Industrial robot failure prediction and health management system

Abstract

The application relates to the technical field of industrial equipment state monitoring, and particularly discloses a fault prediction and health management system based on an industrial robot, which collects robot benchmark operation data and real-time operation data, and constructs a benchmark data set containing individual identity labels; multi-domain feature extraction and weighted fusion are performed on the data, a high-dimensional feature space is obtained through phase space reconstruction; the high-dimensional features are decomposed into mutually orthogonal individual attribute subspaces and degradation state subspaces by using an orthogonal subspace learning algorithm, and pure benchmark degradation features are obtained; a health index is constructed based on the benchmark degradation features, and a segmented continuous degradation model is established; real-time features are projected into the degradation state subspace to obtain real-time degradation features, which are input into the degradation model to invert the remaining service life and output graded early warning information; the application effectively suppresses false abnormal alarms by stripping individual difference interference through orthogonal decomposition of the feature space, and improves the cross-device generalization capability and prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, specifically to a fault prediction and health management system based on industrial robots. Background Technology

[0002] Industrial robots, as core equipment in intelligent manufacturing, are widely used in fields such as automotive welding, 3C electronics assembly, and precision machining. Their operational reliability directly determines the continuous operation capability of the production line and the stability of product quality. As service time accumulates, key robot components such as servo motors, reducers, and transmission bearings inevitably experience performance degradation. If this is not detected and addressed in a timely manner, it will lead to unplanned downtime or even equipment damage, resulting in economic losses.

[0003] In the field of industrial robot fault prediction and health management, traditional data-driven models struggle to distinguish between "individual attribute differences" and "performance degradation characteristics" due to inherent characteristic shifts caused by variations in mounting base stiffness, mechanical assembly tolerances, and load fluctuations among different units of the same model. This results in a high degree of overlap between the two types of signals in the feature space of the monitoring data. Existing technologies typically employ transfer learning or data augmentation to attempt to adapt to different units, but fundamentally, they still cannot completely decouple the two types of information mathematically. This leads to frequent false alarms, such as misjudging resonant frequency drift as bearing wear or misidentifying normal current fluctuations as motor demagnetization. Ultimately, this causes on-site maintenance personnel to lose trust in the early warning system and abandon it. Predictive maintenance systems fail in actual industrial deployments due to high false alarm rates. Summary of the Invention

[0004] The purpose of this invention is to provide a fault prediction and health management system for industrial robots to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: Based on an industrial robot fault prediction and health management system, including: The data acquisition module collects baseline operating data of industrial robots in a healthy state and real-time operating data during service, and builds a baseline operating dataset containing individual identification tags for each robot. The feature extraction module performs multi-domain feature extraction on the benchmark running dataset and the real-time running data respectively to obtain the corresponding benchmark high-dimensional feature space and real-time high-dimensional feature space. The subspace learning module uses the orthogonal subspace learning algorithm to supervise the learning of the benchmark high-dimensional feature space, obtains the projection matrix that orthogonally decomposes the high-dimensional features into individual attribute subspaces and degenerate state subspaces, and obtains the benchmark degenerate features based on the projection matrix. The degradation modeling module constructs health indicators characterizing the degree of equipment degradation based on baseline degradation characteristics, and uses time series analysis to establish a degradation model of the health indicators evolving over operating time. The prediction output module uses a projection matrix to project the real-time high-dimensional feature space onto the degradation state subspace to obtain real-time degradation features. These real-time degradation features are then input into the degradation model to predict the remaining service life and output early warning information.

[0006] As a further aspect of the present invention: obtaining the corresponding baseline high-dimensional feature space and real-time high-dimensional feature space specifically includes: Time-domain dimensionless features are extracted from vibration signals in the baseline running dataset and real-time running data to obtain the first feature set; Spectral analysis was performed on the current signals in the benchmark running dataset and real-time running data to extract sideband features and obtain a second feature set; Wavelet packet decomposition is performed on the vibration signal to extract the energy proportion features of each frequency band, thus obtaining the third feature set; The first feature set, the second feature set, and the third feature set are fused to form a baseline high-dimensional feature space and a real-time high-dimensional feature space.

[0007] As a further aspect of the present invention: the construction of the baseline high-dimensional feature space and the real-time high-dimensional feature space specifically includes: The variance of each feature in the first, second, and third feature sets under the health condition is calculated as a stability index, and the slope of change of each feature in the real-time running data sequence is calculated as a trend index. The fusion weight of each feature is determined based on the ratio of the stability index to the trend index; The corresponding features in the first feature set, the second feature set, and the third feature set are weighted and summed according to the fusion weights to obtain the weighted fusion features; The phase space of the weighted fusion features is reconstructed to expand their dimension, resulting in a baseline high-dimensional feature space and a real-time high-dimensional feature space.

[0008] As a further aspect of the present invention: the step of obtaining the baseline degradation features based on the projection matrix specifically includes: Construct a category relationship matrix between each sample point and individual identity label in the benchmark high-dimensional feature space, and simultaneously construct an evolutionary relationship matrix between each sample point and the running timestamp; The class relation matrix and the evolution relation matrix are jointly diagonalized to obtain a set of mutually orthogonal projection directions. The first projection direction maximizes the inter-class dispersion of the class relation matrix, and the second projection direction maximizes the temporal correlation of the evolution relation matrix under orthogonal constraints. Construct a projection matrix based on the projection direction, such that the projection matrix maps the reference high-dimensional feature space to two orthogonal subspaces; The baseline high-dimensional feature space is projected onto the degenerate state subspace to obtain the baseline degenerate features.

[0009] As a further aspect of the present invention: the construction of the projection matrix according to the projection direction specifically includes: Use the first projection direction as column vectors to construct the basis matrix of the individual attribute subspace, and use the second projection direction as column vectors to construct the basis matrix of the degenerate state subspace; The basis matrices of the individual attribute subspace and the degenerate state subspace are subjected to Schmitt orthogonalization to ensure that the column vectors of the two basis matrices are orthogonal to each other. The two basis matrices after orthogonalization are merged along the column direction to obtain the projection matrix, where the number of rows in the projection matrix is ​​equal to the dimension of the reference high-dimensional feature space, and the number of columns is equal to the sum of the dimensions of the two subspaces. Multiplying the baseline high-dimensional feature space by the projection matrix on the left yields two low-dimensional feature representations corresponding to the individual attribute subspace and the degenerate state subspace, respectively.

[0010] As a further aspect of the present invention: the establishment of a degradation model for the evolution of health indicators over time specifically includes: The baseline degradation features are normalized to eliminate the dimensional differences between the feature dimensions, resulting in normalized degradation features. Calculate the rate of change of each normalized degradation feature at consecutive time points, and sum the cumulative rates of change to obtain the initial health index; The initial health indicators are smoothed by a sliding window filter to remove random fluctuations and obtain smoothed health indicators. Trend fitting was performed on the smoothed health indicators to establish a degradation model describing the monotonic changes of health indicators over time.

[0011] As a further aspect of the present invention: the step of performing trend fitting on the smoothed health indicators to establish a degradation model describing the monotonic change of health indicators over time specifically includes: The smoothed health indicators are divided into multiple consecutive time windows in chronological order, and each time window corresponds to a fitting interval. In each fitting interval, a monotonic nonlinear function is constructed to approximate the changing trend of health indicators in the corresponding interval. Based on the monotonic nonlinear functions corresponding to each fitting interval, a piecewise continuous degradation model is obtained by combining them.

[0012] As a further aspect of the present invention: the prediction of remaining service life and the output of early warning information specifically includes: The cumulative difference of real-time degradation features is calculated to obtain the deviation value between the current time and the health baseline. Input the deviation value into the degradation model to retrieve the phase position of the current moment in the degradation model, and calculate the remaining running time based on the difference between the phase position and the phase corresponding to the failure threshold. The warning level is determined based on the relationship between the remaining running time and the preset warning threshold; Based on the warning level, generate warning information including the remaining running time value and maintenance instruction code.

[0013] The beneficial effects of this invention are: (1) This invention decomposes the high-dimensional feature space into mutually orthogonal individual attribute subspaces and degradation state subspaces through an orthogonal subspace learning algorithm, effectively removing individual feature interference caused by differences in mounting base stiffness, mechanical assembly tolerances, and load fluctuations. After this processing, the real-time features input to the degradation model retain only degradation information related to time evolution, avoiding false anomaly alarms such as misjudging resonant frequency drift as bearing wear or misidentifying normal current fluctuations as motor demagnetization, and improving the generalization ability and early warning reliability of the degradation model among different robots of the same model.

[0014] (2) The invention uses piecewise monotonic nonlinear functions to fit health indicators and constructs a smooth and continuous degradation model through continuity constraints, which can adaptively describe the complete degradation path of equipment from a healthy state to the failure threshold. Compared with single linear or exponential fitting, this method can still maintain high accuracy in the stage where the degradation rate changes. By combining the calculation of the cumulative difference of real-time degradation characteristics with phase inversion, accurate remaining service life and graded early warning information can be output, providing a quantitative basis for spare parts preparation, planned downtime and emergency maintenance, effectively avoiding unplanned downtime and over-maintenance, and reducing the operation and maintenance cost of the whole life cycle. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1As shown, this invention is based on an industrial robot fault prediction and health management system, comprising: The data acquisition module collects baseline operating data of industrial robots in a healthy state and real-time operating data during service, and builds a baseline operating dataset containing individual identification tags for each robot. The feature extraction module performs multi-domain feature extraction on the benchmark running dataset and the real-time running data respectively to obtain the corresponding benchmark high-dimensional feature space and real-time high-dimensional feature space. The subspace learning module uses the orthogonal subspace learning algorithm to supervise the learning of the benchmark high-dimensional feature space, obtains the projection matrix that orthogonally decomposes the high-dimensional features into individual attribute subspaces and degenerate state subspaces, and obtains the benchmark degenerate features based on the projection matrix. The degradation modeling module constructs health indicators characterizing the degree of equipment degradation based on baseline degradation characteristics, and uses time series analysis to establish a degradation model of the health indicators evolving over operating time. The prediction output module uses a projection matrix to project the real-time high-dimensional feature space onto the degradation state subspace to obtain real-time degradation features. These real-time degradation features are then input into the degradation model to predict the remaining service life and output early warning information.

[0019] The data acquisition module collects baseline operating data of the industrial robots in a healthy state and real-time operating data during service. It also constructs a baseline operating dataset for each robot, including individual identification tags. First, sensing units are deployed at key locations on each industrial robot. Specifically, acceleration sensors are installed on the housings of the drive motors at each joint and at the output of the reducer to pick up vibration signals; the three-phase current values ​​of the servo motors are read in real time through the current monitoring circuit built into the robot controller; and temperature data is collected using resistance temperature detectors (RTDs) attached to the motor windings and the heat sink of the control cabinet. The outputs of all sensing units are connected to a field data acquisition terminal, which is connected to the communication interface of the robot controller to synchronously acquire the position feedback and running timestamps of each robot axis.

[0020] Then, during the first week after the robots were installed, debugged, and entered stable operation, multi-source data was collected under no-load and typical load conditions, defined as baseline operating data. During the data collection process, the factory serial number and installation station information of each robot were simultaneously recorded as individual identification tags. The baseline operating data of the same robot were then associated and stored with that robot's individual identification tag to construct a baseline operating dataset.

[0021] After the robot is put into service, vibration, current, and temperature data during the robot's production tasks are continuously acquired at a fixed sampling frequency through the aforementioned sensing units and data acquisition terminals. This data is defined as real-time operating data. This real-time operating data also includes corresponding timestamp information and is associated with the robot's identification identifier for subsequent traceability analysis on an individual device basis.

[0022] In the feature extraction module, multi-domain feature extraction is performed on the benchmark dataset and the real-time dataset to obtain the corresponding benchmark high-dimensional feature space and the real-time high-dimensional feature space, specifically including: First, time-domain analysis is performed on the vibration signals in the baseline and real-time operating datasets. For each continuous vibration sampling sequence, its peak index (the ratio of the maximum absolute value to the root mean square value), impulse index (the ratio of the maximum absolute value to the absolute mean value), and margin index (the ratio of the maximum absolute value to the root mean square amplitude) are calculated. These peak, impulse, and margin indices are combined to form the first feature set. This first feature set characterizes the impact characteristics of the vibration signal in the time domain and is highly sensitive to waveform changes caused by early faults.

[0023] Secondly, spectral analysis is performed on the current signals in the baseline and real-time operating datasets. The current signals are transformed from the time domain to the frequency domain to obtain their spectral distribution. On the spectrum, the sideband regions on both sides of the fundamental frequency component are located, and the amplitude of each sideband and its ratio to the fundamental frequency amplitude are extracted. The amplitudes and ratios of these sidebands are combined to form a second feature set. This second feature set is used to characterize the specific frequency modulation phenomenon generated in the current signal when a local defect occurs in the motor or transmission component.

[0024] Next, wavelet packet decomposition is performed on the vibration signals in the baseline and real-time operation datasets. A multi-level wavelet packet decomposition is performed on the vibration signals, resulting in multiple equal-width frequency bands covering low to high frequencies. The energy value of the wavelet packet coefficients within each frequency band is calculated, and then the percentage of energy in each frequency band relative to the total energy is calculated. The energy percentages of each frequency band are combined to form a third feature set. This third feature set is used to characterize the energy distribution of the vibration signal in different frequency bands, reflecting the state changes of mechanical components within different frequency ranges.

[0025] Then, the features in the first, second, and third feature sets are fused. For the benchmark dataset, the variance of each feature across all sampling times is calculated, and this variance is used as the stability index of that feature. For the real-time data, a continuous time series is taken, and linear regression is performed on each feature in the series. The slope of the regression line is used as the trend index of that feature. For each feature, the product of its stability index and trend index is calculated, and this product is used as the initial weight of that feature. The initial weights of all features are normalized so that the sum of the weights of all features is 1, resulting in the fused weight of each feature.

[0026] Next, according to the aforementioned fusion weights, the corresponding features in the first, second, and third feature sets are weighted and summed. Specifically, for each sampling time, all feature values ​​at that time are multiplied by their respective fusion weights, and then the weighted feature values ​​are summed to obtain the weighted fusion feature value at that time, forming a weighted fusion feature sequence.

[0027] Finally, phase space reconstruction is performed on the weighted fused feature sequence. An embedding dimension is set, and for each point in the sequence, that point and several subsequent consecutive points are combined into a new vector. The first element of the vector is the value of that point itself, the second element is the value of the first point after that point, and so on, until the set embedding dimension is reached. Through the above operations, the one-dimensional weighted fused feature sequence is expanded into a multi-dimensional vector space, which constitutes the baseline high-dimensional feature space. The same weighted fusion and phase space reconstruction operations are performed on the real-time running data to obtain the real-time high-dimensional feature space.

[0028] In the subspace learning module, the orthogonal subspace learning algorithm is used to perform supervised learning on the benchmark high-dimensional feature space, obtaining the projection matrix that orthogonally decomposes the high-dimensional features into individual attribute subspaces and degenerate state subspaces. Based on the projection matrix, the benchmark degenerate features are obtained, specifically including: First, a category relationship matrix is ​​constructed between each sample point and its individual identity label in the benchmark high-dimensional feature space. For each sample point in the benchmark high-dimensional feature space, each sample point corresponds to an individual identity label of a specific robot. For each individual identity category, the mean vector of all sample points within that category is calculated and denoted as the intra-class center. Then, the difference vector between each sample point and its intra-class center is calculated, and all difference vectors are arranged in the column direction to form an intra-class scatter matrix. Simultaneously, the global mean vector of all sample points is calculated. For each individual identity category, the difference vector between the intra-class center and the global mean vector is calculated. This difference vector is multiplied by the square root of the number of sample points in that category to obtain a weighted difference vector. All weighted difference vectors of all categories are arranged in the column direction to form an inter-class scatter matrix. The intra-class and inter-class scatter matrices are combined to form the category relationship matrix.

[0029] Secondly, an evolutionary relationship matrix between each sample point and its runtime timestamp in the benchmark high-dimensional feature space is constructed. For all sample points in the benchmark high-dimensional feature space, each sample point corresponds to a runtime timestamp, which represents the accumulated duration since the robot began operation. The sample points are sorted according to the chronological order of their timestamps to form a time series. For two adjacent sample points in the time series, the difference vector between the latter and the former is calculated, and all adjacent difference vectors are arranged in the column direction to form a temporal difference matrix. Simultaneously, the difference vector between each sample point and sample points that are a fixed step away in time is calculated, and these are also arranged in the column direction to form a temporal delay matrix. The temporal difference matrix and the temporal delay matrix are combined to form the evolutionary relationship matrix.

[0030] Next, the category relation matrix and evolution relation matrix are jointly diagonalized to obtain a set of mutually orthogonal projection directions. During the solution process, the first projection direction vector is sought such that the ratio of between-class scatter to within-class scatter, as defined by the category relation matrix, is maximized along this projection direction. Specifically, this is achieved by solving the generalized eigenvalue problem, multiplying the inverse of the within-class scatter matrix by the between-class scatter matrix, and taking the eigenvector corresponding to the largest eigenvalue of the product matrix as the first projection direction. This first projection direction maximizes the differentiation of sample points with different individual identity labels. After obtaining the first projection direction, a second projection direction is sought. This direction must be orthogonal to the first projection direction, and the temporal correlation, as defined by the evolution relation matrix, is maximized along this direction. The temporal correlation is defined as the ratio of the sum of squares of the differences between adjacent sample points after projection to the total variance of the sample points after projection. By introducing orthogonal constraints, the vector with the largest such ratio is searched in the orthogonal complement space of the first projection direction and taken as the second projection direction. This second projection direction, while orthogonal to the individual attribute subspace, maximizes the preservation of information related to temporal evolution.

[0031] During the joint diagonalization solution process, a weighted fusion coefficient needs to be calculated to ensure the balance of the two relation matrices in the optimization objective. The formula for calculating this weighted fusion coefficient is as follows: ; in, This represents the weighted fusion coefficient, used to balance the contributions of the category relation matrix and the evolution relation matrix in joint optimization. The function represents the trace of a matrix, which is the sum of the elements on the diagonal of the matrix. The trace represents the inter-class scatter matrix, which reflects the overall degree of separation between different individual identity categories. The trace represents the intra-class dispersion matrix, which reflects the overall clustering of sample points within the same class. The time delay matrix represents the trace of the overall change in sample points over time. By calculating this weighted fusion coefficient, the weights of category and temporal information can be adaptively adjusted during joint diagonalization, so that the obtained projection direction can both distinguish different individuals and capture degradation trends.

[0032] Then, a projection matrix is ​​constructed based on the two obtained projection directions. The first projection direction vector is used as the first column, and the second projection direction vector as the second column, forming the initial projection matrix. To ensure the orthogonality of the two subspaces, the initial projection matrix is ​​subjected to Schmitt orthogonalization. Specifically, the first projection direction vector is kept unchanged and used as the first basis vector. The projection component of the second projection direction vector onto the first basis vector is calculated, and this projection component is subtracted from the second projection direction vector to obtain the second basis vector orthogonal to the first basis vector. The second basis vector is normalized to a length of 1. The orthogonalized first and second basis vectors are arranged according to their column directions to form the final projection matrix. The number of rows in this projection matrix is ​​equal to the dimension of the baseline high-dimensional feature space, and the number of columns is 2, corresponding to the individual attribute subspace and the degenerate state subspace, respectively.

[0033] Finally, the baseline high-dimensional feature space is projected onto the degenerate state subspace to obtain the baseline degenerate features. Specifically, the second column vector corresponding to the degenerate state subspace in the projection matrix is ​​extracted to form the degenerate state projection vector. Each sample point in the baseline high-dimensional feature space is treated as a row vector, and this row vector is multiplied by the degenerate state projection vector to obtain the one-dimensional projected coordinates of each sample point in the degenerate state subspace. The projected coordinates of all sample points are arranged in their original order to form the baseline degenerate feature sequence. This baseline degenerate feature retains only information related to time evolution while filtering out information related to individual identity differences, providing clean input data for the subsequent construction of health indicators.

[0034] In the degradation modeling module, health indicators characterizing the degree of equipment degradation are constructed based on baseline degradation characteristics, and a degradation model of the evolution of health indicators over operating time is established using time-series analysis. Specifically, this includes: First, the baseline degradation features are normalized. The baseline degradation feature sequence contains multiple feature values, each with different dimensions and numerical ranges. For each feature dimension, the minimum and maximum values ​​of that dimension over the entire time series are calculated. For each feature value in the sequence, the minimum value of its corresponding dimension is subtracted, and then divided by the difference between the maximum and minimum values ​​of that dimension to obtain the normalized value. After this process, all feature values ​​are mapped to a closed interval between 0 and 1, eliminating dimensional differences and yielding the normalized degradation features.

[0035] Secondly, the rate of change of each normalized degradation feature at consecutive time points is calculated, and the rates of change are accumulated and summed to obtain the initial health index. For each feature dimension in the normalized degradation feature sequence, the difference between two adjacent time points is calculated in chronological order. This difference is divided by the time interval between the two consecutive time points to obtain the average rate of change within that period. The rates of change for all time periods are accumulated in chronological order, i.e., starting from the first time point, the rate of change for each time period is added sequentially to obtain the cumulative change trajectory for that feature dimension. The cumulative change trajectories for all feature dimensions are then weighted and averaged. The weights are related to the degradation contribution of each feature dimension determined in the subspace learning module. The weighted average yields a single time series, defined as the initial health index. This initial health index reflects the total amount of degradation accumulated by the device from its healthy state.

[0036] Next, a sliding window smoothing filter is applied to the initial health indicators to eliminate random fluctuations. A fixed-length time window is set, for example, containing 10 consecutive sampling points. For each time point in the initial health indicator sequence, a subsequence is formed by taking that point and the consecutive points preceding it that share a window length. The arithmetic mean of all points in this subsequence is calculated, and this mean is used as the smoothed value for that time point. The sliding window is then repeated point by point in chronological order to obtain a new time series, which is defined as the smoothed health indicator. This smoothed health indicator eliminates random spikes caused by operating condition fluctuations or measurement noise, while preserving the long-term trends associated with equipment degradation.

[0037] Then, the smoothed health indicators are divided into multiple consecutive time windows in chronological order, with each time window corresponding to a fitting interval. The division is based on the rate of change of the health indicators. When a health indicator monotonically increases or decreases over a continuous period of time and the rate of change is relatively stable, that period is designated as an independent fitting interval. The boundary points of the fitting intervals are usually selected at moments when the rate of change of the health indicators shows a significant inflection point.

[0038] Within each fitting interval, a monotonic nonlinear function is constructed to approximate the trend of health indicator changes within that interval. The form of the monotonic nonlinear function is determined based on the characteristics of the health indicator changes within that interval. If the health indicator shows a continuously accelerating upward trend within the interval, an exponential function is chosen for fitting; if it shows a continuously decelerating upward trend, a logarithmic function is chosen; and if it shows an approximately linear upward trend, a linear function is chosen. During the fitting process, all time points within the interval are used as independent variables, and the corresponding smoothed health indicator values ​​are used as dependent variables. The undetermined coefficients in the function are determined by minimizing the sum of squared errors between the fitted function values ​​and the actual values.

[0039] Finally, based on the monotonic nonlinear functions corresponding to each fitting interval, a piecewise continuous degradation model is obtained. At the connection points of two adjacent fitting intervals, continuity constraints are applied, requiring that the function value of the function in the preceding fitting interval at the connection point equals the function value of the function in the following fitting interval at the connection point, and simultaneously requiring that the first derivative of the function in the preceding fitting interval at the connection point equals the first derivative of the function in the following fitting interval at the connection point. These constraints ensure that the entire degradation model is smooth and continuous over time. This piecewise continuous degradation model fully describes the complete monotonic evolution path of the health index, starting from its initial value, undergoing different degradation stages, until it reaches the failure threshold.

[0040] In the prediction output module, the real-time high-dimensional feature space is projected onto the degradation state subspace using a projection matrix to obtain real-time degradation features. These features are then input into the degradation model to predict the remaining useful life and output early warning information. Specifically, this includes: First, the cumulative difference of the real-time degradation features is calculated to obtain the deviation value between the current time and the health benchmark. The real-time degradation features are a one-dimensional time series obtained by mapping the real-time high-dimensional feature space to the degradation state subspace through the projection matrix in the subspace learning module. The health benchmark is taken as the average value of the benchmark degradation features in the subspace learning module during the initial running period. This average value represents the degradation feature benchmark level of the device in a healthy state. For each newly acquired real-time degradation feature value, the difference between the feature value and the health benchmark is calculated, and the absolute value of the difference is taken as the single-point deviation. From the start of device operation, the single-point deviations at each time moment are accumulated and summed in chronological order to obtain the cumulative total deviation. This cumulative total deviation is divided by the total running time from the start time to the current time to obtain the average deviation intensity per unit time, which is defined as the deviation value between the current time and the health benchmark. This deviation value quantifies the cumulative performance degradation magnitude of the device since it has been in a healthy state.

[0041] Secondly, the deviation value is input into the degradation model to retrieve the current phase position within the model. The remaining operating time is then calculated based on the difference between this phase position and the phase corresponding to the failure threshold. The degradation model is a piecewise continuous function established in the degradation modeling module, describing the complete evolution path of health indicators from their initial values ​​to the failure threshold over time. The failure threshold is a pre-set value; when the health indicator reaches this value, it indicates that the equipment can no longer meet normal operating requirements and needs to be shut down for maintenance. The deviation value calculated at the current moment is used as the current value of the health indicator and substituted into the degradation model. Since the degradation model is piecewise monotonic, for each health indicator value, there is a unique corresponding operating time point in the model, which is the current phase position. The operating time point corresponding to the failure threshold is subtracted from the operating time point corresponding to the current phase position to obtain the time difference between the two. This difference represents the remaining operating time, indicating the duration the equipment can continue operating from the current moment until the expected failure time.

[0042] Then, the warning level is determined based on the relationship between the remaining operating time and the preset warning thresholds. The preset warning thresholds comprise three numerical ranges, each corresponding to a different level of urgency. The first threshold is set at a remaining operating time greater than 720 hours, corresponding to a low-level warning, indicating that the equipment is in the early stages of degradation and only requires routine monitoring. The second threshold is set at a remaining operating time between 720 and 240 hours, corresponding to a medium-level warning, indicating that degradation has entered the development stage and planned maintenance needs to be scheduled. The third threshold is set at a remaining operating time less than 240 hours, corresponding to a high-level warning, indicating that the equipment is nearing failure and requires immediate shutdown. The calculated remaining operating time value is compared with the above three threshold ranges to determine its corresponding range, and thus the corresponding warning level.

[0043] Finally, a warning message containing the remaining uptime and a maintenance instruction code is generated according to the warning level. The maintenance instruction code is a predefined digital code associated with the warning level and recommended maintenance actions. When the local warning level is low, the generated maintenance instruction code is number 1, indicating a recommendation to perform data tracking and monitoring. When a medium-level warning is triggered, the generated maintenance instruction code is number 2, indicating a recommendation to prepare spare parts and schedule an upcoming maintenance. When a high-level warning is triggered, the generated maintenance instruction code is number 3, indicating a recommendation to immediately shut down the system and replace the component. The remaining uptime and maintenance instruction code are combined into a structured data entry, defined as the warning message. This warning message is displayed to on-site maintenance personnel through a human-machine interface or sent to the upper-level production management system via a communication interface to guide subsequent maintenance decisions.

[0044] The working principle of this invention is as follows: First, by using acceleration sensing units deployed at the robot joints, current monitoring loops built into the controller, and resistance temperature detectors (RTDs) to collect baseline operating data of the robot in a healthy state and real-time operating data during service, a baseline operating dataset containing individual identity tags is constructed for each robot. Then, multi-domain feature extraction is performed on the baseline operating dataset and the real-time operating data. Time-domain dimensionless features are extracted from vibration signals, spectral sideband features are extracted from current signals, and energy proportion features of each frequency band are extracted from wavelet packet decomposition of vibration signals. The fusion weights are determined based on the stability index of each feature in a healthy state and the trend index in the real-time operating data. A baseline high-dimensional feature space and a real-time high-dimensional feature space are obtained through weighted summation and phase space reconstruction. Next, the orthogonal subspace learning algorithm is used to supervise the learning of the baseline high-dimensional feature space by constructing a category relation matrix and an evolution relation matrix and performing joint diagonalization. The process involves obtaining a set of mutually orthogonal projection directions, constructing a projection matrix to orthogonally decompose high-dimensional features into individual attribute subspaces and degradation state subspaces, and projecting the baseline high-dimensional feature space onto the degradation state subspace to obtain pure baseline degradation features. The baseline degradation features are then normalized, and a health index is constructed using cumulative summation of change rates and sliding window smoothing filtering. The smoothed health index is divided into multiple continuous time windows according to the rate of change, and a monotonic nonlinear function is fitted within each window to obtain a piecewise continuous degradation model. Finally, the real-time high-dimensional feature space is projected onto the degradation state subspace using the projection matrix to obtain real-time degradation features. The current deviation value is calculated using cumulative difference, input into the degradation model to invert the current phase position, and compared with the failure threshold to obtain the remaining running time. The warning level is then determined based on the relationship between the remaining running time and a preset warning threshold, generating warning information containing the remaining running time value and maintenance instruction code.

[0045] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault prediction and health management system for industrial robots, characterized in that, include: The data acquisition module collects baseline operating data of industrial robots in a healthy state and real-time operating data during service, and builds a baseline operating dataset containing individual identification tags for each robot. The feature extraction module performs multi-domain feature extraction on the benchmark running dataset and the real-time running data respectively to obtain the corresponding benchmark high-dimensional feature space and real-time high-dimensional feature space. The subspace learning module uses the orthogonal subspace learning algorithm to supervise the learning of the benchmark high-dimensional feature space, obtains the projection matrix that orthogonally decomposes the high-dimensional features into individual attribute subspaces and degenerate state subspaces, and obtains the benchmark degenerate features based on the projection matrix. The degradation modeling module constructs health indicators characterizing the degree of equipment degradation based on baseline degradation characteristics, and uses time series analysis to establish a degradation model of the health indicators evolving over operating time. The prediction output module uses a projection matrix to project the real-time high-dimensional feature space onto the degradation state subspace to obtain real-time degradation features. These real-time degradation features are then input into the degradation model to predict the remaining service life and output early warning information.

2. The industrial robot fault prediction and health management system according to claim 1, characterized in that, The process of obtaining the corresponding benchmark high-dimensional feature space and real-time high-dimensional feature space specifically includes: Time-domain dimensionless features are extracted from vibration signals in the baseline running dataset and real-time running data to obtain the first feature set; Spectral analysis was performed on the current signals in the benchmark running dataset and real-time running data to extract sideband features and obtain a second feature set; Wavelet packet decomposition is performed on the vibration signal to extract the energy proportion features of each frequency band, thus obtaining the third feature set; The first feature set, the second feature set, and the third feature set are fused to form a baseline high-dimensional feature space and a real-time high-dimensional feature space.

3. The industrial robot fault prediction and health management system according to claim 2, characterized in that, The aforementioned constituting the baseline high-dimensional feature space and the real-time high-dimensional feature space specifically includes: The variance of each feature in the first, second, and third feature sets under the health condition is calculated as a stability index, and the slope of change of each feature in the real-time running data sequence is calculated as a trend index. The fusion weight of each feature is determined based on the ratio of the stability index to the trend index; The corresponding features in the first feature set, the second feature set, and the third feature set are weighted and summed according to the fusion weights to obtain the weighted fusion features; The phase space of the weighted fusion features is reconstructed to expand their dimension, resulting in a baseline high-dimensional feature space and a real-time high-dimensional feature space.

4. The industrial robot fault prediction and health management system according to claim 1, characterized in that, The process of obtaining the baseline degradation features based on the projection matrix specifically includes: Construct a category relationship matrix between each sample point and individual identity label in the benchmark high-dimensional feature space, and simultaneously construct an evolutionary relationship matrix between each sample point and the running timestamp; The class relation matrix and the evolution relation matrix are jointly diagonalized to obtain a set of mutually orthogonal projection directions. The first projection direction maximizes the inter-class dispersion of the class relation matrix, and the second projection direction maximizes the temporal correlation of the evolution relation matrix under orthogonal constraints. Construct a projection matrix based on the projection direction, such that the projection matrix maps the reference high-dimensional feature space to two orthogonal subspaces; The baseline high-dimensional feature space is projected onto the degenerate state subspace to obtain the baseline degenerate features.

5. The industrial robot fault prediction and health management system according to claim 4, characterized in that, The construction of the projection matrix based on the projection direction specifically includes: Use the first projection direction as column vectors to construct the basis matrix of the individual attribute subspace, and use the second projection direction as column vectors to construct the basis matrix of the degenerate state subspace; The basis matrices of the individual attribute subspace and the degenerate state subspace are subjected to Schmitt orthogonalization to ensure that the column vectors of the two basis matrices are orthogonal to each other. The two basis matrices after orthogonalization are merged along the column direction to obtain the projection matrix, where the number of rows in the projection matrix is ​​equal to the dimension of the reference high-dimensional feature space, and the number of columns is equal to the sum of the dimensions of the two subspaces. Multiplying the baseline high-dimensional feature space by the projection matrix on the left yields two low-dimensional feature representations corresponding to the individual attribute subspace and the degenerate state subspace, respectively.

6. The industrial robot fault prediction and health management system according to claim 1, characterized in that, The establishment of a degradation model for the evolution of health indicators over time specifically includes: The baseline degradation features are normalized to eliminate the dimensional differences between the feature dimensions, resulting in normalized degradation features. Calculate the rate of change of each normalized degradation feature at consecutive time points, and sum the cumulative rates of change to obtain the initial health index; The initial health indicators are smoothed by a sliding window filter to remove random fluctuations and obtain smoothed health indicators. Trend fitting was performed on the smoothed health indicators to establish a degradation model describing the monotonic changes of health indicators over time.

7. The industrial robot fault prediction and health management system according to claim 6, characterized in that, The process of performing trend fitting on the smoothed health indicators to establish a degradation model describing the monotonic changes of health indicators over time specifically includes: The smoothed health indicators are divided into multiple consecutive time windows in chronological order, and each time window corresponds to a fitting interval. In each fitting interval, a monotonic nonlinear function is constructed to approximate the changing trend of health indicators in the corresponding interval. Based on the monotonic nonlinear functions corresponding to each fitting interval, a piecewise continuous degradation model is obtained by combining them.

8. The industrial robot fault prediction and health management system according to claim 1, characterized in that, The prediction of remaining service life and output of early warning information specifically includes: The cumulative difference of real-time degradation features is calculated to obtain the deviation value between the current time and the health baseline. Input the deviation value into the degradation model to retrieve the phase position of the current moment in the degradation model, and calculate the remaining running time based on the difference between the phase position and the phase corresponding to the failure threshold. The warning level is determined based on the relationship between the remaining running time and the preset warning threshold; Based on the warning level, generate warning information including the remaining running time value and maintenance instruction code.

Images