Design and Simulation-Based Validation of an Embedded Acquisition Architecture for In Situ PCB Integrity Monitoring in Biomedical Devices

Abstract

The reliability of biomedical devices is closely linked to the quality and long-term stability of the electronic circuits that support their operation. Printed circuit boards (PCBs), in particular, can be affected by manufacturing imperfections, thermal stress and progressive ageing, which may lead to failures during the device life cycle. In this study, we present the design and simulation-based validation of an embedded acquisition circuit aimed at monitoring PCB electrical integrity in a non-invasive and remote manner. The presented solution is based on Hall-effect current sensing combined with a 16-bit analog-to-digital conversion stage and a digital communication interface managed by a Raspberry Pi. This configuration allows the system not only to acquire integrity-related electrical signals but also to process them locally and transmit them wirelessly for supervision purposes. A lightweight artificial intelligence model is implemented directly on the embedded platform to analyse the acquired signals and to classify different PCB operating conditions in real time. Simulation results show that the system is able to identify small current variations caused by micro-discontinuities and abnormal conductive paths. The classification accuracy exceeds 97% for PCB integrity states, confirming the suitability of the approach for remote monitoring, predictive maintenance and safety support in electromedical devices.

1. Introduction

In recent decades, biomedical devices have seen widespread adoption of increasingly complex and miniaturised electronics, with PCBs forming the functional heart of electro-medical systems for monitoring, diagnostics and therapy [1]. PCBs are the basis of physiological signal control instruments, human–machine interfaces and real-time data communication modules, and their reliability is a fundamental prerequisite for the safety and effectiveness of the device as a whole [2]. However, manufacturing processes and operational constraints introduce structural and functional criticalities on PCBs, such as micro-trace defects, variations in electrical parameters and discontinuities in conductive paths, which can lead to intermittent or permanent failures during the device’s life cycle [3].

The need to monitor PCB integrity is therefore central not only during production, where quality control methodologies such as automated optical inspection and radiographic techniques provide effective inspections, but also during operation, to detect emerging degradation phenomena before they can compromise the functional safety of the device [4]. The literature on Non-destructive testing (NDT) applied to electronics has introduced various techniques, including thermographic methods, acoustic analysis and radiography for the qualitative assessment of PCBs produced using traditional or additive technologies [5]. Although effective in industrial contexts, these techniques generally require laboratory equipment and are not always easily adaptable to embedded solutions for continuous or remote monitoring. In the field of research, methodologies have been developed for the non-destructive diagnosis of electronic paths using approaches based on the frequency domain or on intrinsic characteristics of energy distribution networks, with the aim of identifying anomalies automatically and with noise robustness [6]. Other studies have explored in situ inspection systems that combine magnetic sensor technologies (e.g., GMR sensors coupled with eddy current techniques) with high-frequency acquisition for the detection of micro-defects on PCBs [7]. Although relevant for measurement, these approaches tend to focus on the sensor or the physical phenomenon associated with detection [8]. This choice leaves little room for in-depth analysis of the integrated acquisition circuits needed to obtain reliable, digitizable, and transmissible measurements in remote monitoring environments. The growing interest in predictive maintenance methodologies highlights the need for integrated electronic systems capable of collecting meaningful data from PCBs and communicating it to advanced supervision platforms [9]. Despite the wide range of quality control techniques available, the literature is still limited in areas where cost, compactness, interference and electromagnetic compatibility constraints are particularly stringent [10]. In fact, most of these studies deal with detection techniques, leaving out the electrical acquisition chain, which includes signal conditioning, noise mitigation, high-resolution digitalisation, and data preparation for embedded intelligence [11,12]. Essentially, the acquisition circuit has been relegated to a secondary or implicit component, rather than a critical design element requiring dedicated validation. This paper addresses precisely this gap: it proposes the design, implementation and validation of an embedded acquisition circuit dedicated to the remote monitoring of PCB integrity in biomedical applications. Specifically, the main contributions of this study are threefold: (i) the design and validation of a low-noise, Hall-effect-based acquisition circuit tailored for non-invasive PCB integrity monitoring in biomedical devices; (ii) the integration of the acquisition chain with an embedded processing and communication platform enabling on-board AI-based signal analysis; and (iii) a system-level validation demonstrating the suitability of the introduced architecture for continuous and remote monitoring under embedded constraints. The developed architecture integrates a Hall-effect-based sensing interface with a passive analog conditioning stage, followed by a 16-bit embedded ADC suitable for low-frequency, low-amplitude signal acquisition in biomedical environments. The acquisition chain is interfaced with a Raspberry Pi-based platform enabling local AI-based processing and wireless data transmission. This study focuses on the design and validation of an integrated monitoring architecture centred on a dedicated acquisition circuit for assessing PCB integrity in biomedical devices. By placing the acquisition circuit at the core of the monitoring strategy, the presented approach enables continuous, non-invasive monitoring compatible with embedded PCB systems, overcoming the limitations of thermography- and inspection-based methods. Unlike approaches primarily centred on sensors or imaging, the study addresses the complete embedded signal acquisition and processing chain, with particular attention to circuit-level implementation and applicability in real biomedical devices. The remainder of the paper details the system architecture, acquisition circuit design, embedded processing pipeline, and validation results.

2. Related Works

Monitoring the integrity of electronic circuits used in biomedical devices is the subject of a growing number of studies, motivated by the need to ensure patient safety, operational continuity and compliance with regulatory requirements [13]. In particular, ensuring the integrity of PCBs embedded within biomedical devices poses significant challenges due to miniaturisation, low signal amplitudes, electromagnetic interference constraints and the requirement for continuous, non-invasive monitoring [14]. The solutions proposed in the literature can be traced back to different methodological approaches, including thermography-based inspection [15,16,17,18,19], direct electrical signal monitoring [20,21], and embedded magnetic or electromagnetic sensing techniques [22,23], each addressing the problem with different tools and levels of integration. A first established approach involves the use of infrared thermography for the analysis of electronic biomedical devices [24]. Classic studies have shown that thermal imaging can detect functional anomalies and hotspots associated with overcurrents or soldering defects in electronic circuits [25,26]. In the biomedical field, these techniques have been applied to the reliability assessment of electronic boards used in diagnostic and therapeutic instrumentation, demonstrating good non-invasive detection capabilities [27]. However, several authors point out that these approaches require high-resolution external equipment and are not easily integrated into embedded systems for continuous monitoring [28]. Moreover, thermographic approaches are inherently sensitive to environmental conditions and require optical access to the device under test, limiting their applicability in sealed, miniaturised or implantable biomedical electronics [29,30]. A second group of studies focuses on direct electrical monitoring of circuit parameters in biomedical devices. In this context, the analysis of voltages, currents and impedances assesses the health status of systems such as pacemakers, neurostimulators and physiological monitoring devices [31,32]. These approaches are particularly interesting because they allow direct integration into the medical device, reducing dependence on external instrumentation. However, the literature highlights that the reliability of such approaches is strongly dependent on the quality of the analog front-end and the analog-to-digital acquisition chain, which are often described only at a high level, without detailed circuit-level design choices, noise analysis or simulation-based validation [33]. A further line of research concerns the use of magnetic and electromagnetic sensors for the non-invasive control of biomedical circuits. Hall and magnetoresistive sensors have been used for the indirect measurement of currents and for the detection of discontinuities in the conductive paths of sealed or implantable devices [34,35,36]. The main advantage of these approaches lies in the possibility of taking measurements without direct contact, a crucial feature in biomedical applications. However, many studies focus mainly on the measurement principle or sensor performance, neglecting an in-depth characterisation of the associated conditioning and acquisition electronics [37]. As a consequence, the reproducibility and robustness of these solutions in embedded biomedical environments remain difficult to assess, particularly when low-amplitude signals and long-term stability are required. In recent years, numerous contributions have addressed the issue of remote monitoring and predictive maintenance of biomedical devices, integrating acquisition systems with communication architectures and data analysis platforms [38,39]. These solutions improve the life cycle management of medical devices and support maintenance strategies based on actual operating conditions. However, as highlighted in several review papers, such systems often rely on generic or commercial acquisition modules, without a design aimed at controlling circuit integrity [40,41]. As a result, these architectures are generally not optimised for detecting subtle electrical anomalies related to PCB degradation, but rather focus on high-level functional parameters. Other studies address the issue of reliability through systemic failure analysis and health monitoring strategies for medical electronics. In this context, methodologies have been developed for assessing the degradation of electronic circuits over time, with a particular focus on implantable and portable devices [42,43,44]. Although these works provide a solid theoretical approach, their practical implementation requires dedicated and carefully designed acquisition circuits, an aspect that often remains implicit or is left to standard solutions [45]. Although several studies exploit high-frequency acquisition or external instrumentation for defect detection, these solutions are often unsuitable for embedded, low-power, and continuously operating biomedical systems. Despite the wide variety of approaches to monitoring biomedical devices, there is still no solution that focuses on the design and validation of an integrated acquisition circuit designed to verify PCB integrity and for integration into remote monitoring architectures. Recent studies integrate research on PCB integrity with data analysis and artificial intelligence techniques, which further increase the requirements in terms of signal quality, stability, and reliability of the acquisition chain [46,47,48]. In contrast with the approaches reviewed above, this study explicitly addresses the circuit-level design and simulation-based validation of an embedded acquisition system specifically tailored for low-frequency, non-invasive PCB integrity monitoring, typically in the range of a few hertz to several kilohertz, consistent with the electrical dynamics of biomedical PCBs. Beyond thermography-based approaches, recent works have explored embedded and electrical-signal-based monitoring strategies for PCB and electronic system integrity. Several studies focus on direct current and voltage monitoring combined with lightweight machine learning models for fault detection in embedded systems. Compared to these approaches, the present work distinguishes itself by explicitly addressing the entire acquisition chain, from non-invasive Hall-effect sensing and analog front-end design to digitisation and embedded AI-based processing. While many contributions emphasize either sensing principles or data-driven analysis alone, this study adopts a system-level perspective tailored to the constraints of biomedical electronics. Compared with thermography-based and external instrumentation approaches, the proposed system offers several advantages: (i) it operates directly on electrical signals without requiring optical access or external sensors; (ii) it is fully embedded and suitable for continuous in situ monitoring; (iii) it integrates the complete acquisition chain with on-device (edge) AI inference; and (iv) it is inherently low-power and compatible with biomedical electronic constraints. These features make the proposed architecture more scalable and deployable for real medical devices than inspection-oriented solutions.

3. System Architecture and Acquisition Circuit Design

The architecture design aims to ensure reliability, modularity and compatibility with different types of electronic biomedical devices. This section describes the architecture of the developed system, with particular emphasis on the design of the embedded acquisition circuit and the related data processing chain. Unlike approaches that focus mainly on the principle of sensing or downstream analysis of results, attention is focused on the entire chain of acquisition, digitisation and intelligent processing of the electrical signal. The study starts from the assumption that the electrical signal is the key element for reliable monitoring of the integrity of biomedical devices. From this perspective, the acquisition circuit is not a simple support block, but constitutes the functional core of the presented architecture. The section is organised into subsections that progressively guide the reader from the conceptualisation of the contribution and its positioning with respect to the existing literature, to a detailed description of the hardware architecture and the artificial intelligence-based processing pipeline. The analysis concludes with a performance evaluation of the system and a critical comparison with alternative approaches, highlighting the novelty and robustness of the presented approach.

3.1. Conceptualisation of the Contribution and Positioning with Respect to Previous Studies

This study builds upon previous research on the monitoring of biomedical electronic devices using thermographic imaging and artificial intelligence. In particular, in [49], device status monitoring was mainly based on external thermographic analysis, supported by FEM modelling and automatic anomaly classification. Dependence on external instrumentation, sensitivity to environmental conditions, and the indirect nature of measurement have made it difficult to integrate the system into continuous or integrated monitoring scenarios, which are typical of many biomedical applications. The present study differs by shifting the centre of gravity of the architecture to the embedded acquisition circuit, designed to operate directly on the electrical signals of the biomedical device. In this context, the focus is not on the sensor itself, but on the systematic design of the entire chain of data acquisition, digitisation and processing, considered as an enabling element for reliable circuit integrity monitoring. The architecture maintains the modular philosophy already adopted in the [50] study, reusing its embedded acquisition system and artificial intelligence-based processing pipeline. However, in this case, the thermal sensor is replaced by a dedicated electrical interface, allowing the approach to be extended to a wider class of electronic biomedical devices.

The central role of the electrical acquisition circuit therefore makes it necessary to describe its architecture and main design choices in detail, which is the subject of the following sections. Figure 1 conceptually shows the transition from an external monitoring system, based on indirect measurements and dedicated instrumentation, to an integrated architecture incorporated directly into the electrical signals of the biomedical device.

This comparison clarifies the positioning of the presented contribution with respect to the authors’ previous work and anticipates the key role of the acquisition circuit in the processing and diagnosis chain.

3.2. System Architecture

The system architecture, designed as a modular embedded platform, was created for the reliable monitoring of the integrity of electronic biomedical devices through the direct analysis of internal electrical signals. The main objective is to design a robust and reconfigurable acquisition and processing chain compatible with different biomedical circuits, while maintaining the artificial intelligence-based processing pipeline unchanged. As illustrated in Figure 2, the system architecture is divided into four functionally distinct but closely interconnected subsystems: (i) the non-invasive electrical acquisition interface, (ii) the analog signal conditioning front-end, (iii) the embedded analog-to-digital conversion and local processing unit, and (iv) the communication and remote supervision module. This subdivision allows the critical design issues associated with the acquisition of low-amplitude signals to be isolated from digital processing and data transmission, improving the overall reliability of the system. The biomedical device under test is monitored via electrical access points selected so as not to alter its normal functioning.

The signals acquired consist of local variations in voltage and current, directly related to the integrity of the conductive tracks and components. This monitoring mode allows the detection of both progressive degradation phenomena, such as increased contact resistance or the onset of micro-discontinuities, and more pronounced anomalies associated with incipient faults or abnormal operating conditions. The acquisition circuit is designed to operate on low-amplitude signals, typically ranging from a few tens of millivolts to a few volts, with a frequency band of interest extending from DC to approximately 150 Hz, consistent with the electrical dynamics of PCB degradation phenomena. These amplitude and frequency ranges are consistent with reported electrical dynamics in low-power biomedical electronic subsystems and PCB-level integrity monitoring applications [51,52,53]. The sampling frequency is configured up to 860 samples per second per channel, which is sufficient to capture slow and mid-band electrical variations associated with PCB integrity monitoring, while maintaining low power consumption and I²C compatibility. A key aspect of the architecture is the conceptual and functional separation between the electrical acquisition subsystem and the artificial intelligence-based processing pipeline. This design choice allows the reuse of the same embedded infrastructure and AI models validated in previous work by the authors, modifying only the sensing interface according to the specific application. The result is a highly flexible system, in which the centrality of the embedded acquisition circuit is the distinctive feature compared to approaches based on external sensors or indirect measurements. The main specifications of the embedded acquisition system are summarised in

Table 1. Main specifications of the embedded acquisition system.

Parameter	Value
ADC resolution	16 bits
Sampling frequency	Up to 860 SPS (per channel)
Bandwidth	DC–400 Hz
Programmable gain	±0.256 to ±6.144 V (PGA, ADS1115)
Equivalent input noise	<20 µV RMS
Power consumption	<150 mW
Single-board computer (SBC)	Raspberry Pi
Communication	Wi-Fi

, providing a quantitative overview of the performance achieved in terms of resolution, bandwidth, noise and power consumption.

3.3. Analog Front-End: Detailed Circuit Layout

The analog front-end is the core of the system and guarantees high measurement sensitivity, operational stability and high noise immunity, which are fundamental requirements for reliable monitoring of the integrity of electronic biomedical devices.

This subsystem plays a central role in preserving the diagnostic information contained in the acquired electrical signals, ensuring that variations associated with progressive degradation or abnormal operating conditions are not masked by interference or distortion introduced by the acquisition chain. The preservation of diagnostically relevant information is achieved through a combination of non-invasive Hall-effect sensing, impedance buffering, and analog low-pass filtering. The Hall-effect sensor provides galvanic isolation and converts current variations into proportional voltage signals without loading the PCB under test. The buffering stage ensures that the signal source is not affected by the ADC input impedance. The anti-aliasing and noise mitigation stage is implemented as a first-order passive RC low-pass filter placed at the input of each ADC channel. No second-order or active filtering stages are employed in the analog domain. The anti-aliasing and noise mitigation stage is implemented as a first-order passive RC low-pass filter placed at the input of each ADC channel. No second-order or active filtering stages are employed in the analog domain. The attenuation slope is therefore 20 dB/decade, consistent with a single RC network per channel. At the digital level, windowing and statistical feature extraction further enhance robustness by emphasizing structured signal variations and reducing sensitivity to uncorrelated noise. Any additional or higher-order filtering is performed digitally after analog-to-digital conversion.

As shown in Figure 3, the analog front-end is structured as a signal conditioning chain that includes a non-invasive Hall-effect current sensing interface, an output buffering and impedance matching stage, and a passive RC low-pass filter for anti-aliasing. The diagram highlights the Hall-effect input interface and the signal conditioning network preceding the ADS1115 16-bit ADC. The detailed circuit schematic of the acquisition interface is reported in Figure 4. It is based on four ACS712 Hall-effect current sensors, each connected to a PCB conductive path under monitoring. The ACS712 devices provide galvanically isolated current-to-voltage conversion, enabling non-invasive acquisition of integrity-related electrical variations.

The acquisition channels shown in Figure 4 are electrically independent and implement identical filtering stages. Each Hall-effect sensor output is connected to a dedicated ADC input through an individual first-order RC low-pass filter (1 kΩ–33 nF). No electrical interconnection exists between ADC input nodes, except for the common ground reference. This configuration ensures channel isolation, prevents cross-coupling effects, and guarantees that any detected signal variation can be unambiguously attributed to the corresponding monitored PCB trace. The system integrates four ACS712 Hall-effect current sensors and an ADS1115 16-bit ADC, whose inputs are conditioned by dedicated first-order RC low-pass filters. The ADC communicates with an external Raspberry Pi through the I²C bus via the SDA and SCL lines for data acquisition and remote transmission. A +3.3 V reference rail biases the acquisition network, while 2.2 kΩ pull-up resistors are used on the I²C lines to ensure reliable digital communication. This architecture allows low-amplitude signals to be acquired non-invasively, reducing the impact of load variations and common-mode interference typical of biomedical environments. The main design features of the input stage and Hall-effect current sensor with internal signal conditioning are summarised in

Table 2. Main parameters of the input stage and instrument amplifier.

Parameter	Value	Description
Input impedance	>1 MΩ	Non-invasive interface
Output impedance	<1 kΩ	Compatible with ADC input
CMRR	Inherent to sensor design	Magnetic isolation
Programmable gain	Fixed (sensor sensitivity)	66–185 mVA−1 (ACS712)
Equivalent noise	≈20–30 µV RMS	Low—noise front-end
Useful bandwidth	DC—400 Hz	PCB degradation dynamics

, which summarises the fundamental electrical parameters adopted to ensure a non-invasive interface and high quality of the acquired signal.

In particular, the input impedance greater than 1 MΩ allows the biomedical device to be monitored without altering its operating regime, while the high common-mode rejection reduces the influence of electromagnetic interference and spurious voltages typical of clinical and laboratory environments. The gain adaptation is achieved through the programmable input range of the ADS1115 ADC, which allows the full-scale voltage to be selected according to the expected signal amplitude.

As shown in Table 2, the low equivalent input noise allows the detection of weak electrical variations, particularly relevant in the early stages of degradation, which could be indistinguishable from background noise in the absence of adequate amplification.

Downstream of the amplification stage, the signal undergoes analog anti-aliasing filtering implemented as a first-order RC low-pass filter. Downstream of the amplification stage, the signal undergoes analog anti-aliasing filtering implemented as a first-order RC low-pass filter. In the present design, the RC network shown in Figure 4 (1 k Ω–33 n F) results in a cutoff frequency of approximately 4.8 kHz, according to (1):

$$f_c={\displaystyle \frac{1}{2\mathsf{\pi }\mathrm{R}\mathrm{C}}}$$

(1)

It is important to clarify that the analog filtering stage is strictly first order. The resulting attenuation slope is therefore 20 dB/decade, fully consistent with the use of a single RC network per channel. This choice represents a design trade-off between preserving diagnostically relevant transient components of the electrical signal and attenuating high-frequency electromagnetic interference and sensor noise before digitization. Although PCB degradation phenomena mainly evolve in the low-frequency range, the higher cutoff ensures that short-time perturbations caused by micro-discontinuities and intermittent contacts are not excessively smoothed by the analog stage. The filter is therefore not intended to strictly limit the signal bandwidth to the Nyquist frequency, but rather to provide noise mitigation while preserving waveform features relevant for AI-based classification. As reported in

Table 3. Analog anti-aliasing filter specifications.

Parameter	Value	Description
Type	Low-pass RC filter	Anti-aliasing/noise mitigation
Order	First order	Simple, stable implementation
Cut-off frequency	≈ 4.8 kHz	Preserves transient degradation features
Components	1 k Ω–33 n F	RC network (Figure 4)
Out-of-band attenuation	~20 dB/dec	High frequency suppression
Design rationale	-	Noise reduction + signal preservation

, the use of low-noise passive components with reduced thermal drift contributes to improving the overall signal-to-noise ratio of the acquisition chain and ensures stable performance over time. From a quantitative perspective, the thermal noise of a resistor is given by (2),

$$v_n=\sqrt{4kTRB}$$

(2)

where k is Boltzmann’s constant, T the absolute temperature, R the resistance, and B the bandwidth. By selecting low-value, low-tolerance resistors and stable dielectric capacitors in the RC filter, the equivalent input noise of the front-end is minimized. In addition, the low-pass filtering limits the noise bandwidth seen by the ADC, directly reducing the integrated noise power and improving the effective signal-to-noise ratio of the digitized signal. The presence of internal test nodes (Figure 2) allows for the experimental characterisation of the analog front end and the validation of system performance under operating conditions, facilitating calibration and diagnostic activities during the development and testing phases.

3.4. Analog-to-Digital Conversion and Data Synchronisation

The processing of electrical information through the analog front-end and its transfer to the digital world for further processing with AI depends on the ADC. The proper design of this stage is crucial to preserving signal integrity and ensuring reliable diagnostic performance. A brief description of the main features of the A/D unit employed in the system, in terms of resolution, sampling frequency and timing performances, is presented in

Table 4. Synchronization and timing performance of the embedded acquisition system.

Parameter	Value	Description
Type	SAR ADC	Low power consumption and low latency
Resolution	16 bits	Configurable
Sampling frequency	1–860 SPS	Adaptive
INL/DNL	< ±1 LSB	Linearity
Conversion time	~1.2 ms @860 SPS	Fast acquisition
Power consumption	<20 mW	Embedded-friendly

. The selected ADC is a low-noise SAR device (ADS1115) with 16-bit resolution and a configurable sampling rate up to 860 samples per second per channel. The sampling frequency can be set in the range 1–860 samples per second (SPS), allowing the system to be adapted to the different dynamics of the monitored electrical signals.

The choice of a nominal sampling frequency of 860 SPS ensures adequate time resolution for the system’s band of interest and is consistent with the specifications of the analog anti-aliasing filter described in Section 3.3. The configuration in Table 4 prevents aliasing phenomena and maintains high fidelity of the acquired signal, even in the presence of rapid variations associated with abnormal operating conditions of the biomedical device. A fundamental aspect of the design concerns the temporal synchronisation of the acquired data, which uses a synchronisation mechanism based on a software timer running on the Raspberry Pi’s Linux kernel. This approach ensures regular sampling instants and maintains limited temporal drift.

Table 5. Analog-to-digital converter specifications.

Parameter	Value	Description
Sampling jitter	<1 ms	Temporal stability
Overall latency	<100 ms	Acquisition + buffering
Timestamp resolution	≈1 ms	Temporal accuracy
Synchronisation method	Software timer (Linux version 6.1)	Deterministic
AI support	Local inference on SBC	AI support

summarises the system’s synchronisation and timing performance, reporting the jitter and latency results of the acquisition chain.

With a sampling jitter below 1 ms, the system provides sufficient temporal consistency for AI-based classification of low-frequency electrical integrity signals. Inference is performed locally on the Raspberry Pi using a lightweight CNN implemented in TensorFlow Lite, enabling near real-time classification. This processing paradigm can be properly described as edge computing, since data acquisition, feature extraction, and AI inference are performed locally on the device without reliance on continuous cloud connectivity. The configuration is appropriate for near real-time applications monitoring and possibly predictive maintenance since the overall data acquisition and transfer delay to the embedded processing unit is below 50 ms. Accurate time stamping of the acquired samples is a crucial element for coupling with the artificial intelligence-based processing pipeline. Each data block, associated with a timestamp synchronized to the system clock, allows for the temporal sequences to be accurately reconstructed and the features extracted to be dependable and consistent. It is a consequence of the proposed approach automatically classifying the status of biomedical equipment that the time-dependent signals can be of good quality, which in turn plays a critical role in the machine learning models’ efficacy.

3.5. AI-Based Signal Processing and Experimental Setup

After analog-to-digital conversion, the acquired electrical signals are processed locally on the embedded platform using a lightweight artificial intelligence pipeline designed for low-frequency integrity monitoring. The complexity and variability of electrical signals associated with PCB degradation phenomena requires the choice of an AI approach. For this reason, the study presents a data-driven classification approach that can learn discriminating features directly from the acquired signals, while remaining compatible with the computational constraints of embedded platforms. The digital samples are first filtered and normalised, then segmented into fixed-length, partially overlapping time windows. This operation preserves the temporal structure of the signal while isolating short-time patterns that are relevant for integrity assessment. Each window is provided as input to a compact convolutional neural network (CNN) specifically tailored for execution on resource-constrained hardware. Although the input representation is a compact vector of statistical features (RMS, mean, variance, kurtosis and energy), the CNN operates on sequences of such feature vectors extracted from consecutive time windows. In this way, one-dimensional convolutions are applied along the temporal dimension of the feature sequence, enabling the network to learn local temporal patterns in the evolution of the signal statistics. The adopted CNN architecture consists of two 1D convolutional layers with ReLU activation, followed by a max-pooling layer and a lightweight fully connected layer, for a total number of parameters in the order of a few thousand. The resulting TensorFlow Lite model occupies less than 200 kB of memory, making it suitable for embedded deployment on the Raspberry Pi platform. The CNN architecture adopted consists of a limited number of convolutional layers followed by fully connected light layers, designed to balance classification accuracy and computational efficiency. The model operates on low-dimensional input representations derived from electrical signals in the time domain. Furthermore, the model is trained to discriminate between different PCB integrity states, namely normal operation, progressive degradation, and failure conditions. This architecture was selected as a compromise between expressive power and real-time execution capability on the Raspberry Pi platform. The model is deployed on the Raspberry Pi using TensorFlow Lite, allowing on-board inference with limited memory and computational power. In this way, classification is performed directly at the edge, without the need for continuous data streaming to an external server, and the system can react in near real time to abnormal electrical behaviour.

Figure 5 shows that, after analog-to-digital conversion, the signal is digitally filtered, segmented into time windows, and analysed by a lightweight CNN running on Raspberry Pi using TensorFlow Lite. The digitised signals are divided into windows of N samples, corresponding to a time interval (3)

$$T={\displaystyle \frac{N}{f_s}}$$

(3)

where (f_s) is the sampling frequency. A 50% overlap between consecutive windows is used to avoid loss of transient information and to ensure continuity between segments.

The digitised electrical signals are first processed to extract a reduced and informative set of statistical features, which are then used as the sole input for the classification algorithms. This choice ensures dimensionality reduction, computational efficiency and noise robustness, which are important conditions for reliable execution on embedded platforms. In order to justify the choice of the convolutional neural network adopted, a preliminary comparative evaluation was carried out using the same set of acquired data and identical training and validation protocols. The study considered classification for all algorithms, using the same set of statistical features extracted from the acquired electrical signals, ensuring a fair and consistent comparison. The analysis, as detailed in

Table 6. Performance comparison of AI algorithms.

Algorithm	Input Representation	Accuracy (%)	Inference Time (ms)	Embedded Suitability
Threshold-based method	RMS, mean	72.4	<1	High
k-NN (k = 5)	Statistical features	81.6	18	Medium
SVM (RBF kernel)	Statistical features	86.9	32	Low
Random Forest	Statistical features	88.1	25	Medium
CNN	Statistical features	94.3	7	High

, focuses on classification accuracy, inference latency, and suitability for real-time embedded execution.

As shown in Table 6, traditional threshold-based methods exhibit limited accuracy and are not able to reliably distinguish between early-stage degradation and normal operating conditions. Classical machine learning techniques, such as k-nearest neighbours and support vector machines, improve classification performance but suffer from increased inference time and limited scalability when deployed on embedded hardware.

The CNN achieves the highest classification accuracy while maintaining low inference latency, making it suitable for real-time execution on the Raspberry Pi platform. These results support the choice of a convolutional neural network as a robust and embedded-friendly solution for PCB integrity monitoring. In order to formalise the signal processing and classification strategy adopted, the sampling index within the time window of length N is formulated in (4).

$$RMS=\sqrt{{\displaystyle \frac{1}{N}}\sum _{n=1}^N{x}^2\left[n\right]}$$

(4)

where x[n] represents the discrete-time electrical signal acquired by the PCB under test after analog-to-digital conversion, with n = 1, …, N.

Each signal segment is processed independently to extract a reduced set of statistical characteristics that describe its amplitude and distribution characteristics.

The mean value μ and variance σ² of the signal are calculated using (5) and (6):

$$\mu ={\displaystyle \frac{1}{N}}\sum _{n=1}^{N}x\left[n\right]$$

(5)

$${\sigma }^2={\displaystyle \frac{1}{N}}\sum _{n=1}^N{\left(x\left[n\right]-\mu \right)}^2$$

(6)

To capture higher-order statistical properties related to asymmetry and impulsivity, the kurtosis k of the signal distribution is evaluated using (7):

$$k={\displaystyle \frac{1}{N{\sigma }^4}}\sum _{n=1}^N{\left(x\left[n\right]-\mu \right)}^4$$

(7)

Finally, the signal energy E within each window is defined by (8).

$$E=\sum _{n=1}^N{x}^2\left[n\right]$$

(8)

The extracted features are combined into a feature vector (9).

$$\mathbf{f}={\left[RMS, \mu , {\sigma }^2, k,E\right]}^T$$

(9)

which represents a compact and informative description of the electrical behaviour of the PCB over the considered time window. This feature vector is provided as input to the classification model. The convolutional neural network implements a non-linear mapping (10).

$$\widehat{y}=\mathcal{F}(\mathbf{f};\mathsf{\theta })$$

(10)

where f is the input feature vector, θ denotes the set of trainable network parameters, and ŷ represents the predicted PCB integrity class. The network is trained to minimise the categorical cross-entropy loss function (11)

$$\mathcal{L}=-\sum _{c=1}^C{y}_c\mathrm{l}\mathrm{o}\mathrm{g}\left({\widehat{y}}_c\right)$$

(11)

where C is the number of integrity classes and y_c is the ground-truth label associated with each input feature vector. The use of statistical features as a common input representation allows the comparison to focus on the intrinsic capabilities of the classification algorithms, rather than on differences arising from input dimensionality or signal representation. The features that have been chosen include the Root Mean Square value (RMS), mean, variance, kurtosis, and energy of the signal. These features have been chosen such that they are able to describe both the amplitude and distribution of the electrical signal well. This is important for reliable computation on the microprocessor. The simulation-based validation was carried out on a set of biomedical PCB samples equipped with the acquisition circuit. Controlled fault conditions were introduced on selected conductive paths to reproduce different integrity states. These included normal operation, gradual increases in contact resistance and the presence of micro-discontinuities that generate abnormal current distributions. The acquisition system continuously monitored the electrical signals associated with each PCB under test, and the corresponding data were labelled according to the imposed fault condition. All experiments were performed under repeatable laboratory conditions to evaluate both the robustness of the hardware chain and the stability of the AI-based classification process. The experimental setup used for system validation is shown in Figure 6.

The PCB under test is instrumented with Hall-effect current sensors connected to the acquisition circuit and to the Raspberry Pi, which performs on-board AI inference and wireless data transmission. Three main PCB integrity conditions were considered in the experiments: (i) normal operation, (ii) degraded conductive paths characterised by increased resistance, and (iii) micro-discontinuities or abnormal current paths. Each condition was represented by a sufficient number of signal segments to ensure statistical significance during training and evaluation of the model. The dataset was divided into training, validation and test subsets. The acquired data were kept separate in the subsets to avoid bias and provide a realistic assessment of the classifier’s generalisation ability.

3.6. Results and Discussion

The performance of the embedded monitoring system was evaluated in terms of signal quality, classification reliability and suitability for real-time operation on a resource-constrained platform. The discussion in this section focuses on the behaviour of the acquisition chain and the AI-based processing pipeline under realistic operating conditions, without anticipating the detailed quantitative results that are presented in Section 4. From a hardware point of view, the analog front-end and the ADC stage proved to be stable and repeatable. The Hall-effect current sensors made it possible to monitor the PCB in a completely non-invasive way, without affecting its normal operating conditions. This aspect turned out to be crucial when looking for early signs of degradation, where very small current changes have to be preserved by the acquisition chain rather than being hidden by noise or drift. On the software side, the AI pipeline running on the embedded platform behaved consistently under the different test conditions. The use of windowed signal segments together with a compact CNN allowed the system to recognise short-time patterns linked to degradation phenomena. Deploying the model on the Raspberry Pi with TensorFlow Lite resulted in low inference latency, which makes the solution suitable for continuous monitoring and near real-time integrity assessment. Another important feature of the presented system is its modular structure. By keeping the sensing interface separate from the digital processing stage, the same software and AI components can be reused across different biomedical devices, while only the front-end circuitry needs to be adapted to the specific application. This design choice increases the scalability and reusability of the solution in a wide range of embedded monitoring scenarios.

The results obtained from the simulation-based validation confirm that the architecture is able to discriminate between different PCB integrity states with high reliability. In particular, the system shows good sensitivity to early-stage degradation, which is essential for predictive maintenance and safety-oriented applications in biomedical electronics. The architecture and processing pipeline described in this section provide the technical basis for the experimental evaluation presented in the following section. Section 4 reports the quantitative results obtained from the validation tests, including classification performance, robustness analysis and discussion of the main outcomes.

4. Simulation-Based Validation and Comparative Results

Future work includes hardware prototyping to validate against real-world parasitic, EMI, and ageing effects, as current models exclude manufacturing tolerances. The performance of the embedded acquisition and monitoring architecture was evaluated using a simulation-based validation methodology. It must be stressed that no physical prototype has yet been implemented at this stage. Therefore, all the results presented in this section refer to simulated electrical behaviour and represent an upper-bound estimation of system performance under idealized conditions. The reported performance metrics reflect the behaviour of the proposed architecture in an ideal and controlled environment. While the simulation framework is designed to reproduce realistic PCB degradation phenomena, the results should be interpreted as an upper bound of the achievable performance. Real hardware implementations may introduce additional non-idealities, such as parasitic effects, electromagnetic interference and component variability, which are not fully captured in the present validation. Since the physical prototype has not yet been fabricated, the validation focuses on the electrical behaviour of the acquisition chain and on the signal processing and AI classification stages, using synthetic and numerically generated datasets that reproduce realistic PCB degradation conditions. The simulation environment was designed to model the electrical dynamics of PCB conductive paths under different integrity states, including normal operation, increased contact resistance and micro-discontinuities. The degradation conditions were generated using numerically implemented electrical models developed in a Python(3.11)-based simulation environment. Increased contact resistance was modelled as a progressive variation in series resistance in PCB traces and solder joints, ranging from nominal values up to +300% of baseline resistance. Micro-discontinuities were modelled as time-varying resistive–capacitive interruptions using controlled switches and intermittent high-impedance paths to emulate crack-induced transient conduction instability. The electrical model also included parasitic capacitances and trace inductances derived from standard PCB approximations.

Noise sources include additive white Gaussian noise and sensor-related noise components, emulating Hall sensor and ADC imperfections. Low-frequency noise components (flicker), typical of Hall-effect sensors and analog front-end electronics, are implicitly approximated within the sensor noise model. A more explicit characterisation of flicker noise and long-term drift effects will be addressed in a future hardware-based validation. The simulated signals refer to a typical low-power biomedical electronic subsystem (e.g., patient monitoring or infusion control circuitry), operating in the sub-kilohertz electrical domain. For each integrity condition, a total of 1500 signal segments were generated, each with a duration of 2 s and sampled at 10 kHz, resulting in a balanced dataset of 4500 labelled time-series signals. The dataset was randomly split into training (70%), validation (15%), and test (15%) subsets using a stratified protocol to preserve class proportions. The generated current and voltage signals were used as inputs to the acquisition chain, including Hall-effect sensing, analog conditioning, analog-to-digital conversion and digital processing. The resulting digital signals were then processed by the AI pipeline described in Section 3. The first set of results concerns the electrical signals generated under the three simulated PCB integrity conditions. Figure 7 shows representative current waveforms at the output of the Hall-effect sensors for (a) normal operation, (b) increased resistance and (c) micro-discontinuities. The simulated signals clearly show that micro-discontinuities do not simply reduce the current amplitude, but introduce structured and repeatable distortions in the waveform. These results confirm that the front-end design preserves degradation-related signal signatures across the entire acquisition chain, from sensing to digital processing.

To assess how well the different integrity conditions can be distinguished, time-domain features were extracted from the simulated signals. As illustrated in Figure 8 and Figure 9, the RMS value and signal energy distributions exhibit a clear separation between normal operation and degraded conditions, with micro-discontinuities occupying an intermediate region. This behaviour is consistent with the physical nature of the defects and suggests that the selected time-domain features are both discriminative and robust under signal variability and noise.

The CNN-based classifier was trained and tested using the simulated dataset described in Section 3.5. The dataset consists of 4500 labelled signal segments (1500 per integrity class), each 2 s long and sampled at 10 kHz. Signals were segmented into windows of 512 samples with 50% overlap. The model was trained using the Adam optimizer with an initial learning rate of 10⁻³, a batch size of 64, and a maximum of 80 epochs with early stopping based on validation loss. During training, additive noise was injected to improve generalization and robustness to signal variability. Data normalization and window-based segmentation were applied to improve convergence and reduce overfitting. The main performance metrics obtained on the test set are reported in

Table 7. Classification performance on the simulated dataset.

Metric	Value (%)
Accuracy	97.4
Precision	97.1
Recall	97.6
F1-score	97.3

.

These results demonstrate that the acquisition and processing chain supports reliable integrity classification, even in the presence of subtle degradation patterns. In addition, the average inference time measured on the Raspberry Pi 4 platform was 3.2 ms per signal window, confirming that the architecture is suitable for near real-time integrity monitoring in embedded biomedical systems.

The classification performance is further analysed through the normalized confusion matrix reported in Figure 10, which illustrates the correspondence between the true PCB integrity states and the classes predicted by the CNN-based classifier.

The low number of misclassifications confirms that the combined acquisition and AI processing chain can reliably discriminate between healthy and degraded operating conditions, even when degradation effects are subtle. Beyond nominal operating conditions, a robustness analysis was carried out to evaluate the stability of the proposed system under signal variability and noise perturbations. In real embedded biomedical electronics, several non-ideal hardware effects may impact both the acquisition chain and the AI-based classification performance. Electromagnetic interference (EMI) can introduce low-frequency noise and spurious components in the measured signals, increasing intra-class variability. Hall-effect sensors are also subject to offset drift and sensitivity variations with temperature, which may bias amplitude-related features such as RMS and energy. Temperature changes can further affect the RC filter components and the ADC reference stability, leading to scale and offset errors. In addition, component tolerances in the analog front-end may cause dispersion in gain and bandwidth across devices. Although these effects are not explicitly modelled in the present simulations, they are expected to reduce the effective signal-to-noise ratio and, consequently, classification margins. Future hardware validation will focus on quantifying and compensating for these phenomena. The results are summarized in

Table 8. Robustness evaluation of the embedded electrical monitoring system.

Noise Level (SNR)	Accuracy (%)	F1-Score (%)
30 dB	97.2	97.1
20 dB	96.5	96.2
10 dB	94.1	93.8

.

This robustness analysis reveals that the embedded electrical monitoring system maintains high classification performance even under significant noise and signal variability, which are expected in real biomedical operating environments. To contextualize these results within the authors’ previous research and to highlight the evolution of the methodologies, a comparative analysis with earlier AI-based PCB monitoring approaches is reported in

Table 9. Comparative overview of AI-based PCB monitoring approaches.

Criterion/Feature	Ref. [49]	Ref. [54]	Electrical AI-Based System
Sensing modality	Infrared thermography	Infrared thermography	Electrical current/voltage signals
Sensor/Hardware	IR camera + FEM model	FLIR P660 camera	Hall sensors + ADC + SBC
Physical phenomenon analysed	Heat diffusion	Heat diffusion + statistical effects	Electrical signal propagation
Acquisition type	External instrumentation	External instrumentation	Fully embedded
Signal domain	Spatial (image-based)	Spatial (image-based)	Temporal (waveform-based)
Feature extraction	Thermal gradients, isotherms	Diffusion + Taguchi descriptors	RMS, energy, temporal gradients
AI pipeline	Segmentation + MLP	U-Net + MLP + ANOVA	Signal segmentation + CNN/ML
Statistical validation	No	Yes (Taguchi + ANOVA)	Simulation-based
Classification accuracy	~94–96%	~95–97%	>97% (simulated)
Environmental sensitivity	High	High	Low
Power consumption	High	High	Low (<150 mW acquisition stage)
Suitability for continuous monitoring	Limited	Limited	High
Integration into medical device	Difficult	Difficult	Native/embedded
Real-time capability	Limited	Limited	Near real-time

.

Table 9 summarizes and compares the three AI-based PCB monitoring methodologies developed by the authors over time, highlighting the methodological continuity and the progressive evolution of the sensing and integration strategies. The first two approaches rely on infrared thermography and analyse thermal diffusion patterns to detect defects and degradation phenomena in electronic boards. Although effective from a diagnostic point of view, both systems require external instrumentation and are therefore mainly suited for off-line inspection or laboratory environments. The presented system represents the natural evolution of this research line towards an embedded and fully integrated solution. While preserving the same conceptual AI pipeline—based on signal segmentation, physics-aware feature extraction and supervised classification—the sensing domain is shifted from thermal images to electrical waveforms acquired directly from the PCB. This transition enables continuous, low-power and real-time monitoring of PCB integrity inside biomedical devices, without the need for external equipment and without sensitivity to environmental thermal conditions. The embedded electrical solution offers a more scalable and deployable alternative for medical applications. It should be noted that the present validation is entirely simulation-based.

Although the electrical models were designed to reproduce realistic degradation mechanisms, the current results do not yet account for parasitic effects, manufacturing tolerances, ageing variability, or electromagnetic interference typical of real biomedical devices. These aspects will be addressed in future experimental implementations and hardware validation campaigns.

5. Conclusions and Future Work

This study focuses on the design and simulation-based validation of an embedded monitoring solution for non-invasive PCB integrity assessment in biomedical devices.

The solution operates directly on electrical signals acquired from the board through Hall-effect sensing and analog-to-digital processing, enabling in situ diagnostics without the need for external inspection tools. This enables continuous, low-power monitoring without interfering with the normal operation of the device, which is a critical requirement in medical applications. A central contribution of this work is the shift from external thermographic inspection to embedded electrical waveform analysis.

By exploiting the natural propagation of current in PCB conductive paths, integrity-related information can be extracted directly during operation, enabling diagnostics compatible with real-time and long-term monitoring. The resulting methodology integrates signal segmentation, physics-inspired feature extraction, and CNN-based classification to identify degradation patterns associated with increased contact resistance and micro-discontinuities. The simulation results indicate that the acquisition chain preserves diagnostically relevant signal variability and that the classifier can reliably discriminate between healthy and degraded PCB states. However, since the present validation is entirely simulation-based, the reported performance must be interpreted as indicative of potential capability rather than demonstrated real-world readiness.

Experimental implementation and hardware testing will be required to fully assess robustness, stability, and compliance with biomedical electromagnetic compatibility constraints. Nevertheless, the present validation is entirely simulation-based and does not yet account for layout parasitics, manufacturing tolerances, electromagnetic interference and long-term ageing effects. These aspects will be addressed in future work through the physical implementation of the acquisition circuit and experimental testing on real biomedical PCBs under realistic operating conditions. An accuracy above 97% indicates the effectiveness of combining a tailored front-end with data-driven signal processing to detect subtle integrity variations that are not observable through simple amplitude- or threshold-based methods. Beyond numerical performance metrics, the results support a clear physical interpretation: degradation in conductive paths generates structured and repeatable distortions in the electrical waveforms, and these signatures can be consistently learned and exploited by the classifier. This suggests that temporal electrical features carry diagnostic information comparable to that obtained from thermal diffusion patterns, while being directly accessible through embedded and low-power electronics.

Future studies will focus on the physical implementation of the acquisition circuit and its simulation-based validation on real biomedical printed circuit boards.

In particular, a hardware prototype will be fabricated and experimentally validated on real biomedical PCBs to assess electromagnetic compatibility, thermal stability, long-term drift, and classification robustness under realistic operating conditions.

Key aspects will include miniaturization, power optimization, and electromagnetic compatibility, in line with medical device constraints. A hardware-based validation campaign will allow for a quantitative assessment of classification stability in the presence of real noise sources and parasitic effects.

Only after this experimental phase can the architecture be considered implementable in biomedical applications. Further developments will address the deployment of the AI pipeline on low-power edge platforms to enable autonomous, real-time integrity monitoring directly within the device. Finally, the methodology will be extended to a broader range of biomedical electronics—such as infusion systems, patient monitors, and implantable devices—to assess robustness, scalability, and generalizability across heterogeneous operating conditions. At the current stage, the proposed system should be regarded as a validated design framework rather than a fully demonstrated deployable product. Hardware implementation and experimental testing are the next essential steps to confirm practical applicability.