On-orbit aerospace real-time decision-making system and method

Abstract

The present application relates to the technical field of satellite applications, and relates to an on-orbit aerospace real-time decision-making system and method, for solving the problem of existing Hopfield neural networks being prone to falling into local optima when solving the traveling salesman problem. The system comprises: an MRAM crossbar memory array comprising multiple memory cells, configured to store data to be processed, wherein each memory cell comprises an MTJ device, and the magnetoresistance of the MTJ device for storing data in a radiation environment has a fluctuation value; a weight acquisition module, configured to acquire a real-time magnetoresistance fluctuation value on the basis of the magnetoresistance of each MTJ device in the MRAM crossbar memory array, so as to obtain randomly varying weights; and a decision-making module, configured to process the data to be processed by means of a Hopfield neural network, and configured to use the data as an input of the Hopfield neural network and use the randomly varying weights as improved weights of the Hopfield neural network, and process the data by means of the Hopfield neural network with the improved weights, so as to obtain a globally optimal output value. Since fluctuations are introduced into the weights, this helps to escape local optima during the solving process.

Description

On-orbit aerospace real-time decision system and method TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite applications, and particularly relates to an on-orbit aerospace real-time decision system and method. BACKGROUND

[0002] With the deep exploration of the universe by human beings, the demand for real-time decision of satellites is increasing in the face of disasters, traffic and other sudden conditions. However, in the space environment, electronic devices will be affected by irradiation, resulting in data damage and device failure, thereby affecting the accuracy of processing decisions. Therefore, it is of great significance to develop an on-orbit anti-radiation real-time decision system.

[0003] In the irradiation simulation process, the atomic average displacement per atom (DPA) is an important parameter for measuring the degree of irradiation damage. After determining the irradiation dose, the ratio of the number of displaced atoms in the unit volume of the solid target material to the total number of atoms in the target material before irradiation represents the number of displacements of target atoms during irradiation. For a magnetic tunnel junction, the tunnel magneto-resistance ratio (TMR) is one of its important physical parameters, which is the ratio of the difference between the resistance values of the two ferromagnetic layers of the magnetic tunnel junction in the anti-parallel state and the parallel state to the resistance value in the parallel state, which determines the data read error probability of the magnetic memory.

[0004] In the complex environment of space, due to irradiation damage caused by high-energy particles, the memory may have read errors and thus functional failure. The magnetic tunnel junction is composed of ferromagnetic metals and oxides that are tolerant to irradiation in principle, and the magnetic tunnel junction uses spin mode rather than traditional charge mode to store data, greatly improving the anti-radiation characteristics of the magnetic memory. However, the existing Hopfield neural network is prone to fall into a local optimal solution when solving the traveling salesman problem.

[0005] SUMMARY

[0006] In view of the above analysis, the embodiments of the present application aim to provide an on-orbit aerospace real-time decision system and method to solve the problem that the existing Hopfield neural network is prone to fall into a local optimal solution when solving the traveling salesman problem.

[0007] In one aspect, the embodiment of the present application provides a real-time decision-making system in orbit, comprising: an MRAM cross storage array including a plurality of storage units for storing to-be-processed data, wherein each storage unit includes a magnetic tunnel junction (MTJ) device, the magnetic resistance of the MTJ device for storing data in a radiation environment has a fluctuation value; a weight acquisition module configured to acquire a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross storage array to obtain a randomly changing weight; and a decision-making module configured to process the to-be-processed data by a Hopfield neural network, wherein the to-be-processed data is taken as an input of the Hopfield neural network and the randomly changing weight is taken as an improved weight of the Hopfield neural network, and the to-be-processed data is processed by the Hopfield neural network with the improved weight to obtain a globally optimal output value.

[0008] The technical scheme has the following beneficial effects: the magnetic tunnel junction device with fluctuation in the irradiation environment is used, the weight value is introduced with fluctuation, the neural network has certain randomness in the weight value updating process, which is beneficial to jump out of a local optimal solution and find a global optimal solution in the solving process, thereby avoiding falling into a local optimal solution and improving the speed and accuracy of real-time decision-making.

[0009] Based on the further improvement of the above system, the weight acquisition module includes a reading submodule, a weight submodule and an assignment submodule, wherein the reading submodule is configured to read magnetic resistance encoding data of each storage unit in the MRAM cross storage array in real time; the weight submodule is configured to acquire a real-time magnetic resistance fluctuation value changing over time based on the real-time read magnetic resistance encoding data, and then acquire the randomly changing weight according to the real-time magnetic resistance fluctuation value; and the assignment submodule is configured to assign the randomly changing weight to the weight of the Hopfield neural network in the weight updating process of the Hopfield neural network.

[0010] Based on the further improvement of the above system, the weight submodule is configured to acquire the real-time magnetic resistance fluctuation value by the following formula when the magnetic resistance encoding data satisfies a Gaussian distribution R ~ N(mu, sigma): F = sigma / mu.

[0011] The randomly changing weight is calculated by the following formula: W = W0 + F*(W max -W min );

[0012] Wherein, sigma is the deviation of the magnetic resistance encoding data; mu is the mean of the magnetic resistance encoding data; W0 is the initial weight of the Hopfield neural network; W max is the maximum weight of the Hopfield neural network; and W minThe minimum weight of the Hopfield neural network.

[0013] Based on the further improvement of the above system, the MRAM cross storage array includes N 2 *N 2 storage units for storing the to-be-processed data, the to-be-processed data being the distance between any two cities in N cities; and the Hopfield neural network in the decision module includes N 2 *N 2 weights.

[0014] Based on the further improvement of the above system, the input end of the Hopfield neural network in the decision module is connected with the output end of the MRAM cross storage array, so that the output data in the MRAM cross storage array is provided to the input end of the Hopfield neural network as initial input data; and the input end of the Hopfield neural network in the decision module is connected with the output end of the Hopfield neural network, so that the current output data of the Hopfield neural network is fed back to the input end of the Hopfield neural network as subsequent input data during the processing of the to-be-processed data by the Hopfield neural network, and the current output data of the Hopfield neural network is stored to an additional storage array.

[0015] Based on the further improvement of the above system, the magnetic tunnel junction MTJ device includes: a top electrode; a ferromagnetic reference layer located below the top electrode; a tunneling layer located below the ferromagnetic reference layer; a ferromagnetic free layer located below the tunneling layer; and a bottom electrode located below the ferromagnetic free layer, wherein when the magnetic tunnel junction device is in the radiation environment, the atomic average displacement DPA of the tunneling layer is different from the atomic average displacement of the ferromagnetic reference layer and the ferromagnetic free layer.

[0016] Based on the further improvement of the above system, the materials of the ferromagnetic reference layer and the ferromagnetic free layer include a perpendicular anisotropic magnetic material, wherein the perpendicular anisotropic magnetic material includes any one of CoFeB, Co2FeAl, C O , CoFe, Fe3GeTe2, Ni3GeTe2.

[0017] Based on the further improvement of the above system,

[0018] The on-orbit space real-time decision system is characterized in that it further comprises a plurality of digital-to-analog converters (DACs) and a plurality of analog-to-digital converters (ADCs); the storage unit comprises a 1T1M storage unit, wherein the 1T1M storage unit comprises the magnetic tunnel junction (MTJ) device and a transistor, each of the plurality of ADCs is connected to a gate of the transistor in the corresponding column storage unit; each of the plurality of DACs is connected to a first terminal of the MTJ device in the corresponding row storage unit, so as to convert output data of the MRAM cross storage array into an output voltage or current value as initial input data of the Hopfield neural network; a second terminal of the MTJ device is connected to a drain of the transistor; and a source of the transistor is grounded.

[0019] In another aspect, the embodiments of the present application provide an on-orbit space real-time decision method, comprising: storing to-be-processed data by an MRAM cross storage array, and the magnetic resistance of a magnetic tunnel junction (MTJ) device for storing data in a radiation environment has a fluctuation value, wherein each storage unit in the MRAM cross storage array comprises the MTJ device; obtaining a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross storage array, so as to obtain a randomly changing weight; and processing the to-be-processed data by a Hopfield neural network, for taking the to-be-processed data as input of the Hopfield neural network and taking the randomly changing weight as an improved weight of the Hopfield neural network, and processing the to-be-processed data by the Hopfield neural network with the improved weight to obtain a globally optimal output value.

[0020] Based on the further improvement of the above method, the obtaining of the real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross storage array to obtain a randomly changing weight further comprises: reading magnetic resistance encoding data of each storage unit in the MRAM cross storage array in real time, wherein the magnetic resistance encoding data satisfies a Gaussian distribution; obtaining a real-time magnetic resistance fluctuation value changing over time based on the real-time read magnetic resistance encoding data, and then obtaining the randomly changing weight according to the real-time magnetic resistance fluctuation value; and assigning the randomly changing weight to a weight of the Hopfield neural network in a weight updating process of the Hopfield neural network.

[0021] Compared with the prior art, the present application can at least achieve one of the following beneficial effects:

[0022] 1. The magnetic tunnel junction device with anti-radiation characteristics is used, so that the reading error is greatly reduced, and the accuracy and reliability of data storage and real-time decision are improved;

[0023] 2. The magnetic tunnel junction device is used in a radiation environment with fluctuations, the neural network has a certain randomness in the weight update process due to the introduction of fluctuations, which is beneficial to jump out of the local optimal solution and find the global optimal solution in the solving process, so as to avoid falling into the local optimal solution and improve the speed and accuracy of real-time decision.

[0024] In the present application, the above technical solutions can be combined with each other to realize more preferred combination schemes. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings are included to provide a further understanding of the embodiments, and are incorporated in and constitute a part of this application. The same reference numerals refer to the same elements throughout the drawings.

[0026] Fig. 1 is a diagram of an on-orbit space real-time decision system according to an embodiment of the present application;

[0027] Fig. 2 is a schematic diagram of an on-orbit chip application scenario according to an embodiment of the present application;

[0028] Fig. 3 is a schematic diagram of an MTJ device structure of an MRAM array in an on-orbit space real-time decision system according to an embodiment of the present application;

[0029] Fig. 4 is a hysteresis loop and TMR change of an MTJ device before and after irradiation according to an embodiment of the present application;

[0030] Fig. 5 is a comparison diagram of TMR changes before and after ion irradiation according to an embodiment of the present application;

[0031] Fig. 6 is a DPA distribution diagram in an irradiation simulation MTJ device according to an embodiment of the present application;

[0032] Fig. 7 is a diagram showing the relationship between TMR change and DPA generated by irradiation according to an embodiment of the present application;

[0033] Fig. 8A and Fig. 8B are a single structure diagram and an overall schematic diagram of a Hopfield neural network according to an embodiment of the present application, respectively;

[0034] Fig. 9 is a diagram of average path distance of Traveling Salesman Problem (TSP) solution under different fluctuations according to an embodiment of the present application;

[0035] Fig. 10 is a diagram of an MRAM array in an on-orbit space real-time decision system according to an embodiment of the present application;

[0036] FIG. 11 is a flow chart of a method for real-time decision-making of in-orbit spaceflight according to an embodiment of the present application. DETAILED DESCRIPTION

[0037] The preferred embodiments of the present application will be described in detail with reference to the drawings, in which:

[0038] Referring to FIG. 1, in one embodiment of the present application, a real-time decision-making system for in-orbit spaceflight is disclosed, which comprises: an MRAM cross-memory array 101 comprising a plurality of memory cells for storing data to be processed, wherein each memory cell comprises a magnetic tunnel junction (MTJ) device, and the magnetic resistance of the MTJ device for storing data in a radiation environment has a fluctuation value; a weight acquisition module 102 for acquiring a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross-memory array to obtain a randomly changing weight; and a decision-making module 103 for processing the data to be processed by a Hopfield neural network, wherein the data to be processed is taken as an input of the Hopfield neural network and the randomly changing weight is taken as an improved weight of the Hopfield neural network, and the data to be processed is processed by the Hopfield neural network with the improved weight to obtain a globally optimal output value.

[0039] Compared with the prior art, the real-time decision-making system for in-orbit spaceflight provided in the embodiment uses a magnetic tunnel junction device with fluctuations in a radiation environment. Since the weight has fluctuations, the neural network has a certain randomness in the weight update process, which is beneficial to jumping out of a local optimal solution and finding a globally optimal solution in the solving process, thereby avoiding falling into a local optimal solution and improving the speed and accuracy of real-time decision-making.

[0040] Hereinafter, the MRAM cross-memory array 101, the weight acquisition module 102 and the decision-making module 103 in the real-time decision-making system for in-orbit spaceflight according to the embodiment of the present application will be described in detail with reference to FIG. 1.

[0041] The MRAM cross-memory array 101 comprises a plurality of memory cells for storing data to be processed, wherein each memory cell comprises a magnetic tunnel junction (MTJ) device, and the magnetic resistance of the MTJ device for storing data in a radiation environment has a fluctuation value.

[0042] Referring to FIG. 3, the magnetic tunnel junction (MTJ) device comprises: a top electrode (TE); a ferromagnetic reference layer (RL) located below the top electrode; a tunneling layer (BL) located below the ferromagnetic reference layer; a ferromagnetic free layer (FL) located below the tunneling layer; and a bottom electrode (BE) located below the ferromagnetic free layer, wherein when the magnetic tunnel junction device is in a radiation environment, the atomic average displacement (DPA) of the tunneling layer is different from the atomic average displacement of the ferromagnetic reference layer and the ferromagnetic free layer.

[0043] Materials of the ferromagnetic reference layer and the ferromagnetic free layer include perpendicular anisotropic magnetic materials, wherein the perpendicular anisotropic magnetic materials include CoFeB, Co2FeAl, C O , CoFe, Fe3GeTe2, Ni3GeTe2.

[0044] The storage unit includes a 1T1M storage unit. Referring to FIG. 10, the 1T1M storage unit includes a magnetic tunnel junction (MTJ) device and a transistor. The on-orbit real-time decision system includes a plurality of analog-to-digital converters (ADCs) and a plurality of digital-to-analog converters (DACs). Each analog-to-digital converter (ADC) is connected to the gate of the transistor in the corresponding column storage unit; each digital-to-analog converter (DAC) is connected to the first terminal of the magnetic tunnel junction (MTJ) device in the corresponding row storage unit to convert the output data of the MRAM cross storage array into an output voltage or current value as the initial input data of the Hopfield neural network; the second terminal of the magnetic tunnel junction (MTJ) device is connected to the drain of the transistor; and the source of the transistor is grounded. In addition, the storage unit can also include a 2T1M storage unit.

[0045] The weight acquisition module 102 is configured to acquire a real-time magnetic resistance fluctuation value according to the magnetic resistance of each magnetic tunnel junction (MTJ) device in the MRAM cross storage array, so as to obtain a randomly changing weight.

[0046] Specifically, the weight acquisition module includes a reading submodule, a weight submodule, and an assignment submodule. The reading submodule is configured to read the magnetic resistance encoding data of each storage unit in the MRAM cross storage array in real time, wherein the magnetic resistance encoding data satisfies a Gaussian distribution; the weight submodule is configured to acquire a real-time magnetic resistance fluctuation value that changes over time based on the real-time read magnetic resistance encoding data, and then acquire a randomly changing weight according to the real-time magnetic resistance fluctuation value; and the assignment submodule is configured to assign the randomly changing weight to the weight of the Hopfield neural network in the weight update process of the Hopfield neural network. Wherein the weight matrix of the MRAM cross storage array and the Hopfield neural network is of the same size, for example, 64x64 storage units correspond to 64x64 weight matrix.

[0047] Specifically, the weight submodule is further configured to acquire the real-time magnetic resistance fluctuation value by the following formula when the magnetic resistance encoding data satisfies the Gaussian distribution R~N(μ,σ): F=σ / μ.

[0048] The randomly changing weight is calculated by the following formula: W=W0+F*(W max -W min );

[0049] wherein σ is a bias of the magnetoresistive encoding data; μ is a mean of the magnetoresistive encoding data; W0is an initial weight of the Hopfield neural network; W max is a maximum weight of the Hopfield neural network; and W min is a minimum weight of the Hopfield neural network.

[0050] The decision module 103 is configured to process the to-be-processed data by the Hopfield neural network, to process the to-be-processed data as an input of the Hopfield neural network and to process a random change weight as an improved weight of the Hopfield neural network, and to process the to-be-processed data by the Hopfield neural network with the improved weight to obtain a globally optimal output value. For example, the globally optimal output value is a shortest distance of traversing all cities and returning to a starting point.

[0051] The MRAM cross storage array includes N 2 *N 2 storage units configured to store the to-be-processed data, the to-be-processed data being a distance between any two cities of N cities; and the Hopfield neural network in the decision module includes N 2 *N 2 weights.

[0052] Referring to FIG. 8B, an input node of the Hopfield neural network in the decision module is connected with an output end of the MRAM cross storage array, so that the output data in the MRAM cross storage array is provided to the input node of the Hopfield neural network as initial input data; and the input node of the Hopfield neural network in the decision module is connected with an output node of the Hopfield neural network, so that the current output data of the Hopfield neural network is fed back to the input node of the Hopfield neural network as subsequent input data during processing of the to-be-processed data by the Hopfield neural network, and the current output data of the Hopfield neural network is stored to an additional storage array.

[0053] Referring to FIG. 11, one specific embodiment of the present application discloses a method for real-time decision-making in space, comprising: in step S1101, storing to-be-processed data through an MRAM cross-storage array, and the magnetic resistance of a magnetic tunnel junction (MTJ) device for storing data in a radiation environment has a fluctuation value, wherein each storage unit in the MRAM cross-storage array comprises an MTJ device; in step S1102, obtaining a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross-storage array to obtain a randomly changing weight; and in step S1103, processing the to-be-processed data through a Hopfield neural network, for taking the to-be-processed data as an input of the Hopfield neural network and taking the randomly changing weight as an improved weight of the Hopfield neural network, and processing the to-be-processed data through the Hopfield neural network with the improved weight to obtain a globally optimal output value.

[0054] Specifically, obtaining the real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM cross-storage array to obtain the randomly changing weight further comprises: reading magnetic resistance encoding data of each storage unit in the MRAM cross-storage array in real time; obtaining a real-time magnetic resistance fluctuation value changing over time based on the magnetic resistance encoding data read in real time, and then obtaining the randomly changing weight according to the real-time magnetic resistance fluctuation value; and assigning the randomly changing weight to a weight of the Hopfield neural network in a weight updating process of the Hopfield neural network.

[0055] Hereinafter, with reference to FIGS. 2 to 10, a real-time decision-making system in space according to an embodiment of the present application is described in detail in a specific example manner.

[0056] An MRAM array composed of MTJ devices with anti-radiation characteristics is combined with a peripheral circuit to realize logical operation.

[0057] First, an MRAM array composed of MTJ devices is used in an integrated system in space, and the advantages of spin information storage instead of charge information are used to achieve the purpose of anti-radiation, reduce read-write errors of the device, and ensure the accuracy of data storage and logical calculation.

[0058] Second, the storage information of the MTJ storage unit is read through the tunneling magnetoresistance effect, and according to the difference in the magnetic moment direction of the ferromagnetic reference layer and the ferromagnetic free layer, a high-resistance state and a low-resistance state are exhibited. On the algorithm level, the resistance value is mapped to the synaptic weight of the neural network to complete the logical calculation.

[0059] Further, a small amplitude fluctuation of the device in the irradiation environment is utilized, where the resistance R of the MTJ satisfies a Gaussian distribution, i.e., R ~ N(μ, σ), and there is a fluctuation value F = σ / μ. The fluctuation value is mapped to the synaptic weight, and the fluctuation is introduced in the weight update of the neural network, thereby relieving the problem of falling into a local optimal solution in the solving process of the neural network.

[0060] In addition, the definitions of the elements and methods described above are not limited to the various specific structures, shapes, or manners mentioned in the embodiments, and can be simply changed or replaced by those skilled in the art, for example, (1) using a magnetic tunnel junction array with different thin film structures: since the magnetic tunnel junction itself has anti-radiation characteristics (the magnetic tunnel junction uses a spin mode rather than a traditional charge mode to store data), the TMR change of different thin film structures under irradiation can introduce a small fluctuation to the neural network; (2) using magnetic tunnel junctions with different shapes, such as cuboids and rings, and different shapes of magnetic tunnel junctions can achieve the function of switch control by write current; (3) using different architectures of neural networks: after introducing the fluctuation, the solving effect is better, and different neural networks can be used.

[0061] FIG. 2 is a schematic diagram of an on-orbit chip application scenario according to an embodiment of the present application. On-orbit satellite applications involve various fields such as meteorology, navigation, and communication. The satellite needs to store high-quality weather charts and simulate and analyze current weather condition indicators to obtain weather forecasts. The satellite needs to monitor and identify disaster conditions in disaster-prone areas such as forests, rivers, and volcanoes in real time. The satellite needs to store city traffic route maps and monitor road conditions in real time, solve the combination optimization problem of the route between the departure location and the destination, and finally make real-time decisions. Therefore, the core technologies that the on-orbit satellite needs to provide include storage, simulation analysis, image recognition, and real-time decision processing. At the same time, the on-orbit satellite has problems such as high energy consumption, low speed, and high-energy ion irradiation, which greatly affect real-time decision functions, and therefore it is more necessary to use spin information storage and a storage-computing integrated architecture chip for on-orbit space systems.

[0062] Referring to FIG. 3, the MRAM array has a basic structure of 100 magnetic tunnel junctions (MTJs), including 101 top electrodes (TEs), 102 ferromagnetic reference layers (RLs), 103 tunnel layers (BLs), 104 ferromagnetic free layers (FLs), and 105 bottom electrodes (BEs). The ferromagnetic reference layer and the ferromagnetic free layer are composed of one or more of CoFeB, Co2FeAl, C O , CoFe, or two-dimensional ferromagnetic material Fe3GeTe2, Ni3GeTe2, etc., which have an easy magnetization direction perpendicular to the in-plane, and are conducive to the miniaturization of the device size and the fast magnetization flipping.

[0063] Figure 4 shows the hysteresis loop and TMR change of MTJ devices before and after irradiation. In the irradiation experiment, the incident ions are Bi ions with an initial single nuclear energy of 10.32 MeV / u (LET = 98 MeV·cm 2 / mg), and the injection dose is 1.5 x 101 7 ions / cm 2 . Five groups of MTJ devices are arranged to receive irradiation, the R-H hysteresis loop of each group of devices is measured by magneto-optical Kerr effect (MOKE), and the TMR values before and after irradiation are calculated. As can be seen from Figure 4, the TMR value of the MTJ stack decreases before and after irradiation, and the average TMR decreases by 1.02%.

[0064] Referring to Figure 5, the decrease in TMR value varies with the irradiation parameters to more directly understand the relationship between TMR change and irradiation, in which the change with irradiation dose is most obvious, indicating that the irradiation damage generated in the device by heavy ion irradiation has an impact on the MTJ. Therefore, it is necessary to obtain a parameter DPA that characterizes the degree of irradiation damage, and referring to Figure 5, the TMR change of the MTJ according to the embodiment of the present application is the smallest. As can be seen from Figure 5, compared with devices 1 to 6, the MTJ device according to the embodiment of the present application has strong anti-irradiation characteristics. Figure 6 shows the distribution of DPA in the device under the same irradiation conditions, in which the ferromagnetic free layer, the barrier layer, and the ferromagnetic reference layer in the MTJ structure are of concern. The barrier layer obtains a DPA of 1.24 x 101 -9 , and the ferromagnetic layer obtains a DPA of 2.01 x 101 -9 , indicating that the degree of irradiation displacement damage in the layer is not high, which is consistent with the slight decrease in TMR value in the irradiation experiment.

[0065] Referring to Figure 6, the relationship between the TMR change generated by irradiation and DPA is summarized. The MTJ film is irradiated with 3 MeV Ta ions, and when the TMR value decreases by 13%, the DPA value increases by 3.15 x 101 -3 , which is much larger than the DPA value of the device after irradiation in this work, indicating that the degree of damage in the device in this work is low and has predictable strong anti-irradiation characteristics, that is, even if the DPA value of the device increases more, the TMR value will not decrease greatly. The MTJ film is irradiated with 17.2 KeV He ions, and the simulation result shows that the DPA value is very large, but the TMR decreases by only 1.45%. Assuming that the material, structure, and packaging process are perfect, the MTJ device according to the embodiment of the present application will not produce a large TMR decrease value, thus indicating that the device fluctuates in the irradiation environment, but does not affect the normal use of the device.

[0066] Figure 8A shows a schematic diagram of the structure of a Hopfield neural network. In a Hopfield neural network, the input of a neuron is the output of other neurons at the previous time, and the state of the neuron is updated by a step function after the weighted sum of all inputs. When the network state is stable, the output of all neurons is the output of the neural network. In the implementation process of the Hopfield neural network, the weights are mapped to a crossbar array of programmable MTJ devices, as shown in Figure 8B, which is a schematic diagram of a crossbar array implementing a Hopfield neural network. In the implementation process, the output of other neurons at the previous time is converted into voltage or direct current by a digital-to-analog converter and input to the MTJ device. The product of each input and the conductance or resistance of the MTJ device is obtained by Ohm's law, and the direct current or voltage on the Kirchhoff's law line is accumulated. Finally, the output at this time is obtained after the step function activation and analog-to-digital conversion. Since there is a fluctuation F = σ R / μ R in the resistance of the device in the irradiation environment, a fluctuation W = W0 + F * (W max -W min ) is introduced in the resistance implementation process of the weights of the Hopfield neural network, so that the calculation of the neural network jumps out of the local minimum value and finds the global optimal solution, improving the success rate of finding the optimal path.

[0067] 1. Problem description: The example is to solve the 8-city Traveling Salesman Problem, and the optimal solution corresponds to a unique path.

[0068] The path is obtained from an 8x8 permutation matrix, that is, the final goal of the state update of the neural network is to obtain the state of the 64 points in the permutation matrix, that is, 64 neurons are needed. The structure of the nth neuron is shown in Figure 8A. The state of the neuron is obtained after the activation function, and the weights between neurons are needed in the solving process, so 64*64 MTJ storage units for storing weights are needed, corresponding to Figure 8B.

[0069] 2. Update NN state

[0070] According to the energy function, the permutation matrix is updated by iterative calculation, that is, the path is finally solved. In each iteration, the new state is calculated according to the current state until the stable state is reached. In order to ensure the effectiveness of the order, the constraint condition is added to the energy function, such as excluding loops and ensuring that each city is visited only once.

[0071] 3. Fluctuation

[0072] The Hopfield neural network can stop at a local minimum in the iteration process. In this case, the resistance fluctuation caused by irradiation in the MTJ device is mapped to the weight matrix, so the iteration process can make the neural network calculation jump out of the local minimum, find the global optimal solution, and improve the success rate of finding the optimal path.

[0073] Figure 9 shows the average path of the TSP problem under different fluctuations. The small amplitude fluctuation of the MTJ resistance is mapped to the fluctuation of the neural network weight, and the Hopfield neural network is used to solve the traveling salesman problem between 8 cities. As can be seen from Figure 9, under a large range of fluctuations from 0 to 0.01, the fluctuation 0.0026 existing in the real-time decision system makes the solution result obtain a shorter average path distance 1.703 and a higher solution success rate 72% than the case without fluctuation. The traveling salesman problem (TSP) is a classic combinatorial optimization problem, which is usually described as follows: given a set of cities and the distances between them, find a shortest path that visits each city exactly once and then returns to the starting city.

[0074] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory, a random access memory, etc.

[0075] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed in the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application.

Claims

1. An on-orbit space mission real-time decision system, characterized in that, The method comprises the following steps: An MRAM cross storage array comprising a plurality of storage units for storing data to be processed, wherein each storage unit comprises a magnetic tunnel junction (MTJ) device, and the magnetic resistance of the MTJ device for storing data in a radiation environment fluctuates; a weight acquisition module configured to acquire a real-time magnetic resistance fluctuation value of each MTJ device in the MRAM cross storage array to obtain a randomly changing weight; a decision module configured to process the data to be processed by a Hopfield neural network, wherein the data to be processed is taken as an input of the Hopfield neural network, and the randomly changing weight is taken as an improved weight of the Hopfield neural network, and the data to be processed is processed by the Hopfield neural network with the improved weight to obtain a globally optimal output value. The weight acquisition module comprises a reading submodule, a weight submodule, and an assignment submodule, wherein 2. The on-orbit space real-time decision system of claim 1, wherein, the reading submodule is configured to read magnetic resistance encoding data of each storage unit in the MRAM cross storage array in real time; the weight submodule is configured to acquire a real-time magnetic resistance fluctuation value varying with time based on the real-time magnetic resistance encoding data, and then acquire the randomly changing weight according to the real-time magnetic resistance fluctuation value; and the assignment submodule is configured to assign the randomly changing weight to a weight of the Hopfield neural network in a weight updating process of the Hopfield neural network. The weight submodule is configured to acquire the real-time magnetic resistance fluctuation value by the following formula when the magnetic resistance encoding data satisfies a Gaussian distribution R ~ N(μ, σ):

3. The on-orbit space real-time decision system of claim 2, wherein, F = σ / μ. The randomly changing weight is calculated by the following formula: W = W0 + F * (W max - W min ); wherein σ is a bias of the magnetoresistive encoding data; μ is a mean of the magnetoresistive encoding data; W0is an initial weight of the Hopfield neural network; W max is a maximum weight of the Hopfield neural network; and W min is a minimum weight of the Hopfield neural network.

4. The on-orbit space real-time decision system of claim 1, wherein, The MRAM cross storage array includes N 2 *N 2 storage units for storing the to-be-processed data, the to-be-processed data being distances between any two cities in N cities. and the Hopfield neural network in the decision module comprises N 2 *N 2 weights.

5. The on-orbit space mission real-time decision system of claim 4, wherein, An input node of the Hopfield neural network in the decision module is connected to an output end of the MRAM cross storage array, so that output data in the MRAM cross storage array is provided to the input node of the Hopfield neural network as initial input data; and an input node of the Hopfield neural network in the decision module is connected to an output node of the Hopfield neural network, so that current output data of the Hopfield neural network is fed back to the input node of the Hopfield neural network as subsequent input data in a process of processing data to be processed by the Hopfield neural network, and the current output data of the Hopfield neural network is stored to an additional storage array.

6. The on-orbit space real-time decision system of claim 1, wherein, The MTJ device comprises: a top electrode; a ferromagnetic reference layer located below the top electrode; a tunneling layer located below the ferromagnetic reference layer; a ferromagnetic free layer located below the tunneling layer; and a bottom electrode located below the ferromagnetic free layer, wherein when the MTJ device is in the radiation environment, an atomic average displacement DPA of the tunneling layer is different from atomic average displacements of the ferromagnetic reference layer and the ferromagnetic free layer.

7. The on-orbit space real-time decision system of claim 6, wherein, Materials of the ferromagnetic reference layer and the ferromagnetic free layer include a perpendicular anisotropic magnetic material, wherein the perpendicular anisotropic magnetic material includes any one of CoFeB, Co2FeAl, C O oFe, Fe3GeTe2, Ni3GeTe2.

8. The on-orbit space real-time decision system of claim 1, wherein, a plurality of digital-to-analog converters (DACs) and a plurality of analog-to-digital converters (ADCs); the memory cells include 1T1M memory cells, wherein the 1T1M memory cells include the magnetic tunnel junction (MTJ) devices and transistors, the plurality of ADCs, each of the ADCs being connected to a gate of a transistor in a corresponding column of memory cells; the plurality of DACs, each of the DACs being connected to a first terminal of a magnetic tunnel junction (MTJ) device in a corresponding row of memory cells to convert output data of the MRAM crossbar array into output voltage or current values as initial input data for the Hopfield neural network; a second terminal of the MTJ device being connected to a drain of the transistor; and a source of the transistor being grounded.

9. An on-orbit space mission real-time decision method, characterized in that, comprising: storing to-be-processed data by a MRAM crossbar array, and a fluctuation value of a magnetic resistance of a magnetic tunnel junction (MTJ) device storing the data in a radiation environment, wherein each memory cell in the MRAM crossbar array includes the MTJ device; obtaining a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM crossbar array to obtain a randomly changing weight; and processing the to-be-processed data by a Hopfield neural network, for taking the to-be-processed data as input of the Hopfield neural network and taking the randomly changing weight as an improved weight of the Hopfield neural network, and processing the to-be-processed data by the Hopfield neural network with the improved weight to obtain a globally optimal output value.

10. The on-orbit space real-time decision method according to claim 9, characterized in that, obtaining a real-time magnetic resistance fluctuation value according to the magnetic resistance of each MTJ device in the MRAM crossbar array to obtain a randomly changing weight further includes: reading magnetic resistance encoding data of each memory cell in the MRAM crossbar array in real time; obtaining a real-time magnetic resistance fluctuation value changing over time based on the read magnetic resistance encoding data in real time, and then obtaining the randomly changing weight according to the real-time magnetic resistance fluctuation value; and assigning the randomly changing weight to a weight of the Hopfield neural network in a weight updating process of the Hopfield neural network.

Images