A cloud edge terminal collaborative lightweight large model defect detection system

Abstract

The application belongs to the technical field of lightweight large models, and relates to a cloud edge end cooperative lightweight large model defect detection system, which comprises a feature generation module, an intermediate state feature tensor array is obtained; a state judgment module, a feature state boundary is obtained; an abnormality extraction module, a non-zero effective feature array and a non-zero position index mask are generated; a key packaging module, encrypted feature data flow is generated and sent to a cloud end processing center; a cloud end wake-up module, a frequency domain receptive field operator window in an open state is triggered; a gradient generation module, a target gradient correction tensor is generated; and an edge evolution module, local adaptive update of frequency domain feature gain mapping is realized. The application solves the problems of how to reduce the redundant calculation overhead of the edge end on massive normal background data and how to realize low-delay targeted cloud edge cooperative update of abnormal features.

Description

Technical Field

[0001] This invention belongs to the technical field of lightweight large models and relates to a cloud-edge-device collaborative lightweight large model defect detection system. Background Technology

[0002] In industrial settings such as semiconductor manufacturing and precision optical component production, machine vision defect detection systems are crucial for ensuring yield rates. As production lines evolve towards higher speeds and greater integration, the streams of image data to be inspected are becoming increasingly massive and high-dimensional, with defect morphologies becoming increasingly subtle and diverse. This presents challenges to existing detection technologies. The current pain point lies in accurately identifying subtle anomalies or novel defects while meeting the millisecond-level real-time response time of production lines.

[0003] Chinese invention patent application CN120930709A discloses a lightweight deployment method for large AI models in complex scenarios. This method obtains lightweight network parameters by performing structured pruning and dynamic mixed-precision quantization on a pre-trained network and deploys it on an edge device. Simultaneously, when abnormal features are detected in the edge inference results, gradient update instructions are generated by calculating the difference in probability distribution between the edge device and the cloud-based teacher model output, thereby controlling the edge device to synchronously update the network parameters.

[0004] However, the aforementioned existing technologies still have limitations when dealing with high-throughput industrial image data streams. At the edge, this approach still requires software-level invocation of neural networks to perform complete forward inference on the entire input data. This results in a significant consumption of edge computing resources by massive amounts of defect-free, normal background data. Furthermore, its model's dynamic evolution relies on anomaly probability distribution comparison and overall network gradient synchronization at the software level. This interaction method easily leads to high bandwidth consumption and processing latency when transmitting complex anomaly features, making it difficult to achieve low-overhead filtering of massive amounts of normal samples and low-latency, rapid adaptation of the underlying detection feature library under conditions of limited computing power at the edge.

[0005] Therefore, how to reduce the redundant computational overhead of the edge on massive normal background data and achieve low-latency targeted cloud-edge collaborative updates of abnormal features is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a lightweight, large-model defect detection system that integrates cloud, edge, and device collaboration.

[0007] A lightweight, large-model defect detection system with cloud-edge-device collaboration includes: The feature generation module performs frequency domain transformation on the acquired original pixel matrix signal through the edge computing unit to obtain an intermediate feature tensor array; The state determination module calculates the spatial non-zero density of the intermediate state feature tensor array and compares the spatial non-zero density with a preset static threshold to obtain the feature state boundary. The anomaly extraction module, in response to the spatial non-zero density exceeding a preset static threshold, extracts the non-zero elements and their coordinates in the intermediate state feature tensor array, and generates a non-zero valid feature array and a non-zero position index mask. The key encapsulation module merges and encapsulates the non-zero valid feature array and the non-zero position index mask to generate an encrypted feature data stream and send it to the cloud processing center. The cloud wake-up module parses the received encrypted feature data stream to obtain the non-zero position index mask and maps it to the physical memory address pool of the cloud large model in order to trigger the frequency domain receptive field operator window in the open state. The gradient generation module imports the non-zero valid feature array into the frequency domain receptive field operator window and generates the target gradient correction tensor through residual back derivation based on semantic features. The edge evolution module sends the target gradient correction tensor to the edge computing unit, and realizes local adaptive update of frequency domain feature gain mapping by overwriting the register gain weight state in the corresponding physical register.

[0008] A further aspect of the present invention involves performing frequency domain transformation on the acquired original pixel matrix signal using an edge computing unit to obtain an intermediate state feature tensor array, comprising the following steps: Obtain continuous raw pixel matrix signals and allocate the physical memory area of ​​the edge computing unit as physical cache space; The sensor network of the edge computing unit performs a two-dimensional discrete Fourier transform on the original pixel matrix signal to separate the amplitude and phase components and obtain the frequency domain feature flow. The frequency domain feature stream is written into the physical buffer space in time order to obtain an intermediate feature tensor array containing time, channel, vertical spatial frequency and horizontal spatial frequency dimensions.

[0009] A further aspect of this invention involves performing a two-dimensional discrete Fourier transform on the original pixel matrix signal through the perception network of the edge computing unit to separate the amplitude and phase components, thereby obtaining a frequency domain feature flow. This includes the following steps: Extract pixel intensity values ​​from the original pixel matrix signal of a single frame; Based on spatial coordinates and frequency domain coordinates, frequency domain mapping calculation is performed on pixel intensity values ​​to obtain frequency domain complex values ​​containing amplitude and phase information; The amplitude spectrum and phase spectrum are extracted from the complex values ​​in the frequency domain and combined to obtain the frequency domain feature flow.

[0010] A further aspect of this invention involves calculating the spatial non-zero density of the intermediate state feature tensor array and comparing it with a preset static threshold to obtain the feature state boundary, including the following steps: Establish a monitoring link between the static random access memory array within the physical cache space and the activation value addressing monitor; The frequency of non-zero addressing activations with an absolute value greater than a preset background noise shielding threshold is counted when accessing the intermediate state feature tensor array using the activation value addressing monitor. The spatial non-zero density is calculated based on the non-zero addressing activation frequency and the total number of elements in the intermediate state feature tensor array. In response to a spatial non-zero density greater than a preset static threshold, the internal state register is triggered to flip, generating a characteristic state boundary.

[0011] A further aspect of the present invention, in response to a spatial non-zero density exceeding a preset static threshold, extracts non-zero elements and their coordinates from an intermediate-state feature tensor array to generate a non-zero valid feature array and a non-zero position index mask, comprising the following steps: In response to a spatial non-zero density exceeding a preset static threshold, the current main task is suspended, and a default safety instruction signal is output to the external control interface. Traverse the intermediate state feature tensor array, extract the real values ​​of effective amplitude elements and phase elements whose absolute values ​​are greater than the preset background noise shielding threshold, and combine them into a non-zero effective feature array in the form of a one-dimensional array. The multidimensional coordinate index values ​​of the non-zero valid feature array in the intermediate state feature tensor array are organized into a two-dimensional integer matrix, which serves as a non-zero position index mask containing multidimensional coordinate topological information.

[0012] A further aspect of this invention involves merging and encapsulating the non-zero valid feature array and the non-zero position index mask to generate an encrypted feature data stream and send it to a cloud processing center, including the following steps: The non-zero valid feature array is associated and arranged with the non-zero position index mask to obtain the associated arrangement data containing coordinates and real values. Based on the media access control address, current timestamp, and data packet sequence number of the edge computing unit, perform encrypted digest extraction operation to generate a header fixed-length hash checksum; The first fixed-length hash check code is inserted into the header register bit of the associated permutation data, encrypted with a pre-shared key, and the encrypted feature data stream is generated and sent to the cloud processing center.

[0013] A further aspect of this invention involves parsing the received encrypted feature data stream to obtain a non-zero position index mask and mapping it to the physical memory address pool of a large cloud model to trigger an active frequency domain receptive field operator window, comprising the following steps: Receive encrypted feature data streams, block the full image inference task input into the cloud-based large model, and maintain the dormant state of inactive neurons in the cloud-based large model; The encrypted feature data stream is decrypted and the header fixed-length hash check code is verified, and the non-zero position index mask is parsed out. The coordinate sequence in the non-zero position index mask is converted into a pointer to the physical memory address pool of the multi-head self-attention module of the large model in the cloud, and the frequency domain receptive field operator window in the enabled state is triggered at a fixed point.

[0014] A further aspect of this invention involves importing a non-zero valid feature array into the frequency domain receptive field operator window, and generating a target gradient correction tensor through residual back-derivation based on semantic features, including the following steps: The non-zero valid feature array is injected into the corresponding frequency domain receptive field operator window via the data bus; The parameters of the local frequency band attention mechanism of the large cloud model are called to perform weighted matching between the non-zero effective feature array and the adjacent key vectors in the feature space, and to perform residual inverse derivation operation based on semantic features. Extract the error backpropagation features generated by the residual back derivation operation based on semantic features, and generate a target gradient correction tensor with the same dimension and topological structure as the non-zero effective feature array.

[0015] A further aspect of this invention involves sending the target gradient correction tensor to the edge computing unit and overwriting the register gain weight state in the corresponding physical register to achieve local adaptive updating of the frequency domain feature gain mapping, including the following steps: Receive the target gradient correction tensor, extract the header fixed-length hash check code for reverse tracing and localization, and determine the specific underlying hardware inside the edge computing unit that generates the abnormal features. Instructions are sent to the specific underlying hardware to retrieve the original forward amplitude feature real values ​​pre-stored in the buffer queue; The gradient values ​​in the target gradient correction tensor are used to overwrite the register gain weight state in the physical register, thus completing the local adaptive update of the frequency domain feature gain mapping.

[0016] A further aspect of this invention utilizes the gradient values ​​in the target gradient correction tensor to overwrite the register gain weight states within the physical register, thereby completing the local adaptive update of the frequency domain feature gain mapping. This includes the following steps: By combining the preset hardware learning rate and quantization base, the feature gradient values ​​in the target gradient correction tensor are converted into target integer weight correction increments; Based on gradient ascent logic, the target integer weight correction increment is used to update the state of the original coefficient values ​​in the physical register to generate a new control word. The new control word is overwritten into the physical register, thus completing the overwriting of the register gain and weight state.

[0017] In summary, the present invention has the following beneficial technical effects: 1. The original pixel matrix signal is converted into an intermediate feature tensor array by the edge computing unit, and its spatial non-zero density is calculated using an activation value addressing monitor. By comparing it with a preset static threshold, the determination of normal sparse samples is transformed into a hardware-level boundary comparison operation, triggering subsequent processes only when features are abnormally dense. This mechanism effectively filters massive amounts of normal background data and reduces redundant computational overhead when processing high-throughput data at the edge.

[0018] 2. When identifying anomalies, only the non-zero valid feature array and the non-zero position index mask are extracted and transmitted to the cloud, reducing the bandwidth consumption of image transmission. The cloud maps the physical memory address pool based on the non-zero position index mask and triggers the frequency domain receptive field operator window at a fixed point, replacing the conventional forward inference of the entire image. This targets computing power to the abnormal feature region, improving the processing efficiency of cloud resources and the response speed of cloud-edge collaboration.

[0019] 3. A target gradient correction tensor is generated in the cloud, and the edge computing unit receives it and directly overwrites the register gain weight state in the corresponding physical register. This mechanism eliminates the need for software-level global network gradient synchronization or interruption of the main service for firmware flashing. It directly converts cloud analysis results into adjustments to edge hardware parameters, achieving local adaptive updates of frequency domain feature gain mapping, shortening model evolution latency, and improving the system's adaptability to weak and novel defects. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0021] Figure 1 This discloses a schematic diagram of the framework in the embodiments of this application.

[0022] Figure 2 This discloses a flowchart of an embodiment of this application.

[0023] Figure 3 This application discloses a comparison curve of the frequency distribution of normal background data and abnormal physical defects under different amplitude frequency domain characteristics in the embodiments of this application.

[0024] Figure 4 The document discloses semantic residual gradient distribution curves of minute scratches at different Euclidean distances in embodiments of this application. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 - Figure 4 A preferred description of the present invention is provided below.

[0026] See attached document Figure 1 - Figure 2 This invention proposes a lightweight, large-model defect detection system with cloud-edge-device collaboration, comprising the following modules: The feature generation module performs frequency domain transformation on the acquired original pixel matrix signal through the edge computing unit to obtain an intermediate feature tensor array; The state determination module calculates the spatial non-zero density of the intermediate state feature tensor array and compares the spatial non-zero density with a preset static threshold to obtain the feature state boundary. The anomaly extraction module, in response to the spatial non-zero density exceeding a preset static threshold, extracts the non-zero elements and their coordinates in the intermediate state feature tensor array, and generates a non-zero valid feature array and a non-zero position index mask. The key encapsulation module merges and encapsulates the non-zero valid feature array and the non-zero position index mask to generate an encrypted feature data stream and send it to the cloud processing center. The cloud wake-up module parses the received encrypted feature data stream to obtain the non-zero position index mask and maps it to the physical memory address pool of the cloud large model in order to trigger the frequency domain receptive field operator window in the open state. The gradient generation module imports the non-zero valid feature array into the frequency domain receptive field operator window and generates the target gradient correction tensor through residual back derivation based on semantic features. The edge evolution module sends the target gradient correction tensor to the edge computing unit, and realizes local adaptive update of frequency domain feature gain mapping by overwriting the register gain weight state in the corresponding physical register.

[0027] In one embodiment of the present invention, the feature generation module is configured to perform the following steps: Obtain continuous raw pixel matrix signals and allocate the physical memory area of ​​the edge computing unit as physical cache space; The sensor network of the edge computing unit performs a two-dimensional discrete Fourier transform on the original pixel matrix signal to separate the amplitude and phase components and obtain the frequency domain feature flow. The frequency domain feature stream is written into the physical buffer space in time order to obtain an intermediate feature tensor array containing time, channel, vertical spatial frequency and horizontal spatial frequency dimensions.

[0028] The first step, acquiring the original pixel matrix signal and calling the lightweight perception network of the edge computing unit to generate an intermediate feature tensor array containing frequency domain features, is executed collaboratively by the image acquisition device and the edge computing unit deployed on the industrial production line.

[0029] Image acquisition equipment, such as a semiconductor camera, continuously captures surface images of target industrial products, such as wafers or printed circuit boards, on a production line conveyor belt at a preset frame rate, thereby generating a time-series continuous raw pixel matrix signal. The preset frame rate is set based on the ratio of the physical movement speed of the conveyor belt to the field of view of the image acquisition device. This ensures that two adjacent frames have a 5% to 10% overlap in physical space, thus avoiding the missed detection of minor defects at the edge of the field of view. For example, when the conveyor belt speed is 1 m / s and the longitudinal field of view is 0.1 m, the preset frame rate should be configured to be no less than 10 frames / second to 15 frames / second.

[0030] Simultaneously with the start of the signal acquisition task, the central processing unit (CPU) of the edge computing unit sends an instruction to the internal memory controller to allocate a specified size of physical memory region on-chip as a physical cache space. The computational configuration of this specified physical memory region is based on the total number of elements in the intermediate feature tensor to be processed multiplied by the data format used, such as the 4-byte data width of a single-precision floating-point FP32. An additional 15% to 20% of high-speed static random access memory (SRAM) space is reserved as overhead margin for system addressing and interrupt stacks. The physical cache space is a high-speed SRAM region within the edge computing unit chip. The size of the physical cache space is configured according to the dimensions of the feature tensor array to be processed. For example, during processing... When creating an image sequence of pixels, space can be allocated from 16 megabytes to 64 megabytes.

[0031] The lightweight perception network embedded within the edge computing unit begins to process data from the original pixel matrix signal frame by frame. Unlike traditional deep neural networks that consist of a large number of convolutional layers, the lightweight perception network uses dedicated processing circuits to perform two-dimensional fast Fourier transforms and subsequent weighted processing, aiming to complete feature extraction from the spatial domain to the frequency domain with lower computational overhead.

[0032] For any time A single-frame two-dimensional pixel matrix The lightweight perceptron performs a transformation operation, mapping the original spatial domain pixel intensity information to the frequency domain, producing a frequency domain complex matrix. This step involves converting the spatial domain image into frequency domain features, and the core calculation is implemented using a two-dimensional discrete Fourier transform, expressed by the following formula:

[0033] For the sake of simplicity, time subscripts are omitted in the formula. , The input single-frame raw pixel matrix in spatial coordinates Pixel intensity value at; and These are the horizontal and vertical coordinate indices of the original pixel matrix of a single frame, respectively. is the base of the natural logarithm; Pi; and These represent the width and height of the original pixel matrix in a single frame, respectively. For example, to balance computational efficiency with frequency resolution for minute scratches, the size of the pixel matrix in a single frame can be set. and The resolution is 256, which is 2 to the power of 8, meeting the bit-flipping structure requirements of the radix-2 time-domain decimation algorithm. This can stimulate the parallel efficiency of butterfly operations of the Fast Fourier Transform in the underlying edge hardware and avoid the additional computing power overhead caused by zero-padding. Secondly, the frequency domain resolution of 256×256 achieves an engineering balance between covering the high-frequency energy distribution of typical small scratches on the surface of industrial products, which usually occupy 3 to 5 pixels of connected domain, and reducing the limited computing power load on the edge side. and For frequency domain coordinates, and The range of values ​​for are respectively arrive and arrive ; The imaginary unit; It is the transformed coordinates in the frequency domain. The complex value at that point contains both amplitude and phase information.

[0034] The two-dimensional fast Fourier transform processing module further separates the amplitude and phase components from the frequency domain complex matrix, specifically by performing a process on all... The amplitude spectrum is obtained by modulo operation, and the phase spectrum is obtained by argument operation. Before writing the separated features into the physical buffer space, the lightweight perceptron utilizes its internally integrated lightweight learnable gain layer to perform multiplicative weighting adjustment only on the amplitude frequency domain features.

[0035] The lightweight learnable gain layer is a programmable hardware layer consisting of a physical register array, where each frequency domain coordinate... The dynamically updatable gain coefficients are designed to physically amplify or suppress the spatial frequency components in the target coordinate system pointed to by the non-zero position index mask, while the phase features maintain the cyclic continuity of their spatial topology to constitute the final frequency domain features at any given time. This process occurs over time. Continuous execution results in a feature sequence that is the frequency domain feature flow.

[0036] The gain-adjusted two-dimensional amplitude feature and the transparent two-dimensional phase feature matrix generated at each sampling time point are treated as a data pair and written into the physical buffer space in chronological order. Through continuous stacking, a matrix with time, channel, and frequency coordinates is constructed. Frequency coordinates Four-dimensional intermediate state feature tensor array The intermediate state feature tensor array is a four-dimensional data structure. The four dimensions of the four-dimensional data structure are time series, feature channels (including amplitude and phase), horizontal spatial frequency (corresponding to frequency coordinates), and horizontal spatial frequency (corresponding to frequency coordinates). ) and vertical spatial frequency (corresponding frequency coordinates) ).

[0037] As an example, suppose this method is applied to a semiconductor wafer surface defect detection scenario, where a high-speed industrial camera is continuously taking pictures of wafers on an assembly line. At any given moment, the camera captures a single-frame grayscale pixel matrix with a size of 256×256. The pixel values ​​of the single-frame grayscale pixel matrix are normalized to the [0, 1] interval. The image content is a generally clean wafer surface with a faint vertical scratch in the central region. This matrix is ​​fed into the lightweight perceptual network of the edge computing unit, namely the two-dimensional fast Fourier transform processing module. The two-dimensional fast Fourier transform processing module performs the aforementioned operation and generates a complex matrix with a size of 256×256. Since there is an approximately vertical scratch in the original image, the vertical scratch is spatially represented by a high-frequency change in the horizontal direction. Therefore, in the amplitude spectrum of the calculated complex matrix, along the horizontal frequency axis... Significant energy spikes appear in the high-frequency region of the direction.

[0038] In the embodiments, high-frequency changes or high-frequency regions are uniformly defined as the frequency band region in the frequency domain coordinate system generated by the two-dimensional discrete Fourier transform, where the Euclidean distance between the spatial frequency index coordinate point and the center zero frequency, i.e., the DC component coordinate point, is greater than a preset frequency radius threshold. The preset frequency radius threshold is set based on the Nyquist frequency of the single-frame pixel matrix, i.e., 1 / 4 to 1 / 2 of the maximum spatial frequency. For example, for the aforementioned 256×256 pixel matrix, the preset frequency radius threshold can be set to the range where the frequency coordinate distance from the center zero frequency is greater than 64 to 128 Euclidean distances.

[0039] After the module separates the amplitude spectrum matrix and phase spectrum matrix, it flows through a lightweight learnable gain layer to perform frequency band weighting on the amplitude spectrum, and then reconstructs it with the transparent phase spectrum, denoted as... and This extraction and weighting process is repeated over the next 15 time steps to obtain... These 16 pairs of amplitude and phase matrices are written together into the physical cache space to form an intermediate state feature tensor array with dimensions [16, 2, 256, 256], where dimension 16 represents the time series length and dimension 2 represents the amplitude and phase channels.

[0040] In one embodiment of the present invention, the state determination module is configured to perform the following steps: Establish a monitoring link between the static random access memory array within the physical cache space and the activation value addressing monitor; The frequency of non-zero addressing activations with an absolute value greater than a preset background noise shielding threshold is counted when accessing the intermediate state feature tensor array using the activation value addressing monitor. The spatial non-zero density is calculated based on the non-zero addressing activation frequency and the total number of elements in the intermediate state feature tensor array. In response to a spatial non-zero density greater than a preset static threshold, the internal state register is triggered to flip, generating a characteristic state boundary.

[0041] After generating the intermediate-state feature tensor array, an activation value addressing monitor based on physical cache space is started. The activation value addressing monitor is a hardware circuit module used to non-intrusively monitor data activity in the memory area pre-allocated for storing the intermediate-state feature tensor array. The activation value addressing monitor is integrated as a separate hardware logic unit within the storage controller of the edge computing unit.

[0042] The activation value addressing monitor establishes a monitoring link between itself and the physical cache space containing the intermediate state feature tensor array, which is a specific static random access memory (SRAM) array. This monitoring link is implemented through physical connections to the address and data buses of the SRAM array, enabling the activation value addressing monitor to detect all read and write operations on the SRAM array in real time.

[0043] During each clock cycle of forward inference in the lightweight perceptron, the activation value addressing monitor continuously monitors the voltage levels on the data bus as data is written to or read from the static random access memory array. Since the frequency domain phase spectrum represents the offset angle of a sine wave, it exhibits a uniform random non-zero distribution even in the background noise region where there is no valid signal. Directly statistically analyzing the global non-zero features would lead to sparsity failure. Therefore, the activation value addressing monitor is logically configured to filter and monitor only the amplitude channels in the intermediate state feature tensor array, where the amplitude channel index is 0.

[0044] When data in the amplitude channel is written or read, the hardware counter inside the activation value addressing monitor increments only when the absolute value of the transmitted amplitude data, after denoising, exceeds a preset noise floor threshold. This method counts the frequency of non-zero addressing activations, and the corresponding phase characteristics are gated synchronously associated with this non-zero amplitude mask. The noise floor threshold is set based on statistical prior knowledge from the pre-deployment calibration phase. Specifically, a large number of manually verified defect-free normal sample images are pre-collected and subjected to the same Fast Fourier Transform, and the mean and standard deviation of their high-frequency amplitude frequency domain characteristics are statistically analyzed. Then, following the 3σ criterion—mean plus three times the standard deviation—the noise floor threshold is set to ensure that, in actual detection, signals exceeding the noise floor threshold have a probability of over 99.7% that they are genuine physical defects rather than sensor thermal noise. Through extensive engineering measurements, the noise floor threshold typically falls within the normalized range of 0.01 to 0.05.

[0045] See appendix Figure 3 The characteristic amplitudes of massive amounts of normal background data are highly concentrated in the lower range; while the distribution of abnormal physical defects shifts significantly to the right, with peaks appearing in the higher amplitude range. The dashed line in the figure represents the preset background noise shielding threshold (set to 0.04 in the example). When signals exceeding this threshold are intercepted, the system can effectively shield most of the normal background noise, triggering subsequent processing logic for abnormal physical defects.

[0046] In this paper, the non-zero addressing activation frequency refers to the engineering quantity, which quantifies the frequency of occurrence of non-zero effective features greater than the noise floor threshold in the transmitted amplitude data when accessing the intermediate-state feature tensor array. It is a physical proxy indicator for measuring the density of effective feature information. After completing the traversal scan of the entire intermediate-state feature tensor array amplitude channels, the activation value addressing monitor calculates the spatial non-zero density based on the total activation frequency. The spatial non-zero density is the ratio of the non-zero addressing activation frequency to the total number of elements in the controlled amplitude tensor, and is a normalized value between 0 and 1. The formula for calculating the spatial non-zero density is as follows:

[0047] in, That is, the spatial density is non-zero; The non-zero addressing activation frequency accumulated by the addressing monitor during the traversal period is the activation value addressing monitor. The total number of elements contained in the controlled amplitude channel is equal to the product of the dimensions of each dimension of a single channel in the intermediate state feature tensor array.

[0048] The calculated spatial non-zero density is compared with a preset static threshold. Comparison, preset static threshold This is the key judgment threshold. Its value is set based on prior statistical analysis of the sparsity of features of normal and abnormal samples in the current industrial application scenario (such as semiconductor wafer surface defect detection). After Fourier transform, the frequency domain energy of a normal, defect-free wafer surface is highly concentrated near zero frequency, and the amplitude in the high-frequency region is lower than the noise floor shielding threshold. Therefore, a relatively sensitive noise floor shielding threshold can be set, for example, in the range of 0.01 to 0.05, to ensure that a response is triggered for weak abnormal features. If the spatial non-zero density exceeds the preset static threshold, the activation value addressing monitor sets the value of an internal status register to 1. This action completes the judgment of the feature state boundary.

[0049] As an example, consider the intermediate state feature tensor array from the previous step. Since the activation value addressing monitor only counts the amplitude channels, the total number of controlled elements is... for The activation value addressing monitor was established. Due to a slight vertical scratch on the surface, a cluster of effective amplitude characteristic values ​​exceeding the background noise shielding threshold appeared in the corresponding high-frequency horizontal region. After traversal and statistical analysis, the hardware counter recorded the non-zero addressing activation frequencies. The number of times was 94372. The spatial non-zero density was calculated. In this example, a static threshold is preset. It is set to 0.05. Since 0.09 is greater than 0.05, the internal status register is triggered, and the output characteristic state boundary signal is toggled from logic 0 to logic 1.

[0050] In one embodiment of the present invention, the anomaly extraction module is configured to perform the following steps: In response to a spatial non-zero density exceeding a preset static threshold, the current main task is suspended, and a default safety instruction signal is output to the external control interface. Traverse the intermediate state feature tensor array, extract the real values ​​of effective amplitude elements and phase elements whose absolute values ​​are greater than the preset background noise shielding threshold, and combine them into a non-zero effective feature array in the form of a one-dimensional array. The multidimensional coordinate index values ​​of the non-zero valid feature array in the intermediate state feature tensor array are organized into a two-dimensional integer matrix, which serves as a non-zero position index mask containing multidimensional coordinate topological information.

[0051] In response to the boundary signal of the characteristic state toggling from logic 0 to logic 1, the boundary signal is sent to the central processing unit (CPU) of the edge computing unit as an edge hardware forced interrupt instruction. The CPU suspends the current main task and executes the interrupt service routine. The interrupt service routine sends a clock disable signal to the control logic gates of the lightweight sensing network, pausing subsequent processing. At the same time, it sends a preset default safety instruction signal, such as a 100ms high-level pulse, to the general-purpose input / output interface connected to the external programmable logic controller (PLC), instructing the production line to maintain normal operation to ensure production continuity.

[0052] The central processing unit initiates a data stripping process to scan the intermediate-state feature tensor array stored in the physical cache space, stripping away invalid background dimensions. If and only if the absolute value of an amplitude channel element is greater than the background noise masking threshold, the valid amplitude element and its corresponding phase element with the same spatial topological coordinates are extracted in pairs and copied into a one-dimensional floating-point array to form a non-zero valid feature array, thus eliminating a large amount of useless background.

[0053] To prevent data from being overwritten by newly generated continuous image pipelines while edge computing units are waiting for cloud feedback, the extracted non-zero valid feature array is stored in a circular historical feature buffer queue. Simultaneously, to support chain rule operations in subsequent backpropagation, the processor, based on the extracted valid amplitude elements, synchronously checks the physical register gain state of the corresponding coordinates in the lightweight learnable gain layer. It then reconstructs the corresponding original unweighted forward amplitude features through hardware inverse computation and uses this as the base for the backpropagation gradient transformation. This data is then stored in the circular historical feature buffer queue with address locking, ensuring the physical-temporal consistency of data during subsequent local gradient transformations.

[0054] While replicating the real values, the effective elements are also mapped to four-dimensional coordinates in the tensor array: time, channel, and frequency coordinates. Frequency coordinates They are recorded as coordinate vectors. All coordinate vectors are sequentially organized into a two-dimensional integer matrix, forming a non-zero position index mask. The non-zero position index mask records the spatiotemporal frequency domain positions of non-zero eigenvalues ​​in the original tensor, completely preserving the pose and topological information of the anomalous features.

[0055] As an example, after an interrupt is triggered, the central processing unit suspends the lightweight sensing module and sends a safety signal. The intermediate-state feature tensor array is traversed based on the amplitude channel mask, identifying a total of 94,372 valid non-zero amplitude activation points. These 94,372 amplitude real values ​​and their corresponding 94,372 phase real values ​​are sequentially extracted to form a one-dimensional floating-point array of length 188,744, i.e., the non-zero valid feature array.

[0056] Assuming the first non-zero amplitude value encountered during the iteration is 0.87, its original four-dimensional coordinates are time 5, channel 0, and frequency. Index 128, Frequency With index 34, 0.87 becomes the first element of the fragment, and [5, 0, 128, 34] becomes the first row of the non-zero position index mask. The phase value at the same coordinate, i.e., channel 1, is extracted to form the second element of the fragment, and the corresponding coordinates [5, 1, 128, 34] become the second row of the mask. This process continues to execute, generating a non-zero valid feature array of dimension [188744] and a non-zero position index mask of dimension [188744, 4].

[0057] In one embodiment of the present invention, the key encapsulation module is configured to perform the following steps: The non-zero valid feature array is associated and arranged with the non-zero position index mask to obtain the associated arrangement data containing coordinates and real values. Based on the media access control address, current timestamp, and data packet sequence number of the edge computing unit, perform encrypted digest extraction operation to generate a header fixed-length hash checksum; The first fixed-length hash check code is inserted into the header register bit of the associated permutation data, encrypted with a pre-shared key, and the encrypted feature data stream is generated and sent to the cloud processing center.

[0058] After extraction, the edge computing unit's processor associates the non-zero location index mask with the non-zero valid feature array. For each feature deviation load (floating-point value) in the non-zero valid feature array, its corresponding multi-dimensional coordinate topology information in the non-zero location index mask is matched, forming a data record unit containing four-dimensional coordinates and real values, which is then organized into an associated arrangement of data in two-dimensional matrix form. A header fixed-length hash checksum is generated for verification and addressing. The header fixed-length hash checksum not only serves as a unique identifier for the data packet but also implicitly contains the data source, generation time, and order. The generation formula is as follows:

[0059] in, This is the first fixed-length hash checksum; For secure hashing algorithms; It is the media access control address of the edge computing unit; It is a high-precision Unix timestamp; It is an incrementing sequence counter used to distinguish multiple data packets within the same timestamp; symbol Indicates string concatenation; This indicates that the hash digest will be truncated to a specified length. For example, 64-bit, which is 8 bytes.

[0060] The first fixed-length hash checksum is then inserted at the very beginning of the serialized data stream of the associative permutation data, serving as a header register block. The complete data packet, containing both the first fixed-length hash checksum and the associative permutation data, is then encrypted using a pre-shared key and an Advanced Encryption Standard (AES) algorithm, encapsulating it to generate an encrypted data stream with logically defined flow attributes. This data stream is then transmitted upstream to the cloud processing center via the network interface.

[0061] As an example, the central processing unit merges the first row [5, 0, 128, 34] with the element 0.87 into a quintuple [5, 0, 128, 34, 0.87]. This iterates through all 188,744 valid feature points and... The device's MAC address (00:1A:2B:3C:4D:5E), current timestamp (1678886400.123456), and serial number (1201) are obtained and concatenated into a string. A SHA-256 hash operation is then performed. The first 64 bits are extracted to obtain a hexadecimal header fixed-length hash checksum (0x1A2B3C4D5E6F7890). This 8-byte checksum is then packaged and encrypted with the serialized associated permutation data to generate an encrypted data stream, which is then sent to the cloud.

[0062] In one embodiment of the present invention, the cloud wake-up module is used to perform the following steps: Receive encrypted feature data streams, block the full image inference task input into the cloud-based large model, and maintain the dormant state of inactive neurons in the cloud-based large model; The encrypted feature data stream is decrypted and the header fixed-length hash check code is verified, and the non-zero position index mask is parsed out. The coordinate sequence in the non-zero position index mask is converted into a pointer to the physical memory address pool of the multi-head self-attention module of the large model in the cloud, and the frequency domain receptive field operator window in the enabled state is triggered at a fixed point.

[0063] Upon receiving the data stream, the cloud processing center initiates processing, while the front-end gateway service suspends and queues for the full image inference task to enter the cloud-based large model. In this implementation, the cloud-based large model is designed to process high-dimensional visual features decoupled in the frequency domain. The system calls the data decryption and parsing module to decrypt and extract the header fixed-length hash checksum for source verification. After successful verification, the program separates the non-zero position index mask and the non-zero valid feature array. A strong matching operation is performed, which does not go through the conventional model input layer but instead activates the memory address using externally provided frequency domain coordinates.

[0064] The cloud resource scheduler forcibly maps the frequency coordinate sequence contained in the non-zero position index mask to the physical memory address pool within a multi-head self-attention module of the frequency domain visual converter model, which is based on a network configuration file and has a preset hierarchical depth. Through a preset mapping algorithm, the scheduler triggers the activation of the corresponding frequency domain receptive field operator window. This window refers to a set of local neurons and their connection weights that are activated at a specific point in the large cloud model. These neurons are dedicated to processing the local frequency band feature regions corresponding to the coordinates of the non-zero position index mask, thereby activating frequency domain resources that highly match the extracted frequency coordinates.

[0065] As an example, after identifying the encrypted data stream, the load balancer decrypts and extracts the header checksum 0x1A2B3C4D5E6F7890. After verification, it reconstructs a non-zero position index mask with dimensions [188744, 4]. At this point, the scheduler uses mapping rules, such as mapping time to batches, channels to attention headers, and frequency... and Mapping to the frequency domain feature channel sequence initiates strong matching. For the first row [5, 0, 128, 34], the scheduler instruction activates the operator window corresponding to the mid-frequency band coordinates [128, 34] of the 5th and 0th attention heads in layer 8. This process opens 188,744 frequency domain receptive field operator windows distributed at different tensor locations in batch, while inactive neurons remain dormant.

[0066] In one embodiment of the present invention, the gradient generation module is configured to perform the following steps: The non-zero valid feature array is injected into the corresponding frequency domain receptive field operator window via the data bus; The parameters of the local frequency band attention mechanism of the large cloud model are called to perform weighted matching between the non-zero effective feature array and the adjacent key vectors in the feature space, and to perform residual inverse derivation operation based on semantic features. Extract the error backpropagation features generated by the residual back derivation operation based on semantic features, and generate a target gradient correction tensor with the same dimension and topological structure as the non-zero effective feature array.

[0067] After the frequency domain receptive field operator window is activated, the resource scheduler guides the non-zero valid feature array into the corresponding activated operator window via the data bus. Subsequently, a residual backpropagation operation based on semantic features is performed. These semantic features refer to the abstract feature representation with global topological correlation extracted after the frequency domain features have undergone nonlinear mapping through a large model's multi-layer, multi-head self-attention network. This operation utilizes the local frequency band attention mechanism parameters trained on the frequency domain model, using the injected frequency domain feature values ​​as query vectors, matching them with adjacent semantic concept key vectors, and deriving a value representing the semantic residual gradient. After the calculation is complete, the error backpropagation procedure is triggered. At this point, the error backpropagation does not update the weights of the entire network model in the cloud; instead, it extracts the gradient information generated for the input features during this process and combines it into a new tensor, forming the target gradient correction tensor. The dimension and topological structure of this target gradient correction tensor are consistent with the injected non-zero valid feature array.

[0068] As an example, when 188,744 frequency domain receptive field operator windows are activated, the first element of the non-zero valid feature array, 0.87, is injected into the window corresponding to coordinates [5, 0, 128, 34]. The model determines that 0.87, representing the high-frequency component at this address coordinate, has a semantic residual with the learned typical "micro-scratches" feature key vector, indicating that adjusting this frequency domain feature value would better match the cloud template. The error backpropagation program then calculates the feature gradient, for example, +0.08, for the input feature 0.87. (See Appendix) Figure 4 The figure uses a micro-scratch as an example, with the dashed line representing the preset frequency radius threshold, which is set to 64 in the example. It can be seen that the high-frequency features of the micro-scratch are mainly distributed in regions where the Euclidean distance is greater than the preset frequency radius threshold. When the Euclidean distance is between 80 and 100, the semantic residual gradient reaches a peak of approximately 0.09, indicating that the large cloud model responds most strongly to the abnormal features of this type of micro-scratch in this frequency band, and can generate effective correction gradients for it. The corresponding feature gradients are calculated in parallel for all 188,744 injected values, combined into a target gradient correction tensor with consistent dimension and topology, and then distributed.

[0069] In one embodiment of the present invention, the edge evolution module is configured to perform the following steps: Receive the target gradient correction tensor, extract the header fixed-length hash check code for reverse tracing and localization, and determine the specific underlying hardware inside the edge computing unit that generates the abnormal features. Instructions are sent to the specific underlying hardware to retrieve the original forward amplitude feature real values ​​pre-stored in the buffer queue; The gradient values ​​in the target gradient correction tensor are used to overwrite the register gain weight state in the physical register, thus completing the local adaptive update of the frequency domain feature gain mapping.

[0070] The cloud sends the target gradient correction tensor to the corresponding edge computing unit. The edge computing unit receives and extracts the header fixed-length hash checksum, performs reverse tracing and localization, and matches it with local historical events. The system locates the specific underlying hardware that generated the anomaly features, that is, it locates the addressing coordinates in the lightweight learnable gain layer after the two-dimensional fast Fourier transform module. =128, A physical register array of 34.

[0071] Since the target gradient correction tensor retains topological information in the time dimension, while the physical registers in the lightweight learnable gain layer are only mapped to feature channels and spatial frequency coordinates, the processor of the edge computing unit initiates a cross-temporal gradient aggregation procedure. This procedure retrieves the original forward amplitude feature real values ​​corresponding to each time step locked in the circular historical feature buffer queue by consulting the non-zero position index mask.

[0072] According to the chain rule of calculus, the feature output gradient at each time step is multiplied one by one with the original unweighted forward amplitude feature real value corresponding to that time step to obtain the partial derivative gradient at each time step. This is applied to coordinates pointing to the same channel and frequency. The multiple partial derivative gradients are aggregated and averaged across time series to calculate the true average partial derivative gradient for the gain weight of the physical register.

[0073] The true average partial derivative gradient is combined with a pre-defined hardware learning rate multiplier to convert it into a target weight correction increment for updating the weights. The processor issues instructions to the gain layer, which, based on gradient ascent logic designed to maximize the response to anomalous features, uses the converted target weight correction increment to adjust the register gain weight state in the corresponding physical register by accumulating it.

[0074] By combining cross-temporal aggregation, learning rate conversion, and overwriting of the physical parameters of the gain layer, the lightweight sensing network of the edge computing unit completes the local adaptive update of the frequency domain feature gain mapping without modifying the mathematically meaningful Fourier transform kernel or recompiling the firmware, thus realizing asynchronous evolution closed-loop control based on physical parameter reconstruction.

[0075] As an example, the edge computing unit receives the target gradient correction tensor data packet and confirms its associated event through a checksum. This allows it to locate the corresponding amplitude channel and frequency coordinates within the lightweight learnable gain layer. The target frequency band gain physical register. This register currently stores the original gain coefficient, for example, an integer 128, corresponding to a gain weight of 1.0. During cross-time-series gradient aggregation, the system finds that for this coordinate (128, 34), within a 16-time-window time series, there are 3 time steps, such as... The output feature gradients were returned as non-zero, with values ​​of +0.07, +0.09, and +0.08.

[0076] The original forward amplitude feature real values, pre-stored in the buffer queue, are retrieved. Assuming the values ​​at these three time steps are 0.82, 0.78, and 0.80, respectively, they are multiplied one by one using the chain rule. The partial derivative gradients at these three time steps are then calculated as follows: 0.07 × 0.82 = 0.0574, 0.09 × 0.78 = 0.0702, and 0.08 × 0.80 = 0.064. To ensure the model's expected stability over the entire time series, the processor sums these three partial derivative gradients, totaling 0.1916, and divides this sum by the total time series length of 16, yielding the true average partial derivative gradient with respect to the gain weights, approximately +0.012.

[0077] Using the built-in hardware learning rate multiplier, assuming a learning rate of 0.5, the floating-point weight correction value is +0.006. Given that the register's quantization base is 128, i.e., 128 LSB = 1.0 gain, the system multiplies this floating-point value by the quantization base 0.006 × 128 = 0.768, and after rounding, obtains the target weight correction increment of +1.

[0078] Based on the gradient ascent logic reinforced by positive features, the system then adds this positive increment to the original value 128 to obtain a new value 129, and overwrites 129 as the new control word into the physical register. This operation enables the lightweight network to adjust its frequency response when encountering wafers with similar spatial distribution characteristics in subsequent inference cycles. =128, The signal gain amplitude provided on the 34 channels is slightly higher than the original baseline, which makes the weak scratch defect signals at the bottom layer more significant when entering the threshold decision-maker, thus realizing a hardware-level self-learning closed loop without interrupting the main task.

[0079] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A lightweight, large-model defect detection system with cloud-edge-device collaboration, characterized in that, include: The feature generation module performs frequency domain transformation on the acquired original pixel matrix signal through the edge computing unit to obtain an intermediate feature tensor array; The state determination module calculates the spatial non-zero density of the intermediate state feature tensor array and compares the spatial non-zero density with a preset static threshold to obtain the feature state boundary. The anomaly extraction module, in response to the spatial non-zero density exceeding a preset static threshold, extracts the non-zero elements and their coordinates in the intermediate state feature tensor array, and generates a non-zero valid feature array and a non-zero position index mask. The key encapsulation module merges and encapsulates the non-zero valid feature array and the non-zero position index mask to generate an encrypted feature data stream and send it to the cloud processing center. The cloud wake-up module parses the received encrypted feature data stream to obtain the non-zero position index mask and maps it to the physical memory address pool of the cloud large model in order to trigger the frequency domain receptive field operator window in the open state. The gradient generation module imports the non-zero valid feature array into the frequency domain receptive field operator window and generates the target gradient correction tensor through residual back derivation based on semantic features. The edge evolution module sends the target gradient correction tensor to the edge computing unit, and realizes local adaptive update of frequency domain feature gain mapping by overwriting the register gain weight state in the corresponding physical register.

2. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 1, characterized in that, The original pixel matrix signal is frequency-domain transformed by the edge computing unit to obtain an intermediate state feature tensor array, including the following steps: Obtain continuous raw pixel matrix signals and allocate the physical memory area of ​​the edge computing unit as physical cache space; The sensor network of the edge computing unit performs a two-dimensional discrete Fourier transform on the original pixel matrix signal to separate the amplitude and phase components and obtain the frequency domain feature flow. The frequency domain feature stream is written into the physical buffer space in time order to obtain an intermediate feature tensor array containing time, channel, vertical spatial frequency and horizontal spatial frequency dimensions.

3. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 2, characterized in that, The sensor network of the edge computing unit performs a two-dimensional discrete Fourier transform on the original pixel matrix signal to separate the amplitude and phase components, obtaining the frequency domain feature flow, including the following steps: Extract pixel intensity values ​​from the original pixel matrix signal of a single frame; Based on spatial coordinates and frequency domain coordinates, frequency domain mapping calculation is performed on pixel intensity values ​​to obtain frequency domain complex values ​​containing amplitude and phase information; The amplitude spectrum and phase spectrum are extracted from the complex values ​​in the frequency domain and combined to obtain the frequency domain feature flow.

4. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 2, characterized in that, The spatial non-zero density of the intermediate state feature tensor array is calculated, and the spatial non-zero density is compared with a preset static threshold to obtain the feature state boundary, including the following steps: Establish a monitoring link between the static random access memory array within the physical cache space and the activation value addressing monitor; The frequency of non-zero addressing activations with an absolute value greater than a preset background noise shielding threshold is counted when accessing the intermediate state feature tensor array using the activation value addressing monitor. The spatial non-zero density is calculated based on the non-zero addressing activation frequency and the total number of elements in the intermediate state feature tensor array. In response to a spatial non-zero density greater than a preset static threshold, the internal state register is triggered to flip, generating a characteristic state boundary.

5. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 4, characterized in that, In response to a spatial non-zero density exceeding a preset static threshold, non-zero elements and their coordinates are extracted from the intermediate-state feature tensor array to generate a non-zero valid feature array and a non-zero position index mask, including the following steps: In response to a spatial non-zero density exceeding a preset static threshold, the current main task is suspended, and a default safety instruction signal is output to the external control interface. Traverse the intermediate state feature tensor array, extract the real values ​​of effective amplitude elements and phase elements whose absolute values ​​are greater than the preset background noise shielding threshold, and combine them into a non-zero effective feature array in the form of a one-dimensional array. The multidimensional coordinate index values ​​of the non-zero valid feature array in the intermediate state feature tensor array are organized into a two-dimensional integer matrix, which serves as a non-zero position index mask containing multidimensional coordinate topological information.

6. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 1, characterized in that, The non-zero valid feature array and the non-zero position index mask are merged and encapsulated to generate an encrypted feature data stream, which is then sent to the cloud processing center. This includes the following steps: The non-zero valid feature array is associated and arranged with the non-zero position index mask to obtain the associated arrangement data containing coordinates and real values. Based on the media access control address, current timestamp, and data packet sequence number of the edge computing unit, perform encrypted digest extraction operation to generate a header fixed-length hash checksum; The first fixed-length hash check code is inserted into the header register bit of the associated permutation data, encrypted with a pre-shared key, and the encrypted feature data stream is generated and sent to the cloud processing center.

7. A lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 6, characterized in that, Parsing the received encrypted feature data stream to obtain the non-zero position index mask and mapping it to the physical memory address pool of the large cloud model to trigger the frequency domain receptive field operator window in the enabled state includes the following steps: Receive encrypted feature data streams, block the full image inference task input into the cloud-based large model, and maintain the dormant state of inactive neurons in the cloud-based large model; The encrypted feature data stream is decrypted and the header fixed-length hash check code is verified, and the non-zero position index mask is parsed out. The coordinate sequence in the non-zero position index mask is converted into a pointer to the physical memory address pool of the multi-head self-attention module of the large model in the cloud, and the frequency domain receptive field operator window in the enabled state is triggered at a fixed point.

8. The lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 1, characterized in that, Import the non-zero valid feature array into the frequency domain receptive field operator window, and generate the target gradient correction tensor through residual back derivation based on semantic features, including the following steps: The non-zero valid feature array is injected into the corresponding frequency domain receptive field operator window via the data bus; The parameters of the local frequency band attention mechanism of the large cloud model are called to perform weighted matching between the non-zero effective feature array and the adjacent key vectors in the feature space, and to perform residual inverse derivation operation based on semantic features. Extract the error backpropagation features generated by the residual back derivation operation based on semantic features, and generate a target gradient correction tensor with the same dimension and topological structure as the non-zero effective feature array.

9. A lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 8, characterized in that, The target gradient correction tensor is sent to the edge computing unit, and the local adaptive update of the frequency domain feature gain mapping is achieved by overwriting the register gain weight state in the corresponding physical register. This includes the following steps: Receive the target gradient correction tensor, extract the header fixed-length hash check code for reverse tracing and localization, and determine the specific underlying hardware inside the edge computing unit that generates the abnormal features. Instructions are sent to the specific underlying hardware to retrieve the original forward amplitude feature real values ​​pre-stored in the buffer queue; The gradient values ​​in the target gradient correction tensor are used to overwrite the register gain weight state in the physical register, thus completing the local adaptive update of the frequency domain feature gain mapping.

10. A lightweight large-model defect detection system with cloud-edge-device collaboration according to claim 9, characterized in that, The gradient values ​​in the correction tensor are corrected using the target gradient, and the register gain weight state in the physical register is overwritten to complete the local adaptive update of the frequency domain feature gain mapping, including the following steps: By combining the preset hardware learning rate and quantization base, the feature gradient values ​​in the target gradient correction tensor are converted into target integer weight correction increments; Based on gradient ascent logic, the target integer weight correction increment is used to update the state of the original coefficient values ​​in the physical register to generate a new control word; The new control word is overwritten into the physical register, thus completing the overwriting of the register gain and weight state.

Images