Intelligent brushless motor drive system with fault self-diagnosis function

Abstract

This invention specifically relates to an intelligent brushless motor drive system with self-diagnostic fault function, belonging to the field of motor drive control technology. It includes a residual generation module, used to calculate a reference voltage based on the duty cycle and the bus voltage, calculate the desired current through the motor voltage equation, generate a current residual by subtracting the desired current from the actual current, and generate a speed residual based on the mechanical motion equation, combining them into a residual vector. In this invention, the physical model residual is generated using the existing signals of the drive system through an analytical redundancy method. This solves the technical problem of existing solutions requiring additional sensors or hardware circuits to obtain fault redundancy information, achieving the technical effect of obtaining physically interpretable fault redundancy information without increasing hardware costs.

Description

Technical Field

[0001] This invention relates to the field of motor drive control technology, and in particular to an intelligent brushless motor drive system with fault self-diagnosis function. Background Technology

[0002] Brushless motor drivers are widely used in industrial drives, electric transportation, and robotics. However, during actual operation, drivers may experience various fault types, such as open or short circuits in the power transistors, Hall sensor failures, and short circuits between winding turns, posing a threat to system reliability.

[0003] Existing fault diagnosis solutions typically rely on additional multi-channel sensors or extra hardware circuitry to acquire redundant information in order to support the extraction and identification of fault characteristics.

[0004] The above solution has the following shortcomings: First, the addition of sensors and hardware circuitry increases the system's hardware cost; Secondly, signal transmission and processing delays limit real-time diagnostic capabilities; Third, once a fault is identified, the drive lacks the ability to automatically switch to derating mode and usually shuts down directly, resulting in decreased system availability. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent brushless motor drive system with fault self-diagnosis function to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent brushless motor drive system with self-diagnostic fault function includes: The signal acquisition and feature extraction module is used to acquire the three-phase current, bus voltage, rotor electrical angle, rotor speed and the drive duty cycle of each phase power transistor of the drive system at a preset sampling frequency, and to perform time-domain and frequency-domain feature calculations on the three-phase current and splice them to generate feature vectors. The residual generation module is used to calculate the reference voltage based on the duty cycle and the bus voltage, calculate the desired current through the motor voltage equation, generate a current residual by subtracting the desired current from the actual current, and generate a speed residual based on the mechanical motion equation, and combine them into a residual vector. The fusion diagnosis module is used to concatenate the standardized feature vector and the standardized residual vector into a fusion input vector, which is then input into a pre-trained lightweight one-dimensional convolutional neural network to output the fault category and confidence level. The fault-tolerant control module is used to read the corresponding control parameters from the pre-stored fault-tolerant strategy parameter table according to the fault category, switch the control mode of the driver and output control commands.

[0007] Preferably, the time-domain features include the average current value, variance, peak factor, waveform factor, and current polarity asymmetry of each phase of the three-phase current within the sampling window; the frequency-domain features are obtained by performing a fast Fourier transform on the three-phase currents, including the fundamental frequency amplitude, the second harmonic amplitude, the third harmonic amplitude, and the amplitude ratio of each harmonic to the fundamental frequency; wherein, the current polarity asymmetry is calculated as follows: the difference between the absolute values ​​of the maximum and minimum phase currents within the sampling window, divided by the maximum absolute value of the phase currents within the sampling window.

[0008] Preferably, in the residual generation module, the desired current is obtained as follows: the reference voltage is substituted into the voltage equation of the dq coordinate system, and the measured speed and motor parameters are used as known quantities. The voltage equation of the dq coordinate system is discretized and integrated, and the desired current values ​​of the d-axis and q-axis at each sampling time are solved step by step. The actual current is obtained by applying Park transformation to the three-phase current to obtain the actual d-axis current and the actual q-axis current. The current residual includes the d-axis current residual and the q-axis current residual, which are the difference between the actual d-axis current and the desired d-axis current, and the difference between the actual q-axis current and the desired q-axis current, respectively. The speed residual is the difference between the measured speed and the speed predicted by the model obtained by integrating the mechanical motion equation based on the estimated electromagnetic torque, load torque and damping coefficient.

[0009] Preferably, the stator resistance and inductance parameters in the voltage equation of the dq coordinate system are obtained through online parameter identification. The online parameter identification adopts the recursive least squares method, which takes the measured voltage and current data in each sampling period as input, outputs the updated motor parameter estimates, and substitutes the updated motor parameters into the voltage equation of the dq coordinate system for residual calculation.

[0010] Preferably, the lightweight one-dimensional convolutional neural network includes a first convolutional layer, a first batch of normalized layers, a first activation layer, a second convolutional layer, a second batch of normalized layers, a second activation layer, a global average pooling layer, a random deactivation layer, an intermediate fully connected layer, and an output fully connected layer connected in sequence; wherein, the global average pooling layer averages the output of the second activation layer along the time dimension to obtain a fixed-length feature representation; the output fully connected layer maps the output of the intermediate fully connected layer to a vector with the same dimension as the number of fault categories, and outputs the probability distribution of each fault category through a Softmax function; the fault categories include normal state, single-phase upper bridge arm open circuit, single-phase lower bridge arm open circuit, two-phase open circuit or short circuit, Hall sensor fault, and winding turn-to-turn short circuit.

[0011] Preferably, the lightweight one-dimensional convolutional neural network is trained using a knowledge distillation method, including: inputting training data into a teacher network with a ResNet18 structure for training, obtaining the output probability distribution of the teacher network for each fault category as soft labels; using the soft labels and the real labels together to form a supervision signal, performing distillation training on the lightweight one-dimensional convolutional neural network, wherein the loss function of the distillation training is a weighted sum of the cross-entropy loss between the output probability distribution of the lightweight one-dimensional convolutional neural network and the output probability distribution of the teacher network, and the cross-entropy loss between the output probability distribution of the lightweight one-dimensional convolutional neural network and the real labels, and the sum of the weight coefficients of the two losses is 1.

[0012] Preferably, the fusion diagnostic module further includes: when the confidence level is greater than a first threshold, locking the current fault category as the final diagnostic result; when the confidence level is between a second threshold and the first threshold, marking the current output as a suspected fault and initiating multi-frame voting, counting the occurrences of each fault category in the output of a consecutive preset number of frames, and confirming the fault category as the final diagnostic result if the occurrences of the same fault category exceed the voting threshold; when the confidence level is lower than the second threshold, determining it as a normal state; wherein, the second threshold is less than the first threshold.

[0013] Preferably, in the fault-tolerant control module: when the fault type is a single-phase upper bridge arm open circuit or a single-phase lower bridge arm open circuit, the PWM output signal of the faulty phase bridge arm is blocked, the motor drive mode is switched to a three-phase four-switch fault-tolerant topology mode, and the maximum output torque is limited to a preset proportion of the rated torque; when the fault type is a two-phase open circuit or short circuit, all PWM output signals are turned off and the active short circuit protection mode is switched on, and the main relay is disconnected to cut off the power circuit; when the fault type is a Hall sensor fault, the module automatically switches to a sensorless control mode, and if the current rotor speed is higher than a preset speed threshold, the rotor position information is obtained by using a back EMF zero-crossing detection method; if the current rotor speed is lower than the preset speed threshold, the module enters an open-loop forced pull mode; when the fault type is a winding inter-turn short circuit, the maximum allowable current is reduced to a preset derating proportion of the rated current.

[0014] Preferably, the method further includes an online update module, which is used to collect the currently running samples and extract their feature vector sets at preset time intervals when the driver is in normal operation, and calculate the distribution difference value between the current sample feature set and the normal data feature set in the pre-training stage using the maximum mean difference algorithm; when the distribution difference value is greater than a preset difference threshold, a fine-tuning process is started. The fine-tuning process only updates the weight parameters of the output fully connected layer of the lightweight one-dimensional convolutional neural network, using the most recently collected normal samples as fine-tuning data, and using the mean square error between the input feature vector reconstructed by the intermediate fully connected layer and the output fully connected layer and the original feature vector as the reconstruction error, and updating the parameters with the goal of minimizing the reconstruction error.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention utilizes the existing signals of the drive system itself to generate physical model residuals through analytical redundancy, solving the technical problem that existing solutions require additional sensors or hardware circuits to obtain fault redundancy information. It achieves the technical effect of obtaining physically interpretable fault redundancy information without increasing hardware costs.

[0016] 2. This invention solves the technical problem of blurred classification boundaries under noise interference and parameter uncertainty conditions when relying solely on analytical residuals. It achieves the technical effect of outputting fault categories within each control cycle by concatenating multidimensional feature vectors with physical model residuals and inputting the concatenation into a lightweight one-dimensional convolutional neural network for classification. Furthermore, by automatically matching pre-stored fault-tolerant control strategies and switching drive modes based on fault categories, it solves the technical problem of system unavailability caused by direct driver shutdown after a fault occurs. This achieves the technical effect of enabling the driver to continue operating in a derating manner when recoverable faults such as single-phase open circuits and Hall sensor failures occur. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1

[0021] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0022] In this embodiment, it includes: Brushless motor drivers may experience various fault types during operation, such as open or short circuits in power transistors, Hall sensor failure, and inter-turn short circuits in the windings. Traditional fault diagnosis solutions typically rely on additional multi-sensor systems or extra hardware circuitry to acquire redundant information, leading to increased hardware costs and limiting real-time diagnostic capabilities due to signal transmission and processing delays. Furthermore, once a fault is identified, the driver lacks the ability to automatically switch to a derating operating mode, often resulting in a direct shutdown and impacting system availability.

[0023] The hardware environment involved in this implementation includes: a three-phase inverter (using MOSFET or IGBT power transistors), a microcontroller with analog-to-digital conversion function, current sampling resistors for at least two phases, a bus voltage sampling circuit, and Hall effect or encoder position sensors. All of the above hardware components are inherent to the brushless motor drive system, and this solution does not require additional sensors.

[0024] A method for fault diagnosis and fault-tolerant control of a brushless motor driver according to an embodiment of the present invention includes the following steps: Step S1: Acquire the operating signals of the drive system and extract multi-dimensional feature vectors.

[0025] The drive system is sampled at a preset sampling frequency to obtain the three-phase current. Bus voltage Rotor electrical angle Rotor speed and the duty cycle of the PWM driven by each phase power transistor .

[0026] The time-domain features of the collected three-phase currents are calculated within each sampling window to obtain the average current value, variance, peak factor, and waveform factor. Simultaneously, a Fast Fourier Transform (FFT) is performed on the three-phase currents to extract the fundamental frequency amplitude, second harmonic amplitude, and third harmonic amplitude. The amplitude ratio of each harmonic to the fundamental frequency is calculated as the frequency-domain feature. The time-domain and frequency-domain features are concatenated to generate a feature vector. The feature vector is updated periodically for each sampling window.

[0027] It should be noted that the aforementioned time-domain characteristics also include current polarity asymmetry. The calculation method is as follows: ; in, Phase current within the sampling window The maximum value, Phase current within the sampling window The minimum value, Phase current within the sampling window The maximum absolute value. Current polarity asymmetry is used to characterize the degree of asymmetry between the positive and negative half-cycles of the current waveform. Similarly, for... Calculate the corresponding current polarity asymmetry for each.

[0028] It should be noted that the above time-domain characteristics also include the current zero-crossing distortion slope, which is the ratio of the instantaneous rate of change of the phase current at the zero-crossing point to the reference value of the rate of change at the zero-crossing point under normal operating conditions. The above frequency-domain characteristics also include the bus current ripple mode characteristics. When the hardware has a DC bus current... When sampling channels, for Apply a sliding discrete Fourier transform to extract its pulsating frequency distribution as an additional frequency domain feature, and concatenate it to the feature vector. middle.

[0029] In the feature vector Used before subsequent steps, for feature vectors Each component in the calculation is Z-score normalized to eliminate the impact of differences in the dimensions of different physical quantities on the fusion calculation.

[0030] Step S2: Generate physical model residuals based on analytical redundancy

[0031] Based on the measured PWM duty cycle With bus voltage Calculate the reference voltages for the d-axis and q-axis. Based on the reference voltage and measured rotational speed In addition to the motor parameters, the desired current is calculated using the voltage equation in the dq coordinate system of the brushless motor. The voltage equation in the dq coordinate system is as follows: ; ; in, The voltage along the d-axis. This is the q-axis voltage. For stator resistance, For d-axis inductance, It is the q-axis inductance. Electric angular velocity, It is a permanent magnet flux chain. For d-axis current, This is the q-axis current.

[0032] Furthermore, the aforementioned desired current and The method of obtaining it is: to use the reference voltage , Substituting into the voltage equation in the dq coordinate system above, and using the measured rotational speed within the current sampling period... Using the motor parameters as known quantities, the voltage equation in the dq coordinate system is discretized and integrated, and the expected current values ​​of the d-axis and q-axis at each sampling time are solved step by step.

[0033] Three-phase current Obtain the actual d-axis current through Park transformation and q-axis current The difference between the current and the desired current is used to generate the current residual: ; ; in, The d-axis current residual, This represents the q-axis current residual.

[0034] Based on the mechanical motion equations of the electric motor, using the measured speed And estimating electromagnetic torque Calculate the rotational speed residual The above mechanical motion equations are: ; in, For rotational inertia, For electromagnetic torque, For load torque, The damping coefficient is... For mechanical angular velocity. Rotational speed residual. Defined as measured speed Based on the above mechanical motion equations, the estimated electromagnetic torque Load torque and damping coefficient The difference between the model's predicted rotational speeds obtained through integration. When the deviation between the Hall signal transition moment and the predicted electrical angle exceeds a preset angle threshold, this deviation is included in the rotational speed residual. The calculation.

[0035] Furthermore, the above-mentioned estimated electromagnetic torque The method of obtaining it is: using the d-axis current q-axis current Permanent magnet flux d-axis inductance and q-axis inductance The electromagnetic torque of a permanent magnet synchronous motor is calculated using the formula: ; in, This represents the number of pole pairs of the motor. It is a permanent magnet flux chain. For d-axis inductance, It is the q-axis inductance. This is the actual d-axis current. This represents the actual q-axis current.

[0036] Combine the three residuals into a residual vector. ,in The d-axis current residual, For the q-axis current residual, This represents the rotational speed residual. The residual vector is then... Used before subsequent steps, for the residual vector Each component in the sample is Z-score normalized to eliminate... , (Unit: Ampere) and The impact of dimensional differences (in radians per second) on fusion calculations.

[0037] It should be noted that when a single-tube open-circuit fault occurs, the q-axis current residual... A specific frequency oscillation related to the phase loss will occur. When an inter-turn short circuit occurs in the winding, the d-axis current residual... and q-axis current residual At the same time, an asymmetrical DC offset component appears.

[0038] In this embodiment, to reduce the impact of motor parameter uncertainty on residual accuracy, the stator resistance in the voltage equation of the dq coordinate system is... and inductor The parameters are obtained through online parameter identification. Specifically, the recursive least squares method is used, taking the measured voltage and current data in each sampling period as input, outputting updated motor parameter estimates, and substituting the updated motor parameters into the voltage equation in the dq coordinate system to calculate the residuals, thereby reducing the impact of parameter drift on the accuracy of the residuals.

[0039] Step S3: Fuse the feature vector and residual vector to output the fault classification result.

[0040] The standardized feature vector With the standardized residual vector The vectors are concatenated into a fused input vector, which is then fed into a pre-trained lightweight one-dimensional convolutional neural network to output the fault category and its corresponding confidence level.

[0041] Furthermore, the aforementioned lightweight one-dimensional convolutional neural network comprises the following sequentially connected units: a first convolutional layer, a first batch of normalization layers, a first activation layer, a second convolutional layer, a second batch of normalization layers, a second activation layer, a global average pooling layer, a random deactivation layer, an intermediate fully connected layer, and an output fully connected layer. The data transfer relationship between these units is as follows: the fused input vector is fed into the first convolutional layer; the output of the first convolutional layer is passed to the first batch of normalization layers for normalization, then through the first activation layer (ReLU function) before being passed to the second convolutional layer; the output of the second convolutional layer is processed sequentially by the second batch of normalization layers and the second activation layer; the global average pooling layer averages the output of the second activation layer along the time dimension to obtain a fixed-length feature representation; the random deactivation layer randomly zeros the output of the global average pooling layer; the intermediate fully connected layer performs feature mapping on the output of the random deactivation layer; and the output fully connected layer maps the output of the intermediate fully connected layer to a vector with the same dimension as the number of fault categories, and outputs the probability distribution of each fault category through a Softmax function. The above fault categories are: normal state, single-phase upper arm open circuit, single-phase lower arm open circuit, two-phase open circuit or short circuit, Hall sensor fault, and winding inter-turn short circuit. In the probability distribution of the fully connected layer output, the category with the highest probability is the currently diagnosed fault category, and the corresponding probability value is the confidence level.

[0042] Furthermore, the training data for the aforementioned lightweight one-dimensional convolutional neural network was collected through fault injection experiments, covering a PWM duty cycle range of 0.2 to 0.8 and an engine speed range of 500 to 3000 rpm. Samples for each fault category were obtained by injecting the corresponding faults onto the experimental platform. The aforementioned lightweight one-dimensional convolutional neural network uses the cross-entropy loss function as the training objective and employs the Adam optimization algorithm for parameter updates.

[0043] In this embodiment of the application, in order to reduce the computational load of the model while maintaining classification accuracy to adapt to the real-time operation of the microcontroller, the above-mentioned lightweight one-dimensional convolutional neural network is trained using a knowledge distillation method. Step S3 may specifically include: Step S301: Train the teacher network and obtain soft labels

[0044] Training data is input into the teacher network for training, and the output probability distribution of the teacher network for each fault category is obtained. The teacher network adopts the ResNet18 structure, with a fused input vector as input and a probability distribution of each fault category as output. The output layer is a fully connected layer with the same dimension as the number of fault categories, and the probability distribution is output through the Softmax function. The training objective is the cross-entropy loss function, and the Adam optimization algorithm is used for parameter updates.

[0045] Step S302: Train a lightweight one-dimensional convolutional neural network with soft label supervision to obtain the compressed network parameters.

[0046] The output probability distribution of the teacher network is used as a soft label, which, together with the real label, constitutes the supervision signal. A lightweight one-dimensional convolutional neural network is then trained using distillation to obtain compressed network parameters. The loss function for this distillation training is a weighted sum of the soft label cross-entropy loss and the real label cross-entropy loss, and the Adam optimization algorithm is used for parameter updates.

[0047] Furthermore, the above-mentioned distillation training loss function is calculated as follows: ; in, For the total loss during distillation training, To reduce the cross-entropy loss between the output probability distribution of a one-dimensional convolutional neural network and the output probability distribution of the teacher network, To reduce the cross-entropy loss between the output probability distribution and the true label of a lightweight one-dimensional convolutional neural network, This is the weighting coefficient for the soft-label loss, with a value ranging from 0 to 1. It is used to balance the contribution ratio of the soft-label supervision signal and the real-label supervision signal to the parameter update.

[0048] Within each control cycle, the fault category and confidence level of the output from the lightweight one-dimensional convolutional neural network are acquired. When the confidence level is greater than a first threshold, the current fault category is locked as the final diagnosis result. When the confidence level is between a second threshold and the first threshold, the current output is marked as a suspected fault, and multi-frame voting is initiated: the occurrence frequency of each fault category is counted in the output of a preset number of consecutive frames. If the occurrence frequency of the same fault category exceeds the voting threshold, the fault category is confirmed as the final diagnosis result. When the confidence level is lower than the second threshold, it is determined to be a normal state, and no fault response is triggered.

[0049] It should be noted that the first threshold and the second threshold mentioned above are both preset confidence level judgment boundaries, satisfying the constraint that the second threshold is less than the first threshold. Their specific values ​​are determined during the offline debugging phase based on the actual system's tolerance for false alarm rate and false negative rate, and are pre-stored in the fault tolerance strategy parameter table.

[0050] Step S4: Execute the corresponding fault-tolerant control strategy based on the fault classification results.

[0051] Based on the fault category of the final diagnosis, the corresponding control parameters are read from the pre-stored fault tolerance strategy parameter table, the control mode of the driver is automatically switched, and the corresponding control commands are output.

[0052] When the fault type is a single-phase upper arm open circuit or a single-phase lower arm open circuit, the PWM output signal of the faulty phase arm is blocked, and the motor drive mode is switched from sinusoidal field-oriented control to a three-phase four-switch fault-tolerant topology mode. Specifically, the mapping relationship of the PWM output channel is reconfigured through software, and the remaining normal arms are used to complete the current commutation. At the same time, the reference coordinate system of the Park transform is modified to adapt to the phase loss state, and the maximum output torque is limited to 50% of the rated torque. A fault warning code is generated and reported to the upper controller via the CAN bus.

[0053] When the fault type is two-phase open circuit or short circuit, immediately shut down all PWM output signals and switch to active short circuit protection mode. If the motor is a permanent magnet synchronous type, perform specific phase-to-phase short circuit protection operations. Simultaneously, disconnect the main relay and cut off the power circuit. While performing the protection action, record the instantaneous current waveform data of the fault to a non-volatile storage area for subsequent offline analysis.

[0054] When the fault type is Hall sensor failure, it automatically switches to sensorless control mode. If the current rotor speed... If the current rotor speed exceeds a preset speed threshold, the rotor position information is obtained using a back EMF zero-crossing detection method. If the speed drops below the preset threshold, the motor enters an open-loop forced pull mode to maintain low-speed operation and notifies the external system via an alarm signal.

[0055] When the fault category is a short circuit between winding turns, the maximum allowable current is reduced to 60% of the rated current, and the PWM duty cycle of the cooling fan is increased to 100%. The motor temperature signal is continuously monitored, and if the motor temperature exceeds the safe temperature threshold, a safe shutdown operation is performed, shutting down all outputs.

[0056] It should be noted that the fault-tolerance strategy parameter table mentioned above is pre-stored in EEPROM, and includes parameters such as three-phase four-switch modulation coefficients, sensorless switching speed threshold, derating current ratio, and safe temperature threshold. The fault-tolerance strategy parameter table supports online updates via a remote communication interface without requiring hardware replacement.

[0057] In this embodiment of the application, in order to accommodate the deviation of the analytical model caused by changes in motor parameters due to aging or replacement, the following steps are also included in step S4: Step S5: Collect normally operating samples and update the diagnostic model parameters online.

[0058] When the driver is in normal operation, a preset number of currently running samples are collected at preset time intervals, and their feature vector sets are extracted. The maximum mean difference between the current sample feature set and the normal data feature set from the pre-training stage is calculated. Specifically, the maximum mean difference. The Maximum Mean Discrepancy algorithm is used to calculate the difference between the current sample feature set and the normal data feature set in the pre-training stage. The output is a scalar value that represents the magnitude of the difference in the distribution of the two feature sets.

[0059] When the maximum mean difference When the difference exceeds a preset threshold, a fine-tuning process is initiated. This process updates only the weight parameters of the fully connected layer at the output of the lightweight one-dimensional convolutional neural network. It uses recently collected normal samples as fine-tuning data, performs unsupervised anomaly detection learning through error reconstruction, and runs the Adam optimization algorithm for a preset number of rounds. The fine-tuning process runs in a background thread, without interrupting the real-time diagnostic task.

[0060] Furthermore, the reconstruction error is calculated as follows: the feature vector of a normal sample is input into the intermediate fully connected layer of a lightweight one-dimensional convolutional neural network to obtain intermediate feature representations, which are then reconstructed into a vector of the same dimension as the input feature vector by the output fully connected layer. The mean square error between the reconstructed vector and the original feature vector is then calculated as the reconstruction error. The fine-tuning process aims to minimize the reconstruction error by performing gradient descent updates only on the weight parameters of the output fully connected layer, aligning the feature mapping relationship of the last layer of the network with the distribution of the current normal samples.

[0061] In this embodiment of the application, in order to provide reference information on remaining service life for operation and maintenance after a recoverable failure occurs, the following steps are included in addition to step S4: Step S6: Predict remaining effective life based on fault status and operating parameters

[0062] When the final diagnosis is a recoverable fault, obtain the current fault severity score. Cumulative running time and the time integral of temperature and current , parameter vector The input is fed into a pre-trained Gaussian process regression model, which outputs a predicted value of the remaining effective lifetime. And its uncertainty range. Among them, the fault severity score... The value ranges from 0 to 1, where 0 indicates no fault and 1 indicates the most severe fault; the time integral of temperature and current. It is the integral of the product of motor temperature and the absolute value of phase current over the cumulative operating time, reflecting the degree of accumulation of combined thermoelectric stress; cumulative operating time The unit is hours.

[0063] Furthermore, the time integral of the above temperature and current The calculation method is as follows: ; in, for Motor temperature at any given time for The absolute value of the phase current at time t. The cumulative running time is measured in intervals from 0 to... Under discrete sampling conditions, the above integral is approximated by summation. Within each sampling period, the product of the absolute values ​​of the motor temperature and phase current at the current moment is multiplied by the sampling time interval and then accumulated to the time integral of temperature and current. .

[0064] In the parameter vector Before inputting the data into the Gaussian process regression model, Z-score standardization is performed on each component to eliminate fault severity scores. (Dimensionless), Cumulative running time (Unit: hours) and time integral of temperature and current The impact of dimensional differences (units are amperes, degrees Celsius, and hours) on the regression calculation. The above Gaussian process regression model uses historical fault operation data as the training set, standardized parameter vectors as input, remaining effective lifetime as the regression objective, and optimizes the kernel function hyperparameters using maximum likelihood estimation.

[0065] Furthermore, the above-mentioned fault severity score The method for obtaining it is as follows: based on the confidence level corresponding to the current fault category output by the lightweight one-dimensional convolutional neural network, combined with the residual vector. The norm of is calculated using a weighted summation method, and the formula is: ; in, The confidence level corresponding to the current fault category. For residual vectors The Euclidean norm, This is the maximum value of the Euclidean norm of the residual vector in the historical data, used to normalize the residual norm to the range of 0 to 1. is a weighting coefficient, ranging from 0 to 1, used to balance the contribution of confidence level and residual norm to the severity score of the fault.

[0066] It should be noted that the above recoverable fault types include minor winding turn-to-turn short circuits and single-phase open circuit faults during derating operation.

[0067] The predicted value of remaining effective life Output via display or communication interface. The predicted remaining effective lifetime value. When the torque falls below the preset lifespan threshold, the maximum output torque will be further limited to 20% of the rated torque, guiding the system into maintenance mode.

[0068] In this embodiment of the application, in order to perform statistical analysis and remote knowledge updates on the failure modes of the same type of driver, the following steps are also included in step S4: Step S7: Upload fault event data and obtain updated diagnostic parameters.

[0069] When a fault event occurs, the fault category, a snapshot of the current waveform at the moment of the fault, ambient temperature, load conditions, and timestamp are packaged into a fault event record and uploaded to the cloud server via a wireless communication interface. The cloud server aggregates and statistically analyzes the received fault event records of the same model of driver, grouping the fault event records using a clustering algorithm. The input of the clustering algorithm is the set of feature vectors for each fault event record, and the output is the group label of the fault mode and the statistical characteristics of each group. Based on the aggregated statistical results, high-frequency fault modes and early failure patterns are identified, and updated fault tolerance strategy parameters and diagnostic model parameters are generated accordingly. Other drivers of the same model download the updated fault tolerance strategy parameters and diagnostic model parameters via a remote communication interface.

[0070] Example 2

[0071] An industrial brushless motor driver, equipped with a conveyor belt drive unit in an automated production line, has a rated power of 2.2kW, a rated current of 5.8A, 4 pole pairs, and a rated speed of 1500rpm. After approximately 1800 hours of continuous operation, the conveyor belt began to exhibit intermittent vibrations, leading maintenance personnel to suspect a potential fault. The microcontroller integrated into the driver continuously acquires operating signals at a sampling frequency of 20kHz and executes a fault diagnosis process in real time. The following demonstrates the complete data flow process of this driver, from signal acquisition to fault-tolerant control.

[0072] In step S1, the microcontroller acquires the operating signals of the conveyor belt drive unit at a frequency of 20kHz. Within a certain sampling window, the original acquired values ​​of the three-phase current, bus voltage, rotor electrical angle, rotor speed, and PWM duty cycle of each phase are shown in Table 1.

[0073] Table 1 Original running signals within the sampling window

[0074] The time-domain characteristics of the three-phase currents within the sampling window are calculated. Taking phase A as an example, the time-domain characteristics of the three-phase currents within the window are calculated. The maximum value is 4.73A, the minimum value is -4.51A, and the absolute maximum value is 4.73A. Therefore, the polarity asymmetry of the A-phase current is: ; The value deviating from zero indicates a slight asymmetry in the positive and negative half-cycles of the A-phase current. A Fast Fourier Transform is applied simultaneously to the three-phase currents to extract the fundamental frequency and the amplitudes of each harmonic. These, along with the time-domain features, are then concatenated to generate a feature vector. The components were then Z-score standardized to eliminate dimensional differences, as shown in Table 2. Table 2. Main components of the feature vector (comparison before and after standardization)

[0075] The normalized value of the ratio of the second harmonic amplitude of phase A to the fundamental frequency amplitude reached 2.14, which is significantly higher than the normal operating condition reference, indicating that there is an abnormal frequency domain component in this phase.

[0076] In step S2, the dq-axis reference voltage is calculated based on the duty cycles of the three-phase PWM (A, B, and C phases) and the bus voltage. Then, combined with the measured speed and motor parameters, the desired current is recursively derived through the discretized integration of the voltage equation in the dq coordinate system. Motor parameters are obtained online using the recursive least squares method. The identified parameters and calculated desired currents within this sampling period are shown in Table 3.

[0077] Table 3. Results of Online Identification of Motor Parameters and Calculation of Expected Current

[0078] The actual dq-axis current is obtained by performing a Park transformation on the three-phase currents. This is then subtracted from the desired current to generate the current residual. Simultaneously, the rotational speed residual is calculated based on the mechanical motion equations. First, the electromagnetic torque is calculated using the actual dq-axis current, where the number of pole pairs... Actual q-axis current A, Actual d-axis current A: ; The residual vector was then generated and Z-score normalization was performed. The results are shown in Table 4.

[0079] Table 4. Residual vector calculation results (comparison before and after standardization)

[0080] q-axis current residual The standardized value is 1.87, and the speed residual is... The standardized value is 1.53, both of which deviate from the normal range, consistent with the expected behavior of the q-axis current residual oscillation at a specific frequency when a single tube is open-circuited.

[0081] In step S3, the standardized feature vectors are... With the standardized residual vector The concatenated input vectors are fed into a lightweight one-dimensional convolutional neural network deployed on a microcontroller. The network sequentially passes through two convolutional layers, a batch normalization layer, an activation layer, a global average pooling layer, a random deactivation layer, and two fully connected layers. Finally, the probability distribution of each fault category is output through the Softmax function, as shown in Table 5.

[0082] Table 5. Fault category probability distribution of the output of the lightweight one-dimensional convolutional neural network.

[0083] The category with the highest output probability is single-phase upper arm open circuit, with a corresponding confidence level of 0.847. This confidence level is higher than the preset first threshold (set to 0.80 in this system), directly locking the current fault category as single-phase upper arm open circuit without triggering a multi-frame voting process.

[0084] In step S4, the driver reads the control parameters corresponding to the single-phase upper bridge arm open circuit from the fault-tolerant strategy parameter table, blocks the PWM output of the A-phase upper bridge arm, switches the drive mode to a three-phase four-switch fault-tolerant topology mode, modifies the Park transformation reference coordinate system to adapt to the phase loss state, and limits the maximum output torque to 50% of the rated torque (i.e., 50% of the rated torque of 4.86 N·m, with a limit of approximately 2.43 N·m). Simultaneously, a fault warning code is generated and reported to the conveyor belt control host computer via the CAN bus.

[0085] In step S6, since the single-phase upper arm open circuit is a recoverable fault, the system immediately initiates the remaining effective life prediction. First, the fault severity score is calculated. ,in Let Euclidean norm be the residual vector. To output the maximum probability, The maximum Euclidean norm of the residual vector in the historical data, and the weighting coefficients. Take 0.6: ; ; Cumulative running time The time integral of temperature and current over 1800 hours Calculated by discrete accumulation as A·℃·h. The standardized parameter vector... Input the Gaussian process regression model, and the output results are shown in Table 6.

[0086] Table 6 Input parameters and output results for remaining useful life prediction

[0087] The remaining effective lifespan is predicted to be 214 hours, which is higher than the preset lifespan threshold (48 hours in this system). The system will not trigger further derating and will maintain the current derating operation state with a 50% torque limit. The prediction result will be output to the host controller through the communication interface.

[0088] In step S7, the system packages the fault category (single-phase upper bridge arm open circuit), a snapshot of the A-phase current waveform at the moment of the fault, the ambient temperature (37.2℃), the current load conditions, and the timestamp into a fault event record, and uploads it to the cloud server via the wireless communication interface. The cloud server performs clustering statistics on the historical fault event records of the same model of driver, and identifies that the frequency of A-phase upper bridge arm open circuits in this batch of drivers has significantly increased within the cumulative operating time range of 1600 to 2000 hours. Based on this, updated fault tolerance strategy parameters are generated and distributed to other drivers of the same model.

[0089] The entire data flow process has a clear logical progression. Step S1 starts from the original operating signal of the conveyor belt drive unit and extracts a standardized feature vector containing time-domain and frequency-domain features. Step S2 utilizes the online identified motor parameters to construct a physical model and generate a standardized residual vector carrying physical interpretability. Step S3 concatenates the two types of vectors and outputs the fault category and confidence level through a lightweight one-dimensional convolutional neural network, directly identifying an open circuit in the single-phase upper bridge arm. Step S4 automatically switches the fault-tolerant topology and limits torque based on the diagnostic results, allowing the conveyor belt to maintain operation under derating conditions. Step S6 calculates the fault severity score using the confidence level and residual norm of the diagnostic output, and predicts the remaining effective lifespan by combining the cumulative operating time and thermoelectric integral quantity, providing a quantitative reference for operation and maintenance decisions. Step S7 uploads the fault event to the cloud, supporting cross-device knowledge accumulation and parameter updates. The output data of each step serves as the direct input for the next step, ensuring data consistency and continuity throughout the entire process.

[0090] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0091] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent brushless motor drive system with self-diagnosis function, characterized in that, include: The signal acquisition and feature extraction module is used to acquire the three-phase current, bus voltage, rotor electrical angle, rotor speed and the drive duty cycle of each phase power transistor of the drive system at a preset sampling frequency, and to perform time-domain and frequency-domain feature calculations on the three-phase current and splice them to generate feature vectors. The residual generation module is used to calculate the reference voltage based on the duty cycle and the bus voltage, calculate the desired current through the motor voltage equation, generate the current residual by subtracting the desired current from the actual current, and generate the speed residual based on the mechanical motion equation, and combine them into a residual vector. The fusion diagnosis module is used to concatenate the standardized feature vector and the standardized residual vector into a fusion input vector, which is then input into a pre-trained lightweight one-dimensional convolutional neural network to output the fault category and confidence level. The fault-tolerant control module is used to read the corresponding control parameters from the pre-stored fault-tolerant strategy parameter table according to the fault category, switch the control mode of the driver and output control commands.

2. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, The time-domain features include the average current value, variance, peak factor, waveform factor, and current polarity asymmetry of each phase of the three-phase current within the sampling window; the frequency-domain features are obtained by performing a fast Fourier transform on the three-phase current, including the fundamental frequency amplitude, the second harmonic amplitude, the third harmonic amplitude, and the amplitude ratio of each harmonic to the fundamental frequency; wherein, the current polarity asymmetry is calculated as follows: the difference between the absolute values ​​of the maximum and minimum phase currents within the sampling window, divided by the maximum absolute value of the phase currents within the sampling window.

3. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, In the residual generation module, the desired current is obtained as follows: the reference voltage is substituted into the voltage equation of the dq coordinate system, and the measured speed and motor parameters are used as known quantities. The voltage equation of the dq coordinate system is discretized and integrated, and the desired current values ​​of the d-axis and q-axis at each sampling time are solved step by step. The actual current is obtained by applying Park transformation to the three-phase current to obtain the actual d-axis current and the actual q-axis current. The current residual includes the d-axis current residual and the q-axis current residual, which are the difference between the actual d-axis current and the desired d-axis current, and the difference between the actual q-axis current and the desired q-axis current, respectively. The speed residual is the difference between the measured speed and the speed predicted by the model obtained by integrating the mechanical motion equation based on the estimated electromagnetic torque, load torque and damping coefficient.

4. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, The stator resistance and inductance parameters in the voltage equation of the dq coordinate system are obtained through online parameter identification. The online parameter identification adopts the recursive least squares method, which takes the measured voltage and current data in each sampling period as input, outputs the updated motor parameter estimates, and substitutes the updated motor parameters into the voltage equation of the dq coordinate system to calculate the residual.

5. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, The lightweight one-dimensional convolutional neural network comprises a first convolutional layer, a first batch of normalized layers, a first activation layer, a second convolutional layer, a second batch of normalized layers, a second activation layer, a global average pooling layer, a random deactivation layer, an intermediate fully connected layer, and an output fully connected layer, connected sequentially. The global average pooling layer averages the output of the second activation layer along the time dimension to obtain a fixed-length feature representation. The output fully connected layer maps the output of the intermediate fully connected layer to a vector with the same dimension as the number of fault categories, and outputs the probability distribution of each fault category via a Softmax function. The fault categories include normal state, single-phase upper arm open circuit, single-phase lower arm open circuit, two-phase open circuit or short circuit, Hall sensor fault, and winding turn-to-turn short circuit.

6. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, The lightweight one-dimensional convolutional neural network is trained using a knowledge distillation method, including: inputting training data into a teacher network with a ResNet18 structure for training, obtaining the output probability distribution of the teacher network for each fault category as soft labels; using the soft labels and the real labels together to form a supervision signal, and performing distillation training on the lightweight one-dimensional convolutional neural network. The loss function of the distillation training is a weighted sum of the cross-entropy loss between the output probability distribution of the lightweight one-dimensional convolutional neural network and the output probability distribution of the teacher network, and the cross-entropy loss between the output probability distribution of the lightweight one-dimensional convolutional neural network and the real label, with the sum of the weight coefficients of the two losses being 1.

7. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, The fusion diagnostic module further includes: when the confidence level is greater than a first threshold, locking the current fault category as the final diagnostic result; when the confidence level is between a second threshold and the first threshold, marking the current output as a suspected fault and initiating multi-frame voting, counting the occurrences of each fault category in the output of a consecutive preset number of frames, and confirming the fault category as the final diagnostic result if the occurrences of the same fault category exceed the voting threshold; when the confidence level is lower than the second threshold, determining it as a normal state; wherein, the second threshold is less than the first threshold.

8. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, In the fault-tolerant control module: when the fault type is single-phase upper arm open circuit or single-phase lower arm open circuit, the PWM output signal of the faulty phase arm is blocked, the motor drive mode is switched to three-phase four-switch fault-tolerant topology mode, and the maximum output torque is limited to a preset ratio of the rated torque. When the fault category is two-phase open circuit or short circuit, all PWM output signals are turned off and active short circuit protection mode is switched in, and the main relay is disconnected to cut off the power circuit; when the fault category is Hall sensor fault, it automatically switches to sensorless control mode. If the current rotor speed is higher than the preset speed threshold, the rotor position information is obtained by back EMF zero crossing detection method. If the current rotor speed is lower than the preset speed threshold, it enters open-loop strong pull mode. When the fault category is a short circuit between winding turns, the maximum permissible current will be reduced to a preset derating ratio of the rated current.

9. The intelligent brushless motor drive system with fault self-diagnosis function according to claim 1, characterized in that, It also includes an online update module, which, when the driver is in normal operation, collects the currently running samples and extracts their feature vector sets at preset time intervals, and uses the maximum mean difference algorithm to calculate the distribution difference value between the current sample feature set and the normal data feature set in the pre-training stage; when the distribution difference value is greater than a preset difference threshold, a fine-tuning process is initiated. The fine-tuning process only updates the weight parameters of the output fully connected layer of the lightweight one-dimensional convolutional neural network, using the most recently collected normal samples as fine-tuning data, and using the mean square error between the input feature vector reconstructed by the intermediate fully connected layer and the output fully connected layer and the original feature vector as the reconstruction error, with the goal of minimizing the reconstruction error in parameter updates.

Images