An energy and leakage current monitoring system for abnormality detection in electrical appliances | Scientific Reports

The proposed system focuses on reducing fires caused by electrical appliances in any location through prompt, dependable monitoring and the use of a control scheme. The proposed system’s framework is depicted in Fig. 1; the process involves collaboration among SMDs, gateway systems, cloud servers, databases, detection algorithms, and visualization. SMDs are used for data acquisition, as shown in Fig. 1, and other necessary features are calculated from the data. Each consumer’s data is transmitted via multiple LoRa gateway channels and uploaded to a cloud server at irregular intervals. The proposed algorithm then categorizes the data based on the acceptable range of leakage current and the number of active appliances. The data from different places are stored and analyzed on the cloud platform because of the increasing number of installed SMDs. Afterward, we applied the proposed RBC-MSVM algorithm to identify the system’s abnormalities.

Figure 3

Schematic structure of the safety monitoring device.

Full size image

Figure 2 illustrates an overview of the proposed methodology, demonstrating the flows of sensing data and information to the cloud database. The system is divided into three parts: the appliance, the database, and the analysis. The appliance section is in charge of acquiring data and transmitting it to the data server via the LoRa module. The database section aims to collect and store sensor data in the database. The relationship between different variables was evaluated in the analysis section to identify the high coloration. Envisaging the households’ appliance specifications, we ascertained the acceptable leakage current to classify the system’s abnormalities. The proposed algorithm will determine the present circumstance regarding the system’s existing issue by investigating the historical data. Furthermore, the analysis section displays the real-time load profiles, leakage current profiles, and the system’s condition. In the following subsection, the detailed methodology is described with other relevant information.

Mục Lục

Device modeling and specification

The schematic diagram of the electrical safety monitoring device is shown in Fig. 3. The device is designed for a single-phase connection rated at 220–380 V (AC) and has the following dimensions: width: 37.5 mm, length: 64.6 mm, and height: 38.2 mm. The LoRa device includes several sensors that measure electrical parameters such as total current, terminal voltage, and leakage currents. From the measured data, we calculated the additional data required for each case, such as total power flow, energy consumption, power factor, resistive and capacitive leakage currents, and insulation resistance. Furthermore, we design in such a way that a multi-step warning signal about the permissible range of total current and residual current concerning the CB’s capacity is provided.

The STM32L microcontroller unit (MCU) handles the overall computation and data indexing. Low-Pass filter and Voltage-Divider are being used in the hardware for better analog data acquisition. Moreover, the STM32L MCU is integrated into the LoRa transceiver device in the proposed system to observe and make a difference in normal conditions. The LoRa system is consisted of end devices, gateways, and a network server that form a star topology with the network server at the root, gateways at level one, and end devices as leaves. The sensed and measured information are accumulated into each LoRa packet. One dedicated channel has been assigned for transmitting the LoRa packet in such an interval that the device remains idle for a certain period in normal operation to reduce power consumption. Furthermore, the device transmits data at very short intervals during the transition from normal to critical conditions. The used LoRa module (SX1276), which is connected to the MCU, sends these data packets to the LoRa gateway module via the 902–928 MHz omnidirectional antenna with a maximum gain of 2dBi. The LoRa network operates in the sub-GHz industrial, scientific, and medical band with maximum transmit powers of 21.7 dBm and 14 dBm in the USA and Europe, respectively31. The LoRa modulation (proprietary chirp spread spectrum modulation) uses different types of physical layer packets with different lengths in time, parameterized by the so-called spreading factor (SF), which can take values ${SF \in \mathbb {Z} | 7 \le SF \le 12}.$ The LoRa gateway is used to detect the fault location over a thousand meters because of its proprietary large area coverage32. The SF depends on the communication range’s requirement, where the low value of SF means low coverage and vice versa. To store the transmitted data, the interface between the LoRa gateway and the network server is provided by cellular Internet protocol that uses the standard transmission control protocol (TCP).

Figure 4

Situation for excessive leakage current.

Full size image

Figure 5

Workflow of the proposed safety monitoring device.

Full size image

Mathematical formulation

Figure 4 shows each possible approach of excessive leakage current flow. We demonstrated three scenarios: an insulation fault between the line and the ground, an insulation fault between the line and the neutral, and an appliance fault with the ground. However, Fig. 5 depicts the connection diagram and workflow of the proposed constructed device, which is deployed at the entry point of a low voltage power (i.e., 220–380 V) line in an electrical system (i.e., building, factory, and market). We consider the dynamic characteristic of loads in the proposed systems because electrical appliances are either turned on or off based on the consumer’s demand. The total apparent power of the systems can be defined as follows for N loads:

$$\begin{aligned} S_{T}(t)= \sum _{ap=1}^{\mathbf {N}} \left\{ P_{ap}(t)+jQ_{ap}(t)\right\} , \end{aligned}$$

(1)

where $P_{i}$ and $Q_{i}$ present the active and reactive power of the individual appliance. Therefore, the total currents entering into the loads ($I_{T}(t)=I_{1}(t)+I_{2}(t)+ \cdots$) is as follows:

$$\begin{aligned} I_{T}(t)&=I_{Zr,in}(t)+jI_{Xlc,in}(t) \end{aligned}$$

(2)

$$\begin{aligned} I_{Xlc,in}(t)&= I_{Xl,in}(t)- I_{Xc,in}(t), \end{aligned}$$

(3)

where $I_{Xl,in}(t)$ and $I_{Xc,in}(t)$ are the inductive and capacitive currents of the practical load, respectively and $I_{Zr,in}(t)= I_{T}(t)\cos \delta _{I,i}$ and $I_{Xlc,in}(t)= I_{T}(t)\sin \delta _{I,in}$ are the resistive and inductive current flowing to the circuit, respectively. The $\delta _{I,in}$ is also known as the power angle at normal conditions. Similarly, the total amount of returning current $I_{L,T}$ of the system can be defined as follows:

$$\begin{aligned} I_{L,T}(t)=I_{Zr,ot}(t)+jI_{Xlc,ot}(t), \end{aligned}$$

(4)

where $I_{L,T}(t)$ is defined as the total system current returning to the current sensor. $I_{Zr,ot}(t)= I_{L,T}(t)\cos \delta _{I,ot}$ and $I_{Xlc,o}(t)= I_{L,T}(t)\sin \delta _{I,ot}$ are the resistive and inductive current flowing to the circuit, respectively. Let’s consider a scenario of the system which is explained in Fig. 5.

$$\begin{aligned} \begin{array}{ll} I_{T}(t) = I_{L,T}(t); &{}\quad \text { at normal condition}, \\ I_{T}(t) \ne I_{L,T}(t); &{}\quad \text { at leakage current condition}. \end{array} \end{aligned}$$

The total leakage current ($I_{L}$) flowing out of the connected appliance after considering residual current can be formulated as follows:

$$\begin{aligned} I_{L}(t)&=I_{T}(t)-I_{L,T}(t) \end{aligned}$$

(5)

$$\begin{aligned} I_{L}(t)&= I_{rl}(t)+jI_{cl}(t), \end{aligned}$$

(6)

where the resistive and capacitive leakage currents are defined as $I_{rl}= I_{L}(t)\cos \delta _{L}$ and $I_{cl}(t)= I_{L}(t)\sin \delta _{L}$, respectively and $\delta _{L}$ is the angle between $I_{rl}$ and $I_{L}(t)$. Therefore, insulation impedance ($Z_{L}$) is equal to the LV bus $(V_{LV})$ voltage divided by the leakage current that flows through the insulation.

$$\begin{aligned} Z_{L} = V_{LV}(t)/I_{L}(t). \end{aligned}$$

(7)

Figure 6

Vector diagram for measuring leakage current.

Full size image

The quantity of leakage current is quite minimal when compared to the total load current because it only passes via the large insulating impedance of the faulty appliances during the breakdown of insulation. Figure 6 depicts the vector diagram for measuring leakage current wherein the amount of leakage current has considered as large for better visualization. Since the load current is so high in comparison to the $I_L(t)$, the total consumed energy does not differ considerably in normal conditions.

Figure 7

Hardware architecture of the safety monitoring device.

Full size image

Data acquisition and classification

Figure 7 shows the SMD device layout. There are two current sensors and one voltage sensor. One current sensor measures the total current of the system and the other sensor measures the leakage current of the system. For measuring the voltage, the terminal of the two wires should be placed as shown in Fig. 5. For measuring the current, the current sensor is only placed on the single wire while both of the wires will be entered inside the leakage current sensor. The leakage current sensor actually measures the difference between the two currents which is described in the Mathematical formulation section. For measuring the phase shift between voltage and current, two operational amplifiers are used for zero-cross detection. Thereafter, both outputs are used as input of an XOR gate. The ON-time of XOR output ( i.e. time difference between two phases) is used to determine the phase shift between voltage and current. Finally, the power factor (p.f.) of the system is measured which is used to determine active and reactive components of the current.

$$\begin{aligned} \delta _{I}&= f \times dt_{VI} \times 360, \end{aligned}$$

(8)

$$\begin{aligned} p.f.&= Cos\delta _{I}, \end{aligned}$$

(9)

where f and $dt_{VI}$ are defined as frequency and XOR output ON-time, respectively. For measuring the leakage current, we have used a leakage current sensor which is shown in Fig. 7. By using the leakage current and voltage sensor data, the phase angle ($\delta _{L}$) between leakage current and terminal voltage is calculated, similarly. Thereafter, the resistive and capacitive leakage current are measured for the system, accordingly.

$$\begin{aligned} \delta _{L} = f \times dt_{VI_{L}} \times 360, \end{aligned}$$

(10)

where f and $dt_{VI_{L}}$ are defined as frequency and XOR output ON-time, respectively.

However, to ensure greater system security, three warning types are provided. In this case, the over-current protection warning is designed based on the capacity of the deployed CB, whereas a multi-step warning is designed for leakage current protection by differentiating between resistive and capacitive residual currents. The consecutive state of the system SoS(t) for any consumer is classified by considering the system’s condition.

$$\begin{aligned} SoS(t) =\left\{ \begin{array}{ll} SoS^{N};&{}\quad \text {System runs at normal condition} \\ SoS^{W};&{}\quad \text {System runs at warning condition}\\ SoS^{C};&{}\quad \text {System runs at abnormal condition}. \end{array}\right. \end{aligned}$$

(11)

In the proposed scheme, we account for the two factors for classifying state and the other two factors for determining the type of appliance. Depending on the different threshold value ranges, the status is defined as $SoS\in \left\{ SoS_{I_{T}},SoS_{I_{L}},SoS_{I_{rl}},SoS_{I_{cl}}\right\}$. The dynamic states of the appliances in terms of total current and leakage currents are defined as $SoS_{I_{T}}\in \left\{ SoS_{I_{T}}^{N},SoS_{I_{T}}^{W},SoS_{I_{T}}^{C}\right\}$, $SoS_{I_{L}}\in \left\{ SoS_{I_{L}}^{N},SoS_{I_{L}}^{W},SoS_{I_{L}}^{C}\right\}$ because of the envisaging three-level warning. For tracing the type of devices, the vulnerability of resistive $SoS_{I_{rl}}\in \left\{ SoS_{I_{rl}}^{N},SoS_{I_{rl}}^{W},SoS_{I_{rl}}^{C}\right\}$ and capacitive leakage currents $SoS_{I_{cl}}\in \left\{ SoS_{I_{cl}}^{N},SoS_{I_{cl}}^{W},SoS_{I_{cl}}^{C}\right\}$ will be taken into consideration. Since the amount of current flow is controlled by the number of contracted appliances and their power rating, the threshold range will be determined accordingly. For additional convenience, we have recommended the opportunity of providing different threshold values. The cut off value of the uninterruptible and healthy system can be defined as $Th_{}^{N}\in \left\{ Th_{I_{T}}^{N},Th_{I_{L}}^{N},Th_{I_{rl}}^{N},Th_{I_{cl}}^{N}\right\}$. In the proposed system, we have considered the intermediate state between the secured and interrupting conditions. The set of range of the interim circumstance of the system is expressed as $Th_{}^{W}\in \left\{ Th_{I_{T}}^{W},Th_{I_{L}}^{W},Th_{I_{rl}}^{W},Th_{I_{cl}}^{W}\right\}$. The excessive current flow causes vulnerable state in the system that is known as critical condition $Th_{}^{C}\in \left\{ Th_{I_{T}}^{C},Th_{I_{L}}^{C},Th_{I_{rl}}^{C},Th_{I_{cl}}^{C}\right\}$. Therefore, the sanctioned constraints of distinguishable apprehension for the $I_{T}$ is as follows:

$$\begin{aligned} I_{T,s}^{N}&\le I_{T}(t)\le I_{T,e}^{N},\,\,\,\,\,\, \left\{ I_{T,s}^{N},I_{T,e}^{N} \right\} \in Th_{I_{T}}^{N}, \end{aligned}$$

(12)

$$\begin{aligned} I_{T,s}^{W}< I_{T}(t)&\le I_{T,e}^{W},\,\,\,\,\,\, \left\{ I_{T,s}^{W},I_{T,e}^{W} \right\} \in Th_{I_{T}}^{W}, \end{aligned}$$

(13)

$$\begin{aligned} I_{T,s}^{C}< I_{T}(t)&\le I_{T,e}^{C},\,\,\,\,\,\, \left\{ I_{T,s}^{C},I_{T,e}^{C} \right\} \in Th_{I_{T}}^{C}, \end{aligned}$$

(14)

where $\forall I_{T,s}^{N}\approx 0$, $\forall I_{T,e}^{N}\approx \forall I_{T,s}^{W}$ and $I_{T,e}^{W}\approx \forall I_{T,s}^{C}$.

figure a

However, the problem associated with leakage current may not remain in the overcurrent flowing system. Consequently, it is mandatory to comprise the leakage current detection to describe whether the system is secured or not. Similarly, the apprehensive state for leakage current will be ascertained based on the following constraints:

$$\begin{aligned} I_{L,s}^{N}&\le I_{L}(t)\le I_{L,e}^{N},\,\,\,\,\,\, \left\{ I_{L,s}^{N},I_{L,e}^{N} \right\} \in Th_{I_{L}}^{N}, \end{aligned}$$

(15)

$$\begin{aligned} I_{L,s}^{W}< I_{L}(t)&\le I_{L,e}^{W},\,\,\,\,\,\, \left\{ I_{L,s}^{W},I_{L,e}^{W} \right\} \in Th_{I_{L}}^{W}, \end{aligned}$$

(16)

$$\begin{aligned} I_{L,s}^{C}< I_{L}(t)&\le I_{L,e}^{C},\,\,\,\,\,\, \left\{ I_{L,s}^{C},I_{L,e}^{C} \right\} \in Th_{I_{L}}^{C}, \end{aligned}$$

(17)

where $\forall I_{L,s}^{N}\approx 0$, $\forall I_{L,e}^{N}\approx \forall I_{L,s}^{W}$ and $I_{L,e}^{W}\approx \forall I_{L,s}^{C}$. The probability of having a leakage issue in multiple devices at the same time is relatively high because of a complete electrical environment inspection. Hence, differentiating resistive and capacitive leakage currents accelerates the process of finding the corresponding appliances. For this reason, we introduced the acceptable range of leakage current using the conditional statement for investigating hazardous circumstances. Furthermore, the permissible limit of the leakage current varies with appliance type, application, and condition. Therefore, the constraints for a reliable and healthy system are defined as follows:

$$\begin{aligned} I_{rl,s}^{N}&\le I_{rl}(t)\le I_{rl,e}^{N},\,\,\,\,\,\, \left\{ I_{rl,s}^{N},I_{rl,e}^{N} \right\} \in Th_{I_{rl}}^{N}, \end{aligned}$$

(18)

$$\begin{aligned} I_{cl,s}^{N}&\le I_{cl}(t)\le I_{cl,e}^{N},\,\,\,\,\,\, \left\{ I_{cl,s}^{N},I_{cl,e}^{N} \right\} \in Th_{I_{cl}}^{N}, \end{aligned}$$

(19)

$$\begin{aligned} I_{rl,s}^{W}&< I_{rl}(t)\le I_{rl,e}^{W},\,\,\,\,\,\, \left\{ I_{rl,s}^{W},I_{rl,e}^{W} \right\} \in Th_{I_{rl}}^{W}, \end{aligned}$$

(20)

$$\begin{aligned} I_{cl,s}^{W}&\le I_{cl}(t)<I_{cl,e}^{W},\,\,\,\,\,\, \left\{ I_{cl,s}^{W},I_{cl,e}^{W} \right\} \in Th_{I_{cl}}^{W}, \end{aligned}$$

(21)

$$\begin{aligned} I_{rl,s}^{C}&< I_{rl}(t)\le I_{rl,e}^{C},\,\,\,\,\,\, \left\{ I_{rl,s}^{C},I_{rl,e}^{C} \right\} \in Th_{I_{rl}}^{C}, \end{aligned}$$

(22)

$$\begin{aligned} I_{cl,s}^{C}&\le I_{cl}<I_{cl,e}^{C},\,\,\,\,\,\, \left\{ I_{cl,s}^{C},I_{cl,e}^{C} \right\} \in Th_{I_{cl}}^{C}, \end{aligned}$$

(23)

where $\left\{ \forall I_{rl,s}^{N},\forall I_{cl,s}^{N} \right\} \in \left[ 0 \right]$, $\forall I_{rl,e}^{N}\approx \forall I_{rl,s}^{W}$, $\forall I_{cl,e}^{N}\approx I\forall _{cl,s}^{W}$ $\forall I_{rl,e}^{W}\approx \forall I_{rl,s}^{C}$, and $\forall I_{cl,e}^{W}\approx \forall I_{cl,s}^{C}$. By applying the given condition in Algorithm 1, we have determined the state of total and leakage currents. Therefore, we have applied Algorithm 2 to identify the current status of resistive leakage in the system. The procedure of finding the capacitive leakage current state is identical to that of determining the resistive leakage current condition; we only provide Algorithm 2 here. Since the boundary of the clusters is very close to each other, the classification algorithm may provide less accuracy. By considering this, we have scaled and re-scaled the features based on the following equations.

$$\begin{aligned} E_{k}(C_{k},x_{i})&= (1+C^{2}_{k})*x_{i}, \end{aligned}$$

(24)

$$\begin{aligned} F_{k}(C_{k},x_{i})&= \frac{E_{k}(C_{k},x_{i})*max(x_{i})}{max(E_{k}(C_{k},x_{i}))}, \end{aligned}$$

(25)

where $C_{k}$, $x_{i}$, $F_{k}$ are presented as $k_{th}$ cluster, $i_{th}$ data of the raw feature, and scaled feature which are selected to make up the cluster’s boundary.

figure b

Database and monitoring

The real-time data storing and monitoring added more value to the electrical safety analysis for understanding the system’s circumstances. Since the leakage current problem and the deterioration of the appliance’s insulation occurred over time, a large amount of data is required to accurately determine the condition of the installed equipment as well as the entire system. As a consequence, the cloud database33 is the best option for storing large amounts of data. Cloud computing is a model for providing convenient, on-demand network access to a shared pool of configurable computing resources that can be rapidly provisioned and released with minimal management effort and interaction from service providers. Cloud computing can also help to reduce the administrative burden of program management. The cloud environment enables very diverse data sources to gather information, store it in the cloud database, and feed distinct applications.

In the proposed system, the real-time data packets from the LoRa gateway are sent to the cloud database. On a Windows 10 PC, MySQL version 8.0.19 (Oracle, Co., Austin, TX, USA)34 was used as a database management system in the cloud (Microsoft, Redmond, WA, USA). MySQL is a multi-threaded, robust, and scalable open-source service, the platform used under either Oracle’s GNU General Public License or a standard business permit. However, the sensor data collected by the gateway is not uniform and contains noise. Following that, the database server begins intensive computational processing (such as summation, statistics, and data conversion). Finally, the data from several users are stored in the database, which will be used for further processing (such as feature extraction, training, and prediction).

Figure 8

Flow chart of fault diagnosis system.

Full size image

Fault classification and detection

In the proposed system, RBC has been applied to determine the device and over-current fault. And MSVM has been used as a discriminative classifier of the system conditions. The flow chart of detecting faults is shown in Fig. 8. In our cases, four rules are generated to diagnose the faults describes as follows:

Rule1: IF (Sensing data = yes) AND (Current level = normal) THEN the system goes normal
Rule2: IF (Sensing data = yes) AND (Current level = abnormal) THEN the system goes over-current fault
Rule3: IF (Sensing data = no) AND (Current level = normal) THEN the system goes device fault
Rule4: IF (Sensing data = no) AND (Current level = abnormal) THEN the system goes both device and over-current faults

For better classification accuracy, data cleaning, including duplicate and missing data, is conducted prior to categorizing the faulty condition. We have used Pearson’s correlation coefficient-based technique35 to remove unnecessary and redundant information and minimize complexity and dimensionality in the proposed system. The density of correlation depends on the Pearson correlation coefficient known as Pearson’s r. Let’s consider two variable matrix $S_{T}=[ S_{T_{1}},S_{T_{2}},\cdots , S_{T_{q}}]$ and ${I_{L}=[ I_{L_{1}},I_{L_{2}},\cdots , I_{L_{\mathbf {q}}}]}$, where q and $\mathbf {q}$ are represented as samples: $\bar{\gamma _{S_{T}}}=\frac{1}{q}\sum _{a}^{q}S_{T_{a}}$ and $\bar{\gamma _{I_{L}}}=\frac{1}{\mathbf {q}}\sum _{b}^{\mathbf {q}}I_{L_{b}}$. The Pearson correlation co-efficient can be defined as follows:

$$\begin{aligned} r_{S_{T},I_{L}}= \frac{\sum _{a=1,b=1}^{q,\mathbf {q}}(S_{T_{a}}- \bar{\gamma }_{S_{T}})(I_{L_{b}}- \bar{\gamma }_{I_{L}})}{\sqrt{\sum _{a=1}^{q}(S_{T_{a}}- \bar{\gamma }_{S_{T}})^2}\sqrt{\sum _{b=1}^{\mathbf {q}}(I_{L_{b}}- \bar{\gamma }_{I_{L}})^2}}. \end{aligned}$$

(26)

Similarly, the value of r is calculated by taking into account the other variables, with the feature being selected depending on the greater value of r.

To classify datasets, it tries to create an optimal hyperplane between two classes of the data set19. The hyperplane acts as a decision boundary to categorize the data into different classes. The points nearer to the hyperplane called support vector, are used to determine the optimized hyperplane. For a given training sample $\left\{ (x_i,y_i) \right\} ,\forall i\in \left\{ 1,2,3,….,n \right\}$, where $y_i \in \left\{ +1,-1 \right\}$ represents class labels, optimal hyperplane is determined by the following mathematical expression:

$$\begin{aligned} \theta ^{T}x_{i} + b = 0, \end{aligned}$$

(27)

where $\theta =\left[ \theta _{1},….,\theta _{n} \right]$ is n-dimensional vector of weights and $x_{i}=\left[ x_{1},x_{2},….,x_{n} \right]$ is an n-dimensional input vector, and b is termed as the biasing unit. Here, n represents number of features. The optimization problem associated with finding the hyperplane can be expressed as follows:

$$\begin{aligned} min(\theta )\frac{1}{2}\sum _{i=1}^{n}(\theta )^2=\frac{1}{2}\left\| \theta \right\| ^2 =\frac{1}{2}\theta ^{T}\theta , \end{aligned}$$

(28)

which is subjected to,

$$\begin{aligned} \theta ^{T}x_{i} + b \ge +1&\ \text {if}&\ y_{i}=+1, \end{aligned}$$

(29)

$$\begin{aligned} \theta ^{T}x_{i} + b \le +1&\ \text {if}&\ y_{i}=-1. \end{aligned}$$

(30)

The final nonlinear decision function can be obtained as follows:

$$\begin{aligned} f(x)=sign\left( \sum _{i=1}^{n} \alpha _{i} \left( \theta ^{T}x_{i} \right) +b \right) . \end{aligned}$$

(31)

To come up with a set of complex features, SVM uses a technique called Kernel $k(x_i,x)$. The value $k(x_i,x)$ corresponds to $\varphi (x_{i}).\varphi (x)$ which maps linearly non-separable patterns into a higher dimension feature space. Finally, the decision function can be modified as follows:

$$\begin{aligned} \begin{aligned} f(x)&=sign\left( \sum _{i=1}^{n} \alpha _{i} k(x_{i},x) +b \right) =sign\left( \sum _{i=1}^{n} \alpha _{i} (\varphi (x_{i}).\varphi (x)) +b \right) . \end{aligned} \end{aligned}$$

(32)

Table 1 Kernel function for the proposed system.

Full size table

In this study, we have performed the classification experiment taking account into four kernel functions (linear, polynomial, radial basis function (RBF), sigmoid) described in Table 1. Moreover, we have used one versus rest manner multiclass approach. According to this approach, for a mth class classification problem mth class are trained as positive samples while the rest are treated as negative samples21,36.

Figure 9