Hindawi

Mobile Information Systems

Volume 2022, Article ID 4715102, 11 pages
https://doi.org/10.1155/2022/4715102

Research Article
Data Anomaly Diagnosis Method of Temperature Sensor Based on
Deep Neural Network

Haifeng Chen
School of Information Engineering, Lianyungang Technical College, Lianyungang 222000, Jiangsu, China

Correspondence should be addressed to Haifeng Chen; 1992050004@lygtc.edu.cn

Received 10 July 2022; Revised 8 August 2022; Accepted 16 August 2022; Published 2 September 2022

Academic Editor: R. Mo

Copyright © 2022 Haifeng Chen. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

As one of the indicators of whether all kinds of machinery and electrical appliances work normally during use, temperature has an
important basis for judging the normal work of related machinery. In order to reduce the probability of safety and quality
problems caused by inaccurate temperature measurement in the use of these machines and electrical appliances, this paper uses
RBF neural network and EEMD modal analysis two deep neural network models to build a deep neural network-based tem-
perature sensor data anomaly diagnosis method. This method first excavates a large number of historical temperature data samples
of temperature control sensors in machinery and electrical appliances and analyzes the change law of relevant sample data, so as to
build a data anomaly diagnosis database, and then establish a temperature prediction model based on RBF to predict the
temperature of electrical appliances and mechanical components; Second, real-time temperature monitoring and sampling are
carried out for the normal temperature sensor. Based on the constructed sample database, automatic identification of various
abnormal conditions is realized, and the real measured value of the sensor is reconstructed or estimated under abnormal
conditions. EEMD feature extraction is carried out for the difference between the predicted temperature and the actual tem-
perature; Finally, the RBF temperature anomaly diagnosis and classification model is constructed, and the feature vector sets are
constructed by variance, variance percentage, energy and energy percentage methods, respectively, or jointly, and these vector sets
are used as the input of the fault model for temperature anomaly diagnosis and monitoring. Through the diagnosis of the
measured temperature sensor data, the established model has a good ability of fault diagnosis and classification.

1. Introduction inevitable. At this time, the performance of the automation

system will be degraded at least, and disastrous conse-
Sensor is an indispensable and important part of mea- quences will be caused at most. Accidents that cause
surement and control system, and it is the main tool to equipment damage and huge economic losses due to sensor
collect signals. With the increase and complexity of in- failure also occur from time to time. According to statistics,
dustrial machinery and civil electrical appliances, the the sensor failures in general automation systems account
complexity of measurement and control system is gradually for more than 45% of all failures [2, 3]. However, once the
increasing, the number and types of sensors are greatly sensor fails, the performance of the control system will be
increased, and the working environment is getting worse degraded at first, and then catastrophic consequences will be
and worse. In addition, the performance of the sensor is not caused. Therefore, the demand for on-site fault diagnosis of
only directly related to the running state of the equipment sensors is becoming increasingly strong, and it is very
but also related to the vital security issues. In particular, the necessary to carry out fault diagnosis of sensors.
working condition of sensors that provide control signals In the field of automatic control, the sensor converts the
directly affects the state of the system, so it is more important measured physical signal into electrical signal, which is very
[1]. The sensors composed of precision components are important for the operation of the monitoring system. When
often in a bad working environment, and failure is it fails, it will have a serious impact on the subsequent
2 Mobile Information Systems

monitoring, control, fault diagnosis, and other systems [4]. of sensor output signal is to set up multiple redundant
For example, in the locomotive transmission control system, sensors. Take the average value or multiple values, of course,
temperature is an important indicator of the working en- this will pay a high cost. With the development of sensor
vironment and state of the equipment, such as the oil technology, some intelligent sensors have been developed,
temperature of the traction transformer, the temperature which makes self-detection and self-diagnosis possible. At
sensor temperature of the water inlet and outlet of the present, most sensor neural network fault diagnosis methods
traction converter water cooling system, the air temperature use BP neural network [16–20]. Generally, offline training
of the cabinet, and the temperature of the traction motor. If and online work processing mode are adopted in data
the temperature is too high, it is necessary to reduce the processing. However, BP neural network has the problem of
power of the locomotive, disconnect the main circuit slow convergence and easy to fall into local minima. In
breaker, and other protective actions to ensure the safe addition, the offline processing method requires that when
operation of the equipment. The operation environment of training the network, a large number of training samples
traction drive system device is complex, and corrosion, containing all the characteristics of the sensor must be
temperature, humidity, electric surge, static electricity, and provided to the network. Otherwise, when the field data
other factors will affect its operation state [5, 6]. exceed the coverage of the neural network training samples,
In foreign countries, the research on abnormal diagnosis the network will often output error information due to the
of temperature sensor data has a long history. As early as the lifting capacity of the network.
1950s, scientists used grain monitoring system to ensure In recent years, ANN has been widely used in sensor
food security. Developed countries in Europe and the United fault diagnosis, especially for BP neural network. At present,
States have an early scientific and technological revolution it covers most fields. The principle of BP artificial neural
and have always been in the lead in technical research in network is relatively simple. It was originally evolved from
related fields. They have an efficient temperature monitoring multilayer perceptron and used for network training with
system and standards. On the basis of previous studies, Zhao unique back propagation [20–23]. However, when adjusting
et al. introduced the concept of dynamic radius support the weight of BP neural network, it is different from other
vector data description (DR SVDD) and the idea of kernel neural networks, which often adopts the negative gradient
space angle to detect turboshaft engine faults based on Dr descent method. It is precisely for this reason that the
SVDD [7]. Msi et al. proposed a new fault detection system network is prone to fall into local extremum and poor
for large-scale grid connected photovoltaic power stations. convergence speed for some complex training samples,
The fault detection system makes a cascade comparison which limits its scope of application. Therefore, people turn
between the DC power supply of the actual photovoltaic their attention to another radial basis function (RBF) neural
power station and the simulated photovoltaic power station, network that can solve the above shortcomings. It is superior
distinguishes the fault situation, and identifies the nature of to BP neural network in approximation ability, classification
the fault [8]. Hd et al. proposed a fault detection method ability, and so on [24, 25]. At the same time, Wang et al. [26]
based on asymmetric pole inductance (APIs), which has used the limit learning machine method to classify the faults
certain accuracy in fault detection [9]. Li et al. systematically of the fuel system. Guo et al. [27] combined the circular
studied the application of ANN and hybrid ANN models in model with the limit learning machine (ELM) to form a fault
photovoltaic fault diagnosis. For each application, the target diagnosis method for linear analog circuits. Xia et al. [28]
photovoltaic fault, detectable fault, data type and amount reported an effective diagnosis method for early faults of
used, model configuration, and FDD performance were water chillers by combining nuclear entropy component
extracted and analyzed [10]. Duan et al. Used the correlation analysis (KECA) and voting based ELM (VELM). Li et al.
between atmospheric data and granary historical data to [29] proposed a method based on extreme learning machine
propose SVR method to predict the average temperature of and ADABoost Samme’s nuclear power plant fault diagnosis
grain pile [11]. Kurpaks et al. established a dynamic model of framework. Liu et al. [30] proposed a new gear personalized
water content and greenhouse heat flow, and expounded the fault monitoring model in combination with two different
movement of indoor heat and mass [12]. Chang et al. neural networks (finite element method and elm method) to
proposed the finite difference method to establish a model to solve the problem of gear fault when mechanical devices
predict the moisture content and distribution of grain [13]. work. The above research combines various methods with
Szke et al. used the temperature of the last two days in limit learning machine for fault diagnosis and has achieved
combination with the maximum and minimum temperature good results, but most of them have poor universality.
given by the weather forecast to predict the temperature in Moreover, due to the randomness of the weights and
the next 24 hours [14]. Fan et al. described the actual thresholds of ELM, the results of each time are unstable,
greenhouse system based on the greenhouse temperature which is easy to affect the prediction results. Therefore, many
prediction model of SVR [15]. These scholars have done a lot scholars have studied a fault diagnosis classification model
of research on fault diagnosis, but the abnormal monitoring based on empirical mode decomposition (EEMD). After
of temperature sensor data based on deep neural network is analysis, it is found that the model has strong signal analysis
not comprehensive enough, and there is a lack of practical technology ability and is very suitable for dealing with this
prediction system. kind of problems.
However, in China, the research of sensors is relatively In order to solve the above problems, this paper proposes
lagging behind. The traditional way to improve the reliability an intelligent temperature data anomaly diagnosis method
Mobile Information Systems 3

Temperature sensor
A1 Temperature signal
Constant Control unit
acquisition circuit
current power
supply

B1 +

Same
Constant Uin proportion
PT100 current power operational ADC
supply amplifier
B2 - circuit

A2

Figure 1: Schematic diagram of temperature signal sampling.

based on RBF neural network and empirical mode de- In Figure 1, both ends of PT 100 are connected to 4
composition (EEMD), which is based on historical fault data through A1 and A2 ends 9 mA constant current source, and
samples and normal historical data of temperature sensors. then the temperature signal sampling circuit obtains its
The rule base of sensor fault diagnosis is established by differential input voltage Uin through B1 and B2. Uin gen-
analyzing the characteristics of fault samples, and the fault- erates a single ended voltage signal Uin′ ’through the voltage
tolerant estimation method of fault conditions is obtained. follower and then enters the in-phase proportional opera-
For the fault conditions that cannot be reconstructed from tion amplification circuit to generate a voltage signal that
the fault samples, the normal historical data of the sensor are meets the sampling requirements of the analog to digital
used to establish the grey prediction model of the measured converter (ADC) of the control unit.
value of the sensor, so as to realize the fault-tolerant esti- If the amplification factor of the in-phase proportional
mation of the nonisolated fault samples, and carry out the operation amplification circuit is k, the relationship between
research on the abnormal diagnosis method of temperature the sampling temperature value t of the temperature sensor
sensor data based on deep learning. (PT100) and the output voltage value u of the temperature
signal sampling circuit is shown in formula (1).
2. Fundamental Theory U  (0.385T + 100) × 0.0049K. (1)

2.1. Brief Introduction of Signal Sampling Circuit of Tem-

perature Sensor. Thermal resistance (such as PT100) is a
temperature sensor that converts temperature into resis- 2.2. Basic Principle of RBF Neural Network. MP model de-
tance based on the principle that its resistance value changes scribes neurons from the perspective of logic functional
with temperature. The temperature measurement circuit devices, which is a mathematical simplification of biological
obtains the resistance value (voltage/current) by first ap- neuron information processing mode and establishes the
plying a known excitation current to the thermal resistance theoretical research foundation of neural networks.
and then measuring the voltage at both ends and converts In 1988, broomhead and low e introduced radial basis
the resistance value into the temperature value, so as to function into neural network to form RBF neural network
realize the temperature measurement [31]. [32]. RBF neural network is a three-layer feedforward
There are three connection modes between thermal network. Its basic idea is to use RBF as the “base” of hidden
resistance and temperature measuring circuit: two-wire units to form the hidden layer space and transform the low-
system, three wire system, and four wire system. Because the dimensional input vector into the high-dimensional space
four wire measurement method is not affected by the re- through projection. So that the original linear inseparable
sistance of connecting wires, it has been widely used in the problem becomes linear separable. Figure 2 shows the basic
field. This wiring method is also used in most locomotive structure of RBF neural network.
transmission systems. Therefore, this paper takes the four For the structure of RBF network, its principle is rela-
wire temperature sensor as an example to study its signal tively special. The hidden layer space formed by it can di-
sampling circuit principle, fault diagnosis, and fault-tolerant rectly map the input vector to the hidden layer through
estimation methods. neurons, so as to reduce the weight association of the middle
The typical temperature signal sampling principle is layer. Therefore, the association weight from the input layer
shown in Figure 1. to the hidden layer of this neural network is also relatively
4 Mobile Information Systems

φ1 (||x–c1||) the distance to the origin in a simplified case, that is,

c1
φ(x) � φ(‖x‖). Using Gaussian kernel function as the basis
function of radial basis function neural network, the output
b1
x1 of hidden unit of radial basis function neural network is

c2
y1 �� �� 1 �� ��2
x2 φi x, ci 􏼁 � G􏼐��x − ci ��􏼑 � exp 􏼠− 2 ���xp − ci ��� 􏼡, (2)
2σ i
ym where φ is the radial basis function, x is the sample, ci is the
ith center point of the kernel function, and σ i is the width of
xn bm the ith center point of the function. The selection of the
ck
central point of kernel function is very critical. The improper
φk (||x–ck||) central position cannot make the network correctly reflect
Input layer Hidden layer Output layer
the actual distribution of the input sample space, and the
Figure 2: RBF neural network. input space cannot be well fitted. The width of the central
point of the kernel function controls the radial range of the
function, which is an important factor affecting the per-
special, which is 1. The hidden layer is only responsible for formance of RBF neural network. When the width is too
the nonlinear projection of the input vector of the input small, the dividing line between classes will become blurred,
layer, and the output layer is only responsible for the which will reduce the classification accuracy; When the
weighted sum of the values mapped by the hidden layer. The width is too large, the coverage area of the basis function will
parameters to be learned and optimized in RBF neural become relatively small, thus reducing the generalization
network include the center and variance of radial basis ability of the network.
function and the connection weight from hidden layer to Then the output of RBF neural network is
output layer. The output layer is responsible for optimizing
the weights through linear optimization strategy, and the h
1 ��� ��2
��
learning speed is usually fast. The hidden layer needs to use yj � 􏽘 wij exp 􏼠− �
2 �xp − ci � 􏼡, j � 1, 2, · · · , n, (3)
i�1 2σ i
nonlinear optimization method to adjust the parameters of
the activation function, so its learning speed is relatively where yj represents the output of the RBF neural network,
slow. In the learning process of RBF neural network model, xp represents the pth input sample, ci represents the ith
the learning method of its parameters is not unique, but the center point, σ i represents the width of the ith center point of
selection of radial basis function center is mostly used, the function, wij represents the connection weight coeffi-
mainly including orthogonal least square method and su- cient between the hidden layer neuron i and the output layer
pervised center method, as well as random center method neuron j, h represents the number of nodes in the hidden
and self-organizing selection method. In addition, the layer, and n is the number of output samples or
learning process of RBF neural network includes two stages: classifications.
the first stage is to solve the center and variance of the hidden Based on the above theory, it can be found that the
layer basis function, which is also called unsupervised performance index is less than the given error by increasing
learning process; the second stage is the supervised learning the number of hidden layer elements, so as to continuously
process, which determines the connection weight between improve the fitting accuracy. However, in practical appli-
the hidden layer and the output layer. RBF neural network cations, if the number of hidden layer elements is too large, it
belongs to local approximation network, which omits the may cause the redundancy of the model and numerical ill
learning behavior of hidden layer weights and avoids the conditioned. Therefore, effective methods must be adopted to
time-consuming layer by layer transmission process of er- select the network center and determine the network weight.
rors in the network. Therefore, the learning convergence
speed of the network is very fast. Compared with other (1) Cluster analysis method is used to select RBF Net-
neural networks, RBF neural network can approach any work Center. After selecting the center of RBF
nonlinear function with any accuracy and has the best network, the determination of weight becomes a
approximation performance, classification ability, and problem about parameter linearization. The least
global optimization characteristics. Moreover, it has simple square method can be used to determine the network
topology, small amount of calculation, good applicability of weight. The advantage of this method is that the
the network, the basic parameters such as the structure of the center is easy to determine, the calculation time is
network, the number of hidden layer units, and constants short, and it is not easy to appear numerical ill
can also be dynamically adjusted and determined, with fast conditioned. The disadvantage is that the selection of
convergence speed. RBF center is separated from the determination of
Radial basis function is a real valued function whose weight, and the center obtained cannot be guaran-
value only depends on the distance from the fixed point c, teed to be the best.
and any function satisfying the φ(x, c) � φ(‖x − c‖) char- (2) The least square problem is orthogonalized by or-
acteristic φ. Both are radial basis functions, which can also be thogonalization algorithm. When all centers of RBF
Mobile Information Systems 5

neural network are selected, the weights are also Table 1: Sensor temperature sampling data.
determined. The advantage of orthogonalization Temperature of different sensors at time t/°C
Item
method is that it can ensure to select the best sample
Time/min Sensor 1 Sensor 2 Sensor 3 Sensor 4
point as the center. Its disadvantage is that when the
number of sample points is large, the workload of 1 23.4 23.2 22.9 24.5
2 26.8 27.3 24.1 26.6
orthogonalization is large, the calculation time is
3 31.2 29.8 29.6 30.2
long, and numerical ill condition often occurs. 4 37.4 36.7 41.5 38.4
(3) The ideal method is to combine the clustering 5 42.8 41.8 36.4 42.3
method with the orthogonalization method, that is, ⋮ ⋮ ⋮ ⋮ ⋮
the clustering method is used for primary selection, 56 44.7 48.5 45.3 54.9
and then the orthogonalization method is used for 57 49.8 50.6 44.5 51.4
selection, and the weight of the network is deter- 58 46.8 48.8 56.2 56.9
mined at the same time. 59 49.6 45.2 51.9 47.3
60 50.2 47.5 53.7 48.4

2.3. Set Empirical Mode Decomposition (EEMD). where: xj (t) is the signal after adding random white
Empirical mode decomposition (EMD) is a decomposition noise, j � 1, 2, 3 · · · M, and M is the number of tests.
method that can reflect the instantaneous frequency of data
(2) x(t) is decomposed into a series of IMFs(ci,j ), using
[33], but because the added noise is random, it is easy to
EMD:
cause modal aliasing, endpoint effect, screening iteration
n
stop standard, and other problems. To solve the above
problems, Hafida et al. [34, 35] proposed integrated em- xj (t) � 􏽘 ci,j (t) + rnj (t), (6)
i�1
pirical mode decomposition (EEMD). This method solves
the phenomenon of modal aliasing by adding white noise to where ci,j is the ith IMF of the jth test; rnj is the
fill the discontinuous signal segment. In the process of noise residual of the jth test, and nj is the number of IMFs
signal decomposition, the filtering characteristics of white of the jth test.
noise signal are used to solve the average value of intrinsic
mode component (IMF) for many times to eliminate the (3) If j < M, repeat steps 1 and 2, adding different ran-
interference of white noise on the original signal at dis- dom white noise signals each time.
continuous points [36]. (4) Obtain I � min (N1, N2, . . ., NM) and calculate the
EEMD is an improved method of empirical mode de- overall average of the corresponding components as
composition EMD [35]. The method of decomposing the the final result:
original signal into several intrinsic mode functions (IMF),
􏼐 􏽐M
j�1 ci,j 􏼑
each IMF must meet two conditions: ci . (7)
M
(1) In all decomposed data sets, the number of extreme
points and zero crossings is equal or at most 1
difference; 3. Design of Temperature Sensor Data Anomaly
(2) The average value of the upper and lower envelope Diagnosis Model Based on RBF Neural
determined by the local maximum point and the Network and EEMD Modal Analysis
local minimum point is 0. After EMD decomposi-
tion, the original signal x(t) can be expressed as 3.1. Data Acquisition and Pretreatment
n 3.1.1. Data Acquisition. To fully reflect the authenticity of
x(t) � 􏽘 ci (t) + rn (t). (4) the experiment. The author selected five normal sensors with
i�1 good performance to collect the experimental data. When
the system was put into operation, the sensors were also in
EMD has the problem of mode aliasing, that is, a certain normal operation and no fault occurred. At this time, the
IMF component has similar characteristics (such as fre- sampling data sequence within 1 hour of the sensor obtained
quency, amplitude, etc.) in other components, which cannot through time-sharing data acquisition is shown in Table 1.
fully reveal the characteristic information of the signal.
EEMD adds white noise with a certain frequency and am-
plitude uniformly distributed to the original signal, which 3.1.2. Data Preprocessing. At present, the longest method in
makes the signal centralized and continuous, and weakens data standardization processing is Z-score standardization,
the influence of instantaneous pulse on signal decomposi- which is also one of the default methods of SPASS. This
tion [33, 35]. The specific decomposition steps are as follows: method is to standardize the data through the mean value
(i.e. mean value) and standard deviation (i.e. standard de-
(1) Add random white noise signal r to x(t): viation) of the original data set and preprocess the measured
xj (t) � x(t) + nj (t), (5) data according to the characteristics of the signal transmitted
and output by the temperature sensor. The preprocessing
6 Mobile Information Systems

Sensor 1 Sensor 2
16 14

14 12

12
10
10
8
8
6
6
4
4

2 2

0 0
30 40 50 60 25 30 35 40 45 50 55

Sensor 3 Sensor 4

17.5 14

15.0 12

12.5 10

10.0 8

7.5 6

5.0 4

2.5 2

0.0 0
25 30 35 40 45 50 55 25 30 35 40 45 50 55
Figure 3: Original temperature measurement data.

code is shown in formula (9), and the data processed by this sequence of sampling points and has the property of random
method are shown in Figures 3 and 4. From the law of the distribution.
data in Figures 3 and 4, it can be seen that the data pre- The EEMD method is used to decompose the abnormal
processed by Z-score standardization only change in the data features of the results in Figure 5 layer by layer, and the
direction of the value of the data, but does not change the results are shown in Figure 6. Imf1 refers to the high-fre-
size relationship of the data itself, and the data processed by quency random noise of the sensor, which fluctuates vio-
this method can eliminate the influence of the value of the lently, and has obvious random variability, strong
data itself on network training and feature selection, so as to nonlinearity, and insignificant periodicity; Imf2 is the high-
speed up network learning and accurately assign training frequency periodic component of the output signal of the
weights. sensor under external influence; Imf3 refers to the sharp
x−μ fluctuation of the output signal caused by the abnormal data
xnormalization � , (8) fluctuation of the sensor; and Imf4 represents the low-fre-
δ
quency periodic component of the sensor output signal
caused by the long-term influence of influencing factors.
Follow the same steps to extract the features of the ab-
3.2. EEMD Abnormal Data Feature Extraction. Taking a normal data of the four temperature sensors used in this paper.
sensor with a lot of abnormal data as an example, the re-
sidual signal between the predicted temperature and the
measured temperature is converted into a continuous time 3.3. Model Design. The fault diagnosis modeling of tem-
series, as shown in Figure 5. The abnormal data occur in each perature sensor drift fault, accuracy decline fault, impact
Mobile Information Systems 7

Sensor 1 Sensor 2

16 14

14 12

12
10

10
8
8
6
6

4
4

2 2

0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Sensor 3 Sensor 4

14
17.5

12
12.5

10
15.0

8
10.0

7.5 6

5.0 4

2.5 2

0.0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Figure 4: Normalized temperature measurement data.

fault, and fixed deviation fault is mainly composed of three EEMD, the fault features are extracted from each compo-
parts: temperature prediction model is based on RBF neural nent, and the fault signal is highlighted to improve the
network; EEMD method is used to extract fault features of recognition rate.
temperature difference between prediction and measure-
ment; and the anomaly diagnosis scheme based on RBF 4. Analysis of Experimental Results
neural network is shown in Figure 7.
The residual signal between the predicted temperature In the test experiment, the normal sensor and the faulty
and the measured temperature still has large fluctuations. sensor are sampled once every 1 min, respectively, and 240
When the fault signal is not obvious, it is easy to be hidden samples are collected by four temperature sensors within
by high-frequency fluctuations. The simple threshold dis- 60 min. Figure 8 reflects the fault transmission signal of an
crimination method cannot identify the fault and judge the abnormal temperature sensor within 60 min of sampling,
specific fault type. At the same time, when the temperature and Figure 9 shows the normal output signal of the normal
prediction accuracy is low, the converted continuous time sensor within 60 min of sampling. From the two temperature
series fluctuates greatly, which is easy to cause misjudgment. data transmission signal curves, it can be seen that the failure
Therefore, the temperature residual signal is decomposed by of the temperature sensor generally occurs 20 minutes after
8 Mobile Information Systems

20 50

15
40
Temperature difference (°C)

10

5 30

T (°C)
0 20
-5
10
-10

-15
Abnormal data Normal data 0 10 20 30 40 50 60
-20 t (min)
0 10 20 30 40 50 60
Sample number Figure 8: Normal temperature sensor fault transmission signal
diagram.
Figure 5: Fixed deviation fault residual signal.

50
Signal 0
40
0 10 20 30 40 50 60

IMF1 0 T (°C) 30
0 10 20 30 40 50 60
20
IMF2 0

0 10 20 30 40 50 60 10

IMF3 0

0 10 20 30 40 50 60 0 10 20 30 40 50 60
t (min)
IMF4 0
Figure 9: Abnormal temperature sensor fault transmission signal.
0 10 20 30 40 50 60
Figure 6: EEMD decomposition effect.
Start

Measured
temperature real-time data
data Normal NO
Feature Abnormal
extraction by identification
Parameter
modal of temperature Mean filtering
RBF neural decomposition sensor reading
Abnormal
network
prediction
model
Temperature Real time
Figure 7: Abnormal diagnosis scheme of temperature sensor based judgment classification
on deep neural network.
YES

Abnormal alarm END

the sensor works normally, that is, the sensor temperature
rises to a certain temperature, which leads to large fluctu- Figure 10: Temperature anomaly detection algorithm flow.
ations in the signal value.
At the same time, 240 sets of sample data measured by
the above four sensors are used to evaluate the classification situation of temperature detection. The process of temper-
effect of combined eigenvectors and RBF neural network. ature anomaly detection algorithm is shown in Figure 10.
First, input the collected acceleration data and filter, then The standard system used to evaluate the effectiveness of
read the model training parameters, and then quickly classify temperature sensor anomaly diagnosis algorithm in this
the data through correlation operation to identify and judge paper mainly includes three indicators: accuracy, sensitivity,
whether the sensor temperature is in an abnormal state. and specificity. Accuracy refers to the proportion of accu-
Finally, output the alarm signal according to the actual rately detecting all abnormal data and nonabnormal data.
Mobile Information Systems 9

Input layer Hidden layer Onput layer Table 2: Test experimental results.
Sensor Number TP FN FP TN
S
1 60 0 2 1 57
2 60 0 0 0 60
hw(x)
3 60 54 1 1 4
D 4 60 58 0 1 1

numerical experiments is finally determined to be 280. The

ΔS … abnormal temperature data diagnosis model formed by this
is shown in Figure 11. The test results are shown in Table 2.
Δt
As can be seen from Table 2, a total of 240 samples were
tested. According to the defined formula, the accuracy,
Figure 11: A temperature sensor data anomaly diagnosis model sensitivity, and specificity of the algorithm are 97.5%, 96.8%,
based on the data in this paper. and 97.9%, respectively. The test results confirm the effec-
tiveness, accuracy, and feasibility of the algorithm.

The higher the accuracy is, the better the effect of this
5. Conclusion
method is; sensitivity refers to the proportion that all ab-
normal data are accurately detected. The higher the sensi- With the promotion of household appliances and industrial
tivity, the lower the misjudgment rate; specificity refers to machinery, temperature, as one of the indicators of normal
the proportion that all nonabnormal data are correctly operation of various machinery and electrical appliances in
detected. The higher the specificity, the lower the mis- the process of use, also occurs from time to time due to
judgment rate. equipment damage and huge economic losses caused by
TP + TN sensor failure. This paper analyzes the above problems and
accuracy � , draws the following conclusions:
TP + TN + FP + FN
(1) Based on RBF neural network and empirical mode
TP
sensitivity � , (9) decomposition (EEMD) deep neural network and
TP + FN based on the historical fault data samples and normal
TN historical data of temperature sensors, an intelligent
specificity � . temperature data anomaly diagnosis method is
TN + FP
proposed to solve the temperature anomaly problem
Among them, TP (real case) is the number of abnormal that is easy to occur in temperature sensors at
temperature data events and falls detected, which belongs to present.
correct judgment; FN (false counterexample) refers to the (2) The sensor in actual work is taken as the object, and
number of abnormal temperature data events but not de- the measured data are used as the input of the fault
tected, which belongs to missed judgment; FP (false positive model to diagnose and monitor the temperature
example) is the number of times that no abnormal tem- abnormality. Through the diagnosis and verification
perature data event occurs but abnormal temperature data of the measured temperature sensor data, the
are detected, which is a misjudgment; TN (true counter- established model has a good fault diagnosis and
example) is the number of times when no abnormal tem- classification ability, with an accuracy rate of 97.5%, a
perature data event occurs and no abnormal temperature sensitivity of 96.8%, and a specificity of 97.9%. It
data are detected, which belongs to correct judgment. The shows that the method constructed in this paper can
number of hidden layer nodes of neural network affects the effectively provide data abnormal diagnosis services
generalization ability and complexity of neural network. If for the temperature sensor.
the number of hidden layer nodes is too small, the network
will get less useful information, the model description ability
is insufficient, and the fault tolerance is poor; too many Data Availability
hidden layer nodes will increase the training time, and the The data used to support the findings of this study can be
network may store irregular information in the samples, obtained from the author upon request.
which may lead to the “over fitting” problem and the decline
of generalization ability. At present, there is no perfect
theory for the selection of the number of hidden layer nodes
Conflicts of Interest
of artificial neural network, which is mainly verified by The author declares that there are no conflicts of interest.
numerical value based on previous experience. The input
layer of the RBF neural network constructed in this paper
References
contains 4 neurons, the output layer contains 2 neurons, and
the number of neurons in the hidden layer is 200 according [1] P. Chen, C. C. Chen, C. C. Tsai, and W. F. Lu, “A time-to-
to the initial value, 20 steps, and 20 iterations. The number of digital-converter-based CMOS smart temperature sensor,”
10 Mobile Information Systems

IEEE Journal of Solid-State Circuits, vol. 40, no. 8, pp. 1642– Conference Series Materials Science and Engineering,
1648, 2005. Changchun, China, August 2009.
[2] B. Feizifar, Z. Muller, and O. Usta, “A new algorithm for [18] Y. Jiang and F. Zhang, “Study on BP neural network opti-
detecting failure to clear mode of circuit breakers using mization by improved decay parameter genetic algorithm,”
fundamental frequency component of voltage signals,” IEEE Journal of Physics: Conference Series, vol. 1621, no. 1, Article
Transactions on Power Delivery, no. 99, p. 1, 2019. ID 012054, 2020.
[3] G. Pham and Ritter, “On-chip sensors to detect impending [19] J. Rong and H. Ge, “Hydroelectric generating unit vibration
metallization failure of LDMOS transistors under repetitive fault diagnosis via BP neural network based on particle swarm
thermo-mechanical stress,” IEEE Transactions on Semicon- optimization,” in Proceedings of the 2009 International
ductor Manufacturing: A Publication of the IEEE Components, Conference on Sustainable Power Generation and Supply,
Hybrids, and Manufacturing Technology Society, the IEEE IEEE, Nanjing, China, April 2009.
Electron Devices Society, the IEEE Reliability Society, the IEEE [20] L. Quan and X. Jiang, “Sensor fault diagnosis based on discrete
Solid-State Circuits Council, vol. 29, no. 3, pp. 193–200, 2016. wavelet transform and BP neural network,” Proceedings of
[4] C. D. Breach, F. Wulff, and C. W. Tok, “An unusual me- SPIE - The International Society for Optical Engineering,
chanical failure mode in gold ballbonds at 50μm pitch due to vol. 5998, 2005.
degradation at the Au–Au4Al interface during ageing in air at [21] Z. R. Hou, “Rolling bearing fault diagnosis based on wavelet
175°C,” Microelectronics Reliability, vol. 46, no. 2-4, packet and improved BP neural network for wind turbines,”
pp. 543–557, 2006. Applied Mechanics and Materials, vol. 347-350, pp. 117–120,
[5] H. L. Massie, “Channel Utilization by Remote Locomotive 2013.
Control Systems Using Digital transmission,” in Proceedings [22] Z. Feng, X. Zhang, and Y. He, “Research of pneumatic ac-
of the Vehicular Technology Conference, IEEE, Detroit, MI, tuator fault diagnosis method based on GA optimized BP
USA, September 1976. neural network and fuzzy logic,” in Proceedings of the In-
[6] R. Niwas and B. Singh, “Solidstate controller for threephase ternational Conference on Advances in Neural Networks,
permanent magnet synchronous generatorbased diesel gen- Springer, Berlin, Germany, July 2013.
erator set feeding singlephase loads,” IET Power Electronics, [23] G. Liu, X. Wang, and F. Li, “Fault diagnosis of diesel engine
vol. 8, no. 10, pp. 1844–1852, 2015. fuel system based on genetic algorithm optimizing BP neural
[7] Y. P. Zhao, Y. L. Xie, and Z. F. Ye, “A New Dynamic Radius network model and D-S evidence theory,” An International
SVDD for Fault Detection of Aircraft engine,” Engineering
Journal of Research & Surveys.Part B Applications, vol. 3,
Applications of Artificial Intelligence, vol. 100, 2021.
pp. 707–716, 2012.
[8] A. Msi, B. Yakn, A. Uak, and B. W. Lee, “Real-time fault
[24] W. M. Lin, C. D. Yang, J. H. Lin, and M. Tsay, “a fault
detection system for large scale grid integrated solar photo-
classification method by RBF neural network with OLS
voltaic power plants,” International Journal of Electrical Power
learning procedure,” IEEE Power Engineering Review, vol. 21,
& Energy Systems, vol. 130, 2021.
no. 8, p. 60, 2001.
[9] A. Hd, A. Mh, and B. Wpa, “DC Grids DC Fault Detection
[25] X. Li, Z. Tao, and D. Zhe, “A Recognition Method of Plate
Using Asymmetric Pole inductors,” Electric Power Systems
Shape Defect Based on RBF-BP Neural Network Optimized
Research, vol. 193, 2021.
[10] B. Li, C. Delpha, D. Diallo, and A Migan-Dubois, “Application by Genetic algorithm,” in Proceedings of theThe 26th Chinese
of Artificial Neural Networks to photovoltaic fault detection Control and Decision Conference, Changsha, China, May 2014.
and diagnosis: a review,” Renewable and Sustainable Energy [26] H. Wang, W. Jing, Y. Li, and H. Yang, “fault diagnosis of fuel
Reviews, vol. 138, Article ID 110512, 2021. system based on improved extreme learning machine,” Neural
[11] S. Duan, W. Yang, X. Wang, S. Mao, and Y. Zhang, “Fore- Processing Letters, vol. 53, no. 4, pp. 2553–2565, 2021.
casting of Grain Pile Temperature from Meteorological [27] S. Guo, B. Wu, J. Zhou et al., “An analog circuit fault diagnosis
Factors Using Machine Learning,” IEEE Access, vol. 7, 2019. method based on circle model and extreme learning ma-
[12] S. Kurpaska, Z. Ślipek, B. Bożek, and J. Fraczek, “Simulation of chine,” Applied Sciences, vol. 10, no. 7, p. 2386, 2020.
heat and moisture transfer in the greenhouse substrate due to [28] Y. Xia, Q. Ding, A. Jiang, N. Jing, W. Zhou, and J. Wang,
a heating system by buried pipes,” Biosystems Engineering, “Incipient fault diagnosis for centrifugal chillers using kernel
vol. 90, no. 1, pp. 63–74, 2005. entropy component analysis and voting based extreme
[13] C. S. Chang, H. H. Converse, and J. L. Steele, “Modeling of learning machine,” Korean Journal of Chemical Engineering,
temperature of grain during storage with aeration[J],” vol. 39, pp. 1–11, 2022.
Transactions of the Asae, vol. 36, no. 2, pp. 509–519, 1993. [29] C. Li, R. Yu, and T. Wang, “The Research of Fault Diagnosis of
[14] S. Szke, V. Wertz, and E. Persoons, “Outdoor temperature Nuclear Power Plant Based on ELM-AdaBoost. SAMME,”
modelling for grain cooling control,” IFAC Proceedings Science and Technology of Nuclear Installations, vol. 2020,
Volumes, vol. 28, no. 6, pp. 47–51, 1995. Article ID 6689829, 9 pages, 2020.
[15] L. Fan, Y. Ji, and G. Wu, “Research on temperature prediction [30] X. Liu, H. Huang, and J. Xiang, “A personalized diagnosis
model in greenhouse based on improved SVR,” Journal of method to detect faults in gears using numerical simulation
Physics: Conference Series, vol. 1802, no. 4, Article ID 042001, and extreme learning machine,” Knowledge-Based Systems,
2021. vol. 195, Article ID 105653, 2020.
[16] Y. Wu, Q. Lan, and Y. Sun, “Application of BP Neural [31] K. L. Man and A. Bermak, “A Time Domain Differential
Network Fault Diagnosis in Solar Photovoltaic system,” in CMOS Temperature Sensor with Reduced Supply Sensitivity,”
Proceedings of the 2009 International Conference on Mecha- in Proceedings of the IEEE International Symposium on Cir-
tronics and Automation, IEEE, Changchun, China, August cuits & Systems, IEEE, Seattle, WA, USA, May 2008.
2009. [32] L. Jiao, S. Yang, and F. Liu, “Seventy years of neural network:
[17] Y. DONG, X. Li, and Cheng, “Application of BP neural review and Prospect,” Journal of Computer Science, vol. 39,
network in network fault diagnosis,” in Proceedings of the IOP no. 8, pp. 1697–1717, 2016.
Mobile Information Systems 11

[33] Z. Liu, H. Wang, and S. Peng, “Texture segmentation using

directional empirical mode decomposition,” in Proceedings of
the Image ProcessingICIP ’04. 2004 International Conference
on, IEEE, Singapore, October 2004.
[34] M. HAFIDA and R. A. I. S. Mahgoun, “Gearbox Fault Di-
agnosis Using Ensemble Empirical Mode Decomposition
(EEMD) and Residual signal,” Mechanics & Industry, vol. 13,
2012.
[35] A. A. Tabrizi, L. Garibaldi, A. Fasana, and S. Marchesielo,
“Ensemble Empirical Mode Decomposition (EEMD) and
Teager-Kaiser Energy Operator (TKEO) Based Damage
Identification of Roller Bearings Using One-Class Support
Vector Machine,” in Proceedings of the Ewshm 2014, Nantes,
France, July 2014.
[36] J. C. Nunes, Y. Bouaoune, E. Delechelle, O. Niang, and
P. Bunel, “Image analysis by bidimensional empirical mode
decomposition,” Image and Vision Computing, vol. 21, no. 12,
pp. 1019–1026, 2003.