See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/387645324

Recognizing the Type of Human Activities using Inferred Deep Learning

Networks and and Intuitive Fuzzy Sets

Article · May 2025

DOI: 10.5829/ije.2025.38.05b.20

CITATIONS READS

0 34

3 authors, including:

Reza Shakerian Meisam Yadollahzadeh-Tabari

Shahrekord University Islamic Azad University, Babol
19 PUBLICATIONS 171 CITATIONS 36 PUBLICATIONS 166 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Meisam Yadollahzadeh-Tabari on 01 January 2025.

The user has requested enhancement of the downloaded file.

 IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222

International Journal of Engineering

Journal Homepage: www.ije.ir

Recognizing the Type of Human Activities using Inferred Deep Learning Networks
and and Intuitive Fuzzy Sets
R. Shakerian, M. Yadollahzadeh-Tabari*, S. Y. Bozorgi Rad
Department of Computer Engineering, Islamic Azad University, Babol Branch, Babol, Iran

PAPER INFO A B S T R A C T

Paper history:
Received 24 June 2024 Human Activity Recognition refers to the identification and analysis of the activities carried out by an
Received in revised form 20 August 2024 individual or a group of individuals. This research introduces a highly effective system for recognizing
Accepted 30 August 2024 human activities. The system utilizes data from wearable sensors and employs advanced deep learning
algorithms. The proposed system utilizes a fusion of convolutional neural network and Long Short Term
Memory (LSTM) neural network to extract the sophisticated characteristics of the sensor data and to
Keywords: acquire knowledge of the patterns based on the temporal sequence of the data. Additionally, this article
Deep Learning Neural Network employs a Soft-max based classifier for the final connected layer of the neural network. This classifier
Human Activity Recognition utilizes fuzzy classification to assign the output of LSTM blocks to relevant activity classes. We chose
Long Short Term Memory
to utilize this classifier due to the fact that sensor data associated with human-like movements, such as
Convolutional Neural Network
walking and running, or opening and closing a door, typically exhibit significant similarities. Enhancing
the Soft-max classifier with fuzzy inference capabilities improves its accuracy in distinguishing between
closely related activities. In addition, this article introduces a post-processing module that takes into
account the classification of actions over an extended time period. Through the utilization of the
proposed fuzzy Soft-max classifier and the subsequent post-processing approach, we attained a
commendable accuracy rate of 97.03% on the PAMAP2 dataset.
doi: 10.5829/ije.2025.38.05b.20

Graphical Abstract

1. INTRODUCTION1 actions carried out by an individual or a group of

individuals. Currently, activity detection technologies are
Human Activity Recognition, often known as HAR, extensively utilized in many domains such as aged care,
refers to the process of identifying and assessing the smart cities , and security scenarios (1-3). Activity
detection approaches can be categorized into two groups:

*Corresponding Author Email: m_tabari@baboliau.ac.ir (M.

Yadollahzadeh-Tabari)

Please cite this article as: Shakerian R, Yadollahzadeh-Tabari M, Bozorgi Rad SY. Recognizing the Type of Human Activities using Inferred Deep
Learning Networks and and Intuitive Fuzzy Sets. International Journal of Engineering, Transactions B: Applications. 2025;38(05):1213-1222.
1214 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222

video-based and sensor-based, depending on the type of framework named CNN-LSTM-ELM, where ELM is
processed data (4-6). Conversely, sensor-based activity utilized in the last layer for classification. An key
detection entails the examination of unprocessed data constraint of this method is the lack of stability in its
acquired from wearable sensors (7, 8). Recently, there performance under certain conditions. Bourjandi et al.
has been a growing popularity in the usage of sensor- (17) emploied a centralized coordinate learning approach
based activity detection approaches. This is mainly due that distributes extracted features across the coordinate
to the advancements in sensor technology and the space. This approach results in a significant increase in
increasing concern for privacy among users (9-12). This the angle between features of different classes.
paper propose a deep architecture utilizing stacks of Additionally, they proposed a hybrid loss function to
convolutional neural network and Long Short Term improve the discriminative power of the learned features.
Memory (LSTM) neural network layers to extract the The experiments were conducted using the
sophisticated characteristics of the sensor data and to OPPORTUNITY and PAMAP2 datasets (14, 18-27).
acquire knowledge of the patterns based on the temporal In the domain of activity detection, numerous prior
sequence of the data. As a novelity, this paper propose a deep learning approaches have employed the Soft-max
fuzzy Soft-max based classifier for the final connected function in the final layer (also known as the fully
layer of the neural network. This classifier utilizes the connected layer) for activity classification. Nevertheless,
Intuitionistic Fuzzy Sets to assign the output of LSTM this function lacks effectiveness in accurately discerning
blocks to relevant activity classes which consederably and categorizing various types of activities. This
enhances the overal detection rate. The reminder of this limitation is especially pronounced in real-world HAR
paper is orgnized as follows. In section 2 we have a bereif applications where the boundaries between different
review of previous related works. The proposed method activities are often fuzzy rather than distinct.
with its material described in section 3. A comperhensive Consequently, there is a pressing need to explore
expriments of the proposed method elabirated in section alternative methods like Fuzzy Softmax, which can
4 and the whole paper concluded in section 5. provide a more nuanced and flexible approach to
classification by accommodating the ambiguity and
overlap inherent in activity data.
2. RELATED WORKS

Deep learning approaches, unlike machine learning 3. MATERIALS AND METHOD

methods, incorporate sophisticated and significant
features, resulting in impressive performance in the The data set utilized in this work comprises the signal
domain of activity identification. Cruciani et al. (13) acquired from the user's ambient inertial measurement
employed Convolutional Neural Networks (CNN) as a units (IMUs). The dataset is then partitioned into
technique for learning features in activity detection. The overlapping windows of a predetermined size.
researchers examined how different parameters of a Subsequently, through a series of consecutive filtering
Convolutional Neural Network (CNN) affected the and combining procedures, significant characteristics of
ability to distinguish between activities in the feature the original unprocessed data are identified.
space. Nilchi et al. (2), Ordóñez and Roggen (14) Subsequently, by implementing it on an LSTM network,
introduced novel methodologies that leverage the fusion the corresponding activity classes are ascertained for
of recurrent CNN and LSTM networks for the purpose of each time window during its activity duration. The article
activity detection using wearable sensors. Their proposed proposes a method to estimate the relevance level of the
technique exhibits a higher level of accuracy in outputs of the LSTM network using a fuzzy Soft-max
distinguishing activities compared to a typical CNN. Gil- classifier. Following that, the Soft-max fuzzy
Martín et al. (15) introduced a human activity detection classification method will be elucidated. Figure 1
method. Their proposed approach comprises three demonstrates the general architecture of the proposed
distinct parts. There are three modules involved in this method.
process. The first module involves dividing the
accelerometer data signal into overlapping windows and 3. 1. Preprocess Stage The input data for this paper
extracting information in the frequency domain. The comprises a collection of RSS readings obtained from
second module uses Convolutional Neural Networks sensors worn by an individual at specific time intervals.
(CNN) to detect the activity conducted in each window. Therefore, the input data contains a feature that is
Finally, the third module is responsible for making dependent on time. In order to enhance comprehension of
judgments at the window level. The HMM structure the data and get optimal precision in the classification
allows for integration over longer durations. As per the process, every sample of the input data is partitioned into
authors' report, they have achieved a precision level of a suitably sized overlapping window. The window size is
96.62%. Sun et al. (16), introduced a hybrid deep selected to ensure that each item within it accurately
 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222 1215

The pre-processing
step of the data
obtained from the Application of deep Post-process label of
sensors learning algorithms
predicted activities
and Fuzzy Softmax
per window
classification function
Overlapping
windows

Figure 1. General architecture of the proposed method

represents the user's perceived activity and attention. If the LSTM network consists of m blocks, the output
Figure 2 demonstrates that data segmentation is of the LSTM layer will be in the form of (ℎ1 . ℎ2 . … . ℎ𝑚 ).
accomplished by utilizing a window size of N. The LSTM layer's output will be passed to the dense
Consequently, every input record of IMU data is layer. Supposinmg 𝑍𝑡 as the raw output of the dense
partitioned into D overlapping windows with a size of N. layer, the softmax function would be as Equation 1.
The raw windowed data of the signals are entered into
𝑍𝑡=concatenated_inpu𝑡𝑡⋅𝑊+𝑏 (1)
a sequential structure of CNN blocks to extract the
desired high-level information from it. As shown in
𝑒 𝑍𝑡 (𝑖)
Figure 3, the presented CNN structure uses successive 𝑠𝑜𝑓𝑡𝑚𝑎𝑥𝑡(𝑖) = ∑ 𝑍 (𝑗) (2)
𝑗𝑒 𝑡
blocks of 1-D convolution filters and pooling Layers. The
number and size of each filter was obtained by testing
different values. Then, in order to use LSTM in the last
layer of the proposed CNN architecture, Flatten layer was
used.

3. 2. Proposed Fuzzy Soft-max Function The soft-

max function, which is a generalization of the regression
function, is extensively employed in many probabilistic
classification techniques. This function is extensively
utilized in various practical domains, including machine
learning. This function is very suitable for multi-class
classifications, as it provides the "confidence score" for
each class. The final class is determined by selecting the Figure 2. Dividing the data signal into D overlapping
item in the output list with the highest confidence score. window with size N

Figure 3. CNN structure of the proposed method including successive combination of convolution and pooling layers
1216 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222

This gives the probability distribution over the output hesitancy about an element's membership in a set. In the
classes at time 𝑡. In summary, the LSTM processes input context of soft-max layer, we can modify the fuzzy soft-
sequences to produce hidden states (ℎ𝑡(i)), which are then max as follows using Intuitionistic Fuzzy Sets:
concatenated with other features if any. The concatenated Intuitionistic_fuzzy_Soft − max(𝑖) =
input is passed through a dense layer to produce raw 𝜇(ℎ𝑡(𝑖)).𝑒 𝑍𝑡 (𝑖) (5)
output (𝑧𝑡(i)), and the soft-max activation function is ∑𝑗(𝜇(ℎ𝑡(𝑖))+𝛼⋅𝜈(ℎ𝑡(𝑗))).𝑒 𝑍𝑡 (𝑗)
applied to obtain a probability distribution over the
output classes at each time step. It is possible to Here:
incorporate fuzzy logic (28) into the soft-max layer to 1. 𝜇(ℎ𝑡(𝑖)) is the membership degree assigned to the
handle classification in a fuzzy manner. Traditional soft- i-th element of the LSTM output.
max assigns probabilities to each class in a crisp manner, 2. 𝜈(ℎ𝑡(𝑗)) is the non-membership degree assigned
meaning that each input is classified into a single class to the 𝑖-th element of the LSTM output.
with a certain degree of certainty. 3. α is a balancing parameter that determines the
Fuzzy soft-max, on the other hand, allows for a more influence of the non-membership degree.
nuanced assignment of degrees of membership to Typically, α is between 0 and 1.
multiple classes. Assuming 𝑍𝑡 (𝑖) is the output of the The values of the parameters 𝜇𝐶𝑗 (ℎ𝑖 ) and 𝜗𝐶𝑗 (ℎ𝑖 ) are
dense layer before applying soft-max, and ℎt (𝑖) is the determined based on the weight of the vector V that
output of the LSTM, you can introduce fuzziness into the connects each node of the LSTM layer to the appropriate
soft-max computation as follows: class. 𝑣𝑛𝑗 shows the weight of the edge between input
𝜇(ℎ𝑡(𝑖)).𝑒 𝑍𝑡 (𝑖) node n in the LSTM layer adjacent to the output class 𝐶𝑗 .
𝐹𝑢𝑧𝑧𝑦_𝑆𝑜𝑓𝑡𝑚𝑎𝑥𝑡(𝑖) = ∑ 𝑍 (𝑗) (3)
𝑗 𝜇(ℎ𝑡(𝑖)).𝑒 𝑡
𝑉 = (𝑣11 . 𝑣12 … . 𝑣16 … … . 𝑣𝑚𝑗 ) (6)
Here, 𝜇(ℎ𝑡(𝑖)) is a fuzzy membership function applied
to the elements of the LSTM output ℎ𝑡(i) This fuzzy 𝜇𝐶𝑗 (ℎ𝑖 ) = (𝑣𝑖𝑗 ) (7)
membership function can be designed based on the
characteristics of the LSTM output we want to capture ∑𝑚
𝑜=1.𝑜≠𝑗 𝜇𝑐𝑜 (ℎ𝑖 )
𝜗𝐶𝑗 (ℎ𝑖 ) = (8)
fuzzily. This modified soft-max function then considers (𝑚−1)

both the soft-max probabilities and the fuzzy

membership values from the LSTM output, allowing for 3. 3. Post-processing Phase In this paper's
a fuzzy inference process. architecture, an extra module is incorporated after the
Intuitionistic Fuzzy Sets (IFS), introduced by batch output to enhance the prediction outcomes. The
Atanassov (29), extend the concept of ordinary fuzzy sets objective of this module is to consolidate data from
by including both membership and non-membership previous time periods in order to arrive at a definitive
degrees for each element. This additional information conclusion. This phase can rectify errors resulting from
allows IFS to handle uncertainty and hesitancy more the misidentification of activities and enhance outcomes.
effectively. IFS are more robust in modeling real-world To identify the activity of a certain individual during the
problems where ambiguity and hesitation are inherent. testing phase, the labels associated with N consecutive
IFS includes two main components as: windows from the sequence of user behaviors are
Membership Degree (𝜇𝐴(𝑥)): Indicates the degree to inputted into the voting phase. Next, by applying a sliding
which an element 𝑥 belongs to the set 𝐴. window of size 2L+1 to the output results and identifying
It ranges between 0 and 1, similar to fuzzy sets. the activity that occurs most frequently, the ultimate
Non-Membership Degree (νA(x)): Indicates the degree activity of the individual is determined. Various tactics
to which an element x does not belong to the set A. It also might be employed for the voting procedure. For
ranges between 0 and 1. instance, the chosen window size or frequency threshold
An intuitionistic fuzzy set A in a universe of discourse X can vary, leading to varied outcomes. Provided that the
is defined as: condition is satisfied, the labels of all windows are
𝐴={⟨𝑥,𝜇𝐴(𝑥),𝜈𝐴(𝑥)⟩ ∣ 𝑥 ∈ 𝑋} where 0 ≤ modified to match the label of the activity that has won.
(4)
𝜇𝐴(𝑥)+𝜈𝐴(𝑥) ≤ 1 For instance, in Figure 5, given that L is 3 and the
threshold value is 60%. The activity labeled 2 will be the
We can explore the use of Intuitionistic Fuzzy Sets (IFS) winner for the first window, while the activity labeled 6
in the soft-max layer to incorporate both membership and
will be the winner for the second window. Based on the
non-membership degrees.
tests conducted at this stage, it has been observed that as
Intuitionistic fuzzy sets extend the concept of fuzzy the number of classes in a data set increases, the
sets by introducing an additional parameter, the non- occurrence of windows between classes becomes more
membership degree, which represents the uncertainty or common, leading to a higher risk of error. Hence, the
 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222 1217

appropriate window size should be established based on postures while utilizing the IMU. These activities
the number of classes included in the dataset being included lying, sitting, standing, walking, running,
utilized. cycling, Nordic walking, stair climbing, descending
stairs, vacuuming, ironing, and ironing. Figure 4 displays
the frequency distribution of activities in the PAMAP2
4. IMPLEMENTATION AND RESULTS dataset.

In order to assess the accuracy of the suggested approach 4. 2. Implementation Parameters In this
in forecasting user behavior, we conducted multiple research, we employed the Leave-One-Subject-Out
experimental assessments. We utilized the PAMAP2 (LOSO) cross-validation approach [15]. The training
dataset (30). Below is a concise summary of the dataset. dataset consists of data from seven volunteers, the
Subsequently, the evaluation methodology is elucidated, validation dataset consists of data from one volunteer,
followed by a comprehensive discussion of the outcomes. and the remaining data is utilized for testing. This
The experiments were conducted using Python 3 on a configuration is iteratively replicated 9 times. The
Windows 10 operating system. The system utilizes a findings reported in this study represent the mean values
quad-core processor with a clock speed of 2.30GHz and derived from the cross-validation process. The inclusion
a graphics card of GeForce GTX 950M. The neural of a validation subset, distinct from the testing subset,
networks discussed in this context are executed with the serves the objective of fine-tuning the primary
Keras package (31). parameters of the deep learning neural network. The
architecture takes over an hour to train, however we
4. 1. PAMAP2 Dataset PAMAP2 dataset consists found that the model can be run in less than 10 seconds
of data from 12 distinct forms of physical activity to determine each of the subject's activity.
conducted by 9 individuals (1 female and 8 males) with In this context, we employ a weighted F1 score that is
an average age of 27.22 ± 3.31 years and an average BMI not influenced by the distribution of classes, alongside
of 25.22 ± 2.62 kg/m2. Volunteers utilize three inertial the accuracy measure. Equations 9 and 10 define the
measurement units (IMUs) in conjunction with a heart formulation of the weighted F1 score.
rate oscilloscope for the purpose of monitoring. Contrary
to the majority of datasets concerning human activities,
PAMAP2 exhibits an even distribution of activities. lying
Three inertial measurement units, consisting of an sitting
accelerometer, gyroscope, and magnetometer, are used to standing
record data. Additionally, a heart rate monitor is linked walking
to the hand, chest, and ankle. The sensors do sampling at running
a frequency of 33 Hz. The data is gathered for a duration cycling
Nordic_walking
exceeding 10 hours, resulting in a dataset comprising 52
ascending_stairs
distinct attributes. Each inertial measurement unit
descending_stairs
encompasses 17 attributes pertaining to the individual's vacuum_cleaning
body, including temperature, gyroscope data, magnetism, ironing
and accelerometer data from the body attachment point, rope_jumping
along with a timestamp. Table 1 provides a
comprehensive summary of the properties present in the
Figure 4. Activity frequency distribution diagram in
PAMAP2 dataset. The test participants adhered to a
PAMAP2 dataset
protocol consisting of a minimum of 12 activities or

TABLE 1. An overview of some features of the PAMAP2 dataset

Activityid heartrate handTemperature handAcc16_1 handAcc16_2 handAcc16_3 handAcc6_1 handAcc6_2 handAcc6_3 handGyro1
0 -0/1 -1/2 1 0/9 0/4 1 0/9 0/4 0
0 -0/1 -1/2 1 0/8 0/4 1 0/8 0/4 -0/2
0 -0/1 -1/2 1 0/7 0/4 1 0/8 0/4 -0/3
0 -0/1 -1/2 1 0/7 0/5 1 0/7 0/4 -0/3
0 -0/1 -1/2 1 0/7 0/5 1 0/7 0/4 -0/1
1218 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222

TABLE 2. An overview of proposed Deep Learning model

weighted F1 − score ( 𝐹𝑤 ) = ∑𝑖 2 × parameters
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖 .𝑟𝑒𝑐𝑎𝑙𝑙𝑖 (9)
𝑤𝑖 Parameter Value
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖 +𝑟𝑒𝑐𝑎𝑙𝑙𝑖

𝑇𝑃
Learning Rate 10𝑒 −3
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (10)
𝑇𝑃+𝐹𝑃 Decay Rate 𝜌 = 0.9

𝑇𝑃 Dropout Rate Dropout with p=0.3

𝑟𝑒𝑐𝑎𝑙𝑙 = (11)
𝑇𝑃+𝐹𝑁
Overlapping window size 500
TP, FP, and FN represent the quantities of true Percentage of overlapping windows 50%
positives, false positives, and false negatives,
The number of hidden layer nodes 60
respectively. Equation 9 defines the parameter i as the
class index, representing the data share of the i th class. Convolution layer filter size 40
M represents the total number of samples. In order to Convolution layer pooling size 10
assess the significance of the disparity between two Batch Size 100
outcomes obtained from executing the algorithm during
the LOSO evaluation, we utilized a 95% confidence 𝛼 0.7
interval. If there was no overlap within this interval, it
indicated that the two results were considerably distinct.
The confidence interval is computed using Equation 12.
𝑚𝑒𝑡𝑟𝑖𝑐 (95%) = 𝑚𝑒𝑡𝑟𝑖𝑐 ±
𝑚𝑒𝑡𝑟𝑖𝑐×(100−𝑚𝑒𝑡𝑟𝑖𝑐) (12)
1.96√
𝑀

The amount of optimization parameters in a deep neural

network (DNN) varies depending on the type of layers it
has and significantly affects the time and computational
resources needed to train the networks. To enhance
efficiency, the data is partitioned into smaller batches of
100 units each throughout both the training and testing
phases. By employing this method, the gradient is
Figure 5. Triangle based function for both membership and
computed for the neural network parameters following non-membership value of proposed fuzzy soft-max function
every small batch.
In order to address the issue of overfitting, we employ
the Dropout approach with a probability of p = 0.3 on the to allocate each window to a certain physical activity. In
dense layers (32). In order to get optimal results, we order to enhance evaluation and showcase the
employ a greedy strategy to determine the most suitable experimental findings, we have included the
values for the sliding window size parameters, the size of conventional Soft-max classifier alongside the newly
the CNN network filter, the size of the integration proposed Fuzzy Soft-max classifier. The findings were
parameter in the CNN network, and the value of the evaluated using four established criteria: Accuracy,
parameter 𝛼 (as defined in Equation 5). Table 2 displays Precision, Recall, and F1-Score, which were previously
an overview of proposed Deep Learning model introduced in this section.
parameters. We also used Triangle based function for The findings of the suggested fuzzy Soft-max
both membership and non-membership value of classifier demonstrate a notable improvement of over 7%
proposed fuzzy softmax function in section 3, which is compared to the classic Soft-max classifier.
illustrated in Figure 5. This improvement is considered noteworthy for the
dataset under study. Furthermore, Figure 6 presents the
4. 3. Exprimental Results Following the historical accuracy and test accuracy curve for both the
determination of the optimal parameters for fine-tuning Soft-max and Fuzzy Soft-max classifiers. These graphs
the suggested deep learning structure in the preceding provide valuable insights into the performance of the
section, the effectiveness of the proposed approach is neural network architecture throughout implementation.
now assessed and examined using the PAMAP2 dataset. From the figure, it is evident that the accuracy and
The suggested deep architecture generates a vector of 12 test_accuracy charts have achieved more suitable values
points for each sliding window, with each point for the proposed classifier. Additionally, they exhibit
corresponding to a distinct action. The scores are further greater convergence capability and reduced fluctuation of
examined using the suggested Fuzzy Soft-max classifier values.
 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222 1219

(a) (b)
Figure 6. History chart of ACCURACY and TEST_ACCURACY a) using SOFT-MAX classifier b) using FUZZY SOFT-MAX
classifier

TABLE 3. Results obtained from running the algorithm on the PAMAP2 dataset
Test Accuracy Recall Precision F-measure
Classifier
Type No with No with No with No with
preprocessing preprocessing preprocessing preprocessing preprocessing preprocessing preprocessing preprocessing
Soft-max 85.33±0.3 90.11±0.5 84.9±0.3 89.63±0.5 85.6±0.3 90.45±0.5 85.24±0.3 90.03±0.5
Fuzzy Soft-
92±0.3 97.03±0.5 90.67±0.3 97.2±0.5 93.1±0.3 97.1±0.5 91.67±0.3 97.1±0.5
max

In addition, we employed a post-processing module able to accurately differentiate between various contexts
to mitigate the impact of signal noise while identifying and activities, achieving a high identification rate. By
an individual's activities during the test phase. The labels employing the conventional Soft-max classifier, it is
derived from N consecutive samples of the projected evident that there is occurrence of misidentification in
sequence of activities are aggregated by voting in the activities such as jumping rope, descending stairs,
proposed technique. Subsequently, the final forecast is ironing, and so on.
ascertained by shifting a sliding window across the However, Figure 7(b) demonstrates that the proposed
output results and conducting a majority vote. Fuzzy Soft-Max classifier exhibits a robust ability to
Ultimately, the tag with the highest occurrence rate is handle uncertainty and effectively differentiate between
chosen as the user's ultimate activity. The results indicate identical activities and situations. This classifier exhibits
that the post-processing module has substantially a commendable capacity to detect pairings of activities,
enhanced the performance of our system. The such as (cycling-running) and (walking-running).
improvement can be attributed to the diminishing impact Furthermore, it is evident that the majority of errors
of on the classifier's mis-classifications. We achieved a observed when employing the conventional Soft-Max
test data accuracy of 97.03% as a result of this. To classifier are nearly eliminated when utilizing the
conduct a thorough investigation, we compared the proposed Fuzzy Soft-Max classifier. Through the
outcomes of the Soft-max and Fuzzy Soft-max classifiers implementation of post-processing techniques, we are
for the activity classes of the PAMAP2 dataset. The able to attain a detection rate of about 100% for activities
comparison performed using the confusion matrix such as walking and cycling. Furthermore, activities such
presented in Figure 7, which for both of them the post- as jumping rope and descending stairs exhibit lower rates
processing techniques have been applied. of detection in comparison to other activities.
The postures observed include (lying, sitting, and
standing), (standing-ironing) and the activities (walking- 4. 4. Comparison with Previous Works The
climbing stairs- and descending stairs) exhibit a results of this research were compared with other state-
significantly higher rate of incorrect diagnoses when of-the-art studies that have been previously performed on
compared to one another. This is due to their similar the similar datasets. The F1 score was selected for
behavior and the near proximity of the values received comparison as a valid criterion for measurement. Table 9
from the IMUs. classifying these activities into distinct compares our work with some of state-of-the-art
groups can potentially lead to confusion for the classifier. researches mentioned in the literature.
However, the confusion matrix reveals that the system is
1220 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222

(a) (b)
Figure 7. Confusion matrix results of the proposed method for the activity classes in the PAMAP2 dataset after applying the pre-
processing step a) using the SOFT-MAX classifier b) using the FUZZY SOFT-MAX

TABLE 9. Comparison with previous works activity recognition. International Journal of Engineering,
Transactions A: Basics. 2022;35(1):130-41.
Reference Classification method F1 score
https://doi.org/10.5829/ije.2022.35.01A.12
Agac and Durmaz 2. Nilchi AN, Vafaei A, Hamidi H, editors. Evaluation of security
Deep Learning+ Attention 96%
Incel (33) and fault tolerance in mobile agents. 2008 5th IFIP International
Challa et al. (34) CNN+BI-LSTM 94.91% Conference on Wireless and Optical Communications Networks
(WOCN'08); 2008: IEEE.
Ramzani et al. (35) CNN+Hybrid loss 95.3% 3. Keramati Hatkeposhti R, Yadollahzadeh Tabari M,
Tang et al. (36) Hs-ResNet (CNN) 91.8% GolsorkhtabariAmiri M. Fall detection using deep learning
algorithms and analysis of wearable sensor data by presenting a
Proposed approach BI-LSTM 97.03% new sampling method. International Journal of Engineering,
Transactions A: Basics. 2022;35(10):1941-58.
https://doi.org/10.5829/ije.2022.35.10A.13
5. CONCLUSION 4. Tian Y, Wang X, Yang P, Wang J, Zhang J, editors. A single
accelerometer-based robust human activity recognition via
wavelet features and ensemble feature selection. 2018 24th
This paper introduces a technique for detecting human International Conference on Automation and Computing (ICAC);
activities using deep learning. The method employed 2018: IEEE.
utilizes a combination of deep convolutional networks 5. Mannini A, Sabatini AM. Machine learning methods for
and deep recurrent LSTM networks to accurately forecast classifying human physical activity from on-body accelerometers.
Sensors. 2010;10(2):1154-75.
and classify activities. In this study, a Fuzzy Soft-max https://doi.org/10.3390/s100201154
classifier is introduced to categorize the output of LSTM
6. Fan L, Poh K-L, Zhou P. A sequential feature extraction approach
blocks into the respective activity classes. Our proposed for naïve bayes classification of microarray data. Expert Systems
technique was applied on the PAMAP2 dataset using this with Applications. 2009;36(6):9919-23.
classifier. When comparing the results with the https://doi.org/10.1016/j.eswa.2009.01.075
conventional Soft-max handle function, it is evident that 7. Bustoni IA, Hidayatulloh I, Ningtyas A, Purwaningsih A, Azhari
the offered handle demonstrates a high proficiency in S, editors. Classification methods performance on human activity
recognition. Journal of Physics: Conference Series; 2020: IOP
detecting similar behaviors, such as opening and closing Publishing. 10.1088/1742-6596/1456/1/012027
the door. Our future plans involve utilizing the suggested 8. Gupta A, Semwal VB. Multiple task human gait analysis and
Fuzzy Soft-max classifier to identify more scenarios and identification: Ensemble learning approach. Emotion and
actions associated with motion prediction or fall Information Processing: A Practical Approach. 2020:185-97.
detection. https://doi.org/10.1007/978-3-030-48849-9_12
9. Yu D. Softmax function based intuitionistic fuzzy multi-criteria
decision making and applications. Operational Research.
2016;16:327-48. 10.1007/s12351-015-0196-7
6. REFERENCES
10. Siddiqi MH, Ali R, Rana MS, Hong E-K, Kim ES, Lee S. Video-
based human activity recognition using multilevel wavelet
1. Bourjandi M, Yadollahzadeh Tabari M, Golsorkhtabaramiri M. decomposition and stepwise linear discriminant analysis. Sensors.
Fuzzy centralized coordinate learning and hybrid loss for human 2014;14(4):6370-92. https://doi.org/10.3390/s140406370
 R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222 1221

11. Ahmad Z, Khan NM, editors. Multidomain multimodal fusion for 23. Deschrijver G, Kerre EE. On the position of intuitionistic fuzzy
human action recognition using inertial sensors. 2019 IEEE Fifth set theory in the framework of theories modelling imprecision.
International Conference on Multimedia Big Data (BigMM); Information Sciences. 2007;177(8):1860-6.
2019: IEEE. https://doi.org/10.1016/j.ins.2006.11.005
12. Mobark M, Chuprat S, Mantoro T, editors. Improving the 24. Grzegorzewski P, Mrówka E. Trapezoidal approximations of
accuracy of complex activities recognition using accelerometer- fuzzy numbers. Fuzzy sets and systems. 2005;153(1):115-35.
embedded mobile phone classifiers. 2017 Second International https://doi.org/10.1016/j.fss.2004.02.015
Conference on Informatics and Computing (ICIC); 2017: IEEE. 25. Naim S, Hagras H. A type 2-hesitation fuzzy logic based multi-
13. Cruciani F, Vafeiadis A, Nugent C, Cleland I, McCullagh P, Votis criteria group decision making system for intelligent shared
K, et al. Feature learning for human activity recognition using environments. Soft Computing. 2014;18:1305-19.
convolutional neural networks: A case study for inertial 26. Verma R, Sharma BD. New operations over hesitant fuzzy sets.
measurement unit and audio data. CCF Transactions on Pervasive Fuzzy Information and Engineering. 2013;5:129-46.
Computing and Interaction. 2020;2(1):18-32. https://doi.org/10.1007/s12543-013-0137-1
14. Ordóñez FJ, Roggen D. Deep convolutional and lstm recurrent 27. Xu Z, Yager RR. Some geometric aggregation operators based on
neural networks for multimodal wearable activity recognition. intuitionistic fuzzy sets. International journal of general systems.
Sensors. 2016;16(1):115. https://doi.org/10.3390/s16010115 2006;35(4):417-33.
15. Gil-Martín M, San-Segundo R, Fernandez-Martinez F, Ferreiros- https://doi.org/10.1080/03081070600574353
López J. Improving physical activity recognition using a new 28. Hamidi H. A combined fuzzy method for evaluating criteria in
deep learning architecture and post-processing techniques. enterprise resource planning implementation. Intelligent systems:
Engineering Applications of Artificial Intelligence. Concepts, methodologies, tools, and applications: IGI Global;
2020;92:103679. 2018. p. 639-70.
https://doi.org/10.1016/j.engappai.2020.103679
29. Atanassov KT, Atanassov KT. Intuitionistic fuzzy sets: Springer;
16. Sun J, Fu Y, Li S, He J, Xu C, Tan L. Sequential human activity 1999.
recognition based on deep convolutional network and extreme
learning machine using wearable sensors. Journal of Sensors. 30. Lara OD, Labrador MA. A survey on human activity recognition
2018;2018(1):8580959. https://doi.org/10.1155/2018/8580959 using wearable sensors. IEEE communications surveys &
tutorials. 2012;15(3):1192-209.
17. Bourjandi M, Yadollahzadeh‐Tabari M, Golsorkhtabaramiri M. https://doi.org/10.1109/SURV.2012.110112.00192
Combined deep centralized coordinate learning and hybrid loss
for human activity recognition. Concurrency and Computation: 31. Brownlee J. Deep learning with Python: develop deep learning
Practice and Experience. 2022;34(22):e6870. models on Theano and TensorFlow using Keras: Machine
Learning Mastery; 2016.
18. Reiss A, Stricker D, editors. Creating and benchmarking a new
dataset for physical activity monitoring. Proceedings of the 5th 32. LeCun Y, Bengio Y, Hinton G. Deep learning. nature.
international conference on pervasive technologies related to 2015;521(7553):436-44.
assistive environments; 2012. 33. Agac S, Durmaz Incel O. On the use of a convolutional block
19. Stančić I, Musić J, Grujić T. Gesture recognition system for real- attention module in deep learning-based human activity
time mobile robot control based on inertial sensors and motion recognition with motion sensors. Diagnostics. 2023;13(11):1861.
strings. Engineering Applications of Artificial Intelligence. 34. Challa SK, Kumar A, Semwal VB, Dua N. An optimized deep
2017;66:33-48. https://doi.org/10.1016/j.engappai.2017.08.013 learning model for human activity recognition using inertial
20. Gers FA, Schraudolph NN, Schmidhuber J. Learning precise measurement units. Expert Systems. 2023;40(10):e13457.
timing with LSTM recurrent networks. Journal of machine 35. Ramzani E, Yadollahzadeh‐Tabari M, GolesorkhtabarAmiri M,
learning research. 2002;3(Aug):115-43. Pouyan AA. Diagnosing of Parkinson's disease based on hand
21. Zeng M, Nguyen LT, Yu B, Mengshoel OJ, Zhu J, Wu P, et al., drawing analysis using Bi‐Directional LSTM equipped with
editors. Convolutional neural networks for human activity fuzzy inferential soft‐max classifier. International Journal of
recognition using mobile sensors. 6th international conference on Imaging Systems and Technology. 2024;34(1):e22948.
mobile computing, applications and services; 2014: IEEE. 36. Tang Y, Zhang L, Min F, He J. Multiscale deep feature learning
22. Feizi A. Convolutional gating network for object tracking. for human activity recognition using wearable sensors. IEEE
International Journal of Engineering, Transactions A: Basics. Transactions on Industrial Electronics. 2022;70(2):2106-16.
2019;32(7):931-9. https://doi.org/10.5829/ije.2019.32.07a.05
 ‫‪1222‬‬ ‫‪R. Shakerian et al. / IJE TRANSACTIONS B: Applications Vol. 38 No. 05, (May 2025) 1213-1222‬‬

‫‪COPYRIGHTS‬‬
‫‪©2025 The author(s). This is an open access article distributed under the terms of the Creative Commons‬‬
‫‪Attribution (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, as long‬‬
‫‪as the original authors and source are cited. No permission is required from the authors or the publishers .‬‬

‫‪Persian Abstract‬‬
‫چکیده‬
‫شناسایی فعالیت های انسانی به شناسایی و تجزیه و تحلیل فعالیت های انجام شده توسط یک فرد یا گروهی از افراد اشاره دارد‪ .‬این تحقیق یک سیستم بسیار موثر برای شناخت‬
‫فعالیت های انسانی معرفی می کند‪ .‬این سیستم از داده های حسگرهای پوشیدنی و از الگوریتم های پیشرفته یادگیری عمیق استفاده می کند‪ .‬سیستم پیشنهادی از تلفیقی از شبکه‬
‫عصبی کانولوشن و شبکه عصبی حافظه کوتاه مدت بلند مدت (‪ )LSTM‬برای استخراج ویژگیهای پیچیده دادههای حسگر و کسب دانش از الگوها بر اساس توالی زمانی‬
‫دادهها استفاده میکند‪ .‬عالوه بر این‪ ،‬این مقاله از یک طبقهبندی کننده مبتنی بر ‪ Soft-max‬برای الیه متصل نهایی شبکه عصبی استفاده میکند‪ .‬این طبقهبندیکننده از طبقهبندی‬
‫فازی برای تخصیص خروجی بلوکهای ‪ LSTM‬به کالسهای فعالیت مربوطه استفاده میکند‪ .‬ما استفاده از این طبقهبندیکننده را انتخاب کردیم زیرا دادههای حسگر مرتبط‬
‫با حرکات انسان‪ ،‬مانند‪ ،‬راه رفتن و دویدن‪ ،‬یا باز و بسته کردن یک در‪ ،‬معموالً شباهتهای قابلتوجهی را نشان میدهند‪ .‬تقویت طبقهبندیکننده ‪ Soft-max‬با قابلیتهای‬
‫استنتاج فازی‪ ،‬دقت آن را در تمایز بین فعالیتهای مرتبط نزدیک بهبود میبخشد‪ .‬عالوه بر این‪ ،‬این مقاله یک ماژول پس پردازش را معرفی می کند که طبقه بندی اقدامات را‬
‫در یک دوره زمانی طوالنی در نظر می گیرد‪ .‬از طریق استفاده از طبقهبندیکننده فازی ‪ Soft-max‬از قبل و رویکرد پس از پردازش متعاقب‪ ،‬ما به نرخ دقت قابل ستایش ‪97.03‬‬
‫درصد در مجموعه داده ‪ PAMAP2‬دست یافتیم‪.‬‬

‫‪View publication stats‬‬