An unmanned aerial vehicle control system obstacle avoidance decision method and system

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) control, specifically to an obstacle avoidance decision-making method and system for a UAV control system. The decision-making method includes: processing obstacle information using a spiking neural network to achieve fast and safe heading decision-making; encoding obstacle distance and relative speed information collected by sensors into a composite pulse sequence based on delay and frequency; feeding the pulse sequence into a spiking neural network topologically organized by azimuth angle; a threat assessment zone that can generate a fast pulse response to immediate threats at close range and high speed, and suppress heading decision neurons in dangerous directions to achieve rapid rejection; and a path planning zone that evaluates other relatively safe potential headings. By decoding the pulse firing pattern of output layer neurons within a certain time window, the pattern is converted into a UAV heading deflection command vector, thereby guiding the UAV to flexibly avoid obstacles.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) control, and in particular relates to an obstacle avoidance decision-making method and system for UAV control systems. Background Technology

[0002] As a highly flexible aerial platform, autonomous obstacle avoidance is a crucial technology for the safe operation of unmanned aerial vehicles (UAVs). Traditional UAV obstacle avoidance methods, such as artificial potential field methods and vector field histograms, perform reasonably well in simple environments but cannot cope with complex real-world environments. Methods based on deep neural networks and reinforcement learning demonstrate powerful environmental perception and decision-making capabilities, enabling the extraction of complex features from massive amounts of data and achieving end-to-end control. However, these methods typically rely on high-power computing platforms, whose enormous computational and storage overhead poses a challenge for resource-constrained UAV platforms. Furthermore, traditional frame-based neural network processing inherently suffers from latency when handling rapidly changing scenes, failing to meet the stringent requirements of UAVs for real-time, low-power decision-making.

[0003] Spiking Neural Networks (SNNs), as a third-generation neural network, are considered ideal for building low-power autonomous decision-making systems for unmanned aerial vehicles (UAVs) due to their event-driven, spatiotemporal information encoding, and low-power characteristics. SNNs transmit and process information by simulating the discrete pulses emitted by biological neurons. The asynchronous and sparse computational characteristics of these neurons are highly compatible with low-power hardware, naturally meeting the energy efficiency requirements of UAV platforms. However, at the information encoding level, existing methods mostly employ a single pulse coding scheme, which cannot tightly integrate multidimensional information from sensors into the pulse sequence, limiting the network's accurate perception of environmental threats. Regarding network structure planning, existing SNN models do not reflect specific functional partitioning for the two different decision-making needs in obstacle avoidance tasks: rapid threat avoidance and optimal path exploration. This results in insufficient network response speed when facing sudden dangers or low decision-making efficiency in path planning. Therefore, there is an urgent need for an SNN obstacle avoidance decision-making method that can integrate multidimensional perception information, possess functionally decoupled heterogeneous network topologies, and adapt to environmental changes. Summary of the Invention

[0004] To address the problems of low efficiency in multi-dimensional sensing information fusion, homogeneous network topology, and weak environmental adaptability in existing technologies.

[0005] In the first aspect, the present invention proposes an obstacle avoidance decision-making method for an unmanned aerial vehicle (UAV) control system, comprising the following steps:

[0006] Acquire obstacle point set data detected by UAV sensors, the data including distance and relative velocity information of each obstacle point; perform multi-channel temporal hybrid encoding on the data, and synchronously generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, forming a composite input pulse cluster;

[0007] The composite input pulse cluster is fed into a topologically ordered spiking neural network. The neurons in the input, hidden, and output layers of this spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. Neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed. Through one-to-one strong inhibitory synaptic connections, they inhibit the activity of neurons in the path planning region and the output layer neurons at the corresponding azimuth angle, achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course directions. All neurons in the network employ a recent pulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their pulse firing activity within a preset time window.

[0008] The number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window is decoded. By multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors, a heading deflection command vector for controlling the flight of the UAV is generated.

[0009] Further, the multi-channel temporal hybrid encoding of the data, synchronously generating a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, includes:

[0010] The distance D of the detected obstacle is calculated using the formula... Convert to first pulse delivery delay ,in For encoding time window, This represents the maximum detection range of the drone's sensors.

[0011] when and At that time, the absolute value of the detected relative velocity V is taken as the effective threat velocity. Divide by the maximum relative speed that the sensor can measure. Then multiply the resulting quotient by the maximum firing frequency of the neurons. The pulse firing frequency F is calculated.

[0012] like or When F = 0Hz, then F = 0Hz.

[0013] Further, configuring the neurons within the threat assessment region to generate rapid impulse responses to immediate threats represented by proximity or high relative velocity includes:

[0014] The resting excitation threshold of neurons in the threat assessment area is set to be 2mV lower than that of neurons in the path planning area;

[0015] The initial weights of synaptic connections from input layer neurons to threat assessment region neurons are set to 1.5 times the initial weights of synaptic connections to path planning region neurons.

[0016] Furthermore, the inhibition of activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle through one-to-one strong inhibitory synaptic connections includes:

[0017] When any neuron in the threat assessment region fires a pulse, an inhibitory postsynaptic potential change is applied to the neurons corresponding to the path planning region and the output layer topological position via a pre-set one-to-one inhibitory synapse. The amplitude of this change is dynamically calculated using the following formula and clamped to a non-positive value:

[0018]

[0019] Where k>0 is the inhibition coefficient, To suppress the upper limit of amplitude, The current membrane potential of the inhibited neuron. This is the resting potential of the neuron. It is a positive fundamental suppression constant.

[0020] Furthermore, all neurons in the network employ a recent impulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their impulse firing activity within a preset time window, including:

[0021] Neuron excitation threshold Based on static benchmark threshold With adjustment amount Add them together to get;

[0022] The adjustment amount of the neuron after each firing pulse. Instantly increase the current value by a fixed value β;

[0023] During the pulseless firing period, the adjustment amount is based on the time constant. Exponential decay.

[0024] Further, the decoding of the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window, and the generation of a heading deflection command vector for controlling the UAV flight by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors, includes:

[0025] Preset the corresponding heading deflection angle for each of the N output neurons. , where i=1,2,...,N, and the angles uniformly cover the range from -90 degrees to +90 degrees;

[0026] At the end of the 25ms decision window, the number of pulses fired by each output neuron i is counted. ;

[0027] Calculate the heading deflection command vector The heading deflection unit vector for all neurons Based on the number of pulses fired The vector sum of the weights, i.e. .

[0028] Furthermore, the neurons in the threat assessment region are configured to generate rapid impulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, the activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle is suppressed, thereby achieving rapid rejection of dangerous directions; the path planning region is responsible for assessing potential safe course, including:

[0029] The synaptic connections from the threat assessment region neuron to the corresponding output layer neuron at the topological location of the neuron are set to be strongly inhibitory;

[0030] The synaptic connection from the neuron in the path planning region to the output layer neuron corresponding to the topological position of the neuron is defined as excitability.

[0031] On the other hand, the present invention also provides an obstacle avoidance decision system for an unmanned aerial vehicle (UAV) control system, comprising the following modules:

[0032] The module is used to acquire obstacle point set data detected by the UAV's sensors. The data includes distance and relative speed information of each obstacle point. The data is then subjected to multi-channel temporal hybrid encoding to generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative speed information for each obstacle point, thus forming a composite input pulse cluster.

[0033] An adjustment module is used to input the composite input pulse cluster into a topologically ordered spiking neural network. The input, hidden, and output layer neurons of the spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. Neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, inhibit the activity of neurons in the path planning region and the output layer neurons at the corresponding azimuth angle, achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course directions. Neurons in the network all employ a recent pulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their pulse firing activity within a preset time window.

[0034] The generation module is used to decode the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window. It generates a heading deflection command vector for controlling the flight of the UAV by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors.

[0035] Preferably, the step of performing multi-channel temporal hybrid encoding on the data to synchronously generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point includes:

[0036] The distance D of the detected obstacle is calculated using the formula... Convert to first pulse delivery delay ,in For encoding time window, This represents the maximum detection range of the drone's sensors.

[0037] when and At that time, the absolute value of the detected relative velocity V is taken as the effective threat velocity. Divide by the maximum relative speed that the sensor can measure. Then multiply the resulting quotient by the maximum firing frequency of the neurons. The pulse firing frequency F is calculated.

[0038] like or When F = 0Hz, then F = 0Hz.

[0039] Preferably, configuring the neurons within the threat assessment region to generate rapid impulse responses to immediate threats represented by proximity or high relative velocity includes:

[0040] The resting excitation threshold of neurons in the threat assessment area is set to be 2mV lower than that of neurons in the path planning area;

[0041] The initial weights of synaptic connections from input layer neurons to threat assessment region neurons are set to 1.5 times the initial weights of synaptic connections to path planning region neurons.

[0042] Preferably, the inhibition of activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle through one-to-one strong inhibitory synaptic connections includes:

[0043] When any neuron in the threat assessment region fires a pulse, an inhibitory postsynaptic potential change is applied to the neurons corresponding to the path planning region and the output layer topological position via a pre-set one-to-one inhibitory synapse. The amplitude of this change is dynamically calculated using the following formula and clamped to a non-positive value:

[0044]

[0045] Where k>0 is the inhibition coefficient, To suppress the upper limit of amplitude, The current membrane potential of the inhibited neuron. This is the resting potential of the neuron. It is a positive fundamental suppression constant.

[0046] Preferably, the neurons in the network all employ a recent impulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their impulse firing activity within a preset time window, including:

[0047] Neuron excitation threshold Based on static benchmark threshold With adjustment amount Add them together to get;

[0048] The adjustment amount of the neuron after each firing pulse. Instantly increase the current value by a fixed value β;

[0049] During the pulseless firing period, the adjustment amount is based on the time constant. Exponential decay.

[0050] Preferably, the step of decoding the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window, and generating a heading deflection command vector for controlling the UAV flight by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the results, includes:

[0051] Preset the corresponding heading deflection angle for each of the N output neurons. , where i=1,2,...,N, and the angles uniformly cover the range from -90 degrees to +90 degrees;

[0052] At the end of the 25ms decision window, the number of pulses fired by each output neuron i is counted. ;

[0053] Calculate the heading deflection command vector The heading deflection unit vector for all neurons Based on the number of pulses fired The vector sum of the weights, i.e. .

[0054] Preferably, the neurons in the threat assessment area are configured to generate rapid impulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, the activity of neurons in the path planning area and output layer neurons at the corresponding azimuth angle is inhibited, thereby achieving rapid rejection of dangerous directions; the path planning area is responsible for assessing potential safe course, including:

[0055] The synaptic connections from the threat assessment region neuron to the corresponding output layer neuron at the topological location of the neuron are set to be strongly inhibitory;

[0056] The synaptic connection from the neuron in the path planning region to the output layer neuron corresponding to the topological position of the neuron is defined as excitability.

[0057] This invention employs multi-channel temporal hybrid encoding of obstacle distance and relative velocity information to more comprehensively represent environmental threats. Utilizing a neural network structure based on azimuth topology mapping, and dividing the hidden layers into a threat assessment region and a path planning region in parallel, it achieves coordinated processing of these two functions. The threat assessment region, through a rapid rejection mechanism constructed with strong inhibitory connections, can react instantly to approaching threats, shortening obstacle avoidance decision delays and improving the flight safety of UAVs in complex environments. Simultaneously, the path planning region can evaluate and screen potential safe flight paths after eliminating dangerous directions. The excitation threshold of neurons in the network is adjusted based on recent pulse firing history, ensuring the network's continuous sensitivity to environmental changes and maintaining response stability. Through vector summation decoding, smooth and clear heading commands can be generated, enabling the entire decision-making process to balance rapid response and flight stability. Attached Figure Description

[0058] Figure 1 This is a flowchart of an obstacle avoidance decision-making method for an unmanned aerial vehicle (UAV) control system, provided as an embodiment of the present disclosure. Detailed Implementation

[0059] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0060] In the first embodiment, the present invention proposes an obstacle avoidance decision-making method for an unmanned aerial vehicle (UAV) control system, such as... Figure 1 As shown, it includes the following steps:

[0061] S1, acquire obstacle point set data detected by UAV sensors, the data includes distance and relative speed information of each obstacle point; perform multi-channel time-series hybrid encoding on the data, and synchronously generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative speed information for each obstacle point, forming a composite input pulse cluster;

[0062] Three-dimensional point cloud data of obstacles is acquired using LiDAR or depth cameras. For each point in the point cloud, distance information is obtained by calculating the square root of the sum of the squares of the x, y, and z coordinates. Registration is performed between two consecutive frames of point cloud data. The Iterative Closest Point (ICP) algorithm from the PCL point cloud library is used to first estimate the UAV's pose change, and then, after subtracting the UAV's own motion, the relative velocity information of each obstacle point relative to the UAV is calculated. The azimuth information of each point is obtained by calling the two-parameter arctangent function atan2 to calculate y and x. For each obstacle point, the first pulse delivery delay time is calculated. The calculation is the maximum coding time window. Multiply by the current distance d and the maximum detection distance The merchant, that is The closer the distance, The smaller the value, the earlier the pulse is emitted; the subsequent pulse emission frequency f is related to the effective threat velocity. Proportional relative velocity The absolute value of is only when and Valid for a period of time. The velocity threshold is set to 0.5 m / s; if or It was determined to be a non-threatening target. ,Right now ,in The maximum relative velocity that the sensor can measure. To determine the maximum firing frequency of the neuron, a Poisson distribution pulse generator is used to generate a pulse sequence within the encoding time window based on this frequency. The first pulse and subsequent pulse sequences are then combined to form a composite input pulse for each obstacle point.

[0063] In some embodiments, the multi-channel temporal hybrid encoding of the data, which synchronously generates a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, includes:

[0064] The distance D of the detected obstacle is calculated using the formula... Convert to first pulse delivery delay ,in For encoding time window, This represents the maximum detection range of the drone's sensors.

[0065] when and At that time, the absolute value of the detected relative velocity V is taken as the effective threat velocity. Divide by the maximum relative speed that the sensor can measure. Then multiply the resulting quotient by the maximum firing frequency of the neurons. The pulse firing frequency F is calculated.

[0066] like or When F = 0Hz, then F = 0Hz.

[0067] Assuming the maximum detection range of the sensors carried by the drone The maximum measurable relative speed is 50m. The encoding time window is 20 m / s. The speed threshold is 20ms. =0.5m / s, maximum firing rate of neurons =100Hz. When the sensor detects an approaching obstacle at a certain azimuth angle, with a distance d of 10m, the relative speed is... When the speed is -5m / s, calculate the first pulse delivery delay according to the formula. , =4ms. This means the first pulse representing the obstacle's distance information will be emitted 4ms after the start of the encoding time window. The pulse emission frequency is calculated as f=25Hz, indicating that after the first pulse is emitted, subsequent pulses will be generated at a frequency of 25Hz. If the other obstacle is 2 meters away and has a relative speed of -15m / s, then the delay is... With a time interval of 0.8ms and a frequency f of 75Hz, the pulse sequence will emit one pulse at 0.8ms and another at 14.1ms, generating two pulses within a 20ms encoding window. This encoding strategy ensures that more dangerous signals are encoded as earlier, denser pulse sequences. For an obstacle that is moving away, such as... =+5m / s, or obstacles approaching at low speed, such as =-0.3m / s, pulse frequency f=0Hz, will not generate speed-based threat signals, thus distinguishing between threatening and non-threatening targets.

[0068] S2, the composite input pulse cluster is input into a topologically ordered spiking neural network. The neurons in the input layer, hidden layer, and output layer of the spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. The neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, inhibit the activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle, thereby achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course. The neurons in the network all adopt a threshold mechanism dependent on recent pulse firing history, adjusting their excitation threshold according to their pulse firing activity within a preset time window.

[0069] The horizontal detection range of the UAV, from -90° to +90°, is discretized into N azimuth sectors, with each sector having an angle of 180° / N. A spiking neural network is constructed, consisting of N input neurons, 2N hidden layer neurons, and N output neurons. The obstacle point encoding pulse is input to the index number. Input neurons are divided by the integer part of a single sector angle, and the conversion coefficient between input weight and membrane potential is defined as 0.5 mV / weight unit; neurons in the hidden layer with indices 0 to N-1 constitute the threat assessment region, and neurons with indices N to 2N-1 constitute the path planning region; a full connection is established from input layer neuron i to threat assessment region neuron i and path planning region neuron N+i, with excitability as the weight; a one-to-one connection is established from threat assessment region neuron i to path planning region neuron N+i and to output layer neuron i, with the weight set to a large negative value to achieve a strong inhibitory synapse; Lateral connections of a Mexican hat function are established between neurons in the path planning region, with nearest-neighbor excitation and distant-neighbor inhibition. The nearest-neighbor range is ±2 neurons within the same azimuth angle adjacent sector, and the distant-neighbor range is >2 neurons. The weight of nearest-neighbor excitation decreases Gaussianly with distance, while the weight of distant-neighbor inhibition decreases linearly with distance, in order to find the optimal heading in the uninhibited region. All neurons in the network adopt the leaky integral firing neuron model, namely the ALIF model. This model adds a threshold variable to the standard leaky integral firing model LIF model. After each firing pulse, the excitation threshold of the neuron increases by a fixed value over time. The exponential decay returns to the resting threshold.

[0070] In some embodiments of this application, configuring neurons within the threat assessment region to generate rapid impulse responses to immediate threats represented by proximity or high relative velocity includes:

[0071] The resting excitation threshold of neurons in the threat assessment area is set to be 2mV lower than that of neurons in the path planning area;

[0072] The initial weights of synaptic connections from input layer neurons to threat assessment region neurons are set to 1.5 times the initial weights of synaptic connections to path planning region neurons.

[0073] To achieve rapid threat response, neurons in the threat assessment region are programmed to be more easily activated than those in the path planning region. For example, the resting membrane potential of all neurons in the network is set to -65mV. The resting excitation threshold of neurons in the path planning region is also specified. Setting it to -50mV indicates that a membrane potential accumulation of 15mV is required to trigger a single pulse. This represents the resting excitation threshold of neurons in the threat assessment region. The voltage is set to -52mV, requiring only 13mV of membrane potential accumulation to trigger the pulse. Simultaneously, the synaptic connection weights are also configured differently: assuming initial weights for synaptic connections from the input neuron representing a certain azimuth angle to the corresponding path planning region neuron. If the initial weight is 0.6, then the initial weights of the synaptic connections from the same input neuron to the threat assessment region neuron are... The threshold is then strengthened to 0.9. This dual optimization of threshold and weight ensures that even a single or sparse early input pulse from a high-threat obstacle is sufficient to quickly drive the membrane potential of neurons in the threat assessment zone to the excitation threshold and fire a pulse, thereby enabling priority processing and preemptive response to potential dangers.

[0074] In one embodiment, the inhibition of activity of the path planning region neurons and output layer neurons at the corresponding azimuth angle through one-to-one strong inhibitory synaptic connections includes:

[0075] When any neuron in the threat assessment region fires a pulse, an inhibitory postsynaptic potential change is applied to the neurons corresponding to the path planning region and the output layer topological position via a pre-set one-to-one inhibitory synapse. The amplitude of this change is dynamically calculated using the following formula and clamped to a non-positive value:

[0076]

[0077] Where k>0 is the inhibition coefficient, To suppress the upper limit of amplitude, The current membrane potential of the inhibited neuron. This is the resting potential of the neuron. It is a positive fundamental suppression constant.

[0078] In network topology, the azimuth angle is... The threat assessment area neurons are the same as the azimuth angle. There is a fixed, one-to-one inhibitory connection between neurons in the path planning region. When a neuron in the threat assessment region successfully fires a pulse, it transmits a dynamically regulated inhibitory postsynaptic potential (IPSP) to the target neuron.

[0079] For example, suppose we set a basic suppression constant. The inhibition coefficient k=0.5, and the upper limit of the inhibition amplitude. Resting potential of neurons in the inhibited pathway planning area The excitation threshold is -50mV.

[0080] Scenario 1: Neurons are in a state of cumulative excitation: The membrane potential of a neuron in a certain pathway planning area rises to a certain level due to receiving excitatory input. This value is 10mV higher than the resting potential, approaching the positive threshold. At this point, the corresponding threat assessment area detects the threat and issues a pulse. The calculation is based on the formula: The membrane potential of the neurons in the pathway planning area immediately dropped from -55mV to -62mV, being forcefully pulled away from the excitation threshold.

[0081] Scenario 2: Triggering the upper limit clamping mechanism: If the neuron in the pathway planning area receives a large amount of excitatory input, the membrane potential approaches the threshold, reaching... The theoretical suppression value calculated at this point is 9mV; if the membrane potential rises to -45mV due to other factors, the calculated value is 12mV. At this point, the clamping mechanism of the min function is triggered, and the actual output suppression amplitude is limited to the upper limit, i.e., -10mV.

[0082] Scenario 3: Neurons are in a resting or hyperpolarized state: If the neurons in the pathway planning area have not been recently stimulated or have just fired and are in a hyperpolarized state, for example... .because The max function takes the value 0. Only basic weak inhibition is applied.

[0083] When a certain direction is assessed as dangerous, the more active the path planning neurons in that direction are, the more likely they are to guide the drone towards that dangerous direction, and the stronger the inhibitory signal applied by the system; simultaneously, through To prevent deadlock caused by excessive suppression.

[0084] In an optional embodiment, the neurons in the network all employ a recent impulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their impulse firing activity within a preset time window, including:

[0085] Neuron excitation threshold Based on static benchmark threshold With adjustment amount Add them together to get;

[0086] The adjustment amount of the neuron after each firing pulse. Instantly increase the current value by a fixed value β;

[0087] During the pulseless firing period, the adjustment amount is based on the time constant. Exponential decay.

[0088] To prevent neurons from continuously firing impulses and to increase the network's adaptability to the environment, all neurons integrate a threshold function. A baseline threshold is used. Taking a neuron with a voltage of -50mV as an example, the initial adjustment amount of the neuron... The actual threshold is -50mV, set to 0. After the neuron fires a pulse at t=10ms, the adjustment amount... Increase immediately =5mV, making =5mV, actual excitation threshold It then drops to -45mV, becoming even more difficult to excite. If no new pulse is emitted after t=10ms, According to the time constant =200ms exponential decay. For example, at t=210ms, The value will decay to 1.84mV, at which point the actual threshold is approximately -48.16mV. This mechanism forms a negative feedback loop; the more frequently the neurons fire, the higher the excitation threshold, thus automatically adjusting the firing rate. This mechanism is particularly important in complex environments with dense obstacles, preventing the network output from "saturating" with a few safest headings and prompting the network to explore suboptimal but equally feasible headings, thereby improving the flexibility of UAV decision-making.

[0089] In some embodiments, the neurons in the threat assessment region are configured to generate rapid impulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, inhibit the activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle, thereby achieving rapid rejection of dangerous directions; the path planning region is responsible for assessing potential safe course, including:

[0090] The synaptic connections from the threat assessment region neuron to the corresponding output layer neuron at the topological location of the neuron are set to be strongly inhibitory;

[0091] The synaptic connection from the neuron in the path planning region to the output layer neuron corresponding to the topological position of the neuron is defined as excitability.

[0092] The connection weights from the hidden layer to the output layer reflect the integration of threat veto and path selection. For each azimuth angle... From the threat assessment region neurons To the output neuron The synaptic weight is fixed at -2.0 to ensure Any activity is related to This produces a strong inhibitory effect. And from the neurons in the path planning area... arrive The synaptic weight is +1.0, used to transmit the recommended safe course signal. The inhibition weight is -2.0, and the excitation weight is +1.0. The absolute value of the inhibition weight is twice that of the excitation weight, ensuring the absolute priority of threat veto. Furthermore, to form smooth and continuous course decisions, the neurons in the path planning region... According to the Mexican hat function rule, for topologically neighboring neurons... lateral excitation is generated. For sectors ≤2, the connection weight will vary with distance. The increase in the Gaussian decays the signal, resulting in a linearly decreasing inhibitory effect on distant neurons. This helps the network to generate some excitation in neighboring directions when it finds a safe direction, causing the output impulse activity to form a "hill"-shaped distribution centered on the optimal heading. This makes the decoded heading command vector more stable and smooth, avoiding high-frequency jitter between two almost equally safe headings.

[0093] S3, decode the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window, and generate a heading deflection command vector for controlling the flight of the UAV by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors.

[0094] Set a decision window with a fixed duration, such as 25ms, and use a counter to count the total number of pulses fired by each output layer neuron i within this window. A two-dimensional unit vector is pre-defined for each output neuron i. The direction corresponds to the center direction of the i-th azimuth sector, and the components of this vector are determined by the corresponding angle. The cosine and sine values ​​are determined; the total number of pulses for each output neuron is determined by executing a population vector encoding / decoding algorithm. With the corresponding unit vector Multiply the results of all N neurons and sum them as a vector to obtain the heading deflection command vector. .

[0095] In one embodiment, decoding the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window involves multiplying the preset heading deflection unit vector corresponding to each output neuron by the total number of pulses fired by the neuron within the window and then summing the results to generate a heading deflection command vector for controlling the UAV's flight.

[0096] Preset the corresponding heading deflection angle for each of the N output neurons. , where i=1,2,...,N, and the angles uniformly cover the range from -90 degrees to +90 degrees;

[0097] At the end of the 25ms decision window, the number of pulses fired by each output neuron i is counted. ;

[0098] Calculate the heading deflection command vector The heading deflection unit vector for all neurons Based on the number of pulses fired The vector sum of the weights, i.e. .

[0099] The generation of decision instructions is a vector voting process based on population coding. Assume the output layer consists of N=37 neurons, representing heading angles from -90 degrees to +90 degrees, with a step size of 5 degrees. Within a 25ms decision window, the network undergoes pulse dynamics evolution. At the end of the window, the total number of pulses fired by each output neuron is counted. For example, the result might be: the neuron representing the +5 degree direction fired 12 pulses. =12, meaning the neuron at 0 degrees fired 5 pulses. =5, indicating that the neuron in the -10 degree direction fired 2 pulses. =2. During decoding, the unit vector in each direction is multiplied by the number of pulses, and then summed. Heading deflection command vector. It will point in a direction slightly to the right and forward. The direction of the vector is used as the yaw angle command for the drone.

[0100] In a second embodiment, the present invention also provides an obstacle avoidance decision system for a drone control system, comprising the following modules:

[0101] The module is used to acquire obstacle point set data detected by the UAV's sensors. The data includes distance and relative speed information of each obstacle point. The data is then subjected to multi-channel temporal hybrid encoding to generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative speed information for each obstacle point, thus forming a composite input pulse cluster.

[0102] An adjustment module is used to input the composite input pulse cluster into a topologically ordered spiking neural network. The input, hidden, and output layer neurons of the spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. Neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, inhibit the activity of neurons in the path planning region and the output layer neurons at the corresponding azimuth angle, achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course directions. Neurons in the network all employ a recent pulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their pulse firing activity within a preset time window.

[0103] The generation module is used to decode the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window. It generates a heading deflection command vector for controlling the flight of the UAV by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors.

[0104] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0105] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for obstacle avoidance decision-making in an unmanned aerial vehicle (UAV) control system, characterized in that, Includes the following steps: Acquire obstacle point set data detected by UAV sensors, the data including distance and relative velocity information of each obstacle point; perform multi-channel temporal hybrid encoding on the data, and synchronously generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, forming a composite input pulse cluster; The composite input pulse cluster is fed into a topologically ordered spiking neural network. The neurons in the input, hidden, and output layers of this spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. Neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed. Through one-to-one strong inhibitory synaptic connections, they inhibit the activity of neurons in the path planning region and the output layer neurons at the corresponding azimuth angle, achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course directions. All neurons in the network employ a recent pulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their pulse firing activity within a preset time window. The number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window is decoded. By multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors, a heading deflection command vector for controlling the flight of the UAV is generated.

2. The method according to claim 1, characterized in that, The process of performing multi-channel temporal hybrid encoding on the data, synchronously generating a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, includes: The distance D of the detected obstacle is calculated using the formula... Convert to first pulse delivery delay ,in For encoding time window, This represents the maximum detection range of the drone's sensors. when and At that time, the absolute value of the detected relative velocity V is taken as the effective threat velocity. Divide by the maximum relative speed that the sensor can measure. Then multiply the resulting quotient by the maximum firing frequency of the neurons. The pulse firing frequency F is calculated. like or When F = 0Hz, then F = 0Hz.

3. The method according to claim 1, characterized in that, The step of configuring neurons within the threat assessment region to generate rapid impulse responses to immediate threats represented by proximity or high relative velocity includes: The resting excitation threshold of neurons in the threat assessment area is set to be 2mV lower than that of neurons in the path planning area; The initial weights of synaptic connections from input layer neurons to threat assessment region neurons are set to 1.5 times the initial weights of synaptic connections to path planning region neurons.

4. The method according to claim 1, characterized in that, The method of inhibiting the activity of neurons in the path planning region and output layer neurons at the corresponding azimuth angle through one-to-one strong inhibitory synaptic connections includes: When any neuron in the threat assessment region fires a pulse, an inhibitory postsynaptic potential change is applied to the neurons corresponding to the path planning region and the output layer topological position via a pre-set one-to-one inhibitory synapse. The amplitude of this change is dynamically calculated using the following formula and clamped to a non-positive value: Where k>0 is the inhibition coefficient, To suppress the upper limit of amplitude, The current membrane potential of the inhibited neuron. This is the resting potential of the neuron. It is a positive fundamental suppression constant.

5. The method according to claim 1, characterized in that, The neurons in the network all employ a recent impulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their impulse firing activity within a preset time window, including: Neuron excitation threshold Based on static benchmark threshold With adjustment amount Add them together to get; The adjustment amount of the neuron after each firing pulse. Instantly increase the current value by a fixed value β; During the pulseless firing period, the adjustment amount is based on the time constant. Exponential decay.

6. The method according to claim 1, characterized in that, The decoding of the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window involves multiplying the preset heading deflection unit vector corresponding to each output neuron by the total number of pulses fired by the neuron within the window, and then summing the results to generate a heading deflection command vector for controlling the UAV's flight. This includes: Preset the corresponding heading deflection angle for each of the N output neurons. , where i=1,2,...,N, and the angles uniformly cover the range from -90 degrees to +90 degrees; At the end of the 25ms decision window, the number of pulses fired by each output neuron i is counted. ; Calculate the heading deflection command vector The heading deflection unit vector for all neurons Based on the number of pulses fired The vector sum of the weights, i.e. .

7. The method according to claim 1, characterized in that, The neurons in the threat assessment area are configured to generate rapid impulse responses to immediate threats represented by close range or high relative speed, and the activity of the path planning area neurons and output layer neurons in the corresponding azimuth angle is suppressed through one-to-one strong inhibitory synaptic connections, thereby achieving rapid rejection of dangerous directions. The path planning area is responsible for assessing potential safe course, including: The synaptic connections from the threat assessment region neuron to the corresponding output layer neuron at the topological location of the neuron are set to be strongly inhibitory; The synaptic connection from the neuron in the path planning region to the output layer neuron corresponding to the topological position of the neuron is defined as excitability.

8. An obstacle avoidance decision-making system for an unmanned aerial vehicle (UAV) control system, characterized in that, Includes the following modules: The module is used to acquire obstacle point set data detected by the UAV's sensors. The data includes distance and relative speed information of each obstacle point. The data is then subjected to multi-channel temporal hybrid encoding to generate a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative speed information for each obstacle point, thus forming a composite input pulse cluster. An adjustment module is used to input the composite input pulse cluster into a topologically ordered spiking neural network. The input, hidden, and output layer neurons of the spiking neural network are all topologically mapped according to azimuth angles. The hidden layer is divided into a threat assessment region and a path planning region in parallel. Neurons in the threat assessment region are configured to generate rapid pulse responses to immediate threats represented by close range or high relative speed, and through one-to-one strong inhibitory synaptic connections, inhibit the activity of neurons in the path planning region and the output layer neurons at the corresponding azimuth angle, achieving rapid rejection of dangerous directions. The path planning region is responsible for assessing potential safe course directions. Neurons in the network all employ a recent pulse firing history-dependent threshold mechanism, adjusting their excitation threshold based on their pulse firing activity within a preset time window. The generation module is used to decode the number of pulses fired by the output layer neurons of the spiking neural network within a preset decision window. It generates a heading deflection command vector for controlling the flight of the UAV by multiplying the preset heading deflection unit vector corresponding to each output neuron with the total number of pulses fired by the neuron within the window and then summing the vectors.

9. The system according to claim 8, characterized in that, The process of performing multi-channel temporal hybrid encoding on the data, synchronously generating a first pulse emission delay encoding sequence based on distance information and a pulse frequency encoding sequence based on relative velocity information for each obstacle point, includes: The distance D of the detected obstacle is calculated using the formula... Convert to first pulse delivery delay ,in For encoding time window, This represents the maximum detection range of the drone's sensors. when and At that time, the absolute value of the detected relative velocity V is taken as the effective threat velocity. Divide by the maximum relative speed that the sensor can measure. Then multiply the resulting quotient by the maximum firing frequency of the neurons. The pulse firing frequency F is calculated. like or When F = 0Hz, then F = 0Hz.

10. The system according to claim 8, characterized in that, The step of configuring neurons within the threat assessment region to generate rapid impulse responses to immediate threats represented by proximity or high relative velocity includes: The resting excitation threshold of neurons in the threat assessment area is set to be 2mV lower than that of neurons in the path planning area; The initial weights of synaptic connections from input layer neurons to threat assessment region neurons are set to 1.5 times the initial weights of synaptic connections to path planning region neurons.

Images