Based on precise TDE traceability and a novel model for CMGs, identifying its parameters is another important factor. The more precisely the novel model is identified, the less ambient temperature influences TDE, and the better the bias stability of CMGs. So, testing TDE accurately is the key prerequisite, and a proper TDE test method is essential.
2.2.1. TDE Accurate Acquisition Methodology
- (a)
Heat conduction solutions
To ensure the experimental results’ reliability, heat from ambient temperature should be conducted to CMGs efficiently. Heat conduction solutions are applied on CMGs, like thermal grease and thermal conductive rubber, which keep the ambient temperature the same as the CMG and reduce the deviation between experimental results and their theoretical values.
- (b)
Precise temperature measurement system
Precise ambient temperature represents CMG’s environmental adaptability comprehensively, so it is necessary to apply a precise temperature measurement system. It needs be installed closely on the CMG to reduce the temperature gradient. Its measuring accuracy needs to be over 2 times more accurate than ambient temperature variation, and its measuring frequency needs to be higher than CMG’s output frequency for complete experimental results.
- (c)
Proper temperature control interval
TDE is a statistical characteristic deviation represented based on mathematical models. Once CMGs are manufactured, their mechanical properties are fixed, and their TDE can be expressed with bias, scale factor, and random error. So, CMGs have similar environmental adaptability but unique TDEs, and from their datasheets, TDEs can be roughly estimated as follows:
where ∆
E′ are roughly estimated TDEs,
γ is the parameter “Zero-rate level change vs. temperature”, and
β is the parameter “Sensitivity change vs. temperature”. Theoretically, ∆
E′ is smaller than TDE because some non-statistical errors are ignored and ∆
E′ ≤ ∆
E. CMG’s sensitivity ∆
ES determines the measured minimum of angular velocity, and higher sensitivity brings more accurate angular velocity. When ambient temperature deteriorates suddenly during CMG work, it is probable that TDE is much bigger than ∆
ES, even covering up CMG’s sensitivity to measure inaccurate angular velocity. So, it is essential to vary ambient temperature at a reasonable temperature variation rate. To test TDE precisely, set ∆
ES ≈ ∆
E′, and the temperature control interval can be further expressed as follows:
- (d)
Reasonable temperature control period
According to heat conduction theory, a long enough heat conduction period conducts the heat from position A to position B completely, which ensures that ambient temperature variation at position A is the same as position B. An insufficient heat conduction period causes an incomplete heat conduction process, and the TDE of the CMG is tested imprecisely. Usually, a thermal chamber is used to test TDE. It is designed as a cube of front-door-open and -closed insulation. Also, it integrates a temperature control system with six temperature control units on each plane, and they can be further simplified according to the targeted control efficiency. There is a rate table inside to offer the referenced angular velocity. To reflect the ambient temperature of the CMG better, temperature sensors are set on it.
Figure 5 shows a schematic diagram of the test platform.
In
Figure 5, the heat is transferred through the thermal chamber’s inner wall to control ambient temperature inside the targets, and the heat from temperature control units uniformly transfers to the CMG. Then, the center of the rate table is taken as a reference position, and it is also the central position of the thermal chamber. Given that the sizes of the thermal chamber are
L1 × L2 × L3, the farther the position is away from the inner wall, the longer the period of heat transfers. Under the action of temperature control units, CMGs obtain the heat, and their environmental adaptability changes. The position where the CMG is located is the final area where the ambient temperature stays stable. According to energy conservation law and Newton’s law of cooling, the heat flux density equation in three dimensions is shown as follows:
where
u is the ambient temperature of any position in the thermal chamber at moment
t,
q is its heat flux density,
kh is the heat conductivity coefficient,
is the spatial varying rate of ambient temperature on the
x-axis,
is the spatial varying rate of the ambient temperature on the
y-axis,
is the spatial varying rate of the ambient temperature on the
z-axis,
is the unit vector on the
x-axis,
is the unit vector on the
y-axis, and
is the unit vector on the
z-axis. The heat on the
x-axis heats CMGs, and according to (15), a heat transfer equation can be established as follows:
where
Qx is the heat at position
A on the
x-axis, (
qx)
A is its heat flux density, (
qx)
A+ΔA is its heat flux density of its close position
A + Δ
A near position
A on the
x-axis, Δ
y is the width of the plane where the heat conducts, Δ
z is the height of the plane where the heat conducts, and Δ
t is the period when the heat conducts. Considering that the CMG is heated on the
x-axis,
y-axis, and
z-axis, a heat transfer equation can be established on the
y-axis and
z-axis as follows:
where
Qy is the heat at position
A on the
y-axis and
Qz is the heat of position
A on the
z-axis. Based on (16) and (17), the heat at position
A in three dimensions can be deduced as follows:
where
F(
x,
y,
z,
t) is the heat flux density of the potential heat sources related to its position and time,
v is the 3D unit vector matrix, and
. When the thermal chamber works in a period
ts, according to the specific heat capacity formula, we can obtain the following equation:
where
C is the specific heat capacity of air in a thermal chamber in a closed condition,
m is its mass, ∆
T is temperature variation, and ∆
T = |
Tb-T0|.
T0 is the initial temperature and
Tb is the final temperature. Then, (19) can be transformed as follows:
where
ρ is the air density inside the thermal chamber. Considering that different positions in the heat conduction pathway have different heat flux densities in the heat conduction process, the further the heat is conducted, the smaller the heat flux density becomes, and the lower heat transfer efficiency stays. The CMG is a sensor, and its heat is relatively insignificant to the thermal chamber, so
F(
x,
y,
z,
t) can be ignored in magnitude. Then, (20) can be deduced as follows:
To calculate total heat along the heat conduction pathway, (21) is integrated on the
x-axis,
y-axis, and
z-axis separately, and then we obtain the following equation:
After integration, (22) is transformed and shown as follows:
where
Tx is the control target of the temperature control unit on the
x-axis,
Ty is on the
y-axis, and
Tz is on the
z-axis. Then,
ts can be deduced as follows:
Usually, a thermal chamber is designed as a cube, and its size is
L × L × L. Moreover, the control target on the
x-axis,
y-axis, and
z-axis is set the same as
Tb, and (24) can be simplified further as follows:
Then, the period for heat conduction from the inner wall to the center of the thermal chamber can be calculated with (25). To guarantee that the thermal chamber is heated uniformly, the temperature control period is set as
ts ≤
tp. Taking L3GD20H as an example, it is chosen to test TDE randomly. From its datasheet, ∆
ES = 8.75 mdps/digit,
γ = ±0.04 dps/°C, and its operating temperature range is −40~85 °C. After dimensional transformation,
β is obtained as follows:
According to (14), temperature control interval ∆
T is shown as follows:
To simplify the test steps, ∆
T = 0.5 °C. A thermal chamber SET-Z-021 is utilized to test L3GD20H, and its parameters are
C = 1.005 kJ/(kg×K),
kh = 0.0267 W/m°C,
L = 0.6 m, and
ρ = 1.293 kg/m
3. Given that CMG’s operating temperature range is −40~85 °C, (25) is obtained as follows:
So, temperature control units take 25.766 s as the temperature control period to vary the temperature control interval at 0.5 °C. To simplify test steps and reserve an allowance for stable heat conduction, tp = 30 s. L3GD20H is tested in temperature experiments, and its temperature is obtained with a precise temperature measurement system of measuring accuracy of ±0.03 °C and measuring frequency of 10 Hz. Thus, a temperature experiment is designed as follows:
CMG L3GD20H is installed on the rate table, its measuring direction is parallel to the rate table, and the referenced true value is the angular velocity of the rate table.
Temperature sensors of the precise temperature measurement system are attached close to L3GD20H, the wireless data transmission module transmits the experimental results, and the PC is prepared to receive its temperature TG and its output DG.
Cool the thermal chamber to a minimum operating temperature of −40 °C and keep TG and DG recording for 0.5 h when the ambient temperature stays stable.
Heat the thermal chamber to a maximum operating temperature of 85 °C at a rate of 60 °C/h, 0.5 °C per 30 s. When TG goes up to 85 °C, stop the test when it stays stable for 0.5 h.
Redo steps (2) to (4) five times and record them as the experimental results.
Figure 6 shows a flowchart of the temperature experiment and experimental results.
2.2.2. Implementation of Novel TDE Precise Estimation Model Based on an RBFNN
Under the premise of building a novel TDE precise compensation model and obtaining TDE and TCQs, identifying its precise parameters is another key to estimating the TDE of the CMG accurately. As shown in
Figure 6, the ambient temperature goes up from −40 °C to 85 °C, and at the very beginning, the ambient temperature stays stable for 0.5 h, and it is set as the reference ambient temperature. Moreover, the reference output of the CMG is 0 dps/s because
. ∆
T, ∆
T2, ∆
T1/2, and TDE deduced from
Figure 6 are shown in
Figure 7.
As shown in
Figure 7, when ambient temperature
T varies, ∆
T and ∆
T2 vary in a similar trend. Because of the small numerical amplitude of ambient temperature variation at the beginning, ∆
T1/2 has the same trend with a small range of numerical fluctuations. Under the excitation of ∆
T and ∆
T2, as well as ∆
T1/2, TDE has an approximate varying trend. Therefore, it shows that there is some relevance among ∆
T, ∆
T2, ∆
T1/2, and TDE.
Figure 8 shows the corresponding relationship among them.
As shown in
Figure 8, there is a complex nonlinear relationship among ∆
T, ∆
T2, ∆
T1/2, and TDE, and it is so difficult to estimate precisely in real time. So, it is essential to use a nonlinear model of multiple inputs and a single output as well as high accuracy to fit the nonlinearity. The RBFNN is a feedforward neural network based on the function approximation method with a single hidden layer. It consists of neurons and neural layers. Neurons are applied as basic computing units, and neural layers are utilized as a computing framework. Its neural layers include an input layer, a hidden layer, and an output layer, and neurons are distributed in diverse layers. The neurons in the hidden layer and output layer have kernel functions, and inputs in the RBFNN are transmitted between the input layer and the output layer by neurons and kernel functions. By means of neurons and neural layers, as well as kernel functions, the output in the RBFNN perfectly represents or approximately approaches the targets. Usually, it uses the Gaussian function as the kernel function in the hidden layer and the purelin function as the activation function in the output layer, and they are shown as follows:
where
cj is the center of the Gaussian function and
δj is its width,
a and
b are constant,
x is the input of the functions, and
y are their outputs. Moreover, the RBFNN has two merits as follows:
Figure 9 shows the structure of the RBFNN for the complex nonlinear relationship among ∆
T, ∆
T2, ∆
T1/2, and TDE.
Where
Ii(
i = 1⋯3) is the
ith neuron of the input layer,
Hi(
i = 1⋯
K) is
the ith neuron of the hidden layer, and
O1 is the neuron of the output layer. Based on that, the RBFNN is a good choice to represent the nonlinear relationship among ∆
T, ∆
T2, ∆
T1/2, and TDE accurately. So, (12) can also be deduced as follows:
Then, the parameters of the novel model should be identified as follows:
Two temperature experiments are conducted, and the experimental results are recorded. One of them is a training sample set, and the other one is a verification sample set.
Based on sample data of the training sample set, TDE is calculated by subtracting the reference value of CMGs from the sample data of their actual outputs one by one. ∆T is calculated by subtracting the reference temperature of CMGs from the sample data of their actual temperature one by one. Then, ∆T2 is obtained by multiplying itself, and ∆T1/2 is obtained by calculating the square root of ∆T.
The RBFNN uses ∆T, ∆T2, and ∆T1/2 as its inputs and TDE as its output. It is trained with mathematical tools until the differences between its outputs and targeted TDE meet the requirements.
The compensated results of CMGs are calculated from subtraction between the actual outputs of CMGs and their corresponding estimated outputs of RBFNNs.
The sample data of the training sample set are determined by MEMS Gyros’ output frequency under the target operating conditions. The higher MEMS Gyros’ output frequency is, the more quickly MEMS Gyros outputs, and the higher the sample data of the training sample set. Then, (30) is trained with the experimental results, and its parameters are identified precisely.
Figure 10 shows the primary data and those compensated by the conventional model and novel model in five experiments.
To check its estimation performance, the Mean Square Deviation (
MSD) is introduced as follows:
where
x is the evaluated sample,
x′ is its reference, and the
MSE is the mean square error algorithm. The
MSD is an index that indicates the dispersion degree of the evaluated sample and its reference. The less the
MSD is, the smaller the dispersion degree stays. The bias stability of CMGs is the evaluated sample. So, the
MSD can be implemented with programming software based on mathematical principles. The
MSD of the bias stability of its primary data is
MSD1, the bias stability of those compensated by the conventional model is
MSD2, and the bias stability of those compensated by the novel model is
MSD3, which are shown in
Table 1. An
MSD improvement of the conventional model
Q1 and the novel model
Q2 are shown as follows:
In
Figure 10, CMGs compensated by the conventional and novel models perform stably with small fluctuations around the reference, but CMGs compensated by the novel model run more stably and have smaller fluctuations, which shows that the novel model can estimate TDE more accurately. In
Table 1, although the conventional and novel models reduce the
MSD of the CMG effectively, the novel model reduces the
MSD of the CMG less, and the
MSD after compensation is enhanced by four orders of magnitude. Moreover, CMG’s bias stability is enhanced better by 10.5% evenly than the conventional model. So, the novel model can estimate TDE more accurately to improve the bias stability of the CMG significantly.