Robot Fault Prediction System Based on Artificial Intelligence Algorithms

Abstract

This invention relates to the field of robot health management and intelligent operation and maintenance technology, specifically a robot fault prediction system based on artificial intelligence algorithms. The system includes: a data acquisition module for collecting multi-dimensional sensor time-series data of the robot and constructing an N-dimensional time-series tensor from the collected data; a manifold mapping module for constructing a real-time updated health phase diagram; a risk assessment module for calculating the spatial distance between the current state point and a preset ideal health manifold, quantifying the physical degradation pressure of the equipment, and generating a probability distribution of the remaining effective working life; a decision generation module for acquiring external spare parts supply chain data and generating dynamic maintenance work orders; and a feedback closed-loop module for executing the dynamic maintenance work orders. This invention effectively solves the problem that traditional fixed threshold schemes are difficult to handle complex working conditions and nonlinear characteristics, significantly improving the sensitivity of fault precursor identification.

Description

Technical Field

[0001] This invention relates to the field of robot health management and intelligent operation and maintenance technology, specifically a robot fault prediction system based on artificial intelligence algorithms. Background Technology

[0002] Industrial robots are the core execution units of modern automated production lines. The accurate prediction of their operating status and maintenance decisions are directly related to production continuity and enterprise operating costs. Robot fault prediction systems integrate data from multiple sources such as current, vibration, and temperature sensors to achieve real-time monitoring of equipment health status and optimized resource scheduling.

[0003] Existing technologies primarily rely on fixed threshold alarms or simple linear regression models, failing to fully leverage the temporal characteristics of high-dimensional heterogeneous data for comprehensive decision-making. They are ill-suited to the complex dynamic nonlinearities and variable workloads of industrial robots. Due to a lack of effective topological representation of the physical degradation process, existing solutions struggle to capture microscopic distortions that cause the system state to deviate from its steady-state trajectory, limiting the accuracy of remaining effective working life predictions and failing to provide reliable probability distribution support for subsequent decisions. Furthermore, current operation and maintenance management models generally operate equipment health monitoring and external spare parts supply chain management in isolation, failing to achieve deep integration of failure risk probability and logistics timeliness costs in the time dimension. This information silo leads to a severe imbalance between maintenance costs and operational risks, failing to effectively address downtime losses due to spare parts delivery delays and hindering the waste of spare parts' remaining value caused by premature maintenance. Simultaneously, the lack of a physical closed loop from algorithm prediction to logistics execution, coupled with information delays caused by human intervention, further limits the system's response speed. Therefore, an artificial intelligence-based solution is urgently needed to address the deadlock between maintenance costs, operational risks, and prediction accuracy in robot operation and maintenance through manifold mapping and supply chain data convolution. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a robot fault prediction system based on artificial intelligence algorithms. Specifically, the technical solution of this invention includes:

[0005] The data acquisition module is used to collect multi-dimensional sensor time-series data of the robot, including current, vibration, temperature and position signals, and to construct the collected multi-dimensional sensor time-series data into an N-dimensional time-series tensor.

[0006] The manifold mapping module is used to perform spatial transformation on the N-dimensional time series tensor, mapping the time series data to an energy manifold surface in a high-dimensional phase space, and constructing a real-time updated health phase diagram.

[0007] The risk assessment module is used to calculate the spatial distance between the current state point and the preset ideal health manifold based on the health phase diagram, quantify the physical degradation pressure of the equipment, and generate the probability distribution of the remaining effective working life.

[0008] The decision generation module is used to acquire external spare parts supply chain data, perform convolution calculation on the probability distribution of the remaining effective working life and the spare parts supply chain data, and generate dynamic maintenance work orders based on the calculation results.

[0009] The feedback closed-loop module is used to execute the dynamic maintenance work order and adjust the robot's operation strategy or spare parts allocation instructions according to the dynamic maintenance work order.

[0010] Optionally, the manifold mapping module includes:

[0011] The dimension reduction projection unit is used to project the N-dimensional temporal tensor onto a three-dimensional topological space in real time using a preset manifold learning algorithm, wherein the manifold learning algorithm includes a variant of the t-SNE algorithm or the UMAP algorithm.

[0012] The feature extraction unit is used to extract the projected manifold feature vector. The normal operation state of the robot is defined as the regular motion of data points at the low potential energy valley of the manifold surface, and the fault precursor state is defined as the data points gaining additional thermal kinetic energy and causing distortion of the local curvature of the manifold surface.

[0013] Optionally, the manifold mapping module also includes:

[0014] The entropy calculation unit is used to introduce topological entropy as the core operator to calculate the degree of disorder and divergence trend of the system state in the manifold space, and to identify abnormal situations in which the system state deviates from the low potential energy manifold by monitoring the rate of change of the topological entropy.

[0015] Optional, the risk assessment module includes:

[0016] The distance calculation unit is used to calculate the Wasserstein distance between the current state point and the ideal healthy manifold. The Wasserstein distance is used as a quantitative indicator to output a continuous probability distribution of the remaining effective working lifetime, rather than a single binary fault label.

[0017] Optionally, the decision generation module includes:

[0018] The data fusion unit is used to access the spare parts supply chain data, which includes inventory location information, logistics timeliness information, and spare parts expedited transfer costs.

[0019] The logic decision unit is used to calculate the product of the failure probability and the downtime loss to obtain the expected risk cost, and to calculate the sum of the current spare parts expedited allocation cost and the remaining value loss of early replacement to obtain the preventive maintenance cost.

[0020] The instruction triggering unit is used to compare the expected risk cost with the preventive maintenance cost. When the expected risk cost is greater than the preventive maintenance cost, it automatically triggers a procurement instruction and a shutdown maintenance request.

[0021] Optionally, the system may also include:

[0022] The reverse evaluation module is used to analyze the change in the friction coefficient of the energy manifold surface. Based on the small changes in the friction coefficient of the manifold surface caused by different batches of spare parts, it generates a sub-health fingerprint of the equipment.

[0023] The supplier evaluation unit is used to reverse-evaluate the quality of incoming materials from upstream suppliers in the supply chain based on the sub-health fingerprint of the equipment, in the absence of substantial damage to the equipment.

[0024] Optionally, the data acquisition module includes:

[0025] An edge gateway unit is used to perform high-frequency sampling to filter out sensor noise and extract manifold feature vectors at a preset sampling frequency, wherein the preset sampling frequency is not less than 10kHz.

[0026] Optionally, the computational hardware for the manifold mapping module includes:

[0027] The tensor processing unit provides the video memory resources required to construct the topology graph and performs manifold mapping calculations on dedicated hardware to compress the dimensions of fault feature extraction to a preset number of topology feature values.

[0028] Optionally, the output of the decision generation module is connected to:

[0029] The physical execution unit is used to respond to the dynamic maintenance work order and directly drive the automated warehouse or logistics vehicle to perform spare parts outbound actions, realizing a physical closed loop from prediction to logistics.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. This system transforms heterogeneous sensor data into a high-dimensional energy manifold through a manifold mapping module, accurately restoring the robot's dynamic characteristics. By utilizing topological entropy and local curvature analysis, it can keenly capture microscopic distortions in the system state that deviate from the steady-state trajectory, effectively solving the problem that traditional fixed threshold schemes are difficult to handle complex working conditions and nonlinear characteristics, and significantly improving the sensitivity of fault precursor identification.

[0032] 2. This system optimizes remaining working life prediction and supports scientific decision-making: The risk assessment module calculates the distance between the current state and the ideal healthy manifold, and outputs a continuous probability distribution of remaining effective working life, rather than a single fault label. This probability density-based quantification method provides a high-precision input for subsequent cost trade-offs, enabling the system to leap from simple state monitoring to risk assessment with uncertainty quantification capabilities.

[0033] 3. This system breaks down information silos and achieves a dynamic balance between maintenance costs and risks: The decision generation module performs convolution calculations with the lifetime probability distribution and external spare parts supply chain data to comprehensively weigh downtime losses, expedited logistics costs, and the remaining value of spare parts. This mechanism resolves the conflict between maintenance costs and operational risks, determines the optimal maintenance time by solving for the minimum total cost, minimizes spare parts waste, and reduces the risk of unplanned downtime caused by spare parts delays.

[0034] 4. This system constructs a fully automated closed-loop response from prediction to logistics; through physical execution units, it directly links automated warehouses and logistics vehicles, realizing a physical closed loop from algorithm prediction to spare parts allocation; this architecture eliminates the information delay and misoperation risk in manual processes, greatly shortens the average repair time, and is particularly suitable for intelligent operation and maintenance scenarios with stringent requirements for response speed, such as unmanned factories. Attached Figure Description

[0035] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0036] Figure 1 This is a system block diagram of the system of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0038] Example 1:

[0039] Please see Figure 1 A robot fault prediction system based on artificial intelligence algorithms includes:

[0040] The data acquisition module is used to collect multi-dimensional sensor time-series data of the robot, including current, vibration, temperature and position signals, and to construct the collected multi-dimensional sensor time-series data into an N-dimensional time-series tensor.

[0041] The manifold mapping module is used to perform spatial transformation on the N-dimensional time series tensor, mapping the time series data to an energy manifold surface in a high-dimensional phase space, and constructing a real-time updated health phase diagram.

[0042] The risk assessment module is used to calculate the spatial distance between the current state point and the preset ideal health manifold based on the health phase diagram, quantify the physical degradation pressure of the equipment, and generate the probability distribution of the remaining effective working life.

[0043] The decision generation module is used to acquire external spare parts supply chain data, perform convolution calculation on the probability distribution of the remaining effective working life and the spare parts supply chain data, and generate dynamic maintenance work orders based on the calculation results.

[0044] The feedback closed-loop module is used to execute the dynamic maintenance work order and adjust the robot's operation strategy or spare parts allocation instructions according to the dynamic maintenance work order.

[0045] This embodiment provides a robot fault prediction system based on artificial intelligence algorithms. This system aims to solve the impossible triad deadlock problem between maintenance costs and operational risks in existing technologies. The first part of the system's core architecture is a data acquisition module. In this embodiment, the data acquisition module is implemented through a multi-dimensional sensor group deployed at key joints of the industrial robot, such as the reducer and servo motor output. It collects real-time time-series data from the robot's multi-dimensional sensors and aligns and synchronizes the heterogeneous data with different physical dimensions on the time axis, constructing an N-dimensional time-series tensor. Specifically, let the number of sensor channels be... The sliding time window length is Sampling time is The constructed tensor is represented as a matrix. , of which Row vector represents the first row. One sensor in The sampled numerical sequence within a time period; the core computing engine of the system is the manifold mapping module, which is used to perform nonlinear spatial transformations on the N-dimensional time series tensor;

[0046] Based on Takens' embedding theorem, this module maps time-series data to an energy manifold surface in a high-dimensional phase space. During the phase space reconstruction process, the system uses mutual information to calculate the minimum point of the autocorrelation function to determine the optimal delay time. The proportion of pseudo-nearest neighbors is calculated in each dimension projection using the pseudo-nearest neighbor method. When the proportion is less than 5%, it is determined as the minimum embedding dimension. This ensures that the topological structure of the high-dimensional manifold surface can accurately reproduce the robot's dynamic characteristics;

[0047] Within this space, a real-time updated health phase diagram is constructed; the risk assessment module quantifies the degree of physical degradation of the system; it calculates the spatial geometric distance between the current system state point in phase space and the preset ideal health manifold, defined as the physical degradation pressure, and outputs a continuously defined probability distribution of remaining effective working life, RUL-PDF, denoted as the probability density function. The decision generation module is key to the integration of business value and technology in this system; this module acquires external spare parts supply chain data and distributes the remaining effective working life probability. The convolution calculation is performed with the time dimension of the spare parts supply chain data. The specific convolution calculation formula is as follows:

[0048]

[0049] in, This is an auxiliary factor for the convolution integral time. The cost function for spare parts arrival time deviation, which includes downtime losses and expedited shipping costs. The specific piecewise mathematical operator is defined as follows:

[0050]

[0051] in, The deviation between the arrival time of spare parts and the predicted failure time; when At that time, the cost consists of a fixed rush fee. and downtime losses that accumulate over time Composition; when At that time, the cost is the holding cost of spare parts in inventory. Through this explicit operator, the decision-making module can perform integral convolution operations to solve for the minimum total cost. When the computer discretizes and performs the convolution integral, the upper limit of integration is set to the preset risk assessment horizon endpoint in the high-dimensional phase space. High-dimensional phase space is used to ensure that the optimal maintenance time can be solved within a finite observation window. System solution Corresponding optimal maintenance time The system generates dynamic maintenance work orders based on these orders; the feedback loop module not only executes the dynamic maintenance work orders, but also adjusts the robot's operating strategy based on the execution results of the work orders.

[0052] Example 2:

[0053] The manifold mapping module includes:

[0054] A dimensionality reduction projection unit is used to project the N-dimensional temporal tensor onto a three-dimensional topological space in real time using a preset manifold learning algorithm. The manifold learning algorithm includes variants of the t-SNE algorithm or the UMAP algorithm; a feature extraction unit...

[0055] The system is used to extract the feature vector of the manifold after projection. The normal operation state of the robot is defined as the regular motion of the data point at the bottom of the low potential energy valley on the manifold surface. The fault precursor state is defined as the equivalent disturbance energy generated by the data point deviating from the steady-state trajectory due to the increase of system entropy. This disturbance energy causes the local curvature of the manifold surface to be distorted in physical characterization.

[0056] This embodiment specifies the algorithm for the manifold mapping module; this module includes a dimension reduction projection unit; in this embodiment, the dimension reduction projection unit utilizes a variant of the t-SNE algorithm to project a high-dimensional N-dimensional time series tensor onto a three-dimensional topological space in real time. This algorithm minimizes the distribution of the high-dimensional phase space. With low-dimensional spatial distribution The objective function is achieved by the Kullback-Leibler divergence between the two sides. ,in, Midpoint of high-dimensional phase space and points The joint probability, The joint probability between low-dimensional mapping points is defined; the feature extraction unit is responsible for extracting the projected manifold feature vector; to quantify the distortion caused by local curvature, this embodiment uses eigenvalue analysis of the local covariance matrix; for any data point on the manifold surface... Select its Construct the covariance matrix from the nearest neighbors. ,calculate The three eigenvalues Define local curvature for:

[0057]

[0058] when When the threshold is exceeded, the manifold surface is determined to be distorted; the preset threshold is obtained by extracting continuous data during the calibration phase of the robot's new machine operation. Local curvature of each sampling period Construct a baseline sequence and calculate its standard deviation in a high-dimensional phase space. High-dimensional phase space and mean high-dimensional phase space High-dimensional phase space, setting the threshold to a high-dimensional phase space. In this embodiment, the calculated value for the high-dimensional phase space is approximately 0.05; this high-dimensional phase space is based on statistics. The dynamic calibration method based on the high-dimensional phase space principle ensures that the threshold can automatically adapt to robots under different loads and noise levels. In this embodiment, the calculated value is approximately 0.05. This method, based on statistics... The principle of dynamic calibration method ensures that the threshold can automatically adapt to the robot under different loads and different noise levels, corresponding to the pre-fault state; at this time, the data point is regarded as having gained additional thermal kinetic energy, causing it to deviate from the low potential energy trough.

[0059] Example 3:

[0060] The manifold mapping module also includes an entropy calculation unit, which is used to introduce topological entropy as a core operator to calculate the disorder and divergence trend of the system state in the manifold space, and to identify abnormal situations in which the system state deviates from the low potential energy manifold by monitoring the rate of change of the topological entropy.

[0061] This embodiment further refines the computational logic of the manifold mapping module; to quantify the rate of system state divergence, an entropy calculation unit is added to the module; this unit introduces topological entropy as the core operator; to dynamically monitor the stability of the system, this embodiment defines a theoretical mapping function for real-time topological entropy. The calculation method is as follows: ;

[0062] in, : Represents the length of time Within this range, the minimum number of distinguishable orbits is determined to avoid conflict with the number of sensor channels. Obfuscation, used here This indicates that its value is obtained statistically by the manifold mapping module through the grid covering method; the specific discrete execution steps of the grid covering method are as follows: in the observation window Within, a series of decreasing resolution scales are selected. Statistically determine the number of grids required to cover the manifold trajectory at the corresponding scale. ; Scattered data sequence using the least squares method Perform linear regression fitting, and the slope of the fitted line is used as the topological entropy. Approximate numerical solutions are obtained, thereby realizing the finite representation of limit operations in computers; The preset resolution scale is used to filter out sensor noise floor; its value is determined according to the formula. Confirmed, among which The original resolution benchmark of the full scale of the hardware sensor is used, and the SNR is the signal-to-noise ratio value returned in real time by the data acquisition module, so as to automatically increase the grid scale in low signal-to-noise ratio environment to cover up random signal glitches. : Observation time window length, in seconds, set by system configuration;

[0063] This unit calculates the degree of disorder and divergence of the system state in the manifold space; in a physical sense, when the robot's mechanical parts wear or the mating clearance increases, its motion trajectory will no longer coincide in phase space, resulting in a decrease in the number of distinguishable orbits. An exponential increase, thus The system exhibits a significant increase in topological entropy. This unit identifies abnormal situations where the system state deviates from the low-potential manifold by monitoring the rate of change of topological entropy. In this embodiment, topological entropy provides a purely mathematical metric independent of specific physical dimensions, such as current amperes or vibration decibels. This allows the same algorithm to be applied to robots of different models and loads without modification, greatly improving the system's generalization ability and avoiding the tedious process of individually calibrating thresholds for each device.

[0064] Example 4:

[0065] The risk assessment module includes a distance calculation unit, which calculates the Wasserstein distance between the current state point and the ideal healthy manifold, and uses the Wasserstein distance as a quantitative indicator to output a continuous probability distribution of remaining effective working lifetime, rather than a single binary fault label.

[0066] This embodiment specifies the measurement criteria for the risk assessment module; the system includes a distance calculation unit; this unit calculates the current state point distribution. With the ideal healthy manifold distribution The Wasserstein distance between them, used here To distinguish the probability distribution from the t-SNE algorithm in Example 2 To enable computation in a computer, this embodiment employs the discretized bulldozer distance EMD algorithm to solve the following linear programming problem:

[0067]

[0068] The constraints are:

[0069]

[0070] in: The distance matrix is ​​Euclidean. From Move to The probability mass flow; Distributions and The weights; the unit will calculate the weights. The value serves as a scalar indicator that directly reflects the physical degradation pressure of equipment. And through a preset mapping function Output a continuous probability distribution of remaining effective job lifetime;

[0071] Specifically, mapping function Using the conditional Weibull distribution model, its probability density function is expressed as:

[0072]

[0073] in, The shape parameters are preset based on historical fault data of the same model of robot; As a scale parameter, it is related to physical decay pressure. Satisfying the negative correlation mapping relationship: ,in This represents the average rated lifespan of the equipment. This is the decay acceleration factor. The above formula represents the abstract manifold distance. Real-time conversion into probability distribution curves over time

[0074] Example 5:

[0075] The decision generation module includes:

[0076] The data fusion unit is used to access the spare parts supply chain data, which includes inventory location information, logistics timeliness information, and spare parts expedited transfer costs.

[0077] The logic decision unit is used to calculate the product of the failure probability and the downtime loss to obtain the expected risk cost, and to calculate the sum of the current spare parts expedited allocation cost and the remaining value loss of early replacement to obtain the preventive maintenance cost.

[0078] The instruction triggering unit is used to compare the expected risk cost with the preventive maintenance cost. When the expected risk cost is greater than the preventive maintenance cost, it automatically triggers a procurement instruction and a shutdown maintenance request.

[0079] This embodiment details the internal logic of the decision generation module; the data fusion unit is used to access spare parts supply chain data in real time; before calculating the expected risk cost, the probability density function of remaining effective life output by the risk assessment module needs to be used. To characterize the likelihood of equipment failure at different points in the future:

[0080]

[0081] The formula defines the time from the current moment. To the future The sum of probabilities of failure occurring within a given time period ensures the rigor of the subsequent risk cost integral in terms of both dimensions and probabilistic logic.

[0082] The logic decision unit performs specific cost trade-off calculations; it calculates the expected risk cost using the following formula. :

[0083]

[0084] in: The preset risk assessment horizon endpoint represents the system's preset maximum future observation duration; For predicted future moments The probability density of failure occurrence; It represents downtime loss per unit of time, expressed in yuan per hour; The estimated mean time to repair (MTBL) is expressed in hours. This value is based on the average of historical maintenance records. This parameter is introduced to ensure that the integral result is measured in monetary units (yuan), thus making it comparable to maintenance costs. Simultaneously, preventative maintenance costs are calculated. :

[0085]

[0086] in It is the real-time shipping cost returned by the logistics API; residual value loss. The calculation formula is defined as follows:

[0087]

[0088] in: This refers to the unit price of spare parts. The rated design life of the spare parts is expressed in hours. The remaining life expectancy is calculated based on a probability distribution; the instruction triggering unit is used for real-time comparison: when When this happens, procurement orders and shutdown maintenance requests will be automatically triggered.

[0089] Example 6:

[0090] The system also includes a reverse evaluation module, used to analyze the change in the friction coefficient of the energy manifold surface, and generate a sub-health fingerprint of the equipment based on the small changes in the friction coefficient of the manifold surface caused by different batches of spare parts;

[0091] The supplier evaluation unit is used to reverse-evaluate the quality of incoming materials from upstream suppliers in the supply chain based on the sub-health fingerprint of the equipment, in the absence of substantial damage to the equipment.

[0092] This embodiment describes the system's reverse evaluation module and supplier evaluation unit; the reverse evaluation module analyzes the change in the equivalent friction coefficient of the energy manifold surface; in this embodiment, the friction coefficient of the manifold surface... Mathematically defined as the tortuosity of a state trajectory in phase space, its calculation formula is as follows:

[0093]

[0094] in For a sequence of trajectory points in the manifold space, The numerator is the number of points within the observation window; the numerator is the approximate geodesic length of the trajectory, and the denominator is the Euclidean distance between the starting and ending points. When spare parts, such as lubricating grease, are of poor quality, the microscopic resistance increases, causing high-frequency jitter in the trajectory, which increases the total path length, thus... Value increases; module based on The statistical distribution of the device generates a sub-health fingerprint; the supplier evaluation unit is used for monitoring. The rate of change, if the new batch of spare parts causes The mean has deviated from the historical benchmark by more than If so, the incoming material quality is determined to be abnormal.

[0095] Example 7:

[0096] The data acquisition module includes:

[0097] An edge gateway unit is used to perform high-frequency sampling to filter out sensor noise and extract manifold feature vectors at a preset sampling frequency, wherein the preset sampling frequency is not less than 10kHz.

[0098] This embodiment defines an edge gateway unit at the hardware level of the data acquisition module; this unit is configured to operate an analog-to-digital converter (ADC) at a frequency of not less than 10kHz; to filter out sensor noise, the FPGA inside the edge gateway performs digital signal processing, specifically using an infinite impulse response (IIR) filter or a finite impulse response (FIR) filter, whose difference equation is defined as:

[0099]

[0100] in This is the original sampling sequence, used here. To distinguish it from the manifold space vector in Example 6 , This is the filtered sequence; coefficients The upper limit of summation is determined based on the Butterworth filter design method. Given the filter order, summation index Used to iterate through the filter coefficients, here it is used Avoid the same as in Examples 2 and 4 Symbol conflicts; the cutoff frequency is set to 2kHz to 4kHz to preserve the bearing fault characteristic frequency band and filter out high-frequency electromagnetic interference, and the manifold feature vector is extracted and uploaded.

[0101] Example 8:

[0102] The computational hardware of the manifold mapping module includes a tensor processing unit, which provides the video memory resources required to construct the topology graph and performs manifold mapping calculations on dedicated hardware to compress the dimensions of fault feature extraction to a preset number of topology feature values.

[0103] This embodiment specifically illustrates the computational hardware architecture of the manifold mapping module. Since manifold learning, such as t-SNE, involves a large number of matrix operations and distance calculations, this embodiment sets up a dedicated tensor processing unit, such as a TPU or GPU. This tensor processing unit provides the VRAM resources required to construct the 3D topology map and performs manifold mapping calculations in parallel on dedicated hardware. It compresses the original high-dimensional data to a preset number, which in this embodiment is three topological feature values. Through dedicated hardware acceleration, this embodiment compresses the manifold learning calculation time, which originally required several seconds or even minutes, to the millisecond level, such as within 200ms, meeting the real-time requirements of industrial robot fault prediction, preventing computational delays from causing missed optimal shutdown windows, and ensuring the feasibility of online monitoring.

[0104] Example 9:

[0105] The output of the decision generation module is connected to the physical execution unit, which responds to the dynamic maintenance work order and directly drives the automated warehouse or logistics vehicle to perform spare parts outbound actions, thereby realizing a physical closed loop from prediction to logistics.

[0106] This embodiment describes the physical output connection of the decision generation module, realizing a physical closed loop from information flow to logistics. The system's output is connected to a physical execution unit, which specifically includes a stacker crane controller for an automated storage and retrieval system (AS / RS) or an AGV (Automated Guided Vehicle) scheduling system. The physical execution unit responds to the dynamic maintenance work order. When the decision module determines that maintenance is required, the system sends an instruction directly to the automated storage and retrieval system without manual confirmation. The physical execution unit drives the stacker crane to retrieve the corresponding spare parts box and schedules the AGV logistics vehicle to transport the spare parts to the maintenance station of the alarm robot. This embodiment realizes a fully automated physical closed loop from algorithm prediction to physical logistics, which not only greatly shortens the mean time to repair (MTTR) from fault detection to spare parts arrival, but also eliminates information delays and the risk of misoperation in manual processes, making it particularly suitable for the operation and maintenance scenarios of unmanned factories.

[0107] 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.

Claims

1. A robot fault prediction system based on artificial intelligence algorithms, characterized in that: include: The data acquisition module is used to collect multi-dimensional sensor time-series data of the robot, including current, vibration, temperature and position signals, and to construct the collected multi-dimensional sensor time-series data into an N-dimensional time-series tensor. The manifold mapping module is used to perform spatial transformation on the N-dimensional time series tensor, mapping the time series data to an energy manifold surface in a high-dimensional phase space, and constructing a real-time updated health phase diagram. The risk assessment module is used to calculate the spatial distance between the current state point and the preset ideal health manifold based on the health phase diagram, quantify the physical degradation pressure of the equipment, and generate the probability distribution of the remaining effective working life. The decision generation module is used to acquire external spare parts supply chain data, perform convolution calculation on the probability distribution of the remaining effective working life and the spare parts supply chain data, and generate dynamic maintenance work orders based on the calculation results. The feedback closed-loop module is used to execute the dynamic maintenance work order and adjust the robot's operation strategy or spare parts allocation instructions according to the dynamic maintenance work order. The manifold mapping module includes: The dimension reduction projection unit is used to project the N-dimensional temporal tensor onto a three-dimensional topological space in real time using a preset manifold learning algorithm, wherein the manifold learning algorithm includes a variant of the t-SNE algorithm or the UMAP algorithm. The feature extraction unit is used to extract the feature vector of the projected manifold. The normal operation state of the robot is defined as the regular motion of data points at the low potential energy valley of the manifold surface, and the fault precursor state is defined as the data points gaining additional thermal kinetic energy and causing distortion of the local curvature of the manifold surface. The manifold mapping module further includes: The entropy calculation unit is used to introduce topological entropy as the core operator to calculate the disorder and divergence trend of the system state in the manifold space, and to identify abnormal situations in which the system state deviates from the low potential energy manifold by monitoring the rate of change of the topological entropy. The risk assessment module includes: The distance calculation unit is used to calculate the Wasserstein distance between the current state point and the ideal healthy manifold, and uses the Wasserstein distance as a quantitative indicator to output a continuous probability distribution of the remaining effective working lifetime, rather than a single binary fault label. The decision generation module includes: The data fusion unit is used to access the spare parts supply chain data, which includes inventory location information, logistics timeliness information, and spare parts expedited transfer costs. The logic decision unit is used to calculate the product of the failure probability and the downtime loss to obtain the expected risk cost, and to calculate the sum of the current spare parts expedited allocation cost and the remaining value loss of early replacement to obtain the preventive maintenance cost. The instruction triggering unit is used to compare the expected risk cost with the preventive maintenance cost. When the expected risk cost is greater than the preventive maintenance cost, it automatically triggers a procurement instruction and a shutdown maintenance request.

2. The robot fault prediction system based on artificial intelligence algorithm according to claim 1, characterized in that: The system also includes: The reverse evaluation module is used to analyze the change in the friction coefficient of the energy manifold surface. Based on the small changes in the friction coefficient of the manifold surface caused by different batches of spare parts, it generates a sub-health fingerprint of the equipment. The supplier evaluation unit is used to reverse-evaluate the quality of incoming materials from upstream suppliers in the supply chain based on the sub-health fingerprint of the equipment, in the absence of substantial damage to the equipment.

3. The robot fault prediction system based on artificial intelligence algorithm according to claim 1, characterized in that: The data acquisition module includes: An edge gateway unit is used to perform high-frequency sampling to filter out sensor noise and extract manifold feature vectors at a preset sampling frequency, wherein the preset sampling frequency is not less than 10kHz.

4. The robot fault prediction system based on artificial intelligence algorithm according to claim 1, characterized in that: The computing hardware of the manifold mapping module includes: The tensor processing unit provides the video memory resources required to construct the topology graph and performs manifold mapping calculations on dedicated hardware to compress the dimensions of fault feature extraction to a preset number of topology feature values.

5. The robot fault prediction system based on artificial intelligence algorithm according to claim 1, characterized in that: The output of the decision generation module is connected to: The physical execution unit is used to respond to the dynamic maintenance work order and directly drive the automated warehouse or logistics vehicle to perform spare parts outbound actions, realizing a physical closed loop from prediction to logistics.

Images