A robot joint module electrical fault diagnosis method and system and a storage medium

Abstract

The application relates to a robot joint module electrical fault diagnosis method and system and a storage medium, and relates to the technical field of computers, and comprises the following steps: generating a feature extraction time window set based on a motion state signal; performing time-frequency transformation on an electrical time sequence signal segment of each feature extraction time window to obtain a time-frequency feature map set; performing attention weight processing on each time-frequency feature map based on the motion state signal to obtain an attention weight set; performing weighted processing on each time-frequency feature map based on the attention weight set to obtain a weighted time-frequency feature map set; identifying each weighted time-frequency feature map based on a preset electrical fault identification model to obtain an electrical fault type sub-result set; and determining the cumulative confidence of the electrical fault type sub-result based on the electrical fault type sub-result set to obtain an electrical fault diagnosis result. The application has the effects of improving the accuracy of electrical fault diagnosis, identifying early weak faults and reducing the false alarm rate.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, system and storage medium for diagnosing electrical faults in robot joint modules. Background Technology

[0002] A robot joint module is the core execution unit for enabling multi-degree-of-freedom movement of a robot joint. It can integrate servo motors, reducers, encoders, drivers, and various electrical circuits.

[0003] In related technologies, because robot joint modules operate under complex conditions for extended periods (e.g., frequent start-stop, acceleration / deceleration, forward / reverse rotation, and even overload), their internal electrical systems are prone to aging, damage, and performance degradation (e.g., short circuits between motor winding turns, increased contact resistance leading to loose connections and overheating, and soft breakdown of driver power devices). Electrical fault diagnosis for robot joint modules typically employs a monitoring method based on fixed thresholds: the system monitors parameters such as phase current of the joint motor, bus voltage, and driver temperature in real time. When any parameter exceeds a preset fixed threshold, an alarm is triggered and the system shuts down. Manual diagnosis is then performed to determine the type of electrical fault and initiate maintenance.

[0004] Regarding the aforementioned technologies, the monitoring method based on fixed thresholds has the following problems: First, it only judges the instantaneous amplitude of electrical parameters, without considering the allowable range of fluctuations in electrical parameters under the current motion state, and the threshold setting cannot be combined with the actual situation. If the threshold is set too high, weak faults will not be detected. If the threshold is set too low, false alarms will occur frequently under normal operating conditions. Second, it only compares the amplitude of electrical parameters, lacking the extraction and analysis of the time-frequency domain characteristics of the signal waveform, resulting in the failure to identify electrical faults such as high-frequency noise and occasional pulses that do not cause amplitude exceedances. Third, it cannot distinguish whether the harmonic components in the phase current originate from mechanical vibration or electrical asymmetry in the motor windings, making current fluctuations caused by mechanical vibration easily misjudged as electrical faults. Therefore, it leads to low accuracy in electrical fault diagnosis, failure to identify early faults, and a high false alarm rate, limiting the application of robot joint modules in fault early warning and preventive maintenance, and there is still room for improvement. Summary of the Invention

[0005] To overcome the challenges of low accuracy in electrical fault diagnosis, difficulty in identifying early-stage faults, and high false alarm rates, this application provides a method for diagnosing electrical faults in robot joint modules.

[0006] In a first aspect, this application provides a method for diagnosing electrical faults in a robot joint module, employing the following technical solution: Based on the collected motion state signals of the target joint module, a feature extraction time window set for the current operating segment is generated. Time-frequency transformation processing is performed on the electrical timing signal segments extracted from each feature extraction time window in the aforementioned feature extraction time window set to obtain a time-frequency feature map set. Based on the aforementioned motion state signals, attention weight processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain an attention weight set. Based on the aforementioned attention weight set, weighted processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain a weighted time-frequency feature map set. Based on a preset electrical fault identification model, electrical fault type identification is performed on each weighted time-frequency feature map in the aforementioned weighted time-frequency feature map set to obtain an electrical fault type sub-result set. Based on the aforementioned electrical fault type sub-result set, the cumulative confidence of the aforementioned electrical fault type sub-result in multiple consecutive feature extraction time windows is determined to obtain the electrical fault diagnosis result.

[0007] Secondly, this application provides an electrical fault diagnosis device for robot joint modules, which adopts the following technical solution: The acquisition module is used to acquire motion state signals of the target joint module; the memory is used to store the program of the above-mentioned method for diagnosing electrical faults in a robot joint module; the processor is used to load and execute the program in the memory and implement the above-mentioned method for diagnosing electrical faults in a robot joint module.

[0008] Thirdly, this application provides a computer storage medium capable of storing corresponding programs, which facilitates the implementation of a method for diagnosing electrical faults in a robot joint module, and adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as any of the above-described methods for diagnosing electrical faults in robot joint modules.

[0009] The above-described embodiments of this disclosure have the following beneficial effects: The robot joint module electrical fault diagnosis method of some embodiments of this disclosure can effectively improve the accuracy of electrical fault diagnosis, achieve effective identification of early and weak faults, and reduce the false alarm rate, thereby improving the application effect of robot joint modules in fault early warning and preventive maintenance. Specifically, the reasons for low diagnostic accuracy, difficulty in identifying early faults, and high false alarm rate are: traditional fixed threshold methods only independently judge the instantaneous amplitude of electrical parameters, without considering the influence of motion state on the fluctuation of electrical parameters. They also lack the extraction of time-frequency domain features of the signal, making it impossible to distinguish the source of harmonic components caused by mechanical vibration and electrical faults. Based on this, the robot joint module electrical fault diagnosis method of some embodiments of this disclosure first generates a feature extraction time window set for the current operating segment based on the collected motion state signal of the target joint module. This enables the analysis window of the electrical signal to be synchronously matched with the actual motion state of the joint, avoiding incomplete segment extraction during transient acceleration and deceleration by the fixed time window, and providing a data foundation adapted to the working conditions for subsequent feature extraction. Secondly, time-frequency transformation processing is performed on the electrical timing signal segments extracted from each feature extraction time window in the aforementioned feature extraction time window set to obtain a time-frequency feature map set. This allows the one-dimensional timing signal to be transformed into a two-dimensional time-frequency domain, enabling the capture of fault features that do not cause amplitude exceeding limits (e.g., high-frequency noise, intermittent pulses, and waveform distortion), overcoming the limitations of relying solely on amplitude judgment. Next, based on the aforementioned motion state signal, attention weight processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain an attention weight set. This allows the generation of an attention weight matrix with the same dimension as the time-frequency feature map based on motion state information such as the joint torque change rate, providing a weighting basis related to the motion state for subsequent weighting processing. Furthermore, based on the aforementioned attention weight set, weighting processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain a weighted time-frequency feature map set. This approach enhances transient electrical characteristics during high torque variations, suppresses background noise during steady-state no-load operation, and allows early, subtle fault characteristics (e.g., initial inter-turn short circuits and minor loose connections) to manifest under dynamic conditions. Then, based on a pre-defined electrical fault identification model, each weighted time-frequency feature map in the aforementioned weighted time-frequency feature map set is used for electrical fault type identification, resulting in a sub-set of electrical fault types. This enables the identification and classification of multiple fault types (e.g., winding inter-turn short circuits, power cable loose connections, encoder signal interference, and driver aging). Finally, based on the aforementioned sub-set of electrical fault types, the cumulative confidence of these sub-results across multiple consecutive feature extraction time windows is determined, yielding the electrical fault diagnosis result. Thus, by using a sliding window mechanism to accumulate identification results from multiple time windows for comprehensive arbitration, single-frame misjudgments caused by occasional electromagnetic interference are effectively suppressed, improving the stability and reliability of the diagnostic results.Furthermore, a diagnostic chain is established, encompassing motion-state adaptive signal segmentation, time-frequency domain feature explicitness, motion-driven attention enhancement, dual-stream decoupling identification, and cumulative confidence assessment. This chain enables the detection and identification of early, subtle electrical faults in dynamic and noisy robot joint operating environments. This improves the accuracy of electrical fault diagnosis, reduces the false alarm rate, and thus enhances the effectiveness of robot joint modules in fault early warning and preventative maintenance. Attached Figure Description

[0010] Figure 1 This is a flowchart of some embodiments of the method for diagnosing electrical faults in robot joint modules according to the present disclosure. Detailed Implementation

[0011] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0012] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0013] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0014] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0015] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0016] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] refer to Figure 1 The diagram illustrates a flow 100 of some embodiments of a robot joint module electrical fault diagnosis method according to the present disclosure. The robot joint module electrical fault diagnosis method includes the following steps: Step 101: Based on the collected motion state signals of the target joint module, generate a feature extraction time window set for the current running segment.

[0018] In some embodiments, the execution entity (e.g., an electronic device) of the above-described robot joint module electrical fault diagnosis method can be hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster composed of multiple servers or terminal devices, or as a single server or a single terminal device. When the computing device is software, it can be installed in the hardware devices listed above. It can be implemented as multiple software or software modules to provide distributed services, or as a single software or software module. No specific limitations are made here.

[0019] In some embodiments, the aforementioned execution entity can generate a feature extraction time window set for the current operating segment based on the collected motion state signals of the target joint module. The target joint module can be a robot joint execution unit that requires electrical fault diagnosis. The robot joint execution unit can be an electromechanical integrated component that realizes single-degree-of-freedom rotation or motion, and may include, but is not limited to, servo motors, reducers, encoders, drivers, and electrical interfaces. The motion state signals can be physical quantity signals reflecting the change in the motion state of the target joint module at the current moment, and may include, but are not limited to, joint command speed, joint torque command, and actual joint position. The current operating segment can be the operating condition type determined based on the motion state signals of the target joint module at the current moment. The feature extraction time window can be a time window for extracting a data segment of a specific length from the electrical timing signal.

[0020] In some optional implementations of certain embodiments, the execution entity may generate a feature extraction time window set for the current running segment based on the collected motion state signals of the target joint module, which may include the following steps: The first step is to determine the rate of change of velocity of the target joint module based on at least one joint command velocity from the aforementioned motion state signals. The joint command velocity can be a velocity command value generated by the robot's main controller (motion planner) and sent to the servo driver of the target joint module. The unit of the velocity command value can be radians per second. The rate of change of velocity can be the amount of change in the joint command velocity per unit time, and can be used to characterize whether the target joint module is currently in a dynamic process of acceleration or deceleration.

[0021] In practice, the rate of change of joint command velocity can be obtained by performing a first-order difference operation on the joint command velocity at multiple consecutive sampling times. For example, if the joint command velocity is 2.0 rad / s at the nth sampling time and 2.5 rad / s at the (n+1)th sampling time, and the sampling time interval is 1 ms, then the rate of change of velocity is 500 rad / s. 2 .

[0022] The second step involves determining the current operating segment identification result based on the comparison between the aforementioned rate of change of speed and the preset rate of change threshold. This current operating segment identification result includes: a steady-state operating segment and a transient transition segment. The preset rate of change threshold can be a critical rate of change value used to distinguish between steady-state operation and transient transition. This preset rate of change threshold can be determined through experimental calibration based on the actual dynamic characteristics of the target joint module. The steady-state operating segment can be an operating segment where the target joint module operates at a basically constant speed with a rate of change of speed less than the preset rate of change threshold. The transient transition segment can be an operating segment where the target joint module is in the process of acceleration / deceleration / stopping with a rate of change of speed greater than or equal to the preset rate of change threshold.

[0023] In practice, the absolute value of the rate of change of speed can be compared with a preset rate of change threshold: when the absolute value of the rate of change of speed is less than the preset rate of change threshold, the current operating segment is identified as a steady-state operating segment. When the absolute value of the rate of change of speed is greater than or equal to the preset rate of change threshold, the current operating segment is identified as a transient transition segment. For example, the preset rate of change threshold can be set to 100 rand / s. 2 If the current rate of change of velocity is 30 rand / s 2 If the current operating segment is determined to be a steady-state operating segment, then the current operating segment is considered to be steady-state. If the current velocity change rate is 500 rand / s... 2 If so, the current operating segment is determined to be the aforementioned transient transition segment.

[0024] The third step involves generating a first feature extraction time window with a first preset window length, in response to the identification result of the current operating segment as a steady-state operating segment. The first preset window length can be a fixed time length applicable to steady-state operating conditions. The first feature extraction time window can be a time interval determined by looking back from the current sampling time to the first preset window length. This first feature extraction time window can be used to extract electrical timing signal segments during steady-state operation.

[0025] In practice, if the current operating segment is identified as a steady-state operating segment, the first step is to obtain the value of a pre-configured first preset window length. Then, the current sampling time is taken as the end time. Next, the time obtained by subtracting the first preset window length from the current sampling time is taken as the starting time, and the time interval from this starting time to the end time is determined as the first feature extraction time window. For example, the first preset window length can be set to 20ms, and the current sampling time is 1000ms, thus generating a first feature extraction time window with a starting time of 980ms and an end time of 1000ms. Since the statistical characteristics of electrical signals within a steady-state operating segment are relatively stable, using a fixed-length first preset window can capture a sufficient number of data points to support subsequent periodic harmonic characteristic analysis.

[0026] Fourth step: In response to the above-mentioned current running section identification result being a transient transition section, dynamic segmentation is performed based on the waveform inflection point of the joint torque command in the above-mentioned motion state signal to generate a second feature extraction time window with a second variable window length, wherein the length of the above-mentioned second variable window length is inversely proportional to the torque change slope.

[0027] The aforementioned joint torque command can be the command value output by the current loop controller of the servo driver of the target joint module to control the motor output torque. The aforementioned joint torque command can be expressed as a percentage of the rated torque and the actual current value. The aforementioned waveform inflection point can be the point where the torque change trend changes in the curve of the joint torque command over time (e.g., the inflection point where the torque changes from increasing to decreasing, or from decreasing to increasing). The aforementioned dynamic segmentation can be a process of dividing the time window based on the waveform inflection point rather than a fixed time interval. The aforementioned second variable window length can be a window length determined based on the actual duration of the torque change process within the transient transition segment. The aforementioned second feature extraction time window can be a time interval determined with the starting time of a complete torque change process within the transient transition segment as the starting point and the second variable window length as the length. The aforementioned torque change slope can be the magnitude of the change in the joint torque command per unit time. The aforementioned torque change slope can be used to characterize the severity of the torque change.

[0028] In practice, the inverse relationship between the length of the second variable window and the slope of the torque change can be interpreted as follows: the more drastic the torque change and the steeper the slope, the shorter the duration of the torque change process, and consequently, the shorter the second variable window length. Conversely, the smoother the torque change and the smaller the slope, the longer the duration of the torque change process, and consequently, the longer the second variable window length. This allows the second feature extraction time window to fully cover an acceleration or deceleration process, avoiding the introduction of irrelevant steady-state data due to an excessively long window, or the truncation of complete transient electrical features due to an excessively short window.

[0029] Fifth step: Determine the first feature extraction time window and the second feature extraction time window as the feature extraction time window set.

[0030] In some optional implementations of certain embodiments, the execution entity may, in response to the current running segment identification result being a transient transition segment, dynamically segment based on the waveform inflection point of the joint torque command in the motion state signal to generate a second feature extraction time window with a second variable window length, which may include the following steps: The first step is to determine the time sequence of the joint torque command changes over time within the aforementioned transient transition segment. This time sequence is a numerical sequence of the joint torque commands arranged in chronological order of sampling time. For example, within the time interval of 1ms between the start time (1000ms) and the end time (1050ms) of the transient transition segment, 51 numerical values ​​of the aforementioned joint torque commands can be extracted. Arranging these 51 values ​​in chronological order constitutes the aforementioned time sequence.

[0031] The second step involves performing trend analysis on the aforementioned time series to obtain the inflection point location information for torque changes. This trend analysis can be a process of determining the direction (upward or downward) of change at each data point in the time series. The inflection point location information for torque changes refers to the sampling point number (time point) that identifies the turning point in the torque change trend within the time series.

[0032] In practice, firstly, the difference between two adjacent data points in the time series analysis can be determined. Then, the trend can be determined based on the sign (positive or negative) of the difference. A positive difference indicates an upward trend in torque, while a negative difference indicates a downward trend. When the sign of the difference between two adjacent segments changes from positive to negative (or vice versa), this turning point is identified as the waveform inflection point. The sampling point number (time) corresponding to the waveform inflection point can be used to determine the location of the torque change inflection point.

[0033] The third step, based on the aforementioned torque change inflection point location information, is to determine the start and end times of a complete torque change process. This complete torque change process can be the continuous torque change interval experienced from the initial inflection point of a torque change trend to the next inflection point in the same direction. For example, the interval between the inflection point where the torque begins to rise and the inflection point where the torque begins to fall can constitute a complete torque rise process.

[0034] In practice, two adjacent inflection point positions can be selected from the torque change inflection point position information. The sampling time corresponding to the first inflection point is determined as the process start time, and the sampling time corresponding to the second inflection point is determined as the process end time.

[0035] The fourth step is to determine the time interval between the start time and the end time of the above process, and obtain the second variable window length.

[0036] In practice, the time difference between the process end time and the process start time can be calculated, and this time difference can be used as the second variable window length. For example, the process start time can be set to 1010ms and the process end time to 1025ms. Then the time interval is 15ms, and 15ms can be used as the second variable window length.

[0037] The fifth step involves extracting corresponding segments from the electrical timing signal segment, using the start time of the aforementioned process as the starting point and the second variable window length as the extraction length, to obtain the second feature extraction time window. The aforementioned electrical timing signal segment can be a sequence of electrical quantity data acquired synchronously with the aforementioned motion state signal, and may include: the three-phase current signal of the servo motor and the DC bus voltage signal. For example, if the process starts at 1010ms, the second variable window length is 15ms, and the sampling interval is 1ms, then starting from the data point corresponding to 1010ms of the electrical timing signal segment, 15 data points are continuously extracted. The subsequence formed by these 15 data points is then determined as the electrical data within the second feature extraction time window. This electrical data completely covers the electrical response characteristics during a single torque transient change process.

[0038] Step 102: Perform time-frequency transformation processing on the electrical timing signal segments extracted from each feature extraction time window in the feature extraction time window set to obtain a time-frequency feature map set.

[0039] In some embodiments, the execution entity can perform time-frequency transformation processing on the electrical timing signal segments extracted from each feature extraction time window in the feature extraction time window set to obtain a time-frequency feature map set. The electrical timing signal can be a sequence of physical quantity signals that are acquired synchronously with the motion state signal and reflect changes in the internal electrical state of the target joint module. For example, the three-phase current signal of a servo motor. The time-frequency transformation processing can be a signal processing procedure that converts a one-dimensional time-domain electrical timing signal segment into a two-dimensional time-frequency domain representation. Through the time-frequency transformation processing, the energy distribution characteristics of the electrical timing signal segment at different time positions and different frequency scales can be observed simultaneously. The time-frequency feature map can be a two-dimensional data matrix characterizing the local time-frequency distribution of the signal. The two-dimensional data matrix can be a matrix with a time axis as the first dimension, a frequency axis as the second dimension, and energy intensity as the element value.

[0040] In practice, the electrical time-series signal segment within the time interval corresponding to the feature extraction time window is first obtained. Then, a time-frequency transformation processing operation (e.g., time-frequency analysis using continuous wavelet transform) is performed on the electrical time-series signal segment to convert it into a time-frequency feature map.

[0041] Step 103: Based on the motion state signal, attention weight processing is performed on each time-frequency feature map in the time-frequency feature map set to obtain the attention weight set.

[0042] In some embodiments, the execution entity can perform attention weighting processing on each time-frequency feature map in the time-frequency feature map set based on the motion state signal to obtain an attention weight set. The attention weighting processing can be a set of weight coefficients generated based on the rate of change of joint torque commands in the motion state signal. Each attention weight can be a separate attention weight for each time-frequency feature map. The attention weight refers to a weight matrix with the same dimension as the time-frequency feature map, where the value of each element represents the degree of attention paid to the corresponding time-frequency feature. The attention weight matrix can be a two-dimensional array of weight coefficients generated through the attention weighting processing.

[0043] In practice, firstly, the time series of joint torque commands within the same feature extraction time window corresponding to the time-frequency feature map can be obtained. Secondly, based on the time series, the rate of change corresponding to each sampling moment within the feature extraction time window is calculated. Then, the rate of change is normalized and mapped to a preset numerical range to obtain a normalized weight coefficient sequence. Next, the normalized weight coefficient sequence is matrix expanded to generate an attention weight matrix with the same number of rows and columns as the aforementioned time-frequency feature map.

[0044] Step 104: Based on the attention weight set, perform weighted processing on each time-frequency feature map in the time-frequency feature map set to obtain a weighted time-frequency feature map set.

[0045] In some embodiments, the execution entity may perform weighted processing on each time-frequency feature map in the time-frequency feature map set based on the aforementioned attention weight set, to obtain a weighted time-frequency feature map set. The weighted processing refers to the element-wise multiplication of the attention weight matrix in the attention weight set with the time-frequency feature map. The weighted time-frequency feature map may be a two-dimensional data matrix that preserves the original time-frequency structure and enhances key regions.

[0046] Step 105: Based on the preset electrical fault identification model, identify the electrical fault type for each weighted time-frequency feature map in the weighted time-frequency feature map set to obtain a sub-result set of electrical fault types.

[0047] In some embodiments, the aforementioned execution entity may, based on a preset electrical fault identification model, identify the electrical fault type of each weighted time-frequency feature map in the aforementioned weighted time-frequency feature map set, thereby obtaining a sub-set of electrical fault types. The preset electrical fault identification model may be a pre-constructed and trained mathematical model used to identify electrical fault types. The aforementioned sub-set of electrical fault types may be the fault type determination result identified for a single weighted time-frequency feature map, and may include, but is not limited to: normal state, winding turn-to-turn short-circuit trend, power cable loose connection, encoder signal electromagnetic interference, and driver power device aging.

[0048] In some optional implementations of certain embodiments, the execution entity may identify the electrical fault type of each weighted time-frequency feature map in the weighted time-frequency feature map set based on a preset electrical fault identification model, and obtain an electrical fault type sub-result set, which may include the following steps: The first step involves inputting the weighted time-frequency feature map into the first feature extraction stream of the preset electrical fault identification model to obtain a global harmonic feature vector reflecting the fundamental frequency harmonic characteristics of the winding state. The first feature extraction stream can be an information processing path within the preset electrical fault identification model used to extract global harmonic features. This first feature extraction stream can consist of three sequentially connected convolutional processing layers, a global pooling processing layer, a soft thresholding processing layer, and a feature vectorization processing layer. The global harmonic feature vector can be a one-dimensional array from the weighted time-frequency feature map that reflects the symmetry and balance of the three-phase windings of the servo motor of the target joint module. The fundamental frequency harmonic characteristics of the winding state can be the energy distribution variation law exhibited at the fundamental frequency and its lower harmonic components of the current signal, which is related to changes in the electrical parameters of the motor windings. When a short-circuit tendency occurs in the motor windings, the symmetry of the three-phase current is disrupted, and abnormal energy distribution appears at specific multiples of the fundamental frequency.

[0049] The second step involves inputting the weighted time-frequency feature map into the second feature extraction stream of the preset electrical fault identification model to obtain a local transient impact feature vector reflecting the characteristics of high-frequency decaying oscillations and abrupt pulse edges. The second feature extraction stream can be an information processing path within the preset electrical fault identification model used to extract local transient impact features. This second feature extraction stream can consist of two sequentially connected standard convolutional processing layers, a multi-scale dilated convolutional processing layer, and a feature vectorization processing layer. The local transient impact feature vector can be a one-dimensional array from the weighted time-frequency feature map that reflects the instantaneous high-frequency energy abrupt changes caused by poor contact or slight creepage. The high-frequency decaying oscillations can be the high-frequency oscillation components appearing in the current signal and their waveform characteristics decaying over time when an instantaneous connection or disconnection occurs at an electrical connection point. These waveform characteristics are represented as local energy spikes in the high-frequency region in the weighted time-frequency feature map. The abrupt pulse edge characteristics can be pulse waveform characteristics with steep rising or falling edges in the electrical signal. These pulse waveform characteristics can be represented as vertical stripes appearing simultaneously across the entire frequency band in the weighted time-frequency feature map.

[0050] For example, firstly, a weighted time-frequency feature map with the same dimension of 2048 x 256 can be input into the second feature extraction stream of the preset electrical fault identification model. The basic features are extracted by two standard convolutional processing layers. The multi-scale dilated convolutional processing layer uses three parallel dilated convolutional kernels with dilation rates of 1, 2 and 4 to process the feature map respectively and splice the results. Finally, a local transient impact feature vector with a length of 1024 is obtained.

[0051] The third step involves performing feature fusion processing on the aforementioned global harmonic feature vector and the aforementioned local transient impact feature vector to obtain a fused feature vector. This feature fusion processing can be a process of concatenating the global harmonic feature vector and the local transient impact feature vector along their respective feature dimensions.

[0052] The fourth step involves classifying fault types based on the fused feature vectors to obtain the electrical fault type sub-results corresponding to the weighted time-frequency feature maps. This fault type classification process can involve inputting the fused feature vectors into a classifier module consisting of two fully connected processing layers and a classification output layer connected sequentially. The module calculates the probability values ​​of the fused feature vectors belonging to each preset fault type and outputs the preset fault type with the highest probability value. The fully connected processing layers can establish weighted connections between each element of the input vector and each element of the output vector. The classification output layer can map the input values ​​to the corresponding probability values ​​for each fault type using a normalized exponential function.

[0053] The fifth step is to determine the obtained electrical fault type sub-results as the electrical fault type sub-result set.

[0054] In some optional implementations of some embodiments, the aforementioned preset electrical fault identification model is a two-stream neural network model, wherein the two-stream neural network model includes a first feature extraction stream and a second feature extraction stream. The first feature extraction stream has a built-in soft thresholding layer for adaptively filtering out noise components that are unrelated to electrical fault features. The second feature extraction stream uses a multi-scale dilated convolutional processing layer with different receptive field scales to capture transient impact features over different time spans.

[0055] Optionally, the construction of the aforementioned two-stream neural network model may include the following: The first feature extraction stream may consist of three sequentially connected convolutional processing layers, a global pooling processing layer, a soft thresholding processing layer, and a feature vectorization processing layer. Specifically, the first convolutional processing layer has a kernel size of 3x3, a stride of 1, and 64 output channels. The second convolutional processing layer has a kernel size of 3x3, a stride of 2, and 128 output channels. The third convolutional processing layer has a kernel size of 3x3, a stride of 2, and 32 output channels. The soft thresholding processing layer embeds a thresholding generation sub-network. This thresholding generation sub-network consists of a sequentially connected fully connected processing layer and a normalization processing layer. The second feature extraction stream may consist of two sequentially connected standard convolutional processing layers, a multi-scale dilated convolutional processing layer, and a feature vectorization processing layer. Specifically, the first standard convolutional processing layer has a kernel size of 3x3, a stride of 1, and 64 output channels. The second standard convolutional processing layer has a kernel size of 3x3, a stride of 2, and 128 output channels. The aforementioned multi-scale dilated convolutional processing layer can be composed of three parallel dilated convolutional branches: the first branch has a dilation rate of 1, the second branch has a dilation rate of 2, and the third branch has a dilation rate of 4. The outputs of the three branches are concatenated along the channel dimension to obtain the output of the multi-scale dilated convolutional processing layer. The aforementioned two-stream neural network model also includes a feature fusion module and a classifier module. The feature fusion module concatenates the global harmonic feature vector output from the first feature extraction stream with the local transient impact feature vector output from the second feature extraction stream. The classifier module consists of two fully connected processing layers and a classification output layer connected sequentially.

[0056] Optionally, training the above-mentioned two-stream neural network model may include the following steps: First, constructing an initial two-stream neural network model. This initial two-stream neural network model has the complete network structure described in the model construction section.

[0057] Secondly, obtain the sample dataset for training. This sample dataset includes multiple weighted time-frequency feature map samples labeled with real fault type tags. The weighted time-frequency feature map samples can be obtained by running the target joint module under different operating conditions, simulating five fault types: normal state, winding inter-turn short circuit trend, power cable loose connection, encoder signal electromagnetic interference, and driver power device aging. The corresponding weighted time-frequency feature maps can be generated according to steps 101 to 104 above, and each weighted time-frequency feature map can be labeled with its corresponding real fault type tag.

[0058] Next, the weighted time-frequency feature map samples from the above sample dataset are input into the above initial two-stream neural network model in batches, and forward propagation calculation is performed to obtain the predicted fault type results.

[0059] Then, based on the difference between the predicted fault type results and the actual fault type labels, the value of the cross-entropy loss function is calculated. Here, the cross-entropy loss function is a loss function used to measure the difference between the predicted probability distribution and the actual probability distribution.

[0060] Finally, based on the value of the loss function, the network parameters of the initial two-stream neural network model are iteratively updated using the backpropagation algorithm until the value of the loss function converges to below a preset threshold or the number of iterations reaches a preset maximum number of iterations. The initial two-stream neural network model at this point is then taken as the trained two-stream neural network model.

[0061] In some optional implementations of certain embodiments, the execution entity may input the weighted time-frequency feature map into the first feature extraction stream of the preset electrical fault identification model to obtain a global harmonic feature vector reflecting the fundamental frequency harmonic characteristics of the winding state, which may include the following steps: The first step involves performing multi-layer convolution on the weighted time-frequency feature map to obtain an intermediate feature map. This multi-layer convolution process involves stacking multiple convolutional layers to perform feature transformation and abstraction layer by layer on the input weighted time-frequency feature map. The intermediate feature map can be the output feature map of the intermediate layers after multi-layer convolution processing but before undergoing soft-thresholding shrinkage.

[0062] The second step involves performing global pooling on the intermediate feature map to obtain feature distribution statistics. This global pooling process can be implemented by calculating the statistical values ​​of the spatial location elements for each channel of the intermediate feature map. The feature distribution statistics can be a sequence of statistical values ​​reflecting the overall distribution level of features across all channels of the intermediate feature map. For example, the intermediate feature map can initially have 32 channels. Then, the average of the absolute values ​​of all 512 multiplied by 64 spatial location elements within each channel is calculated, and these 32 averages are used as the feature distribution statistics.

[0063] The third step involves generating an adaptive soft threshold based on the aforementioned feature distribution statistics through the soft thresholding layer in the first feature extraction stream. This soft thresholding layer can be a non-linear processing unit that generates a numerical threshold based on the statistical distribution characteristics of the input feature map. The adaptive soft threshold can take different values ​​for different input weighted time-frequency feature maps, thereby achieving adaptive filtering of noise components under different operating conditions.

[0064] The fourth step involves applying a soft-threshold shrinkage process to the intermediate feature map based on the adaptive soft threshold, resulting in a denoised feature map. This soft-threshold shrinkage process can be a non-linear transformation operation. It sets elements in the intermediate feature map whose absolute values ​​are less than the adaptive soft threshold to zero, and shrinks elements whose absolute values ​​are greater than or equal to the adaptive soft threshold towards zero by the corresponding threshold value.

[0065] The fifth step is to perform feature vectorization on the above denoising feature map to obtain the global harmonic feature vector. In practice, the elements in the denoising feature map can be unfolded sequentially according to a preset arrangement order and arranged end to end into a one-dimensional array to obtain the global harmonic feature vector.

[0066] For example, firstly, a weighted time-frequency feature map with dimensions 2048 x 256 can be input into the first feature extraction stream of a pre-defined electrical fault identification model. After passing through three convolutional processing layers, it becomes a multi-channel feature map of 512 x 64 x 32. Then, it is compressed into 32 statistical values ​​by a global pooling processing layer. A soft thresholding processing layer calculates a threshold of 0.15 and sets features smaller than this value to zero. Finally, a global harmonic feature vector of length 1024 is obtained.

[0067] Step 106: Based on the electrical fault type sub-result set, determine the cumulative confidence of the electrical fault type sub-result in multiple consecutive feature extraction time windows to obtain the electrical fault diagnosis result.

[0068] In some embodiments, the execution entity may determine the cumulative confidence of the electrical fault type sub-results in multiple consecutive feature extraction time windows based on the electrical fault type sub-result set, and obtain the electrical fault diagnosis result.

[0069] In some optional implementations of certain embodiments, the execution entity may determine the cumulative confidence of the electrical fault type sub-results over multiple consecutive feature extraction time windows based on the electrical fault type sub-result set, and obtain the electrical fault diagnosis result, which may include the following steps: The first step is to determine a sliding window with a preset time length. This sliding window is used to accumulate electrical fault type sub-results corresponding to multiple consecutive feature extraction time window sets. The sliding window can be a data observation interval with a fixed time span that can be shifted forward over time. The preset time length can be a pre-defined time span value. The unit of the preset time length can be milliseconds.

[0070] The second step involves determining each electrical fault type within the aforementioned sliding window based on the above electrical fault type sub-result set, thus obtaining the electrical fault type set. This electrical fault type set can be obtained by deduplicating fault types from all electrical fault type sub-results within the current sliding window.

[0071] The third step is to determine the number of each electrical fault type in the aforementioned electrical fault type set based on the sliding window, thus obtaining the fault type count set. The fault type count can be the total number of times a fault type appears within the current sliding window.

[0072] Fourth, based on the above set of fault type frequencies, for each electrical fault type in the above set of electrical fault types, perform the following confidence level acquisition steps: The first sub-step involves determining the cumulative confidence level of each electrical fault type based on the aforementioned set of fault type occurrences. This cumulative confidence level can be the ratio between the number of times a fault type occurs within the current sliding window and the total number of sub-results for all electrical fault types. This cumulative confidence level can be used to characterize the reliability of fault type identification results across multiple consecutive time windows.

[0073] The second sub-step, in response to the cumulative confidence level exceeding a preset diagnostic confidence threshold, determines the electrical fault type as an electrical fault diagnosis result. The preset diagnostic confidence threshold can be a pre-set minimum confidence limit value used to determine whether to output a fault diagnosis signal. This preset diagnostic confidence threshold is typically set to a relatively high value, such as 0.95 or 0.90, to ensure that an alarm is triggered only when the cumulative confidence level reaches a sufficiently high level, thereby avoiding unnecessary downtime caused by single-frame misjudgments due to occasional electromagnetic interference. For example, the preset diagnostic confidence threshold can be set to 0.95. The cumulative confidence level for a loose power cable connection is 0.80, for normal operation it is 0.13, and for encoder signal electromagnetic interference it is 0.07; none of these exceed 0.95. Therefore, no fault diagnosis result is output, and the cumulative confidence level is recalculated after the recognition result of the next feature extraction time window enters the sliding window. Optionally, if the cumulative confidence of the power cable connection continues to rise and reaches 0.98 within 50 consecutive feature extraction time windows, exceeding the preset diagnostic confidence threshold of 0.95, then the power cable connection is determined as an electrical fault diagnosis result, and an alarm signal is output for maintenance.

[0074] The above-described embodiments of this disclosure have the following beneficial effects: The robot joint module electrical fault diagnosis method of some embodiments of this disclosure can effectively improve the accuracy of electrical fault diagnosis, achieve effective identification of early and weak faults, and reduce the false alarm rate, thereby improving the application effect of robot joint modules in fault early warning and preventive maintenance. Specifically, the reasons for low diagnostic accuracy, difficulty in identifying early faults, and high false alarm rate are: traditional fixed threshold methods only independently judge the instantaneous amplitude of electrical parameters, without considering the influence of motion state on the fluctuation of electrical parameters. They also lack the extraction of time-frequency domain features of the signal, making it impossible to distinguish the source of harmonic components caused by mechanical vibration and electrical faults. Based on this, the robot joint module electrical fault diagnosis method of some embodiments of this disclosure first generates a feature extraction time window set for the current operating segment based on the collected motion state signal of the target joint module. This enables the analysis window of the electrical signal to be synchronously matched with the actual motion state of the joint, avoiding incomplete segment extraction during transient acceleration and deceleration by the fixed time window, and providing a data foundation adapted to the working conditions for subsequent feature extraction. Secondly, time-frequency transformation processing is performed on the electrical timing signal segments extracted from each feature extraction time window in the aforementioned feature extraction time window set to obtain a time-frequency feature map set. This allows the one-dimensional timing signal to be transformed into a two-dimensional time-frequency domain, enabling the capture of fault features that do not cause amplitude exceeding limits (e.g., high-frequency noise, intermittent pulses, and waveform distortion), overcoming the limitations of relying solely on amplitude judgment. Next, based on the aforementioned motion state signal, attention weight processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain an attention weight set. This allows the generation of an attention weight matrix with the same dimension as the time-frequency feature map based on motion state information such as the joint torque change rate, providing a weighting basis related to the motion state for subsequent weighting processing. Furthermore, based on the aforementioned attention weight set, weighting processing is applied to each time-frequency feature map in the aforementioned time-frequency feature map set to obtain a weighted time-frequency feature map set. This approach enhances transient electrical characteristics during high torque variations, suppresses background noise during steady-state no-load operation, and allows early, subtle fault characteristics (e.g., initial inter-turn short circuits and minor loose connections) to manifest under dynamic conditions. Then, based on a pre-defined electrical fault identification model, each weighted time-frequency feature map in the aforementioned weighted time-frequency feature map set is used for electrical fault type identification, resulting in a sub-set of electrical fault types. This enables the identification and classification of multiple fault types (e.g., winding inter-turn short circuits, power cable loose connections, encoder signal interference, and driver aging). Finally, based on the aforementioned sub-set of electrical fault types, the cumulative confidence of these sub-results across multiple consecutive feature extraction time windows is determined, yielding the electrical fault diagnosis result. Thus, by using a sliding window mechanism to accumulate identification results from multiple time windows for comprehensive arbitration, single-frame misjudgments caused by occasional electromagnetic interference are effectively suppressed, improving the stability and reliability of the diagnostic results.Furthermore, a diagnostic chain is established, encompassing motion-state adaptive signal segmentation, time-frequency domain feature explicitness, motion-driven attention enhancement, dual-stream decoupling identification, and cumulative confidence assessment. This chain enables the detection and identification of early, subtle electrical faults in dynamic and noisy robot joint operating environments. This improves the accuracy of electrical fault diagnosis, reduces the false alarm rate, and thus enhances the effectiveness of robot joint modules in fault early warning and preventative maintenance.

[0075] Based on the same inventive concept, embodiments of this application provide a robot joint module electrical fault diagnosis device, comprising: The acquisition module is used to acquire motion state signals of the target joint module; A memory for storing a program for diagnosing electrical faults in a robot joint module; The processor and memory can load and execute programs to implement a method for diagnosing electrical faults in robot joint modules.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0077] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a method for diagnosing electrical faults in a robot joint module.

[0078] Computer storage media include, for example, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media that can store program code.

[0079] Based on the same inventive concept, this application provides an intelligent terminal, including a memory and a processor. The memory stores a computer program that can be loaded and executed by the processor to perform an electrical fault diagnosis method for a robot joint module.

[0080] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0081] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for diagnosing electrical faults in a robot joint module, characterized in that, include: Based on the collected motion state signals of the target joint module, a feature extraction time window set for the current running segment is generated. The electrical timing signal segments extracted from each feature extraction time window in the feature extraction time window set are subjected to time-frequency transformation processing to obtain a time-frequency feature map set; Based on the motion state signal, attention weight processing is performed on each time-frequency feature map in the time-frequency feature map set to obtain an attention weight set; Based on the attention weight set, each time-frequency feature map in the time-frequency feature map set is weighted to obtain a weighted time-frequency feature map set; Based on a preset electrical fault identification model, electrical fault type identification is performed on each weighted time-frequency feature map in the weighted time-frequency feature map set to obtain an electrical fault type sub-result set; Based on the electrical fault type sub-result set, the cumulative confidence of the electrical fault type sub-result in multiple consecutive feature extraction time windows is determined to obtain the electrical fault diagnosis result.

2. The method for diagnosing electrical faults in a robot joint module according to claim 1, characterized in that, The step of generating a feature extraction time window set for the current running segment based on the collected motion state signal of the target joint module includes: The velocity change rate of the target joint module is determined based on at least one joint command velocity in the motion state signal; Based on the comparison result between the speed change rate and the preset change rate threshold, the current operating segment identification result is determined, wherein the current operating segment identification result includes: steady-state operating segment and transient transition segment; In response to the current operating segment identification result being a steady-state operating segment, a first feature extraction time window with a first preset window length is generated; In response to the current running segment identification result being a transient transition segment, dynamic segmentation is performed based on the waveform inflection point of the joint torque command in the motion state signal to generate a second feature extraction time window with a second variable window length, wherein the length of the second variable window length is inversely proportional to the torque change slope; The first feature extraction time window and the second feature extraction time window are determined as the feature extraction time window set.

3. The method for diagnosing electrical faults in a robot joint module according to claim 2, characterized in that, The step of dynamically segmenting the waveform based on the inflection point of the joint torque command in the motion state signal to generate a second feature extraction time window with a second variable window length in response to the current running segment identification result being a transient transition segment includes: Determine the time sequence of the joint torque command as a function of time within the transient transition segment; The time series is subjected to trend analysis to obtain the inflection point location information of torque change; Based on the torque change inflection point location information, determine the process start time and process end time corresponding to a complete torque change process; The time interval between the start time and the end time of the process is determined to obtain the second variable window length; Using the start time of the process as the starting point and the second variable window length as the extraction length, the corresponding segment is extracted from the electrical timing signal segment to obtain the second feature extraction time window.

4. The method for diagnosing electrical faults in a robot joint module according to claim 1, characterized in that, The step of identifying the electrical fault type in each weighted time-frequency feature map in the weighted time-frequency feature map set based on the preset electrical fault identification model to obtain the electrical fault type sub-result set includes: The weighted time-frequency feature map is input into the first feature extraction stream of the preset electrical fault identification model to obtain a global harmonic feature vector that reflects the fundamental frequency harmonic characteristics of the winding state; The weighted time-frequency feature map is input into the second feature extraction stream of the preset electrical fault identification model to obtain a local transient impact feature vector that reflects the characteristics of high-frequency damped oscillation and abrupt pulse edge. The global harmonic feature vector and the local transient impact feature vector are subjected to feature fusion processing to obtain a fused feature vector; Based on the fused feature vector, fault type classification is performed to obtain the electrical fault type sub-result corresponding to the weighted time-frequency feature map; The obtained electrical fault type sub-results are identified as electrical fault type sub-result sets.

5. The method for diagnosing electrical faults in a robot joint module according to claim 4, characterized in that, The preset electrical fault identification model is a two-stream neural network model, which includes a first feature extraction stream and a second feature extraction stream. The first feature extraction stream has a built-in soft thresholding layer for adaptively filtering out noise components that are irrelevant to the electrical fault features. The second feature extraction stream uses a multi-scale dilated convolutional processing layer with different receptive field scales to capture transient impact features over different time spans.

6. The method for diagnosing electrical faults in a robot joint module according to claim 4, characterized in that, The step of inputting the weighted time-frequency feature map into the first feature extraction stream of the preset electrical fault identification model to obtain a global harmonic feature vector reflecting the fundamental frequency harmonic characteristics of the winding state includes: The weighted time-frequency feature map is subjected to multi-layer convolution processing to obtain an intermediate feature map; The intermediate feature map is subjected to global pooling to obtain feature distribution statistics; Based on the feature distribution statistics, an adaptive soft threshold is generated through the soft thresholding processing layer in the first feature extraction stream. Based on the adaptive soft threshold, the intermediate feature map is subjected to soft threshold shrinkage processing to obtain a denoised feature map; The noise reduction feature map is processed into feature vectors to obtain global harmonic feature vectors.

7. The method for diagnosing electrical faults in a robot joint module according to claim 1, characterized in that, The step of determining the cumulative confidence of the electrical fault type sub-results across multiple consecutive feature extraction time windows based on the electrical fault type sub-result set, and obtaining the electrical fault diagnosis result, includes: A sliding window with a preset time length is determined, wherein the sliding window is used to accumulate electrical fault type sub-results corresponding to multiple consecutive feature extraction time window sets; Based on the electrical fault type sub-result set, each electrical fault type within the sliding window is determined to obtain the electrical fault type set; Based on the sliding window, the number of each electrical fault type in the electrical fault type set is determined to obtain the fault type count set; Based on the set of fault type occurrences, for each electrical fault type in the set of electrical fault types, the following confidence level acquisition steps are performed: Based on the set of fault type occurrences, determine the cumulative confidence level of the electrical fault type; In response to the cumulative confidence level being greater than a preset diagnostic confidence threshold, the electrical fault type is determined as an electrical fault diagnosis result.

8. A device for diagnosing electrical faults in a robot joint module, characterized in that, include: The acquisition module is used to acquire motion state signals of the target joint module; A memory for storing a program for diagnosing electrical faults in a robot joint module as described in any one of claims 1 to 7; The processor and the program in the memory can be loaded and executed by the processor to implement the method for diagnosing electrical faults in a robot joint module as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer program is stored and can be loaded by a processor and executed as described in any one of claims 1 to 7, which is a method for diagnosing electrical faults in a robot joint module.

Images