A method and system for uniaxial absolute direction perception based on software-defined hardware paradigm

Abstract

This invention relates to a method and system for single-axis absolute orientation sensing based on a software-defined hardware paradigm, belonging to the field of orientation sensing technology. Addressing the technical bias that "single-axis sensors must rely on integral operations," this invention reconstructs a perception paradigm using a software-defined architecture: simplified single-axis hardware acquisition, parallel spatial and astronomical algorithm computation, and the introduction of absolute orientation distribution entropy, which is a fusion optimization of Shannon entropy and gradient variance, calculated as H = H₁ + λH₂. This not only breaks through the integral constraint to achieve instantaneous absolute orientation estimation but also uses the entropy value quantification to trigger dynamic compensation of a generative deep learning model based on perception confidence. Thus, absolute orientation estimation can be achieved without integral operations, overcoming the technical bias that "single-axis sensors must integrate." This invention improves tracking accuracy and reduces hardware costs while supporting remote software upgrades. In scenarios such as solar tracking, it can significantly reduce false triggering and false detection rates, effectively solving the industry problem of the trade-off between accuracy and cost in traditional solutions.

Description

Technical Field

[0001] This invention relates to the field of orientation sensing technology, specifically to a method and system for single-axis absolute orientation sensing based on a software-defined hardware paradigm. Background Technology

[0002] Direction sensing technology has significant application value in navigation, smart devices, and industrial inspection. Current technologies for achieving absolute direction sensing mainly fall into two categories: dual-axis hardware-defined direction sensing and single-axis software-defined direction sensing. Traditional direction sensing technologies have long suffered from a trilemma of accuracy, cost, and flexibility. While dual-axis hardware solutions offer high accuracy, they are costly, bulky, and complex to calibrate. Single-axis solutions, while low-cost, rely on integral calculations for direction estimation, leading to unavoidable cumulative errors. Therefore, those skilled in the art generally believe that the currently predominantly used fixed hardware architecture cannot dynamically configure hardware logic, and upgrades and maintenance depend on hardware modifications, resulting in high costs and long cycles. Furthermore, geomagnetic and solar azimuth calculations are susceptible to environmental interference, and error compensation relies on fixed models with weak adaptive capabilities. In addition, traditional direction estimation suffers from integral cumulative errors, resulting in insufficient accuracy and stability, making it difficult to balance sensing accuracy, hardware cost, and system flexibility. Overcoming the physical limitations of single-axis systems requires adding hardware dimensions or accepting the inherent defects of integral drift. However, with the development of the software-defined hardware paradigm in general computing, there is still no clear technical path to apply it to the field of direction sensing to reconstruct the aforementioned technical contradictions. Therefore, how to combine minimalist hardware with software algorithms and AI technology to clarify the system module composition, build a layered and decoupled architecture, standardize module interaction methods, clarify the subject of defect compensation, improve perception accuracy and system flexibility, and at the same time significantly reduce costs has become a pressing technical challenge in the field of current orientation perception technology.

[0003] Regarding existing dual-axis hardware sensing technologies, represented by patent CN107565680A, the core technical feature is hardware-defined orientation. It directly acquires and calculates orientation data through the physical structure of the dual-axis sensor hardware, achieving absolute direction perception based on the hardware's multi-axis dimensions. The drawbacks of this approach are: firstly, the multi-axis hardware structure is complex, with high procurement and maintenance costs, and accuracy improvements depend on hardware stacking, with costs increasing exponentially with accuracy, resulting in low cost-effectiveness; secondly, the hardware and algorithm are deeply intertwined, the system module division is vague, and there is no unified standard for module interaction, hindering flexible upgrades. Optimizing accuracy or adapting to new scenarios requires replacing the entire hardware system, resulting in poor scalability and hardware bottlenecks in improving perception accuracy. Regarding existing single-axis software-defined orientation technologies, represented by patent CN212392708U, the core technical feature is basic software-defined orientation. It processes the raw data acquired by the single-axis sensor through simple software algorithms to extend the orientation sensing function of the single-axis hardware. This solution reduces hardware costs while alleviating the problems of large size and high power consumption of dual-axis solutions. However, its software algorithm adopts the traditional static compensation method, which can only compensate for steady-state or slowly changing environmental disturbances. In closed-loop control, single-axis sensor systems cannot avoid relying on integral elements, which limits the ability to eliminate steady-state errors and thus restricts dynamic response performance. Under the same conditions, its sensing accuracy is lower than that of dual-axis hardware solutions. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention provides a method and system for single-axis absolute direction sensing based on a software-defined hardware paradigm, thus solving the aforementioned technical challenges. To achieve the above objectives, the technical solution of this invention is summarized as follows: It adopts a three-layer decoupled architecture consisting of a hardware resource layer, a software-defined layer, and an application service layer. The hardware logic is dynamically defined in software through reconfigurable computing units. Error compensation is achieved using generative adversarial networks, and robustness is improved by fusing geomagnetic and solar orientation calculation results. A generative model is used to achieve single-axis sensing without integration for absolute direction estimation, avoiding the accumulation of integral errors. The direction deviation is determined using the comprehensive feature F-value, which can also serve as a basis for assessing perception confidence. Ultimately, this achieves high-precision, highly continuous, and highly flexible single-axis absolute direction sensing.

[0005] In terms of methodology, this invention independently acquires single-axis absolute direction sensing signals through a single-axis sensor module, and completes spatial calculations by combining them with manually preset rules. It eliminates the need for GPS to provide a reference for single-axis sensing, highlighting the core advantages of software-defined hardware and enhancing the inventive value of manually preset rules. The GPS module is only used to provide a geographical reference for astronomical orientation calculations, with a clear division of labor between it and spatial calculations, avoiding logical conflicts. The signal preprocessing steps, through filtering, detrending, and outlier removal operations, effectively remove signal deviations caused by environmental noise and equipment interference, providing a high-quality signal foundation for subsequent calculations and optimizations.

[0006] The algorithm-based spatial calculation module works in conjunction with astronomical orientation estimation. Spatial calculation, based on the temporal variations of the single-axis sensing signal and manually preset rules, expands the spatial dimension of the single-axis signal and performs preliminary orientation estimation. Astronomical orientation estimation combines GPS location data and real-time time data to deduce the absolute orientation of the sun or celestial bodies as a precise benchmark. The two are cross-checked to improve the accuracy of orientation estimation. The AI ​​fusion optimization module, through a generative compensation algorithm, integrates multi-source data (including data collected by the single-axis sensor module, preprocessed signal sequences, and environmental interference parameters) to trace and compensate for various errors in the spatial calculation, astronomical orientation estimation, and sensor acquisition processes, further improving signal quality.

[0007] The absolute orientation distribution entropy is the core original technical feature of this invention. It integrates the Shannon entropy and signal gradient variance of a single-axis sensing signal, serving as a comprehensive feature index for simultaneously quantifying the degree of single-axis absolute orientation deviation and sensing confidence. While Shannon entropy characterizes the statistical distribution stability of the signal, and gradient variance characterizes the drastic nature of signal temporal changes, this invention introduces a scenario-calibrated weighting coefficient λ for the first time, achieving an optimal balance between the two types of features. This allows for precise quantification of single-axis absolute orientation deviation, providing a reliable basis for control module judgment and adjustment. The weighting coefficient λ ranges from 0.3 to 0.7, determined through multiple sets of experimental calibrations in single-axis tracking and orientation sensing scenarios to ensure adaptability to different working conditions and avoid inaccurate deviation judgments due to a single feature. The formula for calculating the absolute orientation distribution entropy is as follows: F is the comprehensive characteristic value of the signal sequence after deep compensation of absolute orientation distribution entropy. The F value possesses a unique dual-function characteristic: on the one hand, it characterizes the drastic change of the signal through gradient variance, enabling single-axis absolute orientation deviation judgment; on the other hand, it characterizes the signal distribution stability through Shannon entropy, enabling quantitative evaluation of the confidence level of the sensing result. The F value integrates orientation estimation and confidence assessment into a single index, eliminating the need for separate calculations and dual threshold judgments, thus achieving a complete replacement and technological upgrade of the traditional single entropy value confidence method. The aforementioned perception confidence level refers to an index that quantitatively evaluates the reliability and usability of the current single-axis absolute orientation estimation result based on the distribution stability and temporal stationarity of the sensor signal. It characterizes the degree of credibility to which the orientation estimation result can be directly used. Its core function is to determine whether the current orientation estimation result can be directly used. High confidence indicates a stable signal and reliable result, which can be directly used; low confidence indicates a disordered signal and unreliable result, requiring generative compensation for reconstruction and optimization. The absolute azimuth distribution entropy provides a reliable basis for azimuth estimation and confidence level deviation judgment, effectively solving the problem of inaccurate deviation judgment caused by a single feature. This is the key to overcoming the technical bias of "single-axis must be integrated" in this invention, and it is also the most essential difference between this invention and the traditional scheme that only judges the direction deviation.

[0008] The control module achieves precise single-axis direction control by comparing the comprehensive feature F value with the preset comprehensive judgment threshold (1.16, 2.00). The core logic is "a single interval carrying dual judgments": the preset comprehensive judgment threshold is not two independent thresholds, but a unified interval containing two dimensionless judgment nodes, 1.16 and 2.00. Both nodes are dimensionless (because the absolute orientation distribution entropy F is a fusion of Shannon entropy and normalized gradient variance, both of which are dimensionless, hence the nodes are dimensionless), and their functions are clearly defined: 1.16 corresponds to the perception confidence credibility judgment (judging signal stability and result availability), and 2.00 corresponds to the direction deviation exceeding the limit judgment (judging the degree of direction deviation). Relying on the dual functional characteristics of the absolute orientation distribution entropy F value simultaneously representing perception confidence and direction deviation, a logical closed loop of first judging credibility and then judging deviation is completed. When F ≤ 1.16, the system synchronously determines that the perception confidence is high and the direction deviation is far below the allowable range, maintaining the current direction. When 1.16 < F ≤ 2.00, the system synchronously determines that the perception confidence meets the requirements and the direction deviation is within the allowable range, maintaining the current direction. When F > 2.00, the system synchronously determines that the perception confidence is insufficient and the direction deviation exceeds the limit, immediately outputting an adjustment command and initiating compensation to drive the single-axis execution module to adjust the direction, ensuring real-time perception and stability. The entire perception process is repeated to achieve continuous perception of single-axis absolute direction, meeting the needs of various single-axis light tracking and orientation positioning scenarios. Furthermore, when F > 2.00 (deviation exceeds the limit), the present invention, through the logical design of step 6), directly jumps back to step 3) for the AI ​​fusion optimization module to re-perform fusion compensation, without the need for additional deviation quantification estimation and correction processes, forming a complete closed loop of "deviation identification - re-compensation - secondary judgment". This solves the technical pain point of deviation exceeding the limit when F > 1.2, while simplifying the overall process. It also works in synergy with absolute orientation distribution entropy and generative compensation algorithms to improve the technical link, ensuring that the adjusted F value is ≤ 2.00 (actual deviation ≤ 0.1°), further enhancing the high precision and stability of single-axis sensing.

[0009] The AI ​​fusion optimization module also includes a generative compensation module, which employs a lightweight generative deep learning model (diffusion model, variational autoencoder, or generative adversarial network) deployed on edge computing units to meet the engineering application requirements of single-axis sensing systems. Model training data is generated from physical simulations using multi-condition samples, and the sample distribution is expanded by domain randomization. High-information-gain samples are actively selected and labeled to ensure model fusion accuracy and generalization ability. The generative compensation module triggers dual-mode operation based on perception confidence, which is determined by the comprehensive feature F-value. When the confidence is below a threshold, dimensionality reconstruction and depth compensation are initiated; when it is above or equal to the threshold, the fusion result is directly output, balancing perception accuracy and computational efficiency.

[0010] The F-value of the comprehensive feature of the depth-compensated signal is calculated by using the absolute orientation distribution entropy. The formula combines the signal statistical distribution stability represented by Shannon entropy with the signal temporal variation severity represented by gradient variance, and introduces a scenario-based calibration weight coefficient λ (0.3~0.7) to achieve optimal balance. It should be clarified that the absolute orientation distribution entropy algorithm and the generative compensation algorithm are independent and have their own functions: the former calculates the comprehensive feature F-value based on the compensated signal to achieve accurate deviation judgment and confidence assessment; the latter completes error source tracing and depth compensation, and outputs a high-quality signal sequence. The two work together to form a complete technical link, ensuring high accuracy, high stability and strong anti-interference capability of single-axis absolute orientation sensing, forming a complete technical closed loop.

[0011] In terms of the system, a layered decoupled architecture is constructed to decouple the hardware logic and algorithm functions. Error compensation and orientation fusion are achieved using generative models, thereby eliminating the inherent defects of single-axis sensors and realizing high-precision, high-continuity, and high-flexibility single-axis absolute orientation sensing. The system is divided into a hardware resource layer, a software definition layer, and an application service layer. Reconfigurable computing units enable dynamic software-based reconfiguration of the hardware logic. An error compensation mechanism is built using generative deep learning, fusing geomagnetic and solar orientation calculation results. Simultaneously, non-integral absolute orientation estimation is achieved based on single-axis sensor data modeling, and dynamic compensation is triggered by the absolute orientation distribution entropy. Finally, the robustness of perception is improved through software autonomous evolution and dynamic reconfiguration.

[0012] The hardware resource layer forms the physical foundation of the system, including single-axis sensor modules configured with preset rules and reconfigurable computing units. The single-axis sensor modules acquire raw absolute direction signals pointing towards the geographic North Pole; the reconfigurable computing units achieve millisecond-level dynamic reconfiguration of hardware logic through software instructions, without interrupting the perception service output. This achieves complete decoupling of hardware resources and perception algorithms, replacing physical hardware replacement with software reconfiguration, and providing an execution platform for the software-defined layer.

[0013] The software-defined layer is the core of the system, defining the functional logic of the hardware resource layer through software instructions. It supports remote online reconfiguration, autonomous iterative optimization, dynamic reconfiguration, hot-swapping, and version rollback. This layer is further divided into a perception definition module, a decision definition module, and an execution definition module. These modules are connected through standardized data interfaces, enabling the separation of perception, decision-making, and execution functions and independent software iteration. Upgrading a single module does not affect the operation of other modules, ensuring the continuous and stable operation of the system. The decision definition module integrates generative compensation functionality, constructing a generative adversarial network based on generative deep learning, including a generator network and a discriminator network. The generator learns the joint probability distribution and statistical characteristics of the systematic errors of geomagnetic azimuth and solar azimuth and environmental interference errors through adversarial training. The discriminator evaluates the compensation accuracy and back-optimizes the generator, improving the reliability of the compensation parameters. The network structure, training strategy, and parameters all support software definition and remote iterative updates, and an iterative convergence judgment mechanism is set up. When two consecutive accuracy improvements are less than a preset threshold, the parameters are automatically fixed, thus forming adaptive error compensation and achieving software-based autonomous evolution of perception accuracy.

[0014] The application service layer is deployed on top of the software-defined layer and is used to output the absolute azimuth angle in the geographic coordinate system. It transforms the calculation and compensation results into a directly usable orientation-aware service, realizing the software-defined service output.

[0015] The system employs a dual-directional calculation and fusion mechanism to enhance reliability: an astronomical azimuth algorithm provides an absolute direction reference, the geomagnetic azimuth is calculated based on geomagnetic sensor data and magnetic declination, and the solar azimuth is calculated based on time, geographical location, and astronomical algorithms; the generative compensation function fuses the two calculation results to eliminate the risk of single-point failure from a single azimuth source and improve system robustness. Attached Figure Description

[0016] Figure 1 The diagram shows the layered architecture of the sensing system, illustrating the three-layer structure and the flow of data and commands: 1. Application Service Layer; 2. Azimuth Data Output Unit; 3. AR Display Module; 4. Data Flow; 5. Software-Defined Layer; 6. Sensor Data Processing Module; 7. Azimuth Calculation Module; 8. Generative Compensation Module; 9. Raw Sensor Data Flow; 10. Hardware Resource Layer; 11. Single-Axis Sensor Module; 12. Reconfigurable Computing Unit; 13. Reconfigurable Control Flow / Software Command Reconfiguration Flow.

[0017] Figure 2The diagram illustrates the layered and decoupled architecture and module connections of the software-defined layer, showing the interaction relationships between the perception, decision-making, and execution modules, interfaces, instructions, and data: In the diagram, 14. Perception Definition Module; 15. Decision Definition Module; 16. Data Filtering Unit; 17. Geomagnetic Azimuth Calculation Unit; 18. Scale Transformation Unit; 19. Solar Azimuth Calculation Unit; 20. Generative Compensation Submodule; 21. Execution Definition Module; 22. Coordinate Transformation Unit; 23. Remote Instruction Receiving Unit; 24. Perception Service Encapsulation Unit; 25. Online Upgrade Interface; 26. Standardized Data Interface; 27. Flow of Remote Upgrade Instructions; 28. Flow of Error Feedback Optimization, located within the Generative Compensation Submodule, pointing from the discriminator network to the generator network.

[0018] Figure 3 The diagram shows the internal algorithm structure of the generative compensation module (8 / 20), illustrating the network composition, training, and compensation process: 20. Generative compensation submodule; 28. Error feedback optimization flow, from the discriminator network to the generator network; 29. ​​Generator network; 30. Discriminator network; 31. Input layer; 32. Input layer; 33. Hidden layer; 34. Hidden layer; 35. Output layer; 36. Output layer; 37. Iterative convergence determination unit; 38. Compensation execution unit; 39. Training data pool; 40. Compensation data flow; 41. Azimuth angle after precise compensation.

[0019] Figure 4 The diagram shows the complete timeline of the sensing method, illustrating the steps and signal flow from data acquisition to external output: Step 1: Data Acquisition; Step 2: Core Principle Implementation; Step 3: Hardware Reconstruction & Sensing Layer Definition; Step 4: Decision Layer Dual Algorithm Solution; Step 5: Decision Layer Generative Compensation; Step 6: Trigger Signal Flow; Step 7: Dynamic Compensation Trigger Signal Flow; Step 8: Preprocessed Data Flow; Step 9: Dynamic Compensation Trigger Module; Step 10: Execution Layer Service Output; Step 11: External Application Terminal; Step 2: External Service Output Flow; Step 3: Output Absolute Location Service to the External Entity.

[0020] Figure 5 This diagram illustrates the principle of non-integral absolute direction estimation and dynamic compensation triggering, showing the data input, model modeling, non-integral direction estimation, confidence determination, and compensation triggering process: In the diagram, 55. Stage 1: Single-axis sensor data input; 56. Stage 2: Generative model modeling; 57. Stage 3: Non-integral absolute direction estimation; 58. Stage 4: Confidence quantification & compensation triggering; 59. Threshold determination unit; 60. Compensation triggering result; 61. Model output flow direction; 62. Motion continuity constraint; 63. Flow direction of the absolute direction and the F-value of the absolute direction distribution entropy. Detailed Implementation

[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0022] For the overall architecture of the perception system, please refer to the appendix. Figure 1 The system in this embodiment includes three layers: Application Service Layer (1): Standard angle data is output through the azimuth data output unit (2), and the AR display module (3) completes visualization; Software Definition Layer (5): This is the core processing layer, where the sensor data processing module (6) completes preprocessing, the azimuth calculation module (7) completes geomagnetic / solar calculation, and the generative compensation module (8) completes error compensation; Hardware Resource Layer (10): The single-axis sensor module (11) collects raw direction data, and the reconfigurable computing unit (12) receives the reconfiguration command from the software definition layer (5) to realize dynamic configuration of hardware logic. The raw data is uploaded to the software definition layer (5) along the solid line (9); the processed data is uploaded to the application service layer (1) along the solid line (4); and the reconfiguration control command is sent to the reconfigurable computing unit (12) along the short dashed line (13).

[0023] For the internal architecture of the software-defined layer, please refer to [link / reference]. Figure 2 The software-defined layer (5) is divided into three modules: perception, decision-making, and execution. The perception definition module (14) uses a data filtering unit (16) to filter out noise and a scaling transformation unit (18) to normalize the data. The decision definition module (15) uses a geomagnetic azimuth calculation unit (17) and a solar azimuth calculation unit (19) to calculate the data respectively. The generative compensation submodule (20) completes error compensation. The execution definition module (21) uses a coordinate transformation unit (22) to convert the data to the geographic coordinate system and a perception service encapsulation unit (24) to encapsulate the output. Interface and instruction flow: The remote instruction receiving unit (23) receives the upgrade instruction and sends it to the online upgrade interface (25) along the short dashed line (27). Each module interacts through the standardized data interface (26). The dotted line (28) is the internal error feedback of the compensation submodule and does not cross modules.

[0024] For the working principle of generative compensation networks, please refer to [link / reference]. Figure 3 The generative compensation module (8 / 20) adopts a generative adversarial network. The generator network (29) receives sensor data and azimuth angle from the input layer (31), extracts features through the hidden layer (33), and outputs adaptive compensation parameters through the output layer (35). The discriminator network (30) receives compensation parameters and true error from the input layer (32), evaluates them through the hidden layer (34), and outputs accuracy evaluation and optimization signals through the output layer (36). The dotted line (28) feeds back the discrimination result to the generator to complete the iterative optimization. The iterative convergence judgment unit (37) solidifies the parameters after meeting the accuracy threshold and sends them to the compensation execution unit (38) along the solid line (40) to output the azimuth angle (41) after accurate compensation.

[0025] For the flowchart of the single-axis absolute orientation sensing method, please refer to [link / reference]. Figure 4 The perception method includes six steps: Step 1 Data acquisition (42): Acquire single-axis raw data, GPS, and motion features; Step 2 Core principle implementation (43): Perform non-integral direction estimation and confidence quantization; Step 3 Hardware reconstruction & perception layer definition (44): Configure reconfigurable hardware and complete filtering and scaling transformation; Send the preprocessed data to Step 4 (45) along the solid line (49); Send the trigger signal to Step 5 (46) along the solid line (47); Step 4 Decision layer dual algorithm solution (45): Geomagnetic + solar azimuth solution; Step 5 Decision layer generative compensation (46): Call Figure 3 The network completes the compensation; Step 6 Execution layer service output (51): Encapsulate the data and output it to the external application terminal (52). The dynamic compensation trigger module (50) determines the compensation process through the dot line (48) based on the preset threshold.

[0026] For the working principle of non-integral absolute direction estimation and dynamic compensation triggering, please refer to [link / reference]. Figure 5 It utilizes generative deep learning models, such as variational autoencoders (VAEs), to explicitly model the multi-peaked posterior probability distribution obtained from mapping single-axis sensing data. It encodes the temporal motion features (velocity, acceleration) of the target object into dynamic process constraints. Through physically guided latent space sampling, it generates an absolute direction probability distribution that conforms to motion continuity. The absolute direction is directly estimated from the probability distribution without the need for integration calculations, thus completely avoiding the accumulation of integration errors. At the same time, it calculates the absolute direction distribution entropy. When the F value is greater than a preset threshold, it determines that the perception confidence is low and immediately triggers a dynamic compensation mechanism, calling the generative compensation function to compensate for errors and improve perception accuracy. Specifically, it is divided into four stages: Stage 1 Data Input (55): Input sensor data, position and motion information, and enter the modeling stage along the solid line (4); Stage 2 Generative Model Modeling (56): Construct a multi-peak probability distribution model and output the direction estimate along the solid line (61); Stage 3 Non-integral Absolute Direction Estimation (57): Complete the non-integral solution under the motion continuity constraint (62) and output the direction and F value along the solid line (63); Stage 4 Confidence Quantification & Compensation Trigger (58): The threshold judgment unit (59) judges whether the compensation condition is met and outputs the compensation trigger result (60) through the dotted line (48).

[0027] Reasoning and demonstration regarding the value range (0.3~0.7) of the weighting coefficient λ in the formula for calculating the absolute orientation distribution entropy. In single-axis absolute orientation sensing, Shannon entropy alone cannot capture temporal dynamic changes, and gradient variance alone cannot reflect distribution stability. To overcome the limitations of single features, this invention employs a formula for calculating the absolute orientation distribution entropy that integrates and optimizes Shannon entropy and gradient variance, achieving complementarity between statistical and temporal characteristics. In this embodiment, extensive simulation tests and real-world scene calibrations were conducted on the weighting coefficient λ in the formula, determining its optimal value range to be 0.3~0.7. This value is not arbitrarily selected but determined through theoretical derivation and experimental verification based on the signal characteristics and physical meaning of single-axis tracking light scenarios. The specific calibration process is as follows:

[0028] 1. When λ < 0.3, the weight of gradient variance Var (∇x) is too low, and the comprehensive feature F value is insufficient to respond to key anomalies such as signal abrupt changes and angle jumps, resulting in a significant increase in the false negative rate in the abrupt change area, and it is impossible to effectively identify real deviations such as cloud cover and mechanical shaking.

[0029] 2. When λ > 0.7, the gradient variance weight is too high, and the comprehensive feature F value is too sensitive to small noise and normal fluctuations, which leads to a significant increase in the false trigger rate in the stable region, frequent malfunctions of the system, and a significant decrease in operational stability.

[0030] 3. When λ is in the range of 0.3 to 0.7, the Shannon entropy represents the overall stability of the signal distribution and the gradient variance represents the degree of instantaneous change of the signal, achieving the optimal balance. It can maintain low false triggering under stable conditions and high sensitivity under abrupt changes, which is fully suitable for the dual requirements of stability and accuracy in single-axis tracking scenarios.

[0031] 4. Extensive simulations and field measurements show that λ can achieve significantly better judgment results than single features in the range of 0.3 to 0.7. Among them, λ=0.5 is the typical value with the best comprehensive performance, which further verifies the rationality and universality of this value range.

[0032] Simulation comparison results show that the schemes using only Shannon entropy or only gradient variance have significantly higher false trigger rates in the stable region and false detection rates in the abrupt change region than the scheme of this invention. After adopting the calculation formula of absolute orientation distribution entropy of this invention and the λ=0.3~0.7 range, the system has stronger anti-interference ability under stable conditions and more timely response under abrupt change conditions, and the overall tracking accuracy and robustness are significantly improved.

[0033] Simulation verification of the method and range for determining the preset threshold. The preset threshold of this invention adopts a dual threshold logic mechanism of first judging the reliability and then judging the deviation, with a value range of (1.16, 2.00), and a preferred value of 1.66.

[0034] 1. Sample collection and grading strategy:

[0035] (1) Stable area samples: Covering interference-free and signal-stable scenarios, 15 groups were collected, with ≥1500 sampling points in each group.

[0036] (2) Samples from areas of sudden change: 15 sets of samples were collected from scenarios covered by cloud cover (sudden change in illumination), mechanical vibration (vibration interference), and signal distortion (electromagnetic interference).

[0037] (3) Transition region samples: 8 sets of signal quality edge scenarios, used to verify threshold boundary sensitivity.

[0038] Total: 38 sets of samples, sampling frequency 100Hz, continuous acquisition time ≥30 seconds / set

[0039] 2. Threshold interval (1.16, 2.00) calibration process:

[0040] (1) Reference calibration: Using a high-precision total station (accuracy ±0.005°) to provide the absolute direction truth, calculate the direction deviation and signal quality index Q for each group of samples.

[0041] (2) First-level threshold traversal: Traverse the range of 0.8 to 3.0 with a step size of 0.02 and record the system performance under different thresholds.

[0042] Evaluation metric A: False rejection rate in the stable region ≤ 5% (avoid over-filtering of normal signals)

[0043] Evaluation metric B: Correct identification rate of mutation regions ≥ 95% (ensuring effective isolation of interference).

[0044] Evaluation metric C: Transition region sensitivity (Area under ROC curve AUC ≥ 0.92)

[0045] 3. Second-level threshold verification: For samples that pass the first level, verify the angle error distribution:

[0046] Requirement: Angle error ≤ 0.1° within the 95% confidence interval.

[0047] Method: Kalman filter posterior estimation + sliding window validation

[0048] 4. Interval Determination: Based on the above indicators, the optimal threshold interval is determined to be (1.16, 2.00).

[0049] (1) Lower limit 1.16: The critical point that guarantees a mutation region recognition rate of ≥95%.

[0050] (2) Upper limit 2.00: A conservative upper limit to avoid over-filtering in the stable region.

[0051] 5. Optimal threshold of 1.66 - Validation results

[0052] In the single-axis tracking system embodiment, the results were verified through 8000 independent repeated experiments (including 2000 cross-scene verifications):

[0053]

[0054] Layering of physical meaning:

[0055]

[0056] Results Analysis: The optimal overall performance is achieved when the preset threshold is 1.66, resulting in: sensing accuracy ≤ 0.06°, system availability > 98%, complete execution of the dual-judgment logic, and no risk of single-level misjudgment. The preset threshold range and preferred value of this invention achieve two levels of protection—quality filtering and accuracy verification—through dual-judgment logic, fundamentally different from traditional single-level threshold schemes, ensuring high-precision output of highly reliable signals.

[0057] Simulation verification and performance comparison of absolute directional distribution entropy. To verify the technical effect of the absolute directional distribution entropy of the present invention, simulation comparison experiments were conducted for two core working conditions: the stable region (sensor stationary and interference-free) and the abrupt change region (cloud obstruction and signal distortion). The comparison objects were two existing schemes that use only Shannon entropy and only gradient variance to verify the superiority of the absolute directional distribution entropy of the present invention.

[0058] The simulation parameters are set to strictly match the actual working conditions of a single-axis tracking scenario, as detailed below.

[0059] Stable region signal: adopts normal distribution N(45, 1.0) to simulate the normal working state of a single-axis sensor when it is stationary and free from interference; Abrupt region signal: adopts mixed distribution (the first 50 sampling points are N(45, 1.0), and the last 50 sampling points are uniformly distributed U(30, 70) to simulate signal distortion and abrupt changes caused by cloud cover.

[0060] Sample size: 5000 independent repeated experiments are set to ensure the significance of the statistical results and avoid the randomness of a single experiment; Number of discrete intervals n: 10 is used to discretize the continuous signal, calculate the signal probability distribution, and provide a basis for Shannon entropy calculation; Weight coefficient λ: 0.5 is set, falling within the range of 0.3~0.7 as defined in the claims of this invention, and is used for feature fusion calculation of the combined formula of this invention.

[0061] Evaluation metrics and simulation results: Three core metrics were selected, namely, the false trigger rate in the stable region (the probability of misjudging a deviation when there is no anomaly), the false detection rate in the abrupt change region (the probability of not detecting an anomaly when there is one), and the distribution separation degree (the degree to which the feature values ​​of the stable region and the abrupt change region can be distinguished). The simulation results are shown in the table below:

[0062]

[0063] Results analysis:

[0064] 1. False Trigger Rate in Stable Region: The false trigger rate of the absolute direction distribution entropy of this invention is only 3%, far lower than that using only Shannon entropy (18%) and only gradient variance (22%). This indicates that under normal operating conditions where the sensor is stationary and there is no interference, this invention can accurately determine "high confidence and small deviation," effectively reducing false triggers in azimuth recognition under stable operating conditions, improving the stability of azimuth recognition, avoiding ineffective adjustments to the tracking system due to false triggers, reducing energy consumption, and improving system stability. 2. Missed Detection Rate in Abrupt Region: The missed detection rate of the absolute direction distribution entropy of this invention is only 2%, far lower than that using only Shannon entropy (12%) and only gradient variance (16%). This indicates that under abnormal operating conditions such as cloud cover and signal distortion, this invention can accurately determine "low confidence and large deviation," effectively reducing false triggers in azimuth recognition under stable operating conditions. 1. Accurately capture azimuth change signals, trigger timely adjustments and compensation, reduce missed detections, ensure that the tracking system responds to azimuth changes in a timely manner, and ensure the continuity and accuracy of single-axis tracking; 2. Distribution separation: The distribution separation of the absolute direction distribution entropy of the present invention (8.55) is comparable to that of gradient variance alone (8.59), and much higher than that of Shannon entropy alone (3.21), indicating that the present invention achieves dual optimization of false trigger rate and missed detection rate while ensuring high discrimination, and takes into account both stability and sensitivity; 3. Parameter compliance verification: The weight coefficient λ=0.5 strictly falls within the range of 0.3~0.7 defined in the claims of the present invention, verifying the feasibility and technical effect of the absolute direction distribution entropy of the present invention, and proving that the scope of protection of the present invention has practical technical support.

[0065] In summary, the simulation data comparison results show that the absolute orientation distribution entropy of the present invention, through the weighted fusion and optimization of Shannon entropy and temporal gradient variance, achieves performance far superior to existing single feature schemes in both the stable region and the abrupt change region, and has significantly improved orientation recognition accuracy, stability and anti-interference ability, effectively solving the defects of existing technologies.

[0066] This invention discloses a method and system for single-axis absolute direction sensing based on software-defined hardware. Its core lies in replacing hardware-dimensional expansion with innovative software architecture. By employing deterministic calculations using spatial and astronomical algorithms, physical interpretability is preserved. Its core solution formula can accurately and in real-time output the solar azimuth angle and the theoretical angle for single-axis tracking, achieving absolute direction estimation without relying on integration steps. This avoids the drawbacks of traditional integration operations, such as integration saturation, overshoot oscillation, and mechanical wear. Event-triggered compensation is performed using a deep learning model to handle uncertainties during tracking. The system's operating state is quantitatively characterized by the absolute direction distribution entropy, enabling adaptive allocation of computing resources. This architecture achieves single-axis absolute direction estimation without integral operations for the first time, overcoming the technical bias that single-axis tracking must use integral steps. It realizes technological understanding and path innovation. Compared with traditional control schemes with integrals, this invention has significant advantages such as high computational efficiency, strong robustness, and no steady-state error accumulation. It achieves a significant improvement in the accuracy and stability of single-axis absolute direction sensing, and has important engineering application value and promotion value.

Claims

1. A method for single-axis absolute direction sensing based on a software-defined hardware paradigm, comprising acquiring raw absolute direction sensing data pointing towards the geographic North Pole through a single-axis sensor module, and obtaining latitude and longitude geographic location data through a GPS module, characterized in that... The absolute direction estimation and compensation mechanism is triggered based on the following principles, and the hardware reconstruction, algorithm solution, error compensation, and service output steps are executed sequentially, specifically including: 1) A single-axis absolute direction sensing signal is acquired through a single-axis sensor module to obtain a continuous signal sequence x1, x2,..., x n Where n is the total number of signal sampling points and i is the sampling point number of 1≤i≤n; location data is collected through the GPS module to provide a geographical reference for subsequent solar / celestial body orientation estimation; the collected single-axis absolute direction sensing signal is preprocessed to remove noise interference in the signal and obtain a denoised signal sequence. The preprocessing process includes filtering, detrending and outlier removal operations to ensure the stability and reliability of the signal. 2) The algorithm solution module performs spatial solution and astronomical orientation calculation on the denoised signal sequence to obtain the spatial orientation component and astronomical reference direction. The spatial solution is based on the temporal changes of the single-axis sensing signal and the manually preset rules, while the astronomical orientation calculation is completed based on the location data and real-time time data collected by GPS. 3) The spatial orientation component, astronomical reference direction and multi-source data are fused by the AI ​​fusion optimization module. The multi-source data includes data collected by the single-axis sensor module, preprocessed signal sequence and environmental interference parameters. Generative compensation algorithm is used to deeply compensate for various errors in the fusion process and output the deeply compensated signal sequence. 4) Calculate the comprehensive feature F value of the signal after depth compensation using the absolute azimuth distribution entropy. The formula for calculating the absolute azimuth distribution entropy is: Wherein, H1 is the Shannon entropy of the one-dimensional directional signal, which is an indicator of information uncertainty or information content; H2 is the variance of the gradient of the one-dimensional directional signal within the sliding window, which characterizes the drastic degree of signal temporal change; λ is the weighting coefficient calibrated for the single-axis orientation perception scenario, with a value range of 0.3 to 0.

7. The specific value is specially calibrated according to the lighting conditions, environmental interference intensity, and signal characteristics of the single-axis tracking and orientation perception scenario; the absolute orientation distribution entropy has the dual function of simultaneously realizing absolute orientation deviation estimation and perception confidence. 5) The control module compares the F value with the preset comprehensive judgment threshold. If the F value exceeds the preset comprehensive judgment threshold, it indicates that the current direction estimation is not reliable enough and the direction deviation exceeds the limit. An adjustment command is output, and the process jumps directly back to step 3. The AI ​​fusion optimization module re-fuses the spatial orientation component, astronomical reference direction, and multi-source data, and strengthens the error depth compensation through a generative compensation algorithm. The depth-compensated signal sequence is re-output, and then step 4 is executed to control the single-axis structure to adjust to the correct absolute direction. If the F value does not exceed the preset comprehensive judgment threshold, it indicates that the current direction estimation is reliable enough and the direction deviation is within the allowable range. The current single-axis direction is maintained unchanged, and one single-axis absolute direction perception process is completed. 6) Repeat steps 1) to 5) above to achieve continuous, real-time sensing of the absolute direction of a single axis, ensuring sensing accuracy and stability.

2. A single-axis absolute direction sensing system based on a software-defined hardware paradigm, employing a single-axis sensor module and combining manually designed direction calculation rules to achieve absolute direction sensing functionality, also includes a GPS module, a data processing module, an algorithm calculation module, and a control module. Each module collaborates based on a hierarchical decoupled architecture, characterized by: A layered decoupling architecture decouples hardware and algorithms. It leverages spatial calculation algorithms to expand the spatial dimension of single-axis signals and perform preliminary absolute direction estimation. Combined with astronomical orientation algorithms, it provides a precise absolute direction benchmark. Then, an AI fusion optimization module employs a generative compensation algorithm. Based on multi-source data including single-axis observation data, preliminary spatial compensation results, and absolute astronomical reference orientation, it generates a dynamic compensation strategy, traces errors, performs deep compensation, and outputs a deeply compensated signal sequence, thereby obtaining a single-axis absolute direction estimate. The layered decoupling architecture is specifically divided into a hardware resource layer, a software definition layer, and an application service layer. Each layer interacts and transmits commands through standardized interfaces. The hardware resource layer integrates a single-axis sensor module, a GPS module, and a single-axis execution module to provide hardware support for the system. The software-defined layer deploys spatial calculation algorithms, astronomical orientation algorithms, generative compensation algorithms, and absolute orientation distribution entropy to decouple the algorithms from the hardware, and includes an algorithm calculation module and an AI fusion optimization module. The application service layer integrates the control logic and interaction functions of the control module. Relying on the control module, it completes single-axis absolute direction perception, deviation judgment, and adjustment command output. The single-axis sensor module is used to collect single-axis absolute direction perception signals and output a continuous signal sequence. The GPS module is used to collect location data to provide a geographical reference for astronomical orientation calculation. The algorithm solution module is used to perform spatial solution and astronomical orientation calculation on the preprocessed signal sequence, output spatial orientation components and astronomical reference direction, and provide a precise reference for single-axis absolute orientation; the control module is used to compare the F value obtained by absolute orientation distribution entropy with a preset threshold to determine the deviation of single-axis absolute orientation and output orientation adjustment command; the single-axis execution module is used to receive the adjustment command from the control module and drive the single-axis structure to adjust to the correct absolute orientation.

3. The method for single-axis absolute orientation sensing based on a software-defined hardware paradigm according to claim 1, characterized in that, The spatial calculation algorithm is based on single-axis sensing signals and temporal changes, combined with manually preset rules, to complete the spatial dimension expansion of single-axis signals and the preliminary calculation of single-axis absolute direction. The astronomical orientation algorithm combines GPS location data and real-time time data to deduce the absolute orientation of the sun or celestial bodies as a precise benchmark. The results obtained from the spatial calculation and astronomical orientation calculation provide basic input data for the absolute orientation distribution entropy. A single-axis sensor module with manually preset rules is used to collect sensing signals, and the modules work together to complete the single-axis absolute direction sensing. The spatial calculation algorithm is based on single-axis sensing signals and temporal changes, combined with manually preset rules, to complete the spatial dimension expansion of single-axis signals and the preliminary calculation of single-axis absolute direction. The astronomical orientation algorithm combines GPS location data and real-time time data to deduce the absolute orientation of the sun or celestial bodies as a precise benchmark. The results obtained from the spatial calculation and astronomical orientation calculation provide basic input data for the absolute orientation distribution entropy.

4. The method for single-axis absolute orientation sensing based on the software-defined hardware paradigm according to claim 1, characterized in that, The generative compensation algorithm integrates multi-source data through an AI model. The source errors include at least one of spatial calculation errors, astronomical orientation calculation errors, single-axis sensor module acquisition errors, and mechanical deviations. The deep-compensated signal output by the generative compensation algorithm is used as the input of the absolute orientation distribution entropy to support the accurate calculation of the F value. The absolute orientation distribution entropy refers to a comprehensive feature index that integrates and optimizes the Shannon entropy and signal gradient variance of the single-axis sensing signal, and is used to simultaneously quantify the degree of single-axis absolute direction deviation and the perception confidence. It has a single formula with dual functions, and can simultaneously realize the quantification of single-axis absolute direction deviation and the assessment of perception confidence.

5. The method for single-axis absolute orientation sensing based on a software-defined hardware paradigm according to claim 1, characterized in that, The preset threshold is a comprehensive judgment threshold, which includes two core judgment nodes, 1.16 and 2.00, both dimensionless values. The two nodes have clearly defined roles and work together to implement a single interval carrying a dual-judgment logic mechanism, providing a unified judgment benchmark for the logic of first judging credibility and then judging deviation, rather than two independent thresholds: 1.16 is the perceived confidence credibility judgment node, used to judge the credibility of the single-axis direction estimation result; 2.00 is the direction deviation exceeding the limit judgment node, which is a dimensionless feature mapping value. Its core corresponds to a single-axis absolute direction deviation ≤ 0.1°, that is, when the comprehensive feature F value ≤ 2.00, the actual direction deviation corresponding to the mapping is ≤ 0.1°. It is labeled as 2.00 instead of 2.00° to maintain consistency with the dimensionless attribute of the absolute orientation distribution entropy F value and avoid dimensional confusion. The value range of the preset comprehensive judgment threshold is (1.16, ...). 2.00), with a preferred value of 1.66 being the optimal intermediate node that balances reliability and deviation control; the perceived confidence level refers to an index that quantifies the reliability and usability of the current single-axis absolute direction estimation result based on the distribution stability and temporal stationarity of the sensor signal, and is used to characterize the degree of reliability to which the direction estimation result can be directly adopted; The smaller the F value, the smaller the absolute direction deviation of the single axis and the higher the perceived confidence. The larger the F value, the greater the absolute direction deviation of the single axis and the lower the perception confidence. Based on the comparison between the F value and the comprehensive judgment threshold, the control module first judges the credibility of the current direction estimation result, then judges whether the direction deviation exceeds the limit, and executes the corresponding control strategy accordingly.

6. A single-axis absolute orientation sensing system based on a software-defined hardware paradigm according to claim 2, characterized in that, The hardware resource layer, software definition layer, and application service layer of the layered decoupled architecture are physically separated through standardized interfaces, supporting remote independent upgrades of the software definition layer and application service layer without replacing the single-axis sensor module. The single-axis sensor module uses a manually preset two-layer mechanism to achieve absolute direction perception. The algorithm layer establishes a mapping relationship between sensor signals and absolute direction through direction calculation rules; the signal layer ensures the quality of the input signal and meets the calculation requirements through sampling frequency, noise threshold, and effective range parameters. The two layers work together to complete the direction perception function. The perception system includes a single-axis sensor module with manually preset rules and various collaborative modules. The layered decoupled architecture achieves hardware and algorithm decoupling to achieve high-precision single-axis absolute direction perception. The algorithm calculation module is equipped with an algorithm switching unit. When the GPS module signal is abnormal, it can call the location data collected and stored before the GPS module signal was abnormal, or rely on the backup positioning reference data reserved by the system, and switch to providing an accurate absolute direction reference only through the astronomical orientation algorithm to maintain the basic perception function of the system.

7. A single-axis absolute orientation sensing system based on a software-defined hardware paradigm according to claim 2, characterized in that, The neural network model used in the generative compensation algorithm of the AI ​​fusion optimization module can adaptively adjust the dynamic compensation strategy according to real-time multi-source data. The AI ​​fusion optimization module includes a generative compensation module, which is configured to: respond to a perceived confidence level lower than a preset threshold, fuse the cross-validation information of the spatial orientation component and the astronomical reference direction based on a generative deep learning model, reconstruct the missing dimension information, and output the compensated absolute direction estimate; and is further configured to: respond to a perceived confidence level higher than or equal to the preset threshold, directly output the fusion result of the spatial orientation component and the astronomical reference direction as the absolute direction estimate. The perception confidence level is determined based on the F value; meanwhile, the control module includes a communication unit for transmitting the single-axis absolute direction perception results, deviation data, and adjustment commands to external devices to achieve data interaction and remote control.

8. A single-axis absolute orientation sensing system based on a software-defined hardware paradigm according to claim 2, characterized in that, The AI ​​fusion optimization module employs a generative deep learning model, which is one of a diffusion model, a variational autoencoder, or a generative adversarial network. This generative deep learning model is deployed on the edge computing unit after lightweight processing. The lightweight processing includes: transferring the generative capabilities of the teacher model to the lightweight student model based on knowledge distillation; compressing the model parameter weights to INT8 precision using quantized perception training; and obtaining the optimal minimum subnetwork structure matching the computing power of the edge computing unit through neural architecture search. The training data construction method for the generative deep learning model includes: generating single-axis sensor output samples under multiple operating conditions based on physical simulation; expanding the sample distribution coverage through a domain randomization strategy; and using an active learning strategy to screen high information gain samples for manual annotation for model training optimization. This ensures that the model can accurately fuse cross-validation information of spatial orientation components and astronomical reference directions, completing the reconstruction of missing dimensions and absolute direction estimation. The acquisition of the single-axis absolute direction estimation does not rely on integral operations.

Images