Classification of Cellulosic Insulation State Based on Smart Life Prediction Approach (SLPA)

Abstract

The state of cellulosic solid kraft paper (CSKP) insulation, to a large extent, is an indication of a transformer’s health. It not only reflects the condition of transformer but also diagnose its residual life. The quantity of 2-furfuraldehyde (2-FAL), carbon dioxide (CO₂), and carbon monoxide (CO) dissolved in the transformer oil are useful diagnostic indicators to predict the state of the CSKP insulation. In this work, the current physical state of the CSKP is determined with the help of easily measurable parameters, like temperature, moisture, and the aging time. Here, the degree of deterioration of CSKP insulation has been determined using an integrated insulation health assessment system. This technique integrates a two-stage system comprising of a neural network (NN) model followed by a Smart Life Prediction Approach (SLPA). A thermo-moisture-aging multi-layer feed-forward NN model has been developed to predict the concentrations of 2-FAL, CO₂, and CO, which are further correlated to estimate the Degree of Polymerization (DP) values adopting an SLPA. The advantage of the proposed integrated system is that it provides an alternative means of paper health assessment based on Dissolved Gas Analysis (DGA) without estimating dissolved gas concentrations in oil, thereby avoiding the use of sophisticated measuring instruments. The optimal configuration of the NN model has been achieved at minimum iterations with an average cross-validation mean square error of 3.78 × 10⁻⁷. The proposed system thereby avoids destructive and offline measurement of DP and facilitates real-time condition monitoring of oil-immersed transformers. The test results of the developed system show considerable reliability in determining insulation health using easily measurable parameters. Furthermore, the system’s performance is compared with reported work and has been found to provide encouraging outcomes.

1. Introduction

Due to our modern lives being heavily technology dependent, it has become exceedingly important to maintain reliable electrical power supply. Transformers, being a critical component of the power system, could severely impact the power system performance if they are left unattended or inadequately maintained. Their breakdown has far-reaching financial consequences [1,2]. This mandates to identify the transformer faults at the incipient stage, which will help to ensure its smooth operation.

Generally, the transformer’s insulation system is composed of mineral oil (MO) and cellulosic solid kraft paper (CSKP). It is possible to diagnose incipient transformer faults by estimating dissolved fault gases in the MO. The overloading condition in the transformer produces thermal stresses, which in turn accelerates the failure of the CSKP insulation [3]. Earlier, the oil-impregnated CSKP failure was attributed only to electrical and thermal stresses [4]. However, CIGRE identified moisture as a most deleterious agent in the rapid ageing of the CSKP insulation [5,6]. Consequently, moisture, along with thermal and electrical stresses, produces a rapid degradation of the CSKP insulation. The moisture measurement is generally carried by karl fisher titration technique (IEC 60814) and loss of mass measurements using accurate microbalances. In addition to the conventional methods of the moisture measurement, different works have been reported, which aim specifically to simplify the measurement of moisture content in transformer oil [7,8,9].

Hence, the presence of moisture shortens the insulation useful service life, which ultimately influences the transformer’s life-span. The degree of polymerization (DP), carbon dioxide (CO₂), carbon monoxide (CO), and 2-furfuraldehyde (2-FAL) are considered as the ageing indicator of the CSKP insulation. DP, measured through a destructive test method, gives a direct assessment of CSKP with the number of linear cellulosic polymer in the chain structure. The number directly varies with the physical strength of cellulose. DP of fresh Kraft paper has the value of 1200, indicating a healthy paper while as a result of aging during its service life, this value gets reduced to 250 and less, reflecting ill paper health. The other parameters referred to as diagnostic parameters are the by-products of CSKP decomposition, dissolved into the transformer MO. The moisture, thermal, and electrical stresses originate from these parameters and are considered the reliable functions of paper ageing. The concentration of these parameters, when attaining a prescribed limiting value, indicates that the CSKP insulation is considered to have failed. Since these by-products are directly associated with DP [10,11,12,13,14,15,16], a strong correlation can be established between diagnostic and destructive parameters. This helps to avoid the destructive procedure of CSKP health assessment.

Recently, experiments have been carried out for assessing the health of the CSKP insulation. Several methods are proposed and reported that follow a convenient non-destructive procedure to determine insulation deterioration without interrupting the transformer operation. Several non-destructive diagnostic testing methods have been used previously for High-Voltage (HV) insulation testing in power system equipment such as cables. The Return Voltage Measurement (RVM) and Voltage Response (VR) methods used for insulation diagnosis are some of the commonly used non-destructive testing methods [17,18]. In [11], Ghoneim stressed the role of furfurals along with other parameters in the classification of DP of the cellulose paper. Several classification techniques are applied to obtain a classification model to predict the state of insulation. The decision tree classifier is found to be the best with 96.2% classification accuracy in estimating the insulation’s state. Lundgaard et al. in [15] proposed an activation energy model based on the kinetic process to study the cellulosic decay under the effect of temperature and moisture. They found that the accelerated aging of cellulose causes more water content in the transformer oil, thereby increasing its acidity. Consequently, 2-FAL can not be treated as a significant aging indicator of the insulating paper. In [19], a multi-stress mathematical model to trace the remnant life of CSKP based on enhanced Arrhenius relation was developed. Further, the performance of the model was compared with the results obtained by the thermal aging experiments performed on oil impregnated paper samples. The values of DP, 2-FAL, CO₂, and CO of these samples in real time are measured at different temperature and moisture levels. The outcome of this work reveals the influence of moisture on the rapid decomposition of the paper insulation. The intelligent methods need to undergo sufficient optimization to give adequate reliability in its field of application. The challenge lies in determining a robust optimal criterion making the adopted intelligent method more suitable for the application [20]. The Artificial Neural Network (ANN) model used in a multivariable system can be generalized for a larger range of dependent parameters provided that the ANN network is well trained and optimized. The intelligent techniques reported in [21,22,23,24,25] give some useful contribution towards the assessment of residual life of cellulose paper. However, most of these works rely on the DP for assessing the insulation state. The DP measurement technique is intrusive, as it requires removal of transformers from the service. For that reason, there is more emphasis needed on the diagnostic parameters for effective condition monitoring of the insulation. The diagnostic parameters are facilitated with the transformer online estimation. Hence, the objective to involve the diagnostic parameters and correlate them with the destructive parameters in identifying insulation condition forms the motivation of the work presented in the paper.

In this work, the condition of the solid paper insulation is estimated based on easily measurable parameters i.e., temperature, moisture, and the aging time. An integrated two-stage intelligent insulation health assessment system is developed, which is composed of an Artificial Neural Network (ANN) model followed by a heuristic smart life prediction approach (SLPA). The integrated system employs the benefits of easily measurable parameters for correctly estimating the state of insulation and thereby avoiding destructive and offline measurement of DP to facilitate the real-time condition monitoring of transformers. Thus, the present system is helpful for the utilities to avoid the use of sophisticated and complex instruments along with skilled technicians. In the first stage, an optimal NN model is developed to determine the concentrations of degrading by-products (2-FAL, CO₂, and CO) of the insulating paper as the function of temperature, moisture, and the aging duration. Therefore, the proposed system offers an alternative to avoid DGA-based measurement all the time. In the second stage of the intelligent system, the outputs obtained in the first stage are correlated to estimate DP through SLPA. In SLPA, the concentrations of the degrading by-products are classified into four intervals as per IEEE standards to obtain four specific ranges of DP, which ultimately shows the physical status of the paper insulation.

2. Cellulosic Insulation Failure Indicators

It is essential to perform condition monitoring of the transformer at regular intervals of time to ensure its healthy operation. All four of the failure indices can be estimated during the condition assessment of the insulation to interpret their correct health status.

2.1. Degree of Polymerization (DP)

This is also referred to as a destructive failure parameter as this requires the transformer to be taken into the offline mode before the paper samples can be drawn and their DP estimated. The Kraft paper used in solid insulation is made of cellulose. This cellulose is a polymer of glucose where the units of glucose are linked to resemble chain-like structure. Glucose is chemically represented as [C₆H₁₀O₅]_n, where ‘n’ is the total of glucose units linked to every chain. However, ‘n’ also signifies the DP; an index to measure the paper degradation. A DP value of 250 or less is suggestive of completely degraded paper insulation.

It is a well-known fact that any failure indicator measured through the destructive test procedure is always considered better than those measured through diagnostic testing. There are also other parameters that determine the CSKP condition and are referred to as diagnostic parameters.

2.2. 2-Furfuraldehyde or Furfuran (2-FAL)

2-FAL is another important failure index found in the dissolved stable state in the transformer oil and is a chemical indicator of the transformer’s winding insulation condition. It is a more dominant deteriorating agent of solid insulation among other furanic compounds. As the furfural is the significant by-product of paper degradation dissolved in transformer oil, its concentration is directly used to indicate the paper aging. 2-FAL can be correlated with DP following several models described by the Equations (1)–(5).

$$DP=\frac{\mathit{log}(2FAL-1.51)}{-0.0035} ; \mathrm{Chendong} \mathrm{Model}$$

(1)

$$DP=\frac{\mathit{log}\left(2FAL*0.88\right)-4.51}{-0.0035}; \mathrm{Stebbin}’\mathrm{s} \mathrm{Model}$$

(2)

$$DP=\frac{7100}{8.8+2FAL}; \mathrm{De} \mathrm{Pablo} \mathrm{Model}$$

(3)

$$DP=\frac{800}{\left[0.186*2FAL\right]+1} ; \mathrm{Modified} \mathrm{De} \mathrm{Pablo}’\mathrm{s} \mathrm{Model}$$

(4)

$$DP=-122.6 ln\left(2FAL\right)+1294.4; \mathrm{Ghoneim}’\mathrm{s} \mathrm{Model}$$

(5)

However, these models have their own limitation, as they require certain assumptions in estimating DP [26].

2.3. Carbon Oxides (CO₂ and CO)

The amount of CO₂ and CO present in the oil is also a useful indicator of the extent of paper ageing. CO and CO₂ are also mainly released in the oil as a by-product of thermal degradation and oxidation of paper cellulose. If their concentration reaches beyond prescribed limits, the solid paper’s life seems to come to end. The CO₂/CO ratio also provides an indication of the condition of the paper. If the ratio attains a value of less than 3, the insulation is in good health while a value greater than 3 indicates a threat of insulation failure [27].

3. Development of an Integrated System for Transformer Health Assessment

Since the state of transformer is identified by the condition of cellulosic insulation, there is a need to have an essential intelligent system to assess insulation health. Therefore, a two-stage integrated insulation health assessment system has been developed.

The first stage of the developed system uses the application of ANN, followed by the SLPA in the second stage. The first stage provides the concentrations of the degrading by-products of insulation using ANN, whereas the second stage predicts the physical status of paper by processing the output of the first stage through the SLPA. The schematic of the two-stage integrated insulation health assessment system is shown in Figure 1.

3.1. Design of the First Stage of Integrated Insulation Health Assessment System

The ANN primarily functions to process information aided by an effective non-linear mapping of the input space and the output space. It finds its importance in its ability to learn from an elaborate arrangement of well-interconnected neurons. The artificial neurons form a layered structure consisting of well-defined input and output layers, which may be separated from one another by one or more hidden layers [28]. The ANN structure is appropriately designed to achieve maximum accuracy while maintaining minimum complexity. However, the inputs selected should completely define the problem characteristics, and correspondingly, the outcome is reflected at the output.

In this work, a thermo-moisture-aging ANN model for transformer paper insulation health assessment has been implemented in the first stage using MATLAB software. The outputs of the proposed model are 2-FAL, CO₂, and CO. The insulating paper is made of cellulose, a polymer of glucose units. The glucose slowly degrades with time during the transformer’s operation. The temperature and moisture accelerated this degradation as the polymer-chains scission takes place rapidly. This causes a decrease in the molecular weight of the insulating paper. In addition, the progressive aging of the paper cellulose in the presence of temperature and moisture produces some of its degrading by-products, mainly furans and carbon oxide gases. These by-products remain dissolved in the transformer oil and are identified as the significant failure indicators of the insulating paper.

The ANN models, with their multilayer feedforward backpropagation structure, are capable of performing with ease of handling. These structures are universal approximators and have accomplished the ability to approximate any kind of nonlinear and continuous functions with high degree of accuracy. The multilayer feedforward architecture is also aided by the benefits of faster training using the Levenberg Marquardt training algorithm through an adapting learning rate [29,30]. The developed ANN model consists of three layers, namely, input layer, hidden layer and output layer. The proposed NN-based models have been implemented using the Levenberg Marquardt backpropagation algorithm. For training and testing purpose, a sigmoid transfer function is used for the hidden layer whereas for the output layer, linear is used. The sigmoid activation function that is highly nonlinear has been used as it demonstrates the best performance for a backpropagation NN [31]. The range of the input parameters was set to be 90 °C–130 °C for temperature, 1–3% for moisture, and an aging duration from 0–50,000 h. The values of the failure indicators thus obtained indicate the status of the CSKP.

An extensive training dataset comprising 500 data points is used to train the proposed network. Additionally, a ten-fold cross-validation approach is used to train and validate the model.

The proposed insulation assessment model has been configured by adjusting the number of neurons in the hidden layer in order to improve the prediction accuracy of the remaining insulation life. The number of hidden layer neurons is decided upon using a trial approach and is determined to be 25, corresponding to a minimum mean square error (MSE) of 1.51 × 10⁻⁹. MSE indicates the least deviation of the input and the target value. Figure 2 shows the MSE for different neuron levels. The implementation procedure of the NN model is shown in the flow diagram of stage I of the Figure 3. Here, the process is started with the grouping of input–output data pair for training and validation of the ANN using the suitable network architecture and training algorithm. The optimum configuration of the neural network has been achieved by identifying the mean square error at a selected number of neurons. The developed network is validated using a ten-fold cross-validation technique with the optimized parameters thereafter. The model is finally tested using the prepared testing dataset to obtain outputs of the stage I (i.e., 2-FAL, CO₂, and CO). These outputs are treated as the inputs of the stage II of SLPA. The optimal parameters used to train the NN model of the integrated system has been summarized in

Table 1. Specification of the proposed ANN model.

Specification of Proposed NN
Network architecture	Multilayer feed forward
Training algorithm	Levenberg-Marquardt
ANN structure	Single hidden layer in between input and output layers
Transfer function for hidden layer	Sigmoid
Transfer function for output layer	Linear
Input vector	[Temperature Moisture Aging-Duration]
Output vector	[2-FAL CO2 CO]
Training data	500
Testing data	42
No. of Neurons	25
MSE	1.51 × 10−9

.

The 2-FAL value of 6 ppm or more is set as the criteria for the end of life (EOL) of paper insulation. Simultaneously, the other failure indicators (i.e., CO₂ and CO), when reaching the concentration of about 3200 ppm and 375 ppm, respectively, suggest a failed insulation. Additionally, to determine the efficiency of the first stage of the developed insulation health assessment system, it is further tested using a testing dataset comprising 15 known data points.

3.2. Design of the Second Stage of Integrated Insulation Health Assessment System: Smart Life Prediction Approach (SLPA)

The IEEE C57.104 (2008) standard highlights the use of carbon oxides and other hydrocarbon fault gases in diagnosing and monitoring the insulation condition of the oil-immersed power transformers. However, several furanic chemical compounds are also found to be present in the oil because of CSKP decomposition. In particular, 2-FAL is the most consistent constituent among the various Furanic compounds and can serve as a major indicator of insulation health. Furthermore, the DP value of the paper cellulose can directly assess the transformer insulation deterioration. However, it needs to be emphasized that DP is determined through an intrusive, complicated, destructive procedure, which cannot be carried out in real-time scenarios. It, therefore, becomes pertinent to replace the destructive parameter with other diagnostic parameters that are varying with insulation aging. Consequently, 2-FAL and carbon oxides, CO and CO₂, released due to thermal degradation, are regarded as remarkable indicators of the insulation aging.

To develop a credible transformer insulation life assessment system, an SLPA has been proposed in stage II of the integrated insulation health assessment system. The smartness of the SLPA is evident from the fact that the method simultaneously uses all the diagnostic parameters to predict DP values based on which conclusions can be drawn on the current state of the paper insulation. The intent behind using the SLPA is to replace the destructive measurement with a non-intrusive diagnostic procedure as well as avoiding the detailed DGA procedure.

The algorithm used to design the SLPA is shown by the flow diagram of stage II of the Figure 3. The algorithm of stage II is mainly composed of different decision-making steps, each defined with a certain logic. These decision-making steps are intelligent enough and able to estimate the specific ranges of the final output of the SLPA (i.e., DP) based on which the current physical states of the insulating paper are identified. Here, the process starts by specifying the outputs obtained at stage I as the inputs of stage II. These inputs (i.e., 2-FAL, CO₂ and CO) have certain values, which are classified in a certain range based on which the ranges of DP are specified. Each range of DP is helpful to diagnose the current state of the insulating paper. Stage II of the flow diagram given in Figure 3 depicts the underlying logic of SLPA in the integrated transformer’s insulation health assessment system.

The decision-making process of the method is developed using logic obtained with the help of IEEE standard and observed experimental data. The logic used in the SLPA is a four-step procedure subject to different ranges of the input parameters. The DP is also classified into four categories corresponding to each range of input parameters. It illustrates that if DP ranges 1350–700, it signifies a healthy condition of the paper. If the DP lies in the range 700–450, it indicates an early deterioration stage whereas an extensive deterioration is indicated for a DP value in the range 450–250. The paper insulation is rendered useless if the DP value falls below 250 [32]. In stage II, the outputs of stage I are identified for the ranges they belong to, and accordingly, the range of DP is specified. Hence, the output of each decision step follows a particular range, which signifies the exact physical state of the insulating paper.

4. Result and Discussion

4.1. Performance Evaluation of the Proposed ANN Model

As stated earlier, the proposed ANN model, the first stage of the intelligent insulation health assessment system, is used to estimate the values of different failure indicators, which can provide conclusive information about the current state of the transformer insulation. This model is validated using a ten-fold cross-validation approach. The process involves the division of the dataset into ten groups of equal size. Nine out of these ten groups are utilized for training the network while the remaining one group is used to validate the trained ANN model.

Ten successive training and validation processes are followed and in each process, a different sample is used for validation, thu ensuring the model to be tested for all the samples under consideration. Figure 4 graphically shows the Mean Square Errors (MSE) obtained in each validation fold. The MSE of the validation process ranges from 3.60 × 10⁻⁷ to 4.12 × 10⁻⁷ with an average MSE of 3.78 × 10⁻⁷.

Further, an additional dataset of 15 sample points is used to test the model.

Table 2. Failure indices obtained from the proposed ANN model.

Sample No.	Temperature (°C)	Moisture (%)	Ageing Duration (h)	[2-FAL CO2 CO] (Actual) [19]	[2-FAL CO2 CO] (Obtained by Proposed Model)
1	90	1.5	740	[0 809 64]	[0 812 62]
2	90	2	18,000	[0.97 2631 368]	[0.97 2628 370]
3	95	1	10,820	[0.094 1493 135]	[0.095 1498 132]
4	95	2	3600	[0.102 1872 166]	[0.103 1878 164]
5	100	1	5059	[0.105 2301 209]	[0.107 2298 214]
6	105	2.5	5600	[5.26 4404 595]	[5.28 4400 594]
7	110	1.5	2100	[0.10 2490 148]	[0.10 2562 146]
8	110	2.5	2517	[0.87 2584 391]	[0.9 2587 387]
9	115	3	570	[0.5 2509 359]	[0.48 2502 353]
10	120	2	4250	[1.32 4342 577]	[1.35 4348 576]
11	125	1.5	3900	[5.25 4200 580]	[5.19 4206 586]
12	125	2	570	[0.9 2980 502]	[0.97 2984 503]
13	127	1	730	[0.09 1817 209]	[0.093 1815 211]
14	129	2.5	1150	[4.35 4210 580]	[4.36 4217 566]
15	130	1	1191	[0.25 2514 376]	[0.26 2421 372]

gives the outcome of the proposed model for all the sample points under test.

The output is a three-element vector corresponding to each failure indices i.e., 2-FAL, CO₂, and CO subjected to thermo-moisture -aging. The values in the output vector array clearly signify the moisture aided insulation deterioration condition. In all the cases, moisture has an effective impact on the failure parameters of the paper insulation. For example, the sample no. 1, with low moisture content, temperature, and aging time is showing healthy condition of insulation with no trace of 2-FAL and lower concentration of CO₂ and CO as well. Furthermore, it is evident when comparing the sample cases 5 and 6 that the values of the failure indices significantly change due to large variations in moisture content while other input parameters remain nearly constant.

It can hence be concluded that an increase in moisture in paper insulation accelerates deterioration. The impact of moisture on insulation deterioration can also be visualized from sample cases 12 and 13. It can be inferred that an increase in moisture by 100% would accelerate the aging parameters by 90.41%, 39.17%, and 58.05%, respectively, despite the temperature decreasing by 1.57%. Additionally, a comparison based on the failure indicators is made individually with the results reported in [19]. The value of EOL for each failure indices, as stated earlier, forms the basis of this comparison.

A comparative study of the obtained results is done with the same inputs as reported in [19]. The comparison shows a consistent estimation of the failure parameters, which further validates the proposed ANN-based CSKP health assessment model.

Table 3. Comparison of estimated life using neural network by treating 2-FAL as a failure index.

Temp (°C)	Moist (%)	Aging Duration (h)	2-FAL (ppm) in [19]	2-FAL (ppm) Using Proposed NN Model	Deviation	Error (%)
90	1	45,200	6	6.13	−0.13	−2.16
90	2	39,811	6	6	0	0
90	3	18,000	6	6.05	−0.05	−0.83
110	1	16,800	6	5.72	0.28	4.66
110	2	5754	6	6	0	0
110	3	3020	6	5.89	0.11	1.83
130	1	3981	6	6	0	0
130	2	1150	6	6.23	−0.23	−3.83
130	3	682	6	6.59	−0.59	−9.83

,

Table 4. Comparison of estimated life using neural network by treating CO₂ as a failure index.

Temp (°C)	Moist (%)	Aging Duration (h)	CO2 (ppm) in [19]	CO2 (ppm) Using Proposed NN Model	Deviation	Error (%)
90	1	34,161	3200	3204	−4	−0.125
90	2	31,876	3200	3203	−3	−0.093
90	3	15,679	3200	3011	189	5.90
110	1	14,550	3200	3189	11	0.34
110	2	4394	3200	3201	−1	−0.031
110	3	2655	3200	3200	0	0
130	1	4280	3200	3214	−14	−0.437
130	2	1242	3200	3176	24	0.75
130	3	521	3200	3089	111	3.46

and

Table 5. Comparison of estimated life using neural network by treating CO as a failure index.

Temp (°C)	Moist (%)	Aging Duration (h)	CO (ppm) in [19]	CO (ppm) Using Proposed NN Model	Deviation	Error (%)
90	1	49,164	375	377	−2	−0.53
90	2	18,989	375	365	10	2.67
90	3	13,313	375	367	8	2.13
110	1	13,892	375	382	−7	−1.86
110	2	5180	375	366	9	2.4
110	3	3064	375	374	1	0.27
130	1	3963	375	395	−20	−5.33
130	2	1169	375	364	11	2.93
130	3	563	375	348	27	7.2

present a comparison of each of the estimated failure indicators.

In all cases, the EOL criterion is a function of aging duration subject to temperature and moisture content. It is observed that from the nine test cases of Table 3 that a maximum absolute deviation in estimating 2-FAL is 9.83%. The absolute average error, i.e., the error of the 2-FAL values between the proposed model and the reported one is 1.12%. Similarly, the maximum absolute deviation in determining the other non-destructive failure parameters CO₂ and CO is determined to be 5.90% and 5.33% from Table 4 and Table 5, respectively. Hence, the average maximum absolute deviation of the failure indices is determined to be 7.02%. Furthermore, the absolute average error for them is estimated to be 1.12%, 1.08%, and 1.09%, respectively. It is evident that all the three parameters have the values of average error in close proximity, which iterates that they are simultaneously able to identify the insulation status. The life estimation of CSKP insulation based on different failure indices differs subjected to the aging time.

4.2. Performance of SLPA for Estimating Transformer Insulation Health

Even though DP measurement is a more reliable diagnosis, their direct measurement is an intrusive and costly process requiring actual Kraft paper samples. This requires interruption in the operation of power transformer. It is shown in the literature [19] that DP is closely correlated with other diagnostic parameters (2-FAL, CO₂, and CO) to indicate the health of paper insulation. Therefore, in proposed SLPA, DP is indirectly estimated on the basis of estimated values of 2-FAL, CO₂ and CO. Hence, an interpretation of the insulation state can be easily drawn without interrupting the transformer operation.

The SLPA take the outputs obtained in the first stage as their inputs. These inputs are checked for the specified categories to which they belong to. Accordingly, the range of DP is identified, which ultimately gives the state of the insulating paper.

Table 6. Validation of insulation status using SLPA.

S. No.	2-FAL	CO2	CO	DP		State of CSKP Insulation
				[Range as per SLPA]	Reported in [19]
1	0	812	62	1350–700	1304	Healthy
2	0.970	2628	370	700–450	490	Moderate
3	0.095	1498	132	1350–700	963	Healthy
4	0.103	1878	164	1350–700	1189	Healthy
5	0.107	2298	214	1350–700	1149	Healthy
6	5.280	4400	594	450–250	266	Extensive
7	0.10	2562	146	1350–700	846	Healthy
8	0.90	2587	387	700–450	500	Moderate
9	0.480	2502	353	700–450	654	Moderate
10	1.350	4348	576	450–250	362	Extensive
11	5.190	4206	586	450–250	268	Extensive
12	1.060	2984	503	700–450	600	Moderate
13	0.093	1815	211	1350–700	785	Healthy
14	4.360	4217	566	450–250	276	Extensive
15	0.260	2421	372	700–450	652	Moderate

shows the interpretation of insulation state based on the values of DP obtained by correlating it to the non-destructive parameters as outlined in the SLPA. The values of the parameters used to validate the method are the outputs of the first stage of the integrated insulation health assessment system.

It can be observed from the table that the transformer insulation states are correctly predicted following the SLPA for most cases as the value of DP lies in their proposed ranges. The insulation of transformer sample case no. 1 has been identified to be in healthy state by its DP value reported in [19], which is also validated using the SLPA based on non-destructive parameters. The insulation status of test cases No. 7, 14, and 15 have been determined to be in moderate, extensive, and moderate deterioration, respectively.

However, the value of one of the non-destructive parameters for these cases falls very minutely out of range of the criteria presented in the SLPA. Though the values for CO₂ in case No. 7 and 15 are found to be 2497 ppm and 2502 ppm, respectively, and the value of CO in case No. 14 is 572 ppm in [19], these limit violations are very close to the limiting values within a tolerance of ±3.23%.

5. Conclusions

Estimating the health status of cellulosic insulation plays an important role in improving the situational awareness at transformer maintenance level. This work attempts to present a two-stage integrated insulation health assessment system. The proposed system offers a DGA-based health monitoring of paper insulation with easily measurable parameters, i.e., temperature, moisture, and aging time. Therefore, the complexity in the measurement of dissolved gases arising due to sophisticated instruments can be avoided. The first stage uses the application of the ANN technique to estimate the degrading by-products (2-FAL, CO₂, and CO) of the paper insulation subjected to temperature, moisture, and aging time. The training and testing of the ANN model led to the convergence of optimal configuration in terms of hidden layer neurons obtaining the lowest MSE. The model has been tested and found to have an average cross-validation MSE of 3.78 × 10⁻⁰⁷. Each failure indicators of the model have also been compared individually to what is reported in the literature subjected to similar inputs. The absolute average errors for 2-FAL, CO₂, and CO have been calculated to be 1.12%, 1.08%, and 1.09%, respectively. In the second stage, an SLPA is developed to correlate the outputs obtained in the first stage with DP. The SLPA correctly predicts the state of insulation as the DP lies in the proposed ranges. This correlation helps to attain a considerably good accuracy in making decisions over the insulation failure status.