CN115455505A - Active construction method of full-freedom permanent magnet - Google Patents
Active construction method of full-freedom permanent magnet Download PDFInfo
- Publication number
- CN115455505A CN115455505A CN202211031315.3A CN202211031315A CN115455505A CN 115455505 A CN115455505 A CN 115455505A CN 202211031315 A CN202211031315 A CN 202211031315A CN 115455505 A CN115455505 A CN 115455505A
- Authority
- CN
- China
- Prior art keywords
- magnetic
- magnet
- magnetic block
- state
- magnetic field
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000010276 construction Methods 0.000 title claims abstract description 20
- 230000009471 action Effects 0.000 claims abstract description 38
- 238000011156 evaluation Methods 0.000 claims abstract description 38
- 230000003993 interaction Effects 0.000 claims abstract description 29
- 238000013528 artificial neural network Methods 0.000 claims abstract description 22
- 238000004519 manufacturing process Methods 0.000 claims abstract description 14
- 239000000463 material Substances 0.000 claims abstract description 8
- 239000013598 vector Substances 0.000 claims description 28
- 238000000034 method Methods 0.000 claims description 27
- 230000002787 reinforcement Effects 0.000 claims description 15
- 230000004927 fusion Effects 0.000 claims description 11
- 238000013507 mapping Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 9
- 230000005347 demagnetization Effects 0.000 claims description 7
- 230000000694 effects Effects 0.000 claims description 7
- 238000003384 imaging method Methods 0.000 claims description 7
- 230000006872 improvement Effects 0.000 claims description 7
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 5
- 239000000696 magnetic material Substances 0.000 claims description 3
- 230000007704 transition Effects 0.000 claims description 3
- 230000001186 cumulative effect Effects 0.000 claims description 2
- 238000005457 optimization Methods 0.000 description 7
- 239000011159 matrix material Substances 0.000 description 4
- 230000002068 genetic effect Effects 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 2
- 238000002595 magnetic resonance imaging Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- QJVKUMXDEUEQLH-UHFFFAOYSA-N [B].[Fe].[Nd] Chemical compound [B].[Fe].[Nd] QJVKUMXDEUEQLH-UHFFFAOYSA-N 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 229910001172 neodymium magnet Inorganic materials 0.000 description 1
- 229910052761 rare earth metal Inorganic materials 0.000 description 1
- 150000002910 rare earth metals Chemical class 0.000 description 1
- 210000004872 soft tissue Anatomy 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Geometry (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention relates to a full-freedom permanent magnet active construction method, which comprises the following steps: step 1, selecting a permanent magnet material and manufacturing a magnetic block; step 2, constructing a full-freedom-degree magnet geometric model; step 3, determining the initial state of a magnet consisting of a small number of magnetic blocks; step 4, constructing a magnetic field evaluation environment; step 5, constructing a magnetic block action decision network based on a deep neural network; step 6, performing interaction between the magnetic block action decision network and a magnetic field evaluation environment, and storing experience generated by the interaction into a playback unit; and 7, manufacturing the permanent magnet according to the converged permanent magnet model. The invention can improve the magnetic field intensity of the permanent magnet and optimize the uniformity of the magnetic field under the condition of the same weight.
Description
Technical Field
The invention belongs to the technical field of information and communication engineering, and relates to an active construction method of a permanent magnet, in particular to an active construction method of a full-freedom permanent magnet.
Background
Magnetic Resonance Imaging (MRI) has the advantages of zero radiation exposure and good soft tissue imaging contrast, and the performance of a main magnet as the most basic component of a Magnetic resonance imager directly affects the quality of a Magnetic resonance image.
The existing main magnet is mainly divided into a superconducting electromagnetic type and a permanent magnet type. The superconducting electromagnetic main magnet can realize a magnetic field with high field intensity and good uniformity, so that magnetic resonance imaging with better quality compared with the traditional permanent magnet is obtained, and the superconducting electromagnetic main magnet is the mainstream market choice at present. In contrast, the permanent magnet is generally made of rare earth permanent magnet materials, such as neodymium iron boron which is abundant in China, so that the permanent magnet is low in manufacturing cost and easy to popularize. In addition, the permanent magnet also has the advantages of low energy consumption, low maintenance cost and the like. However, the magnetic resonance based on the permanent magnet is limited by poor uniformity, and the imaging quality is generally poor, so that the research on the high-uniformity permanent magnet can reduce the deployment cost of single magnetic resonance and improve the total deployment number of magnetic resonance in the social plane, thereby improving the overall throughput of magnetic resonance detection and having important significance for the increasingly serious aging society.
The permanent magnet generally comprises a plurality of magnetic blocks made of permanent magnet materials according to the special designed spatial distribution, and the uniformity of the permanent magnet is determined by the spatial distribution of the magnetic blocks, so that the uniformity of the permanent magnet is improved by a researcher through the research on the optimization of the spatial arrangement of the magnetic blocks. At present, researchers optimize the arrangement of permanent magnet blocks generally based on a genetic algorithm, and the angle of each magnet block is fixed to the angle of a Halbach type magnet, so that the algorithm can only determine a small number of parameters, such as the type of the magnet block at a certain position. The optimization method has two problems, firstly, the genetic algorithm does not involve the extraction of magnetic field characteristics, the random trial and error of the combination of the magnetic blocks is usually adopted, and the complicated remanence demagnetization effect among the magnetic blocks and the influence of each magnetic block on the total magnetic field are difficult to model; in addition, the angle of the magnetic block is fixed, so that the theoretical performance upper limit of the main magnet is limited. Therefore, although the optimization of the magnetic block arrangement based on the genetic algorithm has made a significant breakthrough in the uniformity of the main magnet, there is still a great room for improvement.
Through searching, the patent documents of the prior art which are the same as or similar to the invention are not found.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a full-freedom permanent magnet active construction method which can improve the field intensity of a magnetic field of a permanent magnet and optimize the uniformity of the magnetic field under the condition of the same weight.
The invention solves the practical problem by adopting the following technical scheme:
a full-freedom permanent magnet active construction method comprises the following steps:
step 2, constructing a full-freedom-degree magnet geometric model;
step 3, determining the initial state of the magnet consisting of a small number of magnetic blocks based on the full-freedom-degree magnet geometric model constructed in the step 2, and setting the initial magnetic blocks as original Halbach angles;
step 4, constructing a magnetic field evaluation environment under a reinforcement learning framework;
step 5, constructing a magnetic block action decision network based on a deep neural network, and performing different modes of feature extraction and fusion on the magnetic field state in the magnetic state and the corresponding magnetic block arrangement state to finally realize accurate mapping from the magnetic state to a Q value;
step 6, based on the magnet evaluation environment constructed in the step 4 and the magnetic block motion decision network constructed in the step 5 and based on the deep neural network, carrying out interaction between the magnetic block motion decision network and the magnetic field evaluation environment according to the preset initial state in the step 3 and the set weight constraint condition, and storing experience generated by the interaction in a playback unit;
step 7, inputting the experience obtained in the step 6 into the magnetic block action decision network in the step 5, optimizing the magnetic block action decision network, and manufacturing a permanent magnet according to the converged permanent magnet model;
moreover, the specific method of step 1 is:
selecting permanent magnetic materials according to magnetic energy product, remanence temperature coefficient and machinability, and manufacturing magnetic blocks with different geometric dimensions according to requirements;
moreover, the specific method of the step 2 is as follows:
and constructing a full-freedom magnet geometric model, wherein the direction of the magnetic block is free, the candidate position of the magnetic block in the model is free, and the magnet parameter is free. The magnetic parameters comprise rows, columns and layer numbers, wherein the rows represent magnetic block rings in the axial direction of the magnet, the columns represent magnetic block columns parallel to the axial direction of the magnet, and all adjacent columns with the same aperture form layers together.
The specific method of step 3 is:
in the geometric model of the magnet, the initial state of the magnet consisting of a small number of magnetic blocks is determined, the initial state of the magnet comprises a plurality of magnetic blocks which are distributed relatively uniformly in a plurality of middle rows of a plurality of layers in the geometric model of the magnet, and the initial magnetic blocks are set to be the original Halbach angle.
Moreover, the specific method of the step 4 is as follows:
constructing a magnetic field evaluation environment containing remanence demagnetization effect modeling based on finite element analysis under a frame of reinforcement learning, wherein the magnetic field evaluation environment is used for realizing the transfer of a magnet state and the evaluation of the magnetic field mean value and the uniformity degree improvement degree of an imaging region before and after the decision of the action of the magnetic block;
the degree of uniformity improvement is referred to as the reward R, and useful evaluation criteria include the average magnetic field strength B mean And magnetic field homogeneity B homogeneous 。
Further, the specific steps of step 5 include:
(1) For the magnetic field state, constructing a deep neural network to extract the characteristics of the magnetic field state to obtain magnetic field characteristics;
(2) For the magnetic block arrangement state corresponding to the magnetic field state, the magnetic block state is encoded by using a magnetic block state vector, the length of a magnetic block state sub-vector is n +1, the magnetic block state sub-vector indicates whether the position of the magnetic block is null by using 1 scalar in the magnetic block state sub-vector, and the angle of the magnetic block is encoded by using n scalars;
(3) Inputting the state vectors of the magnetic blocks into a deep neural network to obtain the arrangement characteristics of the magnetic blocks;
(4) Inputting the magnetic field characteristics and the magnetic block arrangement characteristics into a characteristic fusion module of the deep neural network for characteristic interaction and fusion;
(5) A mapping of magnet states to Q values is obtained at the last classification level of the deep neural network, where Q values represent cumulative discount expectations for reward values in deep reinforcement learning.
Further, the specific steps of step 6 include:
(1) In the preset magnet initial state of the step 3, the interaction between the magnetic block action decision network of the step 5 and the magnetic field evaluation environment of the step 4 is carried out by taking the weight W of the magnet as a termination condition, and after the weight upper limit is reached, the magnet initial state of the step 3 is returned and the interaction process is repeated;
(2) Storing interaction experiences of a magnetic block action decision network and a magnetic field evaluation environment generated in an interaction process into a playback unit, wherein each experience comprises a current magnetic data state, magnetic block actions, rewards and a next magnetic data state;
the specific method of step 7 is:
and (5) taking out experience from the playback unit in the step (6) to optimize the magnetic block motion decision network in the step (5), when the magnetic block motion decision network is converged, a permanent magnet model with optimal performance is constructed, and the main magnet can be manufactured according to the actively constructed full-freedom permanent magnet model.
The invention has the advantages and beneficial effects that:
1. the invention provides a full-freedom permanent magnet active construction method aiming at the problems of the current magnetic block arrangement optimization algorithm and a magnet model, and the full-freedom permanent magnet geometric model is constructed under the algorithm framework of deep reinforcement learning so as to break the upper limit of the theoretical performance of the traditional Halbach type permanent magnet. And a three-dimensional magnetic block action decision network is provided to extract multi-modal characteristics of the magnetic field and the magnetic block arrangement, so that a permanent magnet with higher magnetic field intensity and better magnetic field uniformity compared with a Halbach type permanent magnet is actively constructed in a full-freedom magnet geometric model according to the deep fusion characteristics of the multi-modal.
2. The invention provides an active construction method of a full-freedom permanent magnet, which comprises the steps of selecting permanent magnet materials according to the requirements of maximum magnetic energy product, residual magnetic temperature coefficient, machinability and the like, manufacturing magnetic blocks with different geometric sizes, designing a full-freedom magnet geometric model with free directions and positions of the magnetic blocks and free parameters of the geometric model of the magnet, constructing a magnetic field evaluation environment with residual magnetic demagnetization effect modeling based on a finite element analysis method under the framework of reinforcement learning, constructing a magnetic block action decision network based on a deep neural network, and enabling the decision network to have the accurate mapping capability from a magnet state to an incentive value in the continuous interaction of the magnetic block action decision network and the magnetic field evaluation environment by taking the weight of the magnet as constraint. The magnetic field evaluation environment and magnetic block action decision network can actively explore and optimize the arrangement of the magnetic blocks when the reinforcement learning intelligent body fixes the upper limit of the weight of the magnet, so that the magnetic field intensity of a target imaging area is higher and the uniformity is better.
3. The invention provides a full-freedom permanent magnet active construction method based on a deep neural network, which reduces the limitation on a magnet geometric model as much as possible, and meanwhile, the arrangement of magnetic blocks is actively explored by applying a deep reinforcement learning algorithm, so that a permanent magnet model with high field intensity and good uniformity can be constructed under the condition of light weight constraint.
4. The input based on the deep neural network comprises not only the arrangement states of the magnetic blocks but also the magnetic field state, the two states are collectively called as the magnetic state, the characteristics of the magnetic can be described in an all-dimensional and multi-angle manner, and the interaction and fusion of the two kinds of information improve the performance of the magnetic block action decision network.
5. The magnetic field evaluation environment based on finite element analysis has the capability of modeling the remanence demagnetization effect, can accurately evaluate the magnetic field, gives real and stable reward value feedback, and improves the stability and objectivity of magnetic block action decision network training.
6. The full-freedom-degree magnet geometric model constructed by the method has the characteristics of free position, free direction, free layer number and the like, the theoretical optimal performance of the magnet geometric model is greatly improved, and a larger optimization space is provided for a magnetic block action decision network.
Drawings
FIG. 1 is a flow chart of an active construction method of a full-freedom permanent magnet according to the present invention;
fig. 2 is a schematic view of a full-degree-of-freedom magnet geometric model of the present invention.
Detailed Description
The embodiments of the invention are further described in the following with reference to the drawings:
a full-freedom permanent magnet active construction method is disclosed, as shown in FIG. 1 and FIG. 2, and comprises the following steps:
the specific method of the step 1 comprises the following steps:
selecting permanent magnetic materials according to magnetic energy product, remanence temperature coefficient and machinability, and manufacturing magnetic blocks with different geometric dimensions according to requirements;
in the embodiment, permanent magnet materials are selected according to magnetic energy product, temperature coefficient of remanence and machinability, and magnetic blocks with different geometric dimensions, such as a cube with 1 inch side length, are manufactured according to requirements;
step 2, constructing a full-freedom-degree magnet geometric model;
the specific method of the step 2 comprises the following steps:
and constructing a full-freedom magnet geometric model, wherein the direction (orientation) of the magnetic block is free, the candidate position of the magnetic block in the model is free, and the magnet parameter is free. The magnetic parameters comprise rows, columns and the number of layers, wherein the rows represent magnet ring in the axial direction of the magnet, the columns represent magnet column parallel to the axial direction of the magnet, and all adjacent columns with the same aperture form the layers together.
In this embodiment, a full-degree-of-freedom magnet geometric model is constructed in which the direction (orientation) of the magnetic block is free, the candidate position of the magnetic block in the model, and the magnet parameters are free. The magnetic parameters comprise rows, columns and layer numbers, and the specific geometric model parameters are as follows: first layer L 1 Comprises 16 rows and 42 columns, a 2 nd layer L 2 Comprises 14 rows and 44 columns and a third layer L 3 Comprises 12 rows and 48 columns,Fourth layer L 4 Comprising 8 rows and 52 columns.
Subsequent magnetic block arrangement optimization processes are all performed on the magnet geometric model, and the magnetic blocks used for filling the magnet geometric model are all obtained from the step 1.
Step 3, determining the initial state of the magnet consisting of a small number of magnetic blocks based on the full-freedom-degree magnet geometric model constructed in the step 2, and setting the initial magnetic blocks as original Halbach angles;
the specific method of the step 3 comprises the following steps:
in the geometric model of the magnet, the initial state of the magnet consisting of a small number of magnetic blocks is determined firstly, the initial state of the magnet comprises a plurality of magnetic blocks which are distributed relatively uniformly in a plurality of middle lines of a plurality of layers in the geometric model of the magnet, and the initial magnetic blocks are set to be the original Halbach angle.
In this embodiment, in the geometric model of the magnet, an initial state of the magnet composed of a small number of magnetic blocks is first determined, the initial state of the magnet includes a plurality of magnetic blocks, such as 8 magnetic blocks, distributed relatively uniformly in a plurality of middle rows of a plurality of layers in the geometric model of the magnet, and the initial magnetic blocks are set to be at an original Halbach angle.
Step 4, constructing a magnetic field evaluation environment under a reinforcement learning framework;
the specific method of the step 4 comprises the following steps:
under the framework of reinforcement learning, a magnetic field evaluation environment which is based on finite element analysis and comprises a remanence demagnetization effect modeling is constructed, and the magnetic field evaluation environment is used for realizing the state transition of a magnet and the evaluation of the magnetic field mean value and the uniformity degree of an imaging area before and after the decision of the action of the magnetic block;
the degree of improvement in uniformity is referred to as the reward R, and useful evaluation criteria include the average magnetic field strength B mean And magnetic field homogeneity B homogeneous And the like.
In the embodiment, a magnetic field evaluation environment based on finite element analysis and including remanence demagnetization effect modeling is constructed under a reinforcement learning frame, and the magnetic field evaluation environment can realize the state transition of a magnet and the measurement of the magnetic field mean value and the uniformity degree improvement degree of an imaging area before and after the decision of the motion of the magnet. As described aboveThe degree of improvement is called the reward R, and useful evaluation criteria include the average magnetic field strength B mean And magnetic field homogeneity B homogeneous Etc., in this embodiment R = w 1 ×B mean +w 2 ×B homogeneous ，w 1 ＝1000，w 2 ＝1。
Step 5, constructing a magnetic block action decision network based on a deep neural network, and performing different modes of feature extraction and fusion on the magnetic field state in the magnetic state and the corresponding magnetic block arrangement state to finally realize accurate mapping from the magnetic state to a Q value;
the specific steps of the step 5 comprise:
(1) For the magnetic field state, constructing a deep neural network to extract the characteristics of the magnetic field state to obtain magnetic field characteristics;
(2) For the magnetic block arrangement state corresponding to the magnetic field state, the magnetic block state is encoded by using a magnetic block state vector, the length of a magnetic block state sub-vector is n +1, the magnetic block state sub-vector indicates whether the position of the magnetic block is null by using 1 scalar in the magnetic block state sub-vector, and the angle of the magnetic block is encoded by using n scalars;
(3) Inputting the state vector of the magnetic blocks into a deep neural network to obtain the arrangement characteristics of the magnetic blocks;
(4) Inputting the magnetic field characteristics and the magnetic block arrangement characteristics into a characteristic fusion module of the deep neural network for characteristic interaction and fusion;
(5) Obtaining a mapping of magnet state to Q value at the last classification layer of the deep neural network, wherein the Q value represents the accumulated discount expectation of the reward value in deep reinforcement learning;
in this embodiment, a transform-based magnetic block motion decision network is constructed, and different ways of feature extraction are performed on the magnetic field state in the magnetic state and the corresponding magnetic block arrangement state. For the magnetic field state, it is first decomposed into magnetic field blocks, and then each of the decomposed magnetic field blocks is mapped to a one-dimensional vector of length d =512, called the magnetic field block vector, which has c 1 = 1000; for the arrangement state of the magnetic blocks corresponding to the magnetic field state, the magnetic blocks at adjacent positions are subjected to three-dimensional decomposition to form a plurality of magnetic block groups, and the magnetic blocks are encoded by using magnetic block state vectorsA magnetic block state sub-vector length of n +1, where n =255, indicates whether the magnetic block position is null with 1 scalar and the angle of the magnetic block is encoded with n scalars in the magnetic block state sub-vector. Mapping the state vector of the magnetic blocks in each magnetic block group into a one-dimensional vector called magnetic block group vector, wherein the magnetic block group vector has c 2 = 8; a category vector is also set. Hence, c is shared 1 +c 2 +1=1009 input vectors of dimension 512. To maintain spatial information for the data state, 1009 position vectors are constructed and then embedded into the input vector by way of addition. Finally forming a 512 x 1009 matrix. The matrix is input into 6 transform coding layers for different types of feature interaction to predict phase coding. A transform coding layer sequentially comprises a normalization layer, a multi-head attention layer, a normalization layer and a multi-layer perceptron. The dimension of the output matrix of the last transform coding layer is still 512 x 1009, and the feature corresponding to the category vector in the output matrix is input into a mapping head with a normalization layer and a full connection layer to realize the mapping of the magnet state to the Q value, wherein the Q value represents the accumulated discount expectation of the reward value in the deep reinforcement learning. The Q training process combines the time difference and the current reward value (for example, Q at time t) t ＝R t +Q t+1 ) First, the current reward value is mapped to Q in a time differential manner target (ii) a Then according to Q target Calculating L1Loss according to the Q value predicted by the decision network and feeding the L1Loss back to each layer of the decision network, so that parameter optimization of the magnetic block motion decision network can be realized, and the magnetic block motion decision network has the capability of making a high-performance magnetic block motion decision according to the current magnet state;
step 6, based on the magnet evaluation environment constructed in the step 4 and the magnetic block action decision network constructed in the step 5 and based on the deep neural network, carrying out interaction between the magnetic block action decision network and the magnetic field evaluation environment according to the preset initial state and the set weight constraint condition in the step 3, and storing experience generated by the interaction to a playback unit;
the specific steps of the step 6 comprise:
(1) In the preset magnet initial state of the step 3, the interaction between the magnetic block action decision network of the step 5 and the magnetic field evaluation environment of the step 4 is carried out by taking the weight W of the magnet as a termination condition, and after the weight upper limit is reached, the magnet initial state of the step 3 is returned and the interaction process is repeated;
(2) Storing interaction experiences of a magnetic block action decision network and a magnetic field evaluation environment generated in an interaction process into a playback unit, wherein each experience comprises a current magnet data state, a magnetic block action, a reward and a next magnet data state;
in the present embodiment, in the initial state of the magnet, the interaction between the magnetic block motion decision network and the magnetic field evaluation environment is performed under the condition that the weight W =120kg of the magnet is terminated, and after the upper limit of the weight is reached, the magnetic block motion decision network returns to the initial state of the magnet, and the interaction process is repeated. Storing interaction experiences of each step of magnetic block action decision network and magnetic field evaluation environment into a playback unit, wherein each experience comprises a current magnetic block data state, magnetic block actions, rewards and a next magnetic block data state;
step 7, inputting the experience obtained in the step 6 into the magnetic block action decision network in the step 5, optimizing the magnetic block action decision network, and manufacturing a permanent magnet according to the converged permanent magnet model;
the specific method of the step 7 comprises the following steps:
and (5) taking out experience from the playback unit in the step (6) to optimize the magnetic block motion decision network in the step (5), when the magnetic block motion decision network is converged, a permanent magnet model with optimal performance is constructed, and the main magnet can be manufactured according to the actively constructed full-freedom permanent magnet model.
It should be emphasized that the examples described herein are illustrative and not restrictive, and thus the present invention includes, but is not limited to, those examples described in this detailed description, as well as other embodiments that can be derived from the teachings of the present invention by those skilled in the art and that are within the scope of the present invention.
Claims (8)
1. A full-freedom permanent magnet active construction method is characterized in that: the method comprises the following steps:
step 1, selecting a permanent magnet material and manufacturing a magnetic block;
step 2, constructing a full-freedom-degree magnet geometric model;
step 3, determining the initial state of the magnet consisting of a small number of magnetic blocks based on the full-freedom-degree magnet geometric model constructed in the step 2, and setting the initial magnetic blocks as original Halbach angles;
step 4, constructing a magnetic field evaluation environment under the framework of reinforcement learning;
step 5, constructing a magnetic block action decision network based on a deep neural network, and performing different modes of feature extraction and fusion on the magnetic field state in the magnetic state and the corresponding magnetic block arrangement state to finally realize accurate mapping from the magnetic state to a Q value;
step 6, based on the magnet evaluation environment constructed in the step 4 and the magnetic block action decision network constructed in the step 5 and based on the deep neural network, carrying out interaction between the magnetic block action decision network and the magnetic field evaluation environment according to the preset initial state and the set weight constraint condition in the step 3, and storing experience generated by the interaction to a playback unit;
and 7, inputting the experience obtained in the step 6 into the magnetic block action decision network in the step 5, optimizing the magnetic block action decision network, and manufacturing the permanent magnet according to the converged permanent magnet model.
2. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific method of the step 1 comprises the following steps:
selecting permanent magnetic materials according to magnetic energy product, remanence temperature coefficient and machinability, and manufacturing magnetic blocks with different geometric dimensions according to requirements.
3. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific method of the step 2 comprises the following steps:
constructing a full-freedom-degree magnet geometric model, wherein in the magnet geometric model, the direction of a magnetic block is free, the candidate position of the magnetic block in the model is free, and the magnet parameter is free; the magnetic parameters comprise rows, columns and the number of layers, wherein the rows represent magnet ring in the axial direction of the magnet, the columns represent magnet column parallel to the axial direction of the magnet, and all adjacent columns with the same aperture form the layers together.
4. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific method of the step 3 comprises the following steps:
in the geometric model of the magnet, the initial state of the magnet consisting of a small number of magnetic blocks is determined firstly, the initial state of the magnet comprises a plurality of magnetic blocks which are distributed relatively uniformly in a plurality of middle lines of a plurality of layers in the geometric model of the magnet, and the initial magnetic blocks are set to be the original Halbach angle.
5. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific method of the step 4 comprises the following steps:
under the framework of reinforcement learning, a magnetic field evaluation environment which is based on finite element analysis and comprises a remanence demagnetization effect modeling is constructed, and the magnetic field evaluation environment is used for realizing the state transition of a magnet and the evaluation of the magnetic field mean value and the uniformity degree of an imaging area before and after the decision of the action of the magnetic block;
the degree of uniformity improvement is referred to as the reward R, and useful evaluation criteria include the average magnetic field strength B mean And magnetic field uniformity B homogeneous 。
6. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific steps of the step 5 comprise:
(1) Constructing a deep neural network for the magnetic field state, and extracting the characteristics of the magnetic field state to obtain magnetic field characteristics;
(2) For the magnetic block arrangement state corresponding to the magnetic field state, the magnetic block state is coded by using magnetic block state vectors, the length of a magnetic block state sub-vector is n +1, the magnetic block state sub-vector indicates whether the position of the magnetic block is null by using 1 scalar in the magnetic block state sub-vector or not, and the angle of the magnetic block is coded by using n scalars;
(3) Inputting the state vector of the magnetic blocks into a deep neural network to obtain the arrangement characteristics of the magnetic blocks;
(4) Inputting the magnetic field characteristics and the magnetic block arrangement characteristics into a characteristic fusion module of the deep neural network for characteristic interaction and fusion;
(5) A mapping of magnet state to Q value is obtained at the last classification level of the deep neural network, wherein the Q value represents the cumulative discount expectation of the reward value in deep reinforcement learning.
7. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific steps of the step 6 comprise:
(1) In a preset magnet initial state in the step 3, interaction between the magnetic block action decision network in the step 5 and the magnetic field evaluation environment in the step 4 is carried out by taking the weight W of the magnet as a termination condition, and after the upper limit of the weight is reached, the magnet initial state in the step 3 is returned and the interaction process is repeated;
(2) And storing interaction experiences of the magnetic block action decision network and the magnetic field evaluation environment generated in the interaction process into a playback unit, wherein each experience comprises a current magnetic data state, magnetic block actions, rewards and a next magnetic data state.
8. The active construction method of a full-degree-of-freedom permanent magnet according to claim 1, characterized in that: the specific method of the step 7 comprises the following steps:
and (5) extracting experience from the playback unit in the step (6) to optimize the magnetic block action decision network in the step (5), when the magnetic block action decision network is converged, determining that a permanent magnet model with optimal performance is constructed, and manufacturing a main magnet according to the actively constructed full-freedom-degree permanent magnet model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211031315.3A CN115455505B (en) | 2022-08-26 | 2022-08-26 | Full-freedom permanent magnet active construction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211031315.3A CN115455505B (en) | 2022-08-26 | 2022-08-26 | Full-freedom permanent magnet active construction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115455505A true CN115455505A (en) | 2022-12-09 |
| CN115455505B CN115455505B (en) | 2023-11-24 |
Family
ID=84301453
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211031315.3A Active CN115455505B (en) | 2022-08-26 | 2022-08-26 | Full-freedom permanent magnet active construction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115455505B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116702568A (en) * | 2023-08-04 | 2023-09-05 | 天津天达图治科技有限公司 | Magnetic resonance imaging permanent magnet design method, system, equipment and medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106788028A (en) * | 2016-12-20 | 2017-05-31 | 江苏大学 | Bearing-free permanent magnet synchronous motor intensified learning controller and its building method |
| CN112532002A (en) * | 2020-11-12 | 2021-03-19 | 华中科技大学 | Double-stator excitation full-freedom-degree bearingless motor and active control method thereof |
| US20210116888A1 (en) * | 2019-10-21 | 2021-04-22 | Semiconductor Components Industries, Llc | Systems and methods for system optimization and/or failure detection |
| CN113343592A (en) * | 2021-07-28 | 2021-09-03 | 辽宁锐翔通用飞机制造有限公司 | DQN intelligent control method for permanent magnet synchronous motor of new energy airplane |
| CN113364386A (en) * | 2021-05-26 | 2021-09-07 | 潍柴动力股份有限公司 | H-infinity current control method and system based on reinforcement learning of permanent magnet synchronous motor |
| WO2022083029A1 (en) * | 2020-10-19 | 2022-04-28 | 深圳大学 | Decision-making method based on deep reinforcement learning |
| US20220208355A1 (en) * | 2020-12-30 | 2022-06-30 | London Health Sciences Centre Research Inc. | Contrast-agent-free medical diagnostic imaging |
| CN114723840A (en) * | 2022-03-09 | 2022-07-08 | 天津大学 | Slice-adaptive-oriented active undersampling method for magnetic resonance imaging |
-
2022
- 2022-08-26 CN CN202211031315.3A patent/CN115455505B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106788028A (en) * | 2016-12-20 | 2017-05-31 | 江苏大学 | Bearing-free permanent magnet synchronous motor intensified learning controller and its building method |
| US20210116888A1 (en) * | 2019-10-21 | 2021-04-22 | Semiconductor Components Industries, Llc | Systems and methods for system optimization and/or failure detection |
| WO2022083029A1 (en) * | 2020-10-19 | 2022-04-28 | 深圳大学 | Decision-making method based on deep reinforcement learning |
| CN112532002A (en) * | 2020-11-12 | 2021-03-19 | 华中科技大学 | Double-stator excitation full-freedom-degree bearingless motor and active control method thereof |
| US20220208355A1 (en) * | 2020-12-30 | 2022-06-30 | London Health Sciences Centre Research Inc. | Contrast-agent-free medical diagnostic imaging |
| CN113364386A (en) * | 2021-05-26 | 2021-09-07 | 潍柴动力股份有限公司 | H-infinity current control method and system based on reinforcement learning of permanent magnet synchronous motor |
| CN113343592A (en) * | 2021-07-28 | 2021-09-03 | 辽宁锐翔通用飞机制造有限公司 | DQN intelligent control method for permanent magnet synchronous motor of new energy airplane |
| CN114723840A (en) * | 2022-03-09 | 2022-07-08 | 天津大学 | Slice-adaptive-oriented active undersampling method for magnetic resonance imaging |
Non-Patent Citations (2)
| Title |
|---|
| VAN TAI NGUYEN, ET AL.: "Efficient modelling of permanent magnet field distribution for deep learning applications", 《JOURNAL OF MAGNETISM AND MAGNETIC MATERIALS 》 * |
| 王咏涛: "一种新型永磁球形多自由度电机的研究", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116702568A (en) * | 2023-08-04 | 2023-09-05 | 天津天达图治科技有限公司 | Magnetic resonance imaging permanent magnet design method, system, equipment and medium |
| CN116702568B (en) * | 2023-08-04 | 2023-11-10 | 天津天达图治科技有限公司 | Magnetic resonance imaging permanent magnet design method, system, equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115455505B (en) | 2023-11-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Nerfacc: Efficient sampling accelerates nerfs | |
| Hu et al. | A novel channel pruning method for deep neural network compression | |
| CN109584161A (en) | The Remote sensed image super-resolution reconstruction method of convolutional neural networks based on channel attention | |
| CN111898730A (en) | Structure optimization design method for accelerating by using graph convolution neural network structure | |
| CN112560966B (en) | Polarized SAR image classification method, medium and equipment based on scattering map convolution network | |
| CN115455505B (en) | Full-freedom permanent magnet active construction method | |
| CN113838107B (en) | Automatic heterogeneous image registration method based on dense connection | |
| CN110674742A (en) | Remote sensing image road extraction method based on DLinkNet | |
| CN110516724A (en) | Visualize the high-performance multilayer dictionary learning characteristic image processing method of operation scene | |
| CN110619419A (en) | Passenger flow prediction method for urban rail transit | |
| CN111310811B (en) | Large-scene three-dimensional point cloud classification method based on multi-dimensional feature optimal combination | |
| CN114693064A (en) | Building group scheme generation performance evaluation method | |
| CN116935126A (en) | Medical image classification method based on classical-quantum hybrid machine learning | |
| CN117236201A (en) | Diffusion and ViT-based downscaling method | |
| Wang et al. | Global aligned structured sparsity learning for efficient image super-resolution | |
| Zhang et al. | Memory-efficient hierarchical neural architecture search for image restoration | |
| CN116229061A (en) | Semantic segmentation method and system based on image generation | |
| CN110009048A (en) | A kind of construction method and equipment of neural network model | |
| Ma et al. | Db-rnn: A rnn for precipitation nowcasting deblurring | |
| CN117671666A (en) | Target identification method based on self-adaptive graph convolution neural network | |
| Mohamed et al. | Speeding up fractal image compression by genetic algorithms | |
| CN116660756A (en) | Battery capacity attenuation curve generation method based on condition generation countermeasure network | |
| CN112231880A (en) | Superconducting magnet coil design method | |
| CN115130386A (en) | Digital super-surface one-to-many rapid design method based on nested multilayer perceptron | |
| Wang et al. | MFOC‐CliqueNet: A CliqueNet‐Based Optimal Combination of Multidimensional Features Classification Method for Large‐Scale Laser Point Clouds |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |