Intelligent diagnosis system for multiple faults of asynchronous motor
By combining sensor networks and RBP neural networks, especially by adding humidity and ultrasonic sensors and optimizing the learning rate algorithm, accurate diagnosis of various faults in asynchronous motors is achieved, solving the problem of low intelligence in existing technologies and improving diagnostic efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2024-01-22
- Publication Date
- 2026-05-12
AI Technical Summary
Existing motor fault diagnosis systems cannot meet the comprehensive diagnostic needs of multiple faults, have low intelligence, low accuracy, and poor user experience.
A combination of sensor networks and RBP neural networks is used. Signals are collected through M-type sensors, feature parameters are extracted by the signal processing module, and the fault diagnosis module uses the RBP neural network to determine the motor operating status and output the fault type. Humidity sensors and ultrasonic sensors are added to improve the accuracy of fault diagnosis, and the RBP neural network is trained through an optimized learning rate algorithm.
It enables comprehensive and accurate intelligent diagnosis of various faults in asynchronous motors, improving the efficiency and accuracy of fault diagnosis, enhancing user trust and satisfaction, and improving the output accuracy and interpretability of the RBP neural network.
Smart Images
Figure CN117906679B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to fields such as artificial intelligence systems in the production sector and industrial artificial intelligence network systems in the new generation of information technology, and in particular, designs an intelligent diagnostic system for multiple faults of asynchronous motors. Background Technology
[0002] With the widespread use of electric motors in various industrial applications, motor fault diagnosis has become crucial. Current fault diagnosis technologies often focus only on single fault types, failing to meet the comprehensive diagnostic needs for multiple faults. Furthermore, motor diagnostic systems capable of identifying multiple faults often suffer from low intelligence, low accuracy during operation, and a poor user experience. Summary of the Invention
[0003] (a) Technical problems to be solved
[0004] The present invention aims to at least partially solve one of the above-mentioned technical problems.
[0005] (II) Technical Solution
[0006] This invention provides an intelligent diagnostic system for multiple faults in asynchronous motors, comprising:
[0007] The sensor network includes: M types of sensors used to acquire induced signals during the operation of the asynchronous motor, where M ≥ 2;
[0008] The signal processing module, connected to the M-type sensor, is used to process the sensing signals collected by the M-type sensor and extract feature parameters.
[0009] The fault diagnosis module, connected to the signal processing module, is used to take the characteristic parameters output by the signal processing module as data input, and use a neural network model to judge the operating status of the asynchronous motor. When the motor is running normally, it outputs the normal state representation, and when a fault occurs, it outputs the fault type.
[0010] Among them, the fault type is selected from at least two of the following types: air gap eccentricity fault, bearing fault, rotor bar breakage fault, and insulation abnormality fault.
[0011] In some preferred embodiments of the present invention, the neural network model is an RBP neural network; the RBP neural network includes: an input layer for receiving feature parameters output by the signal processing module; T hidden layers for learning and storing the complex characteristics of the input feature parameters, where T≥2; and an output layer for using Softmax as the activation function to output the probability and confidence level of each fault type when a fault occurs, and to determine the fault type based on the probability and confidence level; wherein, the training process of the RBP neural network includes: forward propagation, where the input layer receives training data and outputs the prediction result through the calculation of each hidden layer neuron; and backpropagation, where the error between the prediction result and the actual result is calculated, and the result is determined based on the error and the learning rate. Adjust the weights and biases that represent the connections between neurons in adjacent layers of the neural network.
[0012] In some preferred embodiments of the present invention, during the backpropagation of the training process of the RBP neural network, the learning rate... The calculation is performed based on the modified first-order moment estimate and second-order moment estimate.
[0013] In some preferred embodiments of the present invention, during the backpropagation of the training process of the RBP neural network, the weights and biases are adjusted according to the following formula: ; ,in, Indicates the first Layer weight matrix; Indicates the first The bias vector of the layer; J is the loss function representing the error.
[0014] In some preferred embodiments of the present invention, during backpropagation, the learning rate Obtain it in the following ways:
[0015] Step A: Initialize hyperparameters
[0016] Initialize the following hyperparameters: the exponential decay rate of the first moment estimate. The exponential decay rate estimated by the second moment Initial learning rate The parameters satisfy the following conditions: as well as ;
[0017] Step B: Initialize variables
[0018] Initialize the following variables: time step ; Variables for first-moment estimation ; variables for second-order moment estimation Mean learning rate
[0019] Step C: Iterative Update
[0020] For each time step :
[0021] Sub-step C1: Calculate the loss function Regarding model parameters gradient: ;
[0022] Sub-step C2: Update the first-order moment estimate and second-order moment estimation ;
[0023] - Update first-order moment estimate: ;
[0024] - Update the second moment estimate: ;
[0025] - Corrections are made to the first-order moment estimates and second-order moment estimates:
[0026] - Corrected first-order moment estimate: ;
[0027] - Corrected second-order moment estimate: ;
[0028] Sub-step C3: Calculate the current adaptive learning rate : ;
[0029] Sub-step C4: When t≥2, update the mean learning rate. : ;
[0030] Sub-step C5: Calculate the mean learning rate and adaptive learning rate The smaller value in the range is used as the adaptive mean learning rate: ;
[0031] Sub-step C6: Update parameters : ;
[0032] Sub-step C7: Update time step : ;
[0033] Sub-step C8: Check termination conditions
[0034] - Check if the RBP neural network has converged or reached the predetermined number of iterations. If one of these conditions is met, the iteration ends and the learning rate is output. =L min If the condition is not met, execute sub-step C1 to continue iterating, where the number of iterations is greater than or equal to 50.
[0035] In some preferred embodiments of the present invention, in step A, =0.9; =0.999; =0.001.
[0036] In some preferred embodiments of the present invention, in sub-step C8, the number of iterations is equal to 100.
[0037] In some preferred embodiments of the present invention, during backpropagation, the cross-entropy loss function is used to characterize the error between the predicted result and the actual result output by the neural network:
[0038]
[0039] in, This refers to the number of neurons in the output layer. The label value representing the actual type i is the expected output, i.e., the actual result. This represents the neural network's predicted output for type i.
[0040] In some preferred embodiments of the present invention, T=4; of the four hidden layers: the first hidden layer includes 18 neurons; the second hidden layer includes 14 neurons; the third hidden layer includes 10 neurons; and the fourth hidden layer includes 5 neurons.
[0041] In some preferred embodiments of the present invention, the fault diagnosis module outputs the probability of each fault type according to the following formula: ,in, It is the input to the output layer, and N is the number of fault types. It is the probability of the i-th type of fault.
[0042] In some preferred embodiments of the present invention, the M-type sensor includes: a humidity sensor for acquiring humidity signals within the asynchronous motor.
[0043] In some preferred embodiments of the present invention, T1 ultrasonic sensors are used to collect ultrasonic components of different frequencies in the noise of the asynchronous motor, where T1≥1.
[0044] In some preferred embodiments of the present invention, T1 ultrasonic sensors include: a low-frequency ultrasonic sensor, a medium-frequency ultrasonic sensor, and a high-frequency ultrasonic sensor, which are used to collect the low-frequency component, medium-frequency component, and high-frequency component of the noise generated by the asynchronous motor, respectively.
[0045] In some preferred embodiments of the present invention, the M-type sensor further includes: a displacement sensor, comprising three displacement sensors, respectively used to collect displacement signals of the asynchronous motor in the X, Y, and Z directions; a vibration sensor, comprising three vibration sensors, respectively used to collect vibration signals of the asynchronous motor in the X, Y, and Z directions; a current sensor, comprising three current sensors, respectively connected to the three-phase power supply lines of the asynchronous motor, used to collect current signals of the three-phase power supply; and an insulation resistance sensor, comprising three insulation resistance sensors, respectively disposed at three different insulation positions within the asynchronous motor.
[0046] In some preferred embodiments of the present invention, the feature parameters input to the input layer of the RBP neural network are:
[0047]
[0048] in,
[0049] (1) The three displacement characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the displacement time-domain data collected by three displacement sensors through a signal processing module;
[0050] (2) The vibration signals collected by three vibration sensors in the X, Y, and Z directions are filtered, normalized, and converted from analog to digital by the signal processing module to obtain three vibration characteristic parameters.
[0051] (3) , , The three ultrasonic feature parameters are obtained by feature extraction of the low-frequency, mid-frequency, and high-frequency components of the noise generated by the asynchronous motor collected by three ultrasonic sensors through a signal processing module.
[0052] (4) The humidity characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the humidity signal collected by the humidity sensor through the signal processing module;
[0053] (5) The current characteristic parameters are obtained after the current signals collected by three current sensors are filtered, normalized, and converted from analog to digital by the signal processing module.
[0054] (6) The comparison value is obtained by performing analog-to-digital conversion and comparison operations between the resistance values collected by the three insulation resistance sensors and the corresponding threshold impedance within the signal processing module.
[0055] (III) Beneficial Effects
[0056] As can be seen from the above technical solution, the present invention has at least one of the following beneficial effects compared to the prior art:
[0057] (1) This embodiment achieves comprehensive and accurate intelligent diagnosis of various faults of asynchronous motors through the organic combination of this set of sensors, signal processing and fault diagnosis modules, thereby improving the efficiency and accuracy of fault diagnosis and enhancing user trust and satisfaction.
[0058] (2) The output layer of the RBP neural network adopts the Softmax and confidence interval method to output the probability distribution of each fault type, which facilitates fault identification and classification. This enables the RBP neural network to clearly and reliably feed back the fault information of the motor under high current conditions, improves the output accuracy and interpretability of the RBP neural network, helps users judge whether the output of the neural network is credible and whether further verification or adjustment is needed, and improves user trust and satisfaction.
[0059] (3) In the process of realizing this invention, the applicant discovered that the humidity inside the motor is positively correlated with certain motor faults. Therefore, a humidity sensor was added to the sensor network in this embodiment, and it was used as one of the characteristic parameters to determine the fault type, thereby improving the accuracy of fault determination.
[0060] (4) In the process of realizing this invention, the applicant discovered that more fault information, especially bearing faults, is hidden in the ultrasonic frequency band of motor noise, but ordinary vibration sensors cannot identify this information. Therefore, ultrasonic sensors were added to the sensor network to extract fault-related information from motor noise, thereby improving the accuracy of fault diagnosis.
[0061] (5) A new variable, mean learning rate, was introduced when calculating the mean learning rate. The average learning rate is used to average the learning rate over all past time steps. This average value adaptively adjusts its size as the current learning rate changes. Therefore, when encountering extreme learning rates, the introduction of the average learning rate can smooth the learning rate, mitigating the adverse effects of extremely high learning rates on training and thus reducing the risk of convergence failure. Even without extreme learning rates, if the final chosen learning rate is the average learning rate instead of the current learning rate, this will not have a significant negative impact on model training. This ensures that the learning rate remains adaptive.
[0062] (6) In this invention, the mean learning rate is used. and adaptive learning rate Obtain the adaptive mean learning rate This adaptive mean learning rate, by limiting and adjusting the learning rate, effectively improves the model's training performance. Experimental results further confirm the positive impact of this algorithm on model training under extreme learning rate conditions. Attached Figure Description
[0063] Figure 1 This is a schematic diagram of the structure of the intelligent diagnostic system for multiple faults in asynchronous motors according to an embodiment of the present invention.
[0064] Figure 2 for Figure 1 The diagram shows the RBP neural network of the fault diagnosis module in the intelligent fault diagnosis system for asynchronous motors. Detailed Implementation
[0065] This invention relates to an intelligent diagnostic system for multiple faults in asynchronous motors. By combining a sensor network, an information processing module, and a neural network model, it achieves accurate and rapid diagnosis of various motor faults, improving the reliability and safety of motor operation. Furthermore, by optimizing the learning rate algorithm, the performance and convergence speed of the algorithm are improved, avoiding the impact of extreme learning rates on neural network training.
[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0067] In one exemplary embodiment of the present invention, an intelligent diagnostic system for multiple faults in an asynchronous motor is provided. Figure 1 This is a schematic diagram of the structure of the intelligent diagnostic system for multiple faults in an asynchronous motor according to an embodiment of the present invention. Figure 1 As shown, the asynchronous motor multi-fault intelligent diagnosis system in this embodiment includes:
[0068] The sensor network includes: M types of sensors used to acquire induced signals during the operation of the asynchronous motor, where M ≥ 2;
[0069] The signal processing module, connected to the M-type sensor, is used to process the sensing signals collected by the M-type sensor and extract feature parameters.
[0070] The fault diagnosis module, connected to the signal processing module, is used to take the characteristic parameters output by the signal processing module as data input, and use a neural network model to judge the operating status of the asynchronous motor. When the motor is running normally, it outputs the normal state representation, and when a fault occurs, it outputs the fault type.
[0071] Among them, the fault type is selected from at least two of the following types: air gap eccentricity fault, bearing fault, rotor bar breakage fault, and insulation abnormality fault.
[0072] The following provides a detailed description of each component of the asynchronous motor multi-fault intelligent diagnosis system in this embodiment.
[0073] In this embodiment, the fault types include four types: air gap eccentricity fault, bearing fault, rotor bar breakage fault, and insulation abnormality fault. However, those skilled in the art should understand that in other embodiments of the present invention, there may be more than four fault types, or it may include other fault types besides the four types mentioned above, or only two or three of the four types mentioned above. As long as it includes two of the fault types mentioned above, the present invention can be applied and should be within the protection scope of the present invention.
[0074] Please continue to refer to Figure 1 In this embodiment of the intelligent diagnostic system for multiple faults in asynchronous motors, the sensor network includes six types of sensors: humidity sensors, ultrasonic sensors, displacement sensors, vibration sensors, current sensors, and insulation resistance sensors.
[0075] In the process of developing this invention, the applicant discovered a positive correlation between humidity inside the motor and certain motor faults. Therefore, a humidity sensor was added to the sensor network in this embodiment, and used as one of the characteristic parameters for fault type determination, thus improving the accuracy of fault diagnosis. Specifically, in this embodiment, the sensor network includes a humidity sensor for collecting humidity signals within the asynchronous motor.
[0076] The applicant discovered that more fault information, especially bearing faults, is contained in the ultrasonic frequency band of motor noise, but ordinary vibration sensors cannot identify this information. Therefore, ultrasonic sensors were added to the sensor network in this embodiment to extract fault-related information from motor noise.
[0077] Specifically, in this embodiment, the sensor network includes ultrasonic sensors installed near the motor bearings. These ultrasonic sensors include low-frequency ultrasonic sensors, mid-frequency ultrasonic sensors, and high-frequency ultrasonic sensors, which are used to collect the low-frequency, mid-frequency, and high-frequency components of the noise generated by the asynchronous motor, respectively.
[0078] Furthermore, in this embodiment, the low-frequency ultrasonic sensor is used to extract features of the low-frequency components in the motor noise with a frequency range of 50-200Hz; the medium-frequency ultrasonic sensor is used to extract features of the medium-frequency components in the motor noise with a frequency range of 500-2 kHz; and the high-frequency ultrasonic sensor is used to extract features of the high-frequency components in the motor noise with a frequency range of 5-10 kHz.
[0079] In this embodiment, displacement sensors are arranged at multiple points inside the motor to measure the displacement of the motor rotor in real time. The displacement data generated by these sensors is used to monitor the air gap eccentricity of the rotor. Specifically, the displacement sensors include three displacement sensors, which are used to collect displacement signals of the asynchronous motor in the X, Y, and Z directions, respectively.
[0080] In this embodiment, ultrasonic sensors and vibration sensors are installed near the motor bearing to monitor the bearing's vibration and sound. These two types of sensors can sense the bearing's operating status, including normal operation and abnormal vibration or noise. Specifically, the vibration sensors include three vibration sensors used to collect vibration signals from the asynchronous motor in the X, Y, and Z directions, respectively.
[0081] In this embodiment, current sensors are installed on the motor's power supply line to monitor the motor's current waveform. By analyzing the waveform characteristics of the current, faults such as broken bars in the rotor can be detected. Specifically, the current sensors include three current sensors, each connected to the three-phase power supply line of the asynchronous motor, for collecting the current signals from the three-phase power supply.
[0082] In this embodiment, insulation resistance sensors are installed on the insulation part of the motor to monitor changes in insulation resistance. This helps to detect insulation abnormalities early and prevent electrical faults. Specifically, the insulation resistance sensors include three insulation resistance sensors, each installed at three different insulation locations within the asynchronous motor.
[0083] Of course, in other embodiments of the present invention, there may be other types of sensors, or only some of the sensors in this embodiment may be selected, or the number of a certain sensor may be different from that in this embodiment. All of these can also achieve the present invention and are also within the protection scope of the present invention.
[0084] Please refer to Figure 1 In this embodiment, the signals obtained from the sensor network are all analog signals, as shown by the dotted lines in the figure. The signal processing module needs to process these analog signals, performing operations such as filtering, normalization, and analog-to-digital conversion, as well as some logical operations. After signal processing, the signal output from the signal processing module to the fault diagnosis module, i.e., the characteristic parameters, is a digital signal, as shown by the dashed lines in the figure.
[0085] In practical scenarios, the signal processing module can be divided into different signal processing units according to the type of signal. Specifically, in this embodiment, the signal processing module includes: a displacement signal processing unit, a vibration signal processing unit, an ultrasonic signal processing unit, a humidity signal processing unit, a current signal processing unit, and an insulation signal processing unit.
[0086] The displacement signal processing unit is connected to three displacement sensors and is used to filter, normalize, and perform analog-to-digital conversion on the displacement time-domain data collected by the three displacement sensors to obtain three displacement characteristic parameters. ).
[0087] The vibration signal processing unit is connected to three vibration sensors and is used to filter, normalize, and perform analog-to-digital conversion on the vibration signals collected by the three vibration sensors in the X, Y, and Z directions, respectively, to obtain three vibration characteristic parameters. ).
[0088] The ultrasonic signal processing unit is connected to three ultrasonic sensors and is used to extract features from the low-frequency, mid-frequency, and high-frequency components of the noise generated by the asynchronous motor collected by the three ultrasonic sensors, resulting in three ultrasonic feature parameters. , , ).
[0089] Specifically, the ultrasonic signal processing unit preprocesses the acquired ultrasonic signal, including removing environmental noise, filtering, and downsampling to reduce noise and improve signal quality. Further, the ultrasonic signal processing unit converts the ultrasonic signal to the time domain for analysis. In the time domain, it analyzes the waveform, amplitude, frequency, and other characteristics of the ultrasonic signal. Then, it converts the ultrasonic signal to the frequency domain by applying Fourier transform, providing spectral information of the ultrasonic signal, including frequency components and spectral density. Finally, it extracts feature parameters: ultrasonic feature parameters related to motor faults are extracted from the time-domain or frequency-domain analysis. , , ).
[0090] The humidity signal processing unit is connected to the humidity sensor and is used to filter, normalize and convert the humidity signal collected by the humidity sensor to obtain the humidity characteristic parameters (S).
[0091] The current signal processing unit is connected to three current sensors and is used to filter, normalize, and perform analog-to-digital conversion on the current signals collected by the three current sensors to obtain current characteristic parameters. ).
[0092] An insulation signal processing unit and three insulation resistance sensors are used to perform analog-to-digital conversion on the resistance values collected by the three insulation resistance sensors and compare them with the corresponding threshold impedance to calculate the comparison value. ).
[0093] In this embodiment, the neural network in the fault diagnosis module is an RBP neural network. The input data for this RBP neural network is the output data of each signal processing unit in the signal processing module. During the training phase, it undergoes ensemble learning, optimizing the network weights through the backpropagation algorithm. By training the RBP neural network, a comprehensive fault identification model is established, enabling rapid diagnosis of various faults.
[0094] Figure 2 for Figure 1 The diagram shows a schematic of the RBP neural network in the fault diagnosis module of the intelligent multi-fault diagnosis system for asynchronous motors. Figure 2 As shown, in this embodiment, the RBP neural network includes:
[0095] The input layer is used to receive the characteristic parameters output by M signal processing units;
[0096] T hidden layers are used to learn and store the complex characteristics of the input feature parameters, where T≥2;
[0097] The output layer uses Softmax as the activation function to output the probability and confidence level of each fault type, and determines the fault type based on the probability and confidence level.
[0098] The training process of the RBP neural network includes: forward propagation, where the input layer receives training data and outputs the prediction result through the calculation of each hidden layer neuron; and backpropagation, where the error between the prediction result and the actual result is calculated, and the weights and biases representing the connections between adjacent layer neurons in the neural network are adjusted according to the error and the learning rate.
[0099] In this embodiment, the output layer of the RBP neural network adopts the Softmax and confidence interval method to output the probability distribution of each fault type, which facilitates fault identification and classification. This enables the RBP neural network to clearly and reliably feed back fault information of motor operation under high current conditions, improves the output accuracy and interpretability of the RBP neural network, helps users judge whether the output of the neural network is reliable and whether further verification or adjustment is needed, and improves user trust and satisfaction.
[0100] Corresponding to the output of the signal processing module, in this embodiment, the feature parameters input to the input layer of the RBP neural network are:
[0101]
[0102] in,
[0103] (1) The three displacement characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the displacement time-domain data collected by three displacement sensors through a signal processing module;
[0104] (2) The vibration signals collected by three vibration sensors in the X, Y, and Z directions are filtered, normalized, and converted from analog to digital by the signal processing module to obtain three vibration characteristic parameters.
[0105] (3) , , The three ultrasonic feature parameters are obtained by feature extraction of the low-frequency, mid-frequency, and high-frequency components of the noise generated by the asynchronous motor collected by three ultrasonic sensors through a signal processing module.
[0106] (4) The humidity characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the humidity signal collected by the humidity sensor through the signal processing module;
[0107] (5) The current characteristic parameters are obtained after the current signals collected by three current sensors are filtered, normalized, and converted from analog to digital by the signal processing module.
[0108] (6) The comparison value is obtained by performing analog-to-digital conversion and comparison operations between the resistance values collected by the three insulation resistance sensors and the corresponding threshold impedance within the signal processing module.
[0109] It can be seen that the constructed feature parameter vector contains multi-source data from different types of sensors, providing comprehensive information for the RBP neural network and helping to comprehensively identify various motor faults.
[0110] Please continue to refer to Figure 2 In this embodiment, the RBP neural network includes four hidden layers. The first hidden layer includes 18 neurons; the second hidden layer includes 14 neurons; the third hidden layer includes 10 neurons; and the fourth hidden layer includes 5 neurons.
[0111] The following explains the forward propagation process of training data or real data through the four hidden layers.
[0112] ①Hidden layer 1:
[0113] - Content: Contains 18 neurons for learning complex features of the input data.
[0114] - Function: To perform linear transformations and activation function processing on the input data to extract its nonlinear features.
[0115] - Formula:
[0116]
[0117] in, These are the activation values of neurons in hidden layer 1. It is the activation function Softmax. It is the connection weight. It is the input data. It is a bias.
[0118] ②Hidden layer 2:
[0119] - Contents: Contains 14 neurons for learning the feature representation of hidden layer 1.
[0120] - Function: Performs further nonlinear transformations on the output of hidden layer 1 to extract higher-level features.
[0121] - Formula:
[0122] in, These are the activation values of neurons in hidden layer 2. It is the connection weight. This is the output of hidden layer 1. It is a bias.
[0123] ③Hidden layer 3:
[0124] - Content: Contains 10 neurons for learning high-level feature representations of hidden layer 2.
[0125] - Function: Performs a deeper nonlinear transformation on the output of hidden layer 2 to extract more abstract features.
[0126] - Formula:
[0127] in, These are the activation values of neurons in hidden layer 3. It is the connection weight. It is the output of hidden layer 2. It is a bias.
[0128] ④Hidden layer 4:
[0129] - Content: Contains 5 neurons for learning higher-level feature representations of hidden layer 3.
[0130] - Function: Performs a deeper nonlinear transformation on the output of hidden layer 3 to further extract abstract features.
[0131] - Formula:
[0132] in, These are the activation values of neurons in hidden layer 4. It is the connection weight. It is the output of hidden layer 3. It is a bias.
[0133] The design of the aforementioned hidden layer aims to learn the complex features of the input data through the RBP neural network, thereby achieving comprehensive identification of various types of motor faults.
[0134] In this embodiment, the fault diagnosis module outputs the probability of each type of fault:
[0135] in:
[0136] (1) The number of neurons in the output layer is set to 5 here, corresponding to 5 states [normal, four types of faults]:
[0137] (2) This is the predicted probability of a sample belonging to type i from the output layer of the neural network. This is the output value of the neural network for each type, typically normalized using the Softmax function to represent the probability distribution.
[0138] (3) It is the 4th hidden layer The activation value of each neuron is calculated using its weights and the input:
[0139]
[0140] in, It is the 3rd hidden layer The activation value of each neuron, and so on.
[0141] In this process, the activation values of the hidden layers are calculated using weights and inputs, and then used to calculate the probabilities of the output layer. As mentioned above, during training, the training data passes through four hidden layers and the output layer to obtain the prediction results. Then, in backpropagation, it is necessary to calculate the error between the prediction results and the actual results, and adjust the weights and biases representing the connections between neurons in the neural network based on the error and the learning rate.
[0142] In this embodiment, the cross-entropy loss function is used to represent the error between the predicted result and the actual result output by the neural network:
[0143]
[0144] in, This refers to the number of neurons in the output layer. The label value representing the actual type i is the expected output; This represents the predicted output value of the neural network for type i.
[0145] In this embodiment, by using the cross-entropy loss function, the system aims to minimize the difference between the actual output and the neural network's predicted output, thereby improving the model's accuracy. During backpropagation, optimization algorithms such as gradient descent are used to adjust the neural network parameters to minimize this loss function.
[0146] In this embodiment, the RBP neural network is trained using the backpropagation algorithm. During training, the concept of an adaptive mean learning rate is introduced. Based on the Adam training optimization algorithm, a new variable, the mean learning rate, is added. This improves the network's adaptability to initial weights and local minima, making the network more robust and helping to address multiple faults in complex systems.
[0147] Specifically, in this embodiment, during backpropagation, the learning rate... The learning rate is calculated based on the modified first-moment and second-moment estimates, as well as the mean learning rate. Specifically, during backpropagation, the learning rate... Obtain it in the following ways:
[0148] Step A: Initialize hyperparameters
[0149] Initialize the following hyperparameters: the exponential decay rate of the first moment estimate. The exponential decay rate estimated by the second moment Initial learning rate The parameters satisfy the following conditions: as well as ;
[0150] Step B: Initialize variables
[0151] Initialize the following variables: time step ; Variables for first-moment estimation ; variables for second-order moment estimation Mean learning rate
[0152] Step C: Iterative Update
[0153] For each time step :
[0154] Sub-step C1: Perform backpropagation on the training data and calculate the loss function. Regarding model parameters gradient: ;
[0155] Sub-step C2: Update the first-order moment estimate and second-order moment estimation ;
[0156] - Update first-order moment estimate: ;
[0157] - Update the second moment estimate: ;
[0158] - Corrections are made to the first-order moment estimates and second-order moment estimates:
[0159] - Corrected first-order moment estimate: ;
[0160] - Corrected second-order moment estimate: ;
[0161] Sub-step C3: Calculate the current adaptive learning rate : ;
[0162] Sub-step C4: When t≥2, update the mean learning rate. : ;
[0163] Sub-step C5: Calculate the mean learning rate and adaptive learning rate The smaller value in the range is used as the adaptive mean learning rate: ;
[0164] Sub-step C6: Update parameters : ;
[0165] Sub-step C7: Update time step : ;
[0166] Sub-step C8: Check termination conditions
[0167] - Check if the RBP neural network has converged or reached the predetermined number of iterations. If one of these conditions is met, the iteration ends and the learning rate is output. =L min If the condition is not met, execute sub-step C1 to continue iterating, where the number of iterations is greater than or equal to 50.
[0168] In this embodiment, in step A, =0.9; =0.999; =0.001; In substep C8, the number of iterations is equal to 100. Those skilled in the art should understand that the specific values of the hyperparameters can also be adjusted as needed.
[0169] The characteristics of the learning rate used in this embodiment are explained below:
[0170] (1) Learning rate Calculations are performed based on the modified first-order moment estimates and second-order moment estimates.
[0171] (2) A new variable was introduced when calculating the mean learning rate. The average learning rate is used to average the learning rate over all past time steps. This average value adaptively adjusts its size as the current learning rate changes. Therefore, when encountering extreme learning rates, the introduction of the average learning rate can smooth the learning rate, mitigating the adverse effects of extremely high learning rates on training and thus reducing the risk of convergence failure. Even without extreme learning rates, if the final chosen learning rate is the average learning rate instead of the current learning rate, this will not have a significant negative impact on model training. This ensures that the learning rate remains adaptive.
[0172] (3) In this invention, the mean learning rate is used. and adaptive learning rate Obtain the adaptive mean learning rate This adaptive mean learning rate, by limiting and adjusting the learning rate, effectively improves the model's training performance. Experimental results further confirm the positive impact of this algorithm on model training under extreme learning rate conditions.
[0173] During the training of an RBP neural network, after obtaining the learning rate, the weights and biases representing neurons in adjacent layers can be adjusted according to this learning rate. Specifically, the weights and biases are adjusted according to the following formula:
[0174]
[0175]
[0176] in, Indicates the first Layer weight matrix; Indicates the first The bias vector of the layer; J is the loss function representing the error, and is related to the previous layer. same.
[0177] The RBP application network in this embodiment repeats the forward propagation, loss calculation, backpropagation, and weight update processes as described above until convergence is achieved. Through these steps, the RBP neural network can fully utilize the output information of various sensors in subsequent real-world scenarios to accurately identify multiple types of motor faults.
[0178] In summary, this embodiment, through the organic combination of this complete set of sensors, signal processing, and fault diagnosis modules, achieves comprehensive and accurate intelligent diagnosis of various faults in asynchronous motors, improving the efficiency and accuracy of fault diagnosis, and enhancing user trust and satisfaction.
[0179] This concludes the description of embodiments of the intelligent diagnostic system for multiple faults in asynchronous motors according to the present invention. Based on the above description, those skilled in the art should have a clear understanding of the present invention.
[0180] It should be noted that, unless explicitly stated otherwise, the numerical parameters in the specification and claims of this invention may be approximate values and can be changed according to the content of this invention. Specifically, all figures in the specification and claims indicating the content of composition, reaction conditions, etc., should be understood to be modified by the term "about" in all cases, which means that they include a specific quantity varying by ±10% in some embodiments.
[0181] This invention can be implemented using hardware comprising several different components and a suitably programmed computer. Various component embodiments of the invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. The physical implementation of the hardware structure includes, but is not limited to, physical devices, including, but not limited to, transistors, memristors, DNA computers, microcontrollers, microprocessors, or digital signal processors (DSPs).
[0182] The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such a program implementing the invention can be stored on a computer-readable medium or can take the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.
[0183] The algorithms and displays provided herein are not related to any particular computer, virtual system, or other inherent device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing this system is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of this invention can be implemented using various programming languages; the description of specific languages herein is for the purpose of disclosing the best mode of implementation of the invention.
[0184] Those skilled in the art will understand that in the claims and specification of this invention, the word "comprising" does not exclude the presence of elements (or steps) not listed in the claims. The word "a" or "an" preceding an element (or step) does not exclude the presence of a plurality of such elements (or steps).
[0185] For certain implementations, if they are not key aspects of the present invention and are well-known to those skilled in the art, they have not been described in detail in the accompanying drawings or text due to space limitations. In such cases, reference can be made to relevant prior art for understanding. Furthermore, the purpose of providing the above embodiments is merely to ensure that the present invention meets legal requirements. The present invention can be implemented in many different forms and should not be construed as limited to the embodiments described herein. Moreover, the above definitions of elements and methods are not limited to the various specific structures, shapes, or methods mentioned in the embodiments, and those skilled in the art can make simple modifications or substitutions.
[0186] Similarly, it should be understood that, for the sake of brevity, in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof. However, this method of invention should not be construed as reflecting an intention that the claimed invention requires more features than expressly recited in each claim. Rather, as reflected in the claims, the various inventive aspects consist of fewer than all the features of the preceding single embodiment. Furthermore, embodiments may be used in combination with each other or with other embodiments based on design and reliability considerations; that is, technical features from different embodiments can be freely combined to form more embodiments. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.
[0187] The above specific embodiments have provided a detailed description of the purpose, technical means, and beneficial effects of the present invention. It should be understood that the purpose of the detailed description is to enable those skilled in the art to better understand the present invention, and it is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An intelligent diagnostic system for multiple faults in an asynchronous motor, characterized in that, include: The sensor network includes: M types of sensors used to collect induced signals during the operation of the asynchronous motor, where M ≥ 2; The signal processing module is connected to the M-type sensor and is used to process the sensing signals collected by the M-type sensor and extract feature parameters. The fault diagnosis module is connected to the signal processing module and is used to take the feature parameters output by the signal processing module as data input, judge the operating status of the asynchronous motor through a neural network model, output the normal state representation when the motor is running normally, and output the fault type when a fault occurs. The fault type is selected from at least two of the following types: air gap eccentricity fault, bearing fault, rotor bar breakage fault, and insulation abnormality fault. The neural network model is an RBP neural network. The RBP neural network includes: an input layer for receiving feature parameters output by the signal processing module; T hidden layers for learning and storing the complex characteristics of the input feature parameters, where T ≥ 2; and an output layer that uses Softmax as the activation function to output the probability and confidence level of each fault type when a fault occurs, and determines the fault type based on this probability and confidence level. The training process of the RBP neural network includes: forward propagation, where the input layer receives training data and outputs prediction results through the calculations of neurons in each hidden layer; and backpropagation, where the error between the prediction result and the actual result is calculated, and the result is determined based on the error and the learning rate. Adjust the weights and biases in the neural network that represent the connections between neurons in adjacent layers; In the backpropagation process of the RBP neural network training, the learning rate... The calculations are performed based on the modified first-order moment estimates and second-order moment estimates; the weights and biases are adjusted according to the following formula: ; ,in, Indicates the first Layer weight matrix; Indicates the first The bias vector of the layer; J is the loss function representing the error; The M-type sensors include: a humidity sensor for collecting humidity signals within the asynchronous motor; and T1 ultrasonic sensors for collecting ultrasonic components of different frequencies in the noise of the asynchronous motor, where T1 ≥ 1.
2. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 1, characterized in that, During backpropagation, the learning rate Obtain it in the following ways: Step A: Initialize hyperparameters Initialize the following hyperparameters: the exponential decay rate of the first moment estimate. The exponential decay rate estimated by the second moment Initial learning rate The parameters satisfy the following conditions: as well as ; Step B: Initialize variables Initialize the following variables: time step ; Variables for first-moment estimation ; variables for second-order moment estimation Mean learning rate Step C: Iterative Update For each time step : Sub-step C1: Calculate the loss function Regarding model parameters gradient: ; Sub-step C2: Update the first-order moment estimate and second-order moment estimation ; - Update first-order moment estimate: ; - Update the second-order moment estimate: ; - Corrections are made to the first-order moment estimates and second-order moment estimates: - Corrected first-order moment estimate: ; - Corrected second-order moment estimate: ; Sub-step C3: Calculate the current adaptive learning rate : ; Sub-step C4: When t≥2, update the mean learning rate. : ; Sub-step C5: Calculate the mean learning rate and adaptive learning rate The smaller value in the range is used as the adaptive mean learning rate: ; Sub-step C6: Update parameters : ; Sub-step C7: Update time step : ; Sub-step C8: Check termination conditions - Check if the RBP neural network has converged or reached the predetermined number of iterations. If one of these conditions is met, the iteration ends and the learning rate is output. =L min If the condition is not met, execute sub-step C1 to continue iterating, where the number of iterations is greater than or equal to 50.
3. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 2, characterized in that, In step A =0.9; =0.999; =0.001; and / or In sub-step C8, the number of iterations is 100.
4. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 1, characterized in that, In backpropagation, the cross-entropy loss function is used to characterize the error between the predicted result and the actual result of the neural network output: in, This refers to the number of neurons in the output layer. The label value representing the actual type i is the expected output, i.e., the actual result. This represents the neural network's predicted output for type i.
5. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 1, characterized in that, T=4; 4 hidden layers: the first hidden layer contains 18 neurons; the second hidden layer contains 14 neurons; the third hidden layer contains 10 neurons; the fourth hidden layer contains 5 neurons; and / or The fault diagnosis module outputs the probability of each fault type according to the following formula: ,in, It is the input to the output layer, and N is the number of fault types. It is the probability of the i-th type of fault.
6. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 1, characterized in that, The T1 ultrasonic sensors include: a low-frequency ultrasonic sensor, a medium-frequency ultrasonic sensor, and a high-frequency ultrasonic sensor, used to collect the low-frequency, medium-frequency, and high-frequency components of the noise generated by the asynchronous motor, respectively; and / or The M-type sensor also includes: Displacement sensors include: three displacement sensors, which are used to collect displacement signals of the asynchronous motor in the X, Y, and Z directions, respectively; Vibration sensors include: 3 vibration sensors, used to collect vibration signals of the asynchronous motor in the X, Y, and Z directions respectively; Current sensors include: three current sensors, each connected to the three-phase power supply line of the asynchronous motor, used to collect the current signal of the three-phase power supply; Insulation resistance sensors include: 3 insulation resistance sensors, which are respectively set at three different insulation positions inside the asynchronous motor.
7. The intelligent diagnostic system for multiple faults in asynchronous motors according to claim 6, characterized in that, The feature parameters input to the input layer of the RBP neural network are: in, (1) The three displacement characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the displacement time-domain data collected by three displacement sensors through a signal processing module. (2) The vibration signals collected by three vibration sensors in the X, Y, and Z directions are filtered, normalized, and converted from analog to digital by the signal processing module to obtain three vibration characteristic parameters. (3) , , The three ultrasonic feature parameters are obtained by feature extraction of the low-frequency, mid-frequency, and high-frequency components of the noise generated by the asynchronous motor collected by three ultrasonic sensors through a signal processing module. (4) The humidity characteristic parameters are obtained by filtering, normalizing and analog-to-digital conversion of the humidity signal collected by the humidity sensor through the signal processing module; (5) The current characteristic parameters are obtained after the current signals collected by three current sensors are filtered, normalized, and converted from analog to digital by the signal processing module. (6) The comparison value is obtained by performing analog-to-digital conversion and comparison operations between the resistance values collected by the three insulation resistance sensors and the corresponding threshold impedance within the signal processing module.