Increasing the Accuracy and Optimizing the Structure of the Scale Thickness Detection System by Extracting the Optimal Characteristics Using Wavelet Transform

Abstract

Loss of energy, decrement of efficiency, and decrement of the effective diameter of the oil pipe are among the consequences of scale inside oil condensate transfer pipes. To prevent these incidents and their consequences and take timely action, it is important to detect the amount of scale. One of the accurate diagnosis methods is the use of non-invasive systems based on gamma-ray attenuation. The detection method proposed in this research consists of a detector that receives the radiation sent by the gamma source with dual energy (radioisotopes ²⁴¹Am and ¹³³Ba) after passing through the test pipe with inner scale (in different thicknesses). This structure was simulated by Monte Carlo N Particle code. The simulation performed in the test pipe included a three-phase flow consisting of water, gas, and oil in a stratified flow regime in different volume percentages. The signals received by the detector were processed by wavelet transform, which provided sufficient inputs to design the radial basis function (RBF) neural network. The scale thickness value deposited in the pipe can be predicted with an MSE of 0.02. The use of a detector optimizes the structure, and its high accuracy guarantees the usefulness of its use in practical situations.

1. Introduction

In the oil industry, scale formation inside the transmission pipes causes major problems. The formation of scales makes the internal cross-section of the pipe smaller and reduces the fluid flow, and failure to identify this problem causes malfunctions in pumps and related equipment, emergency shutdowns, damage to oil equipment, increased repair costs, and reduced efficiency. Researchers who considered it necessary to have accurate detection systems to detect the amount of scale inside the pipe always used gamma ray attenuation systems as the gold standard in determining the various parameters of multiphase flows [1,2,3,4,5,6,7,8]. In the study [1], an attempt was made to predict the volume percentage and classification of flow regimes. Using a cesium source, two sodium iodide detectors, and the test pipe, the researchers implemented a two-phase flow in three regimes, stratified, annular, and bubbling, and using the information obtained from both detectors as input for the radial basis function (RBF) network were able to reach their goal. Roshni et al. in [2], at the cost of increasing the computational load to the system, were able to increase the accuracy by using three GMDH networks. Finally, they were able to achieve good accuracy in determining the volume percentage and detecting the type of flow regimes in three-phase flows. In 2016, in similar research, a ⁶⁰Co source and a NaI detector were used for obtaining the kind of flow patterns and volume percentage, which was met with low accuracy and unfavorable results due to the lack of extraction of appropriate characteristics from the received signals [3]. In 2019, authors used Jaya’s optimization algorithm for predicting the volume percentage of a three-phase flow in the stratified pattern [4]. Sattari et al. in [5] proposed a structure with a cesium source and two NaI detectors around the tested pipe to be able to perform volume percentages prediction and classification of flow regimes with high accuracy. The use of time characteristics and the correct selection of characteristics, while reducing the computational load, resulted in a high-precision system. Later research [6] used a GMDH neural network for detecting the kind of flow patterns and predicting the volume percentage. High accuracy was achieved in determining volume percentages, but not considering the amount of scale inside the pipe is one of the gaps in this research. Alamoudi et al. [7] tried for detecting the thickness of scale in the oil pipe. They simulated a two-phase flow in disparate regimes. They considered Gamma peak counts of Ba-133 and Cs-137 from the first transmission photon detector and the total number from the second scattered photon detector as inputs to the RBF neural network, and they were able to foretell the scale thickness with the RMSE of less than 0.22. In another study, the researchers simulated a three-phase flow in the annular pattern by considering the thickness of the scale inside the pipe to inquire into different volume percentages. Finally, the photopeaks ²⁴¹Am and ¹³³Ba recorded in two transmitted detectors were extracted and considered as the input of an RBF neural network. They predicted the amount of intra-pipe scaling with RMSE of less than 0.09 [8]. In recent years, many researchers have turned to using an X-ray source instead of gamma in their structures, the reason for which is to get rid of problems such as the need to use protective clothing by personnel when working with this device (due to the inability to turn it off) can be stated. They differentiated the parameters of multiphase flows using X-ray tubes [9,10,11,12]. For example, in [9], the aforementioned alternative source was used. The training of two existing multilayer perceptron (MLP) neural networks in their proposed structure was also done by the temporal features, which are extracted from the signals received by the detector. In the research [10], three-phase flows were perused, and three patterns were simulated in distinct volume percentages. In this research, three RBF neural networks were trained with the frequency characteristics of the received signals, and the result brought relatively good accuracy. In Ref. [11], an X-ray tube was used for designing a control system. Four petroleum products that are mixed two by two with distinct volumes were simulated with the Monte Carlo N Particle code (MCNP) code. The recorded signals were used as inputs of three MLP neural networks for predicting the volume ratio of the three products. Although the introduced method predicted the type and amount of products, the lack of feature extraction techniques prevented high accuracy. To develop the previous research (Ref. [11]), Balubaid et al. [12] used wavelet transform for feature extraction, which resulted in an effective reduction of computational load and increased accurateness. Numerous studies have been conducted to gauge scale thickness and examine various features [13,14] with more detectors or higher errors. In this study, inspired by the existing history of similar studies, an attempt was made to design a high-precision system to acquire the scale value inside the pipe. For this purpose, a three-phase flow pattern consisting of water, gas, and oil in distinct volume percentages was simulated. A different value of scale thickness was considered in each simulation. A dual-energy gamma source (²⁴¹Am and ¹³³Ba) and a sodium iodide detector were placed on both sides of a test pipe. From the signals received from the detector, the wavelet characteristics were extracted. These extracted characteristics were considered as the inputs of the RBF neural network to obtain the output, which was the scale thickness inside the pipe. The result of this research is the introduction of an effective and accurate method to detect the amount of scale inside the pipe. The contributions made by this study are as follows:

• Examining the wavelet transform’s properties and effectiveness in calculating scale thickness.
• Using a single detector lowers costs and the structure of the detecting system’s complexity.
• Increasing the precision of scale thickness determination by extracting useful features from received signals.
• Using the RBF neural network as a fast-learning network to calculate scale thickness.

2. Simulated Detection System

In recent years, researchers have obtained positive results from radiation-based system simulations with the MCNP code [15,16,17,18,19]. In addition, in this study, the MCNP code has been used. A dual-energy gamma source, a steel test pipe, and a sodium iodide detector are the main parts of the proposed structure. The gamma source consists of radioisotopes ²⁴¹Am and ¹³³Ba, with photon energies of 59 and 356 keV, respectively. In the simulation of the three-phase flow in the stratified regime, as well as the simulation of the sediment scale, a test pipe made of steel is used, which has an internal diameter of 10 cm and a thickness of 0.5 cm. The detector measuring 2.54 cm² is placed at a distance of 30 cm from the source with a zero angle to the assumed line of the source to receive the passing signals. The proposed structure of this study has been validated by the experimental structure implemented in Ref. [1]. This structure of a three-phase flow is simulated in a stratified regime consisting of gas, water, and oil in a volume percentage between 10% to 80%. The deposited scale inside the pipe is considered cylindrical and with a thickness of 0, 0.05, 1, 1.5, 2, 2.5, and 3 cm. The said scale is made of BaSO4 and has a density of 4.5 g per cubic centimeter. Water, oil, and gas considered in this simulation also had densities of 1, 0.826, and 0.00125 g per cubic centimeter. In Ref. [1], a two-phase flow was implemented in the annular regime, and the simulation of the same structure was performed by the MCNP code. The recorded counts obtained from the detectors of the simulated structure and the experimental structure were compared. It was observed that there is an acceptable match between the two. Seven scale thickness values × 36 different volumetric percentages = 252 simulations were performed. Figure 1 shows the proposed structure of this study and how this structure reaches the goal of determining the thickness of the scale. The signals received by the detector are shown in Figure 2. In this figure, the x-axis shows the thickness of the scale, the y-axis represents the source energy, and the z-axis represents the amount of intensity absorbed by the detector.

3. Discrete Wavelet Transform

One of the widely used numerical analyzes is discrete wavelet transform (DWT), in which the wavelet is sampled discretely. In discrete wavelet transform, in addition to frequency features, time features are also available, which is its most important advantage over the Fourier transform. To calculate the DWT of a signal, first, a low-pass filter with the impulse response g is applied to the signal, resulting in the convolution of the two as follows [20,21].

$$y\left[n\right]=\left(x\ast g\right)\left[n\right]={\displaystyle \sum }_{k=-\infty }^{\infty }x\left[k\right]g\left[n-k\right]$$

(1)

In addition, simultaneously, another high pass filter (h) decomposes the signal. Figure 3 clearly shows the signal decomposition process. The output of this process is detail coefficients (output of high-pass filter) and approximate coefficients (output of low-pass filter). At the output of the filters, as seen from the figure, there is a downsampler with 2 in each step. The downsampled results of the low-pass filter provide the approximation (A), and the output of the low-pass samplers of the high-pass filter provides the detail (D). Knowing that at each step, the approximate part can be decomposed over and over again. In this research, the analysis has continued up to four stages. The wavelet operation calculates the wavelet coefficients of a separate set of child wavelets for a given mother wavelet ψ(t). In the case of discrete wavelet transform, the mother wavelet is shifted and scaled by powers of two [20,21].

$${\psi }_{j,k}=\frac{1}{\sqrt{2^j}}\psi \left(\frac{t-k{2}^j}{2^j}\right)$$

(2)

where j denotes the scale parameter and k the shift parameter. Since x(t) is a 2N-length signal, the wavelet coefficients derived from it may be thought of as the projection of x(t) onto a wavelet. A member of the above-mentioned discrete wavelet family is a child wavelet if and only if [20,21]:

$${\gamma }_{jk}={{\displaystyle \int }}_{-\infty }^{\infty }x\left(t\right)\frac{1}{\sqrt{2^j}}\psi \left(\frac{t-k{2}^j}{2^j}\right)dt$$

(3)

Then, j is fixed, and the ${\gamma }_{jk}$ is obtained only in terms of a function of k. In the above equation,${\gamma }_{jk}$ can be obtained as the convolution of the x(t) with the mother wavelet signal, $h\left(t\right)=\frac{1}{\sqrt{2^j}}\psi \left(\frac{-t}{2^j}\right)$, sampled at points 1, 2j. 2j, …, 2N. These are details of discrete wavelet coefficients at level j. Detail coefficients of the filter bank, therefore, match perfectly to a wavelet coefficient of a discrete collection of child wavelets for a given mother wavelet ψ(t), assuming an acceptable choice of h[n] and g[n] (t). From the analyzed signals, the characteristics of average of a₄ and average of d₁ to d₄ were extracted and used in the next steps. Since they were originally proposed as useful characteristics in earlier studies [12,22,23], these features have been utilized to improve the accuracy and structure of the scale thickness detection system.

4. RBF Neural Network

Neurons, also known as nerve cells, are the smallest components of the human brain that send and receive signals from the brain. While there are many similarities between neurons and cells, neurons are completely unique in function and structure. All these neurons, which number up to millions, are somehow connected to each other. However, according to their location and responsibility, neurons have three main parts in terms of size, shape, and structure: Cell body, axon, and dendrite. Neurons are actually computing units that have branches called dendrites through which they receive information from other neurons. The neuron’s center functions as a processing unit, which then sends the processed data via the axon to neighboring neurons. It all happens biochemically, but researchers have offered ways to describe this architecture mathematically. Radial basis function networks are a popular kind of neural network used to solve this problem (RBF). An RBF network is a kind of convolutional neural network that employs radial basis functions as activation functions. Having been designed for a broad range of purposes, this network has quickly become a popular alternative to the more established multi-layer perceptron. To simplify, RBF employs a three-layer design. There is no processing performed in the input layer since it is a puller layer. Through a non-linear adaptation between the input space and another space with higher dimensions, the second layer (the hidden layer) is able to provide more complex results. The linear output and the final weight total are both generated in the third layer. Such an output is helpful if RBF is used to approximate the function, however, a hard limiter or sigmoid function may be applied to the output neurons to generate 1 or 0 if pattern categorization is necessary. RBF’s standout characteristic is the processing that occurs in the hidden layer. This technique’s fundamental concept was to generate clusters from preexisting patterns in the input space. The distance from the cluster center may be calculated if its centers are known. Since this is a non-linear distance measure, the result obtained will be close to 1 if a pattern is found in the vicinity of the cluster’s center. Outside of this region, the value drops dramatically. The non-linear function may be expressed in terms of the radial basis since this area is radially symmetric around the cluster’s center. Radial basis functions often have the following shape [24]:

$$\phi \left(r\right)=\exp \left[-\frac{r^2}{2{\sigma }^2}\right]$$

(4)

The parameter of r in an RBF is the numeral value of the interval from the center of the cluster. Equation (4) shows a typical bell curve. Generally, the measured distance to the cluster center is the Euclidean distance. In the hidden layer, for each neuron, the coordinates of the cluster center are represented by weights. Therefore, when a neuron receives an input pattern X, the interval is obtained using the following equation [24]:

$$r_j=\sqrt{{\displaystyle \sum }_{i=1}^n{\left(x_i-w_{ij}\right)}^2}$$

(5)

This leads to the following output from the j neuron in the hidden layer [24]:

$${\varnothing }_j=\exp [-\frac{{{\displaystyle \sum }}_{i=1}^n{\left(x_i-w_{ij}\right)}^2}{2{\sigma }^2}]$$

(6)

The width or radius of the bell curve is defined by the variable σ and in some cases, it is appointed only experimentally. The hidden layer of an RBF network has units that have weights, and these weights represent the cluster center vector. Weights can be obtained using traditional methods such as the K-Mean algorithm or methods based on the Kohonen algorithm. In any case, the training is done unsupervised, but the expected number of clusters (k) is already selected, then these algorithms obtain the best fit for these clusters. As a strong mathematical tool, numerical computations [25,26,27,28,29,30,31] have been used recently for a variety of engineering challenges, most notably, artificial networks [32,33,34,35,36,37,38,39,40,41,42,43,44]. The available data are split into two groups: training data and test data, in order to address the issue of over-fitting and under-fitting. The data that the neural network observes and uses to fit the data is included in the training data. Test data are used to evaluate the effectiveness of the neural network following the training phase. The planned network will be protected against the issues of over-fitting and under-fitting by the neural network’s suitable reaction to these two sets of data.

5. Result and Discussion

The features introduced in the previous sections were given as inputs to the RBF neural network so that the output, which is a 1 × 252 matrix determines the thickness of the scale inside the pipe. Several neural networks were put into practice, each with a distinct hidden layer of neurons. The structure of the optimal trained network can be seen in Figure 4. This network has five input neurons and one output neuron, which are the extracted characteristics and scale thickness value, respectively. Between the input and output layers, there is a hidden layer with 18 neurons, which is responsible for transferring the input space to the output space. In order to show the proper functioning of the neural network, three fit, regression, and error diagrams were drawn for three categories of all data, training, and testing data in Figure 5. The network output is represented by a blue line in the fitting diagram, whereas the intended output is represented by a black dashed line. In this diagram, the horizontal axis represents the data number, while the vertical axis represents the scale’s thickness. Two MSE and RMSE criteria were taken into account while calculating the error value of the implemented network. These criterion equations are as follows:

$$MSE=\frac{{{\displaystyle \sum }}_{j=1}^N{\left(X_j\left(Exp\right)-X_j\left(Pred\right)\right)}^2}{N}$$

(7)

$$RMSE={\left[\frac{{{\displaystyle \sum }}_{j=1}^N{\left(X_j\left(Exp\right)-X_j\left(Pred\right)\right)}^2}{N}\right]}^{0.5}$$

(8)

The symbols “X (Exp)” and “X (Pred)” reflect the experimental and predicted (ANN) values, respectively. N represents the number of data points. The second diagram shows the desired output as a black line and the neural network’s outputs as blue squares. Regression diagram is the name given to this illustration. The error figure shows the difference between the network output and the desired output for each data.

Table 1. Comparison of target values with neural network outputs.

	Train Targets	Train Outputs	Test Targets	Test Outputs
1	3.0000	2.8507	1.5000	1.5502
2	2.5000	2.6358	0	0.0084
3	2.0000	1.9591	1.0000	1.2299
4	2.5000	2.6328	2.5000	2.6055
5	0	0.0246	1.5000	1.4134
6	2.5000	2.6908	2.0000	2.1296
7	0.5000	0.5317	1.0000	0.8584
8	1.0000	0.9959	2.0000	2.2414
9	0.5000	0.5000	2.5000	2.6178
10	3.0000	2.9519	1.5000	1.7279
11	2.0000	2.2706	3.0000	2.8875
12	3.0000	2.9778	0.5000	0.5834
13	2.5000	2.3107	1.0000	0.8300
14	1.0000	1.2258	2.5000	2.5796
15	0.5000	0.5077	3.0000	2.8871
16	0	0.0522	0.5000	0.6381
17	1.0000	0.7109	2.0000	2.2191
18	2.5000	2.3532	0.5000	0.5247
19	0	0.0231	1.5000	1.2896
20	3.0000	3.0335	0.5000	0.6990
21	0	0.0752	1.0000	1.2803
22	2.0000	2.3439	3.0000	2.9794
23	2.5000	2.7286	0.5000	0.5555
24	1.5000	1.5813	2.5000	2.4208
25	3.0000	2.8800	0	0.0852
26	3.0000	2.9935	1.5000	1.5737
27	1.0000	0.8766	2.5000	2.4476
28	2.5000	2.4182	0.5000	0.5218
29	1.0000	1.1707	3.0000	2.8760
30	3.0000	3.0492	1.0000	1.2615
31	0.5000	0.5551	2.5000	2.5231
32	0	−0.1112	0.5000	0.5853
33	1.0000	0.9888	2.0000	1.8922
34	1.5000	1.6268	1.5000	1.4737
35	3.0000	3.0185	1.0000	0.8781
36	1.0000	1.2172	0	−0.1686
37	3.0000	2.9045	2.5000	2.8822
38	2.0000	2.2647	0	−0.0797
39	0	−0.0343	0	−0.2528
40	2.0000	2.0162	2.5000	2.8470
41	0	0.0399	2.5000	2.3918
42	2.0000	2.3730	0	0.0240
43	0.5000	0.6606	1.5000	1.3381
44	0	−0.1847	0.5000	0.6577
45	0	0.1900	2.5000	2.7493
46	0	0.0007	1.5000	1.7535
47	1.5000	1.5749	2.0000	1.8070
48	3.0000	2.8047	2.0000	2.3398
49	0.5000	0.4150	2.0000	2.0685
50	0	0.1194	2.5000	2.7588
51	2.5000	2.5892	1.5000	1.6536
52	2.5000	2.6717	0.5000	0.6614
53	3.0000	2.8893	1.0000	0.8097
54	0	−0.0654	1.5000	1.2968
55	1.5000	1.4104	0	−0.1245
56	0	0.0571	1.0000	0.9329
57	1.0000	0.7580	0	−0.0362
58	2.0000	2.3764	3.0000	2.8228
59	0	0.0156	1.0000	1.0669
60	3.0000	2.7795	2.0000	1.8055
61	3.0000	3.0747	2.5000	2.6317
62	3.0000	3.0070	0.5000	0.5188
63	1.0000	1.2388	1.5000	1.4472
64	3.0000	3.0872	0.5000	0.6452
65	2.5000	2.4517	1.5000	1.5560
66	2.0000	1.8413	2.0000	1.7887
67	3.0000	2.9910	1.0000	1.0048
68	0.5000	0.4394	1.0000	1.2338
69	0	−0.0480	0.5000	0.6543
70	1.5000	1.4888	1.0000	0.8751
71	2.5000	2.5288	0	−0.1118
72	2.0000	2.0434	1.0000	1.0108
73	1.5000	1.3414	2.0000	1.9147
74	3.0000	2.8135	0.5000	0.5529
75	0.5000	0.4246	2.0000	1.8891
76	0.5000	0.5672	2.0000	2.3440
77	0.5000	0.6651
78	2.5000	2.4111
79	2.0000	1.9201
80	0.5000	0.4180
81	1.5000	1.7401
82	0	0.1030
83	2.0000	2.3196
84	1.0000	0.8052
85	2.5000	2.7879
86	2.0000	1.8781
87	0.5000	0.5666
88	1.5000	1.5677
89	1.0000	1.1954
90	2.5000	2.3980
91	2.0000	2.3241
92	1.5000	1.2039
93	0.5000	0.6220
94	2.5000	2.5530
95	0	−0.0323
96	2.5000	2.4175
97	0.5000	0.3938
98	1.0000	1.0718
99	1.0000	0.7100
100	0.5000	0.6053
101	0	0.0484
102	0.5000	0.6160
103	2.5000	2.3416
104	2.0000	1.8580
105	1.5000	1.6659
106	3.0000	2.9056
107	2.5000	2.3207
108	2.0000	2.0069
109	2.5000	2.3241
110	0.5000	0.5279
111	1.5000	1.6442
112	1.0000	1.0299
113	0	0.0031
114	0	0.0226
115	0	−0.0417
116	3.0000	2.9503
117	0.5000	0.6860
118	1.5000	1.7487
119	2.0000	1.8221
120	1.0000	0.9477
121	1.0000	1.1897
122	2.0000	1.9254
123	1.5000	1.2626
124	2.0000	2.2340
125	1.5000	1.2649
126	3.0000	2.9737
127	2.0000	1.7831
128	0	−0.0127
129	0.5000	0.6844
130	2.5000	2.6940
131	2.5000	2.4228
132	3.0000	2.8849
133	3.0000	2.7392
134	2.5000	2.3055
135	2.0000	2.2171
136	3.0000	2.8540
137	1.0000	0.8989
138	0	−0.1658
139	1.5000	1.6764
140	1.5000	1.6445
141	3.0000	2.8768
142	0.5000	0.5343
143	1.0000	0.9241
144	1.5000	1.5487
145	1.5000	1.3810
146	3.0000	2.9438
147	3.0000	3.0738
148	1.0000	1.3264
149	0.5000	0.5668
150	1.5000	1.7114
151	2.0000	2.1932
152	2.5000	2.4678
153	0.5000	0.6429
154	1.5000	1.7403
155	1.0000	1.2559
156	2.0000	1.8254
157	0	−0.0976
158	3.0000	2.8746
159	1.0000	0.9637
160	3.0000	3.0812
161	3.0000	2.9221
162	1.0000	1.2994
163	3.0000	2.8575
164	0	0.0467
165	1.5000	1.7537
166	1.5000	1.2195
167	2.5000	2.7950
168	1.5000	1.5556
169	2.0000	1.7226
170	3.0000	3.0435
171	0.5000	0.4247
172	2.0000	1.6679
173	0	0.0561
174	1.5000	1.7550
175	1.0000	1.0100
176	0	−0.0189

shows the target value of the network and the value predicted by the network. In addition, in

Table 2. A comparison of the accuracy of the proposed detection system and previous studies.

Ref	Number of Detectors	Extracted Features	Source Type	Type of Neural Network	Maximum MSE	Maximum RMSE
[6]	1	Time features	137Cs	GMDH	1.24	1.11
[5]	2	Time features	137Cs	MLP	0.21	0.46
[45]	1	No feature extraction	60Co	GMDH	7.34	2.71
[46]	2	Frequency features	137Cs	MLP	0.67	0.82
[47]	1	No feature extraction	X-Ray tube	MLP	17.05	4.13
[48]	1	No feature extraction	137Cs	MLP	2.56	1.6
[49]	1	Compton continuum and counts under full energy peaks of 1173 and 1333 keV	60Co	RBF	37.45	6.12
[50]	2	full energy peak (transmission count), photon counts of Compton edge in transmission detector, and total count in the scattering detector	137Cs	MLP	1.08	1.04
[current study]	1	Wavelet features	Dual-energy gamma source	RBF	0.02	0.15

, a comparison is presented in terms of the error rate of the systems introduced in previous research and the detection system in this research. In this research, due to the correct processing of the obtained signals and the training of the neural network with effective characteristics, the error rate has been reported to be significantly low. On the other hand, the effective use of a detector and the reduction of the number of detectors has led to the optimization of the structure and the reduction of the construction cost. According to the present study procedure, the flows through the pipe and the various scale thicknesses inside the pipe were first simulated using the MCNP code, and the signals picked up by the detectors were labeled. Then, five characteristics of the average of a₄ and average of d1 to d₄ were obtained from the signals of each simulation after the received signals had been processed using the wavelet transform. The collected characteristics were used to forecast the scale thickness within the pipe using the RBF neural network. The output of the neural network was compared with the desired output after training to check that it was operating correctly. The main limitation of this study is the need to use protective equipment and clothing when working with radioisotope devices because they have harmful effects on the human body. Using the features and methodology introduced in this research and investigating the performance of different neural networks for productivity in future research are strongly recommended to researchers in this field.

6. Conclusions

Preventing sudden shutdowns and problems in the operation of oil equipment requires detecting the amount of scale inside the oil pipes and dealing with it. In this research, special attention is paid to this need, and an accurate system is used to detect the amount of sediment inside the pipes based on the gamma-ray attenuation technique to measure the thickness parameter of the internal scale of the pipe. The proposed structure for the detection system is simulated using the MCNP code, and a three-phase fluid passage pipe is tested in the stratified regime where scale is present. The dual energy source emits photons on the one hand, and the detector collects them on the other hand of the test pipe. The aforementioned flow was simulated in different volume percentages while different scale values were investigated. Two hundred fifty-two different conditions, including seven scale thickness values from 0 to 3 cm and 36 different volume percentages from 10% to 80%, were simulated. The received signals from the detectors finally led to the extraction of wavelet features, which were the inputs of the RBF neural network. The value of scale thickness was also considered as output in the training phase. The designed neural network had an input layer, a hidden layer, and an output layer, which had 5, 18, and 1 neurons, respectively. The scale thickness in this neural network was predicted by the designed RBF neural network with an MSE of 0.02, which is a much smaller error value according to the results of previous research. Due to its high accuracy, this detection system can be considered a vital part of oil fluid transfer systems, and its use is strongly recommended to oil industry managers. Future research in this area is advised to focus on the application of various neural networks, particularly deep neural networks, in detecting various three-phase current parameters and analyzing the time and frequency characteristics of the received signals.