Artificial intelligence-based robot joint hidden fault detection method and system

Abstract

The application discloses a robot joint hidden fault detection method and system based on artificial intelligence, relates to the technical field of robot fault diagnosis, and comprises the following steps: collecting vibration signal data sequences of robot joints; performing frequency domain decomposition on the vibration signal data sequences to obtain vibration components of multiple frequency bands; generating a vibration spectrum based on the vibration components, calculating energy mean values of each frequency band in the vibration spectrum, and determining characteristic frequency bands; counting time points of energy anomalies in the characteristic frequency bands, analyzing probability distribution characteristics of time interval sequences, and obtaining fault characteristic probabilities; and determining that the robot joints have hidden fault risks according to the change trend of the fault characteristic probabilities and outputting fault risk warning information. The application realizes early warning of hidden faults of the robot joints by analyzing energy anomaly distribution rules of characteristic frequency bands of vibration signals, can effectively prevent equipment faults, and improves equipment operation reliability.

Description

Technical Field

[0001] This invention relates to the field of robot fault diagnosis technology, specifically to a method and system for detecting latent joint faults in robots based on artificial intelligence. Background Technology

[0002] With the widespread application of industrial robots in manufacturing, their operational reliability and safety have become increasingly important. As a key component of the robot, the performance of the robot joint directly affects its overall operational status. Robot joint fault detection mainly relies on regular maintenance and manual inspection, which makes it difficult to detect hidden faults in a timely manner, easily leading to equipment damage and production interruptions.

[0003] Traditional methods for detecting robot joint faults mainly include monitoring single parameters such as temperature and current, and simple fault diagnosis based on threshold judgment. These methods are effective in detecting obvious faults, but their ability to identify early and latent faults is poor. Existing detection methods often ignore the rich fault feature information contained in the vibration signals of robot joints, failing to fully utilize vibration data for fault early warning.

[0004] While existing technologies employ vibration signals for fault diagnosis, most are limited to single analysis in the time or frequency domains, lacking comprehensive consideration of multi-dimensional characteristics. These methods typically use fixed thresholds for judgment, failing to adapt to the dynamic changes in robots under different working conditions and easily leading to misjudgments.

[0005] Therefore, there is an urgent need for an intelligent detection method that can accurately identify latent faults in robot joints in order to improve the accuracy and timeliness of fault warnings. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting latent faults in robot joints based on artificial intelligence. This method can detect latent faults in joints in a timely manner, provide early warnings, avoid equipment damage and production accidents, and improve the reliability and safety of robot operation.

[0007] In a first aspect, the robot joint latent fault detection method based on artificial intelligence provided in the embodiments of the present invention is characterized by comprising the following steps:

[0008] Collect vibration signal data sequences from robot joints;

[0009] Frequency domain decomposition is performed on the vibration signal data sequence to obtain vibration components in multiple frequency bands;

[0010] Vibration spectra are generated based on vibration components, and the average energy of each frequency band in the vibration spectra is calculated to determine the characteristic frequency bands.

[0011] By analyzing the probability distribution characteristics of the time interval sequence of energy anomalies within the statistical characteristic frequency band, the fault characteristic probability can be obtained.

[0012] Based on the changing trend of fault characteristic probabilities, the robot joints are identified as having latent fault risks, and fault risk warning information is output.

[0013] Furthermore, a vibration spectrum is generated based on the vibration components, the average energy of each frequency band in the vibration spectrum is calculated, and the characteristic frequency bands are determined, including:

[0014] Time-frequency conversion is performed based on vibration components to generate vibration spectra containing multiple frequency bands;

[0015] A nonlinear enhancement function is constructed based on the time-domain rate of change sequence of the vibration components to enhance the vibration spectrum;

[0016] Calculate the mean energy of each frequency band of the enhanced vibration spectrum, and generate time-domain weighting coefficients based on energy fluctuation characteristics;

[0017] The energy mean is weighted using time-domain weighting coefficients to determine the characteristic frequency band.

[0018] Furthermore, the nonlinear enhancement function constructed based on the time-domain rate of change sequence of the vibration components includes:

[0019] Calculate the local fluctuation intensity of the time-domain rate of change sequence and construct an initial enhancement function;

[0020] A dynamic adjustment factor is set based on the local fluctuation intensity. Substituting the dynamic adjustment factor into the initial enhancement function generates an adaptive enhancement function.

[0021] Calculate the energy distribution differences of vibration spectra in different frequency bands and establish an energy difference assessment index;

[0022] The method of optimizing the enhancement function based on the energy difference assessment index yields the final nonlinear enhancement function.

[0023] Furthermore, methods for setting dynamic adjustment factors include:

[0024] The fluctuation intensity sequence is decomposed into multiple layers and fluctuation features are extracted. The weights are adjusted by combining a time-varying gain function, and a dynamic adjustment factor is generated by nonlinear combination.

[0025] Furthermore, methods for optimizing the enhancement function based on energy difference assessment indices include:

[0026] Extract the energy gradient sequence between frequency bands to construct the optimization objective function, and perform grouped iterative optimization on the coefficients of the enhancement function.

[0027] The optimized coefficients and interval energy characteristics are non-uniformly compensated to generate the final enhancement function.

[0028] Furthermore, by analyzing the probability distribution characteristics of the time interval sequences of energy anomalies within the statistical characteristic frequency band, the fault characteristic probabilities are obtained, including:

[0029] A time-domain adaptive gain function is constructed to enhance the energy sequence of the characteristic frequency band, and a progressive scale subdivision analysis is performed on the interval sequence of energy anomaly time points to obtain the fault characteristic probability.

[0030] Furthermore, the progressive scaling subdivision analysis includes the following steps:

[0031] Calculate the change in adjacent points in the time interval sequence and construct the fluctuation intensity function;

[0032] The initial scale division position of the time interval sequence is determined based on the fluctuation intensity function, and the subdivision criteria are constructed according to the statistical characteristics within the scale to progressively subdivide the time interval sequence.

[0033] Furthermore, based on the changing trends of fault characteristic probabilities, the robot joints are identified as having latent fault risks, and fault risk warning information is output, including:

[0034] Calculate the difference between adjacent time points in the fault feature probability sequence, construct the fluctuation function of the fault feature probability, and identify the fault change cycle;

[0035] Based on the changing patterns within the cycle, a fault development prediction curve is constructed. Based on the relationship between the prediction curve and the dynamic early warning threshold, the potential for hidden faults in the robot joints is determined, and fault risk warning information is output.

[0036] Secondly, the robot joint latent fault detection system based on artificial intelligence provided in this embodiment of the invention is used to implement the method described in any one of the foregoing embodiments, the system comprising:

[0037] The signal acquisition unit is used to acquire vibration signal data sequences of robot joints;

[0038] The frequency domain decomposition unit is used to perform frequency domain decomposition on the vibration signal data sequence to obtain vibration components in multiple frequency bands.

[0039] The spectrum analysis unit is used to generate vibration spectra based on vibration components, calculate the average energy of each frequency band in the vibration spectra, and determine the characteristic frequency bands.

[0040] The feature extraction unit is used to statistically analyze the time points of energy anomalies within the characteristic frequency band, analyze the probability distribution characteristics of their time interval sequences, and obtain the fault feature probability.

[0041] The early warning judgment unit is used to determine the potential hidden fault risks in robot joints based on the changing trends of fault characteristic probabilities and output fault risk warning information.

[0042] Thirdly, one technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0043] Fourthly, one technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0044] This invention achieves precise extraction of vibration features by collecting robot joint vibration signals and performing frequency domain decomposition, effectively identifying abnormal vibration characteristics in different frequency bands. Based on energy distribution analysis of vibration spectra, it accurately locates characteristic frequency bands, improving the accuracy of fault feature extraction. Through statistical analysis of energy anomaly time points within characteristic frequency bands, a probabilistic model is established, enabling early identification and warning of latent faults. The use of probabilistic distribution characteristic analysis improves the adaptability and reliability of fault detection. Intelligent analysis of vibration signals can detect anomalies in the early stages of fault development, providing timely warnings and effectively preventing fault escalation. It is applicable to different types of robot joints and has significant practical value. Attached Figure Description

[0045] Figure 1 A flowchart illustrating an artificial intelligence-based method for detecting latent joint faults in robots, provided in an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram of the time-frequency characteristic analysis of the vibration spectrum in an embodiment of the present invention;

[0047] Figure 3 This is a flowchart of the robot joint vibration signal energy difference assessment and enhancement process in this embodiment. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown, Figure 1 A flowchart of an artificial intelligence-based robot joint latent fault detection method provided in an embodiment of the present invention is shown. The method includes the following steps:

[0050] Collect vibration signal data sequences from robot joints;

[0051] Frequency domain decomposition is performed on the vibration signal data sequence to obtain vibration components in multiple frequency bands;

[0052] Vibration spectra are generated based on vibration components, and the average energy of each frequency band in the vibration spectra is calculated to determine the characteristic frequency bands.

[0053] By analyzing the probability distribution characteristics of the time interval sequence of energy anomalies within the statistical characteristic frequency band, the fault characteristic probability can be obtained.

[0054] Based on the changing trend of fault characteristic probabilities, the robot joints are identified as having latent fault risks, and fault risk warning information is output.

[0055] Vibration spectra are generated based on vibration components. The average energy of each frequency band in the vibration spectra is calculated, and the characteristic frequency bands are determined, including:

[0056] Time-frequency conversion is performed based on vibration components to generate vibration spectra containing multiple frequency bands;

[0057] A nonlinear enhancement function is constructed based on the time-domain rate of change sequence of the vibration components to enhance the vibration spectrum;

[0058] Calculate the mean energy of each frequency band of the enhanced vibration spectrum, and generate time-domain weighting coefficients based on energy fluctuation characteristics;

[0059] The energy mean is weighted using time-domain weighting coefficients to determine the characteristic frequency band.

[0060] First, vibration signals from the robot's joints during operation are acquired, typically using accelerometers with multiple acquisition points at the joints. The acquired vibration signals undergo preprocessing, including noise reduction and filtering, to improve signal quality. The preprocessed vibration signals are then decomposed into multiple vibration components using methods such as empirical mode decomposition (EMD) or wavelet decomposition. Each vibration component is then transformed into the time-frequency domain using a short-time Fourier transform (SFT). During the transformation, a Hanning window is selected, with a window length of 256 sampling points and an overlap rate of 50%. The resulting time-frequency data is combined to form a vibration spectrum encompassing multiple frequency bands, typically ranging from 0-2000 Hz and divided into 20 bands, each with a width of 100 Hz.

[0061] The time-domain rate of change sequence of vibration components is obtained by calculating the difference between adjacent sampling points. For a given vibration component, the amplitude difference between two adjacent sampling points is divided by the sampling time interval to obtain the rate of change value at that point. This operation is repeated for all sampling points to form a complete rate of change sequence. A nonlinear enhancement function is constructed based on the rate of change sequence, and this function is designed in an exponential form. When the absolute value of the rate of change is large, the corresponding enhancement coefficient is also large; when the absolute value of the rate of change is small, the enhancement coefficient is close to 1. In specific implementation, the enhancement coefficient can be set as an exponential function of the absolute value of the rate of change, with the base set to 1.5 and the upper limit set to 5 to avoid over-enhancement. The enhancement function is applied to each time-frequency point of the vibration spectrum to achieve nonlinear enhancement processing of the spectrum.

[0062] The enhanced vibration spectrum is divided into preset frequency bands, and the energy mean of all time-frequency points within each band is calculated. During the calculation, for a frequency band ranging from f1 to f2, the energy values ​​of all time-frequency points within that band are extracted, and their arithmetic mean is calculated as the energy mean of that band. This operation is repeated for all frequency bands to obtain a complete sequence of frequency band energy mean values. Time-domain weighting coefficients are generated based on energy fluctuation characteristics, achieved by calculating the coefficient of variation of the energy mean within a continuous time window. For a given time window, the ratio of the standard deviation to the mean of the energy mean for each frequency band within the window is calculated to obtain the coefficient of variation. A larger coefficient of variation indicates more severe energy fluctuations and more significant fault characteristics within that time window, and the corresponding weighting coefficient is also larger.

[0063] The average energy of each frequency band is weighted based on the generated time-domain weighting coefficients. For each frequency band, the average energy of its value across different time windows is multiplied by the corresponding weighting coefficient, and then summed to obtain the weighted energy value. The weighted energy values ​​are sorted, and the three frequency bands with the highest energy values ​​are selected as feature frequency bands. For example, under normal operating conditions, the feature frequency bands of the third joint of a six-axis industrial robot are mainly distributed in the 200-300Hz, 500-600Hz, and 900-1000Hz ranges; when minor bearing damage occurs in this joint, the feature frequency bands show a significant shift, concentrating in the 300-400Hz, 600-700Hz, and 1100-1200Hz ranges. By monitoring the changes in feature frequency bands, a fault feature database is established, and a fault identification model is trained using machine learning algorithms such as support vector machines and random forests. The training samples include 200 sets of normal state samples and 100 sets of samples for each type of fault state. The model training uses cross-validation, dividing the dataset into training and test sets in an 8:2 ratio.

[0064] This method for detecting latent joint faults in robots effectively improves the detection capability of weak fault signals by employing vibration signal time-frequency conversion and nonlinear enhancement techniques, overcoming the limitations of traditional methods in early fault identification. The time-domain weighting mechanism fully utilizes the time-varying characteristics of vibration signals, improving the accuracy of characteristic frequency band identification. This method does not rely on high-precision sensors; it can achieve good detection results using ordinary accelerometers, reducing implementation costs. It also has strong generalization ability, applicable to the detection of various types of robot joint faults, including bearing damage, gear wear, and poor lubrication.

[0065] like Figure 2 As shown, Figure 2 This diagram illustrates the time-frequency characteristics of vibration spectra, showcasing a performance comparison of three different methods in the time-frequency domain. The proposed technique demonstrates significantly superior time-frequency resolution compared to traditional short-time Fourier transform and existing wavelet transform techniques within the 0-2000Hz frequency range. In the 200-300Hz band, the proposed technique achieves an energy value of 0.92, approximately 20-35% higher than the 0.68 of short-time Fourier transform and the 0.75 of wavelet transform. Particularly in the 300-500Hz band, this technique clearly distinguishes adjacent frequency components that are easily confused in traditional methods, improving energy resolution by approximately 40%. The nonlinear enhanced EMD technique maintains an energy value above 0.75 in the high-frequency region (900-1200Hz), while traditional methods attenuate to below 0.45 in this region, demonstrating the advantages of this proposed method in broadband analysis.

[0066] The nonlinear enhancement function constructed based on the time-domain rate of change sequence of vibration components includes:

[0067] Calculate the local fluctuation intensity of the time-domain rate of change sequence and construct an initial enhancement function;

[0068] A dynamic adjustment factor is set based on the local fluctuation intensity. Substituting the dynamic adjustment factor into the initial enhancement function generates an adaptive enhancement function.

[0069] Calculate the energy distribution differences of vibration spectra in different frequency bands and establish an energy difference assessment index;

[0070] The method of optimizing the enhancement function based on the energy difference assessment index yields the final nonlinear enhancement function.

[0071] When calculating the local fluctuation intensity, a sliding window method is used to process the time-domain rate of change sequence. The window length is set to 128 sampling points, and the sliding step size is 64 sampling points. Within each window, the standard deviation of the rate of change sequence is calculated as the local fluctuation intensity index for that window. The larger the local fluctuation intensity, the more severe the signal fluctuation in that area, and the more fault characteristic information it may contain. When constructing the initial enhancement function, an exponential function form is used. The function input is the local fluctuation intensity value, and the output is the corresponding enhancement coefficient. Specifically, the local fluctuation intensity value is multiplied by a preset base coefficient of 5, and then the exponent is taken to the base 2 to obtain the initial enhancement coefficient. To prevent over-enhancement, the upper limit of the enhancement coefficient is set to 8.

[0072] When setting a dynamic adjustment factor based on local fluctuation intensity, the differences in signal characteristics under different vibration states are considered. Statistical analysis is performed on the local fluctuation intensity sequence to calculate the global average and maximum values. The dynamic adjustment factor is defined as the ratio of the current local fluctuation intensity to the global average, and is normalized to a value between 0.5 and 1.5. The dynamic adjustment factor is substituted into the initial enhancement function and multiplied by the initial enhancement coefficient to generate an adaptive enhancement function. This adaptive enhancement function can automatically adjust the enhancement level according to the local characteristics of the signal, providing greater enhancement to areas of strong fluctuation and moderate enhancement to areas of gentle fluctuation.

[0073] To calculate the energy distribution differences in vibration spectra across different frequency bands, the vibration signal is converted to the time-frequency domain using a short-time Fourier transform to generate a vibration spectrum. The frequency range is set to 0-5000Hz, divided into 50 equal-width bands, each 100Hz wide. For each band, its energy value is calculated, which is the sum of the squares of the amplitudes at all time-frequency points within that band. An energy ratio sequence between adjacent frequency bands is constructed, calculating the ratio of the energy in each band to the energy in its immediate preceding and following bands. The energy ratio sequence reflects the degree of difference in energy distribution between different frequency bands. When establishing an energy difference assessment index, the variance of the energy ratio sequence is calculated as a quantitative indicator of the overall energy distribution difference. A greater difference indicates that certain frequency bands have prominent energy, potentially containing more obvious fault characteristics.

[0074] When optimizing the enhancement function based on the energy difference assessment index, a mapping relationship is established between the energy difference assessment index and the enhancement function parameters. When the energy difference assessment index is small, it indicates that the energy distribution across frequency bands is relatively uniform and fault characteristics are not obvious. In this case, the base coefficient of the enhancement function is increased to strengthen weak characteristics. When the energy difference assessment index is large, it indicates that there are already obvious characteristic frequency bands. In this case, the enhancement coefficient is appropriately reduced to avoid over-enhancement leading to false characteristics. Through iterative optimization, the enhancement function parameters are adjusted so that the energy difference assessment index reaches the preset target value, resulting in the final nonlinear enhancement function.

[0075] In this embodiment, an adaptive enhancement function is constructed by calculating the local fluctuation intensity of the time-domain rate of change sequence, effectively enhancing the characteristics of weak fault signals and improving the sensitivity of early fault detection. The optimization mechanism based on energy difference assessment ensures the reliability of the enhancement effect and avoids false features introduced by over-enhancement. It can adapt to different operating conditions and different types of fault characteristics, detecting latent faults such as bearing micro-cracks, minor gear wear, and poor lubrication in advance. This provides maintenance personnel with sufficient fault warning time, preventing equipment damage and production interruptions caused by fault escalation, reducing maintenance costs, extending equipment lifespan, and improving production efficiency and safety.

[0076] Methods for setting dynamic adjustment factors include:

[0077] The fluctuation intensity sequence is decomposed into multiple layers and fluctuation features are extracted. The weights are adjusted by combining a time-varying gain function, and a dynamic adjustment factor is generated by nonlinear combination.

[0078] The wave intensity sequence is subjected to multi-level decomposition using wavelet decomposition with an appropriate wavelet basis, employing a three-level decomposition. The first level decomposes the wave intensity sequence into low-frequency approximate components and high-frequency detail components; the second level further decomposes the low-frequency approximate components; the third level decomposes again, ultimately yielding one low-frequency approximate component and three high-frequency detail components at different scales. This decomposition process effectively separates the characteristics of the wave intensity sequence at different frequency scales, facilitating subsequent analysis. Feature extraction is performed on the components obtained from each level of decomposition, including energy features, statistical features, and morphological features. Energy features calculate the energy value of each component, i.e., the sum of squares of the elements in the component sequence; statistical features include mean, standard deviation, skewness, and kurtosis, reflecting the statistical distribution characteristics of the components; morphological features include peak factor, impulse factor, and margin factor, reflecting the waveform characteristics of the components.

[0079] For decomposition components at different levels, corresponding feature weights are designed. For detail components, which contain high-frequency transient information caused by robot joint faults, higher weights are assigned; for approximate components, which mainly contain low-frequency trend information, lower weights are assigned. Reasonable initial weight values ​​are set for each component. The weights are adjusted using a time-varying gain function, which dynamically adjusts the weight values ​​of each component based on the local characteristics of the signal. The time-varying gain function is designed in an exponential form, with the local signal-to-noise ratio (SNR) of each component as input and the weight adjustment coefficient as output. The local SNR is obtained by calculating the ratio of energy within a local window of the signal to the global background energy, with the window length and sliding step size set to appropriate values.

[0080] The specific form of the time-varying gain function is as follows: For the input local signal-to-noise ratio (SNR) value, it is first normalized to a range between 0 and 1; then, the exponent is taken and multiplied by a preset coefficient to obtain the weight adjustment coefficient. The weight adjustment coefficient increases with the increase of the local SNR, so that regions with high SNR receive greater weight gain. The initial weight is multiplied by the weight adjustment coefficient to obtain the time-varying adjusted weight value. A nonlinear combination method is used to generate a dynamic adjustment factor, which combines each component feature with its corresponding time-varying weight. The nonlinear combination uses a weighted geometric mean to avoid the disadvantage of arithmetic mean being easily affected by extreme values.

[0081] For each component's eigenvalue, it is first multiplied by the corresponding time-varying weight, then the logarithm is taken, summed, and the exponent is taken to obtain the combined result. The combined result is normalized to a preset range to form the final dynamic adjustment factor. The dynamic adjustment factor dynamically adjusts with changes in vibration signal characteristics, adaptively enhancing fault features. The dynamic adjustment factor calculated by this method has an average value close to 1 under normal conditions with a small fluctuation range; under fault conditions, the average value increases significantly, and the fluctuation range widens. After applying the dynamic adjustment factor to vibration spectrum enhancement processing, the energy of the fault feature frequency band is significantly enhanced, and the energy ratio is greatly improved, making previously difficult-to-identify minute fault features clearly discernible. By comparing and analyzing the changes in the dynamic adjustment factor under different working conditions, a fault feature pattern library is established. Combined with deep learning methods such as convolutional neural networks for fault classification and recognition, automatic detection and diagnosis of latent faults in robot joints are achieved.

[0082] This implementation effectively enhances weak fault signals by performing multi-level decomposition and feature extraction on the wave intensity sequence, combined with a time-varying gain function and a nonlinear combination method to generate a dynamic adjustment factor. An adaptive mechanism is introduced at the signal processing level, which dynamically adjusts processing parameters according to signal characteristics, improving the ability to extract fault features. The multi-level decomposition technique enables the method to capture fault information at different frequency scales, and the nonlinear combination method enhances the expressive power of the features.

[0083] like Figure 3 The diagram illustrates the process for evaluating and enhancing the energy difference of robot joint vibration signals in this embodiment.

[0084] In this embodiment, the method for optimizing the enhancement function based on the energy difference assessment index includes:

[0085] Extract the energy gradient sequence between frequency bands to construct the optimization objective function, and perform grouped iterative optimization on the coefficients of the enhancement function.

[0086] The optimized coefficients and interval energy characteristics are non-uniformly compensated to generate the final enhancement function.

[0087] When extracting the energy gradient sequence between frequency bands, the energy difference between adjacent frequency bands is calculated and normalized. For two adjacent frequency bands, the energy value of the latter band is subtracted from the energy value of the former band, and then divided by the average of the two band energy values ​​to obtain the normalized energy gradient value. This operation is repeated for all adjacent frequency bands to form a complete energy gradient sequence. The energy gradient sequence reflects the trend of vibration energy variation in the frequency dimension and is of great significance for robot joint fault diagnosis. Common fault types, such as bearing damage and gear wear, will cause abrupt energy changes in specific frequency bands, resulting in a significant increase in the energy gradient value at the corresponding location.

[0088] When constructing the optimization objective function, a composite function containing three indicators is designed based on the characteristics of the energy gradient sequence. The first indicator is the volatility of the energy gradient, obtained by calculating the variance of the energy gradient sequence, reflecting the degree of unevenness in energy distribution. The second indicator is the significance of characteristic frequency bands, obtained by calculating the ratio of the frequency bands with the highest energy to the average energy, reflecting the prominence of fault characteristics. The third indicator is the background noise suppression degree, obtained by calculating the ratio of the average value of low-energy frequency bands to the overall average energy, reflecting the suppression effect on background noise. The three indicators are weighted and summed to form the final optimization objective function, with each weight set to an appropriate value to reflect the relative importance of each indicator.

[0089] When performing iterative optimization of the enhancement function coefficients, the coefficients to be optimized are divided into several groups, each containing coefficients corresponding to several adjacent frequency bands. An alternating optimization strategy is adopted, fixing the coefficients of other groups each time and optimizing only the coefficients of the current group. The optimization process uses a modified gradient descent method, aiming to maximize the value of the objective function. In each iteration, the gradient of the objective function with respect to the coefficients of the current group is calculated, and the coefficient values ​​are updated along the gradient direction. The update step size is adaptive, dynamically adjusted based on the effect of consecutive iterations. To avoid local optima, a perturbation mechanism is introduced during the optimization process. When the effect is not significant after several consecutive iterations, random perturbations are added to the coefficients to help escape local optima.

[0090] For example, in group optimization, the frequency range of the joint vibration signal of an industrial robot is divided into 20 frequency bands, resulting in 20 corresponding enhancement function coefficients. These coefficients are divided into 5 groups, with 4 coefficients in each group. In the first iteration, the coefficients in groups 2 to 5 are fixed, and the coefficients in group 1 are optimized; in the second iteration, the coefficients in groups 1 and 3 to 5 are fixed, and the coefficients in group 2 are optimized; and so on, completing one round of optimization. After multiple rounds of optimization, the optimization process terminates when the change in the objective function value is less than a preset threshold or when the maximum number of iterations is reached, yielding the optimized enhancement function coefficients.

[0091] When applying non-uniform compensation to the optimized coefficients and interval energy characteristics, the importance differences of different frequency bands are considered. For frequency bands with high energy and significant energy gradients, these bands may contain key fault characteristic information and should be given larger compensation coefficients; for frequency bands with low energy and gentle energy gradients, they may mainly contain background noise and should be given smaller compensation coefficients. The non-uniform compensation is implemented as follows: calculate the weighted combination of the energy value and energy gradient of each frequency band as the compensation reference value; normalize the compensation reference value to a reasonable range, multiply it by the optimized coefficients, and obtain the final enhancement coefficient. Through this non-uniform compensation mechanism, the fault characteristic frequency bands are enhanced and the noise frequency bands are suppressed.

[0092] When generating the final enhancement function, a piecewise continuous function form is adopted, with one enhancement coefficient corresponding to each frequency band. To ensure the smoothness of the enhancement function, interpolation is used for transition processing at the frequency band boundaries. When the final enhancement function is applied to the vibration spectrum, the amplitude of each time-frequency point in the spectrum is multiplied by the enhancement coefficient at the corresponding frequency position to obtain the enhanced vibration spectrum.

[0093] By extracting energy gradient sequences between frequency bands to construct an optimization objective function, iteratively optimizing the enhancement function coefficients in groups, and combining this with non-uniform compensation based on interval energy characteristics, the detection capability of weak fault signals is effectively improved. This method introduces energy gradient analysis and group optimization techniques, significantly enhancing the recognition rate and noise resistance of fault features. The non-uniform compensation mechanism fully considers the differences in importance between different frequency bands, achieving precise enhancement of fault features.

[0094] By analyzing the probability distribution characteristics of the time interval sequences of energy anomalies within the statistical characteristic frequency band, the fault characteristic probabilities are obtained, including:

[0095] A time-domain adaptive gain function is constructed to enhance the energy sequence of the characteristic frequency band, and a progressive scale subdivision analysis is performed on the interval sequence of energy anomaly time points to obtain the fault characteristic probability.

[0096] When constructing the time-domain adaptive gain function, the gain function form is designed based on the local statistical characteristics of the signal. This gain function includes a base gain part and an adaptive adjustment part. The base gain part adopts an exponential function, with the input being the ratio of the characteristic frequency band energy value to the background noise energy, and the output being the base gain coefficient. When the ratio is large, the base gain coefficient is close to 1; when the ratio is small, the base gain coefficient is large, used to boost weak signals. The adaptive adjustment part dynamically adjusts the gain level based on the energy fluctuations within a sliding time window. The sliding window length is set to 1 second, and the coefficient of variation of the energy sequence within the window, i.e., the ratio of the standard deviation to the mean, is calculated as a fluctuation index. The larger the fluctuation index, the stronger the signal instability, which may contain fault information, and the larger the corresponding adjustment coefficient.

[0097] The base gain coefficient is multiplied by the adaptive adjustment coefficient to obtain the final time-domain adaptive gain coefficient. This coefficient is then multiplied by the original characteristic frequency band energy sequence to achieve adaptive enhancement of the energy sequence. During enhancement, the gain effect is more pronounced in regions of sudden energy changes, while the gain effect is relatively smaller in regions of stable energy, thus highlighting areas that may contain fault information. The enhanced characteristic frequency band energy sequence is compared with a preset threshold to determine the time point of energy anomalies. The threshold setting adopts an adaptive method, based on the mean of the energy sequence under normal operating conditions plus a certain number of standard deviations, and can be adjusted according to the joint type and operating conditions.

[0098] The interval sequences of energy anomaly time points are extracted, that is, the time difference sequences between adjacent anomaly time points. These interval sequences contain important periodic information about robot joint faults. For typical rotating component faults, such as damage to the inner and outer rings of bearings and pitting on gear tooth surfaces, specific impact periods are generated in the vibration signals, which are reflected in the interval sequences of energy anomaly time points. A progressive scale subdivision analysis is performed on the interval sequences. First, the overall statistical distribution of the interval sequences is analyzed at a macro scale, including characteristics such as mean, variance, skewness, and kurtosis, reflecting the overall characteristics of the interval sequences.

[0099] At the mesoscale, histogram analysis is used to statistically analyze the frequency of occurrence of different interval lengths, identifying frequently occurring interval values ​​that may correspond to fault characteristic cycles. The interval division in the histogram analysis employs an adaptive method, dynamically determining the appropriate number and width of intervals based on the numerical range of the interval sequence and the sample size. At the microscale, a detailed analysis is performed on the vicinity of frequently occurring interval values. Kernel density estimation is used to obtain the continuous probability distribution of the interval values, and the peak positions and peak widths of the distribution are extracted as precise indicators of fault characteristics.

[0100] The results of progressive scale subdivision analysis are combined with the structural parameters of the robot joints to calculate the probability of fault characteristics. The main period of the interval sequence is matched with the characteristic frequencies of components such as bearings and gears within the joint. These characteristic frequencies are pre-calculated based on the geometric and kinematic parameters of the components. A higher matching degree indicates a greater likelihood of failure in that component. The matching analysis employs a fuzzy matching method, allowing for a certain range of frequency deviations while considering factors such as rotational speed fluctuations and signal acquisition errors during actual operation. Based on the matching results and the characteristic indicators from each scale analysis, the final probability of fault characteristics is calculated through a weighted combination.

[0101] During normal operation, the energy sequence of the characteristic frequency band is stable, and energy anomalies are rare and randomly distributed. When minor wear occurs in the rolling elements of the bearing within the joint, periodic energy anomalies appear in the enhanced energy sequence. The main periodic values ​​of the interval sequence closely match the characteristic frequencies of bearing rolling element failures, resulting in a high calculated probability of the fault characteristics and enabling accurate identification of latent faults.

[0102] By constructing a time-domain adaptive gain function to enhance the energy sequence of characteristic frequency bands and performing progressive scale subdivision analysis on the interval sequence of energy anomaly time points, the detection capability of latent faults in robot joints is effectively improved. The time-domain adaptive gain function dynamically adjusts the gain level according to signal characteristics, enhancing weak fault characteristics, suppressing background noise, and improving the signal-to-noise ratio. The progressive scale subdivision analysis extracts interval sequence features from macroscopic to microscopic levels, comprehensively capturing fault cycle information.

[0103] The steps of the progressive scaling subdivision analysis include:

[0104] Calculate the change in adjacent points in the time interval sequence and construct the fluctuation intensity function;

[0105] The initial scale division position of the time interval sequence is determined based on the fluctuation intensity function, and the subdivision criteria are constructed according to the statistical characteristics within the scale to progressively subdivide the time interval sequence.

[0106] When calculating the change between adjacent points in a time interval sequence, for each point in the sequence, calculate the difference between it and the previous point, i.e., the current interval value minus the previous interval value, to obtain the change sequence. To eliminate the influence of sign, take the absolute value of the change to form the absolute change sequence.

[0107] When constructing the fluctuation intensity function, a sliding window method is used to process the sequence of absolute changes. The window length is determined based on the robot joint characteristics and operating parameters, typically ranging from 10% to 20% of the sequence length. At each window location, the mean of the absolute changes within the window is calculated as the fluctuation intensity value at that location. The fluctuation intensity function reflects the severity of local fluctuations in the time interval sequence; regions with larger fluctuation intensity values ​​indicate more unstable time interval changes, potentially corresponding to fault characteristics or changes in operating conditions. The fluctuation intensity function is normalized to ensure its value range is between 0 and 1, facilitating subsequent analysis.

[0108] When determining the initial scale division position of the time interval sequence based on the fluctuation intensity function, significant peaks and troughs in the fluctuation intensity function are identified. Peak positions correspond to regions of drastic change in the time interval sequence and may be the boundary points between different operating conditions or fault modes; trough positions correspond to relatively stable regions and may represent the persistent state of a specific operating condition or fault mode. Peak and trough identification employs an extreme point detection algorithm, requiring that the difference between the fluctuation intensity value of an extreme point and its neighborhood exceeds a preset threshold. Based on the identified peak and trough positions, the time interval sequence is divided into multiple segments of the initial scale, with relatively consistent time interval characteristics within each segment.

[0109] When constructing subdivision criteria based on the statistical characteristics within the scale, statistical indicators of the time interval sequence within each initial segment are calculated, including mean, standard deviation, coefficient of variation, skewness, and kurtosis. The mean reflects the central tendency of the time intervals and is related to the robot joint's running speed; the standard deviation reflects the dispersion of the time intervals and is related to the system's stability; the coefficient of variation, calculated by dividing the standard deviation by the mean, eliminates the influence of dimensions and facilitates comparison between different segments; skewness reflects the asymmetry of the distribution, and kurtosis reflects the peak of the distribution; both indicators are closely related to fault types. Based on these statistical indicators, a multidimensional criterion vector is constructed to evaluate the uniformity within a segment and whether further subdivision is needed.

[0110] When progressively subdividing the time interval sequence, a bisection method is used to process each initial segment. All possible segmentation points within a segment are calculated. For each segmentation point, the segment is divided into left and right parts, and the statistical characteristics of each part are calculated. The necessity of the segmentation and the optimal segmentation position are evaluated based on a criterion vector. If the statistical difference between the two parts after segmentation is significant, the segmentation point is considered valid; if multiple segmentation points are valid, the point that maximizes the statistical difference between the two parts is selected as the optimal segmentation point. The above process is repeated for each segmented sub-segment until a termination condition is met, such as the sub-segment length being less than a preset minimum value, or the statistical characteristics within the sub-segment meeting the uniformity requirement. The result of progressive subdivision is that the time interval sequence is divided into multiple uniformly characteristic subsequences, each of which may correspond to a specific working condition or fault mode of a robot joint.

[0111] During normal operation, the time interval sequence is relatively stable, with low and gradual fluctuations in the fluctuation intensity function value, resulting in a simple progressive subdivision. When pitting corrosion begins to appear in the bearing, small but regular changes occur in the time interval sequence, and the fluctuation intensity function shows a distinct peak at a specific location. The progressive subdivision divides the sequence into multiple subsequences. Analyzing the statistical characteristics of these subsequences reveals that the characteristics of one subsequence highly match the theoretical characteristics of bearing pitting corrosion, successfully achieving early detection of latent faults.

[0112] By calculating the changes in adjacent points within a time interval sequence, a fluctuation intensity function is constructed. Based on this function, a progressive scaling analysis is performed, effectively improving the sensitivity and accuracy of fault detection. The fluctuation intensity function accurately reflects the local fluctuation characteristics of the time interval sequence, providing a reliable basis for scaling. The progressive scaling strategy can adaptively handle different working conditions and fault types, avoiding the limitations of fixed-parameter methods. The scaling criteria based on statistical characteristics comprehensively consider the multidimensional features of the time interval distribution, improving classification accuracy. This method is applicable to the detection of latent faults in various types of robot joints, especially showing significant advantages for fault types that are difficult to identify using traditional methods, such as early bearing damage, gear micro-cracks, and poor lubrication.

[0113] Based on the changing trends of fault characteristic probabilities, the robot joints are identified as having latent fault risks, and fault risk warning information is output, including:

[0114] Calculate the difference between adjacent time points in the fault feature probability sequence, construct the fluctuation function of the fault feature probability, and identify the fault change cycle;

[0115] Based on the changing patterns within the cycle, a fault development prediction curve is constructed. Based on the relationship between the prediction curve and the dynamic early warning threshold, the potential for hidden faults in the robot joints is determined, and fault risk warning information is output.

[0116] When calculating the difference between adjacent time points in a fault characteristic probability sequence, for each time point in the sequence, the fault characteristic probability value at the current time point is subtracted from the fault characteristic probability value at the previous time point to obtain the probability change. This difference calculation method reflects the rate of change of the fault characteristic probability over time; a positive change indicates an increase in the fault probability, while a negative change indicates a decrease. The absolute value of the change reflects the drastic nature of the change. To more comprehensively characterize the change characteristics, the relative rate of change can be calculated simultaneously, i.e., the change is divided by the probability value at the previous time point, resulting in a relative rate of change sequence.

[0117] When constructing the fluctuation function of the fault characteristic probability, a sliding window smoothing method is used to reduce the impact of random fluctuations. The window length is set according to the operating characteristics of the robot joints, with a typical value being the duration of multiple work cycles. At each window position, the root mean square value of the difference sequence within the window is calculated as the fluctuation intensity at that position. The fluctuation function is represented in continuous function form, obtained by spline interpolation of discrete points. The fluctuation function reflects the distribution of the drastic change in the fault characteristic probability over time, helping to identify key time points and stages in fault development.

[0118] When identifying fault change cycles, a periodicity analysis is performed on the fluctuation function. Autocorrelation analysis is used to calculate the autocorrelation coefficient of the fluctuation function under different time delays. When the autocorrelation coefficient exhibits periodic fluctuations with time delay, the corresponding time delay value is the potential cycle length. To improve the accuracy of cycle identification, a spectral analysis method is also employed. The fluctuation function is transformed to the frequency domain using Fourier transform, and the main peaks in the spectrum are identified. The reciprocal of the frequency corresponding to each peak is the cycle length. The results of both autocorrelation and spectral analysis are combined to determine the most probable fault change cycle. When the fluctuation function does not exhibit obvious periodicity, an adaptive piecewise method is used, dividing the entire observation time into multiple stages based on the local characteristics of the fluctuation function.

[0119] When constructing a fault development prediction curve based on the changing patterns within a cycle, the first step is to analyze the changing patterns of fault characteristic probabilities within each cycle or stage. Common changing patterns include linear growth, exponential growth, S-shaped growth, and fluctuating growth. For each pattern, a corresponding mathematical model is selected for fitting, such as a linear model, exponential model, or autoregressive model. Model parameters are determined using the least squares method or maximum likelihood estimation to minimize the fitting error between the model curve and historical data. Model selection employs cross-validation, dividing historical data into training and validation sets. Multiple candidate models are trained on the training set, and the predictive performance of each model is evaluated on the validation set. The model with the best performance is selected as the final model.

[0120] Based on the selected model and parameters, the trend of fault characteristic probability changes over a future period is extrapolated, forming a fault development prediction curve. The time span of the prediction curve is determined according to the maintenance cycle of the robot joints and the fault development rate, typically the length of multiple fault change cycles. The prediction curve not only provides point prediction values ​​of fault characteristic probabilities but also estimates the range of uncertainty in the prediction, represented by confidence intervals. The calculation of confidence intervals takes into account factors such as model fitting error, parameter estimation error, and random disturbances, providing more comprehensive information for decision-making.

[0121] When constructing dynamic warning thresholds, factors such as the importance of robot joints, maintenance costs, and downtime losses are considered. The dynamic threshold consists of a base threshold and an adjustment factor. The base threshold is determined based on the historical distribution of fault characteristic probabilities and an acceptable false alarm rate, typically set as the mean of fault characteristic probabilities under normal conditions plus several standard deviations. The adjustment factor reflects the dynamic change in warning sensitivity over time and operating conditions, calculated from factors such as robot load intensity, cumulative running time, and ambient temperature. Under critical operating conditions, the adjustment factor lowers the threshold to increase warning sensitivity; under non-critical operating conditions, the adjustment factor raises the threshold to reduce false alarms.

[0122] By comparing the relationship between the fault development prediction curve and the dynamic early warning threshold, the potential for latent faults in robot joints is identified. A fault risk warning is triggered when the prediction curve exceeds the dynamic early warning threshold within the prediction time range. The warning information comprises multiple levels: the warning time point (the point at which the prediction curve first exceeds the threshold); the warning level, determined based on the degree and duration of the curve exceeding the threshold, categorized into minor, moderate, and severe levels; the fault type, inferred from the calculated source and fluctuation characteristics of the fault characteristic probability, such as bearing damage, gear wear, or poor lubrication; and recommended measures, providing corresponding handling suggestions based on the warning level and fault type, such as enhanced monitoring, planned maintenance, or emergency repairs.

[0123] By calculating the difference between adjacent time points in the fault feature probability sequence, a fluctuation function is constructed to identify the fault change cycle. Based on the change pattern, the fault development trend is predicted, achieving early and accurate warning of latent faults in robot joints. Fluctuation function analysis can capture subtle changes during fault development, enhancing sensitivity to latent faults. The cycle identification technology considers the impact of the robot's working cycle, improving the predictive accuracy. The prediction model is adaptively constructed based on historical data, exhibiting strong generalization ability. Dynamic warning thresholds balance warning sensitivity and reliability, reducing false alarms and missed alarms. This method can provide reliable warnings in the early stages when fault symptoms are not obvious, allowing maintenance personnel sufficient time to handle the situation, avoiding sudden downtime and equipment damage caused by fault development, significantly reducing maintenance costs, and improving equipment availability and production efficiency. This method is easily integrated into existing robot monitoring systems, with broad application prospects and significant economic benefits.

[0124] Secondly, embodiments of the present invention provide an artificial intelligence-based robot joint latent fault detection system, the system comprising:

[0125] The data acquisition module is used to collect vibration signal data sequences from the robot's joints.

[0126] The frequency domain decomposition module is used to perform frequency domain decomposition on the vibration signal data sequence to obtain vibration components in multiple frequency bands.

[0127] The feature extraction module is used to generate vibration spectra based on vibration components, calculate the average energy of each frequency band in the vibration spectra, and determine the characteristic frequency bands.

[0128] The fault analysis module is used to statistically analyze the time points of energy anomalies within the characteristic frequency band, analyze the probability distribution characteristics of their time interval sequences, and obtain the fault characteristic probability.

[0129] The early warning output module is used to determine the potential hidden fault risks in robot joints based on the changing trends of fault characteristic probabilities and to output fault risk warning information.

[0130] Thirdly, one technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0131] Fourthly, one technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0132] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting latent joint faults in robots based on artificial intelligence, characterized in that, Includes the following steps: Collect vibration signal data sequences from robot joints; Frequency domain decomposition is performed on the vibration signal data sequence to obtain vibration components in multiple frequency bands; Vibration spectra are generated based on vibration components, and the average energy of each frequency band in the vibration spectra is calculated to determine the characteristic frequency bands. By analyzing the probability distribution characteristics of the time interval sequence of energy anomalies within the statistical characteristic frequency band, the fault characteristic probability can be obtained. Based on the changing trend of fault characteristic probabilities, the robot joints are identified as having hidden fault risks, and fault risk warning information is output. Vibration spectra are generated based on vibration components. The average energy of each frequency band in the vibration spectra is calculated, and the characteristic frequency bands are determined, including: Time-frequency conversion is performed based on vibration components to generate vibration spectra containing multiple frequency bands; A nonlinear enhancement function is constructed based on the time-domain rate of change sequence of the vibration components to enhance the vibration spectrum; Calculate the mean energy of each frequency band of the enhanced vibration spectrum, and generate time-domain weighting coefficients based on energy fluctuation characteristics; The energy mean is weighted according to the time-domain weighting coefficients to determine the characteristic frequency band; The nonlinear enhancement function constructed based on the time-domain rate of change sequence of vibration components includes: Calculate the local fluctuation intensity of the time-domain rate of change sequence and construct an initial enhancement function; A dynamic adjustment factor is set based on the local fluctuation intensity, and the dynamic adjustment factor is substituted into the initial enhancement function to generate an adaptive enhancement function. Calculate the energy distribution differences of vibration spectra in different frequency bands and establish an energy difference assessment index; The method of optimizing the enhancement function based on the energy difference assessment index yields the final nonlinear enhancement function; Methods for optimizing enhancement functions based on energy difference assessment indices include: Extract the energy gradient sequence between frequency bands to construct the optimization objective function, and perform grouped iterative optimization on the coefficients of the enhancement function. The optimized coefficients and interval energy characteristics are non-uniformly compensated to generate the final nonlinear enhancement function. By analyzing the probability distribution characteristics of the time interval sequences of energy anomalies within the statistical characteristic frequency band, the fault characteristic probabilities are obtained, including: A time-domain adaptive gain function is constructed to enhance the energy sequence of the characteristic frequency band, and a progressive scale subdivision analysis is performed on the interval sequence of energy anomaly time points to obtain the fault characteristic probability.

2. The method according to claim 1, characterized in that, Methods for setting dynamic adjustment factors include: The fluctuation intensity sequence is decomposed into multiple layers and fluctuation features are extracted. The weights are adjusted by combining a time-varying gain function, and a dynamic adjustment factor is generated by nonlinear combination.

3. The method according to claim 1, characterized in that, The steps of the progressive scaling subdivision analysis include: Calculate the change in adjacent points in the time interval sequence and construct the fluctuation intensity function; The initial scale division position of the time interval sequence is determined based on the fluctuation intensity function, and the subdivision criteria are constructed according to the statistical characteristics within the scale to progressively subdivide the time interval sequence.

4. The method according to claim 1, characterized in that, Based on the changing trends of fault characteristic probabilities, the robot joints are identified as having latent fault risks, and fault risk warning information is output, including: Calculate the difference between adjacent time points in the fault feature probability sequence, construct the fluctuation function of the fault feature probability, and identify the fault change cycle; Based on the changing patterns within the cycle, a fault development prediction curve is constructed. Based on the relationship between the prediction curve and the dynamic early warning threshold, the potential for hidden faults in the robot joints is determined, and fault risk warning information is output.

5. An artificial intelligence-based robot joint latent fault detection system, used to implement the method described in any one of claims 1-4, characterized in that, The system includes: The data acquisition module is used to collect vibration signal data sequences from the robot's joints. The frequency domain decomposition module is used to perform frequency domain decomposition on the vibration signal data sequence to obtain vibration components in multiple frequency bands. The feature extraction module is used to generate vibration spectra based on vibration components, calculate the average energy of each frequency band in the vibration spectra, and determine the characteristic frequency bands. The fault analysis module is used to statistically analyze the time points of energy anomalies within the characteristic frequency band, analyze the probability distribution characteristics of their time interval sequences, and obtain the fault characteristic probability. The early warning output module is used to determine the potential hidden fault risks in robot joints based on the changing trends of fault characteristic probabilities and to output fault risk warning information.

Images