A method, apparatus and system for predicting probe results based on an electrical network reserve pool

Abstract

The application discloses a kind of based on the detection result prediction method of electric network reserve pool, belong to neural network technical field, the method includes: the voltage signal sequence corresponding to historical detection result is passed into the electric network reserve pool based on the input node of multiple type electric characteristic neuron, to map voltage signal sequence into high-dimensional output state matrix;Compared with the dimension increasing method of pure mathematical operation of traditional reserve pool algorithm, this method is completely based on the I-V intrinsic electric characteristic of semiconductor device, for the spontaneous nonlinear behavior of electric network formed by the connection of multiple electric characteristic neurons, it is highly explainable, and it is convenient to physically realize.Further, the output state matrix of electric network reserve pool is dimensioned again, the future detection result is predicted by the fitting of historical detection result, thereby solve the technical problems that extrapolation prediction based on historical data in existing detection technology is difficult to realize.

Description

Technical Field

[0001] This invention belongs to the field of neural network technology, and more specifically, relates to a method, apparatus and system for predicting detection results based on an electrical network reservoir. Background Technology

[0002] In complex exploration missions (such as monitoring sunspot activity, deep-space galaxy surveys, or predicting atmospheric pollutant concentrations), predicting future results based on historical data often faces multi-dimensional technical challenges. From a data perspective, the datasets generated by these missions often exhibit the "3V" characteristics: large volume, complex data types, and inconsistent data quality. Taking sunspot observation as an example, it requires integrating heterogeneous data sources such as multi-band electromagnetic wave data from ground-based telescopes and space probes, magnetic field strength measurements, and coronal mass ejection records. These data not only have inconsistent sampling frequencies but are also frequently affected by instrument noise and atmospheric disturbances. Traditional time series analysis methods (such as the ARIMA model) struggle to effectively capture the implicit nonlinear correlations and long-range dependencies in this type of data.

[0003] From the perspective of system characteristics, these types of probes often exhibit significant chaotic behavior. The diffusion process of atmospheric pollutants is influenced by multiple factors such as turbulence, temperature gradients, and topographic features, resulting in a non-stationary concentration distribution across time and space. Meanwhile, sunspot activity is driven by the not yet fully understood internal dynamics of the sun, exhibiting a coexistence of an 11-year cycle and random outbursts. This interplay of determinism and randomness in system behavior makes extrapolation predictions based on historical data fundamentally difficult. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for predicting detection results based on an electrical network reservoir. The purpose is to utilize electrical neuron units with physical response characteristics to construct a dynamic reservoir structure to upscale and fit historical detection data, thereby obtaining accurate prediction results. This solves the technical problem that extrapolation prediction based on historical detection data is difficult to achieve in existing technologies.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for predicting detection results based on a power grid storage pool is provided, comprising:

[0006] A1: Convert the historical detection results time series of the detection mission into the corresponding voltage signal sequence;

[0007] A2: The voltage signal sequence is fed into a predetermined input node in the electrical network reservoir based on multi-type electrical characteristic neurons, so as to map the voltage signal sequence into a high-dimensional output state matrix;

[0008] The method for constructing the electrical network reservoir includes: constructing a randomly generated set of neurons comprising a certain proportion of linear neurons, nonlinear neurons, and hysteresis neurons; wherein the behavior of the neurons in the set of neurons is defined by the voltage-current characteristics of the semiconductor / insulating dielectric / semiconductor sandwich structure heterojunction; arranging the neurons in a matrix and connecting them laterally or vertically, with overlapping connection positions forming nodes of the electrical network reservoir for signal input and state output;

[0009] A3: Perform a second-order dimensionality increase on the output state matrix, and then perform a linear fit between the second-order dimensionality increase result and the voltage signal converted from the historical detection results, thereby predicting the future results of the detection task.

[0010] Furthermore, when the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is sub-nanometer, the physical meaning of the voltage-current relationship expression corresponds to the memoryless linear transport of electrons in the heterojunction structure, which is defined as the linear neuron.

[0011] Furthermore, the voltage-current relationship expression corresponding to the linear neuron is:

[0012] ;

[0013] Where I represents the current of the electron, V represents the voltage of the electron, and c is the fitting coefficient of the first term.

[0014] Furthermore, when the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is from the sub-nanometer level to the nanometer level, the physical meaning of the voltage-current relationship expression corresponds to the nonlinear transport of electrons in the heterojunction structure, which is defined as the nonlinear neuron.

[0015] Furthermore, the voltage-current relationship expression corresponding to the nonlinear neuron is:

[0016] ;

[0017] Where I represents the current of the electron, V represents the voltage of the electron, c is the first-order fitting coefficient, and d is the third-order fitting coefficient.

[0018] Furthermore, when the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is at the nanometer to micrometer level, the physical meaning of the voltage-current relationship expression corresponds to the transport of electrons in the heterojunction structure in the form of accumulation and release, and is defined as the hysteresis neuron.

[0019] Furthermore, the voltage-current relationship expression corresponding to the hysteresis neuron is:

[0020] ;

[0021] Where I represents the current of the electron, V represents the voltage of the electron, c is the fitting coefficient for the first term, b is the fitting coefficient for the constant term, a is the fitting coefficient for other terms, and Q is the charge.

[0022] According to another aspect of the present invention, a detection result prediction device based on an electrical network reservoir is provided, comprising: connected in sequence:

[0023] The input layer module is used to amplify the data scale of the historical detection results time series of the detection mission to a preset range and convert it into the corresponding voltage signal sequence;

[0024] An electrical network reservoir is used to feed a voltage signal sequence into predetermined input nodes to map the voltage signal sequence into a high-dimensional output state matrix. The method for constructing the electrical network reservoir includes: constructing a randomly generated set of neurons comprising a certain proportion of linear neurons, nonlinear neurons, and hysteresis neurons; wherein the behavior of the neurons in the set is defined by the voltage-current characteristics of a semiconductor / insulating dielectric / semiconductor sandwich heterojunction; and arranging the neurons in a matrix for horizontal or vertical connections, with overlapping connections forming nodes in the electrical network reservoir for signal input and state output.

[0025] The state extension module is used to receive the output state matrix and perform a second-order dimensionality increase on it;

[0026] The output layer module is used to linearly fit the secondary dimensionality increase result with the voltage signal sequence converted from the time series of the historical detection results, thereby obtaining the prediction result of the detection task.

[0027] According to another aspect of the present invention, a detection result prediction system based on an electrical network reservoir is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described task execution result prediction method.

[0028] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described task execution result prediction method.

[0029] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0030] (1) The detection result prediction method based on the electric network reservoir provided in this scheme feeds the voltage signal sequence into the predetermined input node in the electric network reservoir to map the voltage signal sequence into a high-dimensional output state matrix. Compared with the pure mathematical operation-based dimensionality-upgrading method of the existing reservoir algorithm, the dimensionality-upgrading method is spontaneously formed from the nonlinear behavior of the nodes connected by three types of physical electrical characteristics. The method is entirely based on the intrinsic characteristics of the I-V curve of semiconductor devices, which is highly interpretable and easy to implement physically. The detection result prediction method based on the electric network reservoir has the advantage of prediction performance in that it uses physical neurons based on the heterojunction model, which have real physical nonlinearity and memory characteristics. The resulting node state space has higher distinguishability to the input signal. Therefore, by using the dynamic reservoir structure with physical response characteristics to upgrade and fit the historical detection data, accurate prediction results can be obtained. Compared with traditional reservoir calculation methods, under the same scale, the combination of the three types of neurons in this reservoir has stronger nonlinearity, resulting in a more complex internal state. Therefore, the generated data has a higher dimension, indicating that more features are extracted from the data, thus improving the prediction accuracy.

[0031] (2) This invention provides a method for constructing a reservoir based on a multi-type electrical characteristic node electrical network. The characteristics of the electrical characteristic reservoir are described below. The electrical network reservoir is used to map the input time series to a high dimension. It uses three different electrical characteristic neurons of random generation in a certain proportion to be connected horizontally or vertically in a matrix arrangement. The connection positions overlap to form multiple input and output nodes. The electrical characteristics of the neurons are inspired by the transport process of electrons in the heterojunction structure. The voltage and current relationship is extracted from the connection relationship between the nodes of the physical reservoir. It has good interpretability. Moreover, the proportion and distribution of different neurons can be freely customized, which has good flexibility. Under the premise of ensuring performance, the problems of weak interpretability of traditional neurons and poor flexibility of physical reservoirs can be effectively solved.

[0032] (3) The neuron model provided by this scheme is fixed, the proportion is adjustable, and the distribution of different neurons can also be freely customized. Linear neurons, nonlinear neurons and hysteresis neurons are defined by utilizing the change in the thickness of the dielectric layer in the middle of the heterojunction structure to cause the electron transport characteristics in the heterojunction structure; it has both the interpretability of physical reservoir and the flexibility of traditional software algorithm, and realizes a new reservoir design while ensuring performance. Attached Figure Description

[0033] Figure 1 A flowchart of the detection result prediction method based on the power grid storage pool provided in Embodiment 1 of the present invention;

[0034] Figure 2 A flowchart illustrating the method for constructing an electrical network reserve pool based on multiple types of electrical characteristic nodes provided in Embodiment 2 of the present invention;

[0035] Figure 3 This is a schematic diagram of the multi-type electrical properties of neurons provided by the present invention, as shown in Embodiment 2 of the present invention;

[0036] Figure 4 This is a schematic diagram of the structure of the power grid storage tank provided by the present invention, as shown in Embodiment 2 of the present invention;

[0037] Figure 5 This is a schematic diagram of the neuron equation model based on heterojunctions provided in Embodiment 2 of the present invention;

[0038] Figure 6 This is a schematic diagram of the detection result prediction device based on an electrical network storage pool provided in Embodiment 3 of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] Example 1

[0041] like Figure 1 As shown, this embodiment provides a method for predicting detection results based on an electrical network reservoir, including: A1: converting the time series of historical detection results of a detection mission, such as actual missions like sunspot activity monitoring, deep space galaxy exploration, or atmospheric pollutant concentration detection, into corresponding voltage signal sequences; A2: feeding the voltage signal sequences into predetermined input nodes in the aforementioned electrical network reservoir to map the voltage signal sequences into a high-dimensional output state matrix; A3: performing a second dimensionality increase on the output state matrix, and linearly fitting the second dimensionality increase result with the voltage signals converted from the historical detection results, thereby predicting the future results of the detection mission.

[0042] Specifically, the time series is converted into a voltage signal of appropriate amplitude using the described method and fed into the predetermined input nodes of the electrical network. The electrical network reservoir maps the input time series to a high dimension, randomly generating neurons of different electrical characteristics in a certain proportion. These neurons are arranged in a matrix and connected laterally or vertically, with overlapping connections forming multiple input-output nodes to form the electrical network. The input sequence data is input and output to the electrical network at the parameter configuration positions under certain configuration conditions. Then, the generated state sequence is extended to increase its dimension. Finally, the extended state sequence and the historical input sequence are fitted to obtain the fitting result and performance indicators. The structure of the electrical network is inspired by the electrode array structure in physical reservoir devices, and the electrical characteristics of the neurons are inspired by the electron transport process in heterojunction structures. The voltage-current relationship is extracted from the connection relationship between physical reservoir nodes, exhibiting good interpretability. Furthermore, the proportion and distribution of different nodes can be freely customized, providing good flexibility. While ensuring performance, the problems of weak interpretability of traditional neurons and poor flexibility of physical reservoirs can be effectively solved.

[0043] Example 2

[0044] like Figure 2 As shown, this embodiment provides a method for constructing an electrical network reserve pool based on multiple types of electrical characteristic nodes, including:

[0045] S1: Construct a randomly generated set of neurons, including a certain proportion of linear neurons, nonlinear neurons, and hysteresis neurons; wherein, the neurons in the set of neurons are expressed according to the voltage-current relationship when electrons pass through the heterojunction structure, and the heterojunction structure includes semiconductors on both sides and an insulating dielectric layer in the middle.

[0046] S2: Neurons are arranged in a matrix and connected horizontally or vertically, interleaving to form multiple input-output nodes. The number of various neurons is in a fixed proportion. A schematic diagram of the constructed electrical network reservoir is shown below. Figure 3 As shown.

[0047] like Figure 4 As shown, this invention provides neurons with various electrical properties. The parameter configuration of the neurons in this embodiment is described below, which is used to define and describe the neurons exhibiting different electrical properties. In this embodiment, three different types of neurons are used, referred to as linear neurons, nonlinear neurons, and hysteresis neurons.

[0048] For a linear neuron, it satisfies the following formula, which is in the form of a linear function.

[0049]

[0050] Where I is current and V is voltage.

[0051] For a nonlinear neuron, it satisfies the following equation, which is a cubic function.

[0052]

[0053] For hysteresis neurons, the following equation is satisfied, which is in the form of a nonlinear differential equation:

[0054] Where Q is the charge corresponding to the current formed by the aggregation of electrons.

[0055] like Figure 2 A specific embodiment of the present invention for establishing an electrical network reservoir discloses a design for an electrical network reservoir based on nodes with multiple types of electrical characteristics. The electrical network reservoir is in the form of a node array, in which three types of neurons, randomly generated in a certain proportion, are arranged in a matrix and interconnected, with overlapping connection positions forming input and output nodes.

[0056] In this embodiment, a total of 180 neurons with three types of electrical properties are generated. The generated neurons are arranged in a matrix and interconnected, with overlapping connection positions forming input and output nodes. Regarding the connection method of the neurons, the rule adopted in this embodiment is to arrange half of the neurons horizontally and the other half vertically, with adjacent neurons interconnected to form a 10×10 node array. Figure 3 The present invention provides neurons with multiple types of electrical properties.

[0057] Furthermore, when the thickness of the intermediate dielectric layer is sub-nanometer, the physical meaning of the voltage-current relationship corresponds to memoryless linear transport of electrons in the heterojunction structure, which is defined as a linear neuron. Further, when the thickness of the intermediate dielectric layer is from sub-nanometer to nanometer, the physical meaning of the voltage-current relationship corresponds to nonlinear transport of electrons in the heterojunction structure, which is defined as a nonlinear neuron. Further, when the thickness of the intermediate dielectric layer is from nanometer to micrometer, the physical meaning of the voltage-current relationship corresponds to the transport of electrons in the heterojunction structure through accumulation and release, which is defined as a hysteresis neuron.

[0058] like Figure 4The diagram illustrates the electrical properties of the three types of neurons described above. For the heterojunction structure, the overall structure is a sandwich structure composed of semiconductors on both sides and an insulating dielectric layer in the middle. Considering electron transport within this sandwich structure, electron transport is a superposition of two mechanisms: direct tunneling and trap-assisted tunneling. For the direct tunneling mechanism, electrons are thought to directly pass through the potential barrier formed by the insulating dielectric under the influence of an electric field. For the trap-assisted tunneling mechanism, a trap level exists in the insulating dielectric. Electrons are first thermally injected into the trap level, and then tunnel through the remaining, narrower potential barrier. This entire process is modulated by voltage. Applying voltage reduces the conduction band height of the intermediate dielectric layer, increasing the probability of electron tunneling and thus increasing the current. When the barrier width is large, the current through direct tunneling is small; the trap current is mainly generated by trap-assisted tunneling. When the barrier width is small, the direct tunneling current suddenly increases and is much larger than the trap current. At this point, the total current is mainly provided by the direct tunneling current. When the barrier width is extremely small, the barrier disappears, the trap current disappears, and the direct tunneling current becomes the conduction current.

[0059] Figure 5 The model is a neuron equation based on a heterojunction. The characteristic principle equation of the neuron is as follows: when an electron passes through the heterojunction structure, its voltage-current relationship can be expressed by the following unified formula.

[0060]

[0061] Where I is the current, V is the voltage, and Q is the charge corresponding to the current formed by the aggregation of electrons.

[0062] When the thickness of the intermediate dielectric layer is very small (e.g., sub-nanometer), the original unified formula can be simplified to the following linear function. The physical meaning of this is memoryless linear transport of electrons within the heterojunction structure, which is characteristic of linear neurons.

[0063]

[0064] When the thickness of the intermediate dielectric layer increases to a moderate level (e.g., from sub-nanometer to nanometer scale), the original unified formula can be simplified to the following cubic function. The physical meaning here is the nonlinear transport of electrons in the heterojunction structure, a memoryless nonlinear transmission, which is characteristic of nonlinear neurons.

[0065]

[0066] When the thickness of the dielectric layer is very large (e.g., nanometer- to micrometer-scale), the original equation can be simplified to the following nonlinear differential equation. The physical meaning here is the transport of electrons in the heterojunction structure through accumulation and release, a type of transport with memory. This is a characteristic of hysteresis neurons, which are called hysteresis neurons because their images exhibit a loop-like pattern.

[0067]

[0068] It should be understood that the electrical network reservoir can be considered a novel reservoir model applicable to engineering and time-series data processing. The neuron model in this system is inspired by theories of semiconductor physics, exhibiting distinct characteristics. In fact, any other type of natural single reservoir or reservoir array structure possesses similar phenomena and corresponding theories, and can be configured with different neuron response models. From an architectural algorithmic perspective, the difference between neurons lies solely in their equations. Configuration information includes, but is not limited to, the type and parameters of the neuron's response equation. For the electrical network structure, the topology of interconnected neurons forming nodes utilizes horizontal and vertical connections, formed by overlapping connections. Essentially, any topological neuron connection can be used to form nodes. Numerical simulation yields the state matrix, which is then fitted to obtain the prediction result. While existing traditional reservoir architectures have achieved significant performance improvements in prediction, their interpretability is weak, essentially involving a series of operations on matrices without strict physical definitions.

[0069] This architecture, inspired by the electrical transport properties of heterojunctions, directly utilizes these properties as the response characteristics of neurons. Furthermore, while the neuron model is fixed, its proportions are adjustable, and the distribution of different neurons can be freely customized. It combines the interpretability of physical reservoirs with the flexibility of traditional software algorithms, achieving a novel reservoir design while ensuring performance.

[0070] Example 3

[0071] This embodiment provides a device for predicting detection results based on an electrical network storage pool, such as... Figure 6 As shown, the system comprises, in sequence: an input layer module, the aforementioned electrical network storage pool, a state expansion module, and an output layer module. The input layer module amplifies the data size of the historical detection result time series of the detection task to a preset range and converts it into a corresponding voltage signal sequence. The aforementioned electrical network storage pool feeds the voltage signal sequence into predetermined input nodes to map the voltage signal sequence into a high-dimensional output state matrix. The state expansion module receives the output state matrix and performs a second dimensionality increase on it. The output layer module performs linear fitting on the second-dimensionally increased output state matrix to obtain the predicted result of the task execution.

[0072] Specifically, the architecture includes four parts: input layer module 101, electrical network storage pool 102, state extension module 103, and output layer module 104.

[0073] Input layer module 101 is used to convert the time series signal into a voltage signal of appropriate amplitude using the aforementioned method, and then feed it into the predetermined input node of the power network. A scaling factor is introduced here to control the input signal within the range of 0-10V after multiplication.

[0074] Electrical network storage pool 102 is used to map the input time series to a high dimension, such as Figure 2 As shown, firstly, neurons of three different electrical properties are randomly generated in a certain proportion to form a neuron set (e.g., a set of neurons of size 180). These neurons are then connected horizontally or vertically in a matrix arrangement, with overlapping connections forming a two-dimensional node array (e.g., a 10×10 node array). Input time-series data is entered from the lower right corner input node of the storage pool, and output is generated from several selected output nodes (e.g., 80 output nodes). Furthermore, in this example, based on the theory explaining electron transport phenomena in heterojunction structures, key electrical properties (IV relationships) are extracted and slightly simplified to reduce computational resource consumption as the neuron's response to the input signal. In this example, the proportions of the three types of neurons are 0-80%, 0-20%, and 0-20%. Next, the generated electrical network is initialized and iteratively calculated within a certain time. During initialization, all currents, charges, and voltages need to be set to zero, and the top left node is grounded, i.e., the voltage is set to 0. The network is iterated. During each generation of calculation, Kirchhoff's laws are applied to the network to obtain a series of equations (e.g., 180 equations in a 10×10 network) from the voltages and currents of the horizontal and vertical connections. These equations form a system of differential equations. After setting the running time (30s in this example), the RK45 method is used to solve for the numerical solution of each generation. The numerical solutions of each generation are merged into an array, i.e., the voltage sequence. The voltage sequence merging matrix of each generation of nodes during the iteration process is the output state matrix.

[0075] The state extension module 103 is used to extend the states of the generated state sequence, increasing the dimension of the state sequence. For example... Figure 3 As shown, the leakage integral algorithm added to the neuron state extension module can control the update speed of the reservoir state. In the leakage integral processing, the leakage rate α (e.g., 0-1) controls the degree to which the reservoir state is maintained at the previous moment. The smaller the rate, the slower the reservoir state changes at the previous moment, thus resulting in stronger short-term memory in the network. A certain amount of leakage integral processing is beneficial for adjusting the time response capability of both the input and the expected output in the reservoir. The reservoir state after leakage integral processing is as follows:

[0076]

[0077] in This refers to the state of the reservoir under the influence of the leakage rate. It refers to the past step and the current state of the reserve pool.

[0078] Then, weights are assigned to each state in the process of realizing the performance of the reservoir. Here, 15 weight generation functions are set, including linear constant functions and linear functions, nonlinear sigmoid, tanh and softsign functions. Symmetric operations are performed on the functions to form functions with various monotonicity, so as to generate weights with different characteristics. The meanings of the generated weights are mainly three categories, namely (1) the past state is more important, (2) the present state is more important, and (3) all states are equally important to the reservoir operation.

[0079] Next, the selected reserve pool states are parallelized and time-delayed, and weighted according to a set allocation function. Parallelization establishes effective connections between multiple previous states and the current readout layer, increasing the scale of the reserve pool's output nodes and expanding state dimensions. Time delay limits the selection of previous states. The connection coefficients between each state and the readout layer are configured according to the allocation function to allocate states at different times according to a certain priority. Here, the number of parallel states M is 0-9, and the number of time delays N is 1-5 units of time from the previous state. The specific extension is shown in the following formula.

[0080]

[0081] in It is the past state matrix. It is the expanded state matrix. These are the importance weights introduced above. This involves reading out the layer matrix. Specifically, each element in the selected matrix is ​​multiplied by its importance weight, and the matrix is ​​then horizontally concatenated.

[0082] The output layer module 104 is used to linearly fit the output state matrix of the second-dimensional upscaling with the voltage signal converted from the historical detection results, thereby predicting the future results of the detection task to obtain prediction results and performance indicators. Figure 4 As shown, it mainly obtains the readout weight matrix by linearly fitting the high-dimensional state output from the reserve pool with the target value. The ridge regression algorithm is used. The formula for this algorithm is as follows:

[0083]

[0084] in, The transpose of the state matrix. Ridge regression coefficients (taken as 0-1). It is an identity matrix.

[0085] Finally, the prediction results are evaluated using the Non-Maximum Sequence Equations (NMSEs) of the predicted and target sequences. The NMSE calculation formula is as follows:

[0086]

[0087] in, It is the length of the input sequence. It is the variance of the predicted sequence. and It consists of the target sequence and the predicted sequence.

[0088] Example 4

[0089] This embodiment provides a detection result prediction system based on an electrical network reserve pool, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the electrical network reserve pool construction method.

[0090] Example 5

[0091] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for constructing the electrical network reservoir.

[0092] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting detection results based on an electrical network storage pool, characterized in that, include: A1: Convert the historical detection results time series of the detection mission into the corresponding voltage signal sequence; A2: The voltage signal sequence is fed into a predetermined input node in the electrical network reservoir based on multi-type electrical characteristic neurons, so as to map the voltage signal sequence into a high-dimensional output state matrix; The method for constructing the electrical network reservoir includes: constructing a randomly generated set of neurons comprising a certain proportion of linear neurons, nonlinear neurons, and hysteresis neurons; wherein the behavior of the neurons in the set of neurons is defined by the voltage-current characteristics of the semiconductor / insulating dielectric / semiconductor sandwich structure heterojunction; arranging the neurons in a matrix and connecting them laterally or vertically, with overlapping connection positions forming nodes of the electrical network reservoir for signal input and state output; A3: Perform a second-order dimensionality increase on the output state matrix, and linearly fit the second-order dimensionality increase result with the voltage signal converted from the historical detection results, thereby predicting the future results of the detection task; When the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is sub-nanometer, the physical meaning of the voltage-current relationship expression corresponds to the memoryless linear transport of electrons in the heterojunction structure, which is defined as the linear neuron. When the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is from sub-nanometer to nanometer scale, the physical meaning of the voltage-current relationship expression corresponds to the nonlinear transport of electrons in the heterojunction structure, which is defined as the nonlinear neuron. When the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is in the nanometer to micrometer range, the physical meaning of the voltage-current relationship expression corresponds to the transport of electrons in the heterojunction structure in the form of accumulation and release, which is defined as the hysteresis neuron.

2. The method for predicting detection results based on a power grid storage pool as described in claim 1, characterized in that, The voltage-current relationship expression corresponding to the linear neuron is: I = cV; Where I represents the current of the electron, V represents the voltage of the electron, and c is the fitting coefficient of the first term.

3. The method for predicting detection results based on a power grid storage pool as described in claim 1, characterized in that, The voltage-current relationship expression corresponding to the nonlinear neuron is: I = cV+dV 3 ; Where I represents the current of the electron, V represents the voltage of the electron, c is the first-order fitting coefficient, and d is the third-order fitting coefficient.

4. The method for predicting detection results based on a power grid storage pool as described in claim 1, characterized in that, The voltage-current relationship expression corresponding to the hysteresis neuron is: I = aQ + b + cV; Where I represents the current of the electron, V represents the voltage of the electron, c is the fitting coefficient for the first term, b is the fitting coefficient for the constant term, a is the fitting coefficient for other terms, and Q is the charge.

5. A device for predicting detection results based on an electrical network storage pool, characterized in that, Including those connected sequentially: The input layer module is used to amplify the data scale of the historical detection results time series of the detection mission to a preset range and convert it into the corresponding voltage signal sequence; An electrical network reservoir is used to feed a voltage signal sequence into predetermined input nodes to map the voltage signal sequence into a high-dimensional output state matrix. The method for constructing the electrical network reservoir includes: constructing a randomly generated set of neurons comprising a certain proportion of linear neurons, nonlinear neurons, and hysteresis neurons; wherein the behavior of the neurons in the set is defined by the voltage-current characteristics of a semiconductor / insulating dielectric / semiconductor sandwich heterojunction; and arranging the neurons in a matrix for horizontal or vertical connections, with overlapping connections forming nodes in the electrical network reservoir for signal input and state output. The state extension module is used to receive the output state matrix and perform a second-order dimensionality increase on it; The output layer module is used to linearly fit the secondary dimensionality increase result with the voltage signal sequence converted from the time series of the historical detection results, thereby obtaining the prediction result of the detection task. When the thickness of the dielectric layer in the middle of the sandwich structure heterojunction is sub-nanometer, the physical meaning of the voltage-current relationship expression corresponds to the memoryless linear transport of electrons in the heterojunction structure, which is defined as the linear neuron. When the thickness of the dielectric layer in the middle of the sandwich structure heterojunction is from sub-nanometer to nanometer scale, the physical meaning of the voltage-current relationship expression corresponds to the nonlinear transport of electrons in the heterojunction structure, which is defined as the nonlinear neuron. When the thickness of the dielectric layer in the middle of the sandwich-structured heterojunction is in the nanometer to micrometer range, the physical meaning of the voltage-current relationship expression corresponds to the transport of electrons in the heterojunction structure in the form of accumulation and release, which is defined as the hysteresis neuron.

6. A detection result prediction system based on an electrical network reservoir, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for constructing the electrical network storage pool according to any one of claims 1 to 4.

Images