IoT-assisted Human Activity Recognition Using Bat Optimization Algorithm with Ensemble Voting Classifier for Disabled Persons

Abstract

Internet of Things (IoT)-based human action recognition (HAR) has made a significant contribution to scientific studies. Furthermore, hand gesture recognition is a subsection of HAR, and plays a vital role in interacting with deaf people. It is the automatic detection of the actions of one or many subjects using a series of observations. Convolutional neural network structures are often utilized for finding human activities. With this intention, this study presents a new bat optimization algorithm with an ensemble voting classifier for human activity recognition (BOA-EVCHAR) technique to help disabled persons in the IoT environment. The BOA-EVCHAR technique makes use of the ensemble classification concept to recognize human activities proficiently in the IoT environment. In the presented BOA-EVCHAR approach, data preprocessing is generally achieved at the beginning level. For the identification and classification of human activities, an ensemble of two classifiers namely long short-term memory (LSTM) and deep belief network (DBN) models is utilized. Finally, the BOA is used to optimally select the hyperparameter values of the LSTM and DBN models. To elicit the enhanced performances of the BOA-EVCHAR technique, a series of experimentation analyses were performed. The extensive results of the BOA-EVCHAR technique show a superior value of 99.31% on the HAR process.

Main article text

INTRODUCTION

As older people suffer from age-related diseases or may have disorders of body functions, the necessity for smart health assistance structures increases every year ( Gupta et al., 2022). Physical observation is a common technique of observing geriatric patients that can be expensive, necessitates a large number of staff members, and is not possible in light of the huge population aging in the future ( Xu et al., 2023). Many ambient assisted living (AAL) applications like assistive human–computer interaction technologies, care-providing robots, and video surveillance systems need human activity detection ( Ullah et al., 2019). Although the main operators of the AAL system are the old people, the idea can also be implemented for physically and mentally impaired people in addition to individuals suffering from obesity and diabetes, those who require support at home, and persons of any age engrossed in personal fitness monitoring ( Li et al., 2021). Accordingly, the sensor-related realistic monitoring scheme for supporting independent living at home has been the topic of several current research studies in the human action recognition (HAR) field ( Rashid and Louis, 2019). Activity detection is described as the process of inferring sensor datasets for classifying a set of human actions. HAR refers to a speedily developing region of study that can offer valuable data on the fitness, health, and well-being of monitored people outside hospital backgrounds ( Zhang et al., 2020b). Daily action detection using wearable technology plays a vital role in the domain of ubiquitous healthcare.

The HAR problem, concerning machine learning, is considered a classifier task ( Brishtel et al., 2023), where the task of the method is to relate all input dataset points to sort out a predefined number of categories (e.g. human actions) ( Tang et al., 2021). Classification refers to a supervised task that the methods learn in the training phase by linking the input dataset to its respective ground truth label that denotes the activity that was happening when the data was recorded. While the HAR issue tries to ascertain a perfect segmentation of actions ( Li et al., 2019), this task can be hard to attain as humans easily do actions and the segmentation of time series into intervals that denote single actions is not direct. Even semantically, it is hard to describe where one action ends and where the next commences ( Lattanzi et al., 2022). Hence, much work in HAR can be devoted to segregating the measurements into fixed intervals and detecting the action that took place within each interval.

This study presents a new bat optimization algorithm with an ensemble voting classifier for human activity recognition (BOA-EVCHAR) technique to help disabled persons in the Internet of Things (IoT) environment. The BOA-EVCHAR technique makes use of the ensemble classification concept to recognize human activities proficiently in the IoT environment. In the presented BOA-EVCHAR approach, data preprocessing can be generally performed at the beginning level. For the identification and classification of human activities, an ensemble of two classifiers namely long short-term memory (LSTM) and deep belief network (DBN) models are utilized. Finally, BOA is used to optimally select the hyperparameter values of the LSTM and DBN models. To elicit the enhanced performances of the BOA-EVCHAR technique, a series of experimentation analyses were performed. The key contributions of the paper are given as follows:

• Develop a BOA-EVCHAR technique, which combines the BOA with an ensemble voting classifier to address HAR challenges in IoT environments, demonstrating innovation in algorithm design.

• Utilizing the ensemble classification concept, the BOA-EVCHAR technique enhances the accuracy and robustness of human activity recognition by aggregating the outputs of two distinct classifiers, namely LSTM and DBN models, thereby improving the classification performance.

• Employing the BOA to optimize hyperparameter values for both the LSTM and DBN models, resulting in improved model performance and adaptability to different activity recognition scenarios.

The remaining sections of the article are arranged as follows: the Related Studies section 2 offers the literature review and the next section presents the proposed method. Then, the Experimental Evaluation section elaborates on the evaluation of results and the Conclusion section summarizes the work.

RELATED STUDIES

Soni et al. (2023) proposed a DNN consisting of two complex layers and a BiGRU model. The model utilized can routinely detect and extract activity factors by employing some of the construction parameters. The raw information attained by employing mobile phone sensors is transferred to a two-layer BiGRU and later to a two-layer convolutional neural network (CNN). In Basly et al. (2022), a deep temporal residual scheme for day-to-day life activity detection that targets improving spatiotemporal feature depiction to enhance HAR achievement is presented. Also, a deep RCN is implemented for retaining discriminatory visual aspects transmitted to appearance and LSTM-NN for capturing the long-term temporal evolution of activities. The LSTM is regarded to use time dependencies that arise during the action to improve feature extraction from the RCN network by supplementing time data for addressing the dynamic HAR issue as a serial label work. Pesenti et al. (2023) suggested a DL-based method utilizing inertial sensors for giving industrial exoskeletons with HAR and adaptive payload compensation. This model also utilized LSTM networks for accomplishing HAR and for the classification of the weight of objects lifted up to 15 kg.

Islam et al. (2023) presented a new DL-established method called STC-NLSTMNet, with the capacity for extracting spatial and temporal aspects simultaneously and routinely detecting human actions with greater precision. This model majorly consists of DS-Conv blocks, NLSTM, and FAM. In Zhang et al. (2020a), a plain but efficient SGN for a skeleton-based activity detection model is suggested. Also, the model utilized the high-level semantics of joints in the networking for improving the factor depiction capacity. Mekruksavanich and Jitpattanakul (2021) proposed the generic HAR technique for mobile phone sensor data established on LSTM networks for time-sequence domains. Four baseline LSTM networks are relatively investigated for examining the effect of employing diverse types of mobile phone sensor data. Additionally, a fusion LSTM network known as a 4-layer CNN-LSTM is presented for improving the detection accomplishment.

Hamad et al. (2021) suggested a networking construction established on dilated causal convolution and multi-head self-attention systems that completely distribute frequent constructions for making effective calculations and to preserve the time step order. The presented model is investigated for human actions by employing smart home binary sensor information and embeddable sensor data. In Jugunta et al. (2023), an innovative approach is introduced for optimal fetal health classification, integrating the bat algorithm with a hybrid XGB-RNN model to dynamically adapt to the complexities of fetal health data, leveraging XGB for feature selection and boosting and RNN for capturing temporal dependencies. Tan et al. (2022) suggest an ensemble learning algorithm for smartphone sensor-based activity recognition, integrating a stacked GRU-CNN and a DNN with an extra feature vector, enhancing classification through model fusion.

Researchers have proposed diverse deep-learning models for activity recognition. Soni et al. employed a DNN with two complex layers and a BiGRU model for routine activity detection from mobile phone sensor data. Pesenti et al. utilized LSTM networks in a DL-based method for industrial exoskeletons, addressing both HAR and adaptive payload compensation. Islam et al. presented STC-NLSTMNet, a DL-based method combining DS-Conv blocks, NLSTM, and FAM for precise human action detection. Mekruksavanich and Jitpattanakul introduced a generic HAR technique based on LSTM networks, exploring diverse sensor data types, while Hamad et al. proposed a networking construction using dilated causal convolution and multi-head self-attention for effective calculations in HAR from smart home binary sensor data.

THE PROPOSED MODEL

In this study, we have concentrated on the development and design of the BOA-EVCHAR technique for the recognition of human activities to help disabled persons in the IoT environment. The BOA-EVCHAR technique makes use of the ensemble classification concept to recognize human activities proficiently in the IoT environment. The BOA-EVCHAR technique comprises a three-stage process, namely, data preprocessing, ensemble classification, and hyperparameter adjustment using BOA. Figure 1 illustrates the overall flow of the BOA-EVCHAR method.

Figure 1:

Overall flow of the BOA-EVCHAR approach. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; LSTM, long short-term memory.

Data preprocessing

At the beginning level, data preprocessing can be generally performed. The normalized has been utilized for turning the observed data into values between 0 and 1 across the research time ( Sorguli and Rjoub, 2023). A linear scaling normalized was determined using Equation (1).

(1) $${y^{\prime }}_i=\frac{y_i-y_{\text{min}\hspace{0.17em}}}{y_{\text{max}\hspace{0.17em}}-\hspace{0.17em}{y}_{\text{min}\hspace{0.17em}}}$$

where y _min is the input database’s actual value; y _Yi is the normalized value scaled based on the range; and max and min refer to the maximal and minimal values of the features.

Ensemble classification

For the identification and classification of human activities, an ensemble of two classifiers, namely, LSTM and DBN, is utilized. LSTM is a popular recursive neural network that has greater success in different fields, namely, emotion analysis, image processing, and speech detection ( Khataei Maragheh et al., 2022). LSTM has the advantage of low complexity, nonlinear predictability, and faster convergence. The memory cell is the core of LSTM that includes a single gateway system and parameters and is used for making useless or valuable data.

There are three gates in all the cells: forget, input, and output gates. At all the time steps, LSTM takes the present input ( x _t ) and the previous memory cell state ( c _{t −1}) as input and evaluates the output ( o _t ), and the present cell state ( c _t ) and σ indicates the sigmoid function $\left(\sigma \mathrm{(}X\mathrm{)}=\frac{1}{1+e^{-x}}\right)$ .

(2) $$i_t={\sigma }_i(x_t{W}_{xi}+h_{t-1}{W}_{hi}+b_i)$$

(3) $$f_t={\sigma }_f(x_t{W}_{xf}+h_{t-1}{W}_{hf}+b_f)$$

(4) $$o_t={\sigma }_0(x_t{W}_{xo}+h_{t-1}{W}_{ho}+b_0)$$

(5) $$c_t=f_t\odot c_{t-1}+i_t\odot \left(tanh\right(x_t{W}_{xc}+h_{t-1}{W}_{hc}+b))$$

(6) $$h_t=o_t\odot tanh(c_t),$$

where the parameter is determined by: i _t input gate activation vector, x _t input vector, o _t output gate activation vector, f _t forget gate activation vector, h _t hidden layer (HL) output vector, $h_{t-1}\in R^{d_h}$ and d _h HL dimensions, and d input vector dimensions; c _t shows the cell state and c _t−1 denotes the vector value of the memory unit at time t−1. 𝒲 and b parameters denote the weight matrix and bias.

DBN is a generative method with many layers of hidden variables ( Sun, 2023). However, it can be feasible to construct DBNs with comparatively sparse connection, over all models, all the units from all the layers being linked to all the units from all the neighboring layers, and there is no connection within a layer. The DBN is generated by consecutively stacking many RBMs. The learning procedure was separated into two phases, i.e. the RBM was primarily pretrained layer wise and unsupervised; afterward all the networks were supervised by the back-propagation system. To provide the parameter model θ = ( w ^R , b _v , b _h ), the joint probability distribution P( v, h; θ) of the visible layer (VL) and HL is determined using the energy function E( v, h; θ) as follows:

(7) $$\mathrm{(}\nu ,\hspace{0.17em}h;\theta \mathrm{)}=\frac{1}{Z}{e}^{-E\left(v,h;\theta \right)}.$$

Among them, $Z={\displaystyle {\sum }_{v,h}{e}^{-E\left(v,h;\theta \right)}}$ signifies the normalized factor and marginal distribution of the method on v as follows:

(8) $$\mathrm{(}\nu ;\theta \mathrm{)}=\frac{1}{Z}{\displaystyle \sum _h{e}^{-E\left(v,h;\theta \right)}}.$$

For a Bernoulli (VL) distribution–Bernoulli (HL) dispersed RBM, the energy function was expressed as follows:

(9) $$E\mathrm{(}\nu ,\hspace{0.17em}h;\theta \mathrm{)}=-{\displaystyle \sum _{i=1}^m{b}_{vi}{\nu }_i}-{\displaystyle \sum _{j=1}^n{b}_{hj}{h}_j}-{\displaystyle \sum _{i=1}^m{\displaystyle \sum _{j=1}^n{\nu }_i{w}_{ij}^R{h}_j.}}\hspace{0.17em}$$

Particularly, $w_{ij}^R$ signifies the connection weight of RBM, and b _vi and b _hi stand for bias of VL and HL nodes, correspondingly. Afterwards, the conditional probability distribution is determined as follows:

(10) $$\left(h_j=\frac{1}{\nu },\hspace{0.17em}\theta \right)=\sigma \left(b_{hj}+{\displaystyle \sum _{i=1}^m{\nu }_i{w}_{ij}^R}\right),\hspace{0.17em}$$

where σ denotes the sigmoid function.

(11) $$\left({\nu }_i=\frac{1}{h},\hspace{0.17em}\theta \right)=\sigma \left(b_{vi}+{\displaystyle \sum _{j=1}^n{w}_{ij}^R{h}_j}\right).\hspace{0.17em}$$

With computing gradient of log-likelihood function log P( v;θ), the RBM weighted upgrade equation is attained as follows:

(12) $$w_{ij}^R\mathrm{(}\tau +1\mathrm{)}=w_{ij}^R\mathrm{(}\tau \mathrm{)}+\eta \Delta w_{ij}^R$$

$$\Delta w_{ij}^R=E_{\text{data}}\mathrm{(}{\nu }_i{h}_j\mathrm{)}-E_{\text{model}}\mathrm{(}{\nu }_i{h}_j\mathrm{)}.$$

In this equation, τ and η stand for the iteration counts and the rate of learning of the RBM, correspondingly, and E _data( v _ih _j ) and E _model( v _ihi _j ) imply the probability of observed data from the trained set and the expectation on dissemination defined by the method, correspondingly.

Hyperparameter tuning using BOA

In this work, BOA is used to optimally select the hyperparameter values of the LSTM and DBN models. BOA is generated utilizing the main model of frequency tuning dependent upon microbat echolocation ( Gaber et al., 2023). The echolocation features of microbats are idealized as the subsequent three principles in the standard bat algorithm. The simulated bats need the subsequent initialized parameters: the d-dimension searching space, position x _i , velocity v _i , and frequency f ₁. The following are upgraded principles for novel solution $x_j^i$ and velocity $v_j^i$ in all the steps t:

(13) $$f_i=f_{\text{min}}+(f_{\text{min}}-f_{\text{max}})\beta ,\hspace{0.17em}$$

(14) $$v_i^j=v_i^j\mathrm{(}t-1\mathrm{)}+\left[\widehat{x^j}-x_i^j\mathrm{(}t-1\mathrm{)}\right]f_i,$$

(15) $$x_i^j\mathrm{(}t\mathrm{)}=x_i^j\mathrm{(}t-1\mathrm{)}+v_i^j,$$

where β∈[0,1] represents the uniformly distributed arbitrary vector. We realize that the variable f _i was exploited for changing the velocity and that variable $x_i^j\mathrm{(}t\mathrm{)}$ defines the value of position j for bat i at step t dependent upon Equations (13), (14), and (15). The variable x ^j implies the present global optimum position that is defined by relating each m bat’s answers. Figure 2 depicts the steps involved in BOA.

Figure 2:

Steps involved in BOA.

Song and Gorla utilized a random walk system for all the bats to prevent them from falling as local extremum and to boost their arbitrary searching capability. Next, for the choice of solutions in the present optimum position, the random walk was utilized for generating a novel solution to all the bats.

(16) $$x_{\text{new}}=x_{\text{old}}+\epsilon \overline{A}(t),$$

where ∈ε[−1,1] refers to the arbitrary number which controls the walk direction and stride, and $\overline{A}(t)$ denotes the average volume of every bat from step t.

Besides, along with Equation (17), the loudness A _i and the rate of pulse r _i can be upgraded for all the steps in Equation (15). If the prey is exposed, the loudness A _i is normally decreased and the rate of pulse r _i is enhanced.

(17) $$A_i(t+1)=\alpha A_i(t),$$

(18) $$r_i(t+1)=r_i(0)\left[1-\text{exp(}-\gamma t\text{)}\right],\hspace{0.17em}$$

The loudness A _i (0) and the rate of pulse r _i (0) are generally selected arbitrarily in the first phase of the bat algorithm. Generally, A _i (0) ∈ [1,2] and r _i (0) ∈ [0,1] are set.

The BOA system has developed a fitness function for accomplishing better classifier efficiency. It solves a positive value for signifying the candidate solution’s best efficiency. Here, the minimized classifier error rate was supposed to be a fitness function, as defined in Equation (19).

(19) $$\begin{array}{l}fitness(x_i)=ClassifierErrorRate(x_i)=\\ \hspace{1em}\hspace{1em}\frac{Number\hspace{0.17em}of\hspace{0.17em}misclassified\hspace{0.17em}samples}{Total\hspace{0.17em}number\hspace{0.17em}of\hspace{0.17em}samples}*100\end{array}$$

EXPERIMENTAL EVALUATION

The HAR results of the BOA-EVCHAR technique are assessed on the HAR dataset ( HAR dataset, n.d.) comprising 10,299 samples with 6 class labels as mentioned in Table 1.

Table 1:

Details of the database.

Class	No. of samples
Sitting	1777
Standing	1906
Lying	1944
Walking	1722
Walking upstairs	1544
Walking downstairs	1406
Total number of samples	10,299

In Table 2, the HAR results of the BOA-EVCHAR approach under 80:20 of TRP/TSP are given. Figure 3 illustrates the HAR outcomes of the BOA-EVCHAR technique on 80% of TRP. In the sitting class, the BOA-EVCHAR technique attains accu _y , prec _n , spec _y , F _score, and MCC of 99.24, 97.68, 99.54, 97.70, and 97.24%, respectively. Also, in the standing class, the BOA-EVCHAR method attains accu _y , prec _n , spec _y , F _score, and MCC of 99.27, 98.16, 99.58, 98.03, and 97.58%, correspondingly. In addition, in the lying class, the BOA-EVCHAR algorithm attains accu _y , prec _n , spec _y , F _score, and MCC of 98.97, 98.14, 99.58, 97.21, and 96.58%, correspondingly. Next, in the walking class, the BOA-EVCHAR approach attains accu _y , prec _n , spec _y , F _score, and MCC of 99.37, 97.77, 99.53, 98.18, and 97.80%, correspondingly. Finally, in the walking upstairs class, the BOA-EVCHAR method attains accu _y , prec _n , spec _y , F _score, and MCC of 99.09, 97.61, 99.59, 96.93, and 96.40%, correspondingly.

Table 2:

HAR analysis of the BOA-EVCHAR method on 80:20 of TRP/TSP.

Class	Accuracy	Precision	Specificity	F score	MCC
Training phase (80%)
Sitting	99.25	97.22	99.41	97.84	97.38
Standing	99.27	98.16	99.58	98.03	97.58
Lying	98.97	98.14	99.58	97.21	96.58
Walking	99.37	97.77	99.53	98.18	97.80
Walking upstairs	99.09	97.61	99.59	96.93	96.40
Walking downstairs	99.47	97.16	99.55	98.03	97.72
Average	99.24	97.68	99.54	97.70	97.24
Testing phase (20%)
Sitting	99.47	97.77	99.53	98.46	98.14
Standing	99.22	97.42	99.40	97.93	97.45
Lying	98.83	97.75	99.46	97.02	96.30
Walking	99.32	98.98	99.83	97.64	97.26
Walking upstairs	98.79	96.46	99.37	96.00	95.29
Walking downstairs	99.22	96.12	99.32	97.38	96.93
Average	99.14	97.42	99.48	97.40	96.89

Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

Figure 3:

HAR outcome of the BOA-EVCHAR approach on 80% of TRP. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

Figure 4 shows the HAR outcomes of the BOA-EVCHAR method on 20% of TSP. In the sitting class, the BOA-EVCHAR technique gains accu _y , prec _n , spec _y , F _score, and MCC of 99.47, 97.77, 99.53, 98.46, and 98.14%, correspondingly. Also, in the standing class, the BOA-EVCHAR approach attains accu _y , prec _n , spec _y , F _score, and MCC of 99.22, 97.42, 99.40, 97.93, and 97.45%, correspondingly. Moreover, in the lying class, the BOA-EVCHAR method achieves accu _y , prec _n , spec _y , F _score, and MCC of 98.83, 97.75, 99.46, 97.02, and 96.30%, correspondingly. Next, in the walking class, the BOA-EVCHAR method reaches accu _y , prec _n , spec _y , F _score, and MCC of 99.32, 98.98, 99.83, 97.64, and 97.26%, correspondingly. Lastly, in the walking upstairs class, the BOA-EVCHAR method gains accu _y , prec _n , spec _y , F _score, and MCC of 98.79, 96.46, 99.37, 96, and 95.29%, correspondingly.

Figure 4:

HAR outcome of the BOA-EVCHAR approach on 20% of TSP. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

In Table 3, the HAR outcomes of the BOA-EVCHAR method under 70:30 of TRP/TSP are given. Figure 5 illustrates the HAR outcomes of the BOA-EVCHAR technique on 70% of TRP. In the sitting class, the BOA-EVCHAR technique attains accu _y , prec _n , spec _y , F _score, and MCC of 99.14, 98.06, 99.60, 97.51, and 96.99%, respectively. Also, in the standing class, the BOA-EVCHAR technique gains accu _y , prec _n , spec _y , F _score, and MCC of 99.13, 97.64, 99.47, 97.60, and 97.07%, respectively. Moreover, in the lying class, the BOA-EVCHAR technique reaches accu _y , prec _n , spec _y , F _score, and MCC of 99.14, 97.50, 99.42, 97.71, and 97.18%, correspondingly. Then, in the walking class, the BOA-EVCHAR technique attains accu _y , prec _n , spec _y , F _score, and MCC of 99.11, 97.90, 99.58, 97.33, and 96.80%, respectively. Eventually, in the walking upstairs class, the BOA-EVCHAR technique obtained accu _y , prec _n , spec _y , F _score, and MCC of 99.42, 97.30, 99.51, 98.10, and 97.76%, respectively.

Table 3:

HAR analysis of the BOA-EVCHAR method on 70:30 of TRP/TSP.

Class	Accuracy	Precision	Specificity	F score	MCC
Training phase (70%)
Sitting	99.14	98.06	99.60	97.51	96.99
Standing	99.13	97.64	99.47	97.60	97.07
Lying	99.14	97.50	99.42	97.71	97.18
Walking	99.11	97.90	99.58	97.33	96.80
Walking upstairs	99.42	97.30	99.51	98.10	97.76
Walking downstairs	99.43	97.69	99.63	97.94	97.61
Average	99.23	97.68	99.54	97.70	97.23
Testing phase (30%)
Sitting	99.51	99.04	99.80	98.57	98.28
Standing	99.29	97.18	99.32	98.16	97.72
Lying	99.06	97.79	99.48	97.54	96.96
Walking	99.16	97.30	99.46	97.49	96.99
Walking upstairs	99.35	97.99	99.66	97.77	97.39
Walking downstairs	99.48	98.54	99.78	98.06	97.76
Average	99.31	97.97	99.58	97.93	97.52

Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

Figure 5:

HAR analysis of the BOA-EVCHAR approach on 70% of TRP. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

Figure 6 portrays the HAR analysis of the BOA-EVCHAR technique on 30% of TSP. In the sitting class, the BOA-EVCHAR technique gains accu _y , prec _n , spec _y , F _score, and MCC of 99.51, 99.04, 99.80, 98.57, and 98.28%, correspondingly. Also, in the standing class, the BOA-EVCHAR technique reaches accu _y , prec _n , spec _y , F _score, and MCC of 99.29, 97.18, 99.32, 98.16, and 97.72%, correspondingly. In addition, in the lying class, the BOA-EVCHAR technique acquires accu _y , prec _n , spec _y , F _score, and MCC of 99.06, 97.79, 99.48, 97.54, and 96.96%, correspondingly. Then, in the walking class, the BOA-EVCHAR approach attains accu _y , prec _n , spec _y , F _score, and MCC of 99.16, 97.30, 99.46, 97.49, and 96.99%, correspondingly. Eventually, in the walking upstairs class, the BOA-EVCHAR technique attains accu _y , prec _n , spec _y , F _score, and MCC of 99.35, 97.99, 99.66, 97.77, and 97.39%, correspondingly.

Figure 6:

HAR outcome of the BOA-EVCHAR approach on 30% of TSP. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition.

Figure 7 inspects the accuracy of the BOA-EVCHAR method in the training and validation on 70:30 of TRP/TSP. The result specified that the BOA-EVCHAR technique reaches greater accuracy values over higher epochs. Also, the higher validation accuracy over training accuracy displays that the BOA-EVCHAR algorithm learns productively on 70:30 of TRP/TSP.

Figure 7:

Accuracy curve of the BOA-EVCHAR method on 70:30 of TRP/TSP. Abbreviation: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition.

The loss analysis of the BOA-EVCHAR approach in training and validation is given on 70:30 of TRP/TSP in Figure 8. The figure indicates that the BOA-EVCHAR method reaches adjacent values of training and validation loss. The BOA-EVCHAR technique learns efficiently on 70:30 of TRP/TSP.

Figure 8:

Loss curve of the BOA-EVCHAR approach on 70:30 of TRP/TSP. Abbreviation: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition.

The comparison results of the BOA-EVCHAR technique on HAR are reported in Table 4 and Figure 9 ( Duhayyim, 2023). The results illustrate that the BOA-EVCHAR technique exhibits effective performance under all measures. Based on accu _y , an improving accu _y of 99.31% is reached by the BOA-EVCHAR technique whereas the IPODTL-HAR, RF, NNN, SVM, ANN, and LSTM models have yielded decreasing accu _y of 99.05, 86.18, 87.50, 88.81, 91.83, and 93.97%, respectively. Next, based on prec _n , an improving prec _n of 97.97% is reached by the BOA-EVCHAR method whereas the IPODTL-HAR, RF, NNN, SVM, ANN, and LSTM approaches have yielded decreasing prec _n of 97.04, 82.70, 85.86, 88.86, 88.56, and 91.82%, correspondingly. Also, based on spec _y , an improving spec _y of 99.58% is reached by the BOA-EVCHAR method whereas the IPODTL-HAR, RF, NNN, SVM, ANN, and LSTM models have shown decreasing spec _y of 98.49, 80.96, 82.76, 87.44, 92.20, and 94%, correspondingly. Finally, based on the F _score, an improving F _score of 97.93% is reached by the BOA-EVCHAR method whereas the IPODTL-HAR, RF, NNN, SVM, ANN, and LSTM models have exhibited decreasing F _score of 96.39, 80.94, 83.06, 88.80, 90.85, and 92.50%, correspondingly.

Table 4:

Comparative outcome of the BOA-EVCHAR approach with other systems.

Methodology	Accuracy	Precision	Specificity	F score
BOA-EVCHAR	99.31	97.97	99.58	97.93
IPODTL-HAR	99.05	97.04	98.49	96.39
RF algorithm	86.18	82.70	80.96	80.94
NNN algorithm	87.50	85.86	82.76	83.06
SVM algorithm	88.81	88.86	87.44	88.80
ANN algorithm	91.83	88.56	92.20	90.85
LSTM algorithm	93.97	91.82	94.00	92.50

Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; HAR, human action recognition; LSTM, long short-term memory.

Figure 9:

Comparative outcome of the BOA-EVCHAR approach with other systems. Abbreviations: BOA-EVCHAR, bat optimization algorithm with an ensemble voting classifier for human activity recognition; LSTM, long short-term memory.

CONCLUSION

In this study, we have focused on the design and development of the BOA-EVCHAR technique for the recognition of human activities to help disabled persons in the IoT environment. The BOA-EVCHAR technique makes use of the ensemble classification concept to recognize human activities proficiently in the IoT environment. The BOA-EVCHAR technique comprises a three-stage process, namely, data preprocessing, ensemble classification, and hyperparameter adjustment using BOA. In the presented BOA-EVCHAR technique, data preprocessing is generally performed at the beginning level. For the identification and classification of human activities, an ensemble of two classifiers, namely, LSTM and DBN models, is utilized. Finally, the BOA is used to optimally select the hyperparameter values of the LSTM and DBN models. To elicit the enhanced performances of the BOA-EVCHAR technique, a series of experimentation analyses were performed. The extensive results show the superiority of the BOA-EVCHAR technique on the HAR process. The BOA-EVCHAR method may have limitations in terms of interpretability and explainability of the optimized hyperparameter values selected by the bat optimization algorithm. Future studies on the BOA-EVCHAR method could explore its adaptability to diverse IoT environments, its ability to integrate with edge computing for real-time applications, and further optimization strategies for enhanced performance.

CONFLICTS OF INTEREST

The authors declare no conflicts of interest in association with the present study.

REFERENCES

1. Basly H, Ouarda W, Sayadi FE, Ouni B, Alimi AM. 2022. DTR-HAR: deep temporal residual representation for human activity recognition. Vis. Comput. Vol. 38:993–1013

2. Brishtel I, Krauss S, Chamseddine M, Rambach JR, Stricker D. 2023. Driving activity recognition using UWB radar and deep neural networks. Sensors. Vol. 23(2):818

3. Duhayyim MA. 2023. Parameter-tuned deep learning-enabled activity recognition for disabled people. Comput. Mater. Contin. Vol. 75(3):6587–6603

4. Gaber T, Awotunde JB, Folorunso SO, Ajagbe SA, Eldesouky E. 2023. Industrial internet of things intrusion detection method using machine learning and optimization techniques. Wirel. Commun. Mob. Comput. Vol. 2023:3939895

5. Gupta N, Gupta SK, Pathak RK, Jain V, Rashidi P, Suri JS. 2022. Human activity recognition in artificial intelligence framework: a narrative review. Artif. Intell. Rev. Vol. 55(6):4755–4808

6. Hamad RA, Kimura M, Yang L, Woo WL, Wei B. 2021. Dilated causal convolution with multi-head self attention for sensor human activity recognition. Neural Comput. Appl. Vol. 33:13705–13722

7. HAR dataset. n.d.. Human Activity Recognition Using Smartphones. https://archive.ics.uci.edu/dataset/240/human+activity+recognition+using+smartphones

8. Islam MS, Jannat MKA, Hossain MN, Kim WS, Lee SW, Yang SH. 2023. STC-NLSTMNet: an improved human activity recognition method using convolutional neural network with NLSTM from WiFi CSI. Sensors. Vol. 23(1):356

9. Jugunta SB, Rengarajan M, Gadde S, El-Ebiary YAB, Vuyyuru VA, Verma N, et al.. 2023. Exploring the insights of Bat Algorithm-Driven XGB-RNN (BARXG) for optimal fetal health classification in pregnancy monitoring. Int. J. Adv. Comput. Sci. Appl. Vol. 14(11):731–741

10. Khataei Maragheh H, Gharehchopogh FS, Majidzadeh K, Sangar AB. 2022. A new hybrid based on long short-term memory network with spotted hyena optimization algorithm for multi-label text classification. Mathematics. Vol. 10(3):488

11. Lattanzi E, Donati M, Freschi V. 2022. Exploring artificial neural networks efficiency in tiny wearable devices for human activity recognition. Sensors. Vol. 22(7):2637

12. Li H, Shrestha A, Heidari H, Le Kernec J, Fioranelli F. 2019. Bi-LSTM network for multimodal continuous human activity recognition and fall detection. IEEE Sens. J. Vol. 20(3):1191–1201

13. Li B, Cui W, Wang W, Zhang L, Chen Z, Wu M. 2021. Two-stream convolution augmented transformer for human activity recognitionProceedings of the AAAI Conference on Artificial Intelligence; Vol. Vol. 35(1):p. 286–293

14. Mekruksavanich S, Jitpattanakul A. 2021. LSTM networks using smartphone data for sensor-based human activity recognition in smart homes. Sensors. Vol. 21(5):1636

15. Pesenti M, Invernizzi G, Mazzella J, Bocciolone M, Pedrocchi A, Gandolla M. 2023. IMU-based human activity recognition and payload classification for low-back exoskeletons. Sci. Rep. Vol. 13(1):1184

16. Rashid KM, Louis J. 2019. Times-series data augmentation and deep learning for construction equipment activity recognition. Adv. Eng. Inform. Vol. 42:100944

17. Soni V, Jaiswal S, Semwal VB, Roy B, Choubey DK, Mallick DK. 2023. An enhanced deep learning approach for smartphone-based human activity recognition in IoHTMachine Learning, Image Processing, Network Security and Data Sciences: Select Proceedings of 3rd International Conference on MIND 2021; p. 505–516. Springer Nature Singapore. Singapore

18. Sorguli S, Rjoub H. 2023. A novel energy accounting model using fuzzy restricted Boltzmann machine—recurrent neural network. Energies. Vol. 16(6):2844

19. Sun L. 2023. Optimization of physical education course resource allocation model based on deep belief network. Math. Probl. Eng. Vol. 2023:1–8

20. Tan TH, Wu JY, Liu SH, Gochoo M. 2022. Human activity recognition using an ensemble learning algorithm with smartphone sensor data. Electronics. Vol. 11(3):322

21. Tang CI, Perez-Pozuelo I, Spathis D, Brage S, Wareham N, Mascolo C. 2021. Selfhar: improving human activity recognition through self-training with unlabeled dataProceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies; Vol. Vol. 5. p. 1–30

22. Ullah M, Ullah H, Khan SD, Cheikh FA. 2019. Stacked LSTM network for human activity recognition using smartphone data2019 8th European Workshop on Visual Information Processing (EUVIP); Roma, Italy. 28-31 October 2019; p. 175–180. IEEE. New York:

23. Xu Q, Wei X, Bai R, Li S, Meng Z. 2023. Integration of deep adaptation transfer learning and online sequential extreme learning machine for cross-person and cross-position activity recognition. Expert Syst. Appl. Vol. 212:118807

24. Zhang P, Lan C, Zeng W, Xing J, Xue J, Zheng N. 2020a. Semantics-guided neural networks for efficient skeleton-based human action recognitionProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Long Beach, CA, USA. 15-20 June 2019; p. 1112–1121

25. Zhang J, Zi L, Hou Y, Wang M, Jiang W, Deng D. 2020b. A deep learning-based approach to enable action recognition for construction equipment. Adv. Civ. Eng. Vol. 2020:1–14