The research develops a hybrid model for the helicopter turboshaft engine gas temperature sensor readings predicting and correcting, which combines machine learning algorithms and a cyber-physical infrastructure (CPS) to improve the monitoring accuracy and adaptive response to changes in the engine operating mode. into account both the measurement noise stochastic components and the therocouple slow thermodynamic sh -processing, which smooths out high-frequency by the network through weighted fusion with a confidence coefficient k, which ensures the accumulated drift is almost completely eliminated while maintaining performance. The CPS infrastructure is used, which includes a secure communication channel shown that the telemetry delivery time grows linearly fro n amount of S at S experiments have demonstrated the proposed architecture's high efficiency: the adjusted temperature cur with traditional LSTM, RBF network, and three-layer perceptron confirmed the proposed solution's superiority in accuracy (0.991), recall (0.985), and F1 measure (0.986) terms. The obtained results show that the hybrid approach provides reliable engine health diagnostics and can form the basis for developing a ICyberPhyS 5: 2nd International Workshop on Intelligent & CyberPhysical Systems, July 04, 2025, Khmelnytskyi, Ukraine 1 Corresponding author. These authors contributed equally.
Modern helicopter turboshaft engines (TE) place high demands on the operating process, monitoring key parameters for accuracy and efficiency [1]. One of the most critical indicators is the gas temperature in front of the compressor turbine, which directly affects the engine efficiency, its components' service life, and flight safety [2]. Traditional thermocouple sensors [3, 4], despite their production technologies' development, are subject to errors, aging, and an aggressive environmental influence, which entails measurement distortion and a decrease in the engine's actual state assessment reliability. readings is becoming especially relevant. On the one hand, neural networks [5] are capable of taking al flight and test data, forming more accurate estimates of probable temperature values. On the other hand, the CPS infrastructure (cyberphysical systems) use [6] ensures continuous data exchange between physical sensors and computing modules in real time, which allows not only to correct current readings but also to adaptively respond to changes in the engine operating mode.
The neural network technologies and CPS infrastructure integration are becoming critically important for increasing helicopter TE reliability: by continuously combining analytical models and real data, it is possible to promptly identify deviations in the emergencies, and adaptively adjust engine operating parameters. The introduction of such hybrid solutions not only helps reduce the unscheduled repairs and extend the components' service life but also increases the aircraft's overall cost-effectiveness.
Thus, the hybrid model development for predicting and correcting temperature readings directly meets modern requirements for helicopter TE safety, efficiency, and intellectualization control.
In studies [2, 4, 7], the traditional approach to modeling and assessing the gas temperature in helicopter GTEs is based on the heat exchange and gas flow dynamics physical laws. Such models (gas-dynamic, thermophysical) provide an approximation in steady-state modes at the 90...92 % level but demonstrate significant errors (up to 5 %) under transient conditions and due to the thermocouples' own imperfection (drift due to corrosion, thermal shock, etc.). At the same time, classical error compensation algorithms [8] usually rely on linear calibration and do not take into account complex nonlinear interactions, which leads to insufficient accuracy (up to 80%) and a slow response to changes in the engine operating mode.
With the machine learning methods, studies have emerged on the use of neural networks to correct temperature sensor data. In particular, multilayer perceptrons [9] and recurrent neural networks [10] trained on historical data from takeoff, idle, and cruise modes are considered. Such complex nonlinearities, but they often have difficulty generalizing beyond the training dataset and require large labeled data sets, which acquisition in aviation conditions is associated with high costs.
In parallel, the cyber-physical systems (CPS) direction for aviation equipment is actively developing: distributed architectures are being created that are capable of combining telemetry streams from various sensors, onboard computing modules, and headquarters storages in real time [11, 12]. The CPS infrastructure ensures reliable and scalable data transmission, flexible deployment of analysis algorithms (edge computing on board and cloud analytics), and a high degree. However, existing CPS solutions [11,
Research combining machine learning and CPS is still mostly conceptual or demonstrated in laboratory setups: helicopter TE digital twins are created with embedded neural network blocks to predict failures and modes of destabilization [13]. However, most of them focus on large-sized stationary installations (thermal power plants [14], gas pumping stations [15]) and are poorly adapted to the helicopter TE-specific features of limited computing resources of onboard computers, high vibration levels and thermal effects, and strict requirements for data transmission and processing delays.
Among the unanswered questions, the following are particularly relevant: how to automatically conditions change; how to optimally distribute computing resources between onboard and ground systems (edge-cloud), taking into account delays and transmitted data amounts; and how to guarantee the reliability and the CPS infrastructure cybersecurity, ensuring the telemetry protection and model updates even in combat or emergency situations.
Thus, the hybrid model combining the development of a neural network (to account for complex nonlinearities and predict error-free values) and a CPS infrastructure (to collect, transmit, and flexibly process data in real time) allows us to close these gaps. Such a system will make it possible to automatically compensate for thermocouple drift, continuously adapt to new modes, and perform the helicopter TE gas temperature monitoring accuracy and reliability.
This study proposes a hybrid model structural diagram for the helicopter TE gas temperature sensor readings predicting and correcting using neural networks and a CPS infrastructure (Figure 1), predicting adaptation, which ensures resistance to sensor degradation and adaptability to operating conditions.
Gas temperature sensor
preprocessing module (Edge) filters, normalizes, and aggregates the data. It removes anomalies and determines the signal's validity. It prepares inputs for intelligent algorithms. The sensor drift detection module (Edge) compares the current signals with the reference profiles, identifies deviations caused by thermocouple aging or environmental influences, and activates correction or transmits an alarm. The temperature prediction neural network (Edge) uses historical and current combines the measured value, the neural network predicting, and the drift estimate. It produces a final corrected temperature value suitable for use by the engine control system. The output to the engine management system transmits the corrected temperature value to the engine regulation and monitoring system. telemetry between on-board and cloud modules, cloud analytics with data storage for long-term analysis, trend detection and fault diagnostics, and a module for automatic retraining of the neural network on new data with adaptation to new operating modes and subsequent delivery of the updated model back to the on-board computers.
Based on [16], in this study the noise is represented as Gaussian: ymeas (t ) = ytrue (t ) + d (t ) + n(t ).
n(t )
N (0, n2 ).
The drift d(t) is described, for example, by a stochastic diffusion process [17] in the form: dd (t ) dt = d (t ) + + (t ), (t )
N (0,2 ).
Preprocessing (filtering) includes low-pass approximation using an RC filter [18], represented by the expression: dz (t ) dt
+ z (t ) = ymeas (t ), z (0) = ymeas (0), where z(t) is the smoothed signal.
In the drift detection module, a comparison with the reference profile yref(t) is performed as: The study developed a hybrid model mathematical formulation, reflecting all the diagram key blocks (Figure 1). It is assumed that ytrue(t) is the true gas temperature at time t, ymeas(t) is the d(t) is the sensor drift, n(t) is the measurement noise, d (t ) and ynn (t ) are the drift and true temperature estimates issued by the corresponding modules.
The sensor reading model is described by the expression: ed (t ) = z (t ) − yref (t ).
ed (t ),if ed (t ) , d (t ) =
d (t − t ),otherwise.
x(t ) = z (t ), z (t ), P(t ),... , In this case, the threshold logic is defined as:
The neural network predicts the gas temperature according to the feature vector: where P(t) are other engine parameters.
Based on [19, 20], in this study, the prediction neural network is implemented by a recurrent model (LSTM) as:
The neural network is trained by minimizing the loss function: In the correction and merge module, the final score is determined as: h(t ) = LSTM(x(t ),h(t − t ), W),
T 2 L ( W, V,b) = ( ynn (t ) − ytrue (t )) dt + W 2 .
0
neural
network
ycorr (t ) = ynn (t ) + d (t ) + k ( z (t ) − ynn (t )),
drift
local adaptive correction
where k ∈ [0, 1] is the confidence coefficient for the raw data.
(
SSL/TLS x(t ), z (t ), d (t ), ynn (t ) → CloudStorage, and the channel throughput and delay are described by the equation:
Latency (t ) = L0 +
S (t )
B
+ (t ),
D − i−N:i , i Di = i i−N:i
M, where L0 is the basic protocol delay, S(t) is the amount of transmitted packets during the interval t, B is the channel throughput, and (t) is the delay's random component.
Received on-board data D = x(t ), z (t ), d (t ), ynn (t ) are aggregated into a buffer of size N and normalized over a sliding window: where i N:i and i N:i are the mean and standard deviation of the last N points.
Tretrain hours in the cloud, a retraining optimization problem of the form is solved:
Wm,Vin,b M1 iM=1 ynn ti ; W*, U*,V*,b*,by − ycorr (ti ) 2 + W 2 + R (W) , (11) (12) (13) (14) (15) (16) (17) new data, and is the L2where R(W) is a regularizer to prevent overfitting in changing modes, neural network parameters, 0 < regularization coefficient.
The weight update is performed using stochastic gradient descent as:
Wk+1 = Wk − k W L, kli→m k = 0, k = .
k
To quickly respond to new conditions, a forgetting factor is introduced into the loss function, represented as:
M 2 L = M −i ( ynn (ti ) − ycorr (ti )) , i=1
L = Lprev − L* th , where 0 < < 1 specifies the new data priority degree.
After convergence, {W*, V*, b} are sent back on board via the telemetry channel. The transmission condition is represented as: where th is the minimum gain in model quality that requires updating.
Thus, the CPS infrastructure use not only ensures reliable bidirectional data exchange but also implements the neural network adaptive retraining full cycle, taking into account the new measurements and communication channel limitations priority. 3.3. The neural network predicting model development To develop a neural network predicting model (Figure 2), it is assumed that the input vector consists of the filtered temperature value z(t) and its derivative z (t ) , the latest drift estimate d (t − t ) , and the remaining engine operating parameters vector P(t). Thus,
z (t ) xt = Pz((tt)) m.
d (t − t ) ct– ht– xt
Whf Wxf Whi Wxi Whc Wxc Whg Wxg Who Wxo
(sig)
(sig)
(sig) Output gate (sig) ft it ct gt ot tanh ct ht it = (Wi xt + Ui ht−1 + bi ), ft = (Wf xt + U f ht−1 + bf ), ot = (Wo xt + Uo ht−1 + bo ), ct = tanh (Wc xt + Uc ht−1 + bc ), gt = (Wg xt + Ug ht−1 + bg ),
W* nm , U* nn , b* n .
To account for and compensate for gradual shifts in the thermocouple characteristics, a built-in Based on the current discrepancy between the filtered signal and the previous drift estimate, it generates a correction value for updating the memory state and is described by the expression: (18) (19) (20) (21) (22) (23) where ⊙ is the element-wise multiplication, and δt adds to the cell memory information about the current discrepancy between the filtered signal and the drift estimate.
Thus, the output vector is defined as: ct = ft ct−1 + it
ct + t , ht = ot
tanh (ct ), ynn (t ) = V h(t ) + by , k+1 = k − L(k ). where V 1n , by
.
The weights are updated using stochastic gradient descent:
The resulting architecture is an extended LSTM cell that takes as input the preprocessed z(t), its derivative, engine parameters, and a previous drift estimate, within which the standard input, forget, which the combined memory state produces an output hidden vector that is transformed by a linear layer into a predicted gas temperature value.
In this study, the developed hybrid model structural diagram for predicting and correcting gas temperature sensor readings (see Figure 1) is implemented in the Matlab R2014b software environment (Figure 3). gas temperature sensor) reproduces the ytrue(t), sensor drift d(t), and noise component n(t) sum. It includes a source ytrue is an organic sinusoidal (or tabular) signal reflecting the real temperature dynamics; a drift module implemented through an integrator receiving low-amplitude white noise at the input (generated by the Random Number block), which simulates a slowly increasing drift; and a measuring noise generator n(t) in the Gaussian random ymeas(t) = ytrue(t) + d(t) + n(t), which is then fed for filtering and subsequent processing.
The Filtering/Preprocessing block takes the raw reading ymeas(t) as input and applies a continuous 1 RC filter with the transfer function , where is selected based on the smoothing required s + 1 degree, thereby filtering out high-frequency noise and outputting z(t in parallel to the filter output is the Derivative block, which calculates z (t ) for further use in the predictive model. If necessary, a normalization or scaling block can be added on the RC filter top, but the basic circuit is limited to a series connection of the Transfer Fcn (Numerator = [1], Denominator = [ , 1]) and Derivative blocks, which ensures smoothing of the input signal and its derivative formation for the following processing stages.
The Drift Detection block compares the smoothed value z(t) with a preset reference profile yn · en(t), calculating the difference draw(t) = z(t) yn · en(t) and its absolute value e(d(t)) = |draw(t)|; then, via the Compare To Constant block, it is checked whether e(d(t)) exceeds a specified threshold , and if so, the logical flag Drift_Flag = 1 is set, signaling the drift presence, and the drift estimate dest(t) = draw(t) is passed on for correction algorithms.
The neural network temperature prediction (NN Prediction) block takes as input a feature vector including the current smoothed value z(t), its derivative z (t ) , drift estimate d (t ) , and additional engine parameters P(t), and then loads a pre-trained LSTM network (netLSTM) via a custom MATLAB Function and performs a one-step prediction ytrue (t ) = predict (netLSTM, z (t ), z (t ), P (t ),...) ). Inside the MATLAB Function, the first call loads the trainedNet.mat file containing the LSTM weights and architecture using a persistent variable, calculated, which is output as ypred, providing adaptive prediction taking into account current and historical data.
The Correction & Fusion block combines the raw readings ymeas, the predicting ypred and the drift estimate d , computing the corrected predicting ycporerrd = ypred − d , and then mixes it with the raw data via weighted addition: the signal ymeas is multiplied by the confidence coefficient k (Gain k), then the difference ycporerrd k k), and both results are summed (Sum), forming the final value ycorr = k ymeas + (1− k ) ( ypred − d ) , which is output for further use by the control system. ycorr signal generated in the previous step as input and transmits it to the external controller via the output port (Outport), providing integration with the engine control system; in addition, the Scope block is connected in parallel for the ycorr dynamics visual monitoring during modeling, and if necessary, the Drift_Flag logical signal from the drift detection module can be output to a separate Outport (for example, "Drift_Flag_Out") to indicate the sensor alarming state. 4.2. The input data analysis and preprocessing The computational experiment used data on the gas temperature in front of the compressor turbine ymeas(t) of the TV3-117 engine, recorded by the standard Mi-8MTV helicopter sensor (14 dual thermocouples T-102 [5, 19]) in the nominal operating mode. The tests were carried out at the 2500 meters altitude, measuring every 0.25 seconds for 320 seconds. According to the data in Figure 4, the maximum gas temperature reached 1140 K.
The data on the gas temperature, obtained during flight tests of the Mi-8MTV helicopter using on-board monitoring, were cleared of interference and anomalies and then transformed into timeordered series [21]. To unify the scales, z-normalization [22] was used: z ( ymeas )i = ym(ie)as − 1 N ym(ie)as
N i=1 1 N ym(ie)as − 1 N i=1
N ym(ie)as N i=1
To assess the training dataset (see Table 1) representativeness, the k-means cluster analysis method was used [24]. The training dataset was randomly divided into training and test datasets in a ratio of 2:1 (67% are 858 elements and 33% are 422 elements). The training dataset clustering identified eight groups (classes I VIII), which confirms the two datasets' structural similarity (Figure 5). Based on this, the gas temperature dataset amounts were determined: out of a total training dataset of 1280 elements, 858 (67%) constitute the control dataset, and 422 (33%) constitute the test dataset. 4.3. The computational experiment results For reproducible LSTM training, the Adam optimizer was used with learning rate=10 3 (can be reduced to 10 4 if necessary), batch size=32 64, and seeds were fixed (numpy.random.seed(42), torch.manual_seed(42)).
The deterministic mode was enabled (torch.backends.cudnn.deterministic=True, torch.backends.cudnn.benchmark=False), and a fixed number of epochs and a training rate reduction scheme were set (ReduceLROnPlateau with a 0.1 factor and patience=5).
As the computational experiment part, diagrams corrected gas temperature readings (Figure 6), the sensor drift estimate and its predicted estimate (Figure 7), the data transmission delay and the telemetry amount dependence (Figure 8).
In Figure 6, the gas temperature ranges from 1080 K to 1150 K, with the blue thin curve showing the raw data, which fluctuates with about 30 K amplitude and contains significant noise and drift, while the red curve shows the corrected readings: they have a smoother shape, the drift trend is removed, and the noise is smoothed out, allowing us to observe the true temperature dynamics without sensor artifacts.
According to Figure 7, the thermocouple drift in degrees Celsius ranges from about 0 to 2 °C. The true drift is represented by the red solid line and consists of a linear trend increasing at about a 0.02 °C/s rate, superimposed on a sine wave with a 0.5 °C amplitude and about a 62-second period, plus some random noise with a 0.05 °C variance. The blue dotted line is phased out with the true drift by about 1 second and has a slightly smaller sine wave amplitude (about 0.45 °C) due to smoothing, and the noise in its values is limited to about 0.03 °C variance. Thus, the dotted line accurately follows the general growth and the drift oscillation, but with a small delay and reduced sawtooth peaks.
According to Figure 8, the transmitted telemetry amount covers the range from 10 to 2000 kbps, and the measured transmission delay varies from approximately 50 ms to 1400 ms. ms (base L0 = 50 ms plus minor random noise), then with an increase to 500 kbps, the delay increases linearly to approximately 350 ms, and by 2000 kbps, it is already approaching 1400 ms.
The curve shows small fluctuations of ±10 ms due to random components of (t), but the overall S (kilobits), the transmission time increases almost proportionally, showing the channel bandwidth influence.
Despite random fluctuations in , the overall curve shows a linear increase in delay as the data amount increases: first, at S < 100 kbps, a 60 70 ms plateau is visible, then a predictable rise to 400 ms for S 2000 kbps maximum load.
This clearly illustrates the channel capacity limitation and the basic L0 delay impact. 4.4. The results obtained effectiveness evaluation To evaluate the predicted gas temperature values obtained, the various neural network architectures (traditional LSTM network, three-layer perceptron, and neural network on radial basis functions) were compared using accuracy, precision, recall, and F1-score metrics (Table 3), where: Accuracy =
TP + TN TP + TN + FP + FN , Precision =
TP TP + FP
TP TP + FN , F1-score = 2
Precision Recall Precision + Recall
(25)
The comparative analysis results of the four models demonstrate the modified LSTM network's advantage over the others: its accuracy = 0.991 exceeds the traditional LSTM (0.979), the radial basis function network (0.951), and the three-layer perceptron (0.922) indicator; the classification accuracy (precision) of the proposed LSTM is 0.987, while that of the traditional LSTM is 0.973, RBF is 0.944, and that of the perceptron is 0.909; the recall indicator (0.985) for the proposed model also exceeds the traditional LSTM (0.962), RBF (0.934), and the perceptron (0.883).
Finally, the F1-score of the proposed LSTM (0.986) is the highest compared to 0.967 of the traditional LSTM, 0.939 of RBF, and 0.896 of the three-layer perceptron, indicating a consistently higher balance between precision and recall in the proposed architecture.
The neural network model's stability to external interference was analyzed. For this aim, additive interference in the white noise with zero mathematical expectation i = 0.025 form was introduced, which corresponds to 2.5 % [25]. Table 4 shows the calculating results, the standard deviation, and the absolute error without noise and in the white noise when applying the above neural network architectures to the problem of predicting gas temperatures.
From Table 4 it can be seen that the modified LSTM network demonstrates the best performance among the considered architectures: without noise its standard deviation is 0.008, and with noise it is only 0.135, while the absolute error without noise is 0.9%, and with noise it is 1.623%; the traditional LSTM is inferior with deviations of 0.021 (without noise) and 0.216 (with noise) and errors of 2.1% and 3.779%, respectively; the network on radial basis functions shows higher deviations of 0.063/0.335 and errors of 4.9%/8.117%.
In turn, the three-layer perceptron has the worst results, with deviations of 0.119 (without noise) and 0.519 (with noise) and absolute errors of 7.8% and 14.035%, which emphasizes the stability and noise resistance of the proposed LSTM architecture.
A hybrid model for the helicopter TE gas temperature sensor readings predicting and correcting
block diagram has been developed using a neural network and the CPS infrastructure (see Figure 1).
predicting
raw data ymeas(t) are collected, which contain the true
temperature value ytrue(t), drift d(t), and noise n(t), according to model (
The noise is approximated by a Gaussian process (
The CPS infrastructure use ensures secure two-way exchange between the edge and cloud
modules, where data storage, analytics, and the neural network automatic retraining on new data
with regular resending of the updated model on board are implemented (see equations (11) (17)).
(16), while the weights are updated using the stochastic gradient descent method (15), and the
resending occurs when a specified thr
adaptive, drift- and noise-resistant temperature predicting system capable of self-correction and
realtime learning.
readings ymeas(t), including the true value ytrue(t), drift d(t), and noise n(t) (
Figure 8 shows that the telemetry transmission delay in the CPS infrastructure depends on the from 100 to 2000 Kbit, the delay increases from ~100 to ~1400 ms. This indicates the need to optimize the computations between the edge and cloud modules (14) (17).
Table 3 demonstrates that the modified LSTM with a drift gate (20), (21) outperforms traditional
LSTM, RBF networks, and perceptrons in accuracy, recall, and F1-score due to adaptive weight
updating (16), (23) and drift accounting (
As a result, the integration of the extended LSTM cell (19) the CPS infrastructure (11) (17) ensures high accuracy and reliability in predicting and correcting the helicopter TE gas temperature.
Along with the significant results of the study, the following limitations should be noted, presented in Table 5. Table 6 presents prospects for further research.
Description The model was trained and tested on data obtained from TV3117 engines on the Mi-8MTV helicopter in the nominal operating mode and at a 2500 meters fixed altitude. Under other 2 3 4 types and operating
modes Dependence on reference profile and training data quality Limitations on computational resources and latencies in the
CPS infrastructure Simplified a
priori assumptions about the noise and drift nature conditions (different engines, changed flight modes, extreme temperatures, high loads), the noise, drift, and gas temperature dynamics parameters may differ significantly, which will require additional training or adaptation of the neural network architecture and preprocessing algorithms.
The drift detection module requires a predefined reference profile yref(t) to function correctly. If the reference profile does not accurately reflect the actual engine dynamics (e.g., due to component wear or incorrect thermocouple calibration), the algorithm might misinterpret natural temperature variations as drift or, conversely, miss real shifts in readings. Training an LSTM network requires a high-quality labeled data; insufficient or biased representation in the training dataset limits the prediction accuracy outside the training domain.
On board the helicopter, computing resources (CPU, memory) are limited. A complex LSTM model with a "drift gate" and additional calculations for drift estimation may require more resources than are available in real time, which leads to an increase in the prediction and correction delay. Considering that the model shows that the data transmission delay via the CPS channel can reach more than 1 second with a load of about 2000 Kbps, in extreme scenarios (e.g., in emergency modes), the updated model may simply not have time to be delivered on board in time, and the readings correction may be performed with a delay, which reduces its practical effectiveness.
In the study, the measurement noise n(t) is represented as a Gaussian process, and the thermocouple drift d(t) is modeled by a stochastic diffusion process with a specific sinusoidal superposition. In real engine conditions, more complex and nonstationary sources of errors are possible: nonlinear effects of corrosion, thermal shocks, microcracks in the thermocouple, and electromagnetic interference. Accordingly, the algorithm may underestimate or misinterpret anomalies that go beyond the adopted a priori models, which reduces the estimates' accuracy in real field tests.
Action Adaptation and validation of the proposed hybrid drift detection model for other gas turbine engines and for different flight regimes.
Study of the altitude, temperature, and load variations' influence on predicting accuracy and methods development for the model parameters' automatic reconfiguration.
Develop algorithms for automatic generation and adaptive updating of the reference profile (yref) based on current operational data to reduce the impact of outdated or inaccurate calibration curves.
Expanding the training set through field testing and simulating scenarios with different sensor wear states; implementing data augmentation methods to model unexpected temperature fluctuations.
Research into more compact neural network architectures (e.g., lightweight LSTM, TCN, or Transformer-Lite) and weight quantization to reduce memory and CPU costs.
Development of a prototype taking into account the on-board -board
Optimizing computational complexity and implementing it on the onboard computing
system Modeling and accounting for more complex nonlinear sources of noise and drift
integrating it with the general monitoring system data exchange via the CPS channel.
The stochastic models' construction takes into account nonstationary corrosion processes, transient thermal effects (thermal shocks), electromagnetic interference, and destruction of thermocouple insulation.
To investigate methods for anomaly adaptive detection that are not specific to predefined sinusoidal patterns using hybrid approaches (e.g., combining LSTM with variational autoencoders or GANs).
Integrating Develop online learning methods that enable the model to feedback and self- adjust its parameters based on continuously incoming data learning on real measurements and maintenance results. mechanisms in The adaptive algorithm implementation automatically real time reconfigures itself after maintenance or when systematic deviations are detected, reducing the need for manual reconfiguration.
Developing the A single platform creation combining engine thermodynamic concept of a models, drift detection algorithms, and other diagnostic modules (e.g., vibration analysis, wear assessment of parts).
Ways to investigate transmitting aggregated data to groundbased maintenance centers and using cloud computing resources for deeper analytics and unit life prediction.
The helicopter turboshaft engines' gas temperature sensor readings hybrid model for predicting and correcting structural diagrams has been developed and implemented. It combines an extended LSTM el takes into account stochastic descriptions of noise and thermocouple drift, applies an RC filter for preliminary processing, and a builtaccount the accumulated bias.
The CPS infrastructure use ensures encrypted twoguarantees the model adaptability to changes in engine operating modes.
As the computational experiments result, the corrected readings demonstrate effective drift removal and noise spike smoothing compared to the raw data, which confirms the preprocessing algorithms' (RC filter) correctness and correction via weighted fusion. The drift estimation accurately reproduces the linear trend and the true drift sinusoidal oscillations with about 1 second delay and reduced amplitude (~0.45 °C), which proves the proposed detection logic's high sensitivity and adequacy.
The neural network architectures comparative analysis showed that the modified LSTM with a -layer perceptron in all metrics: the accuracy (0.991), recall (0.985), and F1-measure (0.986) of the proposed model are LSTM retains low RMSE (0.135) and absolute error (1.623%), which confirms its resistance to low external interference.
The CPS channel characteristics study showed that the data transmission delay increases linearly emphasizes the need to balance computations between onboard modules and cloud analytics. The th is exceeded, but the obtained latencies indicate potential limitations under extreme loads, requiring further optimization of transmission protocols and the neural network architecture simplification.
0123U104884.
The research was carried out with the grants support of the National Research Fund of Ukraine: project registration number 273/0024 from 1/08/2024 (2023.04/0024), and "Information system development for automatic detection of misinformation sources and inauthentic behaviour of chat users", project registration number 187/0012 from 1/08/2024 (2023.04/0012).
The author(s) have not employed any Generative AI tools.