Industrial robot motor fault early warning method

Abstract

The invention discloses an industrial robot motor fault early warning method, and relates to the technical field of motor testing and state monitoring, and the method comprises the steps: introducing a kurtosis-based real-time statistical feature to adaptively adjust the noise amplitude and the decomposition frequency, enabling noise-added empirical mode decomposition to dynamically reconfigure parameters along with the working condition of a motor, and enabling the noise-added empirical mode decomposition to be more accurate. Stable intrinsic mode function components can still be obtained when the impact load frequently changes; a screening coefficient formed by three types of sub-indexes of correlation, an energy ratio and impact sensitivity is utilized to realize automatic elimination of noise dominant components, and the problems of modal aliasing and residual noise are effectively relieved; on this basis, target components are reconstructed and fractal box dimensions are extracted, so that the features still have strong distinguishing ability in the face of transient abrupt change and multi-scale roughness, and the sensitivity to early weak faults is enhanced; and vibration signal dynamic time warping with current as a time reference is further combined with multi-sensor feature fusion.

Description

Technical Field

[0001] This invention relates to the field of motor testing and condition monitoring technology, and in particular to a method for early warning of motor faults in industrial robots. Background Technology

[0002] Industrial robots are widely used in high-frequency operation scenarios such as spot welding and stamping. Their drive motors often bear transient impact loads when performing welding, stamping and other actions. In order to ensure the long-term stable operation of the equipment, existing fault early warning methods are usually based on online testing and analysis of the stator current signal of the drive motor. By extracting current characteristics, early faults such as winding short circuits and rotor faults can be identified.

[0003] In related technologies, in order to separate fault feature components from motor current signals, improved signal decomposition algorithms are often used, combined with machine learning models for operating status identification, to adapt to conditions such as large load variations and complex operating conditions in industrial environments. However, under the action of high-frequency impact loads, motor current signals exhibit obvious non-stationary characteristics and are prone to mode aliasing, that is, the fault feature components and noise components overlap in the frequency domain or time-frequency domain. Mode aliasing makes it difficult to effectively separate intrinsic mode components during signal decomposition, which in turn causes distortion of feature parameters calculated based on these components, affecting the reliability of fault identification results.

[0004] To mitigate the impact of impact noise, existing solutions have proposed improved strategies such as adaptive noise injection and fractal dimension calculation during signal decomposition to enhance the stability and feature distinguishability of the reconstructed signal. However, since these methods generally rely on preset decomposition parameters and fixed thresholds, they lack dynamic adjustment mechanisms for sudden load changes. When impact loads change frequently, residual noise or feature distortion may still be introduced during signal decomposition and reconstruction, leading to decreased fault warning sensitivity and increased risks of false alarms and missed alarms. This limits the real-time application effectiveness of the above technologies in high-frequency impact operation scenarios of industrial robots. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides an industrial robot motor fault early warning method to solve the problems of existing current signal early warning methods, such as modal aliasing under high-frequency impact loads, difficulty in timely tracking of operating condition changes by fixed parameters, resulting in feature distortion and unstable early warning.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a method for early warning of motor faults in industrial robots, comprising:

[0009] Step S1: Collect the stator current signal of the industrial robot drive motor;

[0010] Step S2: Based on the real-time statistical characteristics of the stator current signal, dynamically adjust the current signal decomposition parameters;

[0011] Step S3: The stator current signal is decomposed using dynamically adjusted signal decomposition parameters to obtain multiple intrinsic mode function components.

[0012] Step S4: Select a target component from the intrinsic mode function components, and calculate at least one fault characteristic parameter based on the target component;

[0013] Step S5: Input the fault feature parameters into a preset machine learning model to obtain a fault warning output corresponding to the operating status of the industrial robot drive motor.

[0014] In a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the real-time statistical features include the kurtosis index of the stator current signal, and the dynamically adjusted current signal decomposition parameters include:

[0015] Based on the range of change of the kurtosis index, at least one of the noise amplitude and the number of signal decompositions is adaptively modified to improve the responsiveness of the intrinsic mode function components to changes in impact load.

[0016] In a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the signal decomposition employs a noisy empirical mode decomposition method to obtain the intrinsic mode function components, and the selection of target components from the intrinsic mode function components includes:

[0017] Calculate the corresponding screening coefficient for each of the intrinsic mode function components;

[0018] Each of the screening coefficients is compared with a preset threshold, and intrinsic mode function components whose screening coefficients satisfy the noise dominance condition are removed. The remaining intrinsic mode function components are taken as the target components.

[0019] In a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the calculation of the fault characteristic parameters includes:

[0020] The target component is reconstructed to obtain a motor fault-sensitive reconstructed signal;

[0021] The fractal box dimension is calculated based on the motor fault-sensitive reconstructed signal, and the fractal box dimension is used as at least one of the fault characteristic parameters.

[0022] In a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the machine learning model is a support vector machine model, and the kernel parameters and penalty factor of the support vector machine model are obtained by training with a particle swarm optimization algorithm.

[0023] As a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the method further includes a multi-sensor data synchronization and fusion step, wherein the multi-sensor data synchronization and fusion step includes:

[0024] Collect vibration signals corresponding to the drive motor of the industrial robot;

[0025] Using the stator current signal as a time reference, dynamic time warping and alignment processing is performed on the vibration signal to obtain a vibration signal sequence aligned with the stator current signal;

[0026] Sub-features are extracted from the stator current signal and the vibration signal sequence, and the sub-features are concatenated to form a joint feature vector. The joint feature vector is then input into the machine learning model as part of the fault feature parameters.

[0027] As a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the method further includes a load state identification and feature smoothing step, wherein the load state identification and feature smoothing step includes:

[0028] The root mean square value of the current is calculated based on the stator current signal, and the root mean square rate of change of the current is obtained.

[0029] When the root mean square rate of change of the current exceeds the preset load change range, it is determined that there is an impact load condition, and the smoothing processing of the fault characteristic parameters is triggered.

[0030] The fault feature parameters are smoothed using an exponentially weighted moving average method, and the smoothed fault feature parameters are then input into the machine learning model.

[0031] As a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the method further includes a model lightweighting and deployment step, which includes:

[0032] A subset of key features is obtained by filtering the fault feature parameters using a feature selection algorithm;

[0033] Based on the aforementioned key feature subset, knowledge distillation is performed on the support vector machine model to distill it into a decision tree classification model.

[0034] The decision tree classification model is deployed on an edge computing device connected to the industrial robot control system.

[0035] As a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the method further includes a model training step, which includes:

[0036] Stator current and vibration signal samples labeled with normal and various fault states were collected under various typical operating conditions.

[0037] The sample signals are subjected to signal decomposition, target component selection and fault feature parameter extraction processes to construct a training sample set with state labels;

[0038] The machine learning model is trained based on the training sample set, and the trained model parameters are stored for use in online fault early warning.

[0039] As a preferred embodiment of the industrial robot motor fault early warning method of the present invention, the fault early warning output includes at least two early warning levels, and the method further includes:

[0040] Based on the classification results and / or fault probabilities output by the machine learning model, prompt-level warning signals and protection-level warning signals are generated respectively.

[0041] The prompt-level warning signal is used to send a prompt to the upper-level monitoring system or operator, and the protection-level warning signal is used to trigger the load reduction control or shutdown protection control of the industrial robot drive motor.

[0042] The beneficial effects of this invention are as follows: This invention introduces real-time statistical features based on kurtosis to adaptively adjust the noise amplitude and decomposition number, enabling the noisy empirical mode decomposition to dynamically reconfigure parameters according to motor operating conditions, and still obtain stable intrinsic mode function components even under frequent changes in impact load; it utilizes a screening coefficient composed of three sub-indices—correlation, energy proportion, and impact sensitivity—to automatically eliminate the dominant noise components, effectively alleviating mode aliasing and residual noise problems; based on this, the target components are reconstructed and the fractal box dimension is extracted, ensuring that the features still have strong discriminative ability when facing transient changes and multi-scale roughness, enhancing sensitivity to early, weak faults; further... By combining dynamic time warping of vibration signals with current as the time reference with multi-sensor feature fusion, the robustness of the algorithm under complex impact conditions and structural resonance is improved. For drastic feature fluctuations caused by sudden load changes, the impact condition is identified by the root mean square rate of change of current, and the features are exponentially weighted and smoothed to reduce false alarms and false negatives. Finally, through feature selection and model distillation, the support vector machine is compressed into a lightweight decision tree that can be deployed on edge computing devices, enabling real-time online early warning near the robot body. This significantly improves the operational reliability and maintenance prediction capabilities of industrial robot drive motors in high-frequency spot welding, stamping, and other scenarios without significantly increasing hardware resources. Furthermore, each processing step of this method is based on quantifiable statistical indicators and signal processing steps, facilitating integration with existing drive systems and monitoring platforms. The engineering implementation path is clear, and the repeatability is good, providing conditions for widespread application in industrial settings. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0044] Figure 1 This is a flowchart illustrating the industrial robot motor fault early warning method in the embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0047] For example, the terms “first” and “second” used in this application are only used to distinguish similar objects and differentiate the first object from another object, rather than to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0048] This application proposes a method for early warning of motor faults in industrial robots, combined with... Figure 1 As shown, the method includes:

[0049] Step S1: Collect the stator current signal of the industrial robot drive motor;

[0050] Step S2: Based on the real-time statistical characteristics of the stator current signal, dynamically adjust the current signal decomposition parameters;

[0051] Step S3: Use dynamically adjusted signal decomposition parameters to decompose the stator current signal to obtain multiple intrinsic mode function components.

[0052] Step S4: Select a target component from the intrinsic mode function components, and calculate at least one fault characteristic parameter based on the target component;

[0053] Step S5: Input the fault feature parameters into the preset machine learning model to obtain the fault warning output corresponding to the operating status of the industrial robot drive motor.

[0054] In this embodiment, the stator current signal can be acquired by the current sampling module inside the industrial robot driver or by current sensors installed on the phase lines of the drive motor. After acquisition, the current of each phase is converted into a discrete sequence and input into subsequent steps. A single calculation is usually based on a sliding time window of fixed length, which can cover several power grid cycles or motor electrical cycles in engineering to balance frequency resolution and real-time performance. Specifically, within each time window, statistical feature calculation, signal decomposition, target component screening, feature extraction, and machine learning inference are performed sequentially to form a frame of fault warning result. The window update step size can be set to one-tenth to one-half of the window length to control the warning refresh cycle to the order of milliseconds to hundreds of milliseconds. For example, the machine learning model can be configured to output discrete state labels and corresponding fault probabilities. The state labels can distinguish different levels such as normal, minor abnormality, and severe fault, while the fault probabilities are used for subsequent threshold comparison and warning classification processing. The model parameters are obtained and stored offline. During online operation, only forward inference is performed without parameter updates to reduce the occupation of edge computing resources. Optionally, when the current sample is severely missing, the sampling value is saturated, or there is an obvious measurement anomaly in the current time window, the result of the time window can be determined to be invalid, and the warning output of the previous valid time window can be adopted or the system can be degraded to the preset safety mode to ensure that the system remains controllable and safe in the event of sensor malfunction.

[0055] In one embodiment, the real-time statistical features include the kurtosis index of the stator current signal, and the dynamic adjustment of the current signal decomposition parameters includes:

[0056] Based on the range of kurtosis index variation, at least one of the noise amplitude and the number of signal decompositions is adaptively modified to improve the responsiveness of intrinsic mode function components to changes in impact load.

[0057] In this embodiment, the kurtosis index in the real-time statistical features can be calculated in batches within a sliding time window. The length of the time window is consistent with the number of sample points participating in the statistics, ensuring that each parameter adjustment is based on the same data range as the actual decomposed signal. Specifically, the range of kurtosis variation can be obtained by pre-collecting a stator current sequence under typical normal operating conditions and calculating its statistical distribution. The mean and standard deviation of this distribution are used as references to divide the normal fluctuation range and the abnormal fluctuation range. Based on this, a rule is constructed that maps the degree of kurtosis deviation to the adjustment amount of noise amplitude and decomposition number. For example, when the kurtosis is close to the theoretical value of Gaussian noise, the default settings of noise amplitude and decomposition number can be maintained; when the kurtosis is significantly greater than the theoretical value and persists for several sliding windows, the noise amplitude or decomposition number can be gradually increased to enhance the separation capability of the impact component. Similarly, the sensitivity coefficient and trade-off coefficient related to kurtosis mapping can be selected as fixed values ​​within a given range based on simulation experiments and field calibration results, making the parameter adjustment under different operating conditions repeatable. Optionally, when a short-term load change occurs but the kurtosis estimation is unstable due to a small number of samples, the maximum step size of a single parameter adjustment can be limited, or the decomposition parameters of the previous time step can be directly maintained to avoid the algorithm over-responding due to statistical fluctuations.

[0058] In one embodiment, signal decomposition employs a noisy empirical mode decomposition method to obtain intrinsic mode function components, and selecting target components from the intrinsic mode function components includes:

[0059] Calculate the corresponding screening coefficients for each intrinsic mode function component;

[0060] Each screening coefficient is compared with a preset threshold, and intrinsic mode function components that meet the noise dominance condition are removed. The remaining intrinsic mode function components are used as target components.

[0061] The screening coefficient calculation process is as follows:

[0062] Step S31, for the first Correlation between the calculation of individual intrinsic mode function components and the original current:

[0063]

[0064] in, For the first The correlation sub-indicator of the components, It is a discrete sequence of stator currents. for The sample mean, For the first A sequence of intrinsic mode function components. for The sample mean, For discrete-time indexing, The number of sample points included in the statistics. For component index;

[0065] Step S32, calculate the energy and proportion of each component:

[0066]

[0067] in, For the first The time-domain energy of the component As a percentage of energy, For the summation index, The total number of components;

[0068] Step S33: First, calculate the envelope, then use kurtosis to measure impact sensitivity, and perform a normalization mapping:

[0069]

[0070]

[0071]

[0072] in, For the first The amplitude envelope of the component, for Hilbert transform, The imaginary unit, For envelope kurtosis, for The sample mean, The normalized shock-sensitive sub-indicator The mapping sensitivity coefficient (can be taken as follows) (dimensional range)

[0073] Step S34: Weighted fusion of the three types of sub-indicators:

[0074]

[0075] in, For the first The screening coefficient of the component. The weights of the three sub-indicators;

[0076] Step S35: Set a threshold based on quantiles and determine noise dominance:

[0077]

[0078] in, To determine the threshold, For quantile operators, For quantile order (suggested) ), For the set of all screening coefficients, The total number of components;

[0079] Step S36: Increase the weight of the impact sensitivity sub-index when the impact condition is significant:

[0080]

[0081]

[0082]

[0083] in, The kurtosis of the original current. The weighting factor (can be taken as follows) ), As weight;

[0084] For all Perform steps S31-S35 (optional S36) to remove The components are selected, and the remaining components are retained as the target component set for subsequent feature calculations.

[0085] Specifically, when implementing the screening coefficients, the total number of intrinsic mode function components can be automatically determined based on the convergence of the noisy empirical mode decomposition and the bandwidth of the current signal, generally not exceeding ten components, to avoid generating too many fragmented components and increasing the computational burden. For example, the correlation sub-index can be calculated by the ratio of the covariance to the variance of the original stator current sequence. In practical engineering, to prevent numerical instability caused by the denominator approaching zero, the correlation sub-index of this component can be directly set to zero or a minimum value when the variance is too small. The energy sub-index can be obtained by the proportion of the sum of squares of each component sample value to the sum of squares of all components. When the sum of the energy of all components is lower than the preset minimum energy threshold, the signal quality of this time window can be judged to be poor, and its weight can be reduced in subsequent feature extraction. Furthermore, the envelope kurtosis corresponding to the impact sensitivity sub-index can be calculated using the analytic signal amplitude obtained by Hilbert transform. Its normalized form compresses extreme large values ​​through exponential mapping, thereby forming a more balanced distinction between impact and non-impact conditions. The mapping sensitivity coefficient can be tuned within a given range based on the identification comparison test under simulated impact conditions and stable conditions. Optionally, when the quantile operator is used to determine the threshold from the entire set of screening coefficients, the quantile order near the median can be selected to balance the ability to remove noise components and retain weak fault components, and can be fine-tuned as needed according to different motor models or process lines. If necessary, when the screening coefficients of all components are below the threshold, resulting in an empty target component set, the component with the largest screening coefficient or energy quantization index can be automatically retained to ensure the continuity of subsequent reconstruction and feature extraction steps.

[0086] In one embodiment, calculating the fault characteristic parameters includes:

[0087] The target component is reconstructed to obtain the motor fault-sensitive reconstructed signal;

[0088] The fractal box dimension is calculated based on the motor fault-sensitive reconstructed signal, and the fractal box dimension is used as at least one of the fault characteristic parameters;

[0089] The process for calculating the fractal box dimension is as follows:

[0090] Step S41, reconstruct using the target component set:

[0091]

[0092] in, This represents the reconstructed motor fault-sensitive discrete signal. Indicates the first Each intrinsic mode function component Represents the target component index set. Indicates component index, Indicates a discrete-time index;

[0093] Step S42: Normalize the amplitude of the reconstructed signal and limit it to a unit interval.

[0094]

[0095] in, Represents the normalized sequence. express The minimum value of the sample. express The maximum value of the sample. This indicates a small positive number that prevents the denominator from being zero.

[0096] Step S43, define on the time-amplitude plane Each scale:

[0097]

[0098] in, Indicates the first A normalized scale for the side length of a square Indicates the initial scale. Indicates the scale ratio (take) ), Indicates scale index. Indicates the total number of scales;

[0099] Step S44, with Simultaneously divide the time axis and amplitude axis, and record the number of squares occupied:

[0100]

[0101]

[0102] in, Indicates sample The time grid index where it is located, Indicates sample The index of the amplitude square, This indicates the floor operator. This indicates the number of sample points used for statistical analysis. The scale is represented as The number of squares occupied by the trajectory. Represents the cardinality of a set;

[0103] Step S45: Perform weighted least squares fitting in the logarithmic domain to obtain the fractal box dimension.

[0104]

[0105]

[0106] in, Represents the logarithm of the scale. Represents the number of squares in a logarithm. Denotes the fractal box dimension. The fitting weights represent the scale points. express The weighted mean, express The weighted mean;

[0107] To improve the robustness of the fit, one can set... in, Indicates the first Weights for each scale This represents the weighted exponent; if you need to automatically select the fitting interval, you can do so in all... Calculate one by one on the subintervals With the coefficient of determination ,Pick The largest interval is used for the final estimate.

[0108] Will As one of the fault feature parameters, it is concatenated with other features to form a joint feature vector, which is then entered into the machine learning model inference process.

[0109] Furthermore, the fractal box dimension can be implemented by constructing coverage statistics on the time amplitude plane by selecting multiple scale points. This is used to characterize the spatial filling degree of the reconstructed signal at different scales. In engineering, a moderate number of scales can be chosen to balance computational cost and fitting accuracy. Specifically, the initial scale can be determined based on the number of sampling points of the normalized signal, ensuring that the squares corresponding to the smallest scale can cover a small number of sampling points rather than a single point. The scale ratio can be a fixed value between zero and one, thereby generating a set of scale sequences that change geometrically. For example, the relationship between the number of occupied squares at each scale and the scale logarithm can be considered approximately linear. To improve the robustness of fitting the linear part, a weighted least squares method can be used, assigning higher weights to scale points with larger square numbers and more stable statistics. The weight index can be selected within an empirical range to suppress the impact of statistical fluctuations at extremely small or extremely large scales on the fitting results. Similarly, when the number of occupied squares at a certain scale is zero or close to zero, this scale point can be skipped in the implementation or replaced with the minimum acceptable value to avoid numerical anomalies where the logarithm cannot be taken. Optionally, when it is necessary to further improve the stability of dimension estimation, the coefficient of determination of the fit can be calculated one by one in all scale sub-intervals, and the sub-interval with higher coefficient of determination can be selected for the final dimension estimation. This reduces the interference of local nonlinearity on the overall estimation and keeps the fractal box dimension comparable in different working conditions.

[0110] Specifically, the above implementation provides a discrete coverage path for one-dimensional time series on the time-amplitude plane. First, the fault-sensitive signal is reconstructed using the target mode, and then amplitude normalization is performed to avoid interference from grid coverage due to dimensional differences. A multi-resolution grid set is constructed through geometric scale sequences, and the spatial filling degree of the trajectory at different scales is characterized by the number of occupied squares. Then, the box dimension corresponding to the slope is obtained by logarithmic domain linear regression. This dimension increases with the degree of impact and irregularity, reflecting the transient changes and multi-scale roughness introduced by mechanical faults. To reduce statistical fluctuations at individual scales, a power-law weight is set according to the number of squares, and the optimal fitting interval can be selected from multiple scale sub-intervals, thereby improving the estimation stability. The obtained fractal box dimension is easy to combine with indicators such as energy and correlation to form complementary feature subsets, which maintains low computational cost and facilitates real-time deployment at the edge.

[0111] In one embodiment, the machine learning model is a support vector machine model, and the kernel parameters and penalty factor of the support vector machine model are obtained by training with a particle swarm optimization algorithm;

[0112] In one embodiment, the method further includes a multi-sensor data synchronization and fusion step, which includes:

[0113] Collect vibration signals corresponding to the drive motor of the industrial robot;

[0114] Using the stator current signal as a time reference, dynamic time warping and alignment processing is performed on the vibration signal to obtain a vibration signal sequence aligned with the stator current signal;

[0115] Sub-features are extracted from the stator current signal and vibration signal sequence, and the sub-features are concatenated to form a joint feature vector. The joint feature vector is then used as part of the fault feature parameters and input into the machine learning model.

[0116] In this embodiment, the vibration signal can be collected by an accelerometer mounted on the base of the industrial robot drive motor or related joint structure. The sampling process can be synchronized with the stator current sampling clock, or it can be aligned with timestamps before dynamic time warping. Specifically, during dynamic time warping, a distance matrix between the vibration signal and the stator current signal can be constructed within a limited time window, constraining the start point, end point, and local stepping method of the path to ensure that the aligned path reflects the time offset of the load impact without causing excessive distortion. For example, the stator current sub-features can include indicators such as energy, screening coefficients, and fractal box dimension obtained in previous steps, while the vibration sub-features can include time-domain statistics or simple frequency band energy indicators. The two types of sub-features can be combined into a one-dimensional vector in a fixed order during splicing to ensure that the feature dimension and arrangement are consistent during the training and online inference stages. Optionally, when the vibration sensor temporarily fails or the vibration signal quality is lower than the preset minimum signal-to-noise ratio requirement, it can degenerate to using only stator current-related features to construct fault feature parameters, while recording the sensor status in the system log. Once the vibration signal recovers, the multi-sensor fusion mode will be automatically reactivated.

[0117] In one embodiment, the method further includes a load state identification and feature smoothing step, the load state identification and feature smoothing step comprising:

[0118] The root mean square value of the current is calculated based on the stator current signal, and the root mean square rate of change of the current is obtained.

[0119] When the root mean square rate of change of current exceeds the preset load change range, it is determined that there is an impact load condition, and the smoothing processing of fault characteristic parameters is triggered.

[0120] The fault feature parameters are smoothed by using an exponentially weighted moving average method, and the smoothed fault feature parameters are then input into the machine learning model.

[0121] Specifically, the root mean square (RMS) value of the current can be calculated within a sliding time window consistent with the aforementioned statistical characteristics. The RMS rate of change can be obtained by comparing the difference between the RMS values ​​of adjacent time windows with the baseline value or the value at the previous moment. When this rate of change exceeds the preset load variation range, it can be considered that a significant load change or impact condition has occurred. For example, the preset load variation range can be obtained through field testing under different robot process cycles, ensuring that the RMS rate of change under normal load fluctuations mostly falls within the range, while the rate of change corresponding to strong impact loads significantly exceeds the range, thus giving the judgment logic high discriminative power. Similarly, the smoothing factor in the exponentially weighted moving average method can be selected based on a trade-off between warning sensitivity and stability. A smaller smoothing factor results in a higher degree of smoothing but a slightly slower response speed, while a larger smoothing factor results in a faster response but is more sensitive to instantaneous fluctuations. It can be tested on a set of collected data and then fixed to a suitable value. Optionally, when the load change rate is continuously detected to exceed the threshold over a long period of time, the influence of the smoothing window can be increased accordingly to prevent frequent load switching from causing large fluctuations in the features; if the features still show abnormally drastic fluctuations after smoothing, the time period can be marked as data quality abnormal, and the necessary protection margin can be added to the early warning logic.

[0122] In one embodiment, the method further includes a model lightweighting and deployment step, which includes:

[0123] A subset of key features is obtained by filtering fault feature parameters using a feature selection algorithm;

[0124] Based on a subset of key features, knowledge distillation is performed on the support vector machine model to distill it into a decision tree classification model.

[0125] The decision tree classification model is deployed on an edge computing device connected to the industrial robot control system to enable real-time fault warning reasoning;

[0126] Furthermore, the decision tree classification model can be scaled up during construction by limiting the maximum tree depth and the number of leaf nodes, ensuring stable operation under the limited storage and computing power of edge computing devices. Specifically, the dimension of the key feature subset output by the feature selection algorithm is typically significantly lower than the dimension of the original features. This allows the decision tree, after knowledge distillation, to perform only a limited number of threshold comparisons in each inference, thus completing a fault warning determination within milliseconds or even less. For example, the edge computing device can be an industrial control computing unit connected to the industrial robot control system, pre-deployed with software programs for performing signal preprocessing, feature calculation, and decision tree inference. The program immediately calls the model to classify the current feature vector after each time window and returns the result to the control system or upper-level monitoring system via a predefined data channel. Optionally, when the edge computing device is under high load or its resource usage is nearing its limit, the warning refresh frequency can be appropriately reduced or the calculation of some redundant features can be decreased to ensure that the core decision tree inference process can still be completed within the control cycle, thereby avoiding adverse effects on robot motion control.

[0127] In one embodiment, the method further includes a model training step, which includes:

[0128] Stator current and vibration signal samples labeled with normal and various fault states were collected under various typical operating conditions.

[0129] Perform signal decomposition, target component selection and fault feature parameter extraction processes consistent with those in claims 1 to 7 on the sample signal to construct a training sample set with state labels;

[0130] A machine learning model is trained based on a training sample set, and the trained model parameters are stored for use in online fault warning.

[0131] In this embodiment, the training sample set can be collected in batches under various typical operating conditions, including normal operating conditions, different types of fault conditions, and combinations of operating conditions under different loads and speeds, to ensure that the samples cover typical working areas. Specifically, the number of fault samples should be roughly balanced with the number of normal samples if possible, or the impact of class imbalance on the training results can be mitigated by reweighting the samples. During training, the particle swarm optimization algorithm is used to search for the kernel parameters and penalty factors of the support vector machine model, so that the model can obtain better classification performance and generalization ability on the cross-validation set. For example, the number of iterations and the swarm size of the particle swarm optimization algorithm can be selected within a reasonable range according to the training time budget and the required accuracy, so that the parameter search converges to a satisfactory solution within an acceptable time. After training, the final model parameters and necessary feature normalization parameters are stored together in a non-volatile storage medium, which is loaded and fixed for use by the edge computing device during online operation. Optionally, when new fault samples are collected during subsequent field operation or when the operating conditions change significantly, the model can be retrained in the offline environment using the expanded training sample set, and the deployment can be updated within the maintenance window to ensure that the early warning performance gradually adapts to long-term changes in equipment status.

[0132] In one embodiment, the fault warning output includes at least two warning levels, and the method further includes:

[0133] Based on the classification results and / or fault probabilities output by the machine learning model, prompt-level warning signals and protection-level warning signals are generated respectively.

[0134] The prompt-level warning signal is used to send a prompt to the upper-level monitoring system or operator, and the protection-level warning signal is used to trigger the load reduction control or shutdown protection control of the industrial robot drive motor.

[0135] For example, the failure probability or internal decision score output by the machine learning model can be compared with two-level warning thresholds. When the score is between the low and high thresholds, a prompt-level warning signal is generated; when the score exceeds the high threshold, a protection-level warning signal is generated; and when the score is below the low threshold, it can be considered a normal state and no warning is triggered. Specifically, the initial setting of the two-level thresholds can be based on the comprehensive analysis results of false alarm rate and false negative rate on the offline validation set, and can be fine-tuned during the on-site trial operation phase based on operator feedback and actual equipment response to form a warning strategy that balances safety and production continuity. Similarly, prompt-level warnings can alert maintenance personnel to pay attention to relevant robots or production line nodes through the upper-level monitoring interface, audible and visual indicators, or information system alarms, while protection-level warnings can trigger pre-configured load reduction strategies or safe shutdown strategies, such as reducing the speed and load of relevant axes or issuing a shutdown command, which is then executed by the control system according to established safety procedures. Optionally, when the model output is invalid, the input features are missing, or the system detects a fault in the edge computing device, it can enter a conservative mode. In this mode, it is assumed by default that there is a potential fault risk, and protection-level or safer control strategies are prioritized to avoid the risk being ignored due to the abnormality of the early warning system.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0137] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for early warning of motor faults in industrial robots, characterized in that, include: Step S1: Collect the stator current signal of the industrial robot drive motor; Step S2: Based on the real-time statistical characteristics of the stator current signal, dynamically adjust the current signal decomposition parameters; Step S3: The stator current signal is decomposed using dynamically adjusted signal decomposition parameters to obtain multiple intrinsic mode function components. Step S4: Select a target component from the intrinsic mode function components, and calculate at least one fault characteristic parameter based on the target component; Step S5: Input the fault feature parameters into a preset machine learning model to obtain a fault warning output corresponding to the operating status of the industrial robot drive motor.

2. The industrial robot motor fault early warning method as described in claim 1, characterized in that, The real-time statistical features include the kurtosis index of the stator current signal, and the dynamically adjusted current signal decomposition parameters include: Based on the range of change of the kurtosis index, at least one of the noise amplitude and the number of signal decompositions is adaptively modified to improve the responsiveness of the intrinsic mode function components to changes in impact load.

3. A method for early warning of motor faults in industrial robots as described in claim 1 or 2, characterized in that, The signal decomposition employs a noisy empirical mode decomposition method to obtain the intrinsic mode function components, and the selection of target components from the intrinsic mode function components includes: Calculate the corresponding screening coefficient for each of the intrinsic mode function components; Each of the screening coefficients is compared with a preset threshold, and intrinsic mode function components whose screening coefficients satisfy the noise dominance condition are removed. The remaining intrinsic mode function components are taken as the target components.

4. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 3, characterized in that, The calculation of the fault characteristic parameters includes: The target component is reconstructed to obtain a motor fault-sensitive reconstructed signal; The fractal box dimension is calculated based on the motor fault-sensitive reconstructed signal, and the fractal box dimension is used as at least one of the fault characteristic parameters.

5. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 4, characterized in that, The machine learning model is a support vector machine model, and the kernel parameters and penalty factor of the support vector machine model are obtained by training with a particle swarm optimization algorithm.

6. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 5, characterized in that, The method further includes a multi-sensor data synchronization and fusion step, which includes: Collect vibration signals corresponding to the drive motor of the industrial robot; Using the stator current signal as a time reference, dynamic time warping and alignment processing is performed on the vibration signal to obtain a vibration signal sequence aligned with the stator current signal; Sub-features are extracted from the stator current signal and the vibration signal sequence, and the sub-features are concatenated to form a joint feature vector. The joint feature vector is then input into the machine learning model as part of the fault feature parameters.

7. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 6, characterized in that, The method further includes a load state identification and feature smoothing step, wherein the load state identification and feature smoothing step includes: The root mean square value of the current is calculated based on the stator current signal, and the root mean square rate of change of the current is obtained. When the root mean square rate of change of the current exceeds the preset load change range, it is determined that there is an impact load condition, and the smoothing processing of the fault characteristic parameters is triggered. The fault feature parameters are smoothed using an exponentially weighted moving average method, and the smoothed fault feature parameters are then input into the machine learning model.

8. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 7, characterized in that, The method further includes model lightweighting and deployment steps, which include: A subset of key features is obtained by filtering the fault feature parameters using a feature selection algorithm; Based on the aforementioned key feature subset, knowledge distillation is performed on the support vector machine model to distill it into a decision tree classification model. The decision tree classification model is deployed on an edge computing device connected to the industrial robot control system.

9. A method for early warning of motor faults in industrial robots as described in any one of claims 1 to 8, characterized in that, The method further includes a model training step, which includes: Stator current and vibration signal samples labeled with normal and various fault states were collected under various typical operating conditions. The sample signals are subjected to signal decomposition, target component selection and fault feature parameter extraction processes to construct a training sample set with state labels; The machine learning model is trained based on the training sample set, and the trained model parameters are stored for use in online fault early warning.

10. A method for early warning of motor faults in an industrial robot as described in any one of claims 1 to 9, characterized in that, The fault warning output includes at least two warning levels, and the method further includes: Based on the classification results and / or fault probabilities output by the machine learning model, prompt-level warning signals and protection-level warning signals are generated respectively. The prompt-level warning signal is used to send a prompt to the upper-level monitoring system or operator, and the protection-level warning signal is used to trigger the load reduction control or shutdown protection control of the industrial robot drive motor.

Images