A multi-channel sensor synchronous sensing and closed-loop control coprocessing circuitry

Abstract

The application discloses a kind of multi-channel sensor synchronous sensing and closed-loop control coprocessor circuit system, belong to digital integrated circuit real-time control field.For the defects of poor sampling timing consistency, weak signal analysis ability, regulation and control lag of existing architecture, the application adopts full hardwired closed-loop pipeline, integrates synchronous sampling, multi-dimensional embedding, feature solving, field selection, waveform synthesis and closed-loop injection main path, and three auxiliary paths of hodge decomposition anti-jitter, state entropy monitoring and collaborative adaptation.Channel synchronization deviation is less than 5 nanoseconds, hodge decomposition output anti-jitter criterion reduces field type misoperation by more than 70%, state monitoring rollback response is less than 50 nanoseconds, and collaborative adaptation compresses regulation and control delay to within 200 nanoseconds.Full hardware solidification, closed-loop response is less than 500 nanoseconds, can be built-in or externally connected to adapt commercial chips, and is suitable for MEMS field control, silicon light regulation and AI acceleration scene.

Description

Technical Field

[0001] This invention relates to the field of sensor signal processing and hardware closed-loop control technology, specifically to a coprocessing circuit system adapted to physical field array sensing and real-time closed-loop control. It is particularly suitable for real-time control scenarios that require multi-channel synchronous sampling, nanosecond-level hardware response and high-precision adaptive adjustment, such as MEMS (Micro-Electro-Mechanical Systems) mechanical field control arrays, silicon photonic array control and AI computing acceleration. It can also be extended to engineering adaptation fields such as chip thermal management and dynamic precision control. Background Technology

[0002] In a numerical computing pipeline architecture composed of multiple ALU arrays, the coprocessor acts as the interaction hub between the external physical system and the chip's computing core. It collects data from multiple sensors through multi-channel sensing interfaces, completes multi-dimensional sensing signal processing, feature quantity calculation, and control waveform synthesis, and feeds back the optimized control signal to the main processor's computing port. This constructs a real-time control link of "physical sensing - signal processing - control decision - closed-loop feedback," making it a key peripheral and collaborative computing unit for high-end tensor computing, intelligent domain control, and spatial real-time computing chips.

[0003] Existing chip coprocessing architectures for tensor computing, AI acceleration, and domain control still have several key technical shortcomings in multi-channel synchronous sensing and high-speed closed-loop control scenarios, failing to meet the requirements of high-precision, low-latency, and high-dynamic real-time control. Specific problems are as follows:

[0004] I. Poor consistency in synchronous sampling of multi-channel sensors leads to severe distortion in multi-dimensional fusion. Traditional chip multi-channel sampling generally adopts a channel-by-channel polling scanning architecture, resulting in significant timing deviations between sampling times of each channel. These deviations can reach 10 to 100 microseconds, far exceeding the effective response time window of real-time control. Timing mismatch directly leads to misalignment of multi-dimensional sensor coordinates and distortion in the fusion of multi-source physical states, causing deviations in subsequent tensor feature solving and failure of control field matching, significantly reducing the accuracy of closed-loop control and system stability.

[0005] II. The lack of dedicated hardware acceleration solutions for high-order geometric signal decomposition algorithms. Hodge decomposition is a core mathematical tool for precise decoupling of multidimensional differential forms and separation of physical signal features. It can uniquely decompose any multidimensional signal into harmonic components, appropriate components, and co-appropriate components, providing crucial theoretical support for high-precision physical field calculations, jitter assessment, and state stability identification. However, this algorithm involves high-dimensional matrix transformations, Hodge star dual mappings, exterior and co-differential operations, and Laplace iterative convergence solutions, resulting in extremely high computational complexity. Existing technologies rely solely on iterative solutions using CPU or MCU software, leading to millisecond-level computational latency, which is unsuitable for high-speed real-time control scenarios ranging from microseconds to nanoseconds. Currently, there is no mature dedicated hardware pipeline architecture in the industry to achieve real-time hardware acceleration for Hodge decomposition.

[0006] Third, the switching of control field patterns relies on static software configuration and lacks hardware adaptive closed-loop control capabilities. The frequency, phase, amplitude, and field pattern of the existing physical field control waveforms all adopt a static control method of software pre-configuration and timed refresh. This method cannot capture dynamic changes and operating condition drift of the external physical system in real time, and cannot adaptively switch to the optimal control field pattern. This results in engineering problems such as control parameter lag, untimely adaptation to operating conditions, system oscillation, and frequent erroneous switching of field patterns, which seriously restrict the quality of closed-loop control in high dynamic scenarios.

[0007] Fourth, the sensing end and the high-precision tensor computation pipeline lack a hardware-level collaborative mechanism. Traditional tensor computation chips and AI accelerators only passively adjust their working precision and iteration step size based on internal computation data. They cannot perceive the dynamic changes of external physical quantities in advance, and there is a significant lag in precision switching and iteration parameter adjustment. It is difficult to achieve deep real-time collaboration between the external physical state and the chip's internal numerical computation pipeline, which limits the convergence stability and dynamic adaptation capability in high-speed iteration scenarios.

[0008] In summary, existing coprocessing architectures have significant technical deficiencies in four aspects: synchronous sampling, algorithm acceleration, adaptive control, and collaborative linkage. There is an urgent need for a fully hardware coprocessing circuit system that can achieve multi-channel synchronous sensing, hardware-based high-order signal decomposition, adaptive field pattern modulation, and front-end collaborative linkage. Summary of the Invention

[0009] Inventive concept

[0010] The core concept of this invention is to build a fully hardware-connected pipeline link that completely covers the physical field control closed-loop process of multi-channel synchronous acquisition, multi-dimensional coordinate embedding, tensor feature solving, adaptive field matching, control waveform generation and drive signal output. The entire process is executed by pure combinational logic hardware and does not rely on software time-sharing scheduling, bus arbitration and microcode decoding delay.

[0011] This system is equipped with two independent parallel hardware-assisted processing paths. The first path is used for real-time generation of field decomposition criteria signals to achieve field stability and suppress oscillations. The second path is used for real-time calculation of system state entropy values ​​to complete the identification of operational mutations and the rollback adjustment of the iterative process. The two auxiliary paths run in parallel with the main field control pipeline, outputting correction signals synchronously at the nanosecond level. At the same time, a native hardware collaborative adaptation interface is added to realize physical layer pre-linkage prediction between the external physical sensing array and the tensor precision calculation pipeline, so as to complete the pre-adjustment of field control parameters in advance.

[0012] The entire field control coprocessing architecture replaces the traditional "CPU + time-sharing coprocessing" architecture with a pure hardware closed loop, solving a series of technical defects in existing field control systems from the underlying hardware path level, such as high processing latency, large multi-channel synchronization deviation, static and single field control mode, easy oscillation during operation, and lack of state prediction and correction capabilities.

[0013] Technical problems to be solved

[0014] This invention aims to address four inherent defects in existing tensor coprocessing hardware architectures:

[0015] (1) Traditional coprocessing links lack a complete hardware-based perception-computation-control closed loop, and high-order Hodge signal decomposition lacks a dedicated hardware acceleration pipeline, resulting in excessively high overall computational latency.

[0016] (2) The multi-channel sensor adopts a polling sampling method, and the channel timing is inconsistent, resulting in serious distortion in the fusion of multi-dimensional physical sensing data;

[0017] (3) The physical field control field pattern relies entirely on static software configuration and cannot be dynamically switched adaptively according to real-time physical conditions;

[0018] (4) There is no hardware-level prediction linkage between the sensor acquisition module and the tensor operation pipeline, resulting in a lag in the dynamic response of the system and poor stability of the closed-loop control.

[0019] Overall technical solution

[0020] To address the aforementioned technical problems, this invention provides a multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system. It employs a purely hard-wired, full-domain pipeline architecture, comprising: a main hardware pipeline, two parallel hardware auxiliary processing units, and a hardware co-adaptive front-end interface. The entire system eliminates software scheduling, bus arbitration, and microcode latency; all functions are completed in real-time through hardware combinational logic and timing pipelines, achieving a fully hardware nanosecond-level closed loop from physical sensing to geometric tensor calculation, adaptive field shape control, and finally, closed-loop control injection.

[0021] Main hardware pipeline

[0022] The main hardware pipeline is a multi-stage serial hardwired straight-through architecture, which includes the following components sequentially along the forward data flow: a multi-channel synchronous sampling interface, a multi-dimensional embedding pipeline, a signal feature calculation unit, a field pattern selection stage, a waveform synthesis stage, and a control signal injection stage. A bidirectional feedback path is established between the multi-channel synchronous sampling interface, the multi-dimensional embedding pipeline, the signal feature calculation unit, the field pattern selection stage, the waveform synthesis stage, and the control signal injection stage, forming a complete iterative closed loop. Only the path from the signal feature calculation unit to the double buffer, from the double buffer to the signal decomposition auxiliary criterion generation hardware unit, and from the signal decomposition auxiliary criterion generation hardware unit to the field pattern selection stage are unidirectional auxiliary branches, which only transmit anti-jitter criterion signals and do not participate in the iterative parameter feedback.

[0023] (1) Multi-channel synchronous sampling interface

[0024] It contains N parallel ADC channels, where N≥3. All channels are triggered for sampling on the same clock edge and are uniformly marked with timestamps. The built-in sample-and-hold array ensures that each channel synchronously acquires the physical quantity at the same moment, and the synchronization deviation between channels is less than 1 ADC clock cycle. The built-in multi-domain phase-locked loop realizes independent clock alignment for multiple sensing domains. After sampling, the built-in hard-wired filtering unit completes real-time filtering and noise reduction of the raw data.

[0025] (2) Multidimensional embedded pipeline

[0026] The multi-channel synchronous sampling interface is composed of N×M parallel multipliers and M multi-input accumulator trees. It weights and fuses N channels of sensor data into an M-dimensional coordinate vector according to a preset mapping relationship, where M≥3. The dimension mapping mode can be dynamically switched through the dimension mode configuration register, and the mode switching delay is less than 1 clock cycle.

[0027] (3) Signal characteristic calculation unit

[0028] Connecting a multidimensional embedding pipeline, including a metric matrix solver and a comprehensive feature scalar calculator, by... Real-time solution of information manifold metric tensors using 3D coordinate vectors With comprehensive feature scalar The overall computational latency is less than 50ns; feature scalar Real-time output to the field type selection level, metric tensor Write to a double-buffered cache for parallel, non-blocking reading by the auxiliary criterion generation unit.

[0029] (4) Field type selection level

[0030] The signal feature calculation unit is connected to a pure combinational logic decision circuit consisting of a multi-level 2-to-1 comparator tree and a field encoder. It adaptively selects three field types based on configurable thresholds, including the resonant field type corresponding to the high feature range, the decoherent field type corresponding to the low feature range, and the coherent field type corresponding to the middle range, and outputs the corresponding frequency, phase, and amplitude control parameters. The field type switching delay is less than 5ns.

[0031] (5) Waveform synthesis stage

[0032] The connection field selection stage uses a phase accumulator and waveform lookup table to construct a direct digital frequency synthesizer, which continuously generates high-precision sine waves based on a real-time clock, with a synthesis accuracy of 16 bits and a signal output update rate of no less than 10 MSPS.

[0033] (6) Control signal injection stage

[0034] The connected waveform synthesis stage includes a gradient calculator and an injection buffer, which solves the control gradient signal in real time and injects it into the external processor's computing port through a hard-wired direct connection, achieving zero bus delay closed-loop feedback.

[0035] Two-way parallel hardware auxiliary unit

[0036] (1) Signal decomposition auxiliary criterion generation hardware unit

[0037] It is connected in parallel with the double-buffered buffer and field selection stage of the signal feature calculation unit, and adopts a multi-level hard-wired serial pipeline architecture, which integrates the Hodge star operator operation subunit, the Hodge-Laplace operator operation subunit, and the harmonic projector operation subunit in sequence. It can complete the full Hodge decomposition in multidimensional differential form within a fixed clock cycle and output three types of results: harmonic component, appropriate component, and co-appropriate component. The hardware uses the proportion value of the harmonic component as the auxiliary judgment signal for field switching, filters instantaneous noise interference, and avoids frequent false switching of field modes and system oscillation.

[0038] (2) System state entropy real-time monitoring unit

[0039] The system synchronously accesses a multi-dimensional embedded pipeline to acquire real-time tensor data, while simultaneously injecting control signals into the control signal injection stage and outputting regulation signals. The unit is composed of a multi-dimensional numerical integrator, an entropy history first-in-first-out (FIFO) cache, a first-order differential operation circuit, and two-stage threshold comparators cascaded in sequence. It solves the global system entropy in real time through hardware circuit integration, caches multiple rounds of historical entropy data, calculates the entropy change gradient in hardware, compares it with preset two-stage thresholds to identify system mutations and divergence conditions, and outputs multi-stage hardware interrupt signals and iterative rollback control instructions based on the judgment results.

[0040] Hardware Co-adaptor Interface

[0041] The system is configured with a hardware co-adaptive interface, which is hardwired to connect to the precision control unit of the high-precision adaptive tensor operation pipeline. The front-end synchronously samples the original sensing data to realize the advance prediction of physical conditions and the pre-switching of calculation precision, eliminate the control delay caused by pipeline lag, and complete the hardware-level advance linkage of sensing, calculation and control.

[0042] The system is a standardized, independent coprocessor hardware module. It can be integrated into various computing chips as an IP core, or connected to existing mature commercial chips as an independent peripheral coprocessor circuit. It can interface with multi-point temperature sensor arrays of chips to achieve full-domain thermal management of the chips. At the same time, it can collect load and temperature disturbance signals in the front end to achieve dynamic prediction and control of the chip's computational quantization accuracy. There is no need to modify the existing main chip core architecture. Adaptation and linkage can be completed only through standardized hardware direct connection interfaces.

[0043] Beneficial effects

[0044] Compared with existing software iteration schemes and traditional static sensing coprocessing architectures, this invention has several significant and corresponding technological innovations and engineering advantages, each of which addresses a specific deficiency in the underlying technology:

[0045] (1) Achieve multi-channel nanosecond-level synchronous sampling and completely eliminate multi-dimensional data fusion distortion. In view of the technical defects of traditional polling sampling channels with large timing deviation (10~100μs) and multi-dimensional sensor data fusion distortion pointed out in the background technology, the present invention adopts global clock edge synchronous triggering, sample-and-hold array and multi-domain phase-locked loop alignment technology, so that the timing deviation of all sensor channels is strictly controlled within a single ADC clock cycle (less than 5 nanoseconds in this preferred embodiment). Compared with the traditional architecture, it achieves a 3 to 4 order of magnitude improvement in accuracy, completely solves the timing misalignment and phase mismatch problem, eliminates multi-dimensional physical field fusion distortion and tensor solution deviation, and provides a highly consistent original data foundation for back-end feature solution and control field matching.

[0046] (2) Achieving full hardware pipeline acceleration for high-order Hodge geometric decomposition, filling a hardware gap in the industry. Addressing the technical shortcomings of Hodge decomposition, as pointed out in the background technology—relying solely on CPU and MCU software iterative solutions, resulting in millisecond-level computational delays and inability to adapt to high-speed real-time control scenarios—this invention, for the first time, fully pipelines high-order algorithms in differential geometry, such as Hodge star mapping, Hodge-Laplace smoothing, and harmonic projection three-component decomposition, into a hardware system. It incorporates a dedicated matrix lookup table array, parallel finite difference operation circuit, and hardware convergence determination module, enabling precise and unique decomposition of multidimensional differential signals within a fixed 32 clock cycles. Compared to traditional software iterative solutions, the computational speed is improved by 3-4 orders of magnitude, completely solving the technical shortcomings of large time delays and inability to perform real-time computation in high-order geometric signal decomposition, and filling the technical gap in dedicated hardware acceleration circuits for Hodge decomposition.

[0047] (3) A unique hardware anti-jitter criterion mechanism based on the proportion of harmonic components significantly improves the smoothness and stability of field control. Addressing the technical shortcomings mentioned in the background technology, such as the reliance on static software configuration for control field patterns, the inability to adapt to dynamic switching, and susceptibility to noise interference leading to frequent false switching and system oscillations, this invention utilizes the inherent properties of the harmonic components, appropriate components, and co-appropriate components output by Hodge decomposition. The hardware calculates the proportion of harmonic components in real time as an auxiliary criterion for field pattern switching. Through a dual-condition hardware locking circuit built into the field pattern selection stage, it accurately distinguishes between real physical condition drift and instantaneous random noise, allowing field pattern switching only when both conditions are simultaneously met: "the main path characteristic quantity crosses the threshold" and "the proportion of harmonic components is lower than the stability threshold." Field pattern switching is executed by pure combinational logic with a delay of less than 5 nanoseconds. Under normal engineering signal-to-noise ratio (SNR≥15dB) conditions, the number of false field pattern switching is reduced by more than 70%, and the standard deviation of field pattern consistency time is compressed from milliseconds to microseconds, significantly improving the smoothness and steady-state accuracy of high-dynamic physical field control.

[0048] (4) Realize nanosecond-level hardware real-time monitoring of system state entropy to ensure the stability of high-speed iterative operation. Addressing the technical deficiencies mentioned in the background technology, such as the lack of real-time monitoring of system operating status and the inability to respond promptly to abnormal conditions (corresponding to monitoring and rollback requirements), this invention utilizes a cascaded architecture of a hardware discrete integration circuit, an entropy history cache array, a first-order gradient difference operation circuit, and a two-stage threshold hardware comparator to solve the system's global entropy value and its changing trend in pure hardware in real time. This accurately identifies three operating conditions: system convergence, slow degradation, and sudden divergence, and outputs graded interrupt and iterative rollback instructions at the nanosecond level. The first-level early warning response assists the system in adaptively adjusting iteration parameters, and the second-level rollback response completes the rapid recovery of the previous cycle's stable state within 50 nanoseconds. Compared to traditional software polling monitoring schemes, the response speed is improved by 2-3 orders of magnitude, effectively suppressing numerical oscillations and divergence problems during high-speed tensor iteration from a hardware perspective.

[0049] (5) Achieving hardware-level advanced collaborative adaptation between sensing and tensor computation, eliminating pipeline lag bias. Addressing the technical shortcomings mentioned in the background art, such as the lack of hardware-level predictive linkage between the sensing acquisition module and the tensor computation pipeline, resulting in lag in system dynamic response, this invention sets up a dedicated hardware collaborative adaptation interface. External operating condition parameters such as temperature, load, and physical field strength are acquired before data latching is performed at the same clock edge of the multi-channel synchronous sampling interface. Through operating condition feature extraction and hardware differential trend discrimination circuitry, the trend of physical quantity changes is predicted in advance, completing the pre-adaptation adjustment of computational accuracy and iteration step size before the operating condition change data enters the main computation pipeline. Under actual temperature change conditions (change rate ≥ 50℃ / s), the control delay is reduced from 1.5 microseconds in the traditional architecture to less than 200 nanoseconds, a reduction of over 86%, completely eliminating the lag bias caused by the passive control of traditional chips, and achieving deep real-time collaboration between external physical states and the internal numerical computation pipeline.

[0050] (6) Constructing a fully hardwired, complete perception-computation-control closed loop to achieve ultra-low latency deterministic response. Addressing the technical shortcomings of traditional coprocessing links, such as the lack of a complete hardware closed loop and high latency in multi-level scheduling, as pointed out in the background technology, this invention constructs a fully hardware-based, direct-through, bidirectional closed-loop pipeline from multi-channel synchronous sampling, multi-dimensional coordinate embedding, tensor feature calculation, adaptive field selection, high-precision waveform synthesis to gradient signal injection. All data acquisition, dimension fusion, tensor calculation, field decision, status monitoring, waveform generation, and closed-loop injection functions are implemented by pure hardware combinational logic and pipeline timing circuits, with no operating system scheduling, no bus arbitration blocking, and no instruction decoding delay throughout the process. The overall closed-loop latency of the system is less than 500 nanoseconds, possessing completely deterministic ultra-low latency characteristics, which can meet the stringent low-latency requirements of high-precision physical field control, high-speed tensor iteration, and real-time spatial computation.

[0051] Terminology Definition

[0052] In this invention, the relevant terms are defined as follows:

[0053] (1) Multi-channel synchronous sampling interface: refers to a hardware synchronous sampling circuit containing at least 3 parallel ADC channels. All channels can be synchronously triggered for sampling on the same clock edge and uniformly marked with timestamps. It has the ability to sample and hold and align with multi-domain clocks, and can realize high-precision synchronous acquisition of multi-physical sensor signals.

[0054] (2) Multidimensional embedding pipeline: refers to a hardware operation pipeline composed of a parallel multiplier array and a multi-input accumulator tree, which is used to weight and fuse multiple discrete sensor data to generate high-dimensional geometric coordinate vectors. It supports dynamic switching of dimension mapping mode at runtime and realizes hardware embedding transformation from physical space to manifold coordinates.

[0055] (3) Signal decomposition auxiliary criterion generation hardware unit: refers to a dedicated Hodge decomposition hardware pipeline circuit, which is configured with Hodge star operation unit, Hodge-Laplace operation unit and harmonic projection operation unit in sequence. It can complete the global decomposition of multidimensional differential form signals and stably output harmonic components, appropriate components and co-appropriate components. The component ratio is used as the hardware anti-shake criterion for field pattern adaptive switching.

[0056] (4) Hardwired: refers to the physical metal direct connection inside the chip, without bus arbitration, microcode decoding, or software scheduling. The core operation and logic decision are all implemented by pure hardware combinational logic. Only registers and caches are configurable storage units, and it has zero scheduling delay and deterministic transmission characteristics.

[0057] (5) Dual-buffered cache: refers to a non-blocking parallel cache structure composed of two independent memory blocks, which respectively undertake the real-time data writing of the pipeline and the parallel data reading of the auxiliary unit. The read and write paths are independently controlled, with no data conflicts and no handshake waiting, ensuring high-speed parallel interaction of tensor data. (6) Characteristic scalar R: a quantitative judgment value used to characterize the real-time state of the physical field and distinguish different field type intervals. It is obtained in real time based on the manifold metric tensor and serves as the basic criterion for field type switching.

[0058] (7) Harmonic component ratio ρ: The ratio of the harmonic component norm to the overall signal norm, with a value range of 0 to 1, is used to distinguish between real-world drift and instantaneous random noise, and serves as an auxiliary hardware criterion for field-type anti-shake.

[0059] (8) Metric tensor Also known as the manifold metric tensor, it is a second-order symmetric covariant tensor and a core mathematical quantity for describing the local geometric structure of a manifold in differential geometry. It is generated in real time by solving multi-dimensional coordinate vectors on the hardware side and is used to characterize the local distance, angle and curvature features of the manifold. It is stored in a double buffer and is read and processed in parallel by the Hodge decomposition hardware unit.

[0060] (9) Manifold dimension n and differential form order k: Hodge decomposition hardware unit configurable parameters; n represents the total dimension of the information manifold space and k represents the order of the input differential form. Together, they determine the dual mapping transformation rule of the Hodge star operator.

[0061] (10) Differential form ω, harmonic differential form ω is the second-order differential form of the input constructed based on sensor data and the metric tensor; To satisfy the steady-state harmonic component differential form of the triple constraints of Laplace operator, exterior differential, and co-differential.

[0062] (11) External differential operator d, codifferential operator Hodge-Laplace operator : Standard operators in differential geometry; d implements the ordering of differential forms. To achieve order reduction of the differential form, the two are superimposed to form the Hodge-Laplace operator. It is used for smoothing and distortion correction of manifold signals.

[0063] (12) Field type: refers to the working mode in the closed-loop control of the physical field. In this invention, there are three types: resonant field type, decoherent field type and coherent field type. The field type selection stage locks the field type based on the dual conditions of characteristic scalar R and harmonic component ratio ρ. Each field type corresponds to a set of independent frequency, phase and amplitude control parameters.

[0064] (13) Iterative rollback: When the real-time monitoring unit of the system state entropy detects that the system values ​​are diverging or the operation is unstable, the hardware automatically performs a multi-level recovery operation: locking the control parameters of the stable field pattern of the previous cycle, re-initializing the registers of each stage of the pipeline with historical stable data, pausing the injection of new data from upstream, until the rollback completion confirmation signal is valid. The whole process is completed automatically by pure hardware logic without the need for software participation.

[0065] (14) Closed-loop iteration cycle: refers to the number of hardware clock cycles required for the main hardware pipeline to complete a complete process of "synchronous sampling → feature fusion → field pattern decision → waveform synthesis → closed-loop feedback". It is determined by the cumulative delay of each pipeline stage. In this invention, it is a fixed value and does not change with the operating conditions.

[0066] (15) Higher-order geometric signal decomposition (Hodge decomposition): refers to the mathematical operation of uniquely decomposing a multidimensional differential signal into three categories: harmonic components, appropriate components and co-appropriate components according to the Hodge decomposition theorem. In this invention, the hardware unit that generates the signal decomposition auxiliary criterion completes the full hardware pipeline solution within a fixed 32 clock cycles. Attached Figure Description

[0067] Figure 1This is a schematic diagram of the overall architecture of the multi-channel synchronous sensing and closed-loop control co-processing circuit system of the present invention. The diagram includes a series-connected main bidirectional closed-loop pipeline, two parallel auxiliary processing units, and a hardware co-adaptation interface. The main pipeline, along the forward data flow, consists of a multi-channel synchronous sampling interface, a multi-dimensional embedding pipeline, a signal feature calculation unit, a field selection stage, a waveform synthesis stage, and a control signal injection stage, connected in series via hardwired connections. Bidirectional feedback paths are established between adjacent modules, forming a complete iterative closed loop. The output of the signal feature calculation unit is unidirectionally coupled to the signal decomposition auxiliary criterion generation hardware unit via a double buffer. The output of this auxiliary unit is unidirectionally coupled to the field selection stage. The output of the multi-dimensional embedding pipeline is synchronously and unidirectionally connected to the system state entropy real-time monitoring unit. The output of this monitoring unit is unidirectionally coupled to both the waveform synthesis stage and the control signal injection stage. The input of the hardware co-adaptation interface is unidirectionally connected to the output of the multi-channel synchronous sampling interface. The output of this interface is unidirectionally coupled to both the waveform synthesis stage and the control signal injection stage. Arrows indicate the direction of signal transmission between modules. Bidirectional arrows represent bidirectional data flow, while unidirectional arrows represent unidirectional data flow, criterion signals, or interrupt control signals.

[0068] Figure 2 This is a schematic diagram of the internal architecture of the signal decomposition auxiliary criterion generation hardware unit of the present invention. This unit consists of a multi-stage computational pipeline formed by sequentially connecting a Hodge star operator subunit, a Hodge-Laplace operator subunit, and a harmonic projector subunit. The diagram fully illustrates the hardware cascading relationship of each subunit, the parallel computation array, the convergence determination circuit, and the signal input / output flow. The diagram also labels the metric tensor data input port, the differential form signal input port, the parallel output ports for the harmonic component / appropriate component / co-appropriate component, and the hardware anti-jitter criterion output port.

[0069] Figure 3 This is a schematic diagram of the system state entropy real-time monitoring unit structure of the present invention. Internally, this unit consists of a multidimensional numerical integrator, an entropy history first-in-first-out (FIFO) buffer, a first-order differential operation circuit, and two-stage threshold comparators cascaded together. The diagram fully illustrates the cascaded relationships of each hardware module, the discrete integral accumulation circuit, the gradient differential path, and the multi-stage interrupt signal output path. The diagram also labels the standardized tensor feature data input port, the first-level early warning interrupt output port, the second-level forced rollback interrupt output port, and the control signal output port. Detailed Implementation

[0070] Example 1: Multi-path hardware pipeline collaborative timing mechanism for the whole machine (corresponding to) Figure 1 Architecture)

[0071] This embodiment, combined with the overall circuit architecture of the present invention, discloses the collaborative timing and parallel coordination mechanism of the multi-path hardware pipeline of the whole machine, and fully elaborates the hardware timing coordination logic of the main closed-loop pipeline, the anti-shake auxiliary branch, the status monitoring branch, and the front-end collaborative interface. Each path achieves parallel collaboration through hardware-fixed timing logic, and there are no timing conflicts, data competitions, or resource interlocking issues throughout the process, ensuring high-speed and stable iterative operation of the system.

[0072] The system adopts a pure hardware timing architecture of "synchronous sampling, parallel computing, hierarchical decision-making, and closed-loop iteration," with all functional modules driven by a unified global system clock. A complete iteration cycle is arranged in a hardware pipeline, divided into the following stages: data sampling and prediction, feature fusion and distribution, main path feature solving and double-buffered isolation, auxiliary branch parallel computing, field pattern dual-condition decision-making, waveform synthesis and closed-loop feedback, and status monitoring parallel acquisition. Each stage works in concert to achieve non-blocking, low-latency parallel data processing. The specific timing flow within a single iteration cycle is as follows.

[0073] Phase 1: Data Synchronization Sampling and Pre-judgment. The multi-channel synchronous sampling interface completes synchronous sampling of the entire physical signal domain on the same clock edge, latching all channel acquisition data. The specific hardware implementation of this interface is an eight-channel parallel ADC sampling system. All channels share the same global system clock, and the sampling switches are synchronously triggered by the same clock edge, achieving simultaneous acquisition of multiple external physical sensor signals at the same time. This eliminates the inherent timing offset problem of traditional time-division polling sampling architectures at the hardware level. Each sampling channel integrates a purely hard-wired real-time filtering unit at the back end, consisting of a shift register, a parallel accumulator array, and a smoothing operation circuit. This unit can complete the filtering operation and output a smoothed result within the same clock cycle after the sampled data is latched, effectively filtering out random noise and high-frequency interference in the original sampled data. Meanwhile, the hardware collaborative adapter interface latches external core operating condition parameters such as temperature, load, and physical field strength from the original sampled data. Through the built-in operating condition feature extraction circuit and hardware differential trend discrimination circuit, it calculates the rate and direction of change of each operating condition parameter in real time, proactively capturing and predicting dynamic changes in external operating conditions. This provides advance output of precision switching control signals and iteration step size adjustment signals for the subsequent adaptive adjustment of the main flow line. During this stage, the main flow line (synchronous sampling) and the front-end collaborative interface (operating condition prediction) work in parallel, without occupying each other's data paths and computing resources.

[0074] Phase Two: Feature Fusion and Data Distribution. The multidimensional embedding pipeline receives the multi-channel sensor data latched in Phase One, performs multi-channel data fusion and feature normalization mapping in parallel, and outputs standardized tensor feature data. This data is transmitted synchronously in two parallel paths: one path is sent to the signal feature calculation unit of the main pipeline for feature calculation and preparatory calculations for field pattern decision; the other path is connected in real time to the system state entropy real-time monitoring unit, providing real-time data support for steady-state monitoring and gradient calculation. In this phase, the main pipeline and the state monitoring branch receive data and start calculations in parallel under the same clock cycle. The two paths do not occupy each other's resources, realizing one-time acquisition and bidirectional distribution of feature data.

[0075] The third stage: Main path feature solving and double-buffered isolation stage. After receiving data in the previous stage, the signal feature calculation unit completes high-speed hardware feature solving calculations and outputs real-time operating condition decision features. This output data is simultaneously transmitted to the next stage along two paths: one path is directly sent to the subsequent field pattern selection stage as the basis for the main path decision; the other path is sent to the signal decomposition auxiliary criterion generation hardware unit after being isolated by a double buffer. The double buffer here plays a role in data isolation and smooth transition across clock domains, ensuring that the data transmission between the main stream and the anti-jitter auxiliary branch does not generate timing spikes or data races.

[0076] Phase 4: Parallel computation phase of the anti-shake auxiliary branch. After receiving the feature data isolated by double buffering, the signal decomposition auxiliary criterion generation hardware unit independently initiates the anti-shake criterion pipeline computation. This unit is a fully hardwired three-stage pipeline architecture. The first stage is a dual mapping computation unit consisting of a configurable parameter lookup table array and a parallel multiplication operation array, which maps input signals of different dimensions to a standardized data format. The second stage is a multi-dimensional differential smoothing computation unit consisting of an external differential hardware computation module, a co-differential hardware computation module, and an adder array, which completes multi-dimensional gradient solving and signal smoothing correction through hardware finite difference parallel computation. The third stage is a steady-state component projection unit consisting of a hardware residual comparison and convergence determination circuit, which uses multi-stage hardware comparators to judge the computational residuals in real time and stably separates and outputs three types of hardware signals—namely, harmonic components (steady-state components), appropriate components (instantaneous fluctuation components), and co-appropriate components (correction components). The proportion of the harmonic components (i.e., steady-state components) is calculated by a built-in hardware divider to generate hardware anti-shake criteria, which are then output to the field selection stage. This anti-shake auxiliary branch is a unidirectional parallel pipeline that reads data from the double buffer cache to perform independent calculations. It does not occupy the main pipeline's computing resources or interfere with the main iteration timing. The overall processing latency is fixed at 32 clock cycles (160 nanoseconds). It can output steady-state anti-shake criteria in advance before the main pipeline completes the field type decision, providing a pre-hardware constraint for field type adaptive switching.

[0077] Phase 5: Dual-Condition Hardware Decision Stage for Field Pattern Selection. The field pattern selection stage synchronously receives real-time characteristic quantities of the main path and anti-jitter criteria of the auxiliary branch, and completes dual-condition hardware decision through pure combinational logic to lock the current effective field pattern operating mode. This dual-condition decision mechanism allows field pattern switching only after both constraints of "main path characteristic quantity meeting the standard" and "steady-state confirmation of auxiliary criteria" are met at the hardware level, thereby eliminating invalid switching problems caused by instantaneous jitter of single parameters and ensuring the reliability and anti-interference capability of adaptive field pattern switching.

[0078] Phase 6: Waveform Synthesis and Closed-Loop Feedback Stage. The waveform synthesis stage generates a high-precision, controllable waveform at high speed based on the field control parameters locked by the field selection stage. The control signal injection stage receives the waveform data output from the waveform synthesis stage, completes the closed-loop gradient calculation, and outputs a feedback control signal through a hard-wired direct-connection path to achieve bidirectional iterative correction and closed-loop control of the main path. At this point, the main closed-loop pipeline completes a full "sampling, fusion, decision, feedback" iterative cycle.

[0079] Phase 7: Parallel Data Acquisition for Status Monitoring. While the main waterline advances step by step, the system state entropy real-time monitoring unit, after receiving standardized tensor feature data in Phase 2, continuously runs state entropy monitoring and gradient calculation in parallel. Throughout the entire iteration cycle, this unit maintains clock synchronization with the main waterline and operates in parallel, outputting real-time system steady-state evaluation results, providing hardware-level support for anomaly protection and long-term stability monitoring of the main waterline. This monitoring branch is completely decoupled from the main waterline in terms of timing; it only reads data and does not interfere with the main waterline control. Its calculation process does not cause any timing disturbance to the main waterline.

[0080] In summary, in this embodiment, the main bidirectional closed-loop pipeline undertakes the overall system iteration update and core closed-loop control functions. The anti-jitter auxiliary branch independently completes steady-state discrimination and noise filtering, the status monitoring branch performs real-time system status monitoring and anomaly protection prediction in parallel, and the front-end collaborative interface performs pre-condition prediction and accuracy pre-adaptation. These four components have clear division of labor, are parallel and complementary, timing-adaptive, and do not interfere with each other, forming a complete hardware-level synchronous perception, feature calculation, anti-jitter correction, status monitoring, adaptive field control, and closed-loop iteration system. This timing mechanism is entirely hardware-fixed, with no software scheduling delay or bus resource scheduling latency, effectively ensuring the nanosecond-level dynamic response performance and long-term iterative operation stability of the entire system.

[0081] Example 2: Hardware Field Shape Stabilization Control Mechanism Based on Steady-State Component Proportion

[0082] This embodiment corresponds to Figure 1The hardware anti-jitter decision logic within the field pattern selection stage of the overall architecture is disclosed, revealing a hardware field pattern anti-jitter control mechanism based on the proportion of steady-state components. This mechanism is independently implemented by the hardware anti-jitter locking circuit built into the field pattern selection stage. The hardware inputs are the main real-time characteristic quantities from the signal characteristic quantity calculation unit and the steady-state component proportion criteria from the signal decomposition auxiliary criterion generation hardware unit. The hardware outputs are the field pattern locking signal and the field pattern switching enable signal. In high-dynamic physical field closed-loop control scenarios, external environmental interference and instantaneous sensor disturbances can easily cause instantaneous jumps in operating condition characteristic parameters, leading to frequent erroneous switching of the field pattern selection stage, which in turn causes system control oscillations, output waveform disturbances, and a decrease in closed-loop control stability. This embodiment effectively suppresses invalid field pattern switching at the hardware level through multi-level hardware criterion linkage. The specific hardware implementation mechanism is as follows.

[0083] The hardware anti-shake locking mechanism described in this embodiment consists of three levels of cascaded hardware logic: main path feature quantity comparison judgment, steady-state component proportion judgment, and dual-condition hardware locking judgment. The entire process is a mixed circuit of pure combinational logic and sequential logic, without the participation of software scheduling.

[0084] Level 1: Main Path Feature Comparison and Decision. This level consists of multiple hardware comparators that receive the operating characteristic parameters output by the signal characteristic calculation unit in real time and compare them with the preset field switching threshold. The comparator output is a binary level signal: when the characteristic value does not cross the switching threshold, it outputs a low level (invalid trigger); when the characteristic value crosses the switching threshold, it outputs a high level (valid trigger). This output is sent to the third-level decision logic as a field switching candidate condition.

[0085] Level 2: Steady-state component proportion determination. This level receives the steady-state component proportion value from the hardware unit that generates the signal decomposition auxiliary criterion in real time. It then compares this value with a preset stability threshold using a built-in hardware comparator. The discrimination circuit outputs a binary level signal: when the steady-state component proportion is higher than the preset stability threshold, the current signal is determined to be dominated by steady-state components with a low proportion of transient disturbance components, and a high-level "stability flag" is output; when the steady-state component proportion is lower than the preset stability threshold, the current signal is determined to have significant dynamic changes or transient disturbances, and a low-level "instability flag" is output. This output is simultaneously fed into the third-level decision logic.

[0086] Level 3: Two-condition hardware lock decision. This level consists of hardware AND gate logic circuits and a lock status register. It receives inputs from both Level 1 and Level 2 and then executes a hardware decision. The decision rules are as follows:

[0087] When both conditions are met simultaneously, it indicates that the external operating conditions have indeed undergone a substantial change. The lock status register is unlocked, allowing the field type selection stage to perform field type switching operations. When the "characteristic quantity crosses the threshold" but the "steady-state component ratio is still higher than the preset stability threshold", it indicates that the current characteristic quantity jump is caused by instantaneous noise disturbance. The system is still near the steady-state operating point. The lock status register continuously outputs a hardware lock signal to forcibly maintain the current field type operating state and prohibit field type switching operations until the steady-state component ratio falls back to the normal range and the characteristic quantity continues to stably cross the threshold before automatically unlocking.

[0088] The complete anti-shake decision process in this embodiment is as follows: (1) The hardware comparator monitors in real time whether the main path characteristic quantity crosses the preset switching threshold; (2) If it crosses the threshold, the steady-state component ratio value is read synchronously and compared with the preset stable threshold; (3) If the steady-state component ratio is higher than the stable threshold, it is determined to be an instantaneous noise disturbance, and the hardware locking circuit outputs a locking signal to prohibit switching; (4) If the steady-state component ratio is lower than the stable threshold, it is determined to be a change in actual operating conditions, the locking state is released, and field type switching is performed; (5) After the field type switching is completed, the locking state register is automatically reset and waits for the next decision cycle to restart. The above decision process is executed in a hardware pipeline throughout, and the total delay of a single decision is fixed at 3 clock cycles (15 nanoseconds), which can realize instant anti-shake locking.

[0089] This pure hardware anti-jitter mechanism effectively filters various high-frequency random disturbances, reducing the number of field pattern erroneous switching times by more than 70% under normal engineering signal-to-noise ratio conditions. Simultaneously, it compresses the consistency standard deviation of field pattern holding time from milliseconds in traditional solutions to within microseconds, completely resolving the technical defects of traditional field control circuits such as noise sensitivity, switching jitter, and system oscillation, significantly improving the smoothness and long-term operational stability of physical field closed-loop control. This anti-jitter mechanism, together with the main path feature quantity calculation unit and the anti-jitter auxiliary branch, constitutes a complete hardware closed-loop link of "main path feature quantity solution, auxiliary branch steady-state criterion generation, field pattern dual-condition hardware locking, and reliable switching of field pattern control parameters."

[0090] Example 3: Hardware Real-time Monitoring and Iterative Rollback Mechanism for System State Entropy

[0091] This embodiment corresponds to Figure 1The system state entropy real-time monitoring unit in the illustrated overall architecture discloses its pure hardware implementation scheme and its collaborative monitoring mechanism with the main pipeline. This unit extracts system operating state information in real time from the standardized tensor feature data output by the multi-dimensional embedded pipeline, independently completing real-time monitoring of the overall system's steady-state characteristics. Through hardware circuitry, it accurately identifies abnormal operating conditions such as computational convergence degradation, sudden changes in external operating conditions, and system numerical divergence, automatically triggering graded early warnings and iterative rollback protection to ensure the stability and security of the high-speed pipeline's iterative computation. This unit consists of a three-level cascaded hardware pipeline: a state feature accumulation and entropy calculation layer, a state change rate detection layer, and a two-level threshold decision and rollback trigger layer. All are implemented using pure combinational logic and sequential hardware, with no software computation involved throughout. Details are as follows.

[0092] Level 1: State Feature Accumulation and Entropy Calculation Layer. This layer consists of a hardware discrete integrator circuit and a parallel accumulation array. It accumulates the standardized tensor feature data received from the multidimensional embedding pipeline output clock cycle by clock cycle, and calculates the overall steady-state feature quantity (i.e., the system state entropy value) in real time through hardware integration operations. This integrator circuit adopts a sliding window accumulation mechanism, and the window width is pre-configured by a hardware register, which can be loaded all at once during system initialization and requires no software intervention during operation. The real-time entropy value output by this layer is simultaneously fed into the Level 2 state change rate detection layer and the Level 3 threshold decision layer.

[0093] The second stage is the state change rate detection layer. This stage consists of a historical data cache array and a first-order gradient difference operation circuit. The historical data cache array uses a circular addressing method to store multi-period historical entropy values ​​output from the first stage, providing complete time-series data support for gradient calculation. The first-order gradient difference hardware circuit reads the current entropy value and historical entropy value from the cache array, and calculates the system state change rate in real time through hardware subtractors and shifters, accurately capturing sudden changes in operating conditions and abnormal convergence of operations. The output of this stage is the state change rate value, which is synchronously sent to the third-stage threshold decision layer.

[0094] Level 3: Two-level threshold decision and rollback trigger layer. This level consists of two hardware comparators, a warning interrupt register, and a rollback state lock register. It receives the real-time entropy value from the first level and the state change rate from the second level before performing a tiered decision. This unit sets two hardware threshold levels—the first level is a slow degradation warning threshold, and the second level is a rapid divergence forced rollback threshold—corresponding to two typical abnormal operating conditions:

[0095] Level 1 Early Warning (Slow Performance Degradation) Mechanism: When the hardware comparator detects that the system's real-time entropy value continuously deviates from the normal operating range, or that the rate of state change shows a slow increasing trend (indicating a decrease in the system's convergence rate and gradual degradation of steady-state performance), this level triggers a Level 1 early warning interrupt. This interrupt signal is hardwired directly to the waveform synthesis stage and the control signal injection stage, assisting the system in adaptively adjusting iteration parameters and optimizing the calculation step size to achieve adaptive correction of minor anomalies. After the Level 1 early warning is triggered, the main waterline continues to operate without interruption, and the monitoring unit continues to maintain real-time monitoring.

[0096] Secondary rollback (rapid numerical divergence) mechanism: When the hardware comparator detects that the real-time entropy value of the system deviates sharply from the normal range in a short period of time, or the rate of change of the state suddenly jumps to above the preset danger threshold (indicating that there is a risk of numerical divergence and loss of control in the system), the secondary forced rollback interrupt is immediately triggered. The interrupt signal is directly connected to the register reset terminal and multiplexer control terminal of each module of the main pipeline via hardwire. The hardware automatically performs the following rollback operations: (1) Lock the stable field control parameters of the previous cycle latched at the output terminal of the current field selection stage, and prohibit the waveform synthesis stage from receiving abnormal data; (2) Switch the multiplexer of the multidimensional embedded pipeline input terminal to the historical state cache readback path, and reinitialize the registers of each stage of the pipeline with the historical state data of the most recent stable iteration cycle corresponding to the current field; (3) Send the rollback status flag to the hardware co-adaptation interface, instructing the external upstream module to suspend the injection of new data until the rollback completion confirmation signal is valid. The entire rollback operation is executed automatically by hardware without the need for software intervention. From the triggering of the secondary comparator to the validity of the rollback completion confirmation signal, the entire response delay does not exceed 10 clock cycles (50 nanoseconds), and can complete the rapid recovery before the system values ​​diverge to an uncontrollable state.

[0097] This hardware monitoring mechanism enables nanosecond-level real-time monitoring and fault self-correction of the system's operating status, eliminating oscillations, divergences, and instability issues during high-speed iterative computations at the hardware level. This significantly improves the long-term reliability and operational adaptability of the coprocessor pipeline. As one of the two auxiliary branches operating in parallel with the main pipeline, this monitoring unit continuously monitors the main pipeline's computational status during each iteration cycle. Its warning and rollback signals are directly connected to key control nodes of the main pipeline via hardwired connections, forming a closed-loop hardware self-healing system of "real-time monitoring, anomaly identification, tiered response, and rapid recovery," ensuring that the entire system maintains a stable and controllable iterative operating state under complex conditions.

[0098] Example 4: Coprocessor hardware adaptation to commercial chips enables intelligent temperature control and dynamic precision management engineering applications.

[0099] This embodiment corresponds to Figures 1 to 3The illustrated overall architecture demonstrates this invention as an independent coprocessor IP and peripheral circuit, supporting an engineering application scheme for existing commercial computing chips. Typical test results for overall hardware verification are also provided. This hardware coprocessor architecture requires no modification to the main chip's core architecture; it achieves full-domain chip status monitoring, intelligent thermal management, and dynamic precision adaptive adjustment solely through direct hardware connections, exhibiting strong compatibility and practical applicability.

[0100] The multi-channel synchronous sampling interface of this invention is connected to an external chip multi-point temperature sensing array and a load status acquisition unit to synchronously acquire chip global temperature distribution data and real-time computational load data. Multi-dimensional embedded pipelines are used to fuse heterogeneous data from multiple measurement points, enabling real-time calculation of chip operating state characteristics. The field-mode selection stage adaptively matches three operating modes—low-power steady state, constant temperature maintenance, and forced cooling under high thermal load—based on the chip's real-time operating state, outputting corresponding high-precision control waveforms to achieve dynamic adaptation and control of chip thermal state and computational load. Simultaneously, a signal decomposition auxiliary hardware unit filters out instantaneous noise spikes from temperature sampling in real time, avoiding system oscillations and power consumption fluctuations caused by frequent start-stop cycles of the heat dissipation drive circuit. The system status monitoring unit monitors the overall steady-state operation of the chip in real time. Under conditions of high chip heat, high load, and computational convergence degradation, it outputs graded warning and protection commands, adaptively adjusting the chip's quantization accuracy and iteration step size, and triggering iteration state rollback protection when necessary, effectively preventing abnormal problems such as chip instability and thermal runaway.

[0101] To systematically verify the functionality and timing correctness of the overall hardware circuit of this invention, multiple sets of standardized hardware test stimuli were configured to simulate conventional steady-state operating conditions, physical field parameter step change conditions, high-frequency noise interference conditions, system convergence degradation and sudden divergence conditions, and high-speed dynamic adaptation conditions, comprehensively verifying all core hardware functions of this invention. The experimental verification results show that the timing matching of each hardware module of this invention is accurate, the logical linkage is orderly, and the data interaction is reliable. There are no data conflicts, timing disorders, or invalid switching throughout the process. The pipeline operates stably overall, and the closed-loop control response is fast and the correction is accurate.

[0102] This hardware coprocessor architecture can be plugged and played to adapt to various commercial computing chips and control chips, significantly improving chip thermal balance, computational convergence stability and dynamic operating condition adaptability. It has a wide range of engineering applications, low implementation cost, and extremely high practical value and promotion prospects.

Claims

1. A multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system, characterized in that, include: The system comprises a multi-channel synchronous sampling interface, a multi-dimensional embedding pipeline, a signal feature calculation unit, a field selection stage, a waveform synthesis stage, a control signal injection stage, a signal decomposition auxiliary criterion generation hardware unit, a system state entropy real-time monitoring unit, and a hardware collaborative adaptation interface. The multi-channel synchronous sampling interface is used to synchronously acquire multiple external physical sensor signals, and features multi-domain clock alignment, synchronous sample-and-hold, and real-time hardware filtering. The multi-dimensional embedding pipeline, connected to the multi-channel synchronous sampling interface, is used to weighted fuse and convert multiple sensor data into high-dimensional coordinate vectors, supporting dynamic dimension mapping switching. The signal feature calculation unit, connected to the multi-dimensional embedding pipeline, performs hardware solution of the manifold metric tensor and the comprehensive feature scalar, and is configured with double buffering to achieve non-blocking tensor data processing. Parallel output; the field pattern selection stage is a pure combinational logic decision circuit, connected to the signal feature calculation unit, used to adaptively select multiple preset physical field patterns and output corresponding control parameters according to real-time feature quantities; the waveform synthesis stage is connected to the field pattern selection stage, used to generate high-precision controllable waveform signals in hardware; the control signal injection stage is connected to the waveform synthesis stage, used to output hardware closed-loop gradient control signals and form closed-loop feedback through hard-wired direct-connection paths; the signal decomposition auxiliary criterion generation hardware unit is connected in parallel with the double buffer and the field pattern selection stage, forming a hardware anti-jitter correction branch; the system state entropy real-time monitoring unit is connected in parallel with the multi-dimensional embedding pipeline and the control signal injection stage, forming a hardware state monitoring and iterative rollback branch; The hardware collaborative adaptation interface is respectively connected to the multi-channel synchronous sampling interface and the external tensor operation pipeline to realize front-end sensing data pre-linkage and accuracy pre-adjustment.

2. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to claim 1, characterized in that, The signal decomposition auxiliary criterion generation hardware unit is a multi-stage hard-wired pipeline structure, which sequentially includes a Hodge star operator subunit, a Hodge-Laplacian operator subunit, and a harmonic projector subunit along the signal flow direction. The Hodge star operator subunit has a built-in dimension mapping lookup table array to realize dual space mapping of different order differential forms. The Hodge-Laplacian operator subunit integrates an external differential hardware operation module, a co-differential hardware operation module, and an adder array. The harmonic projector subunit has a built-in hardware null space solution and residual convergence determination circuit for steady-state component solution output. The signal decomposition auxiliary criterion generation hardware unit outputs multiple types of signal components and outputs hardware field shape anti-jitter criteria to the field shape selection stage according to the component proportion.

3. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to claim 2, characterized in that, The signal decomposition auxiliary criterion generation hardware unit is a pure hardware pipeline architecture with a fixed computation latency of 32 clock cycles and no software iteration involved throughout the process.

4. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to claim 1, characterized in that, The real-time monitoring unit for system state entropy consists of a multidimensional numerical integrator, an entropy value history FIFO cache, a first-order differential operation circuit, and a two-stage threshold comparator cascaded in hardware. The multidimensional numerical integrator adopts a pure combinational logic discrete accumulation structure. The first-order differential operation circuit is used to calculate the gradient of entropy value change in real time in hardware. The two-stage threshold comparator is configured with two levels of discrimination thresholds, and outputs corresponding hardware warning interrupt and iterative forced rollback control signals.

5. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to any one of claims 1 to 4, characterized in that, The multi-channel synchronous sampling interface includes N parallel ADC channels, N≥3, and all channels are synchronously triggered for sampling using the same clock edge and have a unified timestamp; the multi-channel synchronous sampling interface has a built-in multi-domain phase-locked loop and sample-and-hold array, and the synchronization deviation between channels is less than 5 nanoseconds.

6. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to claim 5, characterized in that, The multidimensional embedding pipeline includes an N×M parallel multiplier array and M sets of multi-input accumulator trees, where M≥3; The multidimensional embedding pipeline supports dynamic switching of dimension mapping modes at runtime, with a mode switching delay of no more than one system clock cycle.

7. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to any one of claims 1 to 4, characterized in that, The overall calculation delay of the signal feature calculation unit is less than 50ns; the dual buffer consists of two independent storage blocks, which correspond to the pipeline write path and the auxiliary unit read path, respectively, with hardware isolation between read and write, no data conflict, and no handshake waiting.

8. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to claim 7, characterized in that, The field pattern selection stage has a built-in configurable threshold register, which can configure at least three types of adaptive field patterns: resonant field pattern, coherent field pattern, and decoherent field pattern. The field pattern selection stage uses pure combinational logic to complete the field pattern decision and switching, and the field pattern switching delay is less than 5ns.

9. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to any one of claims 1 to 4, characterized in that, The waveform synthesis stage adopts a phase accumulator and waveform lookup table architecture, with a waveform quantization accuracy of 16 bits and a signal output update rate of no less than 10 MSPS.

10. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to any one of claims 1 to 4, characterized in that, The hardware collaborative adaptation interface is a hard-wired front-end docking structure that synchronously receives the original sampling data and outputs the precision adjustment signal to the external tensor operation pipeline, thereby realizing physical condition prediction and operation precision pre-switching.

11. The multi-channel sensor synchronous sensing and closed-loop control co-processing circuit system according to any one of claims 1 to 4, characterized in that, The control signal injection stage includes a gradient calculator and an output buffer, which are connected to the external processor's computing port via hardwired direct connection to achieve hardware closed-loop feedback with no bus delay.

Images