Intelligent patrol management system based on grading quality control

Abstract

The invention relates to the technical field of security and protection monitoring, in particular to an intelligent patrol management system based on hierarchical quality control, which comprises an acquisition preprocessing module, an intention quality control module, a privacy evidence storage module and a risk gating module. Performing complex wavelet transform on the data packet, constructing a pulse diagram, performing convolution reasoning, and outputting an intention probability and a second quality score; performing hashing, complex approximate homomorphic encryption and interval zero-knowledge proof on the event vector, and writing the gradient and the third mass score into a local evidence storage library; and finally, based on the condition in-danger value and reinforcement learning, dynamically adjusting a threshold value, triggering client prompting, robot re-checking, access control interlocking or alarming, and performing write-back on an execution hash value and a reverse gradient, thereby realizing low-missing-report, low-false-report and auditable closed-loop patrol.

Description

TECHNICAL FIELD

[0001] The present application relates to the security monitoring technical field, and particularly to an intelligent patrol management system based on hierarchical quality control. BACKGROUND

[0002] Property security is upgrading from "manual fixed-time patrol" to "multi-modal intelligent patrol". The traditional system mainly relies on visible light cameras and fixed threshold alarm mechanism: on the one hand, the camera frequently misses detection under black light, backlight or shielding conditions; on the other hand, the fixed threshold is difficult to balance false positives and misses, leading to fatigue response of security personnel, reducing patrol coverage and satisfaction. In addition, the existing distributed monitoring platform usually uploads video and sensor data to the cloud for centralized analysis, which adds transmission delay and privacy leakage risk; online model updating lacks credible feedback and is vulnerable to abnormal data "poisoning". These defects are caused by single modal input, static threshold strategy and weak data security mechanism. SUMMARY

[0003] In view of the many problems existing in the prior art, the present application provides an intelligent patrol management system based on hierarchical quality control. The present application uses multi-modal synchronous collection as the source, uses pulse graph convolution and particle swarm mutual information optimization to extract high-discriminative latent representation, combines homomorphic encryption domain contrast learning and interval zero-knowledge proof to generate credible quality scores, and uses conditional value at risk to drive reinforcement learning to generate dynamic threshold and trigger hierarchical response, thereby realizing a low-delay, high-precision, verifiable and self-evolving patrol closed loop.

[0004] An intelligent patrol management system based on hierarchical quality control, comprising: A collection preprocessing module collects multi-modal sensor data and writes a timestamp, reconstructs a background and calculates an event disturbance intensity, generates a first quality score according to infrared texture consistency and millimeter wave micro-Doppler features, and outputs a disturbance data packet; An intent quality control module performs waveform transformation on the disturbance data packet, constructs a pulse graph in combination with a time-of-flight depth frame and a millimeter wave radar point cloud, generates an entity latent representation through convolution inference and particle swarm mutual information optimization, obtains an intent probability through linear discrimination and forms a second quality score, and outputs an event vector; A privacy storage module calculates a hash value for the event vector and generates a zero-knowledge proof, updates an intent model through homomorphic encryption gradient comparison learning in an encrypted domain, fuses the first quality score and the second quality score to generate a third quality score, and writes the event hash value, the third quality score, the zero-knowledge proof and the encrypted gradient into a local storage library; A risk gating module calculates a conditional value at risk according to a loss sequence recorded in the storage library, generates a dynamic threshold through reinforcement learning using the third quality score and the intent probability, triggers a multi-level response action when the third quality score exceeds the dynamic threshold, and writes an execution hash value and a reverse gradient into the storage library for feedback and updating.

[0005] Preferably, the multimodal sensors in the acquisition and preprocessing module include at least an event camera and a millimeter-wave radar. The event camera is used to output a time-resolved event stream, and the millimeter-wave radar is used to output target distance and velocity information.

[0006] Preferably, the waveform transformation performed by the intention quality control module on the disturbed data packet is a complex wavelet transform, which simultaneously extracts phase information and amplitude information to characterize the direction and energy changes of the micro-motion.

[0007] Preferably, the intention quality control module maps the texture features obtained from complex wavelet transform, the spatial features of the time-of-flight depth frame, and the velocity features of the millimeter-wave radar point cloud into pulse map nodes, wherein the texture nodes, depth nodes, and velocity nodes are connected through distance weights.

[0008] Preferably, the intention quality control module uses a pulse graph neural network implemented with silicon photonics when performing convolutional inference on the pulse graph. The silicon photonic interference structure completes the weight multiplication and accumulation operation and outputs the convolution result in the form of pulses.

[0009] Preferably, the intention quality control module uses the particle swarm mutual information optimization algorithm to update the weights of the entity latent representation. The particle swarm mutual information optimization algorithm uses the classification loss and the weighted sum of mutual information as the objective function and updates the particle velocity and position in parallel within a fixed number of iterations.

[0010] Preferably, the homomorphic encryption performed by the privacy evidence storage module on the gradient and metric of the intent model is a complex approximate homomorphic encryption scheme. The complex approximate homomorphic encryption scheme simultaneously encodes the real part and the imaginary part to maintain the vector multiplication and addition structure in the encrypted domain.

[0011] Preferably, the privacy proof module uses an interval zero-knowledge proof protocol based on elliptic curve commitment when generating zero-knowledge proofs. The interval zero-knowledge proof protocol proves that both the third quality score and the gradient of the intent model are within a preset trust range without revealing the plaintext.

[0012] Preferably, the risk gating module adopts a deep deterministic policy gradient algorithm, using the conditional risk value as the reward signal and the third quality score and intention probability as state variables to update policy parameters in real time and generate dynamic thresholds.

[0013] Preferably, when the third quality score exceeds the dynamic threshold, the risk gating module sequentially triggers multi-level response actions. The multi-level response actions include client prompting, mobile inspection robot verification, access control relay interlocking, and sending an alarm signal to the security center. After triggering each level of response action, the module generates a corresponding execution hash value and reverse gradient, writes them to the local evidence storage library for feedback updates.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By using event camera + millimeter-wave radar synchronous acquisition and generative background reconstruction technology, millisecond-level detection of disturbances was achieved in complex lighting and occlusion scenarios.

[0015] By using complex wavelet waveform transform and pulse graph convolution fusion techniques, high-confidence intent assessment of cross-modal micro-motion direction, depth, and velocity was achieved.

[0016] By employing complex approximate homomorphic encryption and interval zero-knowledge proof techniques, we have achieved trusted storage and comparative learning of gradients and quality scores within the local encryption domain, balancing privacy and adaptability.

[0017] By employing a condition-based value-driven, deep deterministic strategy gradient technique, real-time threshold adjustment and multi-level response closed loops are achieved, significantly reducing false negatives and false positives.

[0018] By employing hash value and inverse gradient writing chain feedback techniques, decision information is fed back across layers, enhancing the model's ability to continuously evolve. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the interaction of the system of the present invention; Figure 2 This is a flowchart of the homomorphic encryption comparison learning and zero-knowledge proof process in this invention; Figure 3 This is a schematic diagram of the risk gating dynamic threshold and multi-level response in this invention. Detailed Implementation

[0020] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0021] like Figure 1 As shown, an intelligent inspection and management system based on hierarchical quality control includes: The data acquisition and preprocessing module acquires multimodal sensor data and writes it to a timestamp, reconstructs the background and calculates the event disturbance intensity, generates the first mass score based on infrared texture consistency and millimeter-wave micro-Doppler features, and outputs the disturbance data packet. The acquisition and preprocessing module first triggers synchronous sampling of multimodal sensors using a unified clock. The event camera outputs a sparse event stream, the time-of-flight depth sensor outputs a 3D point-to-point matrix, the millimeter-wave radar outputs target range and radial velocity point clouds, and the infrared camera outputs texture intensity images. All data are appended with nanosecond-level timestamps at the driving layer to ensure no inter-frame jitter in the subsequent fusion stage. After synchronization, the edge computing unit combines multi-source data from the same moment into a single batch cache based on the timestamps, laying the foundation for subsequent background reconstruction and perturbation estimation.

[0022] Background reconstruction employs a generative diffusion model for edge-side inference, requiring only one inverse diffusion iteration to obtain the unmanned physical background. The model is pre-trained offline using static images captured from the scene, and outputs background frames online with minimal and rapid sampling. Since the event stream contains positive and negative polarity pulses, the system first integrates these pulses on a micro tensor processor to generate a grayscale tensor, then performs an absolute difference operation with the generative background frame at the pixel level to form a difference map of the same size. Based on application requirements, this invention proposes a perturbation intensity measurement formula:

[0023] in Indicates time The scalar of the disturbance intensity. Indicates the event stream integration at the pixel level. grayscale value at Indicates the background frame in pixels grayscale value at that location This represents the number of pixels in the region of interest. The formula, by accumulating pixel-by-pixel differences and then normalizing, allows for comparison of camera outputs at different resolutions and fields of view under a unified dimension. Experiments have shown that when stationary or slowly moving objects appear in the scene, It grows linearly, making it easy to set thresholds.

[0024] The infrared texture consistency index, calculated based on structural similarity, primarily reflects the degree of damage caused by the intruding object to the background thermal texture. Millimeter-wave micro-Doppler breathing characteristics, obtained through short-time Fourier transform, yield the target's main breathing peak frequency, effectively distinguishing real humans from static camouflage. The system normalizes these three indices to the 0-1 range and then uses a weighted average to generate the first quality score. The weights are automatically adjusted during deployment based on historical false positives and false negatives, requiring no manual calibration.

[0025] The acquisition and preprocessing module encapsulates the disturbance intensity, time-of-flight depth frame, millimeter-wave radar point cloud, first quality score, and timestamp into a disturbance data packet, which is then transmitted to the intent quality control module via a high-speed bus. The encapsulation uses a fixed field order and length to ensure that downstream inference requires no parsing overhead. Through structured output, this invention transforms traditional pipeline-style camera data preprocessing into event-driven data fragment preprocessing, significantly reducing the complexity of multimodal alignment.

[0026] To verify the effectiveness, three scenarios—residential passageway, underground parking garage, and outdoor backlighting—were compared. The results show that, after adopting generative background reconstruction and event-driven differential, the average frame time on a single edge device is 21 milliseconds, while the traditional frame difference plus morphological method takes 35 milliseconds; under strong light flickering conditions, the false alarm rate decreased from 12% to 5%. This confirms that the background reconstruction and disturbance assessment scheme of this invention has advantages in both timeliness and accuracy.

[0027] Example 1: Deployed in an underground parking garage for 7 days, the system detected a total of 38 prolonged loitering events. All were accurately assigned a high intent probability by the intent quality control module, and the risk gating module automatically triggered access control relays and audible / visual warnings based on the third quality score. In contrast, the traditional system relying solely on facial whitelist filtering missed 11 detections and generated 29 invalid alarms. This example demonstrates that dynamic quality scoring and tail-end risk gating can effectively reduce missed detections in complex lighting and low-contrast environments.

[0028] Preferably, the multimodal sensors in the acquisition and preprocessing module include at least an event camera and a millimeter-wave radar. The event camera is used to output a time-resolved event stream, and the millimeter-wave radar is used to output target distance and velocity information.

[0029] The acquisition and preprocessing module utilizes a combination of an event camera and millimeter-wave radar because their complementary nature in temporal resolution, spatial resolution, and media penetration capability meets the requirements of the hierarchical quality control framework for continuous and complete perception. The event camera outputs a polarized event stream with microsecond-level temporal accuracy, with each event containing pixel coordinates, a timestamp, and a brightness change symbol. Compared to traditional frame-by-frame grayscale images, the event stream is naturally sparse, has low bandwidth consumption, and does not exhibit blurring under sudden changes in illumination or high-speed motion scenes. This invention leverages the temporal accuracy of the event stream to quickly construct a pseudo-grayscale image under the current illumination conditions through short-window integration. This image is then differiated with the static background frame output by the generative model, enabling the detection of disturbances within the first few milliseconds after a target enters the field of view. The millimeter-wave radar operates in the 77GHz band and outputs a triplet of reflection intensity, radial distance, and radial velocity. Because millimeter-wave wavelengths are much longer than visible light wavelengths, they have good penetration through rain, fog, smoke, and plastic obstructions, providing stable motion information even when vision fails. In addition, millimeter-wave signals produce strong reflections on metal surfaces, which can be used to identify suspicious persons carrying large metal objects at an early stage.

[0030] The module first transforms the millimeter-wave radar coordinate system to the event camera pixel plane using a target scene general extrinsic parameter matrix. Let the radar output target polar coordinates be... radial velocity within quadrant ,in Indicates distance, Indicates azimuth. Indicates pitch angle, Indicates radial velocity. Transformation to the pixel plane uses a calibrated rotation matrix. Translation vector Then multiply by the camera intrinsic parameter matrix The core projection formula is denoted as:

[0031] in These are the mapped pixel coordinates. The three principal elements correspond to the focal length and the principal pixel position, respectively. and Obtained through a single offline extrinsic parameter calibration, ensuring physical alignment of the two sensors. After mapping, the radar point cloud is written into the depth channel with the same resolution as the event stream. The direct effect of this is that the system can retain distance and velocity information even in visually completely black areas, solving the problem of targets being completely occluded by the event camera under nighttime and backlighting conditions.

[0032] On the event stream side, this invention introduces a threshold adaptive integration strategy. This will continuously... Positive events within microseconds are counted as positive energy, and negative events as negative energy. If the difference between positive and negative energy exceeds a set threshold, a grayscale value is generated at the corresponding pixel. This was determined through experiments. At a saturation of 1000µs, the contours of moving targets can still be accurately reconstructed under indoor illumination of 5 lux. This grayscale image is used for subsequent generative background model differencing. To measure the spatial coupling between radar reflection points and event grayscale, this invention defines a coupling index:

[0033] in This represents the number of radar points that fall within the effective pixel area of ​​the event image in the current time window. For the first Normalized grayscale of each pixel This represents the radar reflection intensity falling on this pixel. Coupling degree. A higher value indicates that the target exists simultaneously in both the visual and radar domains, which can improve the confidence level of subsequent intent quality control modules; if If the value remains close to 0 for an extended period, it indicates that the target is visible in the radar domain but not in the visual domain. The risk gating module needs to lower the dynamic threshold to prevent missed detections.

[0034] The acquisition and preprocessing module outputs a disturbance data packet, with fields including timestamp, event disturbance intensity, time-of-flight depth frame, millimeter-wave radar point cloud, first mass score, and coupling index. The structured data packet is transmitted via a low-latency bus, and the intent quality control module can directly determine whether to perform compensation inference for the missing modalities based on the coupling degree index.

[0035] Example 2: Deployed in an underground parking garage for 72 hours with an average illuminance below 10 lux, the system recorded 63 incidents of people staying in the garage. Of these, 59 were simultaneously captured by both the event camera and millimeter-wave radar, with an average coupling degree of approximately 0.72; 4 were captured solely by millimeter-wave radar, with a coupling degree below 0.05. However, the system still utilized radar point cloud data to trigger risk gating, maintaining a false negative rate of 0.0%. Compared to the baseline solution relying solely on vision, the false negative rate was reduced by 14%, demonstrating that a high coupling degree index can effectively enhance the front-end perception reliability of the hierarchical quality control process.

[0036] Furthermore, this invention dynamically adjusts the polarization noise threshold for event cameras. When vehicle headlights flash or pedestrians hold strong light sources, a large amount of high-frequency noise appears in the event stream. The system analyzes the millimeter-wave radar velocity histogram in real time. If it finds that the number of low-speed targets in the scene is zero while the event stream trigger rate increases significantly, it automatically increases the event polarization threshold to reduce noise. This strategy shares the same criterion with the coupling degree index, thus eliminating the need for repeated calculations. Experiments show that with a vehicle headlight flashing frequency of 20Hz, the dynamic threshold can reduce the event noise trigger rate from 120kEPS to 15kEPS ("kEPS" is an abbreviation for kilo-EventsPer Second, commonly used in the event camera field to measure the event stream trigger frequency), while affecting the normal pedestrian event stream density by less than 5%.

[0037] The intent quality control module performs waveform transformation on the disturbed data packets, constructs a pulse map by combining the time-of-flight depth frame and millimeter-wave radar point cloud, generates entity latent representations through convolutional inference and particle swarm mutual information optimization, obtains intent probability through linear discrimination and forms a second quality score, and outputs event vectors. The intent quality control module receives the perturbation data packets output by the acquisition and preprocessing module. Its core objective is to extract potential features reflecting personnel intent and threat levels from multimodal short-time-series data, and to quantify the reliability of these features using a tiered quality control approach. The module first performs waveform transformation on the event stream subframes, employing complex wavelet functions for combined amplitude and phase analysis. The event stream is then delineated at the pixel level. The time series is The complex wavelet coefficients are:

[0038] in For scale parameters, For translation parameters, For complex conjugate wavelets. Through scale sparsification, the intention is for the quality control module to retain only the scale set. The corresponding amplitude and phase tensors are used to capture subtle movements and stroboscopic textures at different time scales without increasing bandwidth.

[0039] Subsequently, the module maps the wavelet tensor, time-of-flight depth frame, and millimeter-wave radar point cloud to a unified pulse map. During the mapping process, each wavelet pooling window is treated as a texture node, the time-of-flight depth voxel as a depth node, and the millimeter-wave reflection point as a velocity node. The three types of nodes are merged into a vertex set. Graph edge weights are established between nodes using a weight function based on physical distance, allowing features from different spatially proximate modalities to complement each other. The convolutional inference stage uses a three-layer pulsed graph convolutional network. Each layer calculates a weighted sum of adjacent features using a pulsed convolution kernel, and then outputs the updated node representation through pulsed activation. Node embeddings are obtained after convolutional inference. .

[0040] To ensure rapid adaptation of convolutional network weights within limited edge computing power, this invention introduces a particle swarm optimization algorithm based on mutual information. The particle swarm size is set to ten, with each particle corresponding to a set of convolutional kernel weights. The objective function is the sum of classification loss and mutual information, where the classification loss term evaluates the intention probability error based on cross-entropy, and the mutual information term measures the amount of information transferred between the entity latent representation and the perturbation label. The weighted sum of these two terms is used to update the particles through velocity-position iterative updates. After multiple iterations, the particle swarm automatically converges to a weight set that balances discriminative power and information integrity. This method reduces the reliance on a large number of labeled samples compared to traditional stochastic gradient descent, and the addition of mutual information reduces the risk of overfitting with small sample sizes.

[0041] After particle swarm optimization, the module performs linear discrimination for each entity in the convolutional embedding space. The parameters of the linear discriminator are solved using a least-squares closed-form algorithm, and the output value is normalized by a logistic function to obtain the intent probability. A high intent probability indicates that an entity is exhibiting abnormal behavior such as lingering, going against traffic, or obstructing a camera; a low intent probability is considered normal passage. To form a second quality score, the intent quality control module fuses the intent probability with the first quality score using a weighted average. If the first quality score indicates low source signal quality, the second quality score automatically decreases, raising the subsequent risk gating threshold; if both the first quality score and intent probability are high, the second quality score significantly increases, prompting the risk gating module to quickly enter a high-level response.

[0042] This module outputs an event vector, with fields including entity latent representation, intent probability, first quality score, second quality score, and timestamp. The event vector is written to shared memory in a fixed byte sequence for the privacy evidence storage module to read, avoiding redundant parsing. Application example: After deploying the intent quality control module in an office building lobby, 48 hours of data analysis yielded 1420 normal passes and 18 suspicious loitering events. Traditional systems based solely on frame difference and thresholds resulted in 6 missed detections and 36 false positives; the intent quality control module of this invention, under the same data conditions, had no missed detections and only generated 4 false positives, with a false positive rate controlled at 0.28%, demonstrating the combined advantages of particle swarm optimization and multimodal pulse graph convolution.

[0043] Example 3 addresses a complex occlusion scenario at a subway station entrance. Due to high pedestrian density during peak hours, the time-of-flight depth frame exhibits significant gaps, and the millimeter-wave radar experiences multipath propagation due to the metal structure. The intent quality control module utilizes wavelet phase information to detect the continuity of the traversal trajectory. By leveraging the complementarity of depth and velocity nodes, it maintains stable convergence of the entity latent representation. In actual testing, with occlusion occupying 60% of the field of view, the correlation coefficient between the second quality score and the trigger rate of the subsequent risk gating module reaches 0.89, higher than the 0.62 of the traditional scheme based on optical flow direction histograms, demonstrating the high reliability of this module under complex field of view conditions.

[0044] Preferably, the waveform transformation performed by the intention quality control module on the disturbed data packet is a complex wavelet transform, which simultaneously extracts phase information and amplitude information to characterize the direction and energy changes of the micro-motion.

[0045] In this invention, complex wavelet transform is used to analyze sparse pulse signals generated by an event camera within an extremely short time window. Its core idea is to simultaneously preserve both amplitude and phase information. Amplitude characterizes the intensity of brightness abrupt changes and reflects energy variations; phase characterizes subtle translations of the waveform along the time axis and directly maps motion direction and velocity trends. Traditional real wavelets only output amplitude and lose phase, thus failing to distinguish between two edges with the same amplitude but opposite directions. Relying solely on amplitude can easily lead to misjudgments when detecting probing or occlusion-like lingering behaviors. This invention treats complex wavelet coefficients as two-dimensional vectors and performs angular encoding on the phase, accurately preserving subtle motion directions, which then work in conjunction with depth and velocity features in subsequent pulse map convolutional layers.

[0046] Event stream in pixels The time series is denoted as The complex wavelet mother function is obtained through scaling and translation using Morlet form. This invention retains only four discrete scales to accommodate hardware caching. The continuous complex wavelet transform is defined as:

[0047] in For pixels In scale Translation The complex coefficients, It is a complex conjugate wavelet. Indicates the time scale, used to distinguish between rapid shaking and slow motion; Indicates time offset, used to locate local mutations; It is a pulse sequence that retains sub-millisecond sampling accuracy even after threshold discretization.

[0048] At the algorithm implementation level, the hardware first performs piecewise parallel accumulation of the event stream; each accumulation buffer is 1024µs long, which can cover the pixel duration of a typical pedestrian's one-step stride. After accumulation, a Fast Fourier Convolution is performed on the on-chip vector instruction set, processing the real and imaginary channels. The output complex coefficient tensor is split into two paths: the magnitude tensor is fed into the texture node channel of the pulse map convolutional network, and the phase tensor is converted into an angle by the inverse tangent function and encoded as a direction vector, which also serves as a texture node feature. Since only four scales are retained, the tensor size is limited to 32KB per frame, allowing for real-time processing on edge devices with limited video memory.

[0049] In terms of performance, complex wavelet transform demonstrates significantly higher resolution for fine-grained motion than real wavelet transform. Experiments compare the "waving hand in place" and "passing normally" actions within a unified event stream. While the amplitude distributions are similar, the phase distributions show significant differences. Using the complex wavelet features of this invention, the intent probability obtained by the intent quality control module is increased by an average of 0.23, while the false alarm rate is reduced by 8%. In a subway security check example, a passenger repeatedly attempts to pass through the turnstile but fails. The complex wavelet phase sequence forms an alternating offset pattern, reducing the cluster radius of the entity latent representation by 35%, thus assigning a high intent probability to the linear discriminator and triggering manual review; traditional methods fail to recognize this due to insufficient amplitude jitter.

[0050] This invention also fully utilizes complex wavelet features through particle swarm mutual information optimization. When individual particles search the weight space, they calculate the mutual information between the complex wavelet phase and velocity features. High mutual information indicates that the directional features and velocity trends are consistent, which helps improve the discriminative power of latent representations. In practice, after adding the mutual information term to the objective function, the variance of entity latent representations within the same behavioral category decreases by 18%, and the cross-category margin increases by 22%, improving the threshold robustness of subsequent risk gating.

[0051] Comprehensive Example: Deployed for 24 hours in a parking lot nighttime scenario, the system recorded 76 normal entry events, 5 reverse approach events, and 2 stationary tailgating events. After combining complex wavelet transform with mutual information optimization, the second quality score averaged 0.81 for reverse approach events and 0.19 for normal entry events, with a discrimination index of 0.62. If real wavelet transform is used instead, the discrimination index drops to 0.35. This demonstrates that complex wavelet transform is an indispensable key step in evaluating the direction of micro-motions.

[0052] Preferably, the intention quality control module maps the texture features obtained from complex wavelet transform, the spatial features of the time-of-flight depth frame, and the velocity features of the millimeter-wave radar point cloud into pulse map nodes, wherein the texture nodes, depth nodes, and velocity nodes are connected through distance weights.

[0053] After completing the complex wavelet transform, the intent quality control module needs to convert the multimodal features into a unified data structure so that subsequent convolutional networks can perform neighborhood aggregation on the same topology. This invention designs this data structure as a pulse graph. A pulse graph is a weighted directed graph with temporal information, where vertices represent local features observed by the sensor, edge weights represent the proximity between vertices in physical or semantic space, and timestamps are encoded through node self-excited pulse signals, thus preserving both spatial and temporal dependencies during convolutional operations.

[0054] Texture nodes are derived from the results of complex wavelet transform. The event camera frame is pooled in an 8×8 pixel sliding window. Each window outputs the mean amplitude and mean phase. The mean amplitude serves as an energy descriptor, measured in normalized luminance difference values, while the mean phase reflects the dominant motion direction within that window, measured in radians. For the center pixel coordinates of each window... This invention creates a texture node in the pulse map. The feature vector of this node is set to two dimensions, where the first dimension stores the average amplitude, the second dimension stores the average phase, and the center moment of the event time window is attached as the pulse trigger time. The time resolution is quantized to the 1000 microsecond level, which preserves directional changes without putting pressure on edge hardware storage.

[0055] Depth nodes are derived from time-of-flight depth frames. These frames are first bilinearly upsampled to match the event resolution, then voxelized to 16×16 pixels, with the median depth within each voxel representing its spatial distance. The voxel center coordinates are transformed using camera intrinsic and extrinsic parameters to obtain 3D coordinates, which are then projected onto the pixel plane, sharing an indexing system with the texture nodes. Each depth node contains a scalar depth feature and a timestamp. The timestamp is aligned with the event window time, ensuring synchronization between the two modalities during convolution operations.

[0056] Velocity nodes are derived from millimeter-wave radar point clouds. After extrinsic parameter mapping, if a point cloud point falls within the pixel window of a texture node, it is classified into the velocity node corresponding to that window. Multiple points in a single frame may fall into the same window, thus clustering is necessary. Density-based clustering is used, and the average radial velocity is taken as the velocity feature after clustering. The velocity feature is measured in meters per second, and cosine correction is applied to the illumination angle to avoid attenuation due to oblique motion. Velocity nodes also record timestamps, maintaining consistency within the millimeter-wave radar pulse repetition cycle.

[0057] Edge weights need to be established for all three types of nodes. This invention uses an exponential decay function based on Euclidean distance to construct edge weights. Let the three-dimensional coordinates of the texture node or depth node be... The coordinates of the other node are The edge weight function is defined as:

[0058] in This is the distance attenuation constant, used to adjust the attenuation rate. This represents the Euclidean norm of the difference between the coordinates of two nodes. Through exponential decay, physically close nodes have larger weights, while distant nodes approach zero, thus achieving higher focus on local features during pulsed convolution.

[0059] At the implementation level, to avoid computational redundancy caused by creating pairwise edges for all nodes, this invention employs a spherical neighborhood search. First, a spatial index is established for each type of node, then only in nodes with a radius of [missing information] are [searched / searched]. Query potential neighbors within the sphere's neighborhood. This depends on the event camera resolution and scene size. For example, in an office environment, a camera installed at a height of 2.5 meters with a radius of 1.5 meters can cover the same workstation area. By limiting the neighborhood, the average degree of the pulse map is reduced to around 12, significantly reducing the amount of subsequent convolution calculations.

[0060] The temporal information of the pulse graph is encoded using node-specific self-excited pulses. Each node carries a timestamp, and during convolution operations, the pulses are only encoded when the time difference is less than a certain threshold. Pulses propagate between nodes. Setting it to 5 milliseconds eliminates spurious dependencies caused by cross-frame time differences. The temporal gating mechanism ensures that the neighborhood features extracted by convolution reflect the same dynamic process, mitigating matching misalignments caused by camera shake or high-speed target movement.

[0061] In terms of performance, the pulse graph structure achieved efficient fusion of multimodal information in experiments. Compared with simple feature concatenation methods, pulse graph convolution improved the overall accuracy by 11%. Test results on a set of railway station gate datasets show that pulse graph convolution can accurately correlate changes in human texture direction with millimeter-wave velocity surges, quickly identifying fare evasion by going against the flow of traffic. The entity latent representation output by the convolution is... In the visualization of random neighborhood embedding, the distribution is clustered. The inter-cluster distance of the three types of behaviors, namely normal passage, reverse approach, and tail entry, is increased by 1.8 times, and the intra-class variance is reduced by 36%, which proves that node construction and edge weight design effectively improve feature separability.

[0062] Example 4: Monitoring of personnel movement was conducted for 30 days in a factory safety passageway, where there were numerous forklift reflections and metal tooling obstructions. Traditional solutions based on single vision or radar had a false alarm rate of 9%. The pulse graph node mapping of this invention incorporates the spatial overlap of depth and velocity nodes into the convolution, enabling the system to determine the actual collision risk when forklifts and pedestrians coexist, achieving a false alarm rate of 0.5%. Comprehensive evaluation shows that the multimodal pulse graph structure is an indispensable intermediate medium for achieving hierarchical quality control.

[0063] Preferably, the intention quality control module uses a pulse graph neural network implemented with silicon photonics when performing convolutional inference on the pulse graph. The silicon photonic interference structure completes the weight multiplication and accumulation operation and outputs the convolution result in the form of pulses.

[0064] After the pulse graph is constructed, the intent quality control module needs to efficiently aggregate node features to capture cross-modal correlations between texture, depth, and velocity. This invention uses a pulse graph neural network implemented with silicon photonics to complete this convolutional inference. The silicon photonics interference structure utilizes the principle of optical interference to realize matrix multiplication and accumulation operations in the waveguide. Compared with digital circuits, it has the advantages of extremely low latency and extremely low power consumption, making it particularly suitable for deployment in power-constrained edge inspection terminals.

[0065] The core unit is a Mach-Zehnder interferometer array. Each interferometer consists of an input coupler, two phase-shifting waveguides, and an output coupler. When two optical signals enter the interferometer arms, the phase difference determines the output intensity, which is proportional to the product of the input amplitude and the phase difference. Therefore, by controlling the voltage of the electro-optic or thermo-optic modulators on the two phase-shifting waveguides, any real weight can be mapped to the phase difference. The arrayed interferometers form a matrix multiplier in space, capable of performing the multiplication and accumulation of all node neighborhood weights in a single operation.

[0066] Record the nodes in the pulse graph The input pulse amplitude is ,node To the node The edge weight is The computation achieved by the silicon photonics convolutional layer is as follows:

[0067] in Represents nodes The set of adjacent nodes, Input vector for the node. This is the output vector of the convolution. and Both represent pulse amplitude vectors.

[0068] The hardware implementation steps are as follows. First, an on-chip pulsed laser generates pulses at a fixed frequency; the pulses are then distributed to the input of the interferometric array via a multiplexer. Node features are encoded by modulating the pulse amplitude, and edge weights are encoded by phase modulation. The interferometric output is converted into a voltage by a low-noise photodetector, and then sampled by an analog-to-digital converter and written to the on-chip static random access memory for access by the particle swarm mutual information optimization algorithm. Since the duration of a single pulse is only about 50 ps, ​​the delay of the entire multiply-accumulate cycle is limited to 2 ps, far lower than the 0.3 ms achieved by a similarly sized digital convolution.

[0069] To ensure the accuracy of the weights, this invention designs a slope calibration process. Step one: Inject a standard pulse sequence into the interferometric array and acquire the output matrix; Step two: Compare the output with the target weight matrix to obtain the error matrix; Step three: Adjust the driving voltage of each phase shifter according to the error matrix; Repeat this process until the largest element of the error matrix is ​​less than a set threshold. Experimental results show that this calibration can improve the weight accuracy to an equivalent of 888 bits.

[0070] The entity latent representation output by the convolution is set to 64 dimensions. After impulse shaping, it is written to the on-chip cache, and then the weights are iteratively updated by the particle swarm optimization algorithm based on mutual information. The particle swarm size is 10, and each particle corresponds to a set of convolution kernel weights. The objective function is defined as the weighted sum of classification cross-entropy and mutual information; after 15 iterations, the objective function converges, and the intention probability error is less than 5%.

[0071] In terms of power consumption, a single-layer 128×128 silicon photonic convolutional array consumes approximately 0.9W. Compared to the 7.2W of a similarly sized edge graphics processor convolutional layer, this represents a power reduction of nearly 8 times. Regarding latency, when the node degree is 16, the silicon photonic convolutional layer has a latency of approximately 2ps, while the electronic implementation achieves a latency of 2×10... 2 This represents a reduction of more than five orders of magnitude, on the order of µs.

[0072] Example 5: The silicon photonics convolutional accelerator card designed in this invention was deployed at an airport security checkpoint for 10 hours. The system processed 1521 passengers, detecting 4 passengers who repeatedly returned to the scanning area due to leftover luggage. These passengers were assigned a high intent probability and triggered manual review. No passengers were missed. In contrast, a pure graphics processor solution was tested on the same hardware platform, where the latter accumulated two missed passengers due to latency.

[0073] Example 6: Monitoring fare evasion at subway entrances. The silicon photonics convolutional layer has a delay of 2 ps, and the total delay of other modules is 35 ms. The system can provide a risk assessment prompt before passengers complete the gate interaction, which is 270 ms earlier than manual monitoring on average, and successfully intercepted 3 fare evasion incidents.

[0074] The experiment also evaluated the system's battery life. In the mobile inspection robot, after replacing the silicon photonics convolutional layer, the robot's continuous operating time increased from 5 hours to 9 hours with a 100Wh battery, meeting the requirements for unattended operation at night. The silicon photonics network was tested in outdoor environments at low temperatures of -10°C and high temperatures of 60°C, and the weight drift was controlled within 1%, indicating that the calibration scheme of this invention can effectively suppress thermal drift.

[0075] Preferably, the intention quality control module uses the particle swarm mutual information optimization algorithm to update the weights of the entity latent representation. The particle swarm mutual information optimization algorithm uses the classification loss and the weighted sum of mutual information as the objective function and updates the particle velocity and position in parallel within a fixed number of iterations.

[0076] The intent quality control module needs to further optimize the weights on the entity latent representation output by the silicon photonics convolutional layer to ensure stable discrimination capability under different scenarios and lighting conditions. This invention introduces a particle swarm optimization algorithm with mutual information to quickly search for high-quality weight vectors in low-sample environments. The particle swarm algorithm is derived from the simulation of bird flock foraging behavior. Each particle corresponds to a set of weights to be optimized, and the particles search for the optimal solution in a continuous space through iterative velocity and position. Traditional particle swarm optimization often uses pure classification loss as the objective, which is prone to getting stuck in local optima under imbalanced samples or noisy labels. To address this, this invention adds a mutual information term to the objective function to measure the amount of information transferred between the entity latent representation and the perturbation label, so that the search direction not only focuses on minimizing error, but also encourages the latent representation to retain the maximum discriminative information.

[0077] The objective function is defined as:

[0078] in The objective function value, For classification cross-entropy loss, Represents the latent representation vector of an entity. Indicates the disturbance category label. For latent representation and label mutual information, and To balance the factors, both are determined offline in this invention using a grid search. Mutual information is obtained through a method based on... The nearest neighbor approximation estimator can be computed in a single scan at the edge.

[0079] The algorithm consists of three stages: initialization, iterative update, and convergence determination. In the initialization stage, 10 sets of weight vectors are uniformly sampled within the weight space, serving as the positions of 10 particles. Each particle is randomly assigned an initial velocity vector. In the iterative update stage, each particle sequentially performs the following steps: First, calculate the objective function value of the weight corresponding to its position in the current batch of samples; second, compare the particle's historical best value with the global best value, updating the individual extreme value and the global extreme value; third, correct the velocity according to the velocity update equation, and then add it to the current position to obtain the new position. The velocity update equation includes an inertia term, an individual cognitive term, and a group social term. This invention sets the inertia weight to 0.4, the individual coefficient to 1.2, and the group coefficient to 2.0, achieving convergence within 15 iterations.

[0080] To adapt to real-time scenarios, this invention uploads the globally optimal weights to a silicon photonics phase shifter array after each iteration, and refreshes the phase shift voltage via a digital-to-analog converter. The response time of the silicon photonics array is 500 nanoseconds, far less than the time consumed in a single iteration of the particle swarm optimization, thus weight refreshing will not be a bottleneck. The algorithm is considered convergent when it reaches a fixed number of iterations or when the objective function decreases by less than 1e-3, at which point the final weights are output and the algorithm proceeds to the next batch.

[0081] The introduction of mutual information terms significantly reduces overfitting in small sample environments.

[0082] In Example 7, with only 20 labeled anomalous samples provided, the model optimized by pure classification loss had a mean squared error of 0.092 on the intent probability in the validation set, which decreased to 0.057 after adding a mutual information term. Further expanding the anomalous samples to 100 narrowed the gap between the two methods, but the mutual information optimization still maintained an advantage of 0.01, demonstrating its robustness to fluctuations in the number of samples.

[0083] In the security patrol robot test, the robot patrolled an unmanned warehouse for 6 hours, collecting 1800 frames of event streams. The particle swarm optimization, through mutual information optimization, achieved a weight distribution that completely separated normal and abnormal behaviors in the intent probability distribution during the third iteration; subsequent iterations mainly fine-tuned the boundaries. Over the six hours, the intent quality control module triggered 27 high-risk events, with 0 false negatives and 2 false positives compared to manual annotation. Removing the mutual information term increased false positives to 7 and false negatives to 1. Power consumption assessment showed that this optimization stage consumed an average of 0.2W on the edge processor, saving over 90% compared to full-precision backpropagation training.

[0084] The algorithm also possesses online adaptability. When a sudden change in lighting conditions causes an overall change in the event flow amplitude, the mutual information estimate immediately decreases. The particle swarm optimization guides weight migration through a new global extremum, restoring the original classification level in just 5 iterations. This invention was validated in an outdoor parking lot during the dawn-dusk transition. Changes in the sunlight angle caused approximately a 30% change in the event amplitude contrast. The particle swarm mutual information optimization completed weight adaptation within 300 milliseconds, while traditional offline models require waiting for the weights to be retrained in the background, resulting in a latency of over 5 minutes.

[0085] like Figure 2 As shown, the privacy evidence storage module calculates the hash value of the event vector and generates a zero-knowledge proof. After homomorphically encrypting the gradient, it compares and learns in the encrypted domain to update the intent model. It merges the first quality score and the second quality score to generate a third quality score, and writes the event hash value, the third quality score, the zero-knowledge proof and the encryption gradient into the local evidence storage library. The privacy evidence storage module is responsible for transforming the event vectors output by the intent quality control module into traceable and privacy-preserving evidence entries, and providing a homomorphic secure channel for subsequent gradient backflow. The event vectors contain latent entity representations. Intent probability First mass score Second mass fraction and timestamp First, a sponge-type hash function is invoked within the trusted execution environment to generate event hash values. The hash function is based on a permutation network design, and its input is... The byte string is output as a fixed-length 256-bit hash value. Hash functions possess one-wayness and weak collision resistance, providing a unique identifier for subsequent indexing and existence proofs without exposing the original characteristics.

[0086] To prove that the quality score and gradient are within a reliable range, the module uses a zero-knowledge proof protocol. The protocol is based on elliptic curve commitments, embedding the value to be proven into elliptic curve points. Through polynomial commitments and the discrete logarithm problem, it ensures that the verifier cannot deduce the plaintext. The proof process generates a proof message. With a size of less than 1000 bytes, it can be quickly verified at the edge.

[0087] The homomorphic encryption part employs a complex approximate homomorphic encryption scheme. This scheme, based on the Qin Yixin error distribution, encodes the real and imaginary parts into the same polynomial coefficients. Let the gradient vector obtained from the backpropagation of the intended model be... Encryption operation Output ciphertext The ciphertext supports addition and scalar multiplication for performing contrastive learning in the encrypted domain. Contrastive learning constructs the contrastive loss using latent representations of Euclidean distance, and Chebyshev polynomial approximations are applied to all distance operations, eliminating the need for decryption. The updated gradient is then... It remains in encrypted form and is written back to the intent quality control module via the same key link. The hardware security unit decrypts it and superimposes it according to the particle swarm velocity rule, enabling the model to be updated online without exposing the gradient plaintext.

[0088] Third mass score The credibility of the entire privacy evidence preservation process is assessed using a weighted fusion method:

[0089] in The three coefficients are determined based on historical false alarms and missed alarms during deployment. This represents the confidence level of the zero-knowledge proof verification result, with 1 for a successful proof and 0 for a failed proof. The weighting coefficients are stored in an encrypted register to prevent tampering. After calculation, the module outputs the proof entry, with fields including the event hash value. Third quality score Zero-knowledge proof and encryption gradient Entries are written to a local immutable evidence repository indexed by timestamps. The evidence repository uses segmented writing and read-only mapping to block deletion and overwriting paths at the hardware layer.

[0090] The hardware implementation consists of three layers: a trusted execution environment processes plaintext sensitive data; an encryption accelerator implements polynomial multiplication with near-homomorphic encryption for complex numbers; and a zero-knowledge proof acceleration unit generates proofs through an elliptic curve dot product pipeline. These three layers are interconnected via shared memory, achieving read / write latency of less than 20 ns. The entire process takes approximately 14 ms to process a single event vector, with 9 ms dedicated to encryption multiplication, 3 ms for proof generation, and 2 ms for hash and quality score fusion.

[0091] Example 8: Two edge nodes were deployed at the mall entrance and exit, running continuously for 96 hours and recording 51,432 event vectors. The privacy evidence storage module completed all hashing and proof generation, with an average latency of 13.8ms per record, meeting real-time requirements. During network outage testing, the evidence storage database continued to be written locally, and automatically synchronized to the central log after network recovery, with no entries lost. Compared with the unencrypted transmission scheme, this invention reduces the potential data leakage risk by 60%.

[0092] Example 9: An external penetration testing team attempted to read the gradient in plaintext via a side-channel attack. The gradient only exists in plaintext within a trusted execution environment, and the evidence entry is stored in ciphertext; the penetration team was unable to reproduce the gradient content. Fault injection testing showed that even after 20 forced power outages, the evidence database entries could still be recovered, and the third quality score differed by less than 1% from the average before the power outage after restarting, demonstrating robustness.

[0093] Preferably, the homomorphic encryption performed by the privacy evidence storage module on the gradient and metric of the intent model is a complex approximate homomorphic encryption scheme. The complex approximate homomorphic encryption scheme simultaneously encodes the real part and the imaginary part to maintain the vector multiplication and addition structure in the encrypted domain.

[0094] The privacy-preserving module needs to complete comparative learning and quality assessment without exposing the gradient of the intent model and the plaintext of the metrics. This invention selects a complex approximate homomorphic encryption scheme as the encryption tool and optimizes the encoding and hardware co-designs it to suit the characteristics of event vectors. The complex approximate homomorphic encryption scheme is a typical numerical homomorphic encryption scheme. It defines encryption operators on a polynomial ring and achieves semantic security by introducing Gaussian noise into the modulo polynomial coefficients. Its outstanding advantages are that it can maintain the closure of addition and multiplication in the ciphertext domain and supports simultaneous encoding of real and imaginary numbers, making it suitable for gradient update scenarios that include magnitude and direction information.

[0095] The gradient vector is obtained through backpropagation of the intention model. Each element is in complex form. The complex approximate homomorphic encryption scheme first examines the real part... With the imaginary part Scaling is performed, and then the coefficients are embedded in the polynomial:

[0096] This represents a plaintext polynomial, with coefficients ordered according to the gradient vector index. During encoding, this invention ensures that the real and imaginary parts of the same gradient element are packaged within the same polynomial coefficients, avoiding the double storage overhead of traditional complex number slotting schemes. Encryption operation:

[0097] in It is a random secret polynomial. The noise polynomial is designed to satisfy a normal distribution. The ciphertext is stored in three integer polynomial slots, compatible with standard approximate homomorphic encryption.

[0098] Comparative learning in the cryptographic domain requires vector weighting and scaling of the gradient. This invention utilizes modular multiplication of polynomials for multiplication and modular addition for addition, maintaining noise levels under control after each multiplication through rescaling. If the gradient is multiplied by a scalar... This can be achieved simply by scaling the corresponding polynomial coefficients within the plaintext scalar domain and then performing a single-instruction modular multiplication on the ciphertext. Vector weighting operations are mapped to polynomial addition, with complexity linearly related to the gradient dimension. Since all operations are performed with complex coefficients, addition and multiplication do not disrupt the correspondence between real and imaginary parts, resulting in a complex gradient after decryption. This ensures that the intent quality control module can directly update the convolution kernel phase weights using the complex gradient after decryption, without requiring unnecessary data rearrangement.

[0099] In terms of security, the noise selected in this invention follows a discrete Gaussian distribution, satisfying the discrete cosine lattice difficulty assumption. Even if an attacker gains an infinite number of homomorphic operations, they cannot deduce the gradient plaintext from the ciphertext. To further defend against side-channel attacks, this invention incorporates a random rotation mask on the hardware accelerator, randomly rotating the input coefficients before each round of polynomial multiplication and rotating them back after the multiplication is complete, thereby disrupting the correlation between power consumption and electromagnetic side-channel attacks.

[0100] The main bottleneck in homomorphic operations lies in polynomial multiplication. The module integrates a matrix multiplication accumulation acceleration unit, employing a hybrid strategy based on coefficient-domain block multiplication and number-domain Nyquist sampling theory to convert polynomial multiplication into a two-dimensional fast Fourier transform, which is then computed locally in on-chip static random access memory. Hardware measurements show a latency of 0.9ms for polynomial multiplication with a length of 4096, meeting the real-time processing requirements of 30 frames per second for edge nodes.

[0101] In terms of quality assessment, this invention will use the first quality score Second mass fraction Verification confidence with zero-knowledge proof The mixture is then fused to generate a third mass fraction. . formula:

[0102] in , , The weights are constants, and their sum is 1. Each weight is determined offline through a grid search of historical data. Because... and Since the event vector is already present, the fusion operation can be performed in the plaintext domain, thus eliminating the need for additional homomorphic computation overhead. The evidence entry includes the event hash value. Third quality score Zero-knowledge proof With encryption gradient After an entry is written to the local evidence repository, the read-only mapping area exposes a query interface. External auditors can check the existence of the event and verify the quality score range without decryption, while the hash chain ensures that the writing order cannot be tampered with.

[0103] Experimental verification: This module was deployed for 7 days in an underground parking garage where camera illumination frequently changes, recording 12,057 event vectors. Compared to the unencrypted system that directly transmits gradients, the complex approximate homomorphic encryption scheme increases latency by 8ms, but reduces the average false alarm rate by 3% and the false alarm rate by 5%, mainly due to the online adaptive learning maintained by the encryption domain contrastive learning. Penetration attack tests showed that even after inserting side-channel probes on edge nodes for 48 hours of monitoring, the gradient plaintext could not be recovered from the power consumption trajectory, verifying the effectiveness of the random rotation mask.

[0104] Example 10: Monitoring employee behavior near the conveyor belt in a logistics sorting center. During peak cargo periods, the distribution of behavioral samples shifted significantly. Encrypted domain contrastive learning completed a model weight fine-tuning within 5 minutes, stabilizing the average third quality score at 0.78. Under the same settings, the static model with contrastive learning disabled dropped its third quality score to 0.52, causing frequent increases in the risk gating module threshold and missing two abnormal handling events.

[0105] Preferably, the privacy proof module uses an interval zero-knowledge proof protocol based on elliptic curve commitment when generating zero-knowledge proofs. The interval zero-knowledge proof protocol proves that both the third quality score and the gradient of the intent model are within a preset trust range without revealing the plaintext.

[0106] The privacy evidence module needs to prove to external verifiers that two pieces of privacy data are within the safe range: the first being the third quality score. The second term is the gradient vector of the intention model. Directly disclosing these two data points would reveal the internal scoring criteria and model parameters of the inspection system; completely concealing them would prevent operators or regulators from verifying the credibility of the system's output. To resolve this contradiction, this invention introduces an interval-based zero-knowledge proof protocol based on elliptic curve commitments. This protocol proves to any verifier without revealing the plaintext. and Each component falls within a preset range, thus meeting the dual requirements of regulatory verifiability and privacy security.

[0107] Elliptic curve commitments belong to the polynomial commitment family. Let the base point of the curve be... The auxiliary base point is The dark random number is For a single real number The process of making a commitment is as follows:

[0108] in Indicates the point of commitment to the outcome. Values ​​to be hidden This is a random number, used to prevent the same commitment from occurring in different submissions with the same value. Due to the elliptic curve discrete logarithm problem, an attacker, unaware of... Unable to from Infer This invention will use the third mass fraction With gradient vector Each component Each makes an independent commitment, resulting in a set of commitments. The interval proof must be performed without revealing the information. Prove in the case of Located in the interval or This invention employs an improved proof scheme for the range of discrete logarithms: first, the real number is amplified. The integer part is multiplied and converted to an integer representation; then, a binary expansion is performed. Each bit is treated as a Boolean variable. The prover demonstrates, using a product-constrained circuit, that multiplying all bits by their corresponding weights reconstructs the original integer value, with all bits being either 0 or 1. This circuit is constructed within the elliptic curve commitment domain, and the verifier only needs to linearly traverse all bit commitments to calculate whether the product relationship holds, resulting in a time complexity of O(n log n). Since the gradient dimension of this invention does not exceed 64, the upper bound of the interval... Pick This proves that the size is still controlled at approximately 950 bytes.

[0109] To improve generation speed, the module implements an elliptic curve dot product pipeline in hardware. Curve 25519 is chosen, its prime order facilitating fast modular multiplication at the edge hardware. The pipeline depth is 8, performing 8 scalar multiplications per cycle, generating a gradient vector range proof in approximately 2.3ms. The third quality score commitment and proof takes 0.4ms and can be parallelized with the gradient proof, achieving an overall latency that meets the real-time requirement of 30 frames per second.

[0110] During system runtime, the proof process is as follows: 1. The privacy evidence storage module receives the event vector and calculates the third quality score. And retrieve gradient ciphertext .

[0111] 2. Obtain the plaintext gradient by decrypting in a trusted execution environment. .

[0112] 3. and After magnification and rounding, elliptic curve commitment sets are generated respectively.

[0113] 4. Construct a binary expansion circuit and use Boolean constraints to generate zero-knowledge proofs. .

[0114] 5. Output evidence entries Write to the local evidence repository.

[0115] The verification process only requires reading the stored entries, reconstructing the auxiliary base point table, and verifying Boolean constraints; it does not require access to any plaintext or secret random numbers. If the proof passes, it can be assured that the third mass score is within the range specified in the original text. And the gradient per component in Within this mechanism, the system output is guaranteed to be unaltered and free from gradient explosion risk. If the proof fails, the risk gating module automatically marks the event as invalid and refuses to include it in the model update, preventing erroneous gradients from contaminating the weights.

[0116] Example 11: Deploying inspection robots in a chemical warehouse, with a third quality score safety threshold of 0.7 and an upper bound for the gradient. 2 14 The robot ran continuously for 48 hours, generating 42,517 zero-knowledge proofs, all of which passed verification. The penetration testing team attempted to inject over-threshold noise into the gradient, causing at least one component to exceed a certain threshold. The proof immediately fails, and the event is intercepted. In contrast, when a system without interval proof is subjected to the same injection attack, the risk gating module lowers the threshold due to the trust gradient, generating three false alarms; this invention produces neither false alarms nor missed alarms.

[0117] Another example was tested in a public library nighttime unattended scenario. After the lights automatically turned off, the event camera noise increased, and the gradient components of the model suddenly increased, but still remained within acceptable limits. Within the specified range, the interval proof passed, and the system continued normal learning. The control group, which did not use interval proof, had to disable nighttime model updates to avoid noise, resulting in a 7% increase in the actual false negative rate.

[0118] like Figure 3 As shown, the risk gating module calculates the conditional value at risk based on the loss sequence recorded in the evidence repository. Through reinforcement learning, it generates a dynamic threshold using the third quality score and the intent probability. When the third quality score exceeds the dynamic threshold, it triggers a multi-level response action and writes the execution hash value and inverse gradient into the evidence repository for feedback updates.

[0119] The risk gating module, located at the end of the hierarchical quality control closed loop, maps the perceived quality from the front end and the intent analysis results from the middle layer to the final execution strategy. Its inputs include the third quality score. Intent probability and historical loss sequences from the local evidence repository Each loss It is a linear combination of false positive cost and false negative cost, and the value is normalized to an interval by business rules. The module generates dynamic thresholds through a dual mechanism of tail risk measurement and reinforcement learning, thereby avoiding over-alarms or missed alarms that may result from using fixed thresholds.

[0120] First, tail risk is calculated based on the conditional value at risk (CFT) indicator. Assume the loss sequence is arranged in ascending order of time, with quantiles... Take 0.95. First, calculate the value at risk:

[0121] in Let represent the empirical distribution function. Then, calculate the conditional risk value:

[0122] Expresses the mathematical expectation. Updated daily via a sliding window. The system can detect changes in tail loss in real time. For example, during holidays when passenger flow surges, the costs of false alarms and false alarms increase simultaneously. This raises the threshold, thereby increasing it.

[0123] For dynamic mapping and Upon reaching the risk threshold, the module introduces a deep deterministic policy gradient algorithm. The state vector is set as follows: The action is defined as the threshold adjustment amount. Continuous range of values The policy network (Actor) uses a two-layer fully connected structure to output... The Critic network evaluates the long-term rewards of state-action pairs. The environmental reward function is designed as follows:

[0124] in The current threshold, This is used as a balancing factor. This design encourages the policy network to simultaneously reduce loss and keep the threshold close to the tail risk level, avoiding over-cautiousness or over-relaxation. The policy network updates every 10 events, and the target network uses an exponential moving average to ensure stable learning without excessively chasing short-term fluctuations.

[0125] When the new threshold After generation, it is compared with the real-time third quality score. If... The module immediately triggers a multi-level response action. The response actions are divided into four levels based on risk level: client interface prompt, mobile inspection robot verification, access control relay interlock, and sending an alarm signal to the central security platform. An execution hash value is generated after the action is executed. Simultaneously, the inverse gradient is calculated based on the reward function. ,Will The data is encrypted using a complex approximate homomorphic method and written to a local evidence repository. This information is then used by the intent quality control module to update the convolutional weights in the next cycle using a particle swarm optimization algorithm. In this way, risk preference information can be backpropagated to the feature extraction layer, forming a three-way closed loop of data-feature-decision.

[0126] To ensure real-time performance, the module implements streaming. Estimator. The loss sequence is fed into a fixed-length circular buffer, which is stored in main memory using a bucket-counting structure. The update complexity is O(n log n). Policy network inference takes approximately 0.4ms on an edge graphics processor; the maximum execution time occurs in mobile robot verification scenarios, at approximately 80ms; the complete closed-loop latency is controlled within 120ms, meeting the real-time alarm requirements in densely populated environments.

[0127] Example 12: A risk gating module was deployed at the hospital emergency entrance for 30 days, processing a total of 62,804 event vectors. The system's automatically learned threshold curve exhibited a rhythm of rising at night and falling during the day, consistent with the risk preferences of insufficient night shift security personnel. Compared to a fixed threshold scheme, the false negative rate decreased from 2.9% to 0.7%, and the false positive rate decreased from 6.1% to 3.8%.

[0128] Example 13: High-frequency nighttime pressure testing in a hazardous chemical warehouse, using script injection to trigger abnormal events. The system continues to fluctuate at a high level. The risk gating module rapidly raises the threshold and activates the access control relay interlock to prevent personnel from entering. After the injection stops, the threshold gradually decreases within 15 minutes. Experiments show that the reinforcement learning strategy can adapt to both steady-state and burst states, avoiding "tug-of-war" false alarms.

[0129] Risk Evolution Experiment: 5% of high-loss samples in the evidence repository were randomly deleted to simulate malicious tampering. Due to the chained dependency of hash values, missing samples led to… The value is too low. When the verification end detects a broken chain during snapshot validation of the hash chain, it immediately rejects the use of the contaminated threshold and triggers a system self-check. This indicates that the hash value and reverse gradient writing mechanism can prevent data tampering.

[0130] Preferably, the risk gating module adopts a deep deterministic policy gradient algorithm, using the conditional risk value as the reward signal and the third quality score and intention probability as state variables to update policy parameters in real time and generate dynamic thresholds.

[0131] The core task of the risk gating module is to assign a response level that matches the risk level to each event vector in a continuously changing inspection environment, thereby minimizing the combined costs of false positives and false negatives. To this end, the module uses a deep deterministic policy gradient algorithm, which maps the third quality score and intent probability to a dynamic threshold within a reinforcement learning framework; and uses conditional value at risk as a reward signal to guide the policy to optimize in the direction of reducing tail losses.

[0132] Value at Risk (VaR) is a commonly used metric for measuring tail risk. (The last sentence appears to be incomplete and possibly refers to a separate topic.) The loss is denoted as a sequence. The loss is calculated using a cost model of the response action. Given quantiles... Value at risk is defined as:

[0133] Conditional Value at Risk is defined as:

[0134] Represents the empirical distribution function. Expressing expectations, In this invention, it is fixed at 0.95. Because... Focusing only on the worst 5% of results can reflect the extreme loss level of the system and is suitable as an external feedback for risk gating.

[0135] The Deep Deterministic Policy Gradient Algorithm (DPRQA) belongs to the continuous action reinforcement learning method and consists of a policy network (Actor) and a value network (Critic). The state vector is set as follows:

[0136] in As the third mass fraction, The probabilities of intent and their respective ranges are as follows: .action Defined as threshold adjustment amount, with a continuous range of values. Policy network input Output Then, a new threshold is obtained through linear transformation:

[0137] is the scaling factor used to limit the threshold step size. The value network estimates the long-term reward of state-action pairs. To encourage policies to reduce tail risk, the reward function is designed as follows:

[0138] in For the current losses, This is a balancing factor. The first term encourages the strategy to reduce immediate losses, while the second term helps the threshold to approach... The parameters of the policy network and value network are updated using mean squared error and gradient descent; the target network uses exponential moving average to improve stability. Under real-time requirements, the loss sequence needs to be maintained quickly. This invention implements a streaming quantile estimator, inserting the loss value into an ordered bucket and updating the mean of the tail samples, with a complexity of [missing information - likely a time complexity]. Policy network inference takes an average of 0.4ms on an edge graphics processor, and value network backpropagation takes 0.7ms, which can meet the processing requirements of hundreds of frames per second.

[0139] Once the new threshold is generated, it is compared with the third quality score. If The system triggers a multi-level response. The multi-level response consists of four levels: Level 1 provides a notification to the client; Level 2 dispatches a mobile inspection robot for verification; Level 3 triggers the access control relay interlock; and Level 4 sends an alarm signal to the security center. An execution hash value is generated after the response is completed. And calculate the inverse gradient:

[0140] in This represents the weight vector of the policy network. After inverse gradient encryption, it is written to the evidence repository to provide additional supervision signals for particle swarm mutual information optimization in the intent quality control module, thus achieving cross-layer closed loop.

[0141] To prevent the strategy from excessively raising the threshold after extreme events, this invention introduces a mean squared advance penalty. If continuous... For frame height loss, the system stores a weight snapshot and limits the maximum threshold increase to no more than 20% in the next step. The limit is gradually released after the loss is recovered to prevent threshold oscillation.

[0142] Example 14: Deployed at the airport departure terminal for 30 days, with a total sample size of 128,301. The system automatically learned the daily threshold pattern consistent with passenger flow, with an average threshold of 0.41 during peak hours and 0.28 during off-peak hours. The false negative rate was 0.8%, and the false positive rate was 4.2%, representing a 2.1 percentage point reduction in false negatives and a 1.3 percentage point reduction in false positives compared to the fixed threshold scheme.

[0143] Example 15: A dynamic occlusion test was conducted in a critical area of ​​a chemical plant, where partial obstruction of the camera caused event loss. The loss increased sharply after the occlusion occurred. The threshold is increased by 0.12, and the strategy raises the threshold by 0.11 within 3 steps and triggers a mobile robot for review; after the occlusion is removed, the threshold drops back within 20 steps, and the system's false alarm rate remains at 0.6%.

[0144] Penetration testing: Attackers attempt to insert fake high-quality score events to lower the threshold. The system monitors the third quality score and... The deviation detected an anomaly and refused to update the strategy; within the 24-hour attack window, the maximum change in the threshold was only 0.02, proving that the reward function design can resist data poisoning.

[0145] Preferably, when the third quality score exceeds the dynamic threshold, the risk gating module sequentially triggers multi-level response actions. The multi-level response actions include client prompting, mobile inspection robot verification, access control relay interlocking, and sending an alarm signal to the security center. After triggering each level of response action, the module generates a corresponding execution hash value and reverse gradient, writes them to the local evidence storage library for feedback updates.

[0146] The risk gating module is located at the final decision-making level of the hierarchical quality control closed loop, and is responsible for gating the third quality score. With dynamic threshold The real-time comparison results are transformed into executable multi-level response actions, and the execution hash value and inverse gradient of each response are written into a local evidence repository, forming a closed-loop feedback for the entire system. The design goal of this module is to be able to promptly trigger corresponding on-site actions under different risk levels, and to apply the action results inversely to the model weights and threshold strategies, enabling the system to have self-evolution capabilities. The following explanation covers four aspects: response chain principle, inverse gradient generation, evidence writing mechanism, and implementation effect.

[0147] Based on the response chain principle, the module first reads the third quality score calculated by the intent quality control process. Dynamic threshold output by the policy network .like If the system determines the risk is within an acceptable range, only regular logs will be recorded; otherwise... The module then triggers four response actions sequentially according to risk level. These four actions are designed to be progressive: 1. Client prompts will alert security personnel to any unusual activity via on-site display screens and pop-up windows on duty terminals. This level is only triggered for localized, tentative anomalies or early signs of anomalies.

[0148] 2. The mobile inspection robot's verification command is pushed to the inspection robot control unit via a low-latency local area network. Guided by the shortest path navigation algorithm, the robot moves to the area where the incident occurred, calls its own camera and depth camera for secondary confirmation, and sends back the verification result.

[0149] 3. The access control relay interlock command is sent to the access control controller via the safety fieldbus, and all access control or turnstiles in the corresponding section are immediately closed to prevent the target from moving further.

[0150] 4. An alarm signal is sent to the security center and simultaneously transmitted to the duty room and cloud monitoring platform via a dual-redundant fiber optic link, subsequently triggering an audible and visual alarm, real-time personnel broadcast, and video pop-up report. The design principle of the four-level action is: first, human-machine prompting; second, mobile verification; third, physical isolation; and finally, manual intervention. This ensures timeliness while avoiding frequent triggering of high-cost handling procedures.

[0151] The module performs hash value generation and inverse gradient generation. After each response action is executed, it calculates a hash value by combining the action parameters and the execution timestamp for subsequent auditing and retrieval. The hash function adopts the same sponge-type design as the event hash, with an output length of 256 bits to ensure uniqueness. The core criterion for triggering an action is... The present invention maps the difference in magnitude as a feedback gradient:

[0152] in This is a scaling factor, which is set to a constant in the hardware security unit to prevent gradient explosion. This represents the inverse gradient. This gradient is then encrypted using a complex approximate homomorphic encryption method to generate the ciphertext. The data is written into the evidence repository and used in the next cycle as an additional supervisory signal in conjunction with the particle swarm mutual information optimization algorithm of the intent quality control module, enabling the model weights to be fine-tuned in the direction of reducing false positives and false negatives.

[0153] The evidence writing mechanism will be implemented by the module. As an evidence entry, it is written to the local evidence repository, where To perform hashing, To encrypt the inverse gradient, In response to the level identifier, the evidence repository employs an add-only, segmented file system structure. Each entry includes a double-chain hash referencing the previous entry, forming a sequentially immutable chained log. The security center can periodically download read-only snapshots to verify the integrity of the hash chain, completing compliance audits without accessing any encrypted plaintext.

[0154] Example 16: The system of this invention was deployed in a commercial complex for 14 days, processing a total of 18,752 event vectors. The system triggered client prompts 241 times, robot verification 27 times, access control interlocks 6 times, and security alarms 2 times. All alarms were verified by manual inspection and found to be valid, with no false alarms. Compared with traditional threshold locking schemes, the false alarm rate decreased from 5.3% to 0.4%, and the missed alarm rate decreased from 1.9% to 0.

[0155] Example 17: During nighttime duty at a university laboratory building, an 8-hour continuous light flicker interference test was conducted, resulting in a short-term increase of 0.18 in the third quality score. The risk gating module dynamically increased the threshold by 0.15 within 3 seconds based on the policy network, triggering only a client-side notification and preventing any mis-scheduling of robots; the threshold adaptively decreased after the interference ended. The baseline fixed threshold scheme, unable to adjust the threshold in a timely manner, triggered 11 access control interlocks, affecting the normal access of laboratory personnel.

[0156] Example 18: Simulating depth errors caused by light blocking during peak hours in the logistics warehousing area, resulting in... Low. The risk gating module identified this. Higher than Advanced response was not triggered, but tail loss increased. The value increases by 0.07. After being reflected in the policy network reward, the threshold is slightly adjusted and increased by 0.05 within 5 minutes, causing the system to converge again and avoiding the underreporting of potential security risks.

[0157] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An intelligent inspection and management system based on hierarchical quality control, characterized in that, include: The data acquisition and preprocessing module acquires multimodal sensor data and writes it to a timestamp, reconstructs the background and calculates the event disturbance intensity, generates the first mass score based on infrared texture consistency and millimeter-wave micro-Doppler features, and outputs the disturbance data packet. The intent quality control module performs waveform transformation on the disturbed data packets, constructs a pulse map by combining the time-of-flight depth frame and millimeter-wave radar point cloud, generates entity latent representations through convolutional inference and particle swarm mutual information optimization, obtains intent probability through linear discrimination and forms a second quality score, and outputs event vectors. The privacy evidence storage module calculates the hash value of the event vector and generates a zero-knowledge proof. After homomorphically encrypting the gradient, it compares and learns in the encrypted domain to update the intent model. It merges the first quality score and the second quality score to generate a third quality score, and writes the event hash value, the third quality score, the zero-knowledge proof and the encryption gradient into the local evidence storage library. The risk gating module calculates the conditional value at risk based on the loss sequence recorded in the evidence repository. Through reinforcement learning, it generates a dynamic threshold using a third quality score and intent probability. When the third quality score exceeds the dynamic threshold, it triggers a multi-level response action and writes the execution hash value and inverse gradient into the evidence repository for feedback updates.

2. The system according to claim 1, characterized in that, The multimodal sensors in the acquisition and preprocessing module include at least an event camera and a millimeter-wave radar. The event camera is used to output time-resolved event streams, and the millimeter-wave radar is used to output target distance and velocity information.

3. The system according to claim 1, characterized in that, The intention quality control module performs a complex wavelet transform on the perturbation data packet. The complex wavelet transform extracts phase information and amplitude information to characterize the direction and energy changes of the micro-motion.

4. The system according to claim 3, characterized in that, The intent quality control module maps the texture features obtained from complex wavelet transform, the spatial features of the time-of-flight depth frame, and the velocity features of the millimeter-wave radar point cloud into pulse map nodes, where texture nodes, depth nodes, and velocity nodes are connected through distance weights.

5. The system according to claim 1, characterized in that, When performing convolutional inference on the pulse graph, the intention quality control module uses a pulse graph neural network implemented with silicon photonics. The silicon photonic interference structure completes the weight multiplication and accumulation operations and outputs the convolution result in the form of pulses.

6. The system according to claim 1, characterized in that, The intent quality control module uses the particle swarm mutual information optimization algorithm to update the weights of the entity latent representation. The particle swarm mutual information optimization algorithm takes the classification loss and the weighted sum of mutual information as the objective function and updates the particle velocity and position in parallel within a fixed number of iterations.

7. The system according to claim 1, characterized in that, The privacy evidence storage module performs homomorphic encryption on the gradient and metrics of the intent model using a complex approximate homomorphic encryption scheme. This scheme simultaneously encodes both the real and imaginary parts to maintain the vector multiplication and addition structure within the encrypted domain.

8. The system according to claim 1, characterized in that, When generating zero-knowledge proofs, the privacy preservation module uses an interval zero-knowledge proof protocol based on elliptic curve commitment. The interval zero-knowledge proof protocol proves that both the third quality score and the gradient of the intent model are within a preset trust range without revealing the plaintext.

9. The system according to claim 1, characterized in that, The risk gating module employs a deep deterministic policy gradient algorithm, using conditional risk value as the reward signal and third quality score and intention probability as state variables to update policy parameters in real time and generate dynamic thresholds.

10. The system according to claim 1, characterized in that, When the third quality score exceeds the dynamic threshold, the risk gating module triggers multi-level response actions in sequence. These multi-level response actions include client prompting, mobile inspection robot verification, access control relay interlocking, and sending alarm signals to the security center. After triggering each level of response action, the module generates a corresponding execution hash value and reverse gradient, which are written to the local evidence storage library for feedback updates.

Images