Embedded EDA Recognition Technology Based on STM32

Microcontroller

Huajuan Qi, Wenqing Xu and Xinwei Yang

Weifang Engineering Vocational College, Weifang, 262500, Shandong, China

Keywords: STM32 Microcontroller, Embedded EDA, EDA Identification Technology, Emotional State Recognition.

Abstract: In this paper, the embedded EDA recognition technology is implemented based on the STM32

microcontroller, which can collect the skin conductance signal in real time, and perform preprocessing of the

signal based on the signal processing and feature extraction module. In the process of research, the system

architecture is carried out, then the modeling work is carried out, and further training and optimization are

made, and then the system is integrated and applied to the actual case. The experimental data shows that the

accuracy of the system is 92%, and the average delay is less than 30 milliseconds, which indicates that the

system has high real-time and accuracy. It is concluded that the embedded EDA recognition technology based

on STM32 can achieve efficient and accurate emotional state recognition in resource-constrained

environments, and further provide a powerful boost for portable health monitoring devices.

1 INTRODUCTION

Emotional state recognition is of great significance in

mental health monitoring, especially in the field of

real-time emotion monitoring, and the study of EDA

signals has attracted great attention. Some researchers

have proposed that statistical analysis methods can be

used to solve these problems, but the effect is

obviously not good when dealing with complex and

high-dimensional. Some researchers have also

proposed that traditional classification algorithms,

such as decision trees and decision trees, can be used

to classify emotional states, but these methods are

quite unsatisfactory in terms of real-time performance

and accuracy. In addition, some researchers have

proposed that multi-layer neural networks can be used

for identification, but such methods have more

superior computing resources and are basically

impossible to achieve in embedded systems. In this

paper, intelligent algorithms such as Support Vector

Machine (SVM) are used to carry out this research,

because SVM has excellent performance in

processing high-dimensional data and complex

pattern classification, especially suitable for

embedded environments with limited resources. The

goal of this study is to implement a real-time and

accurate EDA emotion recognition system based on

STM32 microcontrollers, so that it can be better

applied to portable health monitoring devices.

2 RELATED WORKS

2.1 STM32 Microcontroller Theory

Based on the ARM Cortex-M core architecture, the

STM32 microcontroller is a high-performance, low-

power embedded microcontroller. It integrates a

variety of peripheral modules (Fang, Yang et al.

2024), such as analog-to-digital converters, serial

communication interfaces, etc. STM32 has the most

powerful real-time processing power to efficiently

perform complex tasks with limited resources. It

supports low-power modes and flexible power

management, making it especially suitable for

embedded systems where energy efficiency is critical

(He and Liu, 2023). The main task of STM32 in the

EDA recognition system is to collect and process skin

conductance signals, and further ensure the real-time

performance of the signals and the stability of data

transmission. Its embedded architecture supports

multi-task parallel processing, providing a strong

hardware foundation for EDA signal processing.

Qi, H., Xu, W. and Yang, X.

Embedded EDA Recognition Technology Based on STM32 Microcontroller.

DOI: 10.5220/0013534900004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 1, pages 32-38

ISBN: 978-989-758-763-4

2.2 SVM Algorithm Theory

Support vector machine (SVM) is a supervised

learning algorithm based on statistical learning

theory, which can be classified and regressed tasks.

The SVM is based on the construction of a hyperplane

that splits the data into different classes

(Kavoliunaite-Ragauskiene, 2024) and maximizes

the spacing between the classes. SVM can effectively

deal with complex linear indivisibility problems in

high-dimensional spaces, and by using kernel

functions, SVMs map low-dimensional data to high-

dimensional spaces to find an optimal classification

boundary in this space (Li, Hua, et al. 2023). In the

EDA recognition task, SVM can classify different

emotional states based on the analysis of changes in

skin conductance signals. It has the advantage of

being able to process high-dimensional data and

avoid overfitting. SVM has a global optimal decision

boundary, so it performs well in sentiment

classification (Miller, 2023).

2.3 Theory of Embedded EDA

Recognition Technology

EDA recognition technology is based on monitoring

the electrical activity of the skin to reflect the

physiological and emotional state of the human body.

EDA signaling is reflected by changes in skin

conductance caused by sweat gland activity, which

fluctuates with mood (Ortiz-Clavijo, Gallego-Duque,

et al. 2023), especially when stressed, anxious, or

excited, with a significant increase in conductance.

This technology is based on sensors to collect skin

conductance signals in real time, and combines signal

preprocessing and feature extraction methods to

perform noise and interference and extract effective

emotion recognition features (Wang, Wang, et al.

2023). These features will be classified by the

classifier and, ultimately, the different emotional

states will be identified in real time. The real-time

nature and low power consumption of embedded

systems will enable EDA identification technology to

be widely applied to portable health monitoring

devices (Xiong, Yin, et al. 2023).

3 METHODS

3.1 Introduction of Embedded EDA

Identification Technology Based on

STM32 Microcontroller

The function of the signal acquisition module is to

collect the skin conductance signal in real time and

convert it into a digital signal. It is responsible for

collecting small changes in skin conductance based

on sensors, and then converts the analog signal into a

digital signal using the STM32 interface, ensuring

high-frequency sampling to capture subtle

conductance fluctuations. It provides a stable data

source for subsequent modules to process. The

function of the signal pre-processing module is to

denoise and filter the collected signal to filter out the

environmental noise and high-frequency interference

in the EDA number. The signal is then low-pass

filtered to retain the key low-frequency components.

At the same time, it can also be based on

standardization to make the signal suitable for feature

extraction and classification. It ensures maximum

purity and reliability of the input signal. The function

of the feature extraction module is to extract key

features from the preprocessed signal for recognition,

and it will recognize the rate of change of skin

conductance and the baseline value as the basis for

recognition. In its work, it also calculates the

instantaneous spikes and durations of EDA signals to

extract emotional state cues. It also has the ability to

convert the extracted features into vector inputs that

the model can process, which provides the most

discriminating signal features for the classification

model. The function of the classification model

module is to use the extracted features to classify and

identify emotional states. It uses the trained model to

classify the input signals in real time, and

distinguishes different emotional states based on

algorithms such as support vector machines. The

current emotional state is judged according to the

output probability value of the model, and the

corresponding emotional label such as "nervous" or

"relaxed" can be output according to the classification

results. The function of the Feedback & Storage

module is to provide feedback based on the

recognition results and store historical data. It feeds

back the current emotional state to the user based on

the LED display or sound alarm to provide the user

with a visual sentiment trend chart. The Feedback &

Storage module stores the identified emotional states

and the corresponding EDA signals locally or in the

cloud. This provides strong historical data support for

subsequent data analysis and model optimization. At

the same time, it also helps users to track long-term

mood changes and manage emotions.

3.2 Design of Embedded EDA

Identification Technology

In the embedded EDA recognition system based on

Embedded EDA Recognition Technology Based on STM32 Microcontroller

STM32, this paper uses the SVM algorithm

modeling, which can use EDA signals to classify the

user's emotional or physiological state. SVM is based

on the construction of an optimal hyperplane to

distinguish the EDA features corresponding to

different emotional states. The decision function is

described in Eq. (1).

𝑦 = 𝑠𝑖𝑔𝑛(𝑊 × 𝑋 + 𝑏)

(1)

In this formula, 𝑦is the output, which represents

the result of the assessment of the emotional or

physiological state after classification. For example,

it is possible to assess whether the current user is in a

state of stress or relaxation. 𝑊is a weight vector that

represents the importance of each EDA feature in the

classification, such as the rate of change in skin

conductance and the baseline value. In EDA

signaling, the feature can be an instantaneous

fluctuation of skin conductance, 𝑋 is a higher weight

indicates that the feature more critical to the

classification of emotions. It refers to the eigenvector

of the input EDA signal, such as the baseline, change

amplitude, frequency and other information of skin

conductance. In SVM, these feature vectors are

actually the basis for classification, which𝑏is used to

identify the different emotional states of the user. In

order to offset the top, it was used to correct the effect

of the SVM model on the baseline difference of skin

conductance in different individuals, and then to

ensure the accuracy of its classification. For example,

bias can correct the normal level of skin electrical

activity in different individuals, and allow the model

to judge the emotional state of individuals more

accurately.

In SVM, 𝑊 is the key to the model to optimize the

weights and bias the top, 𝑏 is different emotional

states can be effectively distinguished based on EDA

features. This is followed by the strategy formula for

classification, as shown in Eq. (2).

𝑓

(𝑥)=𝑊⋅𝑋+ 𝑏

(2)

In this formula, 𝑊is the importance of each EDA

feature in sentiment classification is described. For

example, transient changes in skin conductance and

baseline fluctuations differ in different emotional

states, and it is up to it to determine which feature is

more important in the classification. 𝑋is the

eigenvector from the EDA signal. Each dimension is

a characteristic of the EDA signal, such as the

baseline level of skin conductance, the rate of change,

etc. These traits can be used to identify changes in

mood. 𝑏is used to adjust the output results of the

model to accommodate the differences in skin

conductance characteristics of different individuals.

This ensures that the model can accurately classify

the emotional state of different users, even if they

have different EMG baselines.

In SVM, the classification results depend on the

selection of support vectors. Support vectors

represent those key feature points that are near the

classification boundary. In EDA recognition, support

vectors are generally those EDA signals that have

significant characteristic changes, such as

conductance changes when emotions fluctuate

violently. The model uses vectors to optimize its

classification boundaries to ensure that the interval

between emotional states is maximized. See Eq. (3)

for details.

𝑓

(𝑥)=𝛼









𝑦



𝐾(𝑥



, 𝑥)+𝑏

(3)

In this equation, the Lagrangian multiplier𝛼



referred to, representing the contribution of each

support vector to the classification boundary. 𝑦



are

category labels that support vectors that represent

their emotional states, such as high stress, low stress.

𝐾(𝑥



, 𝑥) is a kernel function that calculates the

similarity between the input features and the support

vectors. The kernel function can map the skin

electrical signal to a higher dimensional space, which

in turn prompts the model to truly find the

classification boundary in complex data.

3.3 Embedded EDA Recognition

Technology Training

In the process of model training, SVM uses the loss

function to evaluate the classification accuracy of the

model, and continuously updates the weights and

biases based on the gradient descent method. In EDA

identification, the loss function measures the gap

between the emotional state assessed by the model

and the actual emotional state, and if the gap is large,

the model minimizes the gap based on adjusting the

weights. The descent method adjusts its model

parameters based on changes in its loss function.

Using the update of the weights and biases of each

step, the model will gradually learn to better classify

the emotional features in the galvanic skin signals. In

this process, the learning rate will determine the step

size of the model update, and a larger learning rate

will make the model update faster, however, it will

cause instability. A smaller learning rate ensures a

more stable model, but its training process is slower.

The bias adjustment is used to correct for differences

INCOFT 2025 - International Conference on Futuristic Technology

in skin conductance baselines in different individuals,

allowing the model to maintain classification

accuracy among different individuals. In the whole

process of model training, the loss function, gradient

descent, learning rate and bias will play a role

together, so that the model can gradually improve the

accuracy of emotional state recognition, and finally

achieve efficient real-time EDA emotional

recognition.

In embedded systems, the computational

complexity of the model is very limited, so this paper

needs to reduce the size of the model based on

pruning. For EDA recognition, it means that the

features and data points in the model that have less

effect on the judgment of emotional states should be

reduced, so that the model can be more efficient in

processing the galvanic skin signals. The process of

pruning can be based on removing features that do not

affect the classification results of the model. For

example, in the characteristics of skin conductance, it

is possible that the magnitude of some changes and

the frequency of certain fluctuations have little effect

on the identification of the user's emotional state, such

as stress or relaxation, then pruning technology can

be used to remove this information, thereby

minimizing the amount of model computation and

speeding up the classification. For this, see Eq. (4).

𝑊



= 𝑊⋅𝑀

(4)

In Eq. (4), 𝑊



is the weight matrix after

pruning represented, which removes the

electrodermal features that have less effect on the

classification of emotions. This makes the model

more concise and efficient. 𝑀is the mask matrix,

which identifies which EDA features are important

and which can be removed. During the pruning

process, the model retains features that are useful for

emotion recognition, such as sudden changes in

conductance, and

Deletes Extraneous Features.

The learning rate plays a decisive role in model

optimization. In EDA signal recognition, optimizing

the learning rate can ensure that the model can

quickly learn the correlation between the

electrodermal signal and the emotional state during

the training process. If the learning rate is too high,

the model will be too sensitive to the signal, which

will lead to unstable emotion recognition, such as

overreaction to small fluctuations in skin

conductance. If the learning rate is too low, the update

speed of the model is too slow, so it cannot capture

the changes of its skin electrical signal in time, which

will affect the real-time recognition. Therefore,

reasonable optimization of the learning rate can help

the model quickly adapt to the changes in user

sentiment and improve the recognition accuracy. See

Eq. (5) for details.

𝑊



= 𝑊



−𝜂⋅𝛻𝐿(𝑊)

(5)

In this formula, 𝜂 is the speed at which the model

updated is described. For skin conductance signals, a

moderate learning rate allows the model to quickly

identify mood changes while remaining stable.

represents 𝛻𝐿(𝑊) is the gradient of its loss function,

which is the current classification error of the model

on the skin conductance signal. After optimizing the

learning rate, the model can adjust the weights more

accurately, thereby improving the ability to recognize

mood swings. In embedded systems, such as the

STM32 platform, hardware resources are limited.

Therefore, in the EDA recognition system, the

hardware optimization is to accelerate the processing

of the skin electrical signal. Based on optimized

hardware, the system reduces latency in data

transmission and is able to speed up categorical

responses to emotional states. For example, based on

direct memory access technology, the system can

automatically process a large amount of

electrodermal data in the background and reduce

processing time. Based on this, the system can

respond immediately when the user's sentiment

changes, without having to wait for a long time for

data processing. See Eq. (6) for details.

𝑇



𝑇



1+𝛼 × 𝑁







(6)

In this formula, 𝑇



refers to the processing time

of the optimized system. After hardware

optimization, the system processes the skin signals

more quickly, and can classify the user's emotional

state in a more timely manner. 𝛼denotes the

optimization factor, which means that hardware

acceleration improves the performance of the system.

Based on hardware optimization, the system can

quickly respond to changes in user sentiment and

further improve the real-time performance of its EDA

signal processing.

3.4 Research and Optimization of

Embedded EDA Recognition

Technology

The system integration adopts a modular design, with

signal acquisition first, and skin conductance signals

based on sensors. The acquired signal is immediately

Embedded EDA Recognition Technology Based on STM32 Microcontroller

transmitted to the pre-processing module, which

removes noise and filters to ensure the quality and

purity of the signal. Subsequently, the feature

extraction module extracts key features from the

purified signal (Xue, Jin, et al. 2024), such as the rate

of change and peak of skin conductance, which

provide a basis for the classification of emotional

states. The classification model receives features and

begins to classify them in real time, and also identifies

emotional states, such as nervousness or relaxation,

based on the trained algorithm. Finally, the feedback

module reminds the user according to the

classification results, and stores the data based on

visual and sound feedback emotional states to

facilitate their subsequent analysis. The system

integration process can maintain the continuity of the

data flow, make it truly seamless from collection to

classification feedback, and ensure that the various

modules work together to ensure the real-time

responsiveness of the system (Zheng and Sun, 2024).

4 RESULTS AND DISCUSSION

4.1 Introduction to the Research Case

of EDA Identification Technology

In modern emotion recognition and physiological

state monitoring, EDA is commonly used to track

individual mood fluctuations. EDA technology is

based on measuring changes in skin conductance to

reflect the physiological response of individuals in

different emotional states. Especially in the field of

mental health monitoring, EDA signals can

effectively identify emotional agitation, anxiety,

stress and other states. The embedded EDA

recognition system based on STM32 microcontroller

can provide real-time emotional feedback for

individuals based on real-time collection of

electrodermal activity combined with appropriate

classification algorithms. For example, in stress

management applications, EDA signals will be used

to monitor the individual's stress levels at work and

school, and wearable devices will be used to remind

users to maintain emotional balance in real time, The

results are shown in Table 1.

Table I shows typical changes in skin conductance

in different emotional states such as calm, nervous,

anxious, and pleasant. According to experimental

data, skin conductance increases significantly when

people are in a state of tension and anxiety, while skin

conductance remains at a lower level in a calm state.

The fluctuation of skin conductance increases in the

pleasant state, although the average is lower, the

composition of the microcontroller is shown in Fig. 1.

Table 1: Changes in skin conductance in different

emotional states.

Emotional

state

Mean skin

conductance

(μ

)

Fluctuations

in skin

conductance

remark

calm 1.8 low

The

conductance

remains

stable

nervous 5.2 high

Conductance

is markedly

elevate

anxiety 4.7 medium

The

conductance

is fluctuatin

pleasant 2.5 Higher

Conductivity

is highly

volatile

Figure 1: Composition of STM32 microcontroller.

4.2 The Signal Recognition Process of

Microcontroller

The EDA recognition technology studied the changes

of skin conductance in different emotional states, and

30 people were selected to participate in the

experiment, 4 different emotions were recorded, and

the experiment lasted for 3 days, 5 conductance

fluctuation recognizers, and 3 EDA signal detection

machines. These conductance changes provide the

basis for emotional classification of EDA signals, as

shown in Table 2.

INCOFT 2025 - International Conference on Futuristic Technology

Table 2: Real-time EDA signal acquisition performance

based on STM32 microcontrollers

Performance

metrics

Data values remark

Acquisition

accuracy

±0.02 μS High-precision

signal

uisition

Average latency 30 ms Ideal for real-

time emotion

monitoring

lications

Data transfer

delays

Less than 20

milliseconds

Real-time

response is

guarantee

Table 2 shows the performance metrics for real-

time acquisition of EDA signals in an embedded

system based on STM32 microcontrollers. For

example, the accuracy and real-time nature of data

acquisition, and the delay of data transmission. The

results show that STM32 can maintain high-precision

EDA signal acquisition with an average delay of less

than 30 milliseconds under limited computing

resources, indicating that the system can support real-

time emotion monitoring applications,The EDA

recognition process of the microcontroller is shown in

the pattern..

Figure 2: Embedded recognition of microcontroller.

Figure 2 shows that after embedded recognition

by the microcontroller, data correlation analysis can

be achieved and disturbance signals can be identified.

4.3 Research Effect of Embedded EDA

Recognition Technology

Based on the comprehensive analysis of the three

tables, it can be concluded that the embedded EDA

recognition system based on STM32 microcontroller

in this study has a multi-faceted performance in

emotion recognition. The data showed significant

changes in skin conductance in different emotional

states, with significantly higher conductance values in

stress and anxiety states and lower levels in calm

states. This trend provides a very critical basis for the

classification of EDA signaling, indicating that skin

conductance can effectively reflect emotional state,

The comparison of recognition accuracy of

microcontrollers is shown in Table 3.

Table 3: Comparison of recognition accuracy of different

classification algorithms

Algorithm

Recognition

Accurac

(%)

Computational

complexit

KNN

85 Lower

SVM

92 medium

Random

fores

88 Higher

Table 3 shows the performance of different

machine learning algorithms in the EDA signal

sentiment recognition task. Based on the comparison

of the recognition accuracy of KNN, SVM and

random forest algorithms, it can be found that SVM

has the highest accuracy of 92%, while the accuracy

of KNN and random forest is 85% and 88%,

respectively. It fully shows that SVM can still

maintain high classification performance in STM32

systems with limited resources, and is suitable for

actual embedded EDA identification systems,

Accuracy changes, as shown in Figure 3.

Figure 3: Embedded recognition process of

microcontroller.

The data analysis in Figure 3 shows that., the

system is outstanding in high-precision acquisition

and low-latency transmission. The average latency of

30 milliseconds ensures that the system can be

applied in real-time scenarios, so that the emotion

monitoring results can be fed back to users in time.

This conclusion echoes the trend of skin conductance

change, so that the system can capture emotional

fluctuations in time and respond quickly, showing

good real-time processing ability. In addition, the data

Embedded EDA Recognition Technology Based on STM32 Microcontroller

shows that SVM has the best performance in

classification tasks, with an accuracy rate of 92%,

which is much higher than that of KNN and random

forest. Based on this, it can be seen that SVM has

stronger discrimination ability and stability when

processing multi-dimensional data such as EDA

signals. Combined with the characteristics of skin

conductance fluctuations, SVM can accurately

capture the differences between different emotional

states, and show high reliability in classification

results.

5 CONCLUSIONS

The conclusion of this paper is that the embedded

EDA based on STM32 microcontroller can use the

effective combination of SVM algorithms to

efficiently distinguish different emotional states,

which reflects the superiority of the SVM algorithm

based on STM32 microcontroller for real-time signal

processing. By extracting the key features of EDA

signals and optimizing the weights and biases, the

model can still maintain high classification accuracy

in the face of individual differences. The selection of

support vectors and the optimization of classification

boundaries ensure that the model can effectively

capture the characteristics of different emotional

states of individuals in complex EDA signals, and at

the same time, maximize the interval between

emotional states. In addition, through model

optimization and other steps, this paper also ensures

the effectiveness of embedded EDA recognition

based on STM32 microcontrollers, and greatly

improves the real-time response ability of the system.

Based on this, the model and its system will provide

further support for the application of intelligent health

monitoring. Although this paper has relatively

complete data support, it still has some technical

limitations and limitations. Based on this, further

optimizations can be made at a later date.

REFERENCES

Fang, L. Y., Z. Yang, T. L. Ma, J. Yue, W. Y. Xie, P.

Ghamisi, and J. Li. (2024). Open-World Recognition in

Remote Sensing. Concepts, challenges, and

opportunities. Ieee Geoscience and Remote Sensing

Magazine, 12(2), 8-31.

He, N., and W. H. Liu. (2023). Network Oral English

Teaching System Based on Speech Recognition

Technology and Deep Neural Network. International

Journal of Advanced Computer Science and

Applications, 14(12), 829-839.

Kavoliunaite-Ragauskiene, E. (2024). Right to Privacy and

Data Protection Concerns Raised by the Development

and Usage of Face Recognition Technologies in the

European Union. Journal of Human Rights Practice,

16(2), 658-674.

Li, W., M. L. Hua, Y. Sun, H. S. Li, and Y. H. Lin. (2023).

FACE, FACIAL RECOGNITION TECHNOLOGY

AND PERSONAL PRIVACY. Acta Bioethica, 29(2),

259-268.

Miller, K. W. (2023). Facial Recognition Technology.

Navigating the Ethical Challenges. Computer, 56(1),

76-81.

Ortiz-Clavijo, L. F., C. J. Gallego-Duque, J. C. David-Diaz,

and A. F. Ortiz-Zamora. (2023). Implications of

Emotion Recognition Technologies. Balancing Privacy

and Public Safety. Ieee Technology and Society

Magazine, 42(3), 69-75.

Wang, Q. W., P. X. Wang, and Y. Z. Chang. (2023). Deep

Learning-Based Intelligent Image Recognition and Its

Applications in Financial Technology Services.

Traitement Du Signal, 40(2), 735-742.

Xiong, J. X., H. Yin, and M. S. Pan. (2023). Book

recommendation and purchase of intelligent image

recognition technology under the background of 5G

environment. Journal of Computational Methods in

Sciences and Engineering, 23(2), 995-1005.

Xue, Y. X., Y. T. Jin, S. P. Chen, H. J. Du, and G. Shen.

(2024). Modulation Signal Automatic Recognition

Technology Combining Truncated Migration

Processing and CNN. Ieee Access, 12, 120414-120428.

Zheng, X., and Z. Sun. (2024). Exploration on multimodal

data recognition method for Internet of Things based on

deep learning. Intelligent Decision Technologies-

Netherlands, 18(2), 759-767

INCOFT 2025 - International Conference on Futuristic Technology