Magnetizer for motor
A machine learning system optimizes the magnetization process by adjusting voltage commands based on real-time observations, addressing temperature and resistance issues to maintain consistent magnetization rates and improve motor torque.
Patent Information
- Application Number
- DE102016011524
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2015-09-30
- Filing Date
- 2016-09-23
- Publication Date
- 2024-11-14
- Estimated Expiration
- 2036-09-23
AI Technical Summary
Existing magnetization units for motors face issues with reduced magnetization rate and torque due to increasing temperature and resistance in the magnetizing windings, leading to inconsistent production quality and motor performance.
A machine learning system that includes a state observer, reward calculator, and learning unit to adjust the voltage command based on real-time observations of winding temperature, resistance, and magnetic flux density, using methods like Q-learning and neural networks to optimize the magnetization process.
The system maintains consistent magnetization rates and improves motor torque by dynamically adjusting the magnetization process, enhancing production quality and durability.
Smart Images

Figure 00000000_0000_ABST
Abstract
Description
BACKGROUND OF THE INVENTION 1. Field of the invention
[0001] The present invention relates to a magnetizer for a motor, and more particularly, to a magnetizer for manufacturing a rotor for a motor having the function of determining a voltage command for flowing an optimum current to obtain a desired magnetization rate. 2. Description of the state of the art
[0002] To magnetize permanent magnets, it is necessary to apply a magnetic field whose strength reaches the saturation point of the maximum magnetic flux density of a magnetic material. Magnetizing units are used for this purpose. The magnetizing units include a magnetizing yoke and windings for generating magnetic fields to impart magnetic properties to a magnetic material. As an example of the magnetizing unit, a capacitor discharge magnetizing unit is known, in which a capacitor is charged and energy accumulated in the capacitor is discharged into the magnetizing yoke to generate a strong magnetic field.
[0003] Fig. 1 shows a circuit example of a conventional magnetizing unit (e.g., JP 2006-173447 A). The conventional magnetizing unit includes a phase-locked loop 1001 for controlling the phase of an alternating voltage, a step-up transformer 1002 for multiplying the voltage, a rectifier circuit 1003 for rectifying the multiplied voltage, a capacitor C1 to be charged with the rectified voltage, a switching element SCR1 for magnetizing connected in parallel with the capacitor C1, and a magnetizing winding L1. In the conventional magnetizing unit, after the SCR1 conducts, a magnetizing current is supplied to the magnetizing winding L1 by a firing pulse. Due to a counter electromotive force generated in the magnetizing winding L1 from the time the charged voltage of the capacitor C1 reaches zero, a compensating current flows through a freewheeling diode O1.The compensating current of the freewheeling diode O1 prevents heat generation in the winding and regenerates the counter electromotive force.
[0004] When a rotor made of a magnetic material is magnetized in a conventional magnetizing unit, a current flows at a constant voltage. However, when many rotors are continuously magnetized, the temperature of a copper wire forming the magnetizing winding of the magnetizing unit will gradually rise, thus increasing the resistance, causing a decrease in the current flowing through the magnetizing winding. As a result, the magnetization rate of the rotors is reduced, so that a high magnetization rate cannot be maintained. Furthermore, the reduction in the rotor magnetization rate results in a decrease in the torque of a motor.
[0005] JP 2009 - 4 624 A is directed to a magnetization device for magnetizing a magnetized body of a workpiece in which a plurality of magnetized bodies are arranged. The workpiece can be the rotor for a motor. As the magnetization time increases, the temperature also rises due to the heat generated by the magnetization yoke, and the resistance of the yoke coil increases. Therefore, under constant magnetization conditions, the magnetic field generated by the magnetization yoke gradually decreases. If a magnetization defect is detected in a magnetized body, a control device immediately changes the magnetization condition. For example, a coolant such as air or water can be fed to the magnetization yoke to cool or dry it.to avoid overheating, or to suppress the gradual decrease of the magnetic field by increasing the supply current for the magnetizing yoke.
[0006] The article „ Florentin Woergoetter and Bernd Porr: Reinforcement learning. In: Scholarpedia 3(3): 1448. Last edited: September 10, 2012. URL: http: / / scholarpedia.org / article / Reinforcement learning," discusses machine learning systems for interacting with environmental conditions. Here, it is described that an agent can observe a finite number of states in an observation mode and collect a numerical reward, where negative numbers can represent punishments as a result of reinforcement learning. An action value table can be used, for example, to determine the optimal value function.
[0007] Further technical background information can be found in the following documents: US 2008 / 0 018 288 A1. WANG, Ying-Chung; CHIEN, Chiang-Ju; LEE, Der-Tsai: Reinforcement fuzzy-neural adaptive iterative learning control for nonlinear systems. In: 10th International Conference on Control, Automation, Robotics and Vision (ICARCV 2008) Hanoi, Vietnam, 17-20 Dec. 2008, pp. 733-738. SUTTON, Richard S.; BARTO, Andrew G.: Reinforcement learning: an introduction. Cambridge, MA: MIT Press, 1998. JP H11 - 252 873 A. US 8 788 439 B2. SUMMARY OF THE INVENTION
[0008] The present invention aims to provide a magnetizer that improves the consistency of production quality of a rotor and improves the consistency of torque of a motor.
[0009] A magnetizer for manufacturing a motor according to the present invention is defined in claim 1. Preferred embodiments are defined in the subclaims. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] The objects, features and advantages of the present invention will become more apparent from the following description of an embodiment taken in conjunction with the accompanying drawings in which: Fig. 1 is a circuit diagram of a conventional magnetizing unit; Fig. 2 is a block diagram of a machine learning system according to an embodiment of the present invention; Fig. Figure 3 is a schematic view of a model of a neuron; Fig. Figure 4 is a schematic view of a model of a three-layer neural network; Fig. 5 is a flowchart of the operation of the machine learning system according to the embodiment of the present invention; Fig. 6A is a graph showing a change in the magnetization rate before using the machine learning system according to the embodiment of the present invention; and Fig. 6B is a graph showing a change in magnetization rate after using the machine learning system according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0011] A machine learning system and a magnetizer for a motor according to the present invention will be described below with reference to the drawings.
[0012] Fig. Figure 2 is a block diagram of a machine learning system according to an embodiment of the present invention. A machine learning system 10 according to the embodiment of the present invention includes a state observer 1, a reward calculator 2, and a learning unit 3.
[0013] State observer 1 observes the winding temperature, winding resistance, current value, and rotor magnetic flux density of a magnetizing unit 20 with a magnetizing yoke and windings. However, these parameters are only examples, and another parameter may be observed instead.
[0014] The reward calculator 2 calculates a reward from the rotor magnetic flux density obtained by the state observer 2. A method for calculating the reward will be described later.
[0015] Learning Unit 3 updates an action value table based on a magnetization rate calculated from the rotor magnetic flux density and a target magnetic flux rate, the winding temperature, and the winding resistance. A method for updating the action value table will be described later.
[0016] The machine learning system 10 preferably further comprises a decision maker 4 for determining a voltage command based on the action value table of the learning unit 3.
[0017] In this embodiment, a winding resistance R is corrected taking into account the fact that the winding resistance varies depending on a winding temperature Ta detected in the magnetization unit 20, instead of considering the winding resistance R as a constant. A winding resistance R a after correction is calculated by the following copper wire resistance equation: Ra=(234.5+Ta) / (234.5+T0)×R0, where R0 [Ω] is a reference resistance at a reference temperature T0 [°C] and 234.5 is a copper temperature coefficient. In this embodiment, the corrected winding resistance R a used as a resistor to determine a current to obtain a desired magnetic flux density.
[0018] Next, a method for calculating the reward will be described. For example, the reward calculator 2 may provide a negative reward when the magnetization rate is higher or lower than the specified range of a target magnetization rate, while the reward calculator 2 may provide a positive reward when the magnetization rate is within the specified range of the target magnetization rate. The specified range of the target magnetization rate is assumed to be 95 to 100%, for example. If an obtained magnetization rate at this time is 90%, which is lower than the specified range of the target magnetization rate, the reward calculator 2 provides a negative reward. On the other hand, if an obtained magnetization rate is 98%, which is within the specified range of the target magnetization rate, the reward calculator 2 provides a positive reward.
[0019] The learning unit 3 preferably performs a calculation of a state variable observed by the state observer 1 in a multi-layer structure and updates the action value table in real time. As a method for performing the calculation of the state variable in the multi-layer structure, for example, a multi-layer neural network as shown in Fig. 4 shown.
[0020] Here the machine learning system 10, as in Fig. 2. The machine learning system 10 has the function of learning knowledge, as well as the functions of extracting useful rules, knowledge representations, criteria, and the like from a set of data input to the system through analysis and outputting determination results. There are various methods for this, which are generally divided into "supervised learning," "unsupervised learning," and "reinforcement learning." Furthermore, when these methods are implemented, a method called "deep learning" is used, which teaches feature extraction itself.
[0021] In supervised learning, a large amount of data pairs of a specific input and a result (label) are provided to the machine learning system. The machine learning system learns features from the data set and heuristically obtains a model to predict a result from an input, i.e., the relationship between them. In this embodiment, supervised learning can be used to predict a magnetization rate from winding resistance and temperature, and the like. This can be implemented using an algorithm such as a neural network, as described later.
[0022] In unsupervised learning, a machine learning system is provided with a large amount of input data. The machine learning system learns the distribution of the input data and applies compression, classification, alignment, or the like to the input data without being provided with corresponding output data as supervisors. The features of the dataset can be clustered by analogy. Using this result, while providing a specific criterion, an output is assigned to optimize the criterion, and this enables output prediction. There is also a method called semi-supervised learning, which is an intermediate problem between supervised learning and unsupervised learning. In this method, part of the provided data contains pairs of an input and an output, while the other part contains only inputs.In this embodiment, unsupervised learning uses data that can be obtained without actually operating the magnetization unit to improve learning efficiency.
[0023] Problems in reinforcement learning are defined as follows. - The magnetization unit observes an environmental condition and determines an action. - An environment varies according to a certain rule and in addition to this, an action itself can sometimes vary the environment. - A reward signal is sent back each time an action is performed. - The goal is to maximize total reward (discounted reward) in the future. - Learning is initiated from a state in which the result of an action is unknown or insufficiently known. The magnetization unit can only receive the result as data after the magnetization unit has actually been operated. It is necessary to search for an optimal action through trial and error. - Learning can be started from a favorable starting point by previously performing some learning (using a method of supervised learning or reverse reinforcement learning described above) to mimic the action of a human in an initial state.
[0024] Reinforcement learning is a method for learning a sequence of optimal actions by learning, determining, and classifying actions based on the interactions between the action and an environment; in other words, a learning method for maximizing the total reward received in the future. In this embodiment, this indicates that a sequence of actions with an impact on the future can be obtained. The following description takes Q-learning as an example, but is not limited to it.
[0025] Q-learning is a method of learning a Q(s, a) value to select an action "a" in a given environmental state "s." In other words, in a given state "s," an action "a" with the highest value of Q(s, a) is selected as the optimal action. However, a correct Q(a, s) value with respect to a combination of a state "s" and an action "a" is initially unknown. Thus, an agent selects various actions "a" in a given state "s," and performing each action "a" in the state "s" provides the agent with a reward. Consequently, the agent learns to select a better action, i.e., a correct Q(s, a) value.
[0026] The goal is to maximize the total reward to be received in the future by choosing a sequence of actions. Thus, the agent ultimately aims at Q(s, a) = E[γ t r t] (an expected value is taken as the agent moves from state to state by performing a sequence of optimal actions. The sequence of optimal actions is, of course, unknown and must therefore be discovered during learning.) For example, an update equation for a Q(s, a) value is represented as follows: Q(st, at)←Q(st, at)+α(rt+1+γmaxqQ(st+1, a)−Q(st, at)) where s t represents an environmental state at a time t and a t represents an action at time t. By performing action a t the agent moves to a state s t+1 . The state change provides the agent with a reward r t+1 . A term with “max” stands for the product of a Q-value when an action “a” with the highest Q-value known at that time in the state s t+1is selected, and γ. "γ" is a parameter from 0 < γ ≤1, called the discount factor. "α" is a learning rate in the range 0 < α ≤1.
[0027] This equation shows a method for updating a Q(st, at) value of an action a t in a state s t on the basis of a reward r t+1 which resulted from the action a t was returned. This update equation indicates that if a Q(s t+1 , max a t+1 )-value of an optimal action “max a” in the next state, which is followed by a reward r t+1 plus the action “a”, compared with a Q(s t , a t ) value of an action “a” in a state “s” is higher, Q(s t , a t ) is increased. If this is not the case, Q(s t , a t) is reduced. In other words, the value of an action in a given state is approximated to an optimal action value in the next state, that is, derived from a reward returned as a result of the action and the action itself.
[0028] There are two methods for representing Q(s, a) values in a computer: a method in which the Q values of all state-action pairs (s, a) are stored in a table (action value table), and a method in which a function for approximating Q(s, a) values is prepared. In the latter method, the above-described update equation can be implemented by adjusting a parameter for an approximation function using a stochastic gradient method or the like. A neural network can be used as the approximation function, as described later.
[0029] A neural network can be used as an approximation algorithm for a value function in supervised learning, unsupervised learning, and reinforcement learning. The neural network is composed of, for example, a computing unit, a memory, and the like, which imitate a model of a neuron, as shown in Fig. 3 shown. Fig. Figure 3 is a schematic view of a model of a neuron.
[0030] As in Fig. As shown in Figure 3, a neuron produces an output y in response to multiple inputs x (for example, inputs x1 through x3). A weight (w1 through w3) corresponding to the input x is applied to each of the inputs x1 through x3. Thus, the neuron produces the output y, represented by the following equation. Note that the input x, the output y, and the weight w are all in the form of a vector. y=fk(∑i=1nxiwi−θ), where θ is a bias and f kis an activation function.
[0031] Next, a three-layer neural network combining the neurons described above is constructed with reference to Fig. 4 described. Fig. Figure 4 is a schematic view of a neural network with weights of three layers D1 to D3.
[0032] As in Fig. As shown in Figure 4, multiple inputs x (e.g., inputs x1 to x3) are input from the left side of the neural network, and a result y (e.g., results y1 to y3) is output from the right side thereof.
[0033] To be more specific, inputs x1 to x3 are input to each of the three neurons N11 to N13, while they are weighted accordingly. The weights applied to the inputs are collectively indicated by W1.
[0034] Neurons N11 to N13 each output vectors Z11 to Z13. The vectors Z11 to Z13 are collectively represented by a function vector Z1, which is considered to extract a function from the input vector. The function vector Z1 lies between the weight W1 and a weight W2.
[0035] The vectors Z11 to Z13 are input to each of two neurons N21 and N22, while they are weighted accordingly. The weights applied to the function vectors are collectively indicated by W2.
[0036] Neurons N21 and N22 output vectors Z21 and Z22, respectively. The vectors Z21 and Z22 are collectively represented by a function vector Z2. The function vector Z2 lies between the weight W2 and a weight W3.
[0037] The function vectors Z21 and Z22 are input to each of three neurons N31 to N33 and are weighted accordingly. The weights applied to the function vectors are collectively indicated by W3.
[0038] Finally, neurons N31 to N33 output the results y1 to y3, respectively.
[0039] The neural network has a learning mode and a value prediction mode. In learning mode, the weight W is learned using a training dataset. In value prediction mode, the action of the magnetization unit is determined using the parameter obtained in learning mode (the word "prediction" is used for simplicity, but various tasks can be performed, including detection, classification, inference, and the like).
[0040] The agent can learn data directly, i.e., data obtained by actually operating the magnetization unit in prediction mode, and reflect the learning result in the next action (online learning). The agent can also collaboratively learn a group of data collected in advance and then perform recognition mode using the parameters (batch learning). The agent can perform the learning mode every time a certain amount of data is accumulated.
[0041] The weights W1 to W3 can be learned using an error backpropagation algorithm. Information about an error enters from the right side and propagates to the left side. The error backpropagation algorithm is a method in which the weights are adjusted (learned) with respect to each neuron to minimize the difference between an output y and an actual output y (supervisor) in response to an input x.
[0042] Such a neural network can have more than three layers (which is called deep learning). A computational unit that performs feature extraction from inputs in stages and regression of results can be automatically acquired only from supervisor data.
[0043] To perform the Q-learning described above, as in Fig. Accordingly, as shown in Figure 1, the machine learning system 10 according to this embodiment includes the state observer 1, the learning unit 3, and the decision maker 4. However, a machine learning method applied to the present invention is not limited to Q-learning. For example, in the case of applying supervised learning, a value function corresponds to a learning model, and a reward corresponds to an error.
[0044] As in Fig. As shown in Figure 2, a state of the magnetization unit 20 includes a state indirectly changed by an action and a state directly changed by an action. The state indirectly changed by an action includes the winding resistance, the winding temperature, and the magnetization rate. The state directly changed by an action includes the current value.
[0045] Learning Unit 3 updates the action values corresponding to the current state variable and actions to be taken in the action value table based on the update equation and the reward.
[0046] The machine learning system 10 can be connected to the magnetization unit 20 via a network, and the state observer 1 can detect the current state variable through the network. The machine learning system 10 is preferably located on a cloud server.
[0047] The machine learning system 10 may be incorporated into a magnetizer for a motor. In other words, the magnetizer for a motor may include the machine learning system 10 and the magnetization unit 20 with the magnetization yoke and the windings.
[0048] In the example of Fig. 2, the own action value table is updated using the action value table updated by the learning unit of the own machine learning system, but is not limited to this. In other words, the own action value table may be updated using an action value table updated by a learning unit of another machine learning system that is different from the own machine learning system.
[0049] Next, the operation of the machine learning system according to the embodiment of the present invention will be described. Fig. 5 is a flowchart of the operation procedure of the machine learning system according to the embodiment of the present invention.
[0050] First, in step S101, the state observer 1 observes the various states of the magnetization unit 20. In other words, the state observer 1 observes the winding temperature, the winding resistance, the current value, and the rotor magnetic flux density detected by the magnetization unit 20.
[0051] Next, in step S102, the reward calculator 2 calculates a reward from the observed conditions. For example, the reward calculator 2 provides a negative reward if the magnetization rate exceeds the specified range of the target magnetization rate, while the reward calculator 2 provides a positive reward if the magnetization rate is within the specified range of the target magnetization rate.
[0052] Next, in step S103, the learning unit 3 learns an action value from the observed states and the reward.
[0053] Next, in step S104, the decision maker 4 determines an optimal parameter (an optimal action) based on the conditions and the action value. For example, the decision maker 4 determines the parameter (the winding temperature, the winding resistance, and the current value) to obtain a desired magnetization rate.
[0054] Next, in step S105, the states change according to the parameter (action). In other words, the magnetization unit 20 determines a voltage command to produce a current value to obtain the desired magnetization rate.
[0055] The magnetization rate, which is the state directly changed by an action, reaches its target value in this way. Fig. 6A and Fig. 6B show an example of an adjustment of the parameter according to the present invention. Fig. 6A and Fig. 6B are graphs showing the characteristic of a magnetization rate before and after adjustment of the parameter by the machine learning system according to the embodiment of the present invention. Fig. Figure 6A shows the voltage, current, and magnetization rate of several samples prepared before adjusting the parameter. Fig. Figure 6B shows the voltage, current, and magnetization rate of several samples produced with the parameter adjustment of the present invention. Note that vertical axes are divided by arbitrary scale ticks.
[0056] As in Fig.As shown in Figure 6A, if the parameter is not adjusted, the current and magnetization rate decrease with an increase in the number of patterns. This is because, as the number of patterns increases, the winding temperature rises, and thus the resistance of a copper wire forming the windings increases, and consequently, the current flowing through the windings decreases at a constant voltage.
[0057] On the other hand, the magnetizer for a motor according to the embodiment of the present invention adjusts the voltage command to obtain a current value with which a magnetization rate is within the specified range of the target magnetization rate, taking into account the fact that the resistance of the copper wire increases with an increase in the number of patterns, thereby improving the consistency of the magnetization rate.
[0058] As described above, according to the machine learning system and the magnetizer for a motor according to the present invention, it is possible to automatically adjust a parameter for the magnetizing unit by machine learning with respect to a desired magnetization rate, thereby improving the consistency of the magnetization rate.
[0059] According to the machine learning system and the magnetizer for a motor according to the embodiment of the present invention, it is possible to provide the magnetization unit that improves the consistency of the production quality of a rotor and improves the consistency of the torque of the motor.
Claims
[1] Magnetizer for producing a rotor for a motor, the magnetizer comprising: a machine learning system (10) and a magnetization unit (20) with a magnetization yoke and windings, the machine learning system comprising: - a state observer (1) which observes the winding temperature, the winding resistance, the current value and the magnetic rotor flux density of the magnetization unit with the magnetization yoke and the windings; - a reward calculator (2) arranged to calculate a reward from the magnetic rotor flux density obtained by the state observer; and - a learning unit (3) configured to update an action value table based on a magnetization rate calculated from the rotor magnetic flux density and a target magnetic flux rate, the winding temperature, and the winding resistance, wherein the learning unit (3) performs a calculation operation of a state variable observed by the state observer (1) in a multi-layer structure and updates the action value table in real time. [2] The magnetizer according to claim 1, wherein the machine learning system (10) further comprises a decider (4) for determining a voltage command based on the action value table of the learning unit (3). [3] The magnetizer according to any one of claims 1 to 2, wherein the own action value table is updated using an action value table updated by a learning unit of another machine learning system.
Citation Information
Patent Citations
magnetizer
JP2006173447A
Magnetizing method for permanent magnet for motor
JP1999252873A
Synthetic fibre cable and elevator device with such synthetic fibre cable
JP2008094624A
Method of adjusting parameters of a synchronous motor and variable speed drive using such a method
US20080018288A1
Instance weighted learning machine learning model
US8788439B2
