Decision-making method for hybrid drive of existing parameters

Abstract

The present invention discloses a decision-making method for existing parameter hybrid drive, specifically relating to the field of simulated confrontation, and is used to solve the problem of group coordination failure caused by dynamic conflicts between visual and text parameters. When processing the fusion of high-frequency visual parameters and low-frequency text parameters, adaptive downsampling and phase compensation are used to reduce the feature cancellation effect in nonlinear fusion; at the same time, based on the joint analysis of time-frequency energy distribution and causal correlation strength, an inhibition weight matrix is constructed to weaken policy contradictions, strengthen the policy causal path, and ensure the coordination and consistency of decision instructions; in addition, through the collaboration of hierarchical neural networks and spatiotemporal gating mechanisms, the dynamic fusion of action sequences and policy constraints is achieved, and the response speed and policy consistency are improved; finally, with the help of damping control, the nonlinear oscillation of the instruction stream is suppressed in real time to prevent the group policy from deviating, thereby optimizing the overall decision quality.

Description

Technical Field

[0001] The present invention relates to the field of simulated confrontation, and more particularly, to a decision-making method driven by existing parameter mixtures. Background Art

[0002] The dynamic conflict of multimodal decision parameters during the nonlinear fusion process leads to the failure of group coordination. In adversarial simulation scenarios, the defensive policy parameters generated by the visual perception module based on real-time environmental features and the aggressive policy parameters generated by the text semantic module by parsing historical instructions, when fused through a conventional gating network, produce unexpected feature cancellation effects due to the essential differences in the dynamic characteristics of the parameters. The technical mechanism is as follows: the visual parameters are driven by local environmental changes and have high-frequency fine-tuning characteristics, while the text parameters are constrained by the global policy and exhibit low-frequency steady-state characteristics. The two form a phase misalignment during the temporal evolution process; when the gating network uses a fixed nonlinear function (such as sigmoid) for parameter fusion, the gradient update directions of the high-frequency and low-frequency parameters produce adversarial interference within a specific time window, causing the fused decision vector to continue to oscillate in the policy space. This problem is exponentially amplified in multi-agent collaborative scenarios: parameter oscillations of a single agent propagate in the group through the strategy coupling graph, forming a positive feedback loop, causing the group strategy to deviate from the equilibrium state; the decision-making system needs to repeatedly adjust parameters to correct the deviation, resulting in the exhaustion of computing resources and a surge in response delays, ultimately paralyzing the large-scale collaborative decision-making system.

[0003] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a decision-making method for existing parameter hybrid drive, which reduces the feature cancellation effect in nonlinear fusion through adaptive downsampling and phase compensation; at the same time, based on the joint analysis of time-frequency energy distribution and causal correlation strength, a suppression weight matrix is constructed to weaken policy contradictions, strengthen the policy causal path, and ensure the coordination and consistency of decision instructions; in addition, through the collaboration of hierarchical neural networks and spatiotemporal gating mechanisms, the dynamic fusion of action sequences and policy constraints is realized, and the response speed and policy consistency are improved; finally, with the help of damping control, the nonlinear oscillation of the instruction flow is suppressed in real time to prevent the group policy from deviating, thereby optimizing the overall decision quality to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] S1: Based on the detection of fluctuation characteristics of visual parameter streams and the recognition of policy instruction events in text parameter streams, the dynamic time warping algorithm is used to align the key conflict intervals of multimodal parameters and generate a non-uniformly distributed set of conflict period coordinates;

[0007] S2: Within the designated conflict period coordinates, event-driven adaptive downsampling is performed on high-frequency visual parameters to align their time base with the text parameters. Time-frequency analysis techniques are then used to extract phase compensation vectors to eliminate inter-modal phase differences.

[0008] S3: Conduct a joint analysis of the time-frequency energy distribution and causal correlation strength of the synchronized parameters, construct a counterfactual intervention model to generate a frequency band suppression weight matrix, and specifically weaken the frequency band energy that causes strategic contradictions and strengthen the strategic causal path;

[0009] S4: De-conflicting visual parameters are input into the tactical layer spiking neural network to generate millisecond-level action sequences. Text parameters are input into the policy layer memory-augmented network to generate a constraint framework. Multi-granularity command fusion is achieved through a spatiotemporal gating mechanism.

[0010] S5: Monitor the nonlinear oscillation characteristics of the fused instruction stream in real time, construct a damping controller based on Lyapunov stability theory, and inject a negative feedback correction signal to maintain the convergence of the group strategy when chaotic edge characteristics are detected.

[0011] In a preferred embodiment, step S1 includes the following contents:

[0012] First, the fluctuation feature mutation point is detected by calculating the local fluctuation entropy of the visual parameter flow. The probability distribution of the change amplitude of the visual parameter flow is statistically analyzed in each time window and the information entropy is calculated. When the local fluctuation entropy exceeds the preset entropy threshold, the central time point of the corresponding time window is marked as the visual mutation point.

[0013] Then, the exponential smoothing cumulative deviation method is used to identify the strategic instruction events of the text parameter stream. Exponential smoothing is applied to the text parameter stream to reduce short-term fluctuations. The dynamic deviation between the smoothed parameter and the moving average is calculated. When the dynamic deviation exceeds the preset deviation threshold, the corresponding time point is marked as a strategic instruction event point.

[0014] Then, the dynamic time warping algorithm is used to align the visual mutation point set with the strategy instruction event point set, construct a cumulative distance matrix, and solve the minimum cumulative distance path through dynamic programming to generate a set of aligned time point pairs.

[0015] Finally, a conflict period coordinate set is generated based on the set of aligned time point pairs. A conflict period is defined for each pair of aligned time points and then expanded and merged to form a non-uniformly distributed conflict period coordinate set.

[0016] In a preferred embodiment, step S2 includes the following:

[0017] During each conflict period, the local change rate of the high-frequency visual parameter stream is first calculated to identify the event trigger point. When the local change rate exceeds the preset change rate threshold, the corresponding time point is marked as the event trigger point.

[0018] Then, adaptive downsampling is performed based on the event trigger point, retaining the parameter value at the event trigger point, and linear interpolation is used to generate approximate values for the non-event trigger point, so that the number of sampling points of the downsampled high-frequency visual parameter stream in the conflict period is consistent with the number of sampling points of the low-frequency text parameter stream in the same conflict period, achieving time reference alignment.

[0019] In a preferred embodiment, step S2 further includes the following:

[0020] Then, continuous wavelet transform is performed on the downsampled high-frequency visual parameter stream and low-frequency text parameter stream respectively to generate their respective time-frequency spectra, and the phase difference between the two at each time point and scale is calculated. A phase compensation vector is constructed based on the phase difference, and the negative value of the phase difference at the dominant scale at each time point is taken as the compensation phase.

[0021] Finally, a phase rotation adjustment is applied to the downsampled high-frequency visual parameter stream, and the complex signal at each time point is multiplied by the complex exponential term, where the phase angle of the complex exponential term is the compensation phase value, and the real part is taken to generate a synchronous parameter stream, thereby achieving alignment with the low-frequency text parameter stream in time and phase.

[0022] In a preferred embodiment, step S3 includes the following contents:

[0023] Perform wavelet packet decomposition on the synchronized visual parameter stream to generate a time-frequency energy matrix, and perform short-time Fourier transform on the text parameter stream to generate a main frequency energy feature matrix;

[0024] The cross-coherence matrix is calculated based on the time-frequency energy matrix of the visual parameter stream and the time-frequency energy matrix of the text parameter stream to quantify the energy coupling strength of the physical layer;

[0025] A structural equation model is constructed to analyze the causal effect of visual parameter flow on text parameter flow, and the change value of conditional entropy is calculated through counterfactual reasoning to represent the causal correlation strength at the logical level.

[0026] The cross-coherence matrix, the sign of the causal strength coefficient, and the conditional entropy change value are integrated to generate a dynamic inhibition weight matrix. The exponential decay function is used to suppress the frequency bands with high coherence and negative causal correlation, while applying positive compensation to the strategy's main frequency.

[0027] Finally, time-frequency filtering is performed on the synchronized visual parameter stream to generate a de-conflicting visual parameter stream, which weakens the frequency band energy that causes strategic contradictions and strengthens the strategic causal path.

[0028] In a preferred embodiment, step S4 includes the following contents:

[0029] Dynamic threshold modulation is performed on the de-conflicting visual parameter stream through a tactical layer spiking neural network, and a nonlinear neuron model is used to simulate the tactical action generation process. The temporal evolution of the neuron membrane potential is driven by the product of the de-conflicting visual parameter stream and the phase modulation term. At the same time, the dynamic threshold is adjusted according to the local energy accumulation of the de-conflicting visual parameter stream. When the neuron membrane potential exceeds the dynamic threshold, a pulse signal is triggered and the membrane potential is reset. The pulse sequence is mapped into a millisecond-level action sequence through a multi-layer structure.

[0030] In a preferred embodiment, step S4 includes the following contents:

[0031] The policy layer memory-augmented network performs attention-enhanced bidirectional loop processing on the text parameter stream. The hidden state update is determined by the current text parameter stream input, the hidden state at the previous moment, and the weighted summary of the historical hidden state. The weighted summary of the historical hidden state is calculated by the attention attenuation factor, and finally the constraint framework is generated by the sigmoid activation function.

[0032] In a preferred embodiment, step S4 includes the following contents:

[0033] The dynamic fusion of millisecond-level action sequences and constraint frameworks is achieved through the spatiotemporal gating mechanism. A spatiotemporal gating unit is constructed to calculate the dynamic fusion coefficient, where the dynamic fusion coefficient is determined by the millisecond-level action sequence, the constraint framework, and the spatiotemporal correlation terms between the two. The spatiotemporal correlation terms are calculated by the product integral of the millisecond-level action sequence and the constraint framework within a local time window. The final fused instruction stream is generated by the weighted average of the millisecond-level action sequence and the constraint framework.

[0034] In a preferred embodiment, step S5 includes the following contents:

[0035] The dynamic characteristics of the fused instruction stream are analyzed by phase space reconstruction technology. The embedding dimension and delay parameters are selected to construct the reconstruction vector. The maximum Lyapunov exponent is calculated in the reconstructed phase space to determine the chaotic behavior of the fused instruction stream.

[0036] The damping controller is designed based on Lyapunov stability theory. The Lyapunov function is defined and the dynamic evolution of the damping term is designed to ensure the stability of the fused instruction stream.

[0037] The maximum Lyapunov exponent is monitored through the chaos edge detection mechanism. When its absolute value is less than the chaos edge threshold, the fused instruction stream is determined to be at the edge of chaos, and a damping term is injected as a negative feedback correction signal to generate a corrected instruction stream to maintain the convergence of the swarm strategy.

[0038] The technical effects and advantages of the present invention over the existing parameter hybrid driven decision-making method are as follows:

[0039] The present invention uses an intelligent focusing mechanism for dynamic conflict periods to precisely target core time windows where high-frequency strategic conflicts occur in complex confrontation scenarios, reducing computing resource consumption during non-essential time periods. It employs cross-modal spatiotemporal reference synchronization technology to eliminate timing misalignment and phase deviation between visual and textual parameters, establishing a foundation for precise alignment of multi-source heterogeneous data. A two-dimensional analysis mechanism, integrating time-frequency physical properties with causal logic, enables breakthrough identification of invisible cross-modal conflict frequency bands that are difficult to detect using traditional methods, systematically addressing the deep mutual exclusion between defensive and aggressive strategic parameters at the energy coupling and causal transmission levels. A hierarchical decision-making architecture based on a spiking neural network and a memory-enhanced network achieves a dynamic balance between millisecond-level tactical micro-manipulation responses and minute-level strategic framework stability, ensuring that high-frequency actions execute within macroscopic constraints. A group decision-making oscillation propagation blocking mechanism, informed by nonlinear stability theory, effectively suppresses the chain-like diffusion of local parameter conflicts within the collaborative network through real-time analysis of strategic waveforms and closed-loop negative feedback control. This creates a comprehensive guarantee system for precise and controllable tactical execution, dynamic correction of strategic objectives, and steady-state convergence of group states, providing intelligent decision-making support with both agile response speed and sustained robustness for dynamic confrontation environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flow chart of the decision-making method for existing parameter hybrid drive according to the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] Example 1: Figure 1 The present invention provides a decision-making method for existing parameter hybrid drive, including:

[0043] S1: Based on the fluctuation feature mutation detection of visual parameter stream and the policy instruction event recognition of text parameter stream, the key conflict intervals of multimodal parameters are aligned through the dynamic time warping algorithm to generate a non-uniformly distributed conflict period coordinate set.

[0044] S2: Within the designated conflict period coordinates, event-driven adaptive downsampling is performed on high-frequency visual parameters to align their time base with the text parameters, and time-frequency analysis technology is used to extract phase compensation vectors to eliminate the phase difference between modalities.

[0045] S3: Conduct a joint analysis of the time-frequency energy distribution and causal correlation strength of the synchronized parameters, construct a counterfactual intervention model to generate a frequency band suppression weight matrix, and directionally weaken the frequency band energy that causes strategic contradictions and strengthen the strategic causal path.

[0046] S4: The de-conflicted visual parameters are input into the tactical layer pulse neural network to generate millisecond-level action sequences, and the text parameters are input into the strategy layer memory enhancement network to generate a constraint framework, and multi-granularity command fusion is achieved through the spatiotemporal gating mechanism.

[0047] S5: Monitor the nonlinear oscillation characteristics of the fused instruction stream in real time, construct a damping controller based on Lyapunov stability theory, and inject a negative feedback correction signal to maintain the convergence of the group strategy when chaotic edge characteristics are detected.

[0048] In multimodal decision-making scenarios, visual and textual parameter streams, due to their differing dynamic characteristics, are prone to conflict during temporal evolution, impacting the effectiveness of group collaborative decision-making. Step S1 focuses on detecting sudden fluctuations in the visual parameter stream and identifying policy instruction events in the textual parameter stream. Using a dynamic time warping algorithm, it aligns the key conflicting intervals of multimodal parameters, generating a non-uniformly distributed set of conflicting period coordinates. This set provides a precise time window for subsequent steps to achieve parameter synchronization and conflict resolution.

[0049] Step S1 includes the following contents:

[0050] S1.1, Detection of sudden changes in fluctuation characteristics of visual parameter flow:

[0051] The visual parameter stream is tested for sudden fluctuations. The visual parameter stream refers to the defensive strategy parameters generated by the visual perception module in a time series. First, the amplitude of change in the visual parameter stream between adjacent time points is calculated. Specifically, for each time point, the absolute value of the difference between the parameter value at the current time point and the parameter value at the previous time point is calculated to reflect the dynamic trend of the visual parameter stream. Next, a sliding window mechanism is introduced to analyze local characteristics within each time window. The sliding window covers a certain number of consecutive time points, and the probability distribution of all amplitude changes within the window is statistically analyzed. Specifically, the amplitude change within the window is divided into several intervals, the frequency of occurrence of each interval is calculated, and then the information entropy, called the local fluctuation entropy, is calculated based on these frequencies. The local fluctuation entropy reflects the complexity and uncertainty of the fluctuations of the visual parameter stream within the window. If the local fluctuation entropy of a time window exceeds a preset entropy threshold, the central time point of the window is marked as a mutation point, indicating that the fluctuation characteristics of the visual parameter stream have changed significantly at this time point. Finally, all marked mutation points are collected to form a visual mutation point set, which serves as input for subsequent processing.

[0052] Using local fluctuation entropy to detect mutation points fully captures the dynamic characteristics of the visual parameter flow within a local range. Compared to methods based solely on amplitude thresholds, local fluctuation entropy comprehensively considers the uncertainty and complexity of fluctuations, more accurately identifying significant changes in fluctuation patterns, thereby improving the accuracy of mutation point detection.

[0053] S1.2, Policy instruction event recognition of text parameter stream:

[0054] The text parameter stream is used to identify strategic instruction events. The text parameter stream refers to the aggressive strategic parameters generated by the text semantic module in the time series by parsing historical instructions. First, the dynamic deviation of the text parameter stream is calculated. The specific method is to first apply exponential smoothing to the text parameter stream to generate a smoothed parameter sequence to reduce the interference of short-term fluctuations. Then, the moving average of the smoothed parameter sequence is calculated to reflect the long-term trend of the parameter. For each time point, the absolute difference between the smoothed parameter value and the moving average is calculated, which is called the dynamic deviation. The size of the dynamic deviation reflects the degree to which the text parameter stream deviates from the trend. Then, a deviation threshold is set. When the dynamic deviation at a certain time point exceeds the deviation threshold, the time point is marked as a strategic instruction event point, indicating that the text parameter stream has undergone a significant strategic shift. Finally, all marked event points are collected to form a text event point set.

[0055] Using exponentially smoothed cumulative deviations to identify strategic order events effectively smooths short-term fluctuations in text parameter streams and highlights long-term trend changes. This method is particularly suitable for detecting strategic order events, avoiding misidentification of minor tactical adjustments as critical events, thereby improving the pertinence and accuracy of event identification.

[0056] S1.3, Dynamic Time Warping algorithm aligns conflicting intervals:

[0057] The dynamic time warping algorithm is used to align the visual mutation point set and the text event point set in time series. The visual mutation point set and the text event point set come from the output of the first two steps, representing the key time points of the two parameter streams. First, the visual mutation point set and the text event point set are regarded as two independent time series. Then, the distance between time points is defined as the absolute time difference between the two, and a cumulative distance matrix is constructed based on this. Each element of the cumulative distance matrix represents the sum of the time differences in all possible paths from the starting point to the current point. Subsequently, the minimum cumulative distance path in the cumulative distance matrix is solved by the dynamic programming method. Specifically, from the starting point to the end point of the matrix, the path with the smallest cumulative distance is gradually selected to ensure the lowest overall alignment cost. After the path solution is completed, the path is backtracked to generate the correspondence between the visual mutation point and the text event point, forming a set of aligned time point pairs as the input for subsequent processing.

[0058] The Dynamic Time Warping algorithm can handle the nonlinear temporal misalignment between the visual and text parameter streams, adapting to the differences in their dynamic characteristics. By calculating the minimum cumulative distance path, it ensures the optimal temporal alignment of the alignment results, thereby accurately identifying conflicting intervals.

[0059] S1.4, generate the conflict period coordinate set:

[0060] A conflict period coordinate set is generated based on the set of aligned time point pairs. The set of aligned time point pairs contains the correspondence between visual mutation points and text event points. First, for each pair of aligned time points, the conflict period is defined as the intersection of a certain time range around the visual mutation point and the text event point. Specifically, with each time point as the center, the preset time width is extended forward and backward to obtain the respective time period ranges, and then the overlapping part of the two is taken as the initial conflict period. Next, the conflict periods of all aligned time point pairs are checked. If the time difference between a pair of time points is less than the preset range, the corresponding conflict periods are merged into an extended period to cover the continuous conflict area. Finally, all overlapping conflict periods are further merged to form a non-uniformly distributed conflict period coordinate set as the final output of step S1.

[0061] By expanding and merging time periods, we ensure that the conflict time period coordinate set covers all potential conflict points while avoiding redundancy caused by overly dispersed time periods. This method generates non-uniformly distributed time intervals, providing precise and efficient time windows for subsequent parameter synchronization and conflict resolution, improving targeted processing.

[0062] Step S1 detects mutation points by calculating the local fluctuation entropy of the visual parameter stream. This is combined with exponentially smoothed cumulative deviations to identify policy instruction events in the text parameter stream. A dynamic time warping algorithm is then used to align visual mutation points with text event points, ultimately generating a non-uniformly distributed set of conflict period coordinates. This set provides a reliable time reference for subsequent processing, ensuring the accuracy and efficiency of multimodal parameter conflict resolution.

[0063] In step S1, by detecting sudden changes in the fluctuation characteristics of the visual parameter stream and identifying policy instruction events of the text parameter stream, combined with a dynamic time warping algorithm, a non-uniformly distributed set of conflict period coordinates is generated. This set accurately marks the key time intervals where dynamic conflicts occur in the timing between the visual parameter stream and the text parameter stream. Based on this set, step S2 uses event-driven adaptive downsampling and phase compensation technology to achieve alignment and phase synchronization of the time bases of the high-frequency visual parameter stream and the low-frequency text parameter stream during the conflict period, laying a data foundation for the time-frequency energy distribution and causal correlation analysis in step S3.

[0064] Step S2 includes the following contents:

[0065] S2.1, event-driven adaptive downsampling:

[0066] Event-driven adaptive downsampling is performed on the high-frequency visual parameter stream during the conflict period to adjust its time base to match the sampling density of the low-frequency text parameter stream. The conflict period is determined by the aforementioned step S1, identifying the critical interval where the high-frequency visual parameter stream and the low-frequency text parameter stream dynamically conflict in time. Because the sampling rate of the high-frequency visual parameter stream is much higher than that of the low-frequency text parameter stream, there is a misalignment between the two time points. To this end, the local change rate of the high-frequency visual parameter stream at each time point is first calculated. The specific method is: take the difference between the parameter values of two adjacent time points and divide it by the time interval between these two time points to obtain the speed of change of the parameter value over time. Next, a predefined change rate threshold is set. When the local change rate of a time point exceeds this threshold, the time point is marked as an event trigger point, indicating that a significant change in the high-frequency visual parameter stream occurred at this time point. Then, downsampling is performed based on these event trigger points: the parameter values at all event trigger points are retained, and for the time points between two adjacent event trigger points, linear interpolation is used to generate approximate values. The linear interpolation calculation method is to calculate the parameter value of the intermediate time point based on the time position and parameter values of the two event trigger points according to the time ratio. Specifically, the difference between the parameter values of the two event trigger points is taken and the interpolation result is distributed according to the time distance between the intermediate time point and the trigger point. Finally, by adjusting the number of interpolation points, the total number of sampling points of the downsampled high-frequency visual parameter stream in each conflict period is consistent with the total number of sampling points of the low-frequency text parameter stream in the same conflict period, thus achieving temporal alignment between the two.

[0067] Event-driven adaptive downsampling ensures that key changes in the high-frequency visual parameter stream are preserved, preventing the loss of important features due to downsampling. It also reduces the computational complexity of subsequent processing by reducing sampling points in non-critical areas through linear interpolation. Compared to traditional uniform downsampling, this method is more adaptable to the dynamic characteristics of the parameter stream, and the generated downsampling results precisely match the temporal distribution of the low-frequency text parameter stream.

[0068] S2.2, time-frequency analysis and phase compensation vector extraction:

[0069] Time-frequency analysis techniques are used to extract phase compensation vectors to eliminate the phase difference between the downsampled high-frequency visual parameter stream and the low-frequency text parameter stream, further improving synchronization accuracy. During each conflict period, a continuous wavelet transform (CWT) is first performed on the downsampled high-frequency visual parameter stream and the low-frequency text parameter stream to generate their respective time-frequency spectra. The CWT computation involves convolving the parameter streams with a set of predefined wavelet basis functions. By adjusting the scale and time shift parameters of the wavelet basis functions, the parameter streams are decomposed into coefficients at different time locations and frequency scales. These coefficients reflect the energy distribution of the parameter streams in the time-frequency domain. Next, based on the generated time-frequency spectra, the phase difference between the downsampled high-frequency visual parameter stream and the low-frequency text parameter stream at the same time location and frequency scale is calculated. Specifically, the phase angle of the coefficients at each time point and scale is taken and the difference between the two phase angles is calculated. Then, a phase compensation vector is constructed: for each time point of the downsampled high-frequency visual parameter stream, its dominant scale is determined, that is, the scale with the largest energy among all scales, and the energy is calculated by the square of the amplitude of the coefficient; on this dominant scale, the negative value of the phase difference is taken as the compensation phase value at that time point; finally, the compensation phase values of all time points are combined in chronological order to form a phase compensation vector.

[0070] The continuous wavelet transform (CWT) can simultaneously analyze the characteristics of parameter streams in both time and frequency. Compared to traditional methods that analyze only frequency, it is more suitable for processing the dynamic changes of non-stationary signals. By calculating the phase difference on the dominant scale and generating a phase compensation vector, the phase deviation between the downsampled high-frequency visual parameter stream and the low-frequency text parameter stream at key frequencies can be specifically corrected. This method improves the accuracy of phase analysis, ensures the synchronization and consistency of the two in the time-frequency domain, and provides a high-quality phase correction basis for generating synchronized parameter streams.

[0071] S2.3, Phase adjustment and synchronization parameter generation:

[0072] Based on the extracted phase compensation vector, the downsampled high-frequency visual parameter stream is phase-adjusted to generate the final synchronized parameter stream, achieving full alignment in time and phase with the low-frequency text parameter stream. First, the downsampled high-frequency visual parameter stream is treated as a complex signal, where the real part is the original parameter value and the imaginary part is initially zero. Then, a phase rotation adjustment is applied to the complex signal at each time point. Specifically, the complex signal at that time point is multiplied by a complex exponential term, where the phase angle of the complex exponential term is derived from the compensated phase value at the corresponding time point of the phase compensation vector. This operation is equivalent to rotating the signal in the complex plane by an angle determined by the compensated phase value. After the rotation is completed, the real part of the complex signal is extracted as the phase-adjusted parameter value. Ultimately, the phase-adjusted high-frequency visual parameter stream is aligned with the low-frequency text parameter stream in terms of time (guaranteed by step S2.1) and is also consistent in phase with the low-frequency text parameter stream, forming a synchronized parameter stream.

[0073] Phase rotation adjustment accurately corrects the phase deviation of the downsampled high-frequency visual parameter stream without changing its amplitude, ensuring phase synchronization with the low-frequency text parameter stream. This avoids the amplitude distortion that could be introduced by directly modifying parameter values, preserves the original dynamic characteristics of the parameter stream, and achieves high-precision alignment of the cross-modal parameter streams in time and phase.

[0074] Step S2 synchronizes the high-frequency visual parameter stream with the low-frequency text parameter stream through three sub-steps: first, event-driven adaptive downsampling is used to adjust the time base of the high-frequency visual parameter stream so that its sampling points match those of the low-frequency text parameter stream; then, a continuous wavelet transform is used to extract the time-frequency spectrum and calculate a phase compensation vector to eliminate the phase difference between the two; finally, a phase rotation is performed based on the phase compensation vector to generate a synchronized parameter stream that is aligned in both time and phase with the low-frequency text parameter stream. This process ensures the accuracy and integrity of the synchronized data, providing high-quality input for subsequent processing.

[0075] In step S2, event-driven adaptive downsampling, time-base alignment of the high-frequency visual parameter stream, and phase compensation techniques are used to eliminate the phase difference between the visual parameter stream and the low-frequency text parameter stream during the conflict period, generating synchronized visual parameter streams and text parameter streams. The conflict period coordinate set provided by step S1 identifies the key conflict intervals of the multimodal parameters. Based on this, step S3 conducts a joint analysis of the time-frequency energy distribution and causal correlation strength of the synchronized visual parameter stream and text parameter stream, constructs a counterfactual intervention model, and generates a frequency band suppression weight matrix. This method selectively weakens the frequency band energy that causes strategic contradictions and strengthens the strategic causal path, providing a de-conflicted visual parameter stream for step S4.

[0076] Step S3 includes the following contents:

[0077] S3.1, time-frequency energy distribution extraction:

[0078] The time-frequency energy distribution of the synchronized visual parameter stream and text parameter stream is extracted to capture their dynamic characteristics in time and frequency. Wavelet packet decomposition is first performed on the synchronized visual parameter stream. Wavelet packet decomposition decomposes the signal into different levels, generating a time-frequency energy matrix that records the energy distribution of the visual parameter stream at different time points and frequency levels. The energy value is calculated by squared coefficients of each level after wavelet packet decomposition. The decomposition uses Daubechies wavelet basis functions, and the decomposition level is determined according to pre-set values to balance the requirements of time and frequency resolution while controlling computational complexity. A short-time Fourier transform is then performed on the synchronized text parameter stream to generate its dominant frequency energy feature matrix. The short-time Fourier transform slides a Gaussian window along the time axis, transforming the signal within the window and calculating the energy distribution of the signal at different time periods and frequencies. The window length is adaptively adjusted based on the duration of the conflict period to ensure sufficient frequency resolution during the conflict period.

[0079] Wavelet packet decomposition is suitable for processing non-stationary signals, providing highly accurate time and frequency localization features, particularly well suited for capturing high-frequency fluctuations in visual parameter flows. Short-time Fourier transforms, while maintaining computational efficiency, provide the dominant frequency energy signature of the text parameter flow, facilitating comparison with the visual parameter flow. The combination of these two time-frequency analysis techniques comprehensively characterizes the dynamic behavior of the synchronized visual and text parameter flows during the conflict period.

[0080] S3.2, cross-coherence calculation:

[0081] Based on the time-frequency energy matrix of the visual parameter stream and the main frequency energy feature matrix of the text parameter stream, the cross-coherence of the two is calculated at each time-frequency unit to quantify the energy coupling strength of the physical layer. The cross-coherence calculation process is as follows: for each time-frequency unit, the product of the energy value of the visual parameter stream and the energy value of the text parameter stream is calculated, and then divided by the square root of the sum of the square of the energy of the visual parameter stream and the square of the energy of the text parameter stream to obtain a value between 0 and 1. This value reflects the energy similarity between the visual parameter stream and the text parameter stream in this time-frequency unit. The larger the value, the more similar the energy distribution of the two in this time-frequency unit, and the higher the energy coupling strength. Finally, a cross-coherence matrix is generated, which records the coherence values of all time-frequency units.

[0082] Cross-coherence, a dimensionless similarity metric, effectively captures the energy correlation between visual and textual parameter flows in the time-frequency domain. Compared to traditional correlation metrics, cross-coherence focuses more on the similarity of energy distributions and is suitable for analyzing the physical-layer coupling characteristics of multimodal signals.

[0083] S3.3, Causal Association Strength Analysis:

[0084] A structural equation model was constructed for the synchronized visual and text parameter flows to analyze the causal relationship between them, and the strength of the causal association was quantified through counterfactual reasoning. First, assuming a linear causal influence between the visual parameter flow and the text parameter flow, a linear regression model was constructed with the text parameter flow as the dependent variable and the visual parameter flow as the independent variable. The regression coefficients were estimated using the least squares method. The regression coefficients represent the direction and magnitude of the causal influence. Next, counterfactual reasoning was used to simulate the effects of an intervention on the visual parameter flow. The specific process involved setting baseline and intervention values for the visual parameter flow and calculating the conditional entropy change of the text parameter flow before and after the intervention. The conditional entropy reflects the change in uncertainty in the text parameter flow given the visual parameter flow intervention; a larger entropy change indicates a stronger causal association. Finally, a causal strength coefficient and a conditional entropy change matrix were generated, providing quantitative indicators of causal association at the logical level for generating frequency band suppression weights.

[0085] Structural equation modeling combined with counterfactual reasoning can reveal, at a logical level, the causal influence of visual parameter flows on textual parameter flows. By calculating changes in conditional entropy, this not only quantifies the strength of causal relationships but also provides a basis for identifying the causal pathways that trigger strategic conflicts. This approach helps to selectively reduce the frequency of negative causal associations and improve the effectiveness of conflict resolution.

[0086] S3.4, frequency band suppression weight matrix generation:

[0087] Cross-coherence, causal strength coefficients, and conditional entropy change values are integrated to generate a dynamic suppression weight matrix. This matrix effectively reduces the energy of frequency bands that trigger strategic conflicts and compensates for the energy of the strategic dominant frequency. The specific process involves constructing an exponential decay function whose inputs are the cross-coherence, the sign of the causal strength coefficient, and the conditional entropy change value, and whose output is the suppression weight. For each time-frequency unit, the product of the cross-coherence and the conditional entropy change value is calculated, then multiplied by the sign of the causal strength coefficient (1 for positive causality and -1 for negative causality). The negative value of this product is used as the exponential term of the exponential decay function to calculate the suppression weight. Smaller weights indicate stronger suppression of the time-frequency unit. In particular, for frequency bands with high cross-coherence and negative causal associations, the weights approach zero, resulting in significant energy attenuation. Next, for pre-defined strategic dominant frequencies, positive compensation is applied to the suppression weights. Specifically, the weight of the strategic dominant frequency is increased by a fixed compensation coefficient to strengthen the strategic causal path. Finally, a frequency band suppression weight matrix is generated, recording the weights of all time-frequency units.

[0088] The strategic dominant frequency refers to the frequency component in the text parameter stream that is directly related to the global strategic goal and carries the main strategic intention. The text parameter stream is generated by the text semantic module and reflects the low-frequency steady-state characteristics of the aggressive strategy. The strategic dominant frequency is the key frequency segment that has a dominant influence on group collaborative decision-making, usually manifested as a frequency band with concentrated and stable energy. The strategic dominant frequency is identified through the dominant frequency energy feature matrix extracted by short-time Fourier transform, representing the core dynamic characteristics of the text parameter stream that are consistent with the strategic goal. Strengthening the energy of the strategic dominant frequency is intended to ensure that the de-conflicted visual parameter stream remains consistent with the global strategic direction while eliminating strategic contradictions, thereby avoiding weakening the strategic orientation of the overall decision due to excessive suppression.

[0089] By integrating physical-layer energy coupling metrics (cross-coherence) with logical-layer causal correlation metrics (causal strength coefficient and conditional entropy change), a dynamic suppression weight matrix is generated. This accurately identifies and weakens frequency bands that cause policy conflicts, while simultaneously protecting and strengthening policy-related frequency bands. This two-dimensional constraint mechanism ensures targeted and effective conflict resolution, avoiding the risks of blind suppression or overcompensation.

[0090] S3.5, parameter conflict resolution:

[0091] Based on the frequency band suppression weight matrix, the synchronized visual parameter stream is filtered in the time and frequency domain to generate a de-conflicted visual parameter stream. The specific process is as follows: First, the synchronized visual parameter stream is Fourier transformed to obtain its frequency domain representation; then, the frequency domain signal is multiplied by the corresponding elements of the frequency band suppression weight matrix to adjust the energy of specific frequency bands; finally, the adjusted frequency domain signal is inverse Fourier transformed and converted back to the time domain to obtain the de-conflicted visual parameter stream.

[0092] Time-frequency filtering can precisely suppress frequency bands that trigger strategic conflicts by adjusting the energy of specific frequency bands without changing the overall structure of the visual parameter stream. Compared to traditional time-domain filtering, time-frequency filtering is more suitable for processing non-stationary signals and can flexibly adapt to the dynamic characteristics of the synchronized visual parameter stream, ensuring that the de-conflicted visual parameter stream does not introduce strategic conflicts in the subsequent generation of tactical action sequences.

[0093] In step S3, a counterfactual intervention model is constructed by jointly analyzing the time-frequency energy distribution and causal correlation strength of the synchronized visual parameter stream and text parameter stream. A frequency band suppression weight matrix is generated, and the visual parameter stream is deconflicted to obtain a deconflicted visual parameter stream. Meanwhile, the text parameter stream is passed directly to step S4 as the policy-level input without being processed. Based on this, step S4 uses the deconflicted visual parameter stream to generate millisecond-level action sequences at the tactical level and the text parameter stream to generate a policy-level constraint framework. Multi-granularity instructions are fused between the two through an innovative spatiotemporal gating mechanism, and the fused instruction stream is finally output for use in step S5.

[0094] Step S4 includes the following contents:

[0095] S4.1, tactical layer spiking neural network generates millisecond-level action sequences:

[0096] A tactical-layer spiking neural network (SNN) generates millisecond-level action sequences for the deconflicting visual parameter stream. This tactical-layer spiking neural network uses a dynamic threshold modulation mechanism combined with a nonlinear neuron model to simulate the tactical action generation process. The neuron's membrane potential evolves over time, with its rate of change proportional to the product of the deconflicting visual parameter stream and a phase modulation term, and is also affected by negative feedback from the membrane potential itself. The phase modulation term is derived and normalized from the phase compensation vector in step S2, reflecting the temporal characteristics of the deconflicting visual parameter stream. The threshold is dynamically adjusted based on the energy accumulation of the deconflicting visual parameter stream within a local time window. Specifically, it is a baseline threshold plus a weighted value obtained by integrating the absolute value of the deconflicting visual parameter stream within the local time window. When the membrane potential exceeds the dynamic threshold, the neuron fires a spike signal and resets the membrane potential to a preset value. The tactical-layer spiking neural network consists of a multi-layer structure, including an input layer, two hidden layers, and an output layer. A weighted summation and time delay mechanism is used to map the spike sequence into a millisecond-level action sequence. The process of generating an action sequence is as follows: weighted summation of the pulse signals of each neuron is performed, the weights and delay parameters are determined according to network training, and finally a millisecond-level action sequence is output as a high-frequency action instruction at the tactical level.

[0097] The tactical-layer spiking neural network mimics the efficient information processing of biological neural systems, making it particularly well-suited for high-frequency, time-sensitive tactical action generation tasks. A dynamic threshold modulation mechanism enables the network to adaptively adjust its sensitivity to the deconflicting visual parameter stream, enhancing its responsiveness to key features. Combined with phase modulation, the network better captures the temporal dynamics of the deconflicting visual parameter stream, improving the accuracy and response speed of millisecond-level action sequence generation. This processing approach ensures efficient and real-time tactical actions, providing a precise input foundation for subsequent multi-granularity command fusion.

[0098] S4.2, Policy layer memory-augmented network generation constraint framework:

[0099] A constraint framework is generated for the text parameter stream using a policy-layer memory-augmented network. The policy-layer memory-augmented network employs a bidirectional recurrent structure based on an attention mechanism to extract long-term dependency features from the text parameter stream. The network's hidden state update process is as follows: the current hidden state is determined by a weighted aggregation of the current text parameter stream input, the previous hidden state, and the historical hidden states. This weighted aggregation of historical hidden states is achieved using an attention decay factor, which decreases with increasing time differences, emphasizing the importance of recent hidden states. Finally, a constraint framework is generated from the hidden state using a sigmoid activation function. The output value is constrained between 0 and 1, representing low-frequency constraints at the policy level for subsequent multi-granularity instruction fusion.

[0100] The attention decay factor is a dynamic weight function used in policy-level memory-augmented networks to quantify the influence of past hidden states on the current hidden state. Specifically, the attention decay factor is calculated based on the time difference between the current time point and the past time point. Its value decreases as the time difference increases, and it is typically expressed in an exponential decay form.

[0101] The policy-level memory-augmented network, combined with an attention mechanism, effectively captures long-term dependencies in the text parameter stream, making it suitable for generating policy-level constraint frameworks. The bidirectional recurrent structure allows the network to simultaneously consider both past and future information in the text parameter stream, improving its understanding and prediction of policy instructions. The design of the attention decay factor enables the network to dynamically adjust its emphasis on historical information, enhancing its sensitivity to recent policy changes and ensuring that the constraint framework reflects both long-term trends and adapts to short-term adjustments. This processing approach provides stable strategic guidance for the fused instruction stream, ensuring the coordination of tactical execution and strategic objectives.

[0102] S4.3, spatiotemporal gating mechanism realizes multi-granularity instruction fusion:

[0103] The multi-granularity instruction fusion of millisecond-level action sequences and constraint frameworks is achieved through the spatiotemporal gating mechanism to generate a fused instruction stream. The spatiotemporal gating unit first calculates the dynamic fusion coefficient, which is determined by the millisecond-level action sequence, the constraint framework, and the spatiotemporal correlation terms of the two. The spatiotemporal correlation term is obtained by integrating the product of the millisecond-level action sequence and the constraint framework in the local time window, reflecting the degree of coordination between the two in the short term. Subsequently, the millisecond-level action sequence and the constraint framework are multiplied by the preset weight matrix respectively, and the spatiotemporal correlation term is added. The result is input into the sigmoid function, and the output value is limited to between 0 and 1. Finally, the fused instruction stream is the weighted average of the millisecond-level action sequence and the constraint framework. The weight is dynamically adjusted by the dynamic fusion coefficient, and the fused instruction stream is output as the input for subsequent processing. For example, to generate a fused instruction stream : ,in Dynamically balance the contribution of tactical actions and strategic constraints, For millisecond-level action sequences, The policy constraint framework is a dimensionless eigenvalue sequence.

[0104] The spatiotemporal gating mechanism achieves a flexible balance between millisecond-level action sequences and the constraint framework through dynamic fusion coefficients, adaptively adjusting the contribution ratio of the two in different scenarios. The introduction of spatiotemporal correlation terms enhances the sensitivity of the fusion process to short-term synergies, ensuring that the fused instruction stream can both quickly respond to tactical requirements and adhere to strategic constraints. This multi-granular fusion approach improves the coordination and consistency of instructions, avoids conflicts between tactical actions and strategic objectives, and provides stable and efficient input for subsequent nonlinear oscillation monitoring and swarm strategy convergence.

[0105] Step S4 converts the de-conflicted visual parameter stream into a millisecond-level action sequence using a tactical-level spiking neural network. A policy-level memory-augmented network generates a constraint framework from the text parameter stream. A spatiotemporal gating mechanism is employed to dynamically fuse the millisecond-level action sequence with the constraint framework, ultimately outputting a fused instruction stream. While retaining the high-frequency response at the tactical level, the fused instruction stream is modulated by the policy-level constraint framework, ensuring consistency between tactical execution and policy objectives. This enables step S4 to efficiently process the de-conflicted visual and text parameter streams.

[0106] In step S4, de-conflicting visual parameters are processed by a tactical-layer spiking neural network to generate millisecond-level action sequences. Text parameters are processed by a policy-layer memory-augmented network to generate a constraint framework. Multi-granularity instructions are then fused using a spatiotemporal gating mechanism to generate a fused instruction stream. This fused instruction stream achieves a dynamic balance between tactical actions and policy constraints, possessing both high-frequency response and low-frequency guidance characteristics. However, due to the nonlinear interaction of multimodal parameters and the complexity of group collaborative decision-making, the fused instruction stream may exhibit nonlinear oscillation characteristics, even leading to policy deviations at the edge of chaos. To address this issue, step S5 monitors the nonlinear oscillation characteristics of the fused instruction stream in real time, constructs a damping controller based on Lyapunov stability theory, and injects a negative feedback correction signal when the edge of chaos is detected to ensure the convergence and stability of the group strategy.

[0107] Step S5 includes the following contents:

[0108] S5.1, nonlinear oscillation characteristic monitoring:

[0109] Phase space reconstruction techniques are used to monitor the nonlinear oscillation characteristics of the fused instruction stream. The fused instruction stream is the dimensionless sequence of eigenvalues generated in step S4 above, integrating instruction information related to tactical actions and policy constraints. The purpose of phase space reconstruction is to convert the time series of the fused instruction stream into trajectories in a high-dimensional phase space to reveal the underlying dynamic characteristics of the fused instruction stream. The specific process is as follows: First, appropriate embedding dimensions and delay parameters are selected. The embedding dimension represents the dimension of the phase space, and the delay represents the time interval during reconstruction. These parameters are determined using techniques such as mutual information and pseudo-nearest neighbor methods to ensure that the reconstructed phase space accurately reflects the dynamic behavior of the fused instruction stream. Next, based on the selected embedding dimension and delay, a reconstruction vector is constructed. The reconstruction vector consists of the values of the fused instruction stream at different time lags, forming trajectory points in the phase space. Then, in the reconstructed phase space, the divergence characteristics of the trajectory are calculated. Specifically, the divergence between adjacent trajectory points over time is tracked and the maximum Lyapunov exponent is calculated. The maximum Lyapunov exponent is calculated by taking the logarithmic growth rate limit of the norm of the trajectory deviation vector over time, which represents the rate of trajectory divergence in phase space. If the maximum Lyapunov exponent is positive, the fused instruction stream is considered to exhibit chaotic behavior. If the absolute value of the maximum Lyapunov exponent is close to zero, the fused instruction stream is considered to be on the verge of chaos, which may lead to strategy instability.

[0110] For example:

[0111] Input: Fused Instruction Stream , is the dimensionless eigenvalue sequence generated in step S4, representing the fusion result of multi-granularity instructions.

[0112] deal with:

[0113] Phase space reconstruction technique was used to analyze Dynamic characteristics of . Select the embedding dimension and delay , construct the reconstruction vector:

[0114] in, and Determined by mutual information method and pseudo nearest neighbor method.

[0115] The Lyapunov exponent spectrum is calculated in the reconstructed phase space to measure the trajectory divergence characteristics. Defined as:

[0116] in, represents the deviation vector between adjacent trajectories, represents the Euclidean norm.

[0117] according to The value of determines the system status: if , then the fused instruction stream has chaotic behavior; if near , is on the edge of chaos.

[0118] Output: Maximum Lyapunov exponent , which indicates the degree of chaos of the fused instruction stream.

[0119] Phase space reconstruction technology visualizes and quantifies the nonlinear dynamic characteristics of the fused instruction stream, making it suitable for analyzing chaotic and oscillatory behavior in complex systems. The maximum Lyapunov exponent, a quantitative indicator of chaotic behavior, accurately determines whether the fused instruction stream is unstable, providing a scientific basis for subsequent damping controller construction and negative feedback injection. This monitoring method ensures real-time understanding of the dynamic characteristics of the fused instruction stream, overcoming the limitations of traditional linear analysis methods when dealing with nonlinear systems and improving monitoring accuracy and reliability.

[0120] S5.2, damping controller construction:

[0121] Based on Lyapunov stability theory, a damping controller is constructed for the fused instruction stream to ensure its stability. The design of the damping controller relies on the construction of the Lyapunov function and the realization of stability conditions. The specific process is as follows: First, a Lyapunov function is defined, which takes half the square of the current value of the fused instruction stream to represent the system's energy. Next, the time derivative of the Lyapunov function is calculated, which is the product of the current value of the fused instruction stream and its rate of change. To ensure system stability, that is, the Lyapunov function decreases over time, a damping term is introduced, and a modified instruction stream is designed so that the rate of change of the modified instruction stream is negatively correlated with that of the fused instruction stream. Specifically, the dynamic evolution of the damping term is designed to be a combination of the negative value of the fused instruction stream and the rate of change of the fused instruction stream, thereby achieving negative feedback control of the fused instruction stream. By solving the dynamic evolution relationship of the damping term, the damping term is obtained as a combination of the negative integral term of the fused instruction stream and the fused instruction stream change adjustment term. Finally, the damping term serves as a negative feedback signal for dynamic adjustment of the fused instruction stream in subsequent steps.

[0122] For example:

[0123] Input: Fused Instruction Stream , maximum Lyapunov exponent .

[0124] The damping controller is designed based on Lyapunov stability theory, and the Lyapunov function is defined as:

[0125] Its time derivative is:

[0126] To ensure stability (i.e. ), introducing the damping term , generating the corrected instruction stream:

[0127] And design the damping dynamics:

[0128] in, is the dimensionless damping coefficient, which controls the convergence speed.

[0129] Solve the dynamic equation for the damping term:

[0130] Points earned:

[0131] Output: Damping term , is a dimensionless eigenvalue sequence, representing the negative feedback correction signal.

[0132] Lyapunov stability theory is a classic method for controlling the stability of nonlinear systems. By constructing a Lyapunov function and designing a damping term, it ensures that the dynamic behavior of the fused instruction stream converges to a stable state. The damping controller design adaptively adjusts the evolution of the fused instruction stream, preventing it from falling into chaos or divergence, thereby improving the robustness and reliability of group collaborative decision-making.

[0133] S5.3, Chaos edge detection and negative feedback injection:

[0134] Based on the monitoring results of the maximum Lyapunov exponent, the fused instruction stream is detected to be on the edge of chaos, and a negative feedback correction signal is injected when necessary. The specific processing process is as follows: First, a predefined chaos edge threshold is set. When the absolute value of the maximum Lyapunov exponent is less than the chaos edge threshold, the fused instruction stream is determined to be on the edge of chaos, indicating a risk of policy deviation. Next, a corrected instruction stream is constructed: If the fused instruction stream is detected to be on the edge of chaos, the weighted sum of the fused instruction stream and the damping term is used as the corrected instruction stream; if no chaos edge is detected, the fused instruction stream remains unchanged. The injection strength of the damping term is controlled by a preset adjustment parameter to balance the correction effect with the original characteristics of the fused instruction stream. Finally, the corrected instruction stream is output as a stability-controlled instruction stream for subsequent decision execution.

[0135] Chaos edge detection and negative feedback injection mechanisms enable timely intervention when the fused command stream is about to become unstable, ensuring the convergence and stability of the swarm strategy. By dynamically adjusting and correcting the command stream and injecting damping terms only when necessary, excessive interference with the normal fused command stream is avoided, maintaining the agility of tactical response. This adaptive control strategy improves the multimodal decision-making system's anti-interference capabilities and stability in adversarial scenarios, providing a reliable technical foundation for swarm collaborative decision-making.

[0136] For example:

[0137] Input: Maximum Lyapunov exponent , fused instruction stream , damping term .

[0138] deal with:

[0139] Set the edge of chaos threshold ,when When , it is determined that the fused instruction stream is at the edge of chaos and a correction signal needs to be injected.

[0140] Constructing the corrected instruction stream :

[0141]

[0142] in, To adjust the parameters, control the damping injection intensity.

[0143] Output: Corrected instruction stream , is a dimensionless eigenvalue sequence, representing the instruction stream after stability control.

[0144] Step S5 monitors the nonlinear oscillation characteristics of the fused instruction stream through phase space reconstruction technology, constructs a damping controller using Lyapunov stability theory, and dynamically injects a negative feedback correction signal when the chaotic edge is detected, ultimately generating a corrected instruction stream. The corrected instruction stream effectively suppresses the policy deviation in group collaborative decision-making, ensuring the stability and convergence of multimodal instruction fusion in complex environments. The nonlinear oscillation characteristic monitoring of the first sub-step provides a dynamic basis for subsequent control, the damping controller construction of the second sub-step lays the foundation for stability control, and the chaotic edge detection and negative feedback injection of the third sub-step achieves the precise implementation of dynamic correction. This step-by-step monitoring, control, and correction design enables step S5 to efficiently process the dynamic characteristics of the fused instruction stream, providing a solid guarantee for the robustness and efficiency of the entire decision-making system, and is suitable for group collaborative decision-making that requires high stability and reliability.

[0145] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0146] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0147] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0148] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0149] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A decision-making method driven by hybrid existing parameters, characterized in that: Including steps: S1: Based on the detection of fluctuation characteristics of visual parameter streams and the recognition of policy instruction events in text parameter streams, the dynamic time warping algorithm is used to align the key conflict intervals of multimodal parameters and generate a non-uniformly distributed set of conflict period coordinates; S2: Within the designated conflict period coordinates, event-driven adaptive downsampling is performed on high-frequency visual parameters to align their time base with the text parameters. Time-frequency analysis techniques are then used to extract phase compensation vectors to eliminate inter-modal phase differences. S3: Conduct a joint analysis of the time-frequency energy distribution and causal correlation strength of the synchronized parameters, construct a counterfactual intervention model to generate a frequency band suppression weight matrix, and specifically weaken the frequency band energy that causes strategic contradictions and strengthen the strategic causal path; S4: De-conflicting visual parameters are input into the tactical layer spiking neural network to generate millisecond-level action sequences. Text parameters are input into the policy layer memory-augmented network to generate a constraint framework. Multi-granularity command fusion is achieved through a spatiotemporal gating mechanism. S5: Monitor the nonlinear oscillation characteristics of the fused instruction stream in real time, construct a damping controller based on Lyapunov stability theory, and inject a negative feedback correction signal to maintain the convergence of the group strategy when chaotic edge characteristics are detected.

2. The decision-making method for existing parameter hybrid drive according to claim 1, characterized in that: Step S1 includes the following contents: First, the fluctuation feature mutation point is detected by calculating the local fluctuation entropy of the visual parameter flow. The probability distribution of the change amplitude of the visual parameter flow is statistically analyzed in each time window and the information entropy is calculated. When the local fluctuation entropy exceeds the preset entropy threshold, the central time point of the corresponding time window is marked as the visual mutation point. Then, the exponential smoothing cumulative deviation method is used to identify the strategic instruction events of the text parameter stream. Exponential smoothing is applied to the text parameter stream to reduce short-term fluctuations. The dynamic deviation between the smoothed parameter and the moving average is calculated. When the dynamic deviation exceeds the preset deviation threshold, the corresponding time point is marked as a strategic instruction event point. Then, the dynamic time warping algorithm is used to align the visual mutation point set with the strategy instruction event point set, construct a cumulative distance matrix, and solve the minimum cumulative distance path through dynamic programming to generate a set of aligned time point pairs. Finally, a conflict period coordinate set is generated based on the set of aligned time point pairs. A conflict period is defined for each pair of aligned time points and then expanded and merged to form a non-uniformly distributed conflict period coordinate set.

3. The decision-making method for existing parameter hybrid drive according to claim 2, characterized in that: Step S2 includes the following contents: During each conflict period, the local change rate of the high-frequency visual parameter stream is first calculated to identify the event trigger point. When the local change rate exceeds the preset change rate threshold, the corresponding time point is marked as the event trigger point. Then, adaptive downsampling is performed based on the event trigger point, retaining the parameter value at the event trigger point, and linear interpolation is used to generate approximate values for the non-event trigger point, so that the number of sampling points of the downsampled high-frequency visual parameter stream in the conflict period is consistent with the number of sampling points of the low-frequency text parameter stream in the same conflict period, achieving time reference alignment.

4. The decision-making method for existing parameter hybrid drive according to claim 3, characterized in that: Step S2 also includes the following: Then, continuous wavelet transform is performed on the downsampled high-frequency visual parameter stream and low-frequency text parameter stream respectively to generate their respective time-frequency spectra, and the phase difference between the two at each time point and scale is calculated; A phase compensation vector is constructed based on the phase difference, and the negative value of the phase difference is taken as the compensation phase on the dominant scale at each time point; Finally, a phase rotation adjustment is applied to the downsampled high-frequency visual parameter stream, and the complex signal at each time point is multiplied by the complex exponential term, where the phase angle of the complex exponential term is the compensation phase value, and the real part is taken to generate a synchronous parameter stream, thereby achieving alignment with the low-frequency text parameter stream in time and phase.

5. The decision-making method for existing parameter hybrid drive according to claim 4, characterized in that: Step S3 includes the following contents: Perform wavelet packet decomposition on the synchronized visual parameter stream to generate a time-frequency energy matrix, and perform short-time Fourier transform on the text parameter stream to generate a main frequency energy feature matrix; The cross-coherence matrix is calculated based on the time-frequency energy matrix of the visual parameter stream and the time-frequency energy matrix of the text parameter stream to quantify the energy coupling strength of the physical layer; A structural equation model is constructed to analyze the causal effect of visual parameter flow on text parameter flow, and the change value of conditional entropy is calculated through counterfactual reasoning to represent the causal correlation strength at the logical level. The cross-coherence matrix, the sign of the causal strength coefficient, and the conditional entropy change value are integrated to generate a dynamic inhibition weight matrix. The exponential decay function is used to suppress the frequency bands with high coherence and negative causal correlation, while applying positive compensation to the strategy's main frequency. Finally, time-frequency filtering is performed on the synchronized visual parameter stream to generate a de-conflicting visual parameter stream, which weakens the frequency band energy that causes strategic contradictions and strengthens the strategic causal path.

6. The decision-making method for existing parameter hybrid drive according to claim 5, characterized in that: Step S4 includes the following contents: Dynamic threshold modulation is performed on the de-conflicting visual parameter stream through a tactical layer spiking neural network, and a nonlinear neuron model is used to simulate the tactical action generation process. The temporal evolution of the neuron membrane potential is driven by the product of the de-conflicting visual parameter stream and the phase modulation term. At the same time, the dynamic threshold is adjusted according to the local energy accumulation of the de-conflicting visual parameter stream. When the neuron membrane potential exceeds the dynamic threshold, a pulse signal is triggered and the membrane potential is reset. The pulse sequence is mapped into a millisecond-level action sequence through a multi-layer structure.

7. The decision-making method for existing parameter hybrid drive according to claim 6, characterized in that: Step S4 includes the following contents: The policy layer memory-augmented network performs attention-enhanced bidirectional loop processing on the text parameter stream. The hidden state update is determined by the current text parameter stream input, the hidden state at the previous moment, and the weighted summary of the historical hidden state. The weighted summary of the historical hidden state is calculated by the attention attenuation factor, and finally the constraint framework is generated by the sigmoid activation function.

8. The decision-making method for existing parameter hybrid drive according to claim 7, characterized in that: Step S4 includes the following contents: The dynamic fusion of millisecond-level action sequences and constraint frameworks is achieved through the spatiotemporal gating mechanism. A spatiotemporal gating unit is constructed to calculate the dynamic fusion coefficient, where the dynamic fusion coefficient is determined by the millisecond-level action sequence, the constraint framework, and the spatiotemporal correlation terms between the two. The spatiotemporal correlation terms are calculated by the product integral of the millisecond-level action sequence and the constraint framework within a local time window. The final fused instruction stream is generated by the weighted average of the millisecond-level action sequence and the constraint framework.

9. The decision-making method for existing parameter hybrid drive according to claim 8, characterized in that: Step S5 includes the following contents: The dynamic characteristics of the fused instruction stream are analyzed by phase space reconstruction technology. The embedding dimension and delay parameters are selected to construct the reconstruction vector. The maximum Lyapunov exponent is calculated in the reconstructed phase space to determine the chaotic behavior of the fused instruction stream. The damping controller is designed based on Lyapunov stability theory. The Lyapunov function is defined and the dynamic evolution of the damping term is designed to ensure the stability of the fused instruction stream. The maximum Lyapunov exponent is monitored through the chaos edge detection mechanism. When its absolute value is less than the chaos edge threshold, the fused instruction stream is determined to be at the edge of chaos, and a damping term is injected as a negative feedback correction signal to generate a corrected instruction stream to maintain the convergence of the swarm strategy.

Images