Optimizing robot anomaly detection through stochastic differential approximation and Brownian motion
Optimizing robot anomaly detection through stochastic differential approximation and Brownian motion
Corresponding Author:
Jackrit Suthakorn,
Center for Biomedical and Robotics Technology (BART LAB), Department of Biomedical Engineering,
Faculty of Engineering, Mahidol University
999, Phuttamonthon Sai 4, Salaya, Nakorn Pathom, 73170, Thailand
Email: [email protected]
1. INTRODUCTION
The primary signal and sensor processing issue is anomaly identification during the data stream [1].
This research's primary focus is identifying anomalies within data streams, particularly emphasizing the
element of time. The objective is to enhance the efficiency of representing a specific subset of temporal data
streams through a sequential design of experiments, facilitating accurate and rapid anomaly detection [2], [3].
A significant challenge in time-based anomaly identification, especially in the context of mobile robots used
in rough terrain rescue missions, is the associated cost of transitioning the data stream from one geographical
region to another [4]. This cost primarily arises from the movement of robots. Various approximation
algorithms have been developed to address this issue, with the Brownian motion algorithm being a prominent
choice due to its experience in managing time and cost-effective model construction [5]. However, the
standard Brownian motion algorithm faces challenges, including handling vast datasets, memory limitations,
and the inability to adapt to a behavior-based system with infinite variance [6], [7]. To address the challenges
of real-time anomaly detection in dynamic, resource-constrained environments like rugged terrains, the
authors propose a novel approach leveraging stochastic differential approximation (SDA) and optimistic
optimization of Brownian motion [8]. This method approximates the shortest path characterized by Brownian
motion within a defined time interval, achieving greater accuracy by minimizing path length. The study aims
to provide a cost-effective, adaptive solution that balances energy efficiency, operational cost, and real-time
response while optimizing robot movement for precise anomaly detection.
Brownian motion is defined as the variation in the path length measurement, characterized by its
continuous and natural nature and adaptability to showcase differentiation in erratic motions [9], [10]. In
presenting the minimum Brownian motion using stochastic differential approximation and optimistic
optimization, the path is depicted in time intervals, each further divided into sub-intervals [11]. The critical
objective is to determine how each sub-interval should be represented within the entire time interval. This
involves the selection of a mathematical model, in this case, stochastic differential equations (SDE), and an
optimization technique, specifically an optimistic optimization algorithm [12]. In summary, this paper
introduces an innovative approach to tackle the challenges associated with time-based anomaly identification,
particularly in mobile robots, by presenting the stochastic differential approximation and optimistic
optimization of Brownian motion as a more effective alternative to the traditional Brownian motion
algorithm [13]. The proposed approach provides several key contributions to enhance cost efficiency and
improve anomaly detection accuracy in robotic systems. One of its primary advantages is the significant
reduction in expenses related to the movement of mobile robots, particularly during the process of detecting
anomalies in data streams [14]. By employing a mathematical model that approximates the minimal
Brownian motion path, the method effectively minimizes unnecessary robot movements, thereby conserving
both energy and resources.
Brownian motion, often associated with random movement patterns, is streamlined here to limit
robot movement, focusing on pathways with minimal deviation. This targeted movement strategy not only
saves operational costs but also extends the operational lifespan of robotic components by reducing wear and
tear on the machinery [15]. In addition to cost savings, the approach employs an advanced mathematical
framework based on stochastic differential equations (SDEs) to refine the precision of the minimal path
approximation. By leveraging SDEs, the system can dynamically adjust to unpredictable factors that impact
robot movement and data collection [16]. This added layer of mathematical rigor enhances the accuracy of
the Brownian motion model, ensuring that the robots follow a path close to the minimal distance required to
detect anomalies effectively. The use of SDEs ensures that the approach adapts well to fluctuating conditions,
which are common in real-world applications where robotic systems encounter varied terrains and obstacles
[17]–[20]. This results in a robust system where the accuracy of anomaly detection remains high, even under
challenging conditions.
To improve the approach, an optimistic optimization technique is incorporated to effectively tackle
continuous optimization challenges associated with Brownian motion [16]. This framework plays a vital role
in identifying the optimal paths for robots, striking a balance between minimizing path length and
maximizing the probability of anomaly detection. By focusing on optimizing the robot's path, the method
significantly increases the chances of accurately identifying anomalies without necessitating extensive
movement. This optimization process does not solely focus on identifying anomalies but also emphasizes
efficient resource allocation, minimizing computational power, and ultimately reducing the time and cost
involved in anomaly detection. An additional aspect of this methodology is its emphasis on accurately
pinpointing the most likely minimum path of Brownian motion within a specific time frame, a feature that
significantly enhances the effectiveness of anomaly detection [21]. The proposed approach effectively
narrows down the set of probable paths, focusing on accurately identifying the minimum path that exhibits
Brownian motion characteristics. This ensures that anomalies in extensive data streams are detected promptly
and with minimal resource consumption, reducing unnecessary robotic movements. The time-bound nature of
this identification process further enhances the system's responsiveness, enabling real-time detection of
irregularities. By maintaining high precision in anomaly detection while operating under resource constraints,
the approach becomes highly suitable for applications where both accuracy and operational efficiency are
critical, such as autonomous navigation, disaster response, and environmental monitoring in challenging
terrains.
2. METHOD
Consider the data stream 𝑓 = { 1,2,3 … … 𝑛} where f creates the series of random variables are
𝑓 𝑓 𝑓 𝑓
𝐴1 , 𝐴2 , 𝐴3 … … 𝐴𝑛 occurs with the data sample space A. Each data stream is dependent on either one of the
hypotheses such as 𝐻0 𝑜𝑟 𝐻1 . The observations made by two different probability distribution functions on
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30
IAES Int J Rob & Autom ISSN: 2722-2586 21
𝑓
data sample space A of data stream are 𝐷1 𝑎𝑛𝑑 𝐷2 . Taken 𝐻0 for 𝐷1 where 𝐻0 → 𝐴𝑛 is true for 𝐻0 . Similarly,
𝑓
𝐻1 for 𝐷2 where 𝐻1 → 𝐴𝑛 is true for 𝐻1 where n denotes the set of integers. 𝐻0 is positive when the specific
data stream defines the movement of data stream is normal and 𝐻1 is positive when the movement of data
stream is target. Suppose for the taken data stream, 𝐻1 is true and its probability is 𝜌 and 𝐻0 is true with its
probability (1 − 𝜌) where 𝜌 ∈ (0,1). When these criteria are not satisfied for 𝜌, it shows that target data
stream happens to be too rare case. During the expectation of hypothesis 𝑃𝑖 . The general structure of
Brownian motion is shown in Figure 1.
To reduce the movement cost of the mobile robot during the data stream from one point to another
point, the problem observed and formulated during the data movements from stream 𝑊1 to the next stream
𝑊𝑛 + 1, so it exists with cost 𝜆𝑊 within distribution 𝐷 ∀ 𝑊(𝑖. 𝑒. )𝜆𝑊 ≥ 0 and €[𝜆𝑊 ] = 𝜆̃ is finite [22].
′
Assume that movement cost 𝜆𝑊 and the remarks 𝑂𝑇𝑊 are commonly independent ∀ W,𝑊 ′ &𝑇. The movement
cost is expensive due to the movement of the robot while the data stream is monitored during the new data
stream [23]. If the observations are not taken, then it is said to be zero time and if the observations are
monitored and the time are recorded then it shows the observation for movement cost. This can also be
related to energy.
Problem 1: Consider the Brownian motion (𝐵𝑟(𝑡))𝑡≥0 with a Hurst value Hu > 1/2. Stochastic
linear equations of this type are investigated in the format in (1).
𝐴(𝑡0 ) = 𝐴0 , Whereas 𝑡0 ∈ (0, 𝑇), 𝐴0 is an Unpredictable vector in ℚ𝑛 and the subsequent criteria
hold true with likelihood 1 for the randomly generated functions S and T in (2),
𝜕𝑇(∙,𝑡) 𝜕𝑇(∙,𝑡)
for every 𝑡 ∈ (0, 𝑇) the functions 𝑆(∙, 𝑡), , for all i in [1, 2..., n] are localized Lipschitz.
𝜕𝑥 𝑖 𝜕𝑡
Consider the supplementary partial differential (3) along the route on ℚ𝑛 ∗ ℚ ∗ (0, 𝑇),
𝜕𝐿
(𝑥, 𝑦, 𝑡) = 𝑇(𝐿(𝑥, 𝑦, 𝑡), 𝑡) (3)
𝜕𝑦
𝜕𝐿𝑖
det ( (𝑥, 𝑦, 𝑡)) ≠ 0 (4)
𝜕𝑥 𝑗
For (𝑦, 𝑧, 𝑡) ∈ 𝑁
𝜕2 𝐿 𝜕𝑇
(𝑦, 𝑧, 𝑡) = ∑𝑛𝑗=1 (𝐿(𝑥, 𝑦, 𝑡), 𝑡)𝑇𝑗 (𝐿(𝑥, 𝑦, 𝑡), 𝑡) (5)
𝜕𝑦 2 𝜕𝑦 𝑗
Optimizing robot anomaly detection through stochastic differential approximation … (Branesh M. Pillai)
22 ISSN: 2722-2586
From (5), on [0, T], additionally consider the path wise differential equation (in matrix form) as in (6).
𝜕𝐿 −1 𝜕𝐿
𝑑𝐶(𝑡) = ( 𝐶(𝑡), 𝐵𝑟(𝑡), 𝑡) [𝑆(𝐿(𝐶(𝑡), 𝐵𝑟(𝑡), 𝑡), 𝑡) − 𝐶(𝑡), 𝐵𝑟(𝑡), 𝑡] 𝑑𝑡, (6)
𝜕𝑥 𝜕𝑡
𝐶(𝑡0 ) = 𝐶0 (7)
From (7), a maximum range, has a distinct local solution as (𝑡01 , 𝑡02 ) ⊆ (0, 𝑡) with 𝑡0 𝜖(𝑡01 , 𝑡02 ). Here, a use of
stochastic application of the formula to the randomized function 𝑄(𝑧, 𝑡) = 𝐿(𝐶 (𝑡), 𝑦, 𝑡) and the Brownian
motion Br in (8),
𝑛 𝑡 𝑡 𝑡
𝜕𝐿 𝜕𝐿 𝜕𝐿
= ∑ ∫( (𝐶(𝑠), 𝐵𝑟(𝑠), 𝑠)) 𝑑𝑌𝑗 (𝑠) + ∫ ( (𝐶(𝑠), 𝐵𝑟(𝑠), 𝑠)) 𝑑𝐵𝑟 (𝑠) + ∫ ( (𝐶(𝑠), 𝐵𝑟(𝑠), 𝑠)) 𝑑𝑠
𝜕𝑥 𝑗 𝜕𝑥 𝜕𝑡
𝑗=0 𝑡0 𝑡0 𝑡0
𝑡 𝑡
= ∫𝑡 𝑆(𝐿(𝐶(𝑠), 𝐵𝑟(𝑠), 𝑠), 𝑠)𝑑𝑠 + ∫𝑡 𝑆(𝐿(𝐶(𝑠), 𝐵𝑟(𝑠), 𝑠), 𝑠)𝑑𝐵𝑟(𝑠) . (9)
0 0
For each 𝑍 ∈ ℤ,a treat the path wise differential equation of (6) (represented as a matrix) as an approximation
𝐵𝑟𝑍 (t) of the original process 𝐵𝑟(𝑡) and the (11),
𝐶𝑧 (𝑡0 ) = 𝐶0 (11)
A distinct localized equilibrium 𝐶𝑁 on the maximum period of presence (𝑡1, 𝑡2) ⊆ (𝑡01 , 𝑡𝑜2 ). Using the
stochastic formula, 𝑄(𝑧, 𝑡) = 𝐿(𝐶 (𝑡), 𝑦, 𝑡) for the random function 𝑄(𝑧, 𝑡) and the procedure BN as in (12),
The subsequent path wise condition arises from the proposed theorem
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30
IAES Int J Rob & Autom ISSN: 2722-2586 23
This study effectively demonstrates the connection between the outcomes and approximations
obtained from the provided stochastic differential equation and their association with the path wise
differential equation, as referenced in sources [4]–[7]. More specifically, it showcases how the solution C for
the path wise differential (7) can be inferred from the response A corresponding to the stochastic equation
(1). It's crucial to emphasize that this statement is made under the assumption that equation 6 possesses a
unique global solution, thereby guaranteeing the uniqueness of the local solution for equation 1.
This paper establishes a clear link between the solutions and estimations of stochastic and path wise
differential equations [24], [25]. Additionally, it underscores the significance of unique conditions,
particularly in the context of (6) and its implications for equation 1. Let L represent the solution to (2). It is
well established that the solution is invertible in the neighborhood N of (𝑋0 , 𝑌0 , 𝑡0 ).
𝜕𝐾 𝜕𝐿 −1
(𝑦, 𝑧, 𝑡) = ( (𝐾(𝑦, 𝑧, 𝑡), 𝑦, 𝑡)) (14)
𝜕𝑦 𝜕𝑥
𝜕𝐿 𝜕𝐾
(𝑧, 𝑦, 𝑡) = − ∑𝑛𝑖=0 (𝑧, 𝑦, 𝑡)𝑇 𝑖 (𝑧, 𝑡) (15)
𝜕𝑦 𝜕𝑧 𝑖
𝜕𝐿 𝜕𝐾 𝜕𝐿𝑖
(𝑧, 𝑦, 𝑡) = − ∑𝑛𝑖=0 (𝑧, 𝑦, 𝑡) (𝐾(𝑧, 𝑦, 𝑡), 𝑦, 𝑡) (16)
𝜕𝑡 𝜕𝑧 𝑖 𝜕𝑡
By plugging the parameters of the 𝕊𝑛+1 -valued procedure (𝐴(𝑡), 𝐵𝑟(𝑡)) into the stochastic equation for an
expression K (z, y, t), get (17),
But 𝐴(𝑠) = 𝐿(𝐾(𝐴(𝑠), 𝐵𝑟(𝑠), 𝑠), 𝐵𝑟(𝑠), 𝑠) and (14) holds, hence
satisfies the path-wise (6) on an individual scale. Similarly, it may establish that 𝐴𝑍 (𝑡) ≔ 𝐿(𝐶𝑍 (𝑡), 𝐵𝑟𝑍 (𝑡), 𝑡)
satisfies the path-wise (11) on a particular scale. Let Br approximate a proportionate Brownian motion B z.
Let 𝑆, 𝑇: 𝕊𝑛 ∗ (0, 𝑡) be predetermined in (19) and (20),
𝑡 𝑡
𝐴(𝑡) = 𝐴0 + ∫𝑡 𝑆(𝐴(𝑠), 𝑠)𝑑𝑠 + ∫𝑡 𝑇(𝐴(𝑠), 𝑠)𝑑𝐵𝑟(𝑠). (19)
0 0
𝑡 𝑡
𝐴𝑍 (𝑡) = 𝐴0 + ∫𝑡 𝑆(𝐴𝑍 (𝑠), 𝑠)𝑑𝑠 + ∫𝑡 𝑇(𝐴𝑍 (𝑠), 𝑠)𝑑𝐵𝑟𝑍 (𝑠). (20)
0 0
where 𝑍 ∈ ℤ, accepts, with high probability, a single local solution on the same interval (t 1, t2) (where t0 is
outside of Z but still part of the interval). In addition, the following approximate result has been obtained in
(21).
Optimizing robot anomaly detection through stochastic differential approximation … (Branesh M. Pillai)
24 ISSN: 2722-2586
Problem 2 Find the nearer location from the source point where the data stream W begins and
define optimal path for the data stream W moves from one point to another point.
Proof (a) The anticipated quantity of near-optimal locations for any b is constrained in (22),
ℚ[𝒩𝑏 (𝜂 )] ≤ 6𝜂 2 2𝑏 (22)
1
It fixes the value of 𝜂𝑏 ≜ 𝜂𝜀 ( 𝑏). The speed of growth of 𝒩𝑏 (𝜂𝑏 ), the quantity of 𝜂𝑏 -near-optimal locations
2
in [0, 1] of the form 𝑘/2𝑏 , is measured by the near-optimality dimension in dimension one with the pseudo-
distance𝑙(𝑚, 𝑛) = 𝜂𝜀 (|𝑛 − 𝑚|).It shows that, in general, this quantity grows at a constant rate regarding b.
This implies that there exists a metric where the Brownian is Lipschitz with likelihood at least 1 − 𝜀 and has
a near-optimality aspect = 0 with 𝐸 = 𝒪(𝑙𝑜𝑔(1/𝜀)).
The 𝒪(𝑙𝑜𝑔(1/𝜀)) term, originating from the standard DOO error for deterministic function
optimisation, and a different 𝒪(𝑙𝑜𝑔(1/𝜀)) term, originating from the need to adjust our pseudo-distance
ℒ 𝑡𝑜 𝜀 such that the Brownian is ℒ -Lipschitz with likelihood 1- 𝜀, together represent the finalised study
difficulty bound. Combining these two bounds yields an upper limit on sample complexity of 𝒪(𝑙𝑜𝑔2 (1/𝜀)).
Proof (b) A Brownian motion whose optimum O is reached for the initial time at the location
described as is denoted by U and the Brownian meander 𝑇0+ can be defined as in (23),
𝑂−𝑊(𝑡1−𝑡.𝑡1)
𝑇0+ ≜ (23)
√𝑡1
𝑂−𝑊(𝑡 1+𝑡(1−𝑡1))
𝑇1+ ≜ (24)
√1−𝑡1
Then the theorem 1 declares that 𝑇 + ≦ 𝑇0+ ≦ 𝑇1+ and t1 changes regardless of both 𝑇0+ 𝑎𝑛𝑑 𝑇1+ .
For each positive integer, it establishes a maximum constraint on the predicted amount of 𝜂-near-
optimal positions b>0 and any values of 𝜂 > 0.
2 𝑏 𝑎 𝑏 𝑎
ℚ[𝒩𝑏 (𝜂 )] = ℚ [∑𝑎=0 1 {𝑊 ( 𝑏 ) > 𝑂 − 𝜂}] = ∑2𝑎=0 ℚ [1 {𝑊 ( 𝑏) > 𝑂 − 𝜂}]
2 2
𝑏 𝑎 𝑎 𝑎 𝑎
= ∑2𝑎=0 ℚ [1 {{𝑊 ( 𝑏) > 𝑂 − 𝜂 ∩ ≤ 𝑡1} ∪ {𝑊 ( 𝑏) > 𝑂 − 𝜂 ∩
> 𝑡1}}]
2 2𝑏 2 2𝑏
2𝑏 𝑎 2𝑏
𝑘 𝜂 𝑎 𝑏 – 𝑡1 𝜂 𝑎
+ 2
= ∑ ℚ [1 {𝑇0+ (1 − ) < ∩ ≤ 𝑡1}] + ∑ ℚ [1 {𝑇1 ( )< ∩ 𝑏 > 𝑡1}]
𝑡12𝑏 √𝑡1 2 𝑏 1 − 𝑡1 √1 − 𝑡1 2
𝑎=0 𝑎=0
Since t1 changes regardless of 𝑇0+ 𝑎𝑛𝑑 𝑇1+ , utilizing the above equation with C= (𝑇0+ , 𝑇1+ ), D=t1 and function
in (25),
𝑎
𝑏 𝑘 𝜂 𝑎 𝑏 –𝑡1 𝜂 𝑎
𝑓𝑢𝑛: (𝑐0 , 𝑐1 ), 𝑑 → ∑2𝑎=0 ([1 {𝑐0 (1 − 𝑏) < ∩ ≤ 𝑑}] + 1 {𝑐1 (2 )< ∩ > 𝑑})(25)
𝑡12 √𝑑 2𝑏 1−𝑑 √1−𝑑 2𝑏
= 2 sup{𝛽1 + 𝛽2 + 𝛽3 + 𝛽4 } (26)
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30
IAES Int J Rob & Autom ISSN: 2722-2586 25
𝑣2𝑏
⌊ ⌋
⌊2𝑏 𝜂 2 ⌋ 𝑘 𝜂 2 𝑘 𝜂
𝛽1 = ∑𝑎=0 ℝ [𝑇0+ (1 − )< ], 𝛽2 = ∑𝑎=⌈2𝑏𝜂2 ⌉ ℝ [𝑇0+ (1 − )< ],
𝑣2𝑏 √𝑣 𝑣2𝑏 √𝑣
Given that 1 is the highest possible likelihood, it may simply place a limit on 𝛽1 𝑎𝑛𝑑 𝛽4 as 2𝑏 𝜂 2 , to
obtain that 𝛽1 + 𝛽4 ≤ 2(2𝑏 𝜂 2 ). By accumulating across the Brownian meander distribution parameters, it can
now place upper and lower bounds on the rest of the possibilities occurring in the aforementioned formula.
2
𝑑2 𝑑 exp (− 𝑑 𝑐𝑑 )
) 2(1 − 𝑡) exp (− 2 𝑐
𝑑2
ℝ[𝑇0+ (𝑡) < 𝑐] = 2√2𝜋 ∫ 2𝑡 ∫ 𝑑𝑑 𝑑𝑐 ≤ ∫ 𝑑2 exp (− ) 𝑑𝑑
0 𝑡√𝑡 ∗ 2𝜋 0 𝑡√(1 − 𝑡) ∗ 2𝜋 𝑡√(1 − 𝑡)(𝑡. 2𝜋) 0 2𝑡
3 3 3
2𝑐 2𝑐 2 𝑐
≤ ≤ = ( )
3𝑡√(1 − 𝑡)(𝑡. 2𝜋) 3𝑡(1 − 𝑡)√(1 − 𝑡)(𝑡. 2𝜋) 3√2𝜋 √𝑡(1 − 𝑡)
𝑣2𝑏
⌊ ⌋
2 𝑘 𝜂 ⌊𝑣2𝑏 ⌋−⌈2𝑏 𝜂 2 ⌉ 𝑘 𝜂
𝛽2 + 𝛽3 = ∑𝑎=⌈2𝑏 𝜂2⌉ ℝ [𝑇0+ (1 − )< ]+∑ 𝑣2𝑏
ℝ [𝑇0+ (1 − )< ]
𝑣2𝑏 √𝑣 𝑎=⌈ ⌉ 𝑣2𝑏 √𝑣
2
3 3
𝑣2𝑏
⌊ ⌋ 𝜂 𝜂
2 ⌊𝑣2𝑏 ⌋−⌈2𝑏 𝜂 2 ⌉
2 √𝑣 2 √𝑣
≤ ∑ + ∑
3√2𝜋 3√2𝜋
𝑎=⌈2𝑏 𝜂 2 ⌉ 𝑎 𝑎 𝑣2𝑏 𝑎 𝑎
√(1 − 𝑏 √ 𝑏 ) 𝑎=⌈
2
⌉ √(1 − √ )
( 𝑣2 𝑣2 ) ( 𝑣2𝑏 𝑣2𝑏 )
3
𝑣2𝑏 3
⌊ ⌋ 𝜂 𝜂
2 ⌊𝑣2𝑏 ⌋−⌈2𝑏 𝜂 2 ⌉
1 √𝑣 1 √𝑣
≤ ∑ + ∑ ≤
6√𝜋 6√𝜋 𝑎
𝑎=⌈2𝑏 𝜂 2 ⌉ 𝑎 𝑣2𝑏 √1 −
√√ 𝑏 𝑎=⌈
2
⌉ ( 𝑣2𝑏 )
( 𝑣2 )
3
𝑣2𝑏 2𝑏 𝑣
⌊ ⌋ 𝜂 ⌊𝑣2𝑏 ⌋−⌈ ⌉ 𝜂 3
2 2
1 √𝑣 1 √𝑣
∑ + ∑
6√𝜋 6√𝜋 ⌊2𝑏 𝑣⌋ 𝑎
𝑎=⌈2𝑏 𝜂 2 ⌉ 𝑎 𝑎=⌈2𝑏 𝜂 2 ⌉ √
√√ ( 𝑣2𝑏 + 𝑣2𝑏 )
( 𝑣2𝑏 )
𝑣2𝑏 2𝑏 𝑣
⌊ ⌋ ⌊𝑣2𝑏 ⌋−⌈ ⌉
2 2
(2𝑏 𝜂2 )3/2 1 1
𝛽2 + 𝛽3 ≤ ∑ + ∑
6√𝜋 𝑔3/2 𝑔3/2
𝑎=⌈2𝑏 𝜂 2 ⌉ 𝑎=⌈2𝑏 𝜂 2 ⌉
( )
3 ∝
(2𝑏 𝜂2 )2 1
≤ ∑ 3
3√𝜋
𝑎=⌈2𝑏 𝜂 2 ⌉ 𝑔 2
3
(2𝑏 𝜂 2 )2 3 1
≤ ≤ 2𝑏 𝜂 2 ≤ 2𝑏 𝜂 2 ,
3√𝜋 √⌈2𝑏 𝜂 2 ⌉ √𝜋
Hence, ℚ[𝒩𝑏 (𝜂 )] ≤ 6𝜂 2 2𝑏
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30
IAES Int J Rob & Autom ISSN: 2722-2586 27
All three samples follow a similar trend with their own peculiar characteristics. Sample 1 (black)
exhibits a steady, almost linear decline until it flattens out towards the end of the domain. Sample 2 (red) is
characterized by a significant oscillation before sharply dropping, indicating a variable response or
measurement before reaching a similar level as sample 1 towards the end. Sample 3 (blue) mirrors the
oscillation seen in sample 2, but with a less pronounced initial drop and a deeper final descent. This
comparison allows for an easy assessment of the similarities and differences in behavior, or responses
captured by the three samples across the range of 'n' values. Figure 3 presents a time-series linear
approximation of Brownian motion, reflecting data stream flow dynamics within a unit interval. The
visualization begins at zero and rapidly reaches a maximum at T=0.2, after which it moderately recedes to a
plateau of 0.8, sustained until T=0.6. Subsequently, the series undergoes a pronounced drop to 0.6, stabilizes
momentarily, then descends precipitously to 0.2, culminating in a final downturn back to zero as T
approaches 1. This portrayal suggests a piecewise linear process with distinct, sustained levels before
transitioning, indicative of a system exhibiting stepwise stability before entering new phases. Figures 4 to 6
offer insights into the behavior of data stream movements and their relationship with the parameter λ.
Figure 4 demonstrates a minimum Brownian path with characteristic rise-and-fall patterns over
time, displaying the inherent randomness of the motion through peaks and troughs between T=0 and T=1. In
contrast, Figure 5 shows a decreasing trend in the number of movements starting from 40 at λ=0 and
diminishing as λ increases, suggesting an inverse relationship. Meanwhile, Figure 6 contrasts this by
depicting a direct, linear correlation between λ and the total number of observations, which increases
proportionally from 110 to 145 as λ grows from 0 to 5. This comparison highlights the significance in which
λ values change data stream behavior, with the rise in observations possibly pointing to increased data
collection or detection capabilities as λ grows, despite the decrease in movement frequency. Together, these
figures aid in identifying anomalies between different data stream paths by evaluating the movement cost
over time.
Figure 3. Stochastic differential equation with linear approximation for Brownian motion
Optimizing robot anomaly detection through stochastic differential approximation … (Branesh M. Pillai)
28 ISSN: 2722-2586
4. CONCLUSION
Anomaly detection is achieved by repositioning a data stream from one location to another, with
movement costs evaluated through Brownian motion. The process involves applying a mathematical model
based on the stochastic differential equation of Brownian motion. By optimizing the Brownian motion over
time, the nearest position to the anomaly is identified. The model developed and validated in this work
effectively approximates the minimum path required to detect anomalies. Its strength lies in its ability to
accomplish the task within a relatively short crossing time, while accurately identifying anomalies through
the observation of movement costs. This approach proves to be a valuable tool for efficiently detecting
anomalies, offering both speed and accuracy in the process.
ACKNOWLEDGEMENTS
This project is supported by the Reinventing University System through Mahidol University (IO
864102063000), and This research is partially supported by the e-ASIA Joint Research Program (P-19-
50869) Grant through the National Science and Technology Development Agency (NSTDA) and Mahidol
University, Thailand.
REFERENCES
[1] R. A. Ariyaluran Habeeb, F. Nasaruddin, A. Gani, I. A. Targio Hashem, E. Ahmed, and M. Imran, “Real-time big data processing
for anomaly detection: A Survey,” International Journal of Information Management, vol. 45, pp. 289–307, Apr. 2019, doi:
10.1016/j.ijinfomgt.2018.08.006.
[2] R. Hu, X. Yuan, Y. Qiao, B. Zhang, and P. Zhao, “Unsupervised anomaly detection for multivariate time series using diffusion
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30
IAES Int J Rob & Autom ISSN: 2722-2586 29
model,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Apr.
2024, pp. 9606–9610. doi: 10.1109/icassp48485.2024.10447083.
[3] M. Ma et al., “Jump-starting multivariate time series anomaly detection for online service systems,” in 2021 USENIX Annual
Technical Conference, 2021, pp. 413–426.
[4] Z. Yang, Y. Wang, and F. L. Lewis, “Practical prescribed-time tracking control for uncertain strict-feedback systems with
guaranteed performance under unknown control directions,” Journal of Automation and Intelligence, vol. 2, no. 2, pp. 99–104,
May 2023, doi: 10.1016/j.jai.2023.02.001.
[5] S. He, K. Shi, C. Liu, B. Guo, J. Chen, and Z. Shi, “Collaborative sensing in Internet of Things: A comprehensive survey,” IEEE
Communications Surveys and Tutorials, vol. 24, no. 3, pp. 1435–1474, 2022, doi: 10.1109/COMST.2022.3187138.
[6] K. Rajwar, K. Deep, and S. Das, “An exhaustive review of the metaheuristic algorithms for search and optimization: taxonomy,
applications, and open challenges,” Artificial Intelligence Review, vol. 56, no. 11, pp. 13187–13257, Apr. 2023, doi:
10.1007/s10462-023-10470-y.
[7] S. B. Ramezani et al., “Scalability, explainability and performance of data-driven algorithms in predicting the remaining useful
life: A comprehensive review,” IEEE Access, vol. 11, pp. 41741–41769, 2023, doi: 10.1109/ACCESS.2023.3267960.
[8] Z. Wang, V. Y. F. Tan, and J. Scarlett, “Tight regret bounds for noisy optimization of a Brownian motion,” IEEE Transactions on
Signal Processing, vol. 70, pp. 1072–1087, 2022, doi: 10.1109/TSP.2022.3144939.
[9] A. Ustimenko and A. Beznosikov, “Ito diffusion approximation of universal Ito chains for sampling, optimization and boosting,”
arXiv:2310.06081, Oct. 2023.
[10] L. Kapllani and L. Teng, “Multistep schemes for solving backward stochastic differential equations on GPU,” Journal of
Mathematics in Industry, vol. 12, no. 1, Jan. 2022, doi: 10.1186/s13362-021-00118-3.
[11] D. Higham and P. Kloeden, An introduction to the numerical simulation of stochastic differential equations. Society for Industrial
and Applied Mathematics, 2021. doi: 10.1137/1.9781611976434.
[12] F. Iafrate and E. Orsingher, “Some results on the Brownian meander with drift,” Journal of Theoretical Probability, vol. 33, no. 2,
pp. 1034–1060, Mar. 2020, doi: 10.1007/s10959-019-00891-3.
[13] E. Grong and S. Sommer, “Most probable paths for anisotropic Brownian motions on manifolds,” Foundations of Computational
Mathematics, vol. 24, no. 1, pp. 313–345, Sep. 2024, doi: 10.1007/s10208-022-09594-4.
[14] Y. Chen, J. Dong, and H. Ni, “Ε-strong simulation of fractional Brownian motion and related stochastic differential equations,”
Mathematics of Operations Research, vol. 46, no. 2, pp. 559–594, May 2021, doi: 10.1287/moor.2020.1078.
[15] S. M. J. Khadem, S. H. L. Klapp, and R. Klages, “Search efficiency of discrete fractional Brownian motion in a random
distribution of targets,” Physical Review Research, vol. 3, no. 2, Jun. 2021, doi: 10.1103/PhysRevResearch.3.023169.
[16] J. Urain, M. Ginesi, D. Tateo, and J. Peters, “ImitationFlow: Learning deep stable stochastic dynamic systems by normalizing
flows,” in IEEE International Conference on Intelligent Robots and Systems, Oct. 2020, pp. 5231–5237. doi:
10.1109/IROS45743.2020.9341035.
[17] B. M. Pillai and J. Suthakorn, “Challenges for novice developers in rough terrain rescue robots: A survey on motion control
systems,” Journal of Control Science and Engineering, vol. 2019, pp. 1–12, Jun. 2019, doi: 10.1155/2019/2135914.
[18] D. M. G. A. I. N. Sumanarathna, I. A. S. R. Senevirathna, K. L. U. Sirisena, H. G. N. Sandamali, M. B. Pillai, and A. M. H. S.
Abeykoon, “Simulation of mobile robot navigation with sensor fusion on an uneven path,” in 2014 International Conference on
Circuits, Power and Computing Technologies, ICCPCT 2014, Mar. 2014, pp. 388–393. doi: 10.1109/ICCPCT.2014.7054830.
[19] H. A. N. D. Dayarathna, L. L. G. Prabuddha, K. L. D. N. J. Ariyawansha, M. K. C. D. Chinthaka, A. M. H. S. Abeykoon, and M.
Branesh Pillai, “Sensorless contact position estimation of a mobile robot in pushing motion,” in Proceedings of IEEE
International Conference on Circuit, Power and Computing Technologies, ICCPCT 2013, Mar. 2013, pp. 344–349. doi:
10.1109/ICCPCT.2013.6529009.
[20] B. Sharma, B. M. Pillai, and J. Suthakorn, “Live displacement estimation for rough terrain mobile robot: BART LAB rescue
robot,” in SIBCON 2021 - International Siberian Conference on Control and Communications, May 2021, pp. 1–6. doi:
10.1109/SIBCON50419.2021.9438919.
[21] P. Mörters, Y. Peres, O. Schramm, and W. Werner, Brownian motion, vol. 9780521760188. Cambridge University Press, 2010.
doi: 10.1017/CBO9780511750489.
[22] X. Huang, Z. Li, and F. L. Lewis, “Cost-effective distributed FTFC for uncertain nonholonomic mobile robot fleet with collision
avoidance and connectivity preservation,” Journal of Automation and Intelligence, vol. 2, no. 1, pp. 42–50, Feb. 2023, doi:
10.1016/j.jai.2023.100021.
[23] T. P. Raptis and A. Passarella, “A survey on networked data streaming with Apache Kafka,” IEEE Access, vol. 11, pp. 85333–
85350, 2023, doi: 10.1109/ACCESS.2023.3303810.
[24] X. Fang, J. Li, X. Li, and L. Xie, “2-D distributed pose estimation of multi-agent systems using bearing measurements,” Journal
of Automation and Intelligence, vol. 2, no. 2, pp. 70–78, May 2023, doi: 10.1016/j.jai.2023.06.002.
[25] M. B. Pillai, S. Nakdhamabhorn, K. Borvorntanajanya, and J. Suthakorn, “Enforced acceleration control for DC actuated rescue
robot,” in Proceedings - 2016 22nd International Conference on Electrical Machines, ICEM 2016, Sep. 2016, pp. 2640–2648.
doi: 10.1109/ICELMACH.2016.7732894.
BIOGRAPHIES OF AUTHORS
Optimizing robot anomaly detection through stochastic differential approximation … (Branesh M. Pillai)
30 ISSN: 2722-2586
Jackrit Suthakorn received his Ph.D. in Robotics from The Johns Hopkins
University, USA in 2003, a master's in Control Engineering from Michigan Technological
University, USA in 1998, and a bachelor’s degree in mechanical engineering from Mahidol
University, Thailand in 1995. He established the Center for Biomedical and Robotics
Technology (BART LAB), the first interdisciplinary research laboratory in the Faculty of
Engineering, Mahidol University, Thailand in 2004. In 2006, he and his colleagues
established the Department of Biomedical Engineering at Mahidol University. He was the
first BME Department Chair and remained until 2015. From June 2015 to July 2023, he
served as the Dean of the Faculty of Engineering at Mahidol University. Currently, he is
also the Chair of the IEEE RAS Thailand Chapter. His research interest is aimed towards
Medical Robotics and Field Robotics: Surgical Robotics in various applications,
Rehabilitation Robotics, Hospital Service and Tele-Medicine Robotics, and Rough Terrain
Rescue Robotics. He was one of the founders of the Thai Robotics Society (TRS) in 2004
and the President from 2007 to 2010. Currently, he also serves as a Trustee of the
International RoboCup Federation, and the Vice President of RoboCup Asia-Pacific
(RCAP). He can be contacted at [email protected].
IAES Int J Rob & Autom, Vol. 14, No. 1, March 2025: 19-30