Research on cotton-storage monitoring shows that wireless sensing can supply real-time temperature and humidity information. Zhang et al. [3] developed an RFID-based intelligent inspection and monitoring system for cotton bale storage that uses wireless temperature-humidity sensing and a handheld terminal; in their field tests the positioning error was below 6.7%, the relative temperature error below 8% and the humidity error below 7%. That work establishes the importance of distributed information acquisition, but its objective differs from the present one: the system proposed here additionally closes the loop through adaptive ventilation and mobile actuation.

From the control viewpoint, time-varying process parameters make fixed-coefficient models vulnerable to mismatch. RLS identification provides an online mechanism for tracking parameter changes [5][7], and adaptive state-feedback control built on RLS has been proposed for time-varying systems [8]. Recent thermal-control research combines adaptive identification with predictive control [9]. Two practical difficulties of RLS with forgetting are well known: loss of excitation, which causes the covariance matrix to grow without bound, and estimates that leave the physically admissible region under noise [6][7]. Both are relevant to the present object, because a contactor-controlled fan delivers excitation only while it is running, and both are addressed explicitly in Section 4.2.

The state of the stack, the external disturbances and the control actions are described by the vectors

where T is the temperature, W the humidity-related state, ρ the cotton density, v_a the air velocity in the ventilation channel and ΔP the pressure difference; T_a and RH_a are the ambient temperature and relative humidity, and P_y, v_w and I_s denote precipitation, wind speed and solar irradiance; u_f ∈ {0, 1} is the state of the fan contactor and u_m ∈ {−1, 0, 1} is the motion command of the platform (reverse, stop, forward). The object and its observation by N = 8 sensor nodes are written in the general form

where θ(t) is the vector of slowly varying parameters, T_s the sampling period and ν_i the measurement noise of node i. Only local temperatures and relative humidities are measured; the remaining components of x act on the measured channels through the parameters θ, which is the reason for identifying them online.

The control problem is to construct a supervisory policy u(k) that, whenever a valid node crosses the upper threshold T_on or the risk index exceeds its decision boundary, returns all valid node temperatures below T_on − ΔT_h in finite time, subject to three constraints: mutual exclusion of motion and ventilation, u_f·|u_m| = 0; minimum ON and OFF durations of the fan contactor; and transfer to a safe state on any fault. Because the prototype uses a contactor-controlled fan, ventilation is manipulated through ON/OFF durations and state transitions; continuous fan-speed adaptation by means of a variable-frequency drive (VFD) is identified as a future extension.

The dominant cooling behaviour of the zone observed by node i is described by a first-order heat balance in which ventilation increases the rate of heat exchange with the ambient air:

where k_0,i is the natural (fan-off) heat-exchange coefficient, k_v,i the additional coefficient created by forced ventilation, γ_i the sensitivity to ambient humidity, and q_i a lumped term for internal heat generation and unmodelled exchange with neighbouring zones. For a constant ventilation state over one sampling period, Equation (3) leads to the discrete regression form

in which ε_i is the modelling error. Equation (4) is a regime-wise linear approximation of Equation (3). In the ventilated regime (u_f = 1), in which identification is performed, a_i = exp(−k_c,iT_s) with k_c,i = k_0,i + k_v,i, so that the effective cooling coefficient and the corresponding time constant follow from the estimate of a_i:

The coefficient k_c,i is not assumed constant, because cotton density, channel resistance, external conditions and fan performance change over time. Its current estimate is therefore both a model parameter and a diagnostic indicator of ventilation effectiveness.

For each channel the regression vector is φ ( k ) = [ T ( k ) , u f ( k ) , T a ( k ) , R H a ( k ) , 1 ] T , the parameter vector is θ = [ a , b , c , d , e ] T and the output is y(k) = T(k + 1); the node index is omitted for brevity. The one-step prediction and the prediction error are

and the RLS update with forgetting factor λ is [5][6]

where K(k) is the gain vector and P(k) the covariance matrix. The forgetting factor determines the effective memory of the estimator,

The simulation uses λ = 0.995 as an initial value, which gives N_λ = 200 samples, that is, a memory of 200 - 400 s at the sampling period of 1 - 2 s. This is short in comparison with the rate at which the properties of the stack change and long in comparison with the measurement noise. In the simulations reported in Section 6 the sampling period of the thermal channels is fixed at T_s = 2 s, so that the memory of the estimator is N_λT_s = 400 s ≈ 6.7 min.

Updating is conditional. Equations (7)-(9) are executed for node i only when the indicator

equals one, that is, when the fan is running, the data of the node are valid (v_i = 1, Section 4.3) and a scalar excitation measure ξ_i (in the simulation, the magnitude of the internal-ambient temperature difference, ξ_i(k) = |T_i(k) − T_a(k)|, which determines how informative a sample is about a_i) exceeds the threshold ξ_min. Otherwise, the last reliable estimate is retained, together with the covariance matrix, which prevents the covariance wind-up that exponential forgetting produces in the absence of excitation [6]. Two further remarks apply. First, since identification is restricted to u_f = 1, the regressors associated with b and e coincide in that regime and only their sum is identifiable; this does not affect a ^ and hence does not affect Equation (5). Second, a projection constraint keeps the estimate inside the admissible set a min ≤ a ^ ≤ a max < 1 , equivalent to k c , min ≤ k ^ c ≤ k c , max , which excludes the physically meaningless values that noise or short invalid packets could otherwise produce. The initial estimate, the initial covariance matrix, the threshold ξ_min and the projection limits used in the simulation are given with the other simulation settings in Section 5.

Each received measurement passes three checks, and the validity flag of node i is their product:

where v i to = 1 if the last packet of the node is not older than the timeout τ_to, v i rg = 1 if temperature and relative humidity lie within their admissible ranges, and v i jp = 1 if the increment between consecutive samples does not exceed the admissible jump. The set of active nodes and its cardinality are

and the fused thermal state is computed from active nodes only:

The maximum is used for threshold decisions, so that a local hot spot is not masked by averaging; the mean characterises the general state of the stack. The loss of nodes is expressed by a separate information-uncertainty index,

If N_A falls below a predefined minimum N_min, automatic control is restricted, because the hottest zone may be precisely the one whose node is disconnected. In this way information uncertainty is kept separate from process risk. The case in which the lost node is the critical one while N_A ≥ N_min is treated separately in Section 6.5.

The technological risk index combines normalised temperature, humidity, rate of temperature rise and the difference between internal and external conditions:

where x_j is the corresponding physical quantity and [ x j lo , x j hi ] its normalisation interval. The bilinear term w_TWq_Tq_W accounts for the non-linear amplification of risk when high temperature and high humidity occur simultaneously. The weights are assigned initial values for simulation and must be calibrated against experimental quality and thermal-response data.

The inputs of Equation (16) are formed from the active nodes only. Because the SHT21 nodes measure the relative humidity of the air in the stack and not the moisture content of the cotton, the humidity input is the measured air relative humidity, used as a proxy for the humidity-related state W: x W ( k ) = max i ∈ A ( k ) R H i ( k ) . The temperature input is x_T(k) = T_max(k) from Equation (14); the rate input is the largest rate of rise among the active nodes, x r ( k ) = max i ∈ A ( k ) [ T i ( k ) − T i ( k − m ) ] / ( m T s ) , evaluated over a window of m samples to suppress noise; and the internal-external difference is x_Δ(k) = T_max(k) − T_a(k). Taking the maximum of each input over the active nodes is conservative: the inputs may come from different nodes, but a single critical zone cannot be hidden. The weights, normalisation intervals and decision thresholds used in the simulation are given with the other simulation settings in Section 5. They are assumed settings and not calibrated results; the weights sum to one, so that R_q [0, 1].

The supervisory controller is a finite-state machine with six principal states (Table 2). In MOVING_FORWARD the ventilation contactor is disabled, and in FAN_RUNNING the stepper pulses are blocked; this mutual exclusion prevents simultaneous commands that could create unsafe actuator conflicts.

Table 2. States of the supervisory state machine.

In automatic mode, ventilation is requested when at least one valid sensor exceeds the temperature threshold or when the risk index exceeds its decision boundary:

Ventilation is stopped only after all active sensors have returned below the lower hysteresis boundary, the minimum ventilation time has elapsed and no critical humidity condition remains:

where ΔT_h is the hysteresis width, t_f the elapsed running time of the fan and q_W,cr the critical level of normalised humidity. Hysteresis and minimum ON/OFF times prevent excessive cycling of the contactor. E-STOP and overload protection are implemented in the hardware power circuit, independently of the software state machine.

The role of the RLS estimate in this loop is stated explicitly. In the configuration evaluated in this paper the estimate k ^ c does not modify any supervisory setting (T_on, ΔT_h, R_on, q_W,cr and the minimum ON/OFF times) and does not enter the actuator commands: u_f and u_m are determined only by Equations (18) and (19) and by the transitions of Table 2. RLS is therefore used for adaptive identification and diagnosis. It keeps the channel model of Equation (4) consistent with the current state of the object and provides the indicator k ^ c , which is displayed on the HMI together with its nominal value k_c,nom. When the relative drop D ( k ) = 1 − k ^ c ( k ) / k c , nom remains above the threshold D_al for longer than the estimator memory N_λT_s, a “reduced ventilation effectiveness” warning is sent to the operator; this warning does not change the control sequence. Using k ^ c to adapt the minimum ventilation time or, with a VFD, to compute a continuous fan-speed reference is a possible extension that is not evaluated here.

The platform is moved by a stepper drive with STEP/DIR control through a reduction stage. The number of pulses per metre of travel and the linear resolution are

where α_s is the step angle, m_μ the microstepping division, i_r the reduction ratio and r_w the wheel radius. With α_s = 1.8˚, m_μ = 8, i_r = 6 and r_w = 0.10 m, Equation (20) gives n_p ≈ 15,279 pulses/m and δ_x ≈ 0.065 mm, so that the 0.5 m reverse movement corresponds to approximately 7639 pulses. The actuator moves the platform to the DS-438 limit sensor, after which ventilation is performed; when the thermal condition of Equation (19) is satisfied, the 0.5 m reverse movement is commanded.

Motion without loss of steps requires the tractive force at the wheel to exceed the resistance force with a reserve κ:

where M_m is the available motor torque at the operating speed, η = 0.85 the efficiency of the reduction stage, f_r the rolling-resistance coefficient, m = 290 kg the platform mass (weight ≈ 2.84 kN) and F_add the additional resistance caused by floor irregularity and cable drag. Under the adopted assumptions the design calculation gives a traction reserve of κ ≈ 16% for the 1:6 reduction. This ratio is not claimed to guarantee motion under all real floor conditions: the actual normal load, floor irregularity, cable drag and position of the centre of mass can change the available traction materially, and an experimental comparison of the 1:6, 1:7 and 1:8 ratios is therefore part of the validation plan (Section 7).

Figure 2 shows the temperatures of the eight nodes in the normal scenario. The initial temperatures are non-identical and lie between approximately 35˚C and 44.5˚C. During the first two minutes, while the platform travels to the working position, the temperatures remain at their initial levels; ventilation then starts and all channels cool monotonically. The fan remains active until all valid nodes have returned to 28˚C or below and the minimum operating condition is satisfied, which occurs at approximately 12 min. Because the stop decision is taken on the maximum over the active nodes, a local hot spot cannot be masked by averaging.

Figure 2. Simulated temperature dynamics at the eight monitoring nodes. Scenario S1 with the settings of Table 3; shaded intervals: fan ON.

After the fan is switched off, a second heating episode raises the node temperatures to approximately 30˚C - 31.5˚C, and ventilation is retriggered once the minimum OFF interval has elapsed, at approximately 20 min. The response thus demonstrates hysteresis-based cycling control: two ventilation cycles occur within 30 min, and the contactor is not switched in the band between the thresholds.

The mobile sequence is coordinated by the state machine: the platform moves towards the limit sensor and stops at the working position, ventilation is activated and completed, the two-second reset sequence is executed, and the platform returns by approximately 0.5 m. The simulation therefore evaluates the integrated sequence and not isolated control blocks.

The behaviour of the estimator is shown in Figure 3. The true coefficient of the virtual object equals 0.098 min⁻¹ during the first 15 min and is then reduced stepwise to 0.083 min⁻¹, a change of approximately 15%; by Equation (5) this corresponds to an increase of the cooling time constant from approximately 10.2 min to 12.0 min. Starting from the initial value of 0.060 min⁻¹, the estimate reaches the neighbourhood of the true value in approximately 9 - 10 min, and after the step change it converges to the new value in a comparable time, with residual fluctuations of about ±0.002 min⁻¹. These convergence times amount to roughly 1.4 - 1.5 memory lengths N_λT_s, as expected of an estimator with exponential forgetting.

Figure 3. Online estimation of the effective cooling coefficient using RLS (virtual simulation).

The estimate changes faster during the transient and becomes less dispersed once sufficient excitation has been accumulated. The choice λ = 0.995 is a compromise between adaptation speed and noise sensitivity for a slow thermal process: a smaller λ would respond faster but would amplify noise, whereas a value closer to unity would track real parameter changes more slowly.

The cooling time and the relative energy consumption obtained in the five scenarios are compared in Figure 4 and Table 5; energy is normalised to the nominal scenario S1.

Energy is evaluated from the electrical input power of the two drives in the simulation model:

where P_f is the electrical input power of the fan induction motor computed in the fan subsystem from the stator voltage, current and power factor, P_s the electrical power of the stepper drive and ε_j the relative energy of scenario j reported in Table 5. The evaluation window starts at t₀, the instant of the first ventilation request, and ends at t₁, the instant at which the system returns to IDLE after the reverse movement of the first ventilation cycle; the integral is evaluated numerically with the solver step of the model. The power of the sensor nodes and of the controller is not included. The cooling time in Table 5 is measured from fan switch-on to the fulfilment of the stop condition (19) in the same cycle.

Figure 4. Comparative simulation results for cooling time and relative energy consumption. (a) Cooling time; (b) relative energy consumption, normalised to S1.

Table 5. Cooling time and relative energy consumption in the simulated scenarios (approximate values).

The degradation of the fan (S4) is the most severe disturbance: the model predicts the longest cooling time and the highest relative energy use. A 30% loss of effectiveness lengthens cooling by approximately 71% and raises relative energy use by approximately 48%, so that the consequences of fan degradation are more than proportional to its magnitude. In this scenario RLS detects the changed cooling behaviour and the diagnostic warning of Section 4.4 is raised; the supervisory commands themselves are unchanged, so the longer cooling time results only from the later fulfilment of the stop condition (19). The estimate k ^ c can be used as a diagnostic indicator. In the present contactor-based architecture it does not command a continuous fan speed; with a VFD, the estimated coefficient and the risk index could be used to compute a continuous speed reference. Multiple hot zones (S2) lengthen ventilation because the stop condition is governed by the slowest zone. Increased movement resistance (S5) acts on the positioning channel and not on the thermal one, and accordingly its effect on energy is larger than its effect on cooling time.

The validation protocol is as follows. Identification and evaluation use two separate simulation runs of the channel of node 1 under the nominal settings of Table 3, each of 30 min with T_s = 2 s. The first run was used for tuning the estimator settings (λ, P(0), ξ_min) and for identification. The second run, which starts from a different initial state and has an independent noise realisation, provides the virtual validation series of n = 900 samples shown in Figure 5; no sample of this run was used for tuning, and random point-wise splitting was not used, because successive samples of a time series are correlated. During the evaluation the estimator updates were stopped and the parameters were frozen at the values reached at the end of the identification run. The model output y ^ ( k ) in Equations (24)-(27) is the free-run (independent trajectory) prediction of Equation (4), initialised with the measured temperature at the start of the series and driven only by u_f, T_a and RH_a; the measured temperature is not fed back into the regressor, so the prediction horizon equals the full 30-min series. This differs from the one-step prediction of Equation (6), which uses the measured temperature at every step and is a less demanding test; the values in Table 6 should therefore be read as trajectory-prediction accuracy on virtual data. The ratio MAE/RMSE = 0.81 is close to the value 2 / π ≈ 0.80 expected for Gaussian residuals, which is consistent with a residual dominated by the simulated sensor noise (σ_T = 0.10˚C).

The agreement between the model output y ^ ( k ) and the virtual validation series y(k) of length n was evaluated by four standard indicators:

where y ¯ is the mean of the validation series. The comparison is shown in Figure 5 and the values are given in Table 6.

Figure 5. Comparison of the model output with the virtual validation data.

Table 6. Model adequacy indicators (virtual validation).

The values R² = 0.9689 and RMSE ≈ 0.12˚C indicate good agreement with the virtual validation series. Taken together they imply a standard deviation of the validation series of approximately 0.68˚C, so that the residual amounts to about 18% of the variation of the signal; Figure 5 shows that this residual is dominated by the high-frequency noise superimposed on the virtual data, while the model output reproduces the slow component. Because the virtual object and the identification model share a related mathematical structure, high agreement is expected. The values must therefore not be presented as proof of the physical accuracy of the SHT21 sensor or of a model of the real cotton stack.

The sensor-disconnection scenario demonstrates why simple averaging is insufficient. If the hottest sensor disappears, the remaining measurements artificially reduce the estimated risk. The proposed controller responds by increasing the information-uncertainty index of Equation (15) and by restricting automatic operation when the number of active nodes falls below the critical minimum. For movement faults, a timeout disables the STEP pulses and transfers the system to FAULT; the controller does not attempt to overcome the obstacle by raising the speed, which would increase the loss of steps.

A node-count threshold alone, however, cannot establish that an unobserved zone is safe: if the node that was the hottest disconnects while N_A is still above N_min, all remaining nodes may be below T_on although the lost zone is still hot. The loss of a critical node is therefore handled by a separate rule. A node is classed as critical when, at the moment its validity flag drops, it defined T_max or its last valid temperature was at or above T_on − ΔT_h. For such a node: 1) its last valid temperature is held in Equation (14) instead of being removed, so that T_max cannot decrease because of the loss and the stop condition (19) cannot be met while the zone is unobserved; 2) if ventilation is running it continues, and if the platform is idle a ventilation request is issued as if s_on = 1, so that the platform moves to the working position and the fan is switched on; 3) a “critical node lost” alarm is sent to the operator HMI immediately; 4) if the node is not restored within t_esc, the system escalates: automatic stopping of ventilation is inhibited, the fan remains on with the platform at the working position, manual inspection of the zone is requested, and ventilation can be ended only by the operator in manual mode. The held value is released only after the node has again delivered valid data for a confirmation interval. If N_A falls below N_min, automatic control is restricted as described in Section 4.3, the actuators are kept in their current safe state and the operator is alerted. In scenario S3 the hottest node (node 1) is disconnected between t = 5 and 9 min; its last valid value is held in T_max, t_esc is not reached, and the 60-s confirmation interval after reconnection contributes to the longer cooling time of S3 in Table 5.

The safety model distinguishes the software FAULT state from physical emergency interruption. E-STOP, thermal overload protection and contactor feedback are required to interrupt the power circuit independently of the microcontroller. This separation is essential for a practical mechatronic system, because the software cannot be assumed to remain operational in the situations in which protection is needed.

Table 7 compares the proposed system functionally with portable or local monitoring and with stationary PLC/SCADA monitoring.

The novelty therefore does not lie in the use of a particular sensor, the ESP32 controller or a web interface. It lies in the integration of distributed sensing, adaptive identification, supervisory control, mobile mechanics and safety logic into a closed mechatronic architecture. Existing cotton-storage monitoring work demonstrates the value of wireless temperature-humidity monitoring [3]; the present study extends that concept towards active adaptive control.

Table 7. Functional comparison of monitoring-oriented approaches and the proposed system.