Systems and Intelligent Decision-Making Methods Based on AI Multi-Dimensional Models

Abstract

This invention relates to the field of artificial intelligence computing models and intelligent decision control technology, and discloses a system and intelligent decision-making method based on an AI multi-dimensional model, including a feature perception module, a topology reconstruction engine, a feedback scheduling unit, and a transformation processing unit. The feature perception module extracts feature vectors of each component of a high-dimensional heterogeneous data stream, calculates the fusion value of the gradient, and determines the active feature dimension. The topology reconstruction engine generates a dynamic mask based on the active feature dimension and performs edge pruning on the initial topology graph to construct a task-sensitive topology structure. The feedback scheduling unit uses reinforcement learning operators to adjust the pruning threshold and update the dynamic mask coverage. The transformation processing unit performs weighted mapping operations based on the updated dynamic mask. This invention transforms full matrix operations into sparse topology operations through feature-guided logical pruning, reducing the overhead of invalid digital processing.

Description

Technical Field

[0001] This invention relates to a system and intelligent decision-making method based on an AI multi-dimensional model, belonging to the field of artificial intelligence computing models and intelligent decision control technology. Background Technology

[0002] High-dimensional heterogeneous data processing maps source data to feature space, establishes mathematical models to perform electro-digital transformations and logical calculations. The general processing mode is to ensure the integrity of feature distribution by constructing a global mapping matrix covering all feature dimensions to accommodate feature vectors. When the data source has strong sparsity or non-homogeneous features, traditional AI multi-dimensional decision-making models will generate computational redundancy due to the use of full mapping. The electro-digital processing unit needs to access memory addresses containing zero-value components to perform weighted aggregation operations, resulting in floating-point operation redundancy and memory access overhead. The static indexing mechanism cannot follow the transient evolution of feature axes. When the data distribution changes abruptly, mapping offset deadlock occurs, making it difficult for the model to capture transient features in real time, which seriously affects the real-time processing and the accuracy of intelligent decision-making.

[0003] While hardware improvements provide fundamental support, data processing efficiency is limited by software control logic and scheduling strategies. For example, Chinese invention patent CN119668526B discloses a method, device, and storage medium for improving data reliability in solid-state drives. It monitors the programming progress of migrated data through a timer and actively fills in the data to be supplemented after the timer expires to ensure programming integrity. This technical approach uses a timing compensation mechanism to solve the data reliability problem at the physical storage level. In essence, it is still a passive response scheme based on static time thresholds. When processing high-dimensional transient data streams, it lacks awareness of the internal semantic topology of the data, cannot identify differences in feature dimensions, and cannot adjust computing resources in real time according to the dynamic evolution of the data stream. When faced with a large number of invalid memory accesses and instruction pipeline reordering caused by the sudden changes in heterogeneous data sources, it is difficult to solve the contradiction between redundant operations and transient response at the underlying instruction level.

[0004] Therefore, how to construct a task-sensitive topology structure that can filter invalid feature axis calculation paths while achieving feature transfer self-calibration is the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: a system based on an AI multi-dimensional model, comprising: a feature perception module, a topology reconstruction engine, a feedback scheduling unit, and a transformation processing unit;

[0006] The feature perception module is used to receive high-dimensional heterogeneous data streams, extract feature vectors of each heterogeneous component in the high-dimensional heterogeneous data streams, calculate the fusion value of the first-order gradient and second-order gradient of the feature vectors within a preset sliding window, and determine the dimension whose fusion value exceeds a preset fluctuation threshold as the active feature dimension.

[0007] The topology reconstruction engine is connected to the feature perception module to generate dynamic masks based on the active feature dimensions. The dynamic masks are then used to perform edge pruning on the initial topology graph of the high-dimensional heterogeneous data stream to build a task-sensitive topology structure.

[0008] The feedback scheduling unit is connected to the topology reconstruction engine to obtain the mapping residual of the task-sensitive topology and the computational delay in the electronic digital processing. The mapping residual and computational delay are input into the preset reward function in the reinforcement learning operator. The reward function guides the reinforcement learning operator to adjust the decision threshold of edge pruning, thereby updating the coverage of the dynamic mask and keeping the mask bits in the active state in the dynamic mask aligned with the high-entropy feature axis in the high-dimensional heterogeneous data stream in the time domain.

[0009] The transformation processing unit is used to perform weighted mapping operations only on the active feature dimensions selected by the dynamic mask based on the updated dynamic mask, shield the zero-value operation logic outside the coverage of the dynamic mask, and output the electrical digital processing result.

[0010] Preferably, the feature perception module calculates the activity evaluation factors for each dimension using the following formula: Where S is the activity evaluation factor, The first gradient of the eigenvector. α is the second gradient of the feature vector, and β is the preset weight coefficient; the feature perception module determines the dimension whose activity evaluation factor exceeds the preset fluctuation threshold as the active feature dimension.

[0011] Preferably, the feedback scheduling unit is further configured to set an activation threshold and a release threshold for the edge pruning determination threshold based on the historical update frequency of the feature vector, wherein the activation threshold is greater than the release threshold; when the fusion value of the first-order gradient and the second-order gradient exceeds the activation threshold, the corresponding mask bit is activated to reconstruct the topological edge; when the fusion value is between the activation threshold and the release threshold, the mask bit is kept in an active state; when the fusion value is continuously lower than the release threshold for a period exceeding a preset self-sustaining period, the mask bit is revoked.

[0012] Preferably, the transformation processing unit includes an index scheduling subunit, which is used to calculate the address step size between each active bit in the dynamic mask and generate a run-length index table; according to the run-length index table, it controls the memory addressing pointer to skip the masked feature dimensions in the dynamic mask and execute the loading of feature vectors and electrical-to-digital transformation in a non-contiguous memory access mode.

[0013] Preferably, the feature perception module is also used to extract the projection features of the masked components using the complement of the dynamic mask; the topology reconstruction engine is also used to generate a reference fingerprint that characterizes the stability of the feature space based on the projection features; the feedback scheduling unit is also used to trigger a global mask recalibration instruction when the fluctuation amplitude of the reference fingerprint exceeds a preset stability threshold, so as to correct the evolution deviation of the computational model caused by the source reference drift of the high-dimensional heterogeneous data stream.

[0014] Preferably, the feature perception module is also used to inject an excitation signal with a preset encoding frequency into the high-dimensional heterogeneous data stream; the topology reconstruction engine is also used to generate a shadow mask based on the response features of the excitation signal and calculate the logical divergence between the shadow mask and the current dynamic mask; the feedback scheduling unit is also used to update the edge pruning judgment threshold in advance when the logical divergence exceeds a preset deviation threshold.

[0015] Preferably, the transformation processing unit is also used to count the frequency of state flips of each mask bit in the dynamic mask within a preset period and generate a flip entropy index; the feedback scheduling unit is also used to evaluate the reputation score of each heterogeneous component based on the flip entropy index, and when the reputation score is lower than a preset reputation threshold, to perform weight downgrading processing on the feature vector of the corresponding dimension.

[0016] Preferably, it also includes a phase-locking unit, used to record the mask bit flip pulses corresponding to different heterogeneous data sources, generate the logic sequence of each data channel; calculate the cross-correlation peak value between each logic sequence to determine the relative time delay offset between heterogeneous channels; adjust the processing phase of each heterogeneous channel in the transformation processing unit according to the relative time delay offset, so as to realize the timing alignment of heterogeneous features in the topology reconstruction process.

[0017] Preferably, in the reward function set by the feedback scheduling unit, the positive reward is provided based on the reduction in computation latency, and the negative reward is provided based on the increase in mapping residual. The transformation processing unit adopts a distributed operator array, performs multiplication and accumulation operations on the active feature dimension according to the enable signal of the dynamic mask, and shields the zero-value memory access logic outside the coverage of the dynamic mask.

[0018] An intelligent decision-making method based on an AI multi-dimensional model includes the following steps:

[0019] Step 101: Receive high-dimensional heterogeneous data stream, extract feature vectors of each heterogeneous component in the high-dimensional heterogeneous data stream, calculate the fusion value of the first-order gradient and second-order gradient of the feature vector within a preset sliding window, and determine the dimension with the fusion value exceeding the preset fluctuation threshold as the active feature dimension.

[0020] Step 102: Generate a dynamic mask based on the active feature dimension, and use the dynamic mask to perform edge pruning on the initial topology graph of the high-dimensional heterogeneous data stream to construct a task-sensitive topology structure.

[0021] Step 103: Obtain the mapping residual of the task-sensitive topology and the computation delay in the electronic digital processing. Input the mapping residual and computation delay into the preset reward function in the reinforcement learning operator. Use the reward function to guide the reinforcement learning operator to adjust the decision threshold of edge pruning, thereby updating the coverage of the dynamic mask, so that the mask bits in the active state in the dynamic mask are aligned with the high-entropy feature axis in the high-dimensional heterogeneous data stream in the time domain.

[0022] Step 104: Based on the updated dynamic mask, perform weighted mapping operation only on the active feature dimensions selected by the dynamic mask, shield the zero-value operation logic outside the coverage of the dynamic mask, output the electrical digital processing result, and generate intelligent decision instructions for high-dimensional heterogeneous data streams based on the processing result.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. In high-dimensional heterogeneous data processing, the feature perception module calculates the variation gradient of feature vectors in real time and generates dynamic masks in conjunction with the topology reconstruction engine. This transforms the high-dimensional full matrix mapping operation into a sparse topology operation modulated by the mask, reducing the overhead of invalid electrical and digital processing at the underlying instruction set level. By using feature-guided logic pruning, the computing resources of the electrical and digital processing unit are concentrated on the high-entropy feature axis. When processing heterogeneous data streams, this avoids the memory access redundancy and floating-point operation overload caused by traditional static mapping, improving the system's throughput performance for massive heterogeneous data while maintaining the fidelity of feature mapping.

[0025] 2. A threshold arbitration mechanism is constructed using reinforcement learning operators to obtain task-sensitive topology mapping residuals and dynamically adjust the edge pruning decision threshold to achieve an adaptive balance between computational precision and operational efficiency. The feedback scheduling logic transforms algorithm parameter adjustment into real-time control of topology reconstruction depth, enabling the system to offset transient feature axis migrations when facing nonlinear mutations in heterogeneous data sources by adjusting the dynamic mask coverage. Compared to pre-trained large-scale neural networks, this invention optimizes small-scale scalar parameters to enable the system to have transient adaptive capabilities for handling complex heterogeneous mappings in environments with limited hardware resources.

[0026] 3. To determine the threshold, a dual threshold gate is configured with an enable threshold and a release threshold. A logical self-sustaining cycle is introduced to construct a mask update mechanism with logical lag inertia, which suppresses topology jitter caused by random noise. When the feature variation gradient is in a specific fluctuation range, the mask bit remains active, avoiding cache invalidation and instruction pipeline reordering caused by frequent reconstruction of the logical path. Logical path anchoring based on the lag effect ensures that the backend computing architecture maintains the continuity of instruction execution and timing stability under non-steady-state conditions. Attached Figure Description

[0027] Figure 1This invention provides the overall system architecture design and multi-dimensional information closed-loop flow diagram.

[0028] Figure 2 This is an adaptive optimization curve of the dynamic mask coverage and computation delay D of the present invention;

[0029] Figure 3 This is a flowchart of the internal logic timing interaction and edge pruning process of the topology reconstruction engine of this invention. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] This invention proposes a system based on an AI multi-dimensional model, comprising a feature perception module, a topology reconstruction engine, a feedback scheduling unit, and a transformation processing unit. These modules are connected via an internal data bus or on-chip network to collaboratively complete the data processing flow from feature extraction, logic pruning, parameter optimization to sparse computation. The feature perception module receives high-dimensional heterogeneous data streams and extracts the feature vectors of each heterogeneous component. The feature perception module is equipped with a differential operator circuit to calculate the first-order gradient of the feature vector within a preset sliding window. and second gradient The fusion value, in the specific implementation process, is determined by the formula... The activity evaluation factor S for each dimension is calculated. α and β are preset weight coefficients used to balance the proportion of change rate and change acceleration in the activity determination. The feature perception module uses a numerical comparator to compare the activity evaluation factor S with the preset fluctuation threshold λ. Dimensions with S exceeding λ are identified as active feature dimensions, and dimensions with S less than or equal to λ are marked as silent dimensions.

[0032] The topology reconstruction engine is connected to the feature perception module and is used to generate dynamic masks based on the active feature dimension. The dynamic mask is a binary vector aligned with the feature vector dimension, where the mask bits corresponding to the active feature dimension are set to an active state (e.g., logic high level), while the mask bits corresponding to the silent dimension are set to a masked state (e.g., logic low level). The topology reconstruction engine uses the dynamic mask to perform edge pruning on the initial topology graph of the high-dimensional heterogeneous data stream. The engine cuts the topology edges connected to silent feature nodes through logical AND operations, thereby constructing a task-sensitive topology structure containing only active feature connection paths. The feedback scheduling unit is connected to the topology reconstruction engine and is used to construct a parameter adaptive closed loop based on reinforcement learning. This unit obtains the mapping residual E of the task-sensitive topology structure and the computational delay D in the electronic digital processing process, where the mapping residual... E represents the reconstruction error of the pruned topology on the original data features, and the computation delay D represents the number of clock cycles consumed in processing the current data frame. The feedback scheduling unit inputs the mapping residual E and the computation delay D into the preset reward function in the reinforcement learning operator, and is configured to output a positive reward when the computation delay D decreases and the mapping residual E remains within the allowable range; and output a negative reward when the mapping residual E increases beyond the allowable range. The reward function guides the reinforcement learning operator to adjust the decision threshold for edge pruning (i.e., the aforementioned preset fluctuation threshold λ or the related hysteresis threshold), thereby updating the coverage of the dynamic mask, so that the mask bits in the active state in the dynamic mask are aligned with the high-entropy feature axis in the high-dimensional heterogeneous data stream in the time domain, and establishes the state space S={D,E}. The feedback scheduling unit samples the normalized delay D according to the computation cycle. norm With mapping residual E norm According to the formula The reward value is calculated, and the weights k1 and k2 are determined by grid search based on hardware resource redundancy. The action space is discretized into the decision threshold step size. 0. If the current reward difference ΔR < 0, trigger. The strategy executes a reverse search and updates the judgment threshold. This guides the mask coverage to converge toward the Pareto front of the delay and residual, achieving high-entropy feature axis locking. The initial value of the decision threshold λ0 is dynamically preset based on 1.2 to 1.5 times the ambient noise reference level, ensuring the convergence stability of reinforcement learning.

[0033] To suppress frequent topology jitter caused by signal noise, the feedback scheduling unit configures an enabling threshold δ for edge pruning based on the historical update frequency of the feature vector. on and release threshold δ off And set δ on >δ offWhen performing threshold decision-making, a hysteresis comparison logic is followed: when the fused value of the first-order gradient and the second-order gradient exceeds the opening threshold δ... on When the fusion value drops but remains above the activation threshold δ, the corresponding mask bit is activated to reconstruct the topology edge; when the fusion value drops but remains above the activation threshold δ, the corresponding mask bit is activated to reconstruct the topology edge. on With release threshold δ off During this period, the mask bits remain active; when the fusion value remains below the release threshold δ, off The duration exceeds the preset self-holding period T hold At that time, the mask bit cancellation operation is performed; the transformation processing unit is used to perform weighted mapping operation only on the active feature dimensions selected by the dynamic mask based on the updated dynamic mask, and to shield the zero-value operation logic outside the coverage of the dynamic mask. The transformation processing unit includes an index scheduling subunit, which calculates the address step size between each active bit in the dynamic mask and generates a run-length index table. The direct memory access (DMA) controller or memory management unit controls the memory addressing pointer to skip the masked feature dimensions in the dynamic mask according to the run-length index table, and performs feature vector loading in non-contiguous memory access mode. The operator array in the transformation processing unit performs multiplication-accumulation and other electrical-digital transformations on the loaded active feature data and outputs the electrical-digital processing result. The index scheduling subunit traverses the dynamic mask M and calculates the physical address increment of adjacent active bit indices i,j. ,in For component bit width, the generated run-length index table records the non-zero feature offset sequence. The storage addressing pointer drives the DMA controller to perform discontinuous transfer according to the offset sequence. The transfer step size is aligned with the 64-bit bus width. The physical memory address corresponding to the masked dimension is directly skipped. The discrete feature vector is rearranged into a continuous operator array input stream. Invalid memory slot access cycles are masked to achieve linear matching between storage throughput and computationally active dimensions. The loading process follows a discontinuous memory access protocol. With the distributed multiply-accumulate logic of the operator array, high-dimensional data sparse transformation is completed. The resulting electro-digital processing result is further input into the intelligent decision logic. Based on the feature strength and evolution trend corresponding to the processing result, decision instructions for the application scenario are automatically generated, such as outputting fault warning level or shutdown instruction in industrial monitoring, thereby completing the closed loop from data perception to model reconstruction to intelligent decision-making.

[0034] To address potential data source baseline drift, the feature perception module extracts the projection features of the masked components using the complement of the dynamic mask. The topology reconstruction engine generates a baseline fingerprint characterizing the stability of the feature space based on the projection features. The feedback scheduling unit monitors the statistical characteristics of the baseline fingerprint, such as the mean shift. When the fluctuation amplitude of the baseline fingerprint exceeds a preset stability threshold, a global mask recalibration command is triggered to correct the evolution deviation of the computational model caused by the source baseline drift of the high-dimensional heterogeneous data stream. To address the lag issue during mode switching of heterogeneous data streams, the feature perception module injects an excitation signal with a preset encoding frequency into the high-dimensional heterogeneous data stream. The topology reconstruction engine generates a shadow mask based on the response characteristics of the excitation signal and calculates the logical divergence between the shadow mask and the current dynamic mask. The feedback scheduling unit detects when the logical divergence exceeds a preset stability threshold. When the deviation threshold is reached, it is determined that the data mode is about to change, and the determination threshold for edge pruning is updated in advance. In addition, to improve the system's ability to suppress false features, the transformation processing unit is also used to count the state flip frequency of each mask bit in the dynamic mask within a preset period, and generate a flip entropy index. The feedback scheduling unit evaluates the reputation score of each heterogeneous component based on the flip entropy index, and performs weight downgrading processing on the feature vector of the corresponding dimension when the reputation score is lower than the preset reputation threshold. The system also includes a phase locking unit, which is used to record the mask bit flip pulses corresponding to different heterogeneous data sources, generate the logical sequence of each data channel, calculate the cross-correlation peak value between each logical sequence to determine the relative time delay offset Δt between heterogeneous channels, and adjust the processing phase of each heterogeneous channel in the transformation processing unit according to the relative time delay offset Δt.

[0035] Example 1: In a high-dimensional heterogeneous data processing application scenario for condition monitoring of large industrial units, the system faces the challenge of processing massive amounts of multi-source heterogeneous signals in real time. The input data stream in this scenario consists of high-frequency vibration spectrum features with a sampling rate of 20000Hz, mid-frequency acoustic emission signals, and low-frequency infrared thermal imaging features, forming a high-dimensional heterogeneous feature vector space with a total dimension exceeding 1700. During stable operation of the unit, over 95% of the dimensions in this high-dimensional vector are in a state of low-amplitude random noise, with only a few feature dimensions carrying effective operational information. When a small crack propagates within the unit, or... When transient faults such as loose components occur, the characteristic components of a specific frequency band will experience drastic amplitude changes within milliseconds. The system must complete the entire process from feature capture to logical deduction within a processing latency of no more than 5 milliseconds. If the traditional full matrix mapping method is used, performing fully connected operations on 1700-dimensional data will generate a huge floating-point computational load, resulting in a computation latency D exceeding 50 milliseconds, thus missing crucial transient fault fingerprints. To address this situation, this system uses a feature sensing module to access this high-dimensional heterogeneous data stream in real time. This module utilizes a built-in differential operator circuit to process the input vector v for each feature dimension i. iThe first-order gradient is calculated synchronously within a preset sliding window of length 16. With second gradient The system sets the preset weight coefficients α to 0.6 and β to 0.4, and obtains the activity evaluation factors S for each dimension through register-level operations. i ,Right now In the initial state, a preset fluctuation threshold is set. Set to 1.5 times the reference noise level, when a vibration frequency band experiences abnormal fluctuations due to an early fault, its corresponding activity evaluation factor S... i The value rapidly increases and exceeds λ, based on which the topology reconstruction engine determines the mask bit M of this dimension. i Set to active state and perform edge pruning on the initial topology graph, cutting off all logical edges connected to the silent dimension, constructing a task-sensitive topology structure containing only dozens of active nodes. The index scheduling subunit in the transformation processing unit then generates a run-length index table containing the address steps of the active nodes, and controls the direct memory access controller to load only this part of the valid data into the operator array.

[0036] When changes in the external environment cause an overall increase in background noise, many feature components that originally belonged to the silent dimension experience gradient fluctuations due to noise interference, causing their activity evaluation factor S value to exceed the initial λ. At this time, the coverage of the dynamic mask M increases abnormally, the amount of invalid computation surges, and the computation latency D rises. Since noise is misclassified as a feature, the mapping residual E of the topology structure also increases. The feedback scheduling unit captures this non-ideal state and substitutes it into the reward function. Where k1 and k2 are preset constants, the current reward value R is calculated to decrease. Based on this negative feedback, the built-in reinforcement learning operator performs parameter optimization, outputting instructions to increase the edge pruning decision threshold λ, and simultaneously expand the opening threshold δ. on With release threshold δ off After several processing cycles of iteration, the system automatically converges to a new operating point, i.e., a higher hysteresis interval. Filtering gradient fluctuations caused by background noise, with a wider hysteresis interval δ on -δ off Ensures that even under noise interference, the true fault characteristics, once activated by the mask bits, can be detected within the self-sustaining period T. hold The internal logic connection is kept stable to avoid topology reconstruction oscillation caused by signal jitter. Finally, the transformation processing unit still performs weighted mapping operation only on fault feature axes with high semantic value in a strong noise environment, realizing adaptive sparse processing of high-dimensional heterogeneous data in edge gateways with limited hardware resources.

[0037] Example 2: To verify the effectiveness and stability of the high-dimensional heterogeneous data processing system based on semantic topology reconstruction and reinforcement learning in a real engineering environment, a hardware-in-the-loop simulation test platform containing multi-source heterogeneous sensor data streams was built. This platform simulates a typical high-throughput industrial IoT gateway node. Its input is connected to a mixed signal source from a real factory environment, including 128 vibration acceleration signals (sampling rate 10kHz), 64 temperature and pressure signals (sampling rate 100Hz), and 4 real-time video stream feature vectors captured by industrial cameras (30 frames per second, 512-dimensional features extracted per frame). All signals are mapped to a unified feature vector space with a total dimension of 2048 after analog-to-digital conversion. To simulate channel interference and sensor drift in real working conditions, Gaussian white noise with a signal-to-noise ratio of 15dB is superimposed on the original data stream, and a sensor reference drift disturbance of 5% is randomly introduced.

[0038] In this embodiment, two control groups were set up for comparative testing with the experimental group of the present invention. Control group A (full calculation group) adopted the traditional fully connected matrix operation mode, that is, weighted mapping was performed on all 2048 dimensions without distinction. Control group B (static threshold pruning group) adopted a fixed edge pruning threshold (set to λ=2.0) and did not have a reinforcement learning feedback adjustment mechanism. The experimental group of the present invention fully enabled feature perception, dynamic mask generation and reinforcement learning threshold arbitration mechanism. The test duration was set to 10 minutes, covering the entire process of the device from stable operation to simulated sudden failure (duration 500ms) and then back to stability. During the test, the single frame processing latency D (unit: milliseconds), feature reconstruction residual E (normalized Euclidean distance) and system power consumption (unit: watts) were monitored and recorded in real time. The test data showed that in the stable operation phase, the experimental group of the present invention calculated the activity evaluation factor S through the feature perception module, identified the active feature dimensions that accounted for only about 3% of the total dimensions, and generated the corresponding dynamic mask. The index scheduling subunit controlled the direct memory access controller to skip a large number of silent dimensions and only load the valid data.

[0039] Table 1: Example of performance comparison data during stable operation phase

[0040]

[0041] As shown in Table 1, although control group A had a zero reconstruction residual (theoretically lossless), its processing latency was as high as 48.5ms, approaching the critical value required for system real-time performance, and its power consumption was the highest. While control group B significantly reduced latency and power consumption, its fixed threshold could not accommodate some low-amplitude but semantically important features, such as early weak fault signals, resulting in a high reconstruction residual (0.15) and a risk of feature loss. The experimental group of this invention, while maintaining low latency (3.4ms) and low power consumption (4.3W) comparable to control group B, reduced the reconstruction residual to 0.04. Therefore, the reinforcement learning operator, based on the reward function... Guided by this, the judgment threshold λ is dynamically fine-tuned to find a better balance between filtering noise and preserving weak features.

[0042] At the 300-second mark of the experiment, a simulated transient fault signal was injected into the system, and the background noise level was temporarily increased to 10 dB to simulate severe operating conditions. At this time, because the control group B had a fixed threshold, a large amount of noise was misjudged as active features, causing the processing latency to spike to over 25 ms instantly, resulting in data backlog. However, the feedback scheduling unit of the experimental group of this invention quickly captured the abnormal fluctuation of the mapping residual E, and the reinforcement learning operator activated the threshold δ within 5 clock cycles. on Increase and widen δ on With δ off The hysteresis interval between the two suppresses mask bit flipping caused by noise, enabling the system to still lock the fault characteristic axis under strong noise interference. In addition, the test on sensor reference drift shows that when a 5% random drift disturbance is introduced, the output results of control groups A and B both show DC bias error. However, the experimental group of this invention uses the complement of the dynamic mask to extract the projection features of inactive components and generate a reference fingerprint. When the statistical mean deviation of the reference fingerprint is detected to exceed the preset stability threshold, the system triggers a global mask recalibration command to automatically correct the evolution deviation of the calculation model caused by the source reference drift.

[0043] Example 3: This example combines Figures 1 to 3 This section describes a system and intelligent decision-making method based on an AI multi-dimensional model, such as... Figure 1As shown, the system includes a feature perception module, a topology reconstruction engine, a feedback scheduling unit, and a transformation processing unit. The feature perception module extracts heterogeneous component feature vectors and calculates gradient fusion values ​​by accessing high-dimensional heterogeneous data streams, thereby determining the active feature dimensions and transmitting them to the topology reconstruction engine. The topology reconstruction engine generates a dynamic mask and performs edge pruning on the initial topology graph, thereby constructing a task-sensitive topology structure and outputting it to the transformation processing unit. The transformation processing unit performs active feature weighted mapping based on the dynamic mask and masks zero-value operation logic, outputting the electrical digital processing result. Meanwhile, the feedback scheduling unit obtains the mapping residual and computation delay generated by the transformation processing unit through the feedback path shown by the dashed line, adjusts the edge pruning judgment threshold using internal reinforcement learning operators, and outputs an instruction to update the dynamic mask coverage to the topology reconstruction engine, realizing closed-loop dynamic optimization of heterogeneous data processing overhead.

[0044] like Figure 2 As shown in the figure, the horizontal axis represents the dynamic mask coverage as a percentage, and the vertical axis represents the computation latency D in milliseconds. The multiple hollow dots in the figure represent discrete operating state points of the system under different mask configurations, reflecting the positive correlation between computation latency D and the increase in dynamic mask coverage. The dashed lines running through some state points clearly indicate the reinforcement learning optimization path, demonstrating that the reinforcement learning operator in the feedback scheduling unit guides the system parameters to converge towards a lower computation latency D along this path through continuous iteration, while ensuring the fidelity of feature mapping. Figure 3 As shown, the process involves five core logical objects: active feature dimensions, mask generator, initial topology graph, pruning executor, and task-sensitive topology. The active feature dimensions pass an active dimension index list to the mask generator. After responding, the mask generator initializes a binary mask vector and performs a traversal judgment on each dimension. If it is identified as an active dimension, the corresponding mask position is set to an active state; if it is identified as a silent dimension, it is set to a masked state. The mask generator outputs a complete dynamic mask to the pruning executor. While receiving the dynamic mask, the pruning executor loads the initial topology structure, performs logical AND operations to cut the topology edges connected to silent nodes and retain the connection paths of active nodes. Finally, it outputs the pruned task-sensitive topology structure and sends a topology construction completion feedback signal to the active feature dimension end, thus completing the reconstruction closed loop from high-dimensional full topology to task-sensitive sparse topology.

[0045] Example 4: This example provides a targeted and transparent explanation of the internal logic and parameter setting procedures of two core functional modules: the activity evaluation factor calculation in the feature perception module and the reinforcement learning threshold adjustment in the feedback scheduling unit. Regarding the calculation process of the activity evaluation factor in the feature perception module, the system adopts a discrete difference algorithm based on a time sliding window. The specific operation process is as follows: The system establishes a first-in-first-out (FIFO) queue of length L for each feature dimension to cache the feature values ​​of the most recent L sampling periods. Let L=16. When a new data frame arrives, the queue automatically performs a shift update, and the first-order gradient at the current time t is calculated using the difference operator. and second gradient The system is based on the formula The instantaneous activity evaluation factor is calculated, where the weighting coefficients α and β must be set according to the principle of maximizing the signal-to-noise ratio. The specific parameter calibration procedure is as follows: During the system initialization phase, a test sequence containing known effective signals and background noise is input; α is iterated from 0.1 to 0.9 (step size 0.1), and the instantaneous activity evaluation factor is calculated. The ratio of the average S-value of the effective signal segment to the average S-value of the pure noise segment (i.e., signal-to-noise ratio gain) is calculated for each set of parameters. Finally, the α and β corresponding to the maximum ratio are selected as the system operating parameters. For example, in the calibration of a certain type of vibration sensor, the preferred parameters are determined to be α=0.6 and β=0.4.

[0046] Regarding the implementation mechanism of reinforcement learning threshold adjustment in the feedback scheduling unit, the system constructs a lightweight reinforcement learning model based on Q-learning, and its lifecycle operation mechanism is as follows: The system defines the current state as a two-dimensional vector State. t =(D t E t ), where D t To normalize the computation delay, E t To normalize the mapping residual, at the end of each cycle, the feedback scheduling unit reads these two values ​​from the transform processing unit register, and the action space is discretized into three types of operations: {threshold adjustment} Remain unchanged, threshold lowered }, where step size The initial value of the preset fluctuation threshold is set to 5%. The model selects the action using an ϵ-greedy strategy based on the current Q-Table. t Action t This represents the specific decision instruction selected by the reinforcement learning operator from the action space in the current computation cycle (i.e., time t). In the early stages of training, an exploration rate ϵ = 0.2 is set to encourage exploration; as the number of iterations increases, The reward value decreases linearly to 0.01. After the action is executed, the system calculates the reward value for the next cycle. In this embodiment, it is set To reflect a higher preference for accuracy, the Bellman equation is used. Update the Q table, where the learning rate η is set to 0.1 and the discount factor γ is set to 0.9.

[0047] Example 5: To ensure that the system of the present invention can be stably deployed and achieve the expected adaptive optimization effect in complex and ever-changing industrial environments, this example elaborates on the offline calibration and data filling procedures of the core reinforcement learning model in the system, as well as the pre-calibration and debugging procedures during on-site deployment. This eliminates the black box of parameter setting and environmental adaptability uncertainty that may exist in the technical solution, and ensures the engineering reproducibility of the system. For the offline calibration and data filling of the initial parameters of the reinforcement learning model, it is necessary to build a high-fidelity simulation environment covering typical working conditions or use the historical full dataset. The system establishes an offline training database containing no less than 1,000 hours of labeled historical operating data, covering normal equipment operation, various typical faults such as bearing wear, gear tooth breakage, etc., and data samples at different environmental noise levels such as signal-to-noise ratios from 5dB to 30dB. The initial Q-table or neural network weights of the reinforcement learning model are set to random small values.

[0048] During the offline training phase, historical data is input into the feature perception module and the topology reconstruction engine in time series order to simulate the generation of dynamic masks and topology structures. The feedback scheduling unit uses the time delay D and residual E calculated by the simulation, and applies the reward function. The reward value is calculated, and the Q-table is updated according to the Q-learning algorithm. This process needs to be repeated iteratively until the model converges, meaning that the average reward value fluctuates by less than 5% within 100 consecutive training cycles. The final converged Q-table or network weights will be solidified as the initial knowledge base for system deployment, thus avoiding drastic performance fluctuations caused by exploration strategies during the cold start phase. Furthermore, for pre-deployment calibration and debugging of the system in specific industrial environments, standardized environmental adaptability procedures must be performed. After power-on, the system enters a silent observation mode. In this mode, edge pruning is disabled, and the system performs full feature calculation for at least 24 hours. During this period, the feature perception module statistically analyzes the background noise baseline level, the static variance distribution of each feature dimension, and the average throughput rate of the data stream in the field environment. Based on these statistical data, the system automatically calculates and updates the preset fluctuation threshold. The initial reference value is set to 1.2 to 1.5 times the background noise level, and the system uses the quiescent period data to calibrate the normalization factor of the mapping residual E.

[0049] Example 6: This example describes the standardized calibration and model building process before deploying a high-dimensional heterogeneous data processing system. It covers three key stages: optimization of differential operator parameters in the feature perception module, setting of hysteresis threshold in the topology reconstruction engine, and offline pre-training of the reinforcement learning model. In the optimization stage of differential operator parameters in the feature perception module, the system needs to collect at least one hour of typical data streams of the target scene, including background noise under normal operating conditions and test signals with known amplitude and frequency injected by the user. The system uses Discrete Fourier Transform (DFT) to analyze the power spectral density of the background noise and determine the main frequency band distribution of the noise. The system executes a grid search algorithm to traverse the combination of first-order gradient weight α and second-order gradient weight β (α+β=1) within a preset range. For each parameter combination, the system calculates the signal-to-noise ratio (SNR) of the test signal segment and determines the parameter combination corresponding to the maximum SNR as the optimal differential operator weight in that scene.

[0050] In the hysteresis threshold setting stage of the topology reconstruction engine, the system continuously monitors background noise data in a silent state based on the calibrated feature perception module, records the statistical distribution characteristics of the activity evaluation factor S, and calculates the mean μ of the S value. S and standard deviation σ S And will enable the threshold δ on Set as (where k is the confidence coefficient, typically ranging from 3 to 5), to ensure the false alarm rate is below a preset level, while releasing the threshold δ. off Set as δ on 80% to 90% of the time is used to construct a hysteresis range with sufficient noise tolerance, and a self-sustaining period T. hold The setting is based on the shortest duration of the target signal, typically ranging from 50% to 80% of that shortest duration. For offline pre-training of the reinforcement learning model, the system utilizes a historical fault database to construct a virtual training environment, simulating the generation of high-dimensional feature vector sequences containing various fault modes. The system initializes the Q-table or neural network weights and executes a reinforcement learning training loop in the virtual environment. In each training round, the system selects an action based on the current state, calculates the reward value, and updates the model parameters. The training process continues until the model's average reward value stabilizes over multiple consecutive evaluation cycles, and the response latency to typical faults meets real-time requirements. After training, the model parameters are solidified and loaded into the feedback scheduling unit of the field system as the initial strategy for online operation.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A system based on an AI multi-dimensional model, characterized in that, include: The feature perception module is used to receive high-dimensional heterogeneous data streams, extract feature vectors of each heterogeneous component in the high-dimensional heterogeneous data streams, calculate the fusion value of the first-order gradient and second-order gradient of the feature vectors within a preset sliding window, and determine the dimension whose fusion value exceeds a preset fluctuation threshold as the active feature dimension. The topology reconstruction engine is connected to the feature perception module to generate dynamic masks based on the active feature dimensions. The dynamic masks are then used to perform edge pruning on the initial topology graph of the high-dimensional heterogeneous data stream to build a task-sensitive topology structure. The feedback scheduling unit is connected to the topology reconstruction engine to obtain the mapping residual of the task-sensitive topology and the computational delay in the electronic digital processing. The mapping residual and computational delay are input into the preset reward function in the reinforcement learning operator. The reward function guides the reinforcement learning operator to adjust the decision threshold of edge pruning, thereby updating the coverage of the dynamic mask and keeping the mask bits in the active state in the dynamic mask aligned with the high-entropy feature axis in the high-dimensional heterogeneous data stream in the time domain. The transformation processing unit is used to perform weighted mapping operations only on the active feature dimensions selected by the dynamic mask based on the updated dynamic mask, shield the zero-value operation logic outside the coverage of the dynamic mask, output the electrical digital processing result, and generate intelligent decision instructions for high-dimensional heterogeneous data streams based on the processing result.

2. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The feature perception module calculates the activity evaluation factors for each dimension using the following formula: Where S is the activity evaluation factor, The first gradient of the eigenvector. α is the second gradient of the feature vector, and β is the preset weight coefficient; the feature perception module determines the dimension whose activity evaluation factor exceeds the preset fluctuation threshold as the active feature dimension.

3. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The feedback scheduling unit is also used to configure an activation threshold and a release threshold for the edge pruning decision threshold based on the historical update frequency of the feature vector, wherein the activation threshold is greater than the release threshold; when the fusion value of the first-order gradient and the second-order gradient exceeds the activation threshold, the corresponding mask bit is activated to reconstruct the topological edge; when the fusion value is between the activation threshold and the release threshold, the mask bit is kept in an active state; when the fusion value is continuously lower than the release threshold for a period of time exceeding the preset self-sustaining period, the mask bit is revoked.

4. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The transformation processing unit includes an index scheduling subunit, which is used to calculate the address step size between each active bit in the dynamic mask and generate a run-length index table; according to the run-length index table, it controls the memory addressing pointer to skip the masked feature dimensions in the dynamic mask and executes the loading of feature vectors and electrical-to-digital transformation in a non-contiguous memory access mode.

5. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The feature perception module is also used to extract the projection features of the masked components using the complement of the dynamic mask; the topology reconstruction engine is also used to generate a benchmark fingerprint characterizing the stability of the feature space based on the projection features. The feedback scheduling unit is also used to trigger a global mask recalibration command when the fluctuation amplitude of the reference fingerprint exceeds a preset stability threshold, so as to correct the evolution deviation of the computational model caused by the source reference drift of the high-dimensional heterogeneous data stream.

6. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The feature perception module is also used to inject an excitation signal with a preset encoding frequency into the high-dimensional heterogeneous data stream; the topology reconstruction engine is also used to generate a shadow mask based on the response features of the excitation signal and calculate the logical divergence between the shadow mask and the current dynamic mask; the feedback scheduling unit is also used to update the edge pruning judgment threshold in advance when the logical divergence exceeds the preset deviation threshold.

7. The system based on an AI multi-dimensional model according to claim 1, characterized in that, The transformation processing unit is also used to count the frequency of state flips of each mask bit in the dynamic mask within a preset period and generate a flip entropy index; the feedback scheduling unit is also used to evaluate the reputation score of each heterogeneous component based on the flip entropy index, and when the reputation score is lower than the preset reputation threshold, to perform weight downgrading processing on the feature vector of the corresponding dimension.

8. The system based on an AI multi-dimensional model according to claim 1, characterized in that, It also includes a phase-locking unit, which records the mask bit flip pulses corresponding to different heterogeneous data sources to generate the logic sequence of each data channel; and calculates the cross-correlation peak value between each logic sequence to determine the relative time delay offset between heterogeneous channels.

9. The system based on an AI multi-dimensional model according to claim 1, characterized in that, In the reward function set by the feedback scheduling unit, the positive reward is provided based on the reduction in computation latency, and the negative reward is provided based on the increase in mapping residual. The transformation processing unit adopts a distributed operator array, performs multiplication and accumulation operations on the active feature dimension according to the enable signal of the dynamic mask, and shields the zero-value memory access logic outside the coverage of the dynamic mask.

10. An intelligent decision-making method based on an AI multi-dimensional model, characterized in that, Includes the following steps: Step 101: Receive high-dimensional heterogeneous data stream, extract feature vectors of each heterogeneous component in the high-dimensional heterogeneous data stream, calculate the fusion value of the first-order gradient and second-order gradient of the feature vector within a preset sliding window, and determine the dimension with the fusion value exceeding the preset fluctuation threshold as the active feature dimension. Step 102: Generate a dynamic mask based on the active feature dimension, and use the dynamic mask to perform edge pruning on the initial topology graph of the high-dimensional heterogeneous data stream to construct a task-sensitive topology structure. Step 103: Obtain the mapping residual of the task-sensitive topology and the computation delay in the electronic digital processing. Input the mapping residual and computation delay into the preset reward function in the reinforcement learning operator. Use the reward function to guide the reinforcement learning operator to adjust the decision threshold of edge pruning, thereby updating the coverage of the dynamic mask, so that the mask bits in the active state in the dynamic mask are aligned with the high-entropy feature axis in the high-dimensional heterogeneous data stream in the time domain. Step 104: Based on the updated dynamic mask, perform weighted mapping operation only on the active feature dimensions selected by the dynamic mask, shield the zero-value operation logic outside the coverage of the dynamic mask, output the electrical digital processing result, and generate intelligent decision instructions for high-dimensional heterogeneous data streams based on the processing result.

Images