An aircraft decision recognition method based on multi-wavelet dynamic modeling and pulse neural network

Abstract

A method for aircraft decision recognition based on multi-wavelet dynamic modeling and spiking neural networks is proposed. This method acquires multi-source sensor data, flight trajectory data, and corresponding decision command data. Preprocessing is performed to improve the consistency of the multi-source sensor data. A multi-source time-varying dynamic model for aircraft trajectory prediction is constructed using time-varying autoregressive theory of exogenous inputs. Multi-wavelet basis function expansion is performed on the parameters of the multi-source time-varying dynamic model to obtain the multi-wavelet time-varying dynamic model. A sparse multi-wavelet time-varying optimized model is obtained. A model parameter estimation and identification method is designed, and the model is reconstructed based on the identified parameters. Time-domain and frequency-domain information is extracted from the identified model parameters and used as input to the subsequent intelligent decision model. A control command generation and intelligent decision model is constructed, with key flight state information as input and decision command data as output. After training, the control command generation and intelligent decision model is obtained.

Description

Technical Field

[0001] This invention relates to a decision recognition technology for complex systems oriented towards multi-source information processing, belonging to the field of intelligent information processing technology. Background Technology

[0002] The large-scale, multi-type, multi-source heterogeneous state information and external environmental information generated during the operation of complex systems can effectively characterize the system's own and environmental conditions during operation, providing comprehensive and scientific data support for system control capability assessment, intelligent decision-making and control, and improved design. However, due to data uncertainties such as missing data, low signal-to-noise ratio, and poor consistency often encountered during the acquisition of system experimental data, current computing systems have low utilization rates of comprehensive information, making it difficult to fully extract effective information from large-scale, multi-source heterogeneous experimental data for state assessment and planning decisions of complex systems. Therefore, researching effective system identification and decision recognition methods is crucial for improving the reliability of intelligent decision-making in complex systems. Summary of the Invention

[0003] The technical problem solved by this invention is to provide a method for decision recognition of complex systems with multi-source heterogeneous sensor information processing and intelligent autonomous decision-making under specific task requirements.

[0004] The technical solution of this invention is: an aircraft decision-making and recognition method based on multi-wavelet dynamic modeling and spiking neural networks, comprising:

[0005] Based on the characteristics of the aircraft and the requirements of the flight mission, and according to the set flight parameters, conduct simulation or real flight tests to obtain multi-source sensor data, flight trajectory data and corresponding decision command data;

[0006] Preprocess the acquired multi-source sensor data to improve the consistency of the multi-source sensor data;

[0007] Using the time-varying autoregressive theory of exogenous input, a multi-source data time-varying dynamic model for aircraft trajectory prediction is constructed. The input of the model is preprocessed multi-source sensor data, and the output is flight trajectory data, which is obtained through training.

[0008] Multi-wavelet basis function expansion is performed on the parameters of the time-varying dynamic model for multi-source data to obtain the multi-wavelet time-varying dynamic model;

[0009] We design a model structure selection method for multi-wavelet time-varying dynamics models to obtain a sparse multi-wavelet time-varying optimization model.

[0010] For sparse multi-wavelet time-varying optimization models, a model parameter estimation and identification method is designed, and the model is reconstructed based on the identified parameters.

[0011] Using the reconstructed time-varying model, the time-domain and frequency-domain information is extracted from the identified model parameters as key information for flight status, and then used as input for the subsequent intelligent decision-making model.

[0012] A control command generation and intelligent decision-making model based on a spiking neural network is constructed. Its input is key information about the flight status, and its output is decision command data. The model is trained to obtain a trained control command generation and intelligent decision-making model.

[0013] Preferably, the multi-source sensor data of the aircraft includes flight status data and external environment data; the flight status data includes velocity, position, azimuth angle, pitch angle, and flight status modal parameters; the external environment data includes aerodynamic forces, temperature, and external pressure.

[0014] Preferably, the preprocessing includes sampling interval and data format cleaning and sequence alignment.

[0015] Preferably, the multiple wavelet basis functions are B-spline wavelet basis functions.

[0016] Preferably, the model structure selection method involves selecting important model terms from a multi-wavelet time-varying dynamics model to obtain a sparse multi-wavelet time-varying optimization model composed of important model terms. The method includes:

[0017] Step 1: Assume the first selection of important model items The candidate set is V1 = v i (t), i = 1, 2, ..., H, V1 are all model terms in the multi-wavelet time-varying dynamics model; the flight trajectory data are the output measurements, represented as Y = y(t), t = 1, 2, ..., T, where Y is the set of output measurements for the flight trajectory data. t represents the time index, P(y(t),v i (t) represents y(t) and v i The joint probability density functions of (t), P(y(t)) and P(v) i (t) represent y(t) and v respectively. i (t) marginal probability density function, and

[0018]

[0019] Therefore, the first important model term searched is make

[0020] Step 2: Select the remaining model items as the second most important model items. The candidate set V2 = vi (t), i = 1, 2, ..., M, i ≠ l1, let get

[0021]

[0022] The second most important item in the search is And there are

[0023] This process continues until the search process termination criteria are met.

[0024] Preferably, the signal error ratio is used as an evaluation criterion for terminating a single search process at each individual step;

[0025]

[0026] The formula for calculating the error signal ratio in the second individual step is:

[0027]

[0028] When ||r1(t)|| 2 <ρ or ||r ii (t)|| 2 If <ρ, then the current step is the last search step, where ρ is the set notification tolerance.

[0029] Preferably, the search process termination criterion is: satisfying Or M s =M; where M s M is the cutoff number for the search, with each search counted as one step. M is the maximum number of search steps set by default.

[0030] Preferably, the control command generation and intelligent decision-making model based on spiking neural networks includes an adaptive pulse encoder, a feature extractor, and a pulse decoder;

[0031] The adaptive pulse encoder is an adaptive pulse coding layer based on a neural network to globally generate pulse signals;

[0032] The feature extractor consists of one or more LIF neuron layers and is used to extract spatiotemporal feature information of pulse sequences.

[0033] The pulse decoder outputs identification and classification results based on the spatiotemporal characteristics of the pulse sequence, and the classification results are the decision control instructions.

[0034] Preferably, each neuron in the neural network-based adaptive pulse coding layer first receives multiple real values ​​as input, then processes the information through one or more ANN layers, and finally generates a pulse sequence as output through the LIF neuron layer to globally generate a pulse signal.

[0035] Preferably, the input is in ANN mode and the output is in SNN mode.

[0036] Compared with the prior art, the advantages of this invention are as follows: This invention utilizes system identification and intelligent information processing methods to fully explore the correlation between multi-source sensor data and decision commands in complex systems, constructs a system identification dynamic analysis model for extracting knowledge features of multi-source information and a decision recognition model based on a novel spiking neural network, accurately detects the potential relationship features between multi-source information and further identifies the correlation patterns between them and decision commands, thereby achieving real-time and accurate recognition of decision commands under complex multi-source information.

[0037] (1) This invention conducts in-depth research on aircraft control systems and cutting-edge intelligent information processing methods, and develops a new, stable, reliable, and effective intelligent control method for aircraft that can cope with complex aerospace environments. It has important guiding significance and application value for ensuring that aircraft can accurately complete complex flight missions.

[0038] (2) This invention employs a multi-wavelet method, which possesses excellent approximation capabilities and good local features, to model and solve different types of signals. Complex time-varying parameters are expanded using a limited number of multi-wavelet basis functions, transforming the time-varying characteristics of the initial model into the time-invariant characteristics of the multi-wavelet basis functions. This significantly reduces the number of parameters to be determined and improves the model's generalization performance. Simultaneously, by adjusting the scaling parameters of the multi-wavelet functions, different types of time-varying model parameters can be accurately identified and tracked, thereby quickly and accurately tracking changes in model parameters. This provides a crucial foundation for accurately extracting key time-frequency information of the time-varying model during flight.

[0039] (3) To address the real-time requirements for generating decision-making commands during aircraft flight, this invention combines the working mechanism of SNN (Synthetic Neural Network), which is closer to that of the human brain's neural network, with its low power consumption and high performance characteristics. It proposes to construct an SNN-based control command generation and intelligent decision-making model, learning the potential relationship between key information in the flight state and decision-making control commands to achieve data-driven intelligent decision-making. This project conducts in-depth research on the intelligent control system of aircraft from a data-driven perspective, providing new ideas for exploring novel intelligent flight control system design methods.

[0040] This invention delves into the research of aircraft control systems and cutting-edge intelligent information processing methods, and develops a new, stable, reliable, and effective intelligent control method for aircraft that can cope with complex aerospace environments. It has important guiding significance and application value for ensuring that aircraft accurately complete complex flight missions. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 Neural Network-Based Adaptive Pulse Code Layer

[0043] Figure 3 Simplified diagrams of convolution and pooling operations in spiking neural networks: (a) pulse-based convolution operation; (b) pulse-based average pooling operation.

[0044] Figure 4 Behavioral patterns of neurons in the output layer of a spiking neural network: (a) pulse firing frequency; (b) cumulative membrane potential of neurons. Detailed Implementation

[0045] The following is in conjunction with the appendix Figure 1-4 The invention will be described in detail with examples.

[0046] like Figure 1 The diagram shown is a flowchart of the method of the present invention, and the specific steps are as follows:

[0047] (1) Acquisition and preprocessing of multi-source state data of aircraft

[0048] This invention aims to monitor and control the aircraft's state by acquiring the potential correlation between state information and decision commands within the flight system. Therefore, acquiring multi-source information and standardizing preprocessing methods are primary prerequisites. First, based on the aircraft's characteristics and flight mission requirements, and given flight parameters, simulation or real flight tests are conducted to acquire multi-source sensor data such as flight speed, acceleration, position, and angular velocity, as well as flight trajectory data and corresponding decision commands. The standardization preprocessing technology for flight data mainly achieves format cleaning and sequence alignment of the multi-source sensor data, resolving issues such as poor consistency in sampling intervals and data formats that may exist in multi-source data.

[0049] (2) Construction of a time-varying dynamic system model for aircraft state information

[0050] This time-varying autoregressive with exogenous (TVARX) model characterizes the multi-source data time-varying dynamics model for spacecraft trajectory prediction. The TVARX model is a complex polynomial with linear parameters. The TVARX model has an output feedback term, which adds a significant amount of model information, greatly improving the accuracy and precision of the TVARX model structure.

[0051] Let the flight trajectory information of the aircraft be Y = {y(t)}, t = 1, 2, ..., T, where T is the data length. The N sensor data points of the aircraft are X1 = {x1(t)}, X2 = {x2(t)}, ..., X... N ={x N The multi-source data time-varying dynamic system model for spacecraft orbit prediction, t = 1, 2, ..., T, can be represented as:

[0052]

[0053] Wherein, kernel function a n (t, τ) describes the nth sensing data x n The linear relationship between the output y(t) and the kernel function b(t, τ) represents the time-varying autoregressive model coefficients of the output. These coefficients are functions of time t and the time lag τ between the current time and historical time points. n b(t, τ) and b(t, τ) are the model parameters that need to be estimated in the model; N represents the number of input sensing signals, and M represents the model order of the feedforward process. y This indicates the model order of the feedback process.

[0054] (3) Perform multi-wavelet basis function expansion on the TVARX model parameters to obtain the multi-wavelet time-varying dynamic model;

[0055] Wavelet theory proves that a square-integrable scalar signal f can be approximated by multi-resolution wavelet basis function decomposition, as shown below:

[0056]

[0057] in and ψ j，k (x)=2 j / 2 ψ(2 j x(k), j0, j, k∈z (Z represents integers) are the translation and expansion forms of the scaling function φ(x) and the mother wavelet ψ(x). and β j，k This represents the wavelet expansion coefficients. Furthermore, when the scaling function... When the resolution scale is large enough, i.e., there exists an integer J, then equation (2) can be simplified to:

[0058]

[0059] Therefore, a set of multi-wavelet basis functions {π} with tight support and good generalization ability is selected. μ (t): μ = 1, 2, ..., L} for the model parameter a to be estimated in equation (1) n Expand (t, τ) and b(t, τ):

[0060]

[0061] Where L represents the maximum number of wavelet basis functions. Therefore, the time-varying dynamic model of wavelets can be expressed as:

[0062]

[0063] Furthermore, the above formula can be written in the form of linear parametric regression:

[0064]

[0065] Where, λ i It is the set of parameters to be estimated, v i (t) is the expanded model term in (5), and H is the total number of candidate terms in the model term based on the multi-wavelet basis function expansion. Thus, the problem of estimating the coefficients of the time-varying autoregressive model is transformed into the problem of estimating the time-invariant coefficients based on the basis functions.

[0066] (4) Design a model structure selection method for multi-wavelet time-varying dynamics models to obtain a sparse multi-wavelet time-varying optimization model with low computational complexity and high model generalization.

[0067] The initial multi-wavelet time-varying dynamics model obtained during the above model construction process, due to its complex model structure, cannot well reveal the time-varying relationship between a large amount of multi-source data related to flight states and flight trajectory prediction information. In order to ensure that the proposed modeling algorithm can quickly and accurately obtain a simple and effective model structure, this section selects the model structure for the original model.

[0068] Model structure selection is a crucial step in the system identification process. It involves choosing a few meaningful and important candidate or regression terms from a set of candidate model terms containing a large number of redundant terms to generate a simplified model. For example, the multi-wavelet basis function expansion model described by equation (5) may involve a large number of candidate terms. Furthermore, the candidate set may contain a large number of highly linearly correlated terms, which could lead to overfitting of the identification model, making it impossible to obtain reliable identification results using the maximum likelihood method. Therefore, determining which regression factors should be included in the final model is extremely important.

[0069] For any two time series data X and Y, P(x) and P(y) represent the marginal probability density functions, respectively, and P(x,y) represents the joint probability density function. Therefore, the mutual information I(x,y) between two time series data can be defined as:

[0070]

[0071] For the multi-wavelet basis function expansion for multi-source data of aircraft constructed in formula (5), the process of candidate selection using the proposed FOR-MI algorithm is summarized as follows:

[0072] Step 1: Step 1: Assume the first selection of important model items. The candidate set is V1 = v i (t), i = 1, 2, ..., H, V1 are all model terms in the multi-wavelet time-varying dynamics model; the flight trajectory data are the output measurements, represented as Y = y(t), t = 1, 2, ..., T, where Y is the set of output measurements for the flight trajectory data. t represents the time index, P(y(t),v i (t) represents y(t) and v i The joint probability density functions of (t), P(y(t)) and P(v) i (t) represent y(t) and v respectively. i The marginal probability density function of (t), and

[0073]

[0074] The first important model term is And there are

[0075] The error-to-signal ratio (ESR) is used as a criterion to terminate the search process.

[0076]

[0077] If ||r1(t)|| 2 If <ρ, then the search in this step ends, and the sparse multi-wavelet time-varying optimization model consists of the first important model term.

[0078] Step 2: Select the remaining model items as the second most important model items. The candidate set V2 = v i (t), i = 1, 2, ..., M, i ≠ l1, let We can get

[0079]

[0080] The second most important item is And there are ESR is

[0081]

[0082] If ||r2(t)|| 2 If <ρ, then the search in this step ends. The sparse multi-wavelet time-varying optimization model consists of the first important model term and the second important model term.

[0083] Step 3: The selection of subsequent important model terms is carried out in the same way, up to the Mth term. s Step cutoff (M) s satisfy Or M s =M, where ρ is the set notification tolerance and M is the preset maximum number of search steps.

[0084] Thus, M is obtained. s A meaningful subset of variables or candidate options.

[0085] (5) For sparse multi-wavelet time-varying optimization models, design model parameter estimation and identification methods, and reconstruct the model based on the identified parameters;

[0086] Based on the recorded output trajectory prediction y and a few selected meaningful model terms Estimation results of time-invariant expansion model coefficients It can be obtained through the maximum likelihood estimation method. Furthermore, the kernel function of the time-varying model can be reconstructed according to formula (4), and then the representation model can be reconstructed according to formula (1).

[0087] Based on the characteristics of time-varying model parameters, time-domain and frequency-domain information from the aircraft flight state characterization model is extracted as flight state information and fed into the intelligent decision-making model.

[0088] (6) Control command generation and intelligent decision-making model construction based on spiking neural network

[0089] To address the characteristics of multi-wavelet time-varying nonlinear systems in aircraft state data, an intelligent decision-making model based on a spiking neural network (SNN) is constructed to achieve flight state recognition and control command generation. Compared to traditional perceptron neural networks, the neurons in the SNN use pulse firing sequences for information processing and employ pulse time encoding to encode input variables as pulse firing times. To simulate the information processing mode of biological neurons, the SNN uses nonlinear exponential family functions instead of the traditional perceptron neuron processing mode, which can arbitrarily approximate nonlinear functions and simulate biological neural network signals. This mechanism greatly improves the computational power of neurons. Each sample is fed into the SNN after pulse encoding. Compared to traditional neuron models, the spiking neuron model has the advantages of lower power consumption and is more suitable for parallel computing. The proposed method for generating control commands and constructing an intelligent decision-making model based on a spiking neural network is as follows:

[0090] a. Adaptive pulse encoder

[0091] Spiking neural networks simulate biological neurons receiving pulse sequences as input. Different input data often require different encoding schemes to encode external stimuli into discrete pulse sequences. Related research indicates that using encoding methods independent of the spiking neural network model can lead to a loss of some original information features and an increase in simulation cycle time, resulting in gradient vanishing. To address these issues, for traditional non-pulse stream input data, an adaptive pulse coding layer based on a neural network is added to globally generate pulse signals, such as... Figure 2 As shown. Each neuron in this layer first receives multiple real values ​​as input (ANN working mode), then the information is processed through one or more ANN layers, and finally a pulse sequence is generated as output through the LIF neuron layer (SNN working mode).

[0092] b. Feature Extractor

[0093] Once the input data is encoded as a pulse sequence, a spiking neural network can be constructed in a manner similar to that used in constructing traditional deep neural network models. The feature extraction layer defines one or more LIF neuron layers to extract the spatiotemporal feature information of the pulse sequence. The connections between different neuron layers can be fully connected, convolutional, or recurrent, and operations such as max pooling and average pooling can be used between different layers to reduce the feature dimensionality. Figure 3This paper describes two-dimensional convolution and average pooling operations performed by LIF neurons over three simulation steps. For the pulse-driven convolution operation, in each simulation cycle, the input pulse sequence is convolved with the weight kernel, generating an input current that accumulates postsynaptic membrane potential through the LIF spiking neuron layer. Once the membrane potential exceeds a threshold, the neuron fires a pulse sequence while simultaneously resetting the membrane potential; otherwise, the membrane potential leaks and continues to propagate to the next simulation cycle. For the pulse-driven average pooling operation, the average membrane potential of the neuron is obtained using fixed weight kernel parameters, and a pulse sequence is fired, reducing the feature dimensionality while preserving spatial topological information.

[0094] c. Pulse Decoder

[0095] The function of the decoder is to decode the signals generated by the output layer spiking neurons according to the actual task. Figure 4 The firing frequency and membrane potential of the output layer spiking neurons are shown respectively. The higher the pulse firing frequency, the greater the degree of response to different categories, such as... Figure 4 As shown in (a), ideally, neurons in the output layer corresponding to the true category fire the highest frequency of impulses, while the remaining neurons are mostly in a resting state. Correspondingly, in Figure 4 (b) shows the change in the membrane potential of the output neuron during the simulation period.

[0096] In summary, the spiking neural network model with an encoding-feature extraction-decoding hierarchical structure designed in this section takes the form of: first, the input data is processed by an encoder to generate a pulse sequence; then, a feature extractor in an intermediate layer fully extracts the spatiotemporal features of the pulse sequence; and finally, a decoder designed according to the task requirements outputs the final identification and classification result. It is worth noting that the network parameters, such as those for convolutional layers, pooling layers, and fully connected layers, are shared across all simulation cycles.

[0097] (7) Conduct experimental verification and performance evaluation of the model.

[0098] The parts of this invention not described in detail are common knowledge to those skilled in the art.

Claims

1. A method for aircraft decision recognition based on multi-wavelet dynamic modeling and spiking neural networks, characterized in that... include: Based on the characteristics of the aircraft and the requirements of the flight mission, and according to the set flight parameters, conduct simulation or real flight tests to obtain multi-source sensor data, flight trajectory data and corresponding decision command data; Preprocess the acquired multi-source sensor data to improve the consistency of the multi-source sensor data; Using the time-varying autoregressive theory of exogenous input, a multi-source data time-varying dynamic model for aircraft trajectory prediction is constructed. The input of the model is preprocessed multi-source sensor data, and the output is flight trajectory data, which is obtained through training. Multi-wavelet basis function expansion is performed on the parameters of the time-varying dynamic model for multi-source data to obtain the multi-wavelet time-varying dynamic model; We design a model structure selection method for multi-wavelet time-varying dynamics models to obtain a sparse multi-wavelet time-varying optimization model. For sparse multi-wavelet time-varying optimization models, a model parameter estimation and identification method is designed, and the model is reconstructed based on the identified parameters. Using the reconstructed time-varying model, the time-domain and frequency-domain information is extracted from the identified model parameters as key information for flight status, and then used as input for the subsequent intelligent decision-making model. A control command generation and intelligent decision-making model based on a spiking neural network is constructed. Its input is key information about the flight status, and its output is decision command data. The model is trained to obtain a trained control command generation and intelligent decision-making model.

2. The method according to claim 1, characterized in that: The multi-source sensor data of the aircraft includes flight status data and external environment data; the flight status data includes velocity, position, azimuth angle, pitch angle, and flight status modal parameters; the external environment data includes aerodynamic forces, temperature, and external pressure.

3. The method according to claim 1, characterized in that: The preprocessing includes sampling interval and data format cleaning and sequence alignment.

4. The method according to claim 1, characterized in that: The aforementioned multiple wavelet basis functions employ B-spline wavelet basis functions.

5. The method according to claim 1, characterized in that: The model structure selection method involves selecting important model terms from a multi-wavelet time-varying dynamics model to obtain a sparse multi-wavelet time-varying optimization model composed of important model terms. The method includes: Step 1: Assume the first selection of important model items The candidate set is V1 = v i (t), i = 1, 2, ..., H, V1 are all model terms in the multi-wavelet time-varying dynamics model; the flight trajectory data are the output measurements, represented as Y = y(t), t = 1, 2, ..., T, where Y is the set of output measurements for the flight trajectory data. t represents the time index, P(y(t),v i (t) represents y(t) and v i The joint probability density functions of (t), P(y(t)) and P(v) i (t) represent y(t) and v respectively. i (t) marginal probability density function, and Therefore, the first important model term searched is make Step 2: Select the remaining model items as the second most important model items. The candidate set V2 = v i (t), i = 1, 2, ..., M, i ≠ l1, let get The second most important item in the search is And there are This process continues until the search process termination criteria are met.

6. The method according to claim 5, characterized in that: The signal error ratio is used as an evaluation criterion for terminating a single search process at each individual step. The formula for calculating the error signal ratio in the second individual step is: When ||r1(t)|| 2 <ρ or ||r ii (t)|| 2 If <ρ, then the current step is the last search step, where ρ is the set notification tolerance.

7. The method according to claim 5, characterized in that: The search process termination criterion is: satisfied. Or M s =M; where M s M is the cutoff number for the search, with each search counted as one step. M is the maximum number of search steps set by default.

8. The method according to claim 1, characterized in that: The control command generation and intelligent decision-making model based on spiking neural network includes an adaptive pulse encoder, a feature extractor, and a pulse decoder. The adaptive pulse encoder is an adaptive pulse coding layer based on a neural network to globally generate pulse signals; The feature extractor consists of one or more LIF neuron layers and is used to extract spatiotemporal feature information of pulse sequences. The pulse decoder outputs identification and classification results based on the spatiotemporal characteristics of the pulse sequence, and the classification results are the decision control instructions.

9. The method according to claim 8, characterized in that: Each neuron in the neural network-based adaptive pulse coding layer first receives multiple real values ​​as input, then the information is processed through one or more ANN layers, and finally a pulse sequence is generated as output through the LIF neuron layer to generate a pulse signal globally.

10. The method according to claim 9, characterized in that: The input is in ANN mode; the output is in SNN mode.

Images