The AI Lends a Hand

---

Ener­gy net­work oper­a­tors and asset man­agers already have to com­pen­sate for the grow­ing loss of per­son­nel knowl­edge and the increas­ing main­te­nance require­ments of aging oper­at­ing equip­ment. At the same time, chal­lenges and bur­dens are increas­ing:

The urgent expan­sion of grids, the feed-in of renew­able ener­gies and the decen­tral­iza­tion of ener­gy sup­ply in the face of ever increas­ing ener­gy demands hard­ly seem man­age­able. A key role is there­fore being played by new tech­nolo­gies that help mon­i­tor and diag­nose assets. Such intel­li­gent data-dri­ven sys­tems make it pos­si­ble to use the exist­ing infra­struc­ture more effi­cient­ly and extend its ser­vice life. The per­fect inter­ac­tion of all automa­tion lev­els is cru­cial for this: from the sen­sor to the data node on the trans­former to the lev­el of glob­al asset man­age­ment.

“This is no longer a dream of the future; it is already pos­si­ble today,” explains Tobias Gru­ber, MR prod­uct man­ag­er who is involved in the devel­op­ment of algo­rithms and math­e­mat­i­cal train­ing meth­ods. “Rein­hausen already offers a large port­fo­lio of self-learn­ing sen­sors and sys­tems. And their capa­bil­i­ties are con­stant­ly increas­ing through mutu­al net­work­ing and an ever-grow­ing learn­ing curve.” The dig­i­tal­iza­tion of assets is a nec­es­sary step for the rea­sons men­tioned. But this trans­for­ma­tion process — the inde­pen­dent fur­ther devel­op­ment of the sys­tems them­selves – will bring about changes in all indus­tries, and also bring ben­e­fits that are not yet imag­in­able today. Since it is obvi­ous that AI is here to stay, it makes sense to use it to sup­port the ener­gy indus­try at all lev­els.

At the process lev­el, var­i­ous sen­sors record numer­ous sig­nals — reli­ably and pre­cise­ly — direct­ly at the trans­former and for­ward them to a cen­tral com­mu­ni­ca­tion node at the field lev­el for con­sol­i­da­tion. This pro­vides reli­able infor­ma­tion on the main­te­nance and health sta­tus of the respec­tive trans­form­ers. These eval­u­a­tions then come togeth­er at the con­trol lev­el where main­te­nance strate­gies for the entire fleet can be devel­oped. How­ev­er, this only works if the sen­sor tech­nol­o­gy is mod­u­lar­ly expand­able and man­u­fac­tur­er-inde­pen­dent so that it can react agile­ly to future needs in fleet main­te­nance and renew­al. And with such a high lev­el of data exchange between the lev­els, cyber­se­cu­ri­ty is of par­tic­u­lar impor­tance.

1. INTELLIGENT SENSOR TECHNOLOGY

Sen­sors col­lect infor­ma­tion – that much is obvi­ous. Smart sen­sors can also eval­u­ate it. But what does that mean in con­crete terms? Which data can be bet­ter eval­u­at­ed using machine learn­ing than with pre­vi­ous meth­ods? An impor­tant stan­dard mea­sure, for exam­ple, is the analy­sis of the insu­lat­ing oil in the gas phase (DGA – dis­solved gas analy­sis). In all com­mon analy­sis meth­ods, the gas is extract­ed, which expos­es it each time to fluc­tu­at­ing exter­nal fac­tors such as ambi­ent tem­per­a­ture, air pres­sure, humid­i­ty and much more. Ensur­ing that these val­ues always remain con­stant for mea­sure­ment would con­sume vast sums of mon­ey and sig­nif­i­cant­ly increase the com­plex­i­ty of DGA sys­tems.

This is where arti­fi­cial intel­li­gence comes into play. Using train­ing data sets, the algo­rithm in the smart DGA sys­tem learns the rela­tion­ships between the gas con­cen­tra­tion in the insu­lat­ing oil, the sen­sor sig­nal and the pos­si­ble dis­turb­ing influ­ences. With the aid of math­e­mat­i­cal-sta­tis­ti­cal meth­ods from the machine learn­ing tool­box, the DGA sen­sor is thus “trained” dur­ing its devel­op­ment phase. The new intel­li­gent spe­cial­ist is then installed on the trans­former and imme­di­ate­ly knows what to do. “With the lab­o­ra­to­ry analy­sis of the oil as a ref­er­ence point, the DGA sen­sor then also learns about its trans­former and con­tin­u­ous­ly recal­i­brates itself,” says Gru­ber. “In this way, effects such as sen­sor drift, aging of the insu­lat­ing oil and the like can be com­pen­sat­ed for, and con­sis­tent mea­sure­ment repeata­bil­i­ty ensured.”

Gas analy­sis for on-load tap-chang­ers looks a bit more dif­fi­cult, because in con­trast to the trans­former, few­er data sets on gas for­ma­tion rates and typ­i­cal gas con­cen­tra­tions have been devel­oped. “Ana­lyz­ing on-load tap-chang­ers requires a lot of expert knowl­edge about how the par­tic­u­lar tap-chang­er type works and how it oper­ates,” Gru­ber tells us. That’s why it doesn’t just come down to ana­lyz­ing the key gas­es. Rather, in this regard, sta­tis­ti­cal-math­e­mat­i­cal algo­rithms use a much broad­er data set. In addi­tion to gas con­cen­tra­tions, infor­ma­tion about the on-load tap-chang­er such as the num­ber of switch­ing oper­a­tions, the oil quan­ti­ty, or the time since the last oil change are also used as input vari­ables. The result is a diag­no­sis with an indi­ca­tion of prob­a­bil­i­ties, which is far ahead of pure gas analy­sis.

“Ana­lyz­ing on-load tap-chang­ers requires a lot of expert knowl­edge about how the par­tic­u­lar tap-chang­er type works and how it oper­ates.”

Tobias Gru­ber, Maschi­nen­fab­rik Rein­hausen

For a more com­pre­hen­sive assess­ment of on-load tap-chang­ers, MSENSE® VAM® is the online diag­nos­tic tool that per­forms a vibro-acoustic mea­sure­ment by detect­ing vibra­tions that occur dur­ing the switch­ing process of an on-load tap-chang­er. The asso­ci­at­ed algo­rithm rep­re­sents the switch­ing oper­a­tion in a lim­it val­ue curve, which is cal­cu­lat­ed more pre­cise­ly with each switch­ing oper­a­tion. Thus, the sys­tem iter­a­tive­ly learns dur­ing the switch­ing oper­a­tions what the acoustic sig­na­ture of a cor­rect­ly oper­at­ing on-load tap-chang­er looks like. “And it works every­where,” Gru­ber explains, “for new as well as for exist­ing on-load tap-chang­ers of all man­u­fac­tur­ers and designs.”

One thing that no one wants to find in their trans­former is mois­ture. In addi­tion to many pos­si­ble test points, it is above all the break­down volt­age of the insu­lat­ing oil that gives con­crete indi­ca­tions of pos­si­ble con­t­a­m­i­na­tion. How­ev­er, the meth­ods pro­posed by cur­rent stan­dards for ana­lyz­ing this volt­age are cum­ber­some and expen­sive. Here, too, online mon­i­tor­ing by self-learn­ing algo­rithms is bet­ter: They learn the rela­tion­ship between rel­a­tive oil mois­ture, oil tem­per­a­ture and break­down volt­age. In this process, data for break­down volt­age lev­els at dif­fer­ent oil tem­per­a­tures and humidi­ties are deter­mined exper­i­men­tal­ly using a ref­er­ence method, e.g. IEC 60156, and these data are then used to train a math­e­mat­i­cal mod­el. The mod­el is val­i­dat­ed and opti­mized using test data that is inde­pen­dent of the train­ing data and again, a trained arti­fi­cial expert is on hand to reli­ably iden­ti­fy where prob­lems might be brew­ing.

2. INTELLIGENT DATA EVALUATION

Mon­i­tor­ing sys­tems on pow­er trans­form­ers already play a major role, espe­cial­ly in terms of sta­ble fault detec­tion. How­ev­er, this alone is not enough. They con­tin­u­ous­ly report infor­ma­tion and data which must be eval­u­at­ed and inter­pret­ed in rela­tion to each oth­er – and this for every trans­former. This is no easy task, espe­cial­ly when sev­er­al sen­sors are capa­ble of the same func­tion. “For exam­ple, if tem­per­a­ture, par­tial dis­charge and DGA sen­sors are all capa­ble of detect­ing wind­ing faults, their state­ments are gen­er­al­ly not com­pa­ra­ble which can result in con­tra­dic­to­ry state­ments, and a sim­ple diag­no­sis is not pos­si­ble.”

That’s why at MR all sen­sors are inte­grat­ed into the ETOS® and eval­u­at­ed by the self-learn­ing algo­rithms of the ISM® com­put­ing unit with­in the ETOS®. “A Bayesian net­work checks which fault pat­terns best match the warn­ings that occur – and don’t occur. In addi­tion, a pri­ori prob­a­bil­i­ties for typ­i­cal trans­former faults and sen­sor reli­a­bil­i­ty are con­sid­ered.” The result is a prob­a­bil­i­ty esti­mate for all known trans­former prob­lems, with the most like­ly prob­lems dis­played to the cus­tomer along with a list of rea­sons for the find­ings. This allows the pro­fes­sion­al to inter­pret the diag­nos­tic results in a sim­pli­fied and effi­cient man­ner. “The Bayesian net­work Asset Intel­li­gence for pow­er trans­form­ers pro­vides a guide to min­i­mize risk and take cor­rec­tive action quick­ly,” says Tobias Gru­ber.

3. INTELLIGENT FLEET MANAGEMENT

Most trans­former fleets cur­rent­ly con­sist of mul­ti­ple prod­ucts that come from var­i­ous sup­pli­ers and dif­fer in age and equip­ment. Asset man­age­ment must do jus­tice to them all. The best way to do this is with a cen­tral solu­tion that uses and eval­u­ates all online and offline data – from sen­sor data to man­u­al­ly entered mea­sured val­ues fol­low­ing an on-site inspec­tion. Rein­hausen has devel­oped TESSA® APM® for this pur­pose. The soft­ware behind the plat­form eval­u­ates data, rec­og­nizes trends and inde­pen­dent­ly issues rec­om­men­da­tions for action. In the process, the intel­li­gent soft­ware con­tin­u­ous­ly learns dur­ing oper­a­tion and “gets to know” its con­nect­ed assets. Because at some point the rel­e­vant spe­cial­ist may be miss­ing and it will be TESSA® APM® that says, for exam­ple, “Take a look at trans­former 5A, it’s a bit sen­si­tive with the oil”.

The fact that the soft­ware com­bines all data, can be used remote­ly and warns pro­phy­lac­ti­cal­ly in the event of mal­func­tions, saves asset man­agers valu­able time and also fills the per­son­nel sup­ply gap. A tri­ad of intel­li­gent sen­sors at the process lev­el, data eval­u­a­tion at the field lev­el on the trans­former, and fleet man­age­ment at the con­trol lev­el thus pre­pare net­work oper­a­tors and OEMs for the chal­lenges of today and tomor­row.