Data center operation and maintenance robot control method based on large model

Abstract

The invention discloses a data center operation and maintenance robot control method and system based on a large model, and relates to the technical field of intelligent robot control, and the method comprises the steps: employing extended Kalman filtering to generate a current state estimation vector, generating a prediction state through discrete Kalman filtering, combining a combined filtering mechanism to fuse the current state and the prediction state, and obtaining a prediction result. The method comprises the steps of obtaining a global state estimation vector of a robot, inputting the global state estimation vector into a Gaussian mixture model for abnormal probability prediction, determining the generation of a target state of the robot according to a prediction result, optimizing a KFOSMC parameter based on the target state and the global state estimation vector, and generating a robot control signal by using the optimized parameter. Current state estimation and prediction states are fused through a joint filtering mechanism, efficient and stable operation of the robot in a complex data environment is ensured, abnormal probability prediction is carried out through a Gaussian mixture model, the possibility of fault occurrence is accurately predicted in a variable environment, and the defect that a traditional method depends on threshold judgment is overcome.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot control technology, and in particular to a control method for data center operation and maintenance robots based on a large model. Background Technology

[0002] With the rapid development of robotics, artificial intelligence, and automation technologies, robots are increasingly widely used in industries such as manufacturing, healthcare, and logistics. In the field of data center operations and maintenance, robots, as important automated equipment, can greatly improve operational efficiency and equipment management through integrated multi-sensor systems, precise motion control systems, and intelligent algorithms. Currently, data center operations and maintenance robots generally use state estimation methods to obtain the interaction state between the robot and the environment. However, existing technologies still face many challenges. Although existing technologies have made progress in robot state estimation and control signal generation, their shortcomings significantly limit the actual effectiveness in data center operations and maintenance. First, in terms of multi-sensor data fusion, traditional methods often employ simple weighted averaging or independent filtering mechanisms, which fail to fully utilize the complementarity of data from different sources, resulting in insufficient accuracy in global state estimation, especially when dealing with nonlinear dynamic systems. Second, existing anomaly probability prediction techniques (such as detection based on a single GMM) are less adaptable to complex fault modes and are prone to misjudgment or omission due to limitations in model assumptions, affecting the robot's timely response to abnormal states. Furthermore, in the process of control signal optimization, the parameter adjustment of traditional sliding mode control relies heavily on human experience and lacks adaptive optimization capabilities based on real-time states, leading to low efficiency in adjusting the robot's motion trajectory in dynamic environments. These problems not only reduce the reliability of robots in data center operations and maintenance but also limit their autonomous decision-making capabilities under high loads or abnormal scenarios. Our invention significantly improves the accuracy of state estimation, the reliability of anomaly prediction, and the adaptability of control signals by fusing extended Kalman filtering and discrete Kalman filtering results through a joint filtering mechanism and combining Gaussian mixture models with optimized KFOSMC parameters. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a control method for data center operation and maintenance robots based on a large model. It solves the problems of existing traditional methods that use simple weighted averaging or independent filtering mechanisms, which make it difficult to make full use of the complementarity of data from different sources, resulting in insufficient accuracy of global state estimation. Secondly, existing anomaly probability prediction technology is prone to misjudgment or omission due to the limitations of model assumptions, which affects the robot's timely response to abnormal states.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a control method for a data center operation and maintenance robot based on a large model, comprising,

[0007] Multiple sensors are deployed on the robot to collect and preprocess data. The position of the end effector is calculated using the forward kinematics formula, and the cylindrical coordinates are calculated based on the position of the end effector.

[0008] Based on the preprocessed data, feature vectors are generated using the BERT method, and linearized matrices are generated using Koopman theory.

[0009] An extended Kalman filter is used to generate the current state estimation vector, and a discrete Kalman filter is used to generate the predicted state. The current state and the predicted state are fused by a joint filtering mechanism to obtain the robot's global state estimation vector. The global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. The generation of the robot's target state is determined based on the prediction result. The KFOSMC parameters are optimized based on the target state and the global state estimation vector, and the robot control signal is generated using the optimized parameters.

[0010] The robot control signals are mapped to PWM signals for instruction execution. The state of the instruction execution process is monitored, and the control signals are corrected through state feedback. The robot's motion trajectory and dynamic obstacle avoidance are optimized based on the corrected control signals, and adaptive optimization is performed.

[0011] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the following steps are included: deploying multiple sensors on the robot to collect and preprocess data; calculating the position of the end effector using forward kinematics formulas; calculating cylindrical coordinates based on the position of the end effector; and deploying multiple types of sensors on the robot to collect data, preprocessing the collected data, and calculating the end effector position coordinates P of the robot joints using forward kinematics formulas. n Calculate the position coordinates P of the end effector. n The distance r between the projection points on the x-axis and y-axis planes, and the coordinate P n The cylindrical coordinates Y of the end effector are obtained by integrating the calculation results of the angle θ between the y-axis and the x-axis in the y-plane and the height h in the z-axis. n =[r n ,θ n ,h n ];

[0012] The various types of sensors include temperature sensors, high-definition cameras, MEMS microphones, data center APIs, joint encoders, and lidar sensors.

[0013] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the following steps are taken: The preprocessed data is used to generate feature vectors using the BERT method, and a linearized matrix is ​​generated using Koopman theory. The preprocessed data is then input into the BERT model, and the data is processed through the Transformer architecture in the BERT model to obtain the data feature vector V. BERT Using Koopman theory, the linearized Koopman operator K is calculated. g The dynamic mode decomposition method is used to decompose the feature vectors V at multiple time points. BERT Construct the state observation matrix, and use the least squares method to extract the Koopman operator K from the state observation matrix. g Approximately linearized matrix D t .

[0014] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the following steps are employed: An extended Kalman filter is used to generate a current state estimation vector, and a discrete Kalman filter is used to generate a predicted state. A joint filtering mechanism is then used to fuse the current and predicted states to obtain the robot's global state estimation vector. This global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. Based on the prediction results, the generation of the robot's target state is determined. Based on the target state and the global state estimation vector, the KFOSMC parameters are optimized. The optimized parameters are used to generate the robot control signal. An extended Kalman filter is then used to calculate the current state estimation vector X based on the observation data obtained from the sensors. t The current state estimation vector X t With observation data Z t The residuals are calculated by comparison, and the residual values ​​are used to correct the state estimation vector X. t The optimized state estimation vector X″ is obtained. t , the feature vector V BERT With the state estimation vector X″ t The fusion process is performed to obtain the fused state estimation vector K;

[0015] Discrete Kalman filtering is used to predict the robot's motion state, resulting in the state vector X′ for the next moment. t+1 A joint filtering mechanism is employed, using a weighted average method to fuse K and X′. t+1 The estimation results yield the global state estimation vector A. t ;

[0016] The global state estimation vector is input into a Gaussian mixture model to calculate the probability density function Q, and the eigenvector V is evaluated using the log-likelihood method. BERTThe anomaly probability W is set, and an anomaly probability threshold L is defined. If the anomaly probability W is less than the threshold L, the current state is considered to be without anomalies. A is estimated based on the current global state. t Generate target state M t Calculate the state error e between the current global state and the target state. t The sliding mode control gain matrix G in KFOSMC is optimized using gradient descent. i The state error e is calculated using first-order difference. t error change rate e′ t Based on the optimized KFOSMC parameter sliding mode control gain matrix G i and the rate of change of error e′ t Generate control signal U t ;

[0017] If the anomaly probability W is greater than or equal to the threshold L, the current state is considered to be abnormal, and the system enters anomaly mitigation mode. A corrected target state is generated through a weighted backoff mechanism, and the error E between the current state and the corrected target state is calculated. t and the rate of change of error E′ t With sliding mode control gain matrix G′ i The control signal U′ is generated through the sliding mode control algorithm. t .

[0018] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the step of mapping the robot control signal to a PWM signal for instruction execution refers to adopting a three-stage PWM control method, which maps the control signal to a duty cycle, generates a PWM signal based on the calculated duty cycle and the set PWM frequency, and transmits the PWM signal to the actuator. The robot actuator precisely adjusts its actions according to the PWM signal and regulates the robot's motion state through digital control.

[0019] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the following steps are included: Monitoring the state of the instruction execution process and correcting the control signal through state feedback; establishing communication with the robot controller via Ethernet TCP; monitoring controller data and storing relevant information in real time through a ROS node during robot instruction execution; extracting state information for each motion instruction using data tagging and indexing methods based on the robot's execution data; calculating the robot's load vibration signal at each moment using the torque signal obtained from the controller; inputting the load vibration signal into a Hidden Markov Model for anomaly probability prediction; if an anomaly is predicted, the robot will enter an anomaly mitigation mode and generate a corrected control signal U′. t .

[0020] As a preferred embodiment of the data center operation and maintenance robot control method based on a large model described in this invention, the step of optimizing the robot's motion trajectory and dynamic obstacle avoidance according to the modified control signal and performing adaptive optimization refers to optimizing the path from the current position to the target position using the A* algorithm based on the modified control signal, planning the global path using cylindrical coordinates, calculating the actual distance g(O) from the current position to the target position, calculating the heuristic estimated cost N(O) using Euclidean distance, defining the path evaluation function J(O), estimating the position and velocity of dynamic obstacles using real-time data collected by sensors combined with the Dynamic Window Algorithm (DWA), calculating obstacle avoidance strategies in real time to adjust the robot's motion trajectory, and optimizing path errors and control constraints using a model predictive control algorithm.

[0021] Secondly, the present invention provides a control method and system for a data center operation and maintenance robot based on a large model, including a robot data acquisition and kinematics calculation module for data acquisition and preprocessing, and calculating the position of the end effector;

[0022] The data feature generation and linearization module is used to generate feature vectors using the BERT method and to generate linear matrices using the Koopman theory.

[0023] The state estimation and prediction module is used to generate the current state estimation vector using extended Kalman filtering, predict the state using discrete Kalman filtering, and fuse the current state and the predicted state through a joint filtering mechanism.

[0024] The anomaly prediction and control signal generation module is used to input the global state estimation vector into the Gaussian mixture model to predict the anomaly probability and generate control signals.

[0025] The feedback correction and adaptive optimization module is used to monitor and correct the control signal in real time, and to perform adaptive optimization based on the corrected target state and control signal.

[0026] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the control method for a data center operation and maintenance robot based on a large model as described in the first aspect of the present invention.

[0027] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the control method for a data center operation and maintenance robot based on a large model as described in the first aspect of the present invention.

[0028] The beneficial effects of this invention are as follows: By combining extended Kalman filtering with discrete Kalman filtering, the accumulation of errors caused by sensor errors and dynamic environmental changes is reduced, improving the accuracy and reliability of global state estimation. The joint filtering mechanism effectively integrates the current state estimation and the predicted state, further optimizing the state estimation process and ensuring the efficient and stable operation of the robot in the complex environment of the data center. The introduction of Gaussian mixture model for anomaly probability prediction accurately predicts the possibility of failure in a variable environment, avoiding the shortcomings of traditional methods that rely on threshold judgment. The combination of feature vectors generated by the BERT method and linearized models of Koopman theory greatly improves the robot's adaptability to dynamically changing environments. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of a data center operation and maintenance robot control method based on a large model in Example 1.

[0031] Figure 2 This is a schematic diagram of a data center operation and maintenance robot control system based on a large model, as shown in Example 1.

[0032] Figure 3 This is a flowchart of the control signal generation process in Example 1.

[0033] Figure 4 This is a flowchart of the execution of control signals and status monitoring in Example 1. Detailed Implementation

[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0036] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0037] Example 1, referring to Figures 1 to 4 This is the first embodiment of the present invention, which provides a control method for a data center operation and maintenance robot based on a large model, including the following steps:

[0038] S1. Deploy multiple sensors on the robot to collect and preprocess data, use the forward kinematics formula to calculate the position of the end effector, and calculate the cylindrical coordinates based on the position of the end effector;

[0039] Based on the preprocessed data, feature vectors are generated using the BERT method, and linearized matrices are generated using Koopman theory.

[0040] Specifically, multiple sensors are deployed on the robot to collect and preprocess data. The position of the end effector is calculated using forward kinematics formulas. Based on the position of the end effector, cylindrical coordinates are calculated. This involves deploying various types of sensors on the robot to collect data, preprocessing the collected data, and then using forward kinematics formulas to calculate the end effector position coordinates P of the robot joints. n :

[0041] P n =P n-1 +l n ·[-sin(q n-1 +q n ),cos(q n-1 +q n )],

[0042] Among them, P n Let P be the position coordinates of the nth end effector. n-1 The coordinates of the joint preceding the end effector (referring to the moving point in the robot that connects the various "links"), l n q represents the length of the link where the end effector is located (the length of the link between two adjacent joints in a robot). n-1 and q n P is the angle of the joint where the end effector is located. n-1 l n q n-1 and q n It is calculated based on the position of the previous joint. Starting from the base position P0, the coordinates of each joint position P are calculated through continuous recursion. n ;

[0043] Calculate the position coordinates P of the end effector n The distance r between the projection points on the x-axis and y-axis planes:

[0044]

[0045] Calculate the coordinates P of the end effector n The angle θ between the x-axis and the x-axis plane:

[0046]

[0047] Collect end effector coordinates P n At a height h along the z-axis, where h = z, the calculation results are integrated to obtain the cylindrical coordinates Y of the end effector. n =[r n ,θ n ,h n ];

[0048] The various types of sensors include temperature sensors, high-definition cameras, MEMS microphones, data center APIs, joint encoders, and lidar sensors (LiDAR).

[0049] By using multi-sensor collaborative acquisition and preprocessing, combined with forward kinematics recursive calculation, the accuracy of end effector position calculation is effectively improved, avoiding the shortcomings of traditional single sensor or simple fusion methods in positioning inaccurate in complex environments. In addition, the introduction of cylindrical coordinates provides a more intuitive geometric expression for subsequent path planning and dynamic obstacle avoidance. Compared with the lack of efficient spatial representation in the background technology, this further enhances the robot's motion control capability and environmental adaptability in data center operation and maintenance.

[0050] Furthermore, feature vectors are generated using the BERT method based on the preprocessed data, and a linearized matrix is ​​generated using Koopman theory. The preprocessed data is then input into the BERT model, where the Transformer architecture within the BERT model processes the data to obtain the data feature vector V. BERT :

[0051] V BERT =BERT(Transformer(a)),

[0052] Where 'a' represents the preprocessed data;

[0053] Using Koopman theory, the robot's original nonlinear dynamic system is transformed into a linear system, and the linearized Koopman operator K is calculated. g :

[0054]

[0055] Where g is the observation function, a function that maps the robot's state vector to a high-dimensional feature space, g(λ) t )=[1,α t ,ω t ,α t 2 ,α t ω t ,ω t 2 ,...], where 1 is a constant term, α t 2 and ω t 2 Let α be the squared term of the state variable. t ω t Let F be the interaction term of two state variables, and F be the dynamic evolution function (e.g., the robot's position and velocity form the state vector λ). t =[α t ,ω t If F is an integer, then F is represented as:

[0056]

[0057] Where, α t Let ω be the position of the robot at time t. t Let ι be the velocity of the robot at time t. t To represent the robot's acceleration at time t, Δt is the time step, and ° is the function that transforms the robot based on its current state at each time step. The dynamic evolution function F is applied to calculate the state at the next time step, and the observation function g is applied to this new state to obtain the final observation value.

[0058] The Dynamic Mode Decomposition (DMD) method is used to decompose the feature vectors V at multiple time points. BERT Construct the state observation matrix. The observation matrix of known data at the next time step (S t For time t1 to t n The state matrix between, S t+1 Let time t n to t n+1 The state matrix between, For the total number of times at each moment, the Koopman operator K is extracted from the state observation matrix using the least squares method. g Approximately linearized matrix D t :

[0059] S t+1 =K g S t ,

[0060] S t+1 ≈D t S t ,

[0061] Among them, D t Let be the linearized matrix of the state transition from time t to time t+1;

[0062] Fit matrix D using the least squares method t , so that matrix D t Able to minimize the actual value S t+1 and predicted value DS t The difference between them:

[0063] Min||S t+1 -D t S t ||,

[0064] The optimal linearized state transition matrix D is calculated using the normal equation method. t :

[0065]

[0066] in, For S t The Moore-Penrose pseudoinverse is calculated using the singular value decomposition (SVD) method to solve for the generalized inverse of a matrix.

[0067] Traditional methods face difficulties in modeling and state prediction accuracy when dealing with nonlinear dynamic systems. This invention utilizes the Transformer architecture of the BERT model to extract high-dimensional feature vectors from preprocessed data, effectively capturing complex data patterns and overcoming the limitations of traditional feature extraction methods. Furthermore, it linearizes the nonlinear dynamic system using Koopman theory and generates accurate linearized matrices using Dynamic Mode Decomposition (DMD) and least squares methods. This not only simplifies system modeling but also improves the reliability of state transition prediction. Compared to the shortcomings of linear approximation or empirical adjustments in background technologies, this significantly enhances the adaptability and control accuracy of data center operation and maintenance robots in dynamic environments.

[0068] S2. The current state estimation vector is generated by using extended Kalman filtering, and the predicted state is generated by using discrete Kalman filtering. The current state and the predicted state are fused by a joint filtering mechanism to obtain the robot's global state estimation vector. The global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. The generation of the robot's target state is determined based on the anomaly prediction result. The KFOSMC parameters are optimized based on the target state and the global state estimation vector. The optimized parameters are used to generate the robot control signal.

[0069] The robot control signals are mapped to PWM signals for instruction execution.

[0070] Specifically, an extended Kalman filter is used to generate the current state estimation vector, and a discrete Kalman filter is used to generate the predicted state. A joint filtering mechanism is then used to fuse the current and predicted states to obtain the robot's global state estimation vector. This global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. Based on the anomaly prediction results, the generation of the robot's target state is determined. The KFOSMC parameters are optimized based on the target state and the global state estimation vector. The optimized parameters are used to generate the robot's control signals. The extended Kalman filter is then used to calculate the current state estimation vector X based on the observation data obtained from sensors (such as vision and LiDAR). t :

[0071] X t =X′ t +ε t (Z t -f(X′ t )),

[0072] Among them, Z t X′ is the observation data obtained through the sensor. t f(X′) is the current state vector, calculated from the state vector of the previous time step. t ) represents the observation model, such as:

[0073]

[0074] Where, α t , ζ t v represents the robot's position. t Let σ be the speed of the robot. t ε is the robot's angular velocity. t The Kalman gain is obtained by minimizing the error covariance of the state estimation, reflecting the error weights between the prediction and the observation.

[0075] The current state estimation vector X t With observation data Z t The residuals are calculated by comparison, and the residual values ​​are used to correct the state estimation vector X. t The optimized state estimation vector X″ is obtained. t (Based on the magnitude of the residuals and the weights of the Kalman gain, the error covariance matrix of the state estimate is updated, thereby correcting the state estimate vector. By continuously iterating with new observation data and calculated residuals, the error covariance matrix and the state estimate vector are updated until the error between the estimated state and the true state is less than a set threshold, thus obtaining the optimal state estimate vector X.) t(Approximately the true value), the feature vector V BERT With the state estimation vector X″ t By fusing the data, the current state estimate of the robot can be further optimized:

[0076] K = l(X″) t V BERT )+κ t ,

[0077] Where K is the fused state estimation vector of the robot at the current moment, l(·) is the fusion module, which uses extended Kalman filtering to integrate state estimation vectors from different sources, κ t To observe the noise, the deviation caused by sensor error is obtained by using the Kalman filtering algorithm;

[0078] The Discrete Kalman Filter (DKF) is used to predict the robot's motion state. The state prediction equation is as follows:

[0079] X′ t+1 =D t X′ t +B t u t ,

[0080] Where, X′ t+1 B is the predicted state vector for the next time step. t The control input matrix is ​​obtained by deriving the effect of the control input on state changes, u t To control inputs (such as speed and acceleration), adjustments are made based on the robot's target state, current state, and task requirements;

[0081] A joint filtering mechanism is employed, using a weighted average method to fuse K and X′. t+1 The estimation results yield the global state estimation vector A. t :

[0082]

[0083] Where, p n (t) represents the nth global covariance after joint filtering, which is obtained through weighted merging. and The matrix is ​​obtained by adding the matrices and then inverting them. and The inverse of the covariance matrix is ​​calculated using Gaussian elimination for the Extended Kalman Filter (EKF) and Discrete Kalman Filter (DKF).

[0084] The global state estimation vector is input into the Gaussian mixture model to calculate the probability density function:

[0085]

[0086]

[0087] Where Q is the output probability value of the formula, and τ is all the parameters of the model, including π. k μ k ∑k parameters, π k The mixing coefficient of the k-th Gaussian distribution is obtained using the EM (Expectation-Maximization) algorithm, μ. k Let be the mean vector of the k-th Gaussian distribution, estimated from the expected value of each Gaussian distribution using training data; ∑k is the covariance matrix of the k-th Gaussian distribution, estimated using the EM algorithm; η(A) t |μ k ∑k) is the probability density function of a multidimensional Gaussian distribution (normal distribution), is the transpose operation used to convert a column vector into a row vector, and d is the dimension of the data, i.e., the dimension of each data point A. t The number of features, where I is the total number of variables;

[0088] The eigenvector V is evaluated using log-likelihood. BERT The probability of an anomaly W:

[0089] W = 1 - sigmoid(logQ(V) BERT ∣τ)),

[0090] Wherein, logQ(V) BERT |τ) is the log-likelihood value calculated by the Gaussian mixture model, and sigmoid is the activation function, which means mapping the log-likelihood value to the range [0,1] to obtain an anomaly probability between 0 and 1;

[0091] An anomaly probability threshold L is set using cross-validation. If the anomaly probability W is less than the threshold L, the current state is considered to be without anomalies. A is then estimated based on the current global state. t Generate target state M t (The state the robot needs to achieve):

[0092] M t =δ planner (A t (task)

[0093] Where, δ planner As a path planner, it estimates and calculates the robot's target state based on the task objective and the current state. The task is the objective that the robot needs to complete in the current environment (including target position, target posture, and behavioral objective), which is obtained through task definition.

[0094] Calculate the state error e between the current global state and the target state. t :

[0095] e t =M t -A t ,

[0096] Optimize the sliding mode control gain matrix G in KFOSMC using gradient descent. i :

[0097] G i (t+1) =G i (t) +ρ·sign(e t ),

[0098] Among them, G i (t+1) Let G be the i-th optimized control gain matrix. i (t) The current gain matrix is ​​obtained through an iterative process of error feedback and gain adjustment, where ρ is the learning rate, determining the parameter adjustment step size, and is obtained through cross-validation. sign(e t ) represents the direction of the current error, when the error e t When >0, sign(e) t ) = 1 indicates that the gain adjustment direction is positive, meaning the system needs to reduce the current error. When the error e t When <0, sign(e) t ) = -1 indicates that the gain adjustment direction is reversed, meaning the system needs to adjust the gain to make the error change in the direction of reduction. t When = 0, sign(e) t ) = 0;

[0099] The state error e is calculated using the first-order difference. t error change rate e′ t Based on the optimized KFOSMC parameter sliding mode control gain matrix G i and the rate of change of error e′ t Generate control signal U t :

[0100] U t =-G i ·sat(e t )+j i ·e′ t ,

[0101] Where sat(·) is a saturation function, for example:

[0102]

[0103] `max val` is the maximum value of the defined saturation range, limiting the range of the control signal and avoiding over-control. i The gain of the i-th error derivative is obtained by manual adjustment based on performance requirements;

[0104] If the anomaly probability W is greater than or equal to the threshold L, the current state is considered to be abnormal, and the system enters anomaly mitigation mode. A corrected target state is generated through a weighted backoff mechanism, and the error E between the current state and the corrected target state is calculated. t and the rate of change of error E′ t With sliding mode control gain matrix G′ i The control signal U′ is generated through the sliding mode control algorithm. t :

[0105] U′ t =sat(G′ i ·E t +β(E t ·E′ t )),

[0106] Wherein, β(E t ·E′ t ) is a nonlinear correction term, which is a function of the error and the rate of change of the error, and provides sufficient correction when the error and the rate of change are large.

[0107] EKF utilizes multi-sensor data to calculate the current state estimation vector, improving state estimation accuracy through residual correction and overcoming the low efficiency and insufficient accuracy of multi-source data fusion in existing technologies. Secondly, DKF predicts the state at the next moment and dynamically adjusts the prediction results based on control input, enhancing the reliability of nonlinear system state prediction and compensating for the limited prediction capabilities of existing technologies. The joint filtering mechanism, through weighted averaging of the current and predicted states, generates a global state estimation vector, further improving the robustness and consistency of state estimation and addressing the shortcomings of insufficient data integration in traditional methods. Furthermore, inputting the global state into GMM for anomaly probability prediction can more accurately capture complex fault modes, significantly improving accuracy compared to the high misjudgment rate of single-model methods in background technologies. This invention improves the sensitivity and accuracy of anomaly detection; it generates target states based on prediction results and optimizes KFOSMC parameters, and adaptively adjusts sliding mode control gain through gradient descent, overcoming the shortcomings of background technologies where control parameters rely on human experience and lack dynamic optimization, making the control signal more adaptable to real-time changing environments; finally, in anomaly mitigation mode, a weighted backoff mechanism is used to generate corrected target states and optimize control signals, enhancing the robot's response capability and stability under abnormal conditions. Compared to the low efficiency of anomaly handling in background technologies, this invention provides a more efficient adaptive control strategy. Overall, this invention achieves significant improvements in state estimation accuracy, anomaly detection capability, and control signal optimization, providing a more intelligent and reliable control solution for data center operation and maintenance robots.

[0108] Furthermore, the robot control signals are mapped to PWM signals for instruction execution. A three-stage PWM control method is used, mapping the control signal to a duty cycle. This duty cycle determines the switching time of the PWM signal, controlling the movement of the robot actuator. A higher duty cycle means a stronger control signal output, and vice versa. The duty cycle calculation formula is as follows:

[0109]

[0110] Among them, T on To control the on-time of the signal within one cycle, T total The time of one complete cycle of the entire PWM signal is determined by the hardware configuration, U max The maximum control signal output value represents the maximum amplitude controlled by the actuator and is set by the hardware system. Duty Cycle is the duty cycle, usually expressed as a percentage.

[0111] The PWM signal is generated based on the calculated duty cycle and the set PWM frequency (set according to the actuator's requirements) and the timer and counter modules.

[0112] The PWM signal is transmitted to the actuator, which precisely adjusts the action according to the PWM signal and regulates the robot's motion state through digital control; for example, if the signal requires the robot to turn, the actuator will adjust the joint angle according to the corresponding duty cycle.

[0113] By calculating the duty cycle, the control signal is precisely converted into switching time, realizing a direct correlation between signal strength and actuator motion. A high duty cycle enhances output force, while a low duty cycle weakens output, ensuring the control signal matches actual needs and improving execution accuracy. Secondly, PWM signals are generated based on timer and counter modules, and the frequency setting supported by hardware makes the signal generation process stable and efficient, overcoming the delay or inconsistency problems that may occur in signal transmission in traditional methods. In addition, the PWM signal is transmitted to the actuator and the motion state is adjusted through digital control, realizing precise adjustment of joint angles and other movements. Compared with the low trajectory adjustment efficiency caused by the lack of adaptive optimization in existing technologies, this invention significantly improves the response speed and control robustness of robot movements, thereby enhancing the reliability and adaptability of dynamic task execution in data center operations and maintenance.

[0114] S3. Monitor the status of the instruction execution process, correct the control signal through status feedback, optimize the robot's motion trajectory and dynamic obstacle avoidance based on the corrected control signal, and perform adaptive optimization.

[0115] Specifically, the robot monitors the state of the command execution process and corrects the control signals through state feedback. Communication is established with the robot controller via Ethernet TCP to obtain real-time machine state data (such as joint position, velocity, and torque). Through ROS nodes, controller data is monitored and relevant information is stored in real time during robot command execution. Based on the robot's execution data, state information for each motion command (such as PTP motion and linear motion) is extracted using data tagging and indexing methods. Using the torque signal obtained from the controller, the robot's health index (HIs) is calculated. The health assessment formula is as follows:

[0116]

[0117] Among them, H γ (t) represents the load vibration signal of the robot at time t, γ is the signal identifier, and H const The load is constant, obtained by analyzing the robot's torque data under normal conditions. H represents the amplitude of the load torque oscillation, which is obtained by performing spectral analysis on the torque signal. The characteristic frequency, f, is obtained from the acquired torque signal using the Fourier transform method. c The specific frequencies caused by robot component failures are obtained through vibration analysis fault diagnosis methods.

[0118] The load vibration signal is input into a Hidden Markov Model (HMM) for anomaly probability prediction. If an anomaly is predicted, the robot will enter an anomaly mitigation mode and generate a corrected control signal U′. t This guides the robot to perform corrective actions.

[0119] Communication with the robot controller is established via Ethernet TCP, and data such as joint position, velocity, and torque are acquired in real time through ROS nodes, achieving efficient status monitoring and overcoming the shortcomings of traditional methods in terms of lagging or incomplete data acquisition. Secondly, the status information of each motion command is extracted based on data labeling and indexing methods, making status analysis more targeted and improving the level of data processing refinement, thus compensating for the low efficiency of multi-source data fusion in existing technologies. In addition, health indicators are calculated using torque signals, and key features are extracted through spectrum analysis and vibration diagnosis, significantly improving the comprehensiveness and accuracy of health assessment compared to the limitations of single feature processing in background technologies. Finally, the load vibration signal is input into a hidden Markov model for anomaly probability prediction, and a corrected control signal is generated based on the result. Compared to the weak anomaly detection capability and insufficient control optimization in background technologies, this invention enhances the robustness of fault identification and the adaptability of control signals, thereby improving the reliability and safety of data center operation and maintenance robots in dealing with abnormal scenarios.

[0120] Furthermore, the robot's motion trajectory and dynamic obstacle avoidance are optimized based on the corrected control signal, and adaptive optimization is performed. This involves using the A* algorithm to optimize the path from the current position to the target position based on the corrected control signal, combining cylindrical coordinates to plan the global path, and calculating the actual distance g(O) from the current position to the target position.

[0121]

[0122] Where R represents all path segments from the starting point to the current node, and Δr υ For the radial distance change at step υ, r υ Let Δθ be the radial distance at step υ, where υ is a step in the path, υ is the loop variable ranging from 1 to R, and Δθ is the distance at step υ. υ For the angle change at step υ, (r υ ·Δθ υ Let φ be the arc length of the υth step;

[0123] The heuristic cost N(O) is calculated using Euclidean distance, representing the estimated distance from the current node O to the target point:

[0124]

[0125] Where, r goal and z goalThese are the radial distance and height in cylindrical coordinates of the target location, r. o z is the radial distance from the current node O. o The height of the current node O;

[0126] Define the path evaluation function:

[0127] J(O) = g(O) + N(O),

[0128] Where J(O) is the total cost of node O, representing the total path cost from the starting point through the current node O to the target;

[0129] By combining real-time data collected by sensors with the Dynamic Window Algorithm (DWA), the position and velocity of dynamic obstacles are estimated, and obstacle avoidance strategies are calculated in real time to adjust the robot's trajectory. (The real-time position and velocity information of surrounding obstacles is obtained through sensor data. The DWA algorithm calculates a safe movement window, which is the speed range within which the robot can safely move in the current state. This range is determined by the robot's maximum speed, current speed, and acceleration limits. Based on these limits and the predicted future positions of obstacles, DWA evaluates each speed command to determine whether the robot will collide with an obstacle. If a speed command would cause the robot to collide with an obstacle at a future time, the speed command is discarded, and the robot will...) By continuously adjusting speed and trajectory, DWA (Driving-Based Automation) ensures that the robot can flexibly avoid obstacles in complex environments and complete tasks safely and effectively. It optimizes path error and control constraints through Model Predictive Control (MPC) algorithms. (MPC minimizes the error between the current position and the target trajectory and controls constraints during the process, including preventing the robot from entering obstacle areas, and setting maximum and minimum limits for speed and acceleration. Control constraints ensure that the robot does not collide while performing tasks and that its movement does not exceed its physical capabilities. By adjusting the weights of these two components, the robot's path tracking accuracy is dynamically controlled during movement, achieving optimal obstacle avoidance and path tracking.)

[0130] By combining the A algorithm with cylindrical coordinates to plan a global path and optimizing the optimal path from the current position to the target by calculating the actual distance, the efficiency and accuracy of path planning are significantly improved, overcoming the shortcomings of traditional methods in path selection in complex environments. Secondly, using Euclidean distance as a heuristic cost estimation method and defining a path evaluation function further enhances the search efficiency and target orientation of the A algorithm. In addition, by using real-time sensor data combined with the Dynamic Window Algorithm (DWA) to estimate obstacle positions and velocities and adjust the trajectory, the flexibility and safety of dynamic obstacle avoidance are achieved, making up for the lack of adaptability to dynamic environments in existing technologies. Finally, the Model Predictive Control (MPC) algorithm is used to optimize path errors and control constraints, and a balance between path tracking accuracy and obstacle avoidance capability is achieved through dynamic weight adjustment. Compared with the lack of adaptability in control optimization in the background technology, this significantly improves the robot's autonomy and task completion efficiency in complex scenarios in data center operations and maintenance.

[0131] This embodiment also provides a data center operation and maintenance robot control system based on a large model, including:

[0132] The robot data acquisition and kinematics calculation module is used to acquire and preprocess data, and calculate the position of the end effector.

[0133] The data feature generation and linearization module is used to generate feature vectors using the BERT method and to generate linear matrices using the Koopman theory.

[0134] The state estimation and prediction module is used to generate the current state estimation vector using extended Kalman filtering, predict the state using discrete Kalman filtering, and fuse the current state and the predicted state through a joint filtering mechanism.

[0135] The anomaly prediction and control signal generation module is used to input the global state estimation vector into the Gaussian mixture model to predict the anomaly probability and generate control signals.

[0136] The feedback correction and adaptive optimization module is used to monitor and correct the control signal in real time, and to perform adaptive optimization based on the corrected target state and control signal.

[0137] This embodiment also provides a computer device applicable to a data center operation and maintenance robot control method based on a large model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the data center operation and maintenance robot control method based on a large model as proposed in the above embodiment.

[0138] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0139] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the control method and system for a data center operation and maintenance robot based on a large model as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0140] In summary, this invention combines extended Kalman filtering with discrete Kalman filtering to reduce error accumulation caused by sensor errors and dynamic environmental changes, thereby improving the accuracy and reliability of global state estimation. The joint filtering mechanism effectively integrates the current state estimation with the predicted state, further optimizing the state estimation process and ensuring the robot's efficient and stable operation in complex data center environments. Furthermore, the introduction of a Gaussian mixture model for anomaly probability prediction accurately predicts the likelihood of fault occurrence in volatile environments, avoiding the shortcomings of traditional methods that rely on threshold judgments. The combination of feature vectors generated by the BERT method and the linearized model based on Koopman theory significantly enhances the robot's adaptability to dynamically changing environments.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A control method for a data center operation and maintenance robot based on a large model, characterized in that: include, Multiple sensors are deployed on the robot to collect and preprocess data. The position of the end effector is calculated using the forward kinematics formula, and the cylindrical coordinates are calculated based on the position of the end effector. Based on the preprocessed data, feature vectors are generated using the BERT method, and linearized matrices are generated using Koopman theory. An extended Kalman filter is used to generate the current state estimation vector, and a discrete Kalman filter is used to generate the predicted state. The current state and the predicted state are fused by a joint filtering mechanism to obtain the robot's global state estimation vector. The global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. The generation of the robot's target state is determined based on the prediction result. The KFOSMC parameters are optimized based on the target state and the global state estimation vector, and the robot control signal is generated using the optimized parameters. The robot control signals are mapped to PWM signals for instruction execution. The state of the instruction execution process is monitored, and the control signals are corrected through state feedback. The robot's motion trajectory and dynamic obstacle avoidance are optimized based on the corrected control signals, and adaptive optimization is performed.

2. The data center operation and maintenance robot control method based on a large model as described in claim 1, characterized in that: The process involves deploying multiple sensors on the robot to collect and preprocess data, calculating the position of the end effector using forward kinematics formulas, and then calculating cylindrical coordinates based on the end effector's position. This involves deploying multiple types of sensors on the robot to collect data, preprocessing the collected data, and then using forward kinematics formulas to calculate the end effector position coordinates P of the robot joints. n Calculate the position coordinates P of the end effector. n The distance r between the projection points on the x-axis and y-axis planes, and the coordinate P n The cylindrical coordinates Y of the end effector are obtained by integrating the calculation results of the angle θ between the y-axis and the x-axis in the y-plane and the height h in the z-axis. n =[r n ,θ n ,h n ]; The various types of sensors include temperature sensors, high-definition cameras, MEMS microphones, data center APIs, joint encoders, and lidar sensors.

3. The data center operation and maintenance robot control method based on a large model as described in claim 2, characterized in that: The preprocessed data is used to generate feature vectors using the BERT method, and then... The Koopman theory generates a linearized matrix by inputting preprocessed data into the BERT model, through... BERT In the model, the Transformer architecture processes the data to obtain the data feature vector V. BERT Using Koopman theory, the linearized Koopman operator K is calculated. g The dynamic mode decomposition method is used to decompose the feature vectors V at multiple time points. BERT Construct the state observation matrix, and use the least squares method to extract the Koopman operator K from the state observation matrix. g Approximately linearized matrix D t .

4. The data center operation and maintenance robot control method based on a large model as described in claim 3, characterized in that: The process involves using an extended Kalman filter to generate a current state estimation vector, and a discrete Kalman filter to generate a predicted state. A joint filtering mechanism is then used to fuse the current and predicted states to obtain the robot's global state estimation vector. This global state estimation vector is input into a Gaussian mixture model for anomaly probability prediction. Based on the prediction results, the generation of the robot's target state is determined. The KFOSMC parameters are optimized based on the target state and the global state estimation vector. The optimized parameters are then used to generate the robot's control signals. Finally, an extended Kalman filter is used to calculate the current state estimation vector X based on the sensor's observation data. t The current state estimation vector X t With observation data Z t The residuals are calculated by comparison, and the residual values ​​are used to correct the state estimation vector X. t The optimized state estimation vector X″ is obtained. t , the feature vector V BERT With the state estimation vector X″ t The fusion process is performed to obtain the fused state estimation vector K; Discrete Kalman filtering is used to predict the robot's motion state, resulting in the state vector X′ for the next moment. t+1 A joint filtering mechanism is employed, using a weighted average method to fuse K and X′. t+1 The estimation results yield the global state estimation vector A. t ; The global state estimation vector is input into a Gaussian mixture model to calculate the probability density function Q, and the eigenvector V is evaluated using the log-likelihood method. BERT The anomaly probability W is set, and an anomaly probability threshold L is defined. If the anomaly probability W is less than the threshold L, the current state is considered to be without anomalies. A is estimated based on the current global state. t Generate target state M t Calculate the state error e between the current global state and the target state. t The sliding mode control gain matrix G in KFOSMC is optimized using gradient descent. i The state error e is calculated using first-order difference. t error change rate e′ t Based on the optimized KFOSMC parameter sliding mode control gain matrix G i and the rate of change of error e′ t Generate control signal U t ; If the anomaly probability W is greater than or equal to the threshold L, the current state is considered to be abnormal, and the system enters anomaly mitigation mode. A corrected target state is generated through a weighted backoff mechanism, and the error E between the current state and the corrected target state is calculated. t and the rate of change of error E′ t With sliding mode control gain matrix G′ i The control signal U′ is generated through the sliding mode control algorithm. t .

5. The data center operation and maintenance robot control method based on a large model as described in claim 4, characterized in that: The process of mapping robot control signals to PWM signals for instruction execution involves employing a three-stage PWM control method. The control signals are mapped to a duty cycle. Based on the calculated duty cycle and the set PWM frequency, a PWM signal is generated using a timer and counter module. This PWM signal is then transmitted to the actuator, which precisely adjusts its actions according to the PWM signal. The robot's motion state is also regulated through digital control.

6. The data center operation and maintenance robot control method based on a large model as described in claim 5, characterized in that: The process of monitoring the state of command execution and correcting control signals through state feedback involves establishing communication with the robot controller via Ethernet TCP. Through ROS nodes, controller data is monitored and relevant information is stored in real time during robot command execution. Based on the robot's execution data, state information for each motion command is extracted using data tagging and indexing methods. The load vibration signal of the robot at each moment is calculated using the torque signal obtained from the controller. This load vibration signal is then input into a Hidden Markov Model for anomaly probability prediction. If an anomaly is predicted, the robot enters an anomaly mitigation mode, generating a corrected control signal U′. t .

7. The data center operation and maintenance robot control method based on a large model as described in claim 6, characterized in that: The process of optimizing the robot's trajectory and dynamic obstacle avoidance based on the corrected control signal, and performing adaptive optimization, refers to using the A* algorithm to optimize the path from the current position to the target position based on the corrected control signal, combining cylindrical coordinates to plan the global path, calculating the actual distance g(O) from the current position to the target position, using Euclidean distance to calculate the heuristically estimated cost N(O), defining the path evaluation function J(O), estimating the position and velocity of dynamic obstacles using real-time data collected by sensors combined with the Dynamic Window Algorithm (DWA), calculating obstacle avoidance strategies in real time to adjust the robot's trajectory, and optimizing path errors and control constraints through model predictive control algorithms.

8. A control method and system for a data center operation and maintenance robot based on a large model, wherein the control method for a data center operation and maintenance robot based on a large model is described in any one of claims 1 to 7, characterized in that: This includes a robot data acquisition and kinematics calculation module, used for data acquisition and preprocessing, and calculating the position of the end effector; The data feature generation and linearization module is used to generate feature vectors using the BERT method and to generate linear matrices using the Koopman theory. The state estimation and prediction module is used to generate the current state estimation vector using extended Kalman filtering, predict the state using discrete Kalman filtering, and fuse the current state and the predicted state through a joint filtering mechanism. The anomaly prediction and control signal generation module is used to input the global state estimation vector into the Gaussian mixture model to predict the anomaly probability and generate control signals. The feedback correction and adaptive optimization module is used to monitor and correct the control signal in real time, and to perform adaptive optimization based on the corrected target state and control signal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the data center operation and maintenance robot control method based on any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the control method for a data center operation and maintenance robot based on a large model as described in any one of claims 1 to 7.

Images