, apparatus for establishing state micro-analysis method, the intellectuality of raising state micro-analysis can effectively improve the accuracy and the efficient of the micro-analysis of equipment health status.

There are a lot of method for diagnosing faults to be suggested now like expert system, SVMs, Bayesian network or the like; And the BP neural network has a wide range of applications in electric system pattern-recognition, nonlinear optimization and correlation predictive field with Nonlinear Mapping and self-organization, the self-learning capability of its height.Neural network is the complex network that extensively connects with a large amount of simple processing units; In order to simulating human cerebral nerve network structure and behavior; It is " microcosmic " behavioral study in; Its essence is exactly the information processing function that is used for simulating human brain, and it need not to set up any physical model, has self-organization, self-learning capability; Through scene a large amount of master sample study and training, constantly adjust its weights and threshold value, the knowledge implicit expression of obtaining is distributed on the whole network, and realizes its schematic memory.When the input of network during near training sample. the output of network is just exported near sample; When input and output do not match, can increase new sample on the original basis and continue study, improve precision, to reach the mapping relations of the input and output that meet the requirements.Though neural network has kilter micro-analysis ability, the BP algorithm the convergence speed is slow, is prone to be absorbed in local extremum.

The artificial fish school intelligent optimizing algorithm has good overcome local extremum and the ability that obtains global extremum, and it is applied to having obtained gratifying result in the Robust PID Controller Parameter Optimization at first.

To the problem that above-mentioned prior art exists, the present invention utilizes the artificial fish school optimization algorithm to confirm the parameter of BP neural network, utilizes the state micro-analysis result of converting equipment fault case library to train the BP neural network then; The data message of the converting equipment furfural that newly gets access to, Wei Shui, voltage breakdown, acid number, hydrogen, acetylene, methane, ethene, acetylene, the time of putting into operation, nearest 3 years fault frequency, nearest 3 years fault levels is input to carries out the micro-analysis of converting equipment state in the BP neural network; With definite converting equipment state, thereby reach the purpose of converting equipment state microcosmic situation being carried out intellectual analysis.The present invention not only can reduce blindness and select the accuracy of BP neural network parameter to state micro-analysis result's influence and the micro-analysis of raising converting equipment state.Meanwhile, the present invention can realize not relying on the problem of electric power expert to micro-analysis of converting equipment state and judgement.

Summary of the invention

In order to improve intelligence, to analyze converting equipment state microcosmic situation efficiently, reliably, the present invention proposes a kind of converting equipment state micro-analysis method based on BP neural network and artificial fish school optimization, comprise the steps:

Based on the converting equipment micro-analysis method of BP neural network and artificial fish-swarm, characteristic of the present invention is that step is:

1), obtains the data message of converting equipment furfural, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, acetylene content, the time of putting into operation, nearest 3 years fault frequency, nearest 3 years fault levels respectively;

2), utilize artificial fish school optimization AFSO algorithm that BP neural network weight and threshold value are optimized, confirm the value of weights and threshold value;

3) the state micro-analysis result who, extracts converting equipment fault case library trains the BP neural network;

4), the data message of the converting equipment furfural that will newly get access to, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, the time of putting into operation, nearest 3 years fault frequency, nearest 3 years fault levels is input to and carries out the micro-analysis of converting equipment state in the BP neural network, to confirm the converting equipment state.

Wherein the BP neural network is made up of input node, output node and latent node, and the input vector of supposing input layer is X=(x ₁, x ₂..., x _m), then the output vector of input layer equals the input vector of latent layer, and then the input vector of latent layer is expressed as:

${\mathrm{Net}}_j=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_i-{\mathrm{\θ}}_j,j=\mathrm{1,2},..,k$

Wherein, w _IjBe the input layer vector x _iTo latent course amount o _jThe connection weight w _Ij, O={o ₁, o ₂.., o _j.Net _jBe j neuron input vector of latent layer; θ _jBe j neuron threshold value of latent layer.Therefore, the computing formula of the input vector of latent layer is:

$o_j=f\left({\mathrm{Net}}_j\right)=f(\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_i-{\mathrm{\θ}}_j)j=\mathrm{1,2},..,k$

F () is an action function, and action function is selected the Sigmoid function among the present invention, and its expression formula is:

f(x)＝1/(1+e ^-x)

In like manner, the input vector of output layer is:

${\mathrm{Net}}_l=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{jl}}{o}_j-{\mathrm{\θ}}_l,l=\mathrm{1,2},..,n.$

Wherein, w _JlBe latent course amount o _jTo output layer vector y _l, Y={y ₁, y ₂.., y _lThe connection weights.Net _lBe l neuron input vector of output layer; θ _lBe l neuron threshold value of output layer.Therefore, the computing formula of the output vector of output layer is:

$y_l=f\left({\mathrm{Net}}_l\right)=f(\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{jl}}{o}_j-{\mathrm{\θ}}_l)l=\mathrm{1,2},..,n$

F () is an action function in the following formula, and action function remains the Sigmoid function.

Training set A={ (X _t, T _t) | t=1,2 ..., p}, X _t={ x _T1, x _T2..., x _TmBe t input of training set data; T _t={ t _T1, t _T2..., t _TnIt is corresponding training set desired output vector.Utilize formula $o_j=f\left({\mathrm{Net}}_j\right)=f(\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{Ij}}{x}_i-{\mathrm{\θ}}_j)j=\mathrm{1,2},..,k$ With $y_l=f\left({\mathrm{Net}}_l\right)=f(\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{Jl}}{o}_j-{\mathrm{\θ}}_l)l=\mathrm{1,2},..,n$ Calculate p input sample, the output quantity of actual BP neural network is Y _p={ y _P1, y _P2..., y _Pn.Thereby, with respect to the error functions in the training sample be:

$E=(\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{(t_l-y_l)}^2/n)$

Utilize above-mentioned formula that BP neural network weight and threshold value are adjusted, choose different training sample patterns, constantly carry out above-mentioned iterative process, till reaching requirement.

Though neural network has kilter micro-analysis ability, the BP algorithm the convergence speed is slow, is prone to be absorbed in local extremum.Because artificial fish intelligent optimization algorithm has and good overcomes local extremum and the ability that obtains global extremum, and the realization of this optimized Algorithm need not the characteristics such as Grad of objective function, so it has certain adaptive ability to the search volume.To the problem that the BP algorithm exists, the present invention proposes to utilize artificial fish school optimization, and (Artificial Fish School Optimization, AFSO) algorithm obtains the BP neural network parameter.

Artificial fish school intelligent optimizing AFSO algorithm is a kind of intelligent optimization algorithm based on simulation shoal of fish behavior, and it has adopted method for designing from bottom to top, and the form and the character in optimizing space is not had specific (special) requirements.The individual model that the present invention is configured to artificial fish respectively with the weights and the threshold value of BP neural network; Artificial then fish individuality is selected suitable behavior adaptively in the process of optimizing, the weights of final BP neural network and the result of global optimum of threshold value obtain through each individual local optimal searching in the shoal of fish.

A. food concentration

Food concentration is to weigh the good and bad sign in the current present position of artificial fish, and its effect is similar to the fitness in the particle swarm optimization algorithm.In the present invention, food concentration is the good and bad sign of weights and the current present position of threshold value of weighing the BP neural network.The present invention is according to the weights of BP neural network and the threshold value singularity as the individual model of artificial fish, and the food concentration evaluation function that is adopted is:

$f\left(X_i\right)=\frac{1}{N}(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{X}_i^2-N{\stackrel{\‾}{X}}^2)$

In the formula, X _i,

Represent the average of all current present positions of artificial fish in the artificial current present position of fish of i bar in the shoal of fish and the shoal of fish respectively, N representes the size of the shoal of fish.The B foraging behavior: foraging behavior in the present invention is exactly that weights and the threshold value of BP neural network followed a kind of behavior that the many directions of food are moved about according to food concentration as fish.Artificial fish current state X _iAt its sensing range (d _{I, j}=‖ X _i-X _jSelect a dbjective state X at random in the ‖≤Visual) _jIf, the food concentration Y of this state _jBe superior to current state X _iFood concentration Y _i, then press X _Inextm=X _Im+ Random (AF_Step) (X _Jm-X _Im)/d _{I, j}Y _i＜Y _j, to this state X _jDirection takes a step forward; Otherwise, X _iSelection mode X at random again in its sensing range _j, judge whether to satisfy the condition of advancing.Make repeated attempts after try_number time,, then press X as still not satisfying the progress bar part _Inextm=X _Im+ Random (AF_Step) Y _i>=Y _j, move at random and move a step.Wherein, m=1,2; X _Jm, X _ImAnd X _InextmRepresent state vector X respectively _i, X _jAnd next step state vector of artificial fish X _InextmM component.Random number between Random (AF_Step) expression [0, AF_Step].Below symbol implication in various identical therewith.The C behavior of bunching: the behavior of bunching in the present invention refers to according to the weights of BP neural network and formed many artificial fishes of threshold value individual, in the process of moving about, closes on the center of fellows as far as possible towards periphery and moves, to avoid overcrowding.Artificial fish is to current state X _iSearch for its sensing range d _{I, j}Interior number of partners n _fIf, (n _f/ N)＜and δ, this shows that there is more food and not too crowded at the partner center, then is calculated as follows partner center X _c

$X_{\mathrm{cm}}=\left(\underset{j=1}{\overset{n_f}{\mathrm{\Σ}}}{X}_{\mathrm{jm}}\right)/n_f$

In the formula, X _CmExpression center state vector X _cM element; X _JmRepresent j (j=1,2 ..., n _f) individual partner's m component.Calculate the food concentration Y of this center _cIf, Y _i＜Y _c, show that then partner center degree of safety is higher and not too crowded, take a step forward to partner's center direction by following formula; Otherwise execution foraging behavior.

X _inextm＝X _im+Random(AF_Step)(X _cm-X _im)/d _i，c

The D behavior of knocking into the back: the behavior of knocking into the back in the present invention refer to according to the weights of BP neural network and threshold value and many artificial fish individualities that form to the behavior that the person chases that do not enliven most that closes on.Explore artificial fish current state X _iThe optimum neighbours X of sensing range internal state _MaxIf, Y _i＜Y _Max, and X _MaxPartner's number n in the field _f, satisfy (n _f/ N)＜and δ, this shows X _MaxNear more food and not too crowded is arranged, then press following formula to X _MaxLocality take a step forward; Otherwise execution foraging behavior.

X _inextm＝X _im+Random(AF_Step)(X _maxm-X _im)/d _i，max

In the formula, X _MaxmExpression state vector X _MaxM component.

The E bulletin board is to be used for writing down the individual state of optimum artificial fish.Every artificial fish individuality is in searching process; Each action finishes and all will check the state of oneself state and bulletin board; If oneself state is superior to the bulletin board state, just the state with bulletin board replaces with oneself state, like this; Bulletin board is just noted historical optimum state, the i.e. optimal value of the weights of BP neural network and threshold value.

Through we can find out to the behavior description of artificial fish, every artificial fish is explored its present located environmental aspect, thereby selects a suitable behavior, makes and advances the fastest to optimal direction.Finally, artificial fish concentrate at several local extremums around, and generally can assemble more artificial fish around the more excellent extremal region of value.

The practical implementation step of artificial fish school intelligent optimizing algorithm is following.

Step 1: the initialization shoal of fish.Artificial fish current state, individual extreme value, shoal of fish optimal location and the relevant shoal of fish parameter set when adopting initialization of the present invention;

Step 2: utilize formula

to calculate the food concentration of artificial fish current state;

Step 3: at first to every artificial fish in the shoal of fish calculate foraging behavior respectively, the individual extreme value of the bunch behavior and the three kinds of behaviors such as behavior of knocking into the back;

Step 4: the individual extreme value to three kinds of behaviors of every artificial fish compares, and carries out the behavior of corresponding optimum extreme value;

Step 5:, then stop iterative computation if continuous five shoal of fish optimal locations do not upgrade; Otherwise continue to select new behavior and turn to step 3.

With utilizing the artificial fish school optimization algorithm to confirm that the weights of BP neural network and threshold data are input in the BP neural network, and utilize the state micro-analysis result of converting equipment defective (fault) case library to train the BP neural network; The data message of the converting equipment furfural that newly gets access to, Wei Shui, voltage breakdown, acid number, hydrogen, acetylene, methane, ethene, acetylene, the time of putting into operation, nearest 3 years defective (fault) frequency, nearest 3 years defective (fault) grades is input to carries out the micro-analysis of converting equipment state in the BP neural network; With definite converting equipment state, thereby reach the purpose of converting equipment state microcosmic situation being carried out intellectual analysis.

Further specify content of the present invention below in conjunction with accompanying drawing and instance.

Description of drawings

Fig. 1 is the converting equipment state micro-analysis method flow diagram that the present invention is based on BP neural network and artificial fish school optimization.

Embodiment

Shoal of fish size to each interative computation of the weights of BP neural network and threshold value among the present invention is set at: N=18; Artificial fish scope of activities is set at: scope_x11=352, scope_x12=0, scope_x21=288, scope_x22=0; The step-length that artificial fish is moved is set at: AF Step=0.005; The sensing range of artificial fish is set at: Visual=0.01; Maximum exploration number of times was set at when artificial fish was looked for food at every turn: try_number=30; The crowding factor of the shoal of fish is set at: δ=0.618; BP neural network input quantity is 14, and output quantity is 4, the three-layer network structure.

The present invention carries out initialization at random to the weights and the threshold value of BP neural network, and forms the data set E={X that comprises the artificial fish of N bar respectively _i, wherein

Represent the position of the artificial fish of i bar, and its utilization is carried out initialization to fish-swarm algorithm.The individual extreme value of every artificial fish in the shoal of fish

I=1,2 ..., N, and shoal of fish optimal location P _Gmost={ P _GmostInitial value all equal the weights of BP neural network and the initial position of threshold value.The original state of bulletin board is identical with the initial optimum state of the shoal of fish.

1, food concentration is to weigh the good and bad sign in the current present position of artificial fish

Food concentration is to weigh the good and bad sign in the current present position of artificial fish, and its effect is similar to the fitness in the particle swarm optimization algorithm.In the present invention, food concentration is the good and bad sign of weights and the current present position of threshold value of weighing the BP neural network.The present invention is according to the weights of BP neural network and the threshold value singularity as the individual model of artificial fish, and the food concentration evaluation function that is adopted is:

$Y=f\left(X_i\right)=\frac{1}{N}(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{X}_i^2-N{\stackrel{\‾}{X}}^2)$

In the formula, X _i,

Represent the average of all artificial fish current locations in the artificial current present position of fish of i bar in the shoal of fish and the shoal of fish respectively, N representes the size of the shoal of fish.

2, artificial fish school intelligent optimizing algorithm

In the present invention; Utilize the artificial fish school optimization algorithm optimization to confirm the weights and the threshold value of BP neural network; The blindness of BP neural network weight and threshold value is selected to make the BP neural network be absorbed in local extremum avoiding, existing method is not very desirable shortcoming calculating aspect consuming time and the effect.This optimized Algorithm is imitated the looking for food, bunching of the shoal of fish and the behavior of knocking into the back through constructing artificial fish, thereby realizes seeking the weights and the optimum purpose that is provided with of threshold value of BP neural network.

(1) foraging behavior: foraging behavior in the present invention is exactly that weights and the threshold value of BP neural network followed a kind of behavior that the many directions of food are moved about according to food concentration as fish.Artificial fish current state X _iAt its sensing range (d _{I, j}=‖ X _i-X _jSelect a dbjective state X at random in the ‖≤Visual) _jIf, the food concentration Y of this state _jBe superior to current state X _iFood concentration Y _i, then press X _Inextm=X _Im+ Random (AF_step) (X _Jm-X _Im)/d _{I, j}, Y _i＜Y _jTo this state X _jDirection takes a step forward; Otherwise, X _iSelection mode X at random again in its sensing range _j, judge whether to satisfy the condition of advancing.Make repeated attempts after try_number time,, then press X as still not satisfying the progress bar part _Inextm=X _Im+ Random (AF_Step), Y _i>=Y _jMove at random and move a step.Wherein, m=1,2; X _Jm, X _ImAnd X _InextmRepresent state vector X respectively _i, X _jAnd next step state vector of artificial fish X _InextmM component.Random number between Random (AF_Step) expression [0, AF_Step].Below symbol implication in various identical therewith.

(2) behavior of bunching: it is individual that the behavior of bunching in the present invention refers to many artificial fishes that form according to the weights of BP neural network and threshold value, in the process of moving about, closes on the center of fellows as far as possible towards periphery and move to avoid overcrowding.Artificial fish is to current state X _iSearch for its sensing range d _{I, j}Interior number of partners n _fIf, (n _f/ N)＜and δ, this shows that there is more food and not too crowded at the partner center, then presses

Calculate partner center X _c, wherein, X _CmExpression center state vector X _cM element; X _JmRepresent j (j=1,2 ..., n _f) individual partner's m component.Calculate the food concentration Y of this center _cIf, Y _i＜Y _c, show that then partner center degree of safety is higher and not too crowded, by formula X _Inextm=X _Im+ Random (AF_Step) (X _Cm-X _Im)/d _{I, c}Direction takes a step forward to partner's center; Otherwise execution foraging behavior.

(3) behavior of knocking into the back: the behavior of knocking into the back in the present invention refers to according to the weights of BP neural network and threshold value and many artificial fish individualities that form chase after the behavior of catching to the person that do not enliven most who closes on.Explore artificial fish current state X _iThe neighbours X that its sensing range internal state is optimum _MaxIf, Y _i＜Y _Max, and X _MaxPartner's number n in the field _f, satisfy (n _f/ N)＜and δ, this shows X _MaxNear more food and not too crowded is arranged, then by formula X _Inextm=X _Im+ Random (AF_Step) (X _Maxm-X _Im)/d _{I, max}(wherein, X _MaxmExpression state vector X _MaxM component) to X _MaxLocality take a step forward; Otherwise execution foraging behavior.

(4) bulletin board is to be used for writing down the individual state of optimum artificial fish.Every artificial fish individuality is in searching process; Each action finishes and all will check the state of oneself state and bulletin board; If oneself state is superior to the bulletin board state, just the state with bulletin board replaces with oneself state, like this; Bulletin board is just noted historical optimum state, the i.e. weights of BP neural network and the optimal value in the threshold value.

Through we can find out to the behavior description of artificial fish, every artificial fish is explored its present located environmental aspect, thereby selects a suitable behavior, makes and advances the fastest to optimal direction.Finally, artificial fish concentrate at several local extremums around, and generally can assemble more artificial fish around the more excellent extremal region of value.The practical implementation step of artificial fish school intelligent optimizing algorithm is following.

Step 1: the initialization shoal of fish.Artificial fish current state, individual extreme value, shoal of fish optimal location and the relevant shoal of fish parameter set when adopting initialization of the present invention;

Step 2: utilize formula

to calculate the food concentration of artificial fish current state;

Step 3: at first to every artificial fish in the shoal of fish calculate foraging behavior respectively, the individual extreme value of the bunch behavior and the three kinds of behaviors such as behavior of knocking into the back;

Step 4: the individual extreme value to three kinds of behaviors of every artificial fish compares, and carries out the behavior of corresponding optimum extreme value;

Step 5:, then stop iterative computation if continuous five shoal of fish optimal locations do not upgrade; Otherwise continue to select new behavior and turn to step 3.

As shown in Figure 1, this figure has provided the converting equipment state micro-analysis method flow diagram based on BP neural network and artificial fish school optimization.This method comprises following step:

1, obtains the data message of converting equipment furfural, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, acetylene content, the time of putting into operation, nearest 3 years defective (fault) frequency, nearest 3 years defective (fault) grades respectively;

2, utilize artificial fish school optimization AFSO algorithm that BP neural network weight and threshold value are optimized, confirm the value of weights and threshold value;

3, the state micro-analysis result who extracts converting equipment defective (fault) case library trains the BP neural network;

The data message of the converting equipment furfural that 4, will newly get access to, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, acetylene content, the time of putting into operation, nearest 3 years defective (fault) frequency, nearest 3 years defective (fault) grades is input to and carries out the micro-analysis of converting equipment state in the BP neural network, to realize intelligence, to confirm the converting equipment state efficiently.

Claims (1)

1. based on the converting equipment micro-analysis method of BP neural network and artificial fish-swarm, it is characterized in that step is:

1), obtains the data message of converting equipment furfural, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, acetylene content, the time of putting into operation, nearest 3 years fault frequency, nearest 3 years fault levels respectively;

2), utilize artificial fish school optimization AFSO algorithm that BP neural network weight and threshold value are optimized, confirm the value of weights and threshold value;

3) the state micro-analysis result who, extracts converting equipment fault case library trains the BP neural network;

4), the data message of the converting equipment furfural that will newly get access to, Wei Shui, voltage breakdown, acid number, hydrogen content, acetylene content, methane content, ethylene contents, the time of putting into operation, nearest 3 years fault frequency, nearest 3 years fault levels is input to and carries out the micro-analysis of converting equipment state in the BP neural network, to confirm the converting equipment state.

Images