Brushless direct current motor fault diagnosis method and system based on multi-source information fusion

Abstract

The invention discloses a brushless direct current motor fault diagnosis method and system based on multi-source information fusion, and relates to the technical field of motor fault diagnosis, and the method comprises the following steps: synchronously collecting multi-source data during the operation of a brushless direct current motor; preprocessing the collected multi-source data, and extracting feature vectors from the preprocessed data; constructing a multi-modal feature adaptive weighted fusion algorithm to process the feature vector, and calculating a comprehensive health index; constructing an improved residual attention classification network, and inputting the feature vectors into the network for fault classification and positioning; calculating a fault severity index based on the comprehensive health index and a fault classification result; and outputting a diagnosis result containing the comprehensive health index, the fault type, the fault position and the fault severity index. By fusing multi-source information such as electrical characteristics, thermodynamic characteristics and control signals, diagnosis, classification and positioning of various faults of the brushless direct current motor are realized.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, specifically to a method and system for diagnosing brushless DC motor faults based on multi-source information fusion. Background Technology

[0002] With rapid social development and industrial progress, the demand for brushless DC motors has increased, leading to their rapid rise in the global market. Their advantages, such as simple structure, lightweight, high efficiency, low energy consumption, low noise, and strong controllability, have made them widely used in home appliances, automobiles, medical equipment, aerospace, and industrial control.

[0003] However, to meet daily production demands, motors often operate for extended periods or under heavy loads, making them susceptible to various factors that can lead to malfunctions such as overvoltage, undervoltage, Hall sensor failure, and phase loss. These issues can cause excessive motor vibration, overheating, abnormal starting, and even controller and motor burnout. Motor failures can result in production stoppages, service interruptions, energy waste, and losses. Therefore, fault detection, diagnosis, and prediction for motors are crucial. Diagnosing and classifying the types of motor faults and performing timely preventative maintenance are essential for ensuring long-term reliable motor operation, minimizing downtime losses, and maintaining efficient industrial production.

[0004] However, existing diagnostic technologies still face numerous limitations and challenges in practical applications. First, existing methods often rely on single or a few signal sources, failing to fully utilize the multi-source heterogeneous data generated during motor operation. This results in incomplete information and a tendency for misjudgment or omission under complex operating conditions or concurrent faults. At the feature processing level, traditional methods often simply concatenate or fuse features extracted from different signals with fixed weights, lacking the ability to dynamically adjust feature contributions based on fault type and operating status. This prevents adaptive focusing on the most relevant fault symptoms, leading to insufficient robustness and generalization of the diagnostic model. Diagnostic models based on shallow machine learning or fixed rule thresholds have limited ability to distinguish complex nonlinear fault modes, especially exhibiting low accuracy in classifying early, weak faults and fault types with similar symptoms, making it difficult to meet the requirements of high-reliability applications.

[0005] For example, Chinese patent CN116955995A discloses a fault diagnosis method for a three-phase brushless DC motor inverter. The method involves acquiring the waveform signals of current and voltage when the inverter transformer is operating; decomposing the current and voltage waveform signals; obtaining upper and lower envelopes based on the numerical distribution characteristics of the waveform signals; obtaining the difference sequence of the current envelope; obtaining the gradient difference of the current envelope based on the gradient characteristics of the upper and lower envelopes; obtaining the current trend fluctuation coefficient; obtaining the current envelope window fluctuation coefficient; obtaining the differential variation coefficient; obtaining the threshold selection coefficient; obtaining the adaptive threshold based on wavelet transform; obtaining the denoised current and voltage waveforms; and comparing the current and voltage waveforms during normal operation with the denoised current and voltage waveforms to obtain feature similarity, thereby achieving fault diagnosis of the inverter and improving the accuracy and robustness of the fault detection algorithm.

[0006] Therefore, there is an urgent need for a fault diagnosis method that can integrate multi-source information, intelligent weighted features, and achieve accurate classification and severity quantification assessment. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing brushless DC motor fault diagnosis technologies, such as relying solely on data from a single type of sensor, being insensitive to early or minor faults, and having difficulty distinguishing different fault types with similar symptoms. It provides a brushless DC motor fault diagnosis method and system based on multi-source information fusion.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The fault diagnosis method for brushless DC motors based on multi-source information fusion includes the following steps:

[0010] Step S1: Synchronously collect multi-source data during the operation of the brushless DC motor;

[0011] Step S2: Preprocess the collected multi-source data and extract feature vectors composed of time-domain features and frequency-domain features from the preprocessed data;

[0012] Step S3: Construct a multimodal feature adaptive weighted fusion algorithm to process the feature vector and calculate the comprehensive health index;

[0013] Step S4: Construct an improved residual attention classification network, and input the feature vector into the network for fault classification and localization;

[0014] Step S5: Calculate the fault severity index based on the comprehensive health index and fault classification results;

[0015] Step S6: Output the diagnostic results, which include the comprehensive health index, fault type, fault location, and fault severity index.

[0016] Further, step S1 specifically includes: acquiring three-phase current and DC bus current through the current detection unit; acquiring three-phase voltage, DC bus voltage and power supply voltage through the voltage detection unit; acquiring motor stator temperature and controller power transistor temperature through the temperature detection unit; acquiring Hall sensor signals and encoder signals through the position / speed detection unit; and acquiring PWM control signals and protection circuit status signals through the control signal monitoring unit.

[0017] Further, in step S2, the preprocessing includes filtering and denoising the acquired raw signal; the time-domain features include: mean, variance, kurtosis, skewness, root mean square value, waveform factor, and peak factor; the frequency-domain features include: obtaining the spectrum through fast Fourier transform, and extracting the spectral peak, frequency centroid, and mean square frequency.

[0018] Furthermore, in step S3, the multimodal feature adaptive weighted fusion algorithm includes the following steps:

[0019] The time-domain features and frequency-domain features in the feature vector are normalized respectively;

[0020] Calculate the deviation of each feature, where the deviation represents the degree of difference between the current feature value and the preset normal state reference value;

[0021] Based on the deviation of each feature and the preset standard deviation parameter, the exponential weight coefficient is calculated.

[0022] The normalized features are weighted and summed, where the weights are determined by both the adaptive weight coefficient and the exponential weight coefficient.

[0023] A control signal compensation term is introduced, and the control signal features are added to the fusion result after being processed by the evaluation function;

[0024] The final output is the weighted sum of the normalized weighted eigenvalues ​​and the control signal compensation term, which serves as the comprehensive health index.

[0025] Furthermore, in step S4, the improved residual attention classification network specifically includes a feature encoder, a residual attention module, and a classifier;

[0026] The feature encoder adopts a three-layer fully connected network architecture. The first layer receives a 10-dimensional input feature vector and outputs a 64-dimensional feature. The second layer maps the 64-dimensional feature to 128-dimensional features. The third layer compresses the 128-dimensional feature to 64-dimensional features. The ReLU activation function is used between each layer and batch normalization is added.

[0027] The residual attention module contains two cascaded attention blocks. Each attention block contains a global average pooling layer, a bottleneck structure consisting of two fully connected layers, a sigmoid activation function, a multiplier for feature weighting, a residual connector, and a nonlinear transformation layer containing a fully connected layer and a ReLU activation function.

[0028] The classifier consists of two fully connected layers with parallel outputs. The first branch outputs a 7-dimensional vector corresponding to the probability distribution of seven fault types, and the second branch outputs a 6-dimensional vector corresponding to the fault probability distribution of six power transistor positions. Both branches use the Softmax activation function for probability normalization.

[0029] Furthermore, in step S5, calculating the fault severity index specifically includes the following steps:

[0030] The degree of health decline is calculated based on the difference between the baseline value of the comprehensive health index under normal conditions and the current comprehensive health index.

[0031] Based on the fault classification results, extract feature subsets related to specific fault types;

[0032] Calculate the relative deviation between the current fault-related characteristics and the corresponding characteristics under normal conditions;

[0033] The severity index of the failure is obtained by weighting the degree of decline in health status with the degree of relative deviation of characteristics.

[0034] Furthermore, in step S6, the fault types include: power supply fault, power transistor short circuit fault, motor phase loss fault, Hall position detection fault, inter-turn short circuit fault, bearing wear fault, and rotor demagnetization fault.

[0035] Furthermore, in step S6, the fault location specifically includes: the power transistor of the upper arm of phase A, the power transistor of the lower arm of phase A, the power transistor of the upper arm of phase B, the power transistor of the lower arm of phase B, the power transistor of the upper arm of phase C, and the power transistor of the lower arm of phase C.

[0036] A brushless DC motor fault diagnosis system based on multi-source information fusion, implemented based on any one of the brushless DC motor fault diagnosis methods based on multi-source information fusion, is characterized by comprising:

[0037] The multi-source data acquisition module is used to acquire multi-source data during the operation of the brushless DC motor.

[0038] The preprocessing and feature extraction module is used to preprocess and extract features from the collected multi-source data;

[0039] The multimodal data fusion module is used to calculate the comprehensive health index by applying an adaptive weighted fusion algorithm of multimodal features;

[0040] The fault classification module is used for fault classification and localization using an improved residual attention classification network;

[0041] The severity assessment module is used to calculate the fault severity index based on the comprehensive health index and fault classification results;

[0042] The diagnostic results output module is used to output diagnostic results that include comprehensive health index, fault type, fault location, and fault severity index.

[0043] Furthermore, the multi-source data acquisition module includes a current detection unit, a voltage detection unit, a temperature detection unit, a position / velocity detection unit, and a control signal monitoring unit, with each unit employing synchronous sampling technology.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. This invention overcomes the limitations of single-signal diagnosis by integrating multi-source information such as electrical, thermodynamic and control signals, and improves the accuracy of fault diagnosis and early detection capability.

[0046] 2. This invention uses a multimodal feature adaptive weighted fusion algorithm to dynamically adjust feature weights according to different operating conditions, so that the system can maintain diagnostic accuracy under load fluctuations and noise environments.

[0047] 3. By constructing an improved residual attention classification network, this invention can focus on the features most relevant to the fault, improve the ability to distinguish similar faults, and reduce false alarms and false negatives.

[0048] 4. This invention provides comprehensive fault assessment capabilities, which can not only detect and classify faults, but also assess the severity of faults and locate their positions, providing more comprehensive information support for maintenance decisions. Attached Figure Description

[0049] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0050] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0051] Figure 2 This is a system schematic diagram according to an embodiment of the present invention;

[0052] Figure 3 This is a diagram of the improved residual attention classification network architecture according to an embodiment of the present invention;

[0053] Figure 4 This is a diagram of the residual attention module architecture according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, the fault diagnosis method for brushless DC motors based on multi-source information fusion includes the following steps:

[0056] Step S1: Synchronously collect multi-source data during the operation of the brushless DC motor;

[0057] Step S2: Preprocess the collected multi-source data and extract feature vectors composed of time-domain features and frequency-domain features from the preprocessed data;

[0058] Step S3: Construct a multimodal feature adaptive weighted fusion algorithm to process the feature vector and calculate the comprehensive health index;

[0059] Step S4: Construct an improved residual attention classification network, and input the feature vector into the network for fault classification and localization;

[0060] Step S5: Calculate the fault severity index based on the comprehensive health index and fault classification results;

[0061] Step S6: Output the diagnostic results, which include the comprehensive health index, fault type, fault location, and fault severity index.

[0062] Step S1 specifically includes: acquiring three-phase current and DC bus current through the current detection unit; acquiring three-phase voltage, DC bus voltage and power supply voltage through the voltage detection unit; acquiring motor stator temperature and controller power tube temperature through the temperature detection unit; acquiring Hall sensor signals and encoder signals through the position / speed detection unit; and acquiring PWM control signals and protection circuit status signals through the control signal monitoring unit.

[0063] In step S2, the preprocessing includes filtering and denoising the acquired raw signal; the time-domain features include: mean, variance, kurtosis, skewness, root mean square value, waveform factor, and peak factor; the frequency-domain features include: obtaining the spectrum through fast Fourier transform, and extracting the spectral peak, frequency centroid, and mean square frequency.

[0064] In step S3, the multimodal feature adaptive weighted fusion algorithm includes the following steps:

[0065] The time-domain features and frequency-domain features in the feature vector are normalized respectively;

[0066] Calculate the deviation of each feature, where the deviation represents the degree of difference between the current feature value and the preset normal state reference value;

[0067] Based on the deviation of each feature and the preset standard deviation parameter, the exponential weight coefficient is calculated.

[0068] The normalized features are weighted and summed, where the weights are determined by both the adaptive weight coefficient and the exponential weight coefficient.

[0069] A control signal compensation term is introduced, and the control signal features are added to the fusion result after being processed by the evaluation function;

[0070] The final output is the weighted sum of the normalized weighted eigenvalues ​​and the control signal compensation term, which serves as the comprehensive health index.

[0071] The formula for calculating the comprehensive health index is as follows:

[0072]

[0073] in, This represents the overall health index, with a value range of [0,1]. A higher value indicates a healthier system. Indicates the total number of feature types. This represents the adaptive weight of the i-th class feature. Represents the feature vector of the i-th class. This represents the normalization function, using Min-Max normalization. This represents the deviation of the i-th type of feature. The standard deviation of the i-th feature is represented by the standard deviation of the i-th feature. This represents the control signal compensation coefficient, with a value range of [0.1, 0.3] and a default value of 0.2. This represents the weighting coefficient of the j-th type of control signal. Indicates the number of control signal types. This represents the vector of the j-th type of control signal. This represents the control signal evaluation function;

[0074] Among them, the deviation of the i-th type of feature The calculation formula is:

[0075]

[0076] in, This represents the baseline vector of the i-th feature class under normal conditions. The L2 norm of a vector is represented by the Euclidean norm.

[0077] The formula for calculating the control signal evaluation function is as follows:

[0078]

[0079] in, The optimal reference value for the j-th control signal represents the ideal value of this control signal under typical, healthy operating conditions. It is obtained through statistical measurements performed when the motor system is operating normally under rated conditions. The tolerance parameter for the j-th control signal is set by analyzing the engineering tolerance of the control signal within the allowable fluctuation range and based on the physical characteristics of the motor and controller and engineering experience.

[0080] The adaptive weight update formula for the i-th type of feature is:

[0081]

[0082] in, This indicates the time step, i.e., the number of iterations. This represents the learning rate, which controls the speed at which the weights are updated; its value is 0.05. This represents the loss function, using the mean squared error as the loss function. This represents the health index compensation coefficient, which accelerates weight convergence, and has a value of 0.2. This represents the difference between the actual health index and the predicted health index.

[0083] The control signals are signals obtained directly from the controller or drive circuit, and typically include: PWM duty cycle; actual switching frequency of the power transistor; protection circuit status flags, such as binary states of overcurrent, overvoltage, undervoltage, and overtemperature flags (0-normal, 1-triggered); and control modes, such as open-loop, closed-loop speed control, and closed-loop torque control.

[0084] The weighting coefficients of the j-th type of control signal are used to measure the relative importance of different control signals to the system health, and are determined based on prior knowledge and system characteristics, specifically as follows:

[0085] Prolonged high PWM duty cycle has a significant impact on lifespan, so it is assigned a high weight of 0.4.

[0086] Protection circuit triggering is direct evidence of a serious event and should be assigned the highest weight of 0.5;

[0087] Slight fluctuations in switching frequency have a relatively small impact and are assigned a low weight of 0.1.

[0088] The initial weights must satisfy the normalization condition. The sum of them equals 1.

[0089] like Figure 3 As shown, in step S4, the improved residual attention classification network specifically includes a feature encoder, a residual attention module, and a classifier;

[0090] The feature encoder adopts a three-layer fully connected network architecture. The first layer receives a 10-dimensional input feature vector and outputs a 64-dimensional feature. The second layer maps the 64-dimensional feature to 128-dimensional features. The third layer compresses the 128-dimensional feature to 64-dimensional features. The ReLU activation function is used between each layer and batch normalization is added.

[0091] The residual attention module contains two cascaded attention blocks. Each attention block contains a global average pooling layer, a bottleneck structure consisting of two fully connected layers, a sigmoid activation function, a multiplier for feature weighting, a residual connector, and a nonlinear transformation layer containing a fully connected layer and a ReLU activation function.

[0092] The classifier consists of two fully connected layers with parallel outputs. The first branch outputs a 7-dimensional vector corresponding to the probability distribution of seven fault types, and the second branch outputs a 6-dimensional vector corresponding to the fault probability distribution of six power transistor positions. Both branches use the Softmax activation function for probability normalization.

[0093] The training process of the improved residual attention classification network is as follows: a training set is constructed using historical data samples containing normal states and various fault states, each sample containing the 10-dimensional feature vector and the corresponding real fault label; the network parameters are updated using the Adam optimizer, the initial learning rate is set to 0.001, and batch training is adopted with a batch size of 32; early stopping is used during training to prevent overfitting.

[0094] The input data is a 10-dimensional feature vector, specifically including: 7-dimensional time-domain features, namely mean, variance, kurtosis, skewness, root mean square value, waveform factor and peak factor; and 3-dimensional frequency-domain features, namely spectral peak value, frequency centroid and mean square frequency.

[0095] The input features are normalized to the [0,1] interval using the Min-Max method;

[0096] The feature encoder employs a three-layer fully connected network architecture, responsible for mapping the raw features to a high-dimensional representation space:

[0097] The first encoding layer is a fully connected layer with 10-dimensional input and 64-dimensional output. It uses ReLU activation function, batch normalization for regularization, and a dropout rate of 0.2 for initial feature expansion and nonlinear transformation.

[0098] The second encoding layer is a fully connected layer with 64-dimensional input and 128-dimensional output. It uses ReLU activation function, batch normalization for regularization, and a dropout rate of 0.2 for further feature abstraction and dimensional expansion.

[0099] The third encoding layer is a fully connected layer with 128 input dimensions and 64 output dimensions. It uses ReLU activation function and batch normalization for regularization, which is used for feature compression and refinement to prepare feature representations of appropriate dimensions for the attention module.

[0100] The residual attention module contains two cascaded attention blocks. Each block has the same internal structure but independent parameters, which is used to achieve multi-level attention refinement and capture feature relationships at different levels of abstraction.

[0101] The classifier employs a dual-branch parallel architecture to achieve joint diagnosis of fault type and location:

[0102] The fault type classification branch has 64 input dimensions and 7 output dimensions; the corresponding categories include: power supply fault, power transistor short circuit fault, motor phase loss fault, Hall position detection fault, inter-turn short circuit fault, bearing wear fault, and rotor demagnetization fault.

[0103] The fault location classification branch has 64 input dimensions and 6 output dimensions, and the corresponding locations include: power transistors of the upper arm of phase A, power transistors of the lower arm of phase A, power transistors of the upper arm of phase B, power transistors of the lower arm of phase B, power transistors of the upper arm of phase C, and power transistors of the lower arm of phase C.

[0104] The output layer includes fault type output and fault location output, which are 7-dimensional probability vectors and 6-dimensional probability vectors, respectively. The category corresponding to the maximum probability value is the diagnostic result.

[0105] In the residual attention classification network, the calculation formula for the residual attention module includes:

[0106]

[0107] in, This represents the attention weight matrix, with the same dimensions as the input features and a value range of [0,1]. This represents the Sigmoid activation function. and This represents the weight matrix and bias vector of the attention layer. This indicates a global average pooling operation. This represents the feature map obtained after transformation by the feature encoder. Represents element-wise multiplication. This represents the feature map after the attention mechanism is added. and This represents the weights and biases of the residual layer. This indicates a modified linear unit activation function. This represents the output feature map of the residual attention module.

[0108] like Figure 4As shown, the residual attention module consists of three main parts connected sequentially: an attention weight generation path, a feature reweighting path, and a nonlinear transformation path. It receives the input feature map from the feature encoder. Output the feature map processed by the attention mechanism. ;

[0109] The attention weight generation pathway includes:

[0110] 1. Global average pooling, applied to the input feature map. Compression is performed along the spatial dimension to obtain a scalar value output vector for each channel. For each channel, the average value of the feature values ​​at all spatial locations of that channel is calculated.

[0111] 2. Bottleneck structure, consisting of two fully connected layers. The first fully connected layer compresses the number of channels to 1 / r of the original number, where the compression ratio r is a preset positive integer; the second fully connected layer restores the number of channels to the original dimension, and then passes them through the ReLU activation function.

[0112] 3. The Sigmoid activation function passes the output of the bottleneck structure through the Sigmoid function to obtain the weight value of each channel, which is between 0 and 1, forming the attention weight vector A.

[0113] Among them, the feature reweighting pathway includes:

[0114] 1. Combine the attention weight A with the original input feature map When performing channel-wise multiplication, important channel features are enhanced, while unimportant channel features are suppressed.

[0115] 2. Compare the weighted feature map with the original input feature map. Perform residual connections to obtain weighted feature maps. This ensures that the original information is not lost due to attention weighting, while enhancing important features;

[0116] The nonlinear transformation path includes: the weighted feature map Perform a linear transformation on a fully connected layer, then add a bias, and finally pass it through the ReLU activation function to obtain the final output. .

[0117] The structure of this residual attention module is suitable for brushless DC motor fault diagnosis tasks. It can automatically focus on the electrical characteristics most relevant to a specific fault, thereby improving diagnostic accuracy and robustness.

[0118] In step S5, calculating the fault severity index specifically includes the following steps:

[0119] The degree of health decline is calculated based on the difference between the baseline value of the comprehensive health index under normal conditions and the current comprehensive health index.

[0120] Based on the fault classification results, extract feature subsets related to specific fault types;

[0121] Calculate the relative deviation between the current fault-related characteristics and the corresponding characteristics under normal conditions;

[0122] The severity index of the failure is obtained by weighting the degree of decline in health status with the degree of relative deviation of characteristics.

[0123] The formula for calculating the fault severity index is as follows:

[0124]

[0125] in, This represents the severity index of the k-th type of fault, with a value ranging from [0,1]. A larger value indicates a more severe fault. This represents the overall health index under normal conditions, with a benchmark value typically set at 0.95. This indicates the current overall health index. This represents a subset of features associated with the k-th type of fault. This represents the feature subset associated with the k-th type of fault under normal conditions. This represents the weighting coefficient, used to balance the impact of the overall health index and specific failure characteristics, and is typically set to 0.7.

[0126] In step S6, the fault types include: power supply fault, power tube short circuit fault, motor phase loss fault, Hall position detection fault, inter-turn short circuit fault, bearing wear fault, and rotor demagnetization fault.

[0127] In step S6, the fault location specifically includes: the power transistor of the upper arm of phase A, the power transistor of the lower arm of phase A, the power transistor of the upper arm of phase B, the power transistor of the lower arm of phase B, the power transistor of the upper arm of phase C, and the power transistor of the lower arm of phase C.

[0128] like Figure 2 As shown, a brushless DC motor fault diagnosis system based on multi-source information fusion is implemented based on any one of the brushless DC motor fault diagnosis methods based on multi-source information fusion, characterized in that it includes:

[0129] The multi-source data acquisition module is used to acquire multi-source data during the operation of the brushless DC motor.

[0130] The preprocessing and feature extraction module is used to preprocess and extract features from the collected multi-source data;

[0131] The multimodal data fusion module is used to calculate the comprehensive health index by applying an adaptive weighted fusion algorithm of multimodal features;

[0132] The fault classification module is used for fault classification and localization using an improved residual attention classification network;

[0133] The severity assessment module is used to calculate the fault severity index based on the comprehensive health index and fault classification results;

[0134] The diagnostic results output module is used to output diagnostic results that include comprehensive health index, fault type, fault location, and fault severity index.

[0135] The multi-source data acquisition module includes a current detection unit, a voltage detection unit, a temperature detection unit, a position / velocity detection unit, and a control signal monitoring unit, and each unit adopts synchronous sampling technology.

[0136] The fault diagnosis system can be integrated into the brushless DC motor controller, with an embedded processor as the core. The multi-source data acquisition module is implemented through a high-precision ADC and digital input interface. The preprocessing and feature extraction module, multimodal data fusion module, fault classification module, severity assessment module and diagnostic result output module run on the embedded processor in the form of software programs, and meet the real-time requirement of a diagnostic cycle of less than 10 milliseconds.

[0137] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A fault diagnosis method for brushless DC motors based on multi-source information fusion, characterized in that, Includes the following steps: Step S1: Synchronously collect multi-source data during the operation of the brushless DC motor; Step S2: Preprocess the collected multi-source data and extract feature vectors composed of time-domain features and frequency-domain features from the preprocessed data; Step S3: Construct a multimodal feature adaptive weighted fusion algorithm to process the feature vector and calculate the comprehensive health index; Step S4: Construct an improved residual attention classification network, and input the feature vector into the network for fault classification and localization; Step S5: Calculate the fault severity index based on the comprehensive health index and fault classification results; Step S6: Output the diagnostic results, which include the comprehensive health index, fault type, fault location, and fault severity index.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: acquiring three-phase current and DC bus current through the current detection unit; acquiring three-phase voltage, DC bus voltage and power supply voltage through the voltage detection unit; acquiring motor stator temperature and controller power tube temperature through the temperature detection unit; acquiring Hall sensor signals and encoder signals through the position / speed detection unit; and acquiring PWM control signals and protection circuit status signals through the control signal monitoring unit.

3. The method according to claim 2, characterized in that, In step S2, the preprocessing includes filtering and denoising the acquired raw signal; The time-domain features include: mean, variance, kurtosis, skewness, root mean square value, waveform factor, and peak factor; the frequency-domain features include: obtaining the spectrum through fast Fourier transform, and extracting the spectral peak, frequency centroid, and mean square frequency.

4. The method according to claim 3, characterized in that, In step S3, the multimodal feature adaptive weighted fusion algorithm includes the following steps: The time-domain features and frequency-domain features in the feature vector are normalized respectively; Calculate the deviation of each feature, where the deviation represents the degree of difference between the current feature value and the preset normal state reference value; Based on the deviation of each feature and the preset standard deviation parameter, the exponential weight coefficient is calculated. The normalized features are weighted and summed, where the weights are determined by both the adaptive weight coefficient and the exponential weight coefficient. A control signal compensation term is introduced, and the control signal features are added to the fusion result after being processed by the evaluation function; The final output is the weighted sum of the normalized weighted eigenvalues ​​and the control signal compensation term, which serves as the comprehensive health index.

5. The method according to claim 4, characterized in that, In step S4, the improved residual attention classification network specifically includes a feature encoder, a residual attention module, and a classifier; The feature encoder adopts a three-layer fully connected network architecture. The first layer receives a 10-dimensional input feature vector and outputs a 64-dimensional feature. The second layer maps the 64-dimensional feature to 128-dimensional features. The third layer compresses the 128-dimensional feature to 64-dimensional features. The ReLU activation function is used between each layer and batch normalization is added. The residual attention module contains two cascaded attention blocks. Each attention block contains a global average pooling layer, a bottleneck structure consisting of two fully connected layers, a sigmoid activation function, a multiplier for feature weighting, a residual connector, and a nonlinear transformation layer containing a fully connected layer and a ReLU activation function. The classifier consists of two fully connected layers with parallel outputs. The first branch outputs a 7-dimensional vector corresponding to the probability distribution of seven fault types, and the second branch outputs a 6-dimensional vector corresponding to the fault probability distribution of six power transistor positions. Both branches use the Softmax activation function for probability normalization.

6. The method according to claim 5, characterized in that, In step S5, calculating the fault severity index specifically includes the following steps: The degree of health decline is calculated based on the difference between the baseline value of the comprehensive health index under normal conditions and the current comprehensive health index. Based on the fault classification results, extract feature subsets related to specific fault types; Calculate the relative deviation between the current fault-related characteristics and the corresponding characteristics under normal conditions; The severity index of the failure is obtained by weighting the degree of decline in health status with the degree of relative deviation of characteristics.

7. The method according to claim 6, characterized in that, In step S6, the fault types include: power supply fault, power tube short circuit fault, motor phase loss fault, Hall position detection fault, inter-turn short circuit fault, bearing wear fault, and rotor demagnetization fault.

8. The method according to claim 7, characterized in that, In step S6, the fault location specifically includes: the power transistor of the upper arm of phase A, the power transistor of the lower arm of phase A, the power transistor of the upper arm of phase B, the power transistor of the lower arm of phase B, the power transistor of the upper arm of phase C, and the power transistor of the lower arm of phase C.

9. A brushless DC motor fault diagnosis system based on multi-source information fusion, implemented based on the brushless DC motor fault diagnosis method based on multi-source information fusion as described in any one of claims 1-8, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source data during the operation of the brushless DC motor. The preprocessing and feature extraction module is used to preprocess and extract features from the collected multi-source data; The multimodal data fusion module is used to calculate the comprehensive health index by applying an adaptive weighted fusion algorithm of multimodal features; The fault classification module is used for fault classification and localization using an improved residual attention classification network; The severity assessment module is used to calculate the fault severity index based on the comprehensive health index and fault classification results; The diagnostic results output module is used to output diagnostic results that include comprehensive health index, fault type, fault location, and fault severity index.

10. The system according to claim 9, characterized in that, The multi-source data acquisition module includes a current detection unit, a voltage detection unit, a temperature detection unit, a position / velocity detection unit, and a control signal monitoring unit, and each unit adopts synchronous sampling technology.

Images