A method for detecting orbit maneuver based on multi-threshold pulse neural network

Abstract

The present disclosure provides a multi-threshold pulse neural network-based orbit maneuver detection method, comprising: obtaining orbit elements and radar cross section of a space target, and encoding the orbit elements and the radar cross section into a pulse sequence; detecting the pulse sequence based on a multi-threshold pulse neural network to obtain an orbit maneuver detection result of the space target. The present disclosure forms the orbit data of the space target to be detected into a pulse sequence with time characteristics, combines the framework of the multi-threshold pulse neural network, expands the gradient non-zero region, solves the gradient disappearance problem of the traditional pulse neural network, improves the feature expression and gradient propagation efficiency, significantly improves the orbit maneuver detection recall rate, effectively reduces the missed detection rate and false alarm rate, and finally realizes more accurate and robust orbit maneuver detection, providing reliable technical support for space safety guarantee.

Description

Technical Field

[0001] This disclosure relates to the field of space observation technology, and in particular to a method, device, storage medium and electronic equipment for detecting orbital maneuvers based on a multi-threshold spiking neural network. Background Technology

[0002] With the rapid development of space technology, the number of satellites and other space targets in Earth orbit has increased rapidly. These targets include not only functional satellites, but also defunct satellites, rocket debris, and various other fragments. To ensure the safety of on-orbit operations, accurate monitoring and prediction of orbital changes of space targets have become crucial. Especially for space targets performing orbital maneuvers (such as orbit changes and collision avoidance maneuvers), timely and accurate detection of their orbital changes is of great significance in avoiding potential collision risks.

[0003] However, traditional orbit prediction methods mainly rely on extrapolation from physical models and historical data. This approach has limitations when dealing with complex nonlinear dynamic behaviors, especially when faced with frequent and unpredictable orbital maneuvers, often failing to provide accurate and reliable predictions. Furthermore, due to the sparsity and noise interference of observational data, extracting useful features from limited information becomes a significant challenge. Summary of the Invention

[0004] The purpose of this disclosure is to provide a method, apparatus, storage medium, and electronic device for detecting orbital maneuvers based on a multi-threshold spiking neural network, in order to solve the problems existing in the prior art.

[0005] The embodiments of this disclosure adopt the following technical solution: an orbital maneuver detection method based on a multi-threshold spiking neural network, comprising: acquiring the orbital root number and radar cross-section of a space target, and encoding the orbital root number and radar cross-section into a pulse sequence; detecting the pulse sequence based on the multi-threshold spiking neural network to obtain the orbital maneuver detection result of the space target.

[0006] In some embodiments, the hidden layer of the multi-threshold spiking neural network includes a plurality of alternately connected fully connected layers and multi-threshold spiking units, each of the multi-threshold spiking units including... Each leakage integral distribution unit has a different trigger threshold.

[0007] In some embodiments, the multi-threshold spiking neural network is trained based on the following steps: acquiring sample pulse sequences, each sample pulse sequence having a real maneuver label; updating the trigger threshold of the leakage integral firing unit in each multi-threshold pulse firing unit according to the predicted maneuver result obtained by the multi-threshold spiking neural network detecting the sample pulse sequence and the real maneuver label corresponding to the sample pulse sequence, until a first convergence condition is met; updating the connection weights of the multi-threshold spiking neural network according to the predicted maneuver result obtained by the multi-threshold spiking neural network detecting the sample pulse sequence and the real maneuver label corresponding to the sample pulse sequence, until a second convergence condition is met.

[0008] In some embodiments, the acquisition of sample pulse sequences, each of which has a real maneuver tag, includes: acquiring the number of six sample orbits and sample radar cross-sections of multiple sample space targets from an open-source space satellite database; calculating the parameter changes of each sample space target at adjacent time points; forming a sample pulse sequence by using Poisson encoding for all the number of six sample orbits, all the sample radar cross-sections, and all the parameter changes of each sample space target; and setting a real maneuver tag for the sample pulse sequence based on the actual orbital maneuvering of the sample space target.

[0009] In some embodiments, updating the trigger threshold of the leakage integral firing unit in each of the multi-threshold pulse firing units, based on the predicted maneuver result obtained by detecting the sample pulse sequence using the multi-threshold spiking neural network and the corresponding real maneuver label, until a first convergence condition is met, includes: determining the pulse code of each multi-threshold pulse firing unit based on the forward propagation logic of the multi-threshold pulse firing unit and the current trigger threshold of the leakage integral firing unit; quantifying the error between the predicted maneuver result and the real maneuver label using a loss function; calculating the gradient of the loss function relative to the model parameters of the multi-threshold pulse firing unit according to the chain rule, wherein the model parameters include at least: the base threshold of the leakage integral firing unit, the adaptive threshold of the leakage integral firing unit, and the threshold decay time constant; updating the model parameters using an optimization algorithm based on the gradient of the loss function relative to the model parameters of the multi-threshold pulse firing unit; and repeating the above steps until the number of iterations meets a first preset number of iterations.

[0010] In some embodiments, updating the connection weights of the multi-threshold spiking neural network (MTN) based on the predicted maneuver result obtained by detecting the sample pulse sequence using the MTN and the corresponding real maneuver label, until a second convergence condition is met, includes: quantifying the error between the predicted maneuver result and the real maneuver label using a loss function; calculating the gradient of the loss function relative to the connection weights according to the chain rule; updating the connection weights using an optimization algorithm based on the gradient of the loss function relative to the connection weights; and repeating the above steps until the number of iterations meets a second preset number of iterations.

[0011] In some embodiments, after obtaining the orbital root count and radar cross-section of the space target, the method further includes: standardizing the orbital root count and radar cross-section.

[0012] This disclosure also provides an orbital maneuver detection device based on a multi-threshold spiking neural network, comprising: an acquisition module for acquiring the orbital root number and radar cross-section of a space target, and encoding the orbital root number and radar cross-section into a pulse sequence; and a detection module for detecting the pulse sequence based on the multi-threshold spiking neural network to obtain the orbital maneuver detection result of the space target.

[0013] This disclosure also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described orbital maneuver detection method based on a multi-threshold spiking neural network.

[0014] This disclosure also provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the described orbital maneuver detection method based on a multi-threshold spiking neural network.

[0015] The beneficial effects of the embodiments disclosed herein are as follows: the orbital data of the space target to be inspected is formed into a pulse sequence with time characteristics, and combined with the framework of a multi-threshold spiking neural network, the gradient non-zero region is expanded, the gradient vanishing problem of traditional spiking neural networks is solved, the efficiency of feature expression and gradient propagation is improved, the recall rate of orbital maneuver detection is significantly improved, the false alarm rate and false alarm rate are effectively reduced, and finally more accurate and robust orbital maneuver detection is achieved, providing reliable technical support for space safety assurance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the orbital maneuver detection method based on a multi-threshold spiking neural network in the first embodiment of this disclosure; Figure 2 This is a flowchart of the training steps of the multi-threshold spiking neural network in the first embodiment of this disclosure; Figure 3 This is a schematic diagram of the MLF unit in the first embodiment of this disclosure; Figure 4 This is a schematic diagram of a multi-threshold spiking neural network in the first embodiment of this disclosure; Figure 5 This is a schematic diagram of the track motion detection device based on a multi-threshold spiking neural network in the second embodiment of this disclosure. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0019] To address the problems existing in the prior art, the first embodiment of this disclosure provides a method for detecting orbital maneuvers based on a multi-threshold spiking neural network, the flowchart of which is shown below. Figure 1 As shown, it mainly includes steps S10 and S20: S10: Obtain the orbital six-root number and radar cross-section of the space target, and encode the orbital six-root number and radar cross-section into a pulse sequence; S20 uses a multi-threshold spiking neural network to detect pulse sequences and obtain the orbital maneuver detection results of space targets.

[0020] The multi-threshold spiking neural network used in this embodiment aims to utilize the spatial and temporal features extracted from the orbital element time series. The core of the network consists of multiple hidden layers, each of which includes multiple alternating fully connected layers (FC) and multi-threshold spiking units (MLF). Each MLF unit includes... Trigger threshold Different Leaky Integrate-and-Fire (LIF) units generate pulses when the membrane potential exceeds their own threshold. The MLF unit expands the non-zero region of the approximate derivative, improves gradient propagation efficiency, and enhances the neuron's ability to express features.

[0021] Figure 2 The flowchart of the training steps of the multi-threshold spiking neural network in this embodiment is shown, mainly including steps S100 to S300: S100, acquire sample pulse sequences, each sample pulse sequence has a real motion tag.

[0022] Specifically, the six root numbers of the sample orbits and the sample radar cross-section (RCS) of multiple sample space targets are first obtained from open-source space satellite databases (such as the Spacetrack website). The six root numbers include orbital inclination, right ascension of the ascending node, eccentricity, argument of perigee, mean perigee, and semi-major axis. The change in the major axis is an important indicator of the orbital maneuver of the space target, while the RCS is related to the physical characteristics of the satellite itself and can help determine the satellite's status.

[0023] Secondly, the parameter changes of each sample space target at adjacent time points are calculated, especially the parameter changes before and after the orbit change. These parameter changes help the model understand the trend of change. These features are mainly obtained by calculating the differences in orbital parameters. For example, for the change in orbital inclination... , This refers to the track inclination angle after the track change. The orbital inclination angle before the orbital change; the change in right ascension at the ascending node. , The right ascension of the ascending node after the orbital change. The right ascension of the ascending node before the orbital change; for the change in eccentricity , Eccentricity after orbital change Eccentricity before orbital change; for the change in perigee angle , The perigee argument after the orbital change. The perigee argument before the orbital change; the change in the mean perigee angle. , The angle of approach after the orbital change. For the angle of approach before the orbital change; for the average change in motion , The average motion after the orbital change. This represents the average motion before the orbital change.

[0024] Next, the orbital root count, the changes in the orbital root count before and after the orbital change, and the radar RCS encoding are converted into sample pulse sequences using Poisson encoding. This encoding step ensures that the time series data can be input into the network model in an event-driven manner. Finally, realistic maneuver labels are assigned to the sample pulse sequences based on the actual orbital maneuvers of the target in the sample space.

[0025] It is important to note that the obtained sample orbital root counts and sample radar cross-sections can first undergo standardization processing, including but not limited to: scaling parameter values ​​to the same range, normalization, and removing outliers. When actually detecting orbital maneuvers of space targets, the orbital root counts and radar cross-sections of the current space target can also be processed using the same standardization method before encoding.

[0026] S200: Based on the predicted maneuver results obtained by detecting the sample pulse sequence using a multi-threshold spiking neural network and the actual maneuver labels corresponding to the sample pulse sequence, update the trigger threshold of the leakage integral release unit in each multi-threshold pulse release unit until the first convergence condition is met.

[0027] In this embodiment, an MLF unit is introduced. Combined with the multi-threshold design of the LIF unit, the forward propagation of the MLF unit is described by the following formula: In the The first layer In the MLF unit, the first LIF units in time membrane potential at time It is obtained by superimposing the membrane potential and the input signal at the previous moment:

[0028] in, The membrane potential vector, For pulse output vector, The attenuation coefficient is... For synaptic current input, when the membrane potential exceeds the corresponding threshold, the LIF unit generates a pulse (binary output):

[0029] in, This is the current trigger threshold vector. This is a step function. The final output of the MLF unit is the sum of all neuron impulses, used to pass signals to the next layer.

[0030] in, It is a pulse encoder. For the first MLF units at time The output of .

[0031] As can be seen from the above formula, single-level triggering At this time, the non-zero region of the approximate derivative (rectangular function) is finite. A large number of LIF units become dormant units with a derivative of 0 because the membrane potential falls into the saturation region, thus blocking gradient propagation. Among these, the rectangular function... The formula for approximating the derivative of the impulse activity is as follows:

[0032] in, It is the width parameter of the rectangle function. It is a sign function. When the membrane potential Falling in the range When the sign function is 1, the approximate derivative is: When the membrane potential exceeds this range, the sign function takes the value of 0, and the approximate derivative is 0. The MLF unit expands the non-zero region through K non-overlapping threshold distributions, such as... Figure 3 As shown, reducing the proportion of dormant units ensures effective gradient propagation and outputs pulse codes. It can distinguish features of different intensities and improve expressive ability.

[0033] Then based on the loss function Quantify the predicted maneuver results With the real motor tag The error between them is calculated, and the gradient of the loss function with respect to the model parameters of the MLF unit is calculated according to the chain rule. The model parameters include at least the base threshold of the leakage integral distribution unit. Adaptive threshold of leakage integral distribution unit and threshold decay time constant .

[0034] Specifically, regarding the basic threshold Its gradient can be expressed as:

[0035] in, Indicates the first Each neural network predicts the output. Indicates the first The membrane potential of each LIF unit Indicates the first The threshold of each LIF unit, This is the basic trigger threshold for the LIF unit.

[0036] For parameters related to adaptive thresholds The gradient can be expressed as:

[0037] in, , This indicates the pulse emitted by the LIF unit, with a value of either 0 or 1. This is the adaptive threshold parameter, used to achieve adaptive threshold adjustment.

[0038] for Its gradient calculation needs to take into account the decay effect of the threshold over time:

[0039] in, , For the first The threshold after LIF cell decay Indicates time, used to describe the process of threshold decay over time.

[0040] Then, based on the gradient of the loss function with respect to the model parameters of the multi-threshold pulse firing unit, an optimization algorithm is used to update the model parameters. In this embodiment, the Adam algorithm is used for the optimization. The update formula for the above model parameters is as follows:

[0041]

[0042]

[0043] in, The learning rate is set to a certain value in this embodiment. Then, based on the updated model parameters, the above steps are repeated to dynamically adjust the trigger threshold of each LIF unit, process the input pulse sequence and perform forward propagation, evaluate model performance, and optimize parameters based on error feedback until the model reaches the first preset number of iterations on the validation set.

[0044] S300: Based on the predicted maneuver results obtained by detecting the sample pulse sequence using a multi-threshold spiking neural network and the actual maneuver labels corresponding to the sample pulse sequence, update the connection weights of the multi-threshold spiking neural network until the second convergence condition is met.

[0045] Multi-threshold spiking neural networks for input features Processing is carried out, among which For the sample size, For feature dimensions, such as Figure 4 As shown, the input is first extended to T time steps by a repetition-based direct coding scheme through the coding layer to generate a coded pulse sequence. The core of the network consists of multiple hidden layers. Each hidden layer contains a linear transformation (i.e., a fully connected layer), an MLF unit, and dropout regularization. The MLF unit implements a biomimetic leaky integral distribution mechanism with multi-threshold dynamic characteristics, and its definition is as follows:

[0046]

[0047] in, Let be the membrane potential at time t. Let W be the attenuation constant and W be the weight matrix. It is a step function. This represents the k-th trigger threshold. The final layer aggregates the pulse outputs through time averaging and applies a linear classifier to generate maneuver probabilities. This design combines event-driven sparsity with multi-scale temporal feature extraction, demonstrating higher efficiency and characterization capabilities when detecting orbital maneuvers in noisy and uncertain environments.

[0048] When training a multi-threshold spiking neural network, the error between the predicted maneuver result and the actual maneuver label is first quantified by the loss function. Then, the gradient of the loss function with respect to the connection weights is calculated according to the chain rule. Subsequently, the connection weights are updated using an optimization algorithm based on the gradient of the loss function with respect to the connection weights. The above steps are repeated until the number of iterations meets the second preset number of iterations.

[0049] This embodiment forms a pulse sequence with time characteristics from the orbital data of the space target to be inspected. Combined with the framework of a multi-threshold spiking neural network, it expands the non-zero gradient region, solves the gradient vanishing problem of traditional spiking neural networks, improves feature expression and gradient propagation efficiency, significantly improves the recall rate of orbital maneuver detection, effectively reduces the false alarm rate and false alarm rate, and ultimately achieves more accurate and robust orbital maneuver detection, providing reliable technical support for space safety.

[0050] Based on the same inventive concept, the second embodiment of this disclosure provides a track motion detection device based on a multi-threshold spiking neural network, the structural schematic of which is shown below. Figure 5As shown, it mainly includes: an acquisition module 10, used to acquire the orbital root number and radar cross-section of the space target, and encode the orbital root number and radar cross-section into a pulse sequence; and a detection module 20, used to detect the pulse sequence based on a multi-threshold pulse neural network to obtain the orbital maneuver detection result of the space target.

[0051] The detection device in this embodiment is a device embodiment corresponding to the detection method in the first embodiment. The functions actually implemented by the above modules have been described in detail in the first embodiment, and will not be repeated here.

[0052] This embodiment forms a pulse sequence with time characteristics from the orbital data of the space target to be inspected. Combined with the framework of a multi-threshold spiking neural network, it expands the non-zero gradient region, solves the gradient vanishing problem of traditional spiking neural networks, improves feature expression and gradient propagation efficiency, significantly improves the recall rate of orbital maneuver detection, effectively reduces the false alarm rate and false alarm rate, and ultimately achieves more accurate and robust orbital maneuver detection, providing reliable technical support for space safety.

[0053] Based on the same inventive concept, the third embodiment of this disclosure provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the orbital maneuver detection method based on a multi-threshold spiking neural network of the first embodiment of this disclosure.

[0054] Based on the same inventive concept, the fourth embodiment of this disclosure provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the orbital maneuver detection method based on a multi-threshold spiking neural network described in the first embodiment of this disclosure.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for detecting orbital maneuvers based on a multi-threshold spiking neural network, characterized in that, include: Obtain the orbital six-point number and radar cross-section of the space target, and encode the orbital six-point number and radar cross-section into a pulse sequence; The orbital maneuver detection results of the space target are obtained by detecting the pulse sequence based on a multi-threshold spiking neural network.

2. The track motion detection method according to claim 1, characterized in that, The hidden layer of the multi-threshold spiking neural network includes multiple alternately connected fully connected layers and multi-threshold spiking units, each of the multi-threshold spiking units including... Each leakage integral distribution unit has a different trigger threshold.

3. The track motion detection method according to claim 2, characterized in that, The multi-threshold spiking neural network is trained based on the following steps: Acquire sample pulse sequences, each of which has a real motion tag; Based on the predicted maneuver results obtained by detecting the sample pulse sequence using the multi-threshold spiking neural network and the actual maneuver labels corresponding to the sample pulse sequence, the trigger threshold of the leakage integral release unit in each of the multi-threshold spiking units is updated until the first convergence condition is met. Based on the predicted maneuver result obtained by detecting the sample pulse sequence using the multi-threshold spiking neural network and the actual maneuver label corresponding to the sample pulse sequence, the connection weights of the multi-threshold spiking neural network are updated until the second convergence condition is met.

4. The track motion detection method according to claim 3, characterized in that, The acquired sample pulse sequences, each of which has a real motion tag, include: Obtain the six root numbers of continuous sample orbits and sample radar cross-sections of multiple sample space targets from an open-source space satellite database; Calculate the parameter changes of each of the sample space targets at adjacent time points; For each sample space target, all the sample orbital six-root numbers, all the sample radar cross-sections, and all the parameter variations are encoded using Poisson to form a sample pulse sequence. The sample pulse sequence is labeled with a real maneuver tag based on the actual orbital maneuver of the target in the sample space.

5. The track motion detection method according to claim 3, characterized in that, The step of updating the trigger threshold of the leakage integral firing unit in each of the multi-threshold pulse firing units by updating the predicted maneuver result obtained by detecting the sample pulse sequence based on the multi-threshold spiking neural network and the real maneuver label corresponding to the sample pulse sequence, until the first convergence condition is met, includes: Based on the forward propagation logic of the multi-threshold pulse emission unit and the current trigger threshold of the leakage integral emission unit, the pulse code of each multi-threshold pulse emission unit is determined; The error between the predicted maneuver result and the actual maneuver label is quantified by a loss function; The gradient of the loss function with respect to the model parameters of the multi-threshold pulse firing unit is calculated according to the chain rule. The model parameters include at least: the base threshold of the leakage integral firing unit, the adaptive threshold of the leakage integral firing unit, and the threshold decay time constant. The model parameters are updated using an optimization algorithm based on the gradient of the loss function with respect to the model parameters of the multi-threshold pulse firing unit. Repeat the above steps until the number of iterations meets the first preset number of iterations.

6. The track motion detection method according to claim 3, characterized in that, The step of updating the connection weights of the multi-threshold spiking neural network by updating the predicted maneuver result obtained from detecting the sample pulse sequence based on the multi-threshold spiking neural network and the real maneuver label corresponding to the sample pulse sequence until the second convergence condition is met includes: The error between the predicted maneuver result and the actual maneuver label is quantified by a loss function; The gradient of the loss function with respect to the connection weights is calculated according to the chain rule; The connection weights are updated using an optimization algorithm based on the gradient of the loss function with respect to the connection weights. Repeat the above steps until the number of iterations meets the second preset number of iterations.

7. The track motion detection method according to claim 1, characterized in that, After obtaining the orbital root number and radar cross-section of the space target, the following is also included: The six orbital elements and the radar cross-section are standardized.

8. A track maneuver detection device based on a multi-threshold spiking neural network, characterized in that, include: The acquisition module is used to acquire the orbital six-axis number and radar cross-section of a space target, and encode the orbital six-axis number and radar cross-section into a pulse sequence; The detection module is used to detect the pulse sequence based on a multi-threshold spiking neural network to obtain the orbital maneuver detection result of the space target.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the orbital maneuver detection method based on a multi-threshold spiking neural network as described in any one of claims 1 to 7.

10. An electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program on the memory, it implements the steps of the orbital maneuver detection method based on a multi-threshold spiking neural network as described in any one of claims 1 to 7.

Images