A lightweight industrial internet of things data collection method, system, device and medium for a semiconductor production line

Abstract

The present application relates to the technical field of industrial internet of things, in particular to a lightweight industrial internet of things data acquisition method, system, device and medium for semiconductor production line; relying on edge computing power sinking architecture, constructing edge lightweight real-time processing, cloud deep modeling backtracking hierarchical collaborative system; designing semiconductor process exclusive edge lightweight embedded processing algorithm, realizing high-precision noise reduction, adaptive down-sampling feature aggregation and high-proportion lossless compression of high-frequency time series data without improving the cost of main control hardware computing power; natively supporting semiconductor special interface and general industrial interface, adapting different manufacturer equipment through pluggable design, adapting narrow embedded installation working conditions of semiconductor production line, synchronously meeting the core technical demands of mass production high-frequency acquisition, low-latency transmission, high-precision measurement and non-loss of full-time series data from three technical dimensions of architecture, algorithm and hardware.

Description

Technical Field

[0001] This invention relates to the fields of semiconductor manufacturing and industrial Internet of Things (IIoT) technology, and more specifically, to a lightweight industrial IoT data acquisition method, system, device, and medium for semiconductor production lines. Background Technology

[0002] Semiconductor advanced process mass production sites face five rigid technical constraints: high-frequency sampling, strong electromagnetic interference, heterogeneous and dense equipment, limited deployment space, and the inability to lose data. Core process equipment needs to maintain high-frequency closed-loop sampling of 1kHz to 10kHz for extended periods, with stringent millisecond-level requirements for timing synchronization accuracy, acquisition signal-to-noise ratio, and end-to-end transmission latency. Currently, mainstream industrial IoT acquisition architectures generally adopt a front-end transparent transmission and cloud-based centralized computing power processing mode, without setting up hardware-level edge computing power links near the data source. The full back transmission of high-frequency raw timing data directly leads to bandwidth congestion, transmission jitter, and latency degradation in the factory's industrial control network. Cloud-based centralized computing resources are under constant high load and cannot meet the timing requirements of real-time closed-loop control in semiconductor processes. Meanwhile, existing general-purpose acquisition terminals embed a single filtering algorithm with fixed parameters, which is not optimized for cleanroom radio frequency electromagnetic coupling interference, long-term temperature drift of sensors over a wide temperature range, and instantaneous impact noise from high-power machine start-up and shutdown. This makes it impossible to simultaneously complete multi-dimensional noise reduction and nonlinear anomaly identification. Actual measured steady-state acquisition errors generally exceed ±0.5%, failing to meet the precision metrological quality control standards for WAT electrical testing and CP sorting. Conventional acquisition hardware has low integration, requiring external protocol conversion and signal conditioning modules to adapt to heterogeneous semiconductor devices with multiple protocols, occupying additional deployment space in small cabinets. External cabling also introduces secondary coupling interference. Furthermore, the terminals lack integrated industrial-grade high-capacity non-volatile time-series storage hardware and time-series alignment and continuation logic. In scenarios of short-term network disconnection or network segment maintenance switching, critical process original time-series data is directly and irreversibly lost, causing batch yield traceability and compliance audit failures across the entire chain. In summary, existing solutions suffer from unbalanced computing power configuration, poor algorithm and process adaptability, insufficient hardware integration, and lack of data fault tolerance. The industry urgently needs an integrated closed-loop technical solution that combines dedicated lightweight edge acquisition hardware coupling technology with embedded algorithms.

[0003] Semiconductor production lines are characterized by strong equipment heterogeneity (covering equipment from multiple manufacturers such as lithography machines, etching machines, and coating machines, with interface protocols including SECS / GEM, EtherNet / IP, Profinet, RS485, etc.), limited installation space, harsh environment (existing electromagnetic interference, temperature fluctuations, and dust pollution), and high requirements for data real-time performance and accuracy (sampling delay of key process parameters must be ≤10ms, and error ≤±0.5%).

[0004] Existing industrial IoT data acquisition devices have the following drawbacks: Insufficient versatility: Lack of native adaptation for semiconductor-specific interfaces (such as SECS / GEM), requiring additional protocol conversion modules, resulting in increased size and cost; Low lightweight design: Traditional data acquisition gateways are large in size (usually ≥20cm×15cm×10cm) and have high power consumption (operating power consumption ≥15W), which cannot meet the installation needs of densely populated production line equipment areas; Weak edge processing capabilities: It only forwards data and does not perform local preprocessing for the massive process data of semiconductor production lines, resulting in high bandwidth occupancy and heavy cloud computing pressure. Poor anti-interference performance: It is not optimized for the strong electromagnetic environment of semiconductor production lines, and packet loss and excessive error are prone to occur during data acquisition. Lack of adaptive capability: It cannot dynamically adjust the acquisition parameters according to the device's operating status and network quality, resulting in wasted power consumption or insufficient data integrity. Summary of the Invention

[0005] This invention addresses the problems of unbalanced computing power and latency in existing cloud-based centralized data acquisition architectures, insufficient anti-interference accuracy of general algorithms, low hardware integration, and data loss during network outages. It proposes a lightweight industrial IoT data acquisition method, system, equipment, and medium for semiconductor production lines. Based on an edge-based computing power sinking architecture, it constructs a layered collaborative system of lightweight edge real-time processing and deep cloud-based modeling and backtracking. It specifically designs edge-based lightweight embedded processing algorithms dedicated to semiconductor processes, achieving high-precision noise reduction, adaptive downsampling feature aggregation, and high-ratio lossless compression of high-frequency time-series data without increasing the computing power cost of the main control hardware. It natively supports both semiconductor-specific interfaces and general industrial interfaces, and its pluggable design adapts to equipment from different manufacturers and the confined embedded installation conditions of semiconductor production lines. From the three technical dimensions of architecture, algorithm, and hardware, it simultaneously meets the core technical requirements of mass production high-frequency acquisition, low-latency transmission, high-precision measurement, and no loss of full-time-series data.

[0006] The specific implementation details of this invention are as follows: A lightweight industrial IoT data acquisition method for semiconductor production lines includes the following steps: Step S1: Collect semiconductor device data according to the set sampling frequency and store it in the data buffer area; Step S2: Based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, the semiconductor device data is denoised in parallel and input into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain the process data package; Step S3: Based on the set time sequence tags, upload the process data package to the cloud-based industrial IoT platform with low latency through dual redundant communication.

[0007] To better realize the present invention, step S2 further includes the following steps: Step S21: Based on the 3σ criterion algorithm, construct a semiconductor process timing gradient correlation anomaly discrimination model to remove abnormal data from the semiconductor device data in the buffer area.

[0008] Step S22: Based on the constructed lightweight Kalman recursive filter link, the slow-varying baseline drift noise of the semiconductor data is reduced to obtain the first denoised semiconductor data; Step S23: Based on the deployed multi-layer wavelet soft threshold denoising branch, filter out the instantaneous strong pulse interference of the semiconductor data to obtain the second denoised semiconductor data; To better implement the present invention, step S21 further includes the following steps: Step S211: According to the 3σ criterion, abnormal data in the process timing data in the buffer area are removed. Outlier detection and removal are performed to obtain the effective semiconductor process timing data sequence. Step S212: Use the first-order forward difference method to perform differential operations on the semiconductor effective process timing data of two adjacent sampling points one by one to obtain the timing differential gradient; Step S213: Call the sliding window to smooth the temporal difference gradient, calculate the window mean corresponding to the temporal difference gradient, and obtain the smoothed gradient sequence; Step S214: Determine the smoothed gradient sequence according to the set gradient mutation determination threshold, and obtain the gradient mutation feature set; Step S215: Based on the prior features of the semiconductor device start-up and shutdown timing log, verify the gradient mutation feature set with correlation to obtain the gradient mutation feature set after binary precise differentiation.

[0009] To better implement the present invention, the specific operation of step S215 is as follows: based on the prior features of the semiconductor equipment start-up and shutdown timing ledger, the gradient mutation feature set is verified by correlation. If the gradient mutation time falls within the preset start-up and shutdown timing interval of the equipment, and the mutation gradient value and change direction match the inherent gradient features of the start-up and shutdown stage, it is determined to be a pseudo-anomaly of the inherent impact of the equipment start-up and shutdown. If the gradient mutation time does not fall within the start-up and shutdown timing interval, or the mutation feature does not match the inherent features of the start-up and shutdown stage, it is determined to be an anomaly of the actual process parameter drift.

[0010] To better implement the present invention, step S211 further includes the following steps: Step S2111: Based on the number of sampling points within the window, extract the process timing data sequence within the current sliding analysis window from the edge buffer. Step S2112: Based on the preliminary mean and preliminary variance of the calculated process time series data sequence, obtain a temporary threshold, remove extreme outliers that exceed the temporary threshold, and calculate the robust mean and robust variance. Step S2113: Obtain the anomaly detection threshold range based on the robust mean and robust variance; Step S2114: Based on the anomaly discrimination threshold range, process timing data within the window are discriminated one by one to obtain the semiconductor valid process timing data sequence.

[0011] To better implement the present invention, step S21 further includes the following steps: Step S221: Construct the state equation based on the obtained optimal estimation vector of the true internal state of the semiconductor process system, the fixed coefficient matrix of the one-step transition of the system state, and the Gaussian white noise sequence of the inherent weak process of the process system; Step S221: Construct the observation equation based on the obtained observation mapping coefficient matrix, the obtained optimal estimation vector of the true internal state of the semiconductor process system, and the sensor observation noise sequence; Step S223: Simplify the fixed process noise covariance matrix according to the state equation, and simplify the observation noise covariance matrix according to the observation equation to obtain the semiconductor data of the noise-reduced slow-varying baseline drift noise.

[0012] To better realize the present invention, step S23 further includes the following steps: Step S231: Based on the original noise coefficients of the high-frequency layer of wavelet decomposition and the threshold coefficients of the multi-layer adaptive soft threshold denoising, call the symbol discrimination function to calculate the soft threshold; Step S232: Suppress and reconstruct the high-frequency noise figure based on the soft threshold to obtain semiconductor data with transient strong pulse interference filtered out.

[0013] Based on the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines, and to better realize this invention, a lightweight industrial IoT data acquisition system for semiconductor production lines is further proposed, which is implemented based on the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines; including an acquisition module, a lightweight noise reduction and anomaly identification module, and an edge-cloud layered collaborative transmission module. The acquisition module is used to acquire semiconductor device data according to a set sampling frequency and store it in a data buffer. The lightweight noise reduction anomaly identification module is used to reduce semiconductor equipment data in parallel based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, and input it into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain process data packets. The edge-cloud layered collaborative transmission module is used to upload process data packets to the cloud-based industrial IoT platform with low latency through dual-redundant communication, based on the set time sequence tags.

[0014] Based on the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines, and to better realize this invention, an electronic device is further proposed, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines is realized.

[0015] Based on the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines, and to better realize this invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the aforementioned electronic device, the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines is realized.

[0016] The present invention has the following beneficial effects: (1) This invention optimizes the allocation of computing resources from the topology level through a hierarchical edge computing architecture, which greatly reduces the uplink bandwidth occupation of industrial control networks and the concurrent computing load in the cloud. The end-to-end measurement and control latency is controllable at the millisecond level, which meets the closed-loop linkage timing requirements of semiconductor precision manufacturing processes.

[0017] (2) The present invention has strong hardware adaptability through self-developed lightweight composite algorithm cluster. It can achieve wideband noise reduction, adaptive sampling and accurate anomaly detection without additional computing power modules. The accuracy of data acquisition and measurement is significantly better than that of general acquisition equipment, and it effectively avoids process misjudgment and false alarm caused by environmental interference.

[0018] (3) The present invention has a compact hardware topology with integrated isolation cabling, strong electromagnetic compatibility protection capability, and is suitable for small embedded installation scenarios in clean rooms. It does not require on-site cabling modification, has low deployment and construction costs, and no secondary interference is introduced.

[0019] (4) The hardware-level timing fault-tolerant storage coupled differential transmission logic of the present invention ensures the integrity and traceability of key process data timing at all times, and fully meets the technical specifications of the entire chain of wafer mass production quality control and compliance audit.

[0020] (5) The present invention features partitioned dynamic power consumption control hardware link adaptation process condition adaptive scheduling, which ensures high stability of equipment during long-term continuous operation and significantly reduces the frequency of manual intervention in operation and maintenance. Attached Figure Description

[0021] Figure 1A schematic flowchart of a lightweight industrial IoT data acquisition method for semiconductor production lines provided by the present invention. Detailed Implementation

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] Example 1: This embodiment specifically includes the following steps: Step S1: Collect semiconductor device data according to the set sampling frequency and store it in the data buffer area; Step S2: Based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, the semiconductor device data is denoised in parallel and input into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain the process data package; Step S2 specifically includes the following steps: Step S21: Based on the 3σ criterion algorithm, construct a semiconductor process timing gradient correlation anomaly discrimination model to remove abnormal data from the semiconductor equipment data that needs noise reduction.

[0025] Step S21 specifically includes the following steps: Step S211: According to the 3σ criterion, abnormal data in the process timing data in the buffer area are removed. Outlier detection and removal are performed to obtain the effective semiconductor process timing data sequence. Step S211 specifically includes the following steps: Step S2111: Based on the number of sampling points within the window, extract the process timing data sequence within the current sliding analysis window from the edge buffer. Step S2112: Based on the preliminary mean and preliminary variance of the calculated process time series data sequence, obtain a temporary threshold, remove extreme outliers that exceed the temporary threshold, and calculate the robust mean and robust variance. Step S2113: Obtain the anomaly detection threshold range based on the robust mean and robust variance; Step S2114: Based on the anomaly discrimination threshold range, process timing data within the window are discriminated one by one to obtain the semiconductor valid process timing data sequence.

[0026] Step S212: Use the first-order forward difference method to perform differential operations on the semiconductor effective process timing data of two adjacent sampling points one by one to obtain the timing differential gradient; Step S213: Call the sliding window to smooth the temporal difference gradient, calculate the window mean corresponding to the temporal difference gradient, and obtain the smoothed gradient sequence; Step S214: Determine the smoothed gradient sequence according to the set gradient mutation determination threshold, and obtain the gradient mutation feature set; Step S215: Based on the prior features of the semiconductor device start-up and shutdown timing log, verify the gradient mutation feature set with correlation to obtain the gradient mutation feature set after binary precise differentiation.

[0027] The specific operation of step S215 is as follows: based on the prior characteristics of the semiconductor equipment start-up and shutdown timing ledger, the gradient mutation feature set is verified by correlation. If the gradient mutation time falls within the preset start-up and shutdown timing interval of the equipment, and the mutation gradient value and change direction match the inherent gradient characteristics of the start-up and shutdown stage, it is determined to be a pseudo-anomaly of the inherent impact of the equipment start-up and shutdown. If the gradient mutation time does not fall within the start-up and shutdown timing interval, or the mutation characteristics do not match the inherent characteristics of start-up and shutdown, it is determined to be an anomaly of the actual process parameter drift.

[0028] Step S22: Based on the constructed lightweight Kalman recursive filter link, the slow-varying baseline drift noise of the semiconductor data is reduced to obtain the first denoised semiconductor data; Step S22 specifically includes the following steps: Step S221: Construct the state equation based on the obtained optimal estimation vector of the true internal state of the semiconductor process system, the fixed coefficient matrix of the one-step transition of the system state, and the Gaussian white noise sequence of the inherent weak process of the process system; Step S221: Construct the observation equation based on the obtained observation mapping coefficient matrix, the obtained optimal estimation vector of the true internal state of the semiconductor process system, and the sensor observation noise sequence; Step S223: Simplify the fixed process noise covariance matrix according to the state equation, and simplify the observation noise covariance matrix according to the observation equation to obtain the semiconductor data of the noise-reduced slow-varying baseline drift noise.

[0029] Step S23: Based on the deployed multi-layer wavelet soft threshold denoising branch, filter out the instantaneous strong pulse interference of the semiconductor data to obtain the second denoised semiconductor data; Step S23 specifically includes the following steps: Step S231: Based on the original noise coefficients of the high-frequency layer of wavelet decomposition and the threshold coefficients of the multi-layer adaptive soft threshold denoising, call the symbol discrimination function to calculate the soft threshold; Step S232: Suppress and reconstruct the high-frequency noise figure based on the soft threshold to obtain semiconductor data with transient strong pulse interference filtered out.

[0030] Step S3: Based on the set time sequence tags, upload the process data package to the cloud-based industrial IoT platform with low latency through dual redundant communication.

[0031] Working Principle: This embodiment relies on an edge-based computing power sinking architecture to construct a lightweight real-time edge processing and cloud-based deep modeling and backtracking hierarchical collaborative system. It features a specially designed lightweight embedded processing algorithm for semiconductor manufacturing processes, achieving high-precision noise reduction, adaptive downsampling feature aggregation, and high-ratio lossless compression of high-frequency time-series data without increasing the computing power cost of the main control hardware. It natively supports both semiconductor-specific interfaces and general industrial interfaces, and its pluggable design adapts to different manufacturers' equipment and the confined embedded installation conditions of semiconductor production lines. From the three technical dimensions of architecture, algorithm, and hardware, it simultaneously meets the core technical requirements of mass production high-frequency acquisition, low-latency transmission, high-precision measurement, and no loss of full-time-series data.

[0032] Example 2: This embodiment is based on Embodiment 1 above, and will be described in detail with a specific embodiment.

[0033] Step 1: Hardware layered PCB mounting, debugging and algorithm firmware solidification. The main control chip, protocol core, sampling unit, communication unit and storage unit are integrated according to the four-layer isolation wiring process. The edge algorithm cluster is modularly compiled and burned to the main control unit. The core parameters of the algorithm and the process threshold range are calibrated offline. Step 2: On-site embedded installation and network synchronization. Deploy the device near the semiconductor equipment data source, connect to the native interface of the equipment in the partition, connect to the production line power supply, and complete bus clock synchronization, communication link handshake and storage partition formatting. Step 3: In-situ parallel high-frequency acquisition: The original timing signals and device status messages of the semiconductor process are acquired in parallel and synchronously through the DMA channel and directly written into the on-chip cache; Step 4: Edge-side full-process algorithm preprocessing, calling the solidified algorithm cluster, and sequentially executing improved 3σ criterion anomaly removal, composite noise reduction, process time-series gradient correlation anomaly discrimination, adaptive downsampling and feature aggregation; A dual-branch parallel coupled architecture of Kalman baseline denoising and wavelet impulse denoising is used, in which Kalman filtering constructs the state equation X. k =AX k-1 +W k With observation equation Z k =HX k +V k The process noise covariance matrix Q and the observation noise covariance matrix R are fixed by offline calibration; wavelet soft threshold denoising uses the formula w'=sign(w)(|w|-λ) to suppress and reconstruct the high-frequency noise coefficient, and λ is adaptively fine-tuned according to the electromagnetic environment intensity of the production line.

[0034] The specific steps for anomaly removal using the improved 3σ criterion are as follows: Step 41: Divide the sliding analysis window and extract the process timing data sequence {x} within the window. 1, x2,...,x n The window size dynamically adapts to the sampling frequency. Step 42: Calculate the initial mean μ0 and initial variance σ0² of the original sequence, set temporary thresholds [μ0-2σ0, μ0+2σ0], remove extreme outliers, and recalculate the robust mean μ and robust variance σ². Step 43: Set the anomaly detection threshold range [μ-3σ, μ+3σ], and identify and remove abnormal data within the window point by point; Step 44: Iteratively update the robust mean μ and robust variance σ², and adaptively adjust the window size and threshold coefficient.

[0035] The process timing gradient correlation anomaly detection is achieved through timing difference gradient calculation of adjacent sampling points. The specific steps are as follows: Step 4a: Select the effective process timing data sequence {y1, y2, ..., y} after composite noise reduction and anomaly removal. m}, determine the sampling time interval Δt0; Step 4b: Use the first-order forward difference method, according to formula g j =(y j+1 -y j ) / Δt0 calculates the temporal difference gradient g between adjacent sampling points point by point. j ; Step 4c: Smooth the gradient sequence using a sliding window of length 3 to 5, and calculate the window mean. The smoothed gradient sequence is obtained; Step 4d: Set the gradient mutation detection threshold G th Extract gradient mutation features and record mutation parameters; Step 4e: Combine the prior features of the equipment start-up and shutdown sequence log with the verification gradient mutation features to distinguish between pseudo-anomalies of equipment start-up and shutdown and real process parameter drift anomalies.

[0036] The adaptive downsampling and feature aggregation, based on multi-process threshold linkage discrimination logic, increases the downsampling rate and extracts the mean μ, variance σ², and peak value max(x) in the process steady-state interval. i The data includes multi-dimensional features of step gradient Δx / Δt, retains 10kHz full high-frequency sampling data in the critical fluctuation range of the process, and has a data compression ratio of ≥80%.

[0037] Step 5: Hierarchical priority dual-link transmission, sorting data according to priority and encapsulating it with a lightweight protocol, relying on dual redundant communication channels to transmit to the cloud, real-time monitoring of link status and seamless switching; The data is sorted by three priority levels: abnormal event data, key closed-loop process data, and normal steady-state characteristic data. Under normal circumstances, wired Ethernet transmission is used, and in the event of a link failure, it seamlessly switches to a wireless backup channel. The transmission protocol is the lightweight MQTT protocol.

[0038] Step 6: Timing fault tolerance fallback and closed-loop verification. When the network is down, the unuploaded data is written to the local Flash partition for archiving. After the link is restored, the data is re-uploaded according to the timing priority and the data integrity is verified. When the network is out of service and the cache is interrupted, the main controller generates a time-series index timestamp and archives the data in a three-dimensional structured manner according to the batch number, device ID, and time-series stamp. After the link is restored, the time-series breakpoints in the cloud are compared, and the data is retransmitted in batches with differential transmission and the data integrity is verified packet by packet. After the retransmission is completed, redundant cache is cleared.

[0039] Step 7: Full-condition adaptive control, dynamically adjust the main control computing power, sampling rate and communication power, periodically check the hardware condition, and achieve 7×24-hour unattended steady-state operation.

[0040] Working Principle: This embodiment abandons the traditional full-transmission cloud-based centralized computing architecture and innovatively adopts a near-edge front-end computing power sinking topology. Lightweight acquisition hardware is deployed in close proximity to sensors, PLCs, and chamber controllers. The hardware natively integrates edge real-time computing power channels to complete the entire process of filtering, noise reduction, anomaly removal, time-series aggregation, and redundancy compression of high-frequency raw time-series data on-site. Only structured and effective feature data and abnormal event fragments are output for uplink interaction. This constructs a hierarchical computing power division system that enables millisecond-level real-time measurement and control on the edge side and offline big data deep analysis and modeling on the cloud side. This reduces uplink bandwidth load, compresses transmission latency, and eliminates cloud computing power concurrency bottlenecks from the root of the topology.

[0041] This embodiment targets a wide-range high-frequency sampling scenario of 1kHz~10kHz, employing an embedded, dedicated algorithm cluster with low computational overhead to overcome the limitations of traditional fixed sampling and single-filtering techniques. It achieves adaptive dynamic downsampling and multi-dimensional feature extraction based on multi-process threshold linkage discrimination logic, balancing steady-state data compression efficiency with transient process anomaly fidelity preservation. It integrates a lightweight Kalman recursive filter and a multi-layer wavelet threshold denoising architecture to accurately suppress composite field strength interference. An improved 3σ criterion algorithm is introduced, combined with semiconductor process data timing continuity optimization statistical logic, to accurately remove abnormal data from the buffer, avoiding extreme value interference in subsequent feature extraction. A process timing gradient correlation anomaly discrimination model is constructed, replacing the traditional threshold comparison coarse-grained judgment logic with point-by-point calculation of temporal differential gradients between adjacent sampling points, smoothing preprocessing, abrupt feature extraction, and correlation verification with equipment start-up and shutdown timing. This achieves accurate binary differentiation between weak disturbance real anomalies and strong environmental pseudo-interference. It is complemented by a timing index closed-loop cache and a differential timing alignment breakpoint resumption mechanism, with pure algorithmic logic providing a safety net to ensure the timing integrity of data throughout the entire process.

[0042] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0043] Example 3: This embodiment describes a lightweight industrial IoT data acquisition device and process based on any one of the above embodiments 1-2, using a specific embodiment.

[0044] Step 1: Lightweight and highly reliable edge acquisition hardware integration topology solution; This device adopts a four-layer layered isolated PCB integrated architecture, with core hardware links physically isolated according to functional zones to avoid electromagnetic crosstalk throughout. The main control unit uses an industrial-grade 480MHz floating-point MCU, with independent on-chip partitions for DMA parallel acquisition cache, edge algorithm real-time computation, and link scheduling backend. A natively bound floating-point acceleration unit ensures lightweight, parallel, low-latency iterative computation of algorithms. The front-end signal acquisition link integrates a dedicated semiconductor hardware decoding protocol core, natively parsing SECS / GEM high-speed process messages. It also features an onboard industrial Ethernet bus transceiver chip and a 16-bit synchronous high-precision ADC array, enabling time-division parallel acquisition of digital process messages and analog operating condition signals. All front-end acquisition points are connected in series with second-order passive filter protection circuits to suppress transient pulse interference at the ports. The edge computing power interaction link directly couples to the local 8GB wide-temperature industrial-grade Flash memory chip, independently constructing a timing closed-loop storage hardware channel without occupying main control unit memory resources. The communication transmission link integrates redundant industrial-grade wired Ethernet and anti-interference Wi-Fi 6 dual physical channels. The link layer incorporates a hardware-level heartbeat latency detection circuit, providing real-time output of link error rate and latency fluctuation quantification indicators, triggering seamless link switching commands. The power supply link adopts a wide-voltage 9~24V graded voltage regulation topology, with front-end reverse connection and surge protection circuits, and multiple independent LDOs at the back end for zoned power supply, dynamically allocating power to the acquisition, algorithm, communication, and storage units as needed. The overall hardware is encapsulated in a compact electromagnetically shielded aluminum shell, meeting the protection requirements for embedded installations in the field.

[0045] Step 2: Implementation of adaptive threshold-linked downsampling and multi-dimensional feature extraction algorithm techniques; For four core timing parameters—semiconductor chamber temperature, vacuum level, RF power, and special gas flow rate—a multi-parameter joint steady-state discrimination threshold range [T] is constructed. h [T1], [P] h [P1] Real-time point-by-point comparison of the original time-series signal x(t) sampled at high frequency in situ with the process threshold boundary. When all measured values ​​of process parameters fall stably within the threshold range for a long period of time, it is determined to be the steady-state range of the process. An adaptive downsampling rate N-based enhancement strategy is executed, and the time-series sequence within the window is simultaneously subjected to point-by-point mathematical statistical operations to extract effective process features: mean μ = ∑x i / N、Variance σ²=∑(x i -μ)² / N, peak maximum value max(x iThe step gradient Δx / Δt is only packaged and uploaded as a structured feature dataset. When any single process parameter goes out of bounds or there is a sudden change in gradient, it is determined to be a critical process fluctuation / anomaly initiation range. The original 10kHz full high-frequency time-series sampling data is automatically latched without downsampling compression, thus fully preserving transient process details. The entire algorithm runs in a lightweight iterative manner, with single-window computation latency strictly controlled within 1ms. The overall redundant data compression ratio is no less than 80%, balancing the dual requirements of transmission bandwidth optimization and process anomaly tracing and fidelity.

[0046] Furthermore, the parameters are explained as follows: x(t) is the original high-frequency timing sampled discrete signal sequence of the semiconductor process obtained by in-situ synchronous acquisition of the edge-end ADC; T h T1 and T2 are the upper and lower critical thresholds for safe process temperature of the chamber, respectively, and P is the lower threshold for safe process temperature of the chamber. h P1 and P2 represent the upper and lower critical thresholds for safe closed-loop pressure in the chamber vacuum, respectively. These threshold parameters are calibrated offline based on the lithography, etching, and CVD mass production process logs. N represents the total number of effective sampling points within the sliding analysis window, which is also used as an adaptive dynamic downsampling adjustable factor. i σ² represents the actual process physical quantity sampled at the i-th discrete sampling time within the sliding time window; μ is the steady-state mean value of the process time within the window, used to feedback the degree of steady-state baseline offset of the chamber; σ² is the time-series fluctuation variance quantification index within the window, used to characterize the severity of process micro-perturbations; max(x i ) represents the instantaneous peak extreme value feature of the window, used to capture instantaneous power impact and pressure peak transient conditions; Δx / Δt represents the first-order differential gradient of the time series between adjacent sampling points, used to quickly identify the trend of process step switching and abnormal drift. All parameters are calculated in real time on the embedded local basis, without secondary backtracking calculation in the cloud.

[0047] Step 3: Implementation of a composite lightweight noise reduction and process-specific anomaly detection algorithm coupling technology; First, a lightweight Kalman recursive filter is deployed to achieve baseline smooth noise reduction, and the state equation X is constructed. k =AX k-1 +W k Observation equation Z k =HX k +V kThe system simplifies the fixed process noise covariance matrix Q and observation noise covariance matrix R, reducing the computational overhead of embedded hardware. It iteratively outputs the smoothed optimal time-series estimate, filtering out slow-varying baseline drift noise caused by sensor temperature drift. A multi-layer wavelet soft-threshold denoising branch is simultaneously connected in parallel to suppress high-frequency electromagnetic coupling pulse noise with layered coefficients. The soft-threshold calculation formula is w'=sign(w)(|w|-λ), accurately filtering out instantaneous strong pulse interference. After two-stage denoising, the data is input into the semiconductor process time-series gradient correlation anomaly discrimination model. The time-series differential gradient between adjacent sampling points is calculated. Combined with the prior features of the process equipment start-up and shutdown time-series ledger, a gradient mutation correlation criterion is constructed. This binary distinction differentiates between inherent impact pseudo-anomalies during equipment start-up and shutdown and true process parameter drift anomalies, ultimately outputting high-precision and effective process time-series data. The steady-state acquisition end-to-end error is controlled within ≤±0.1%.

[0048] Furthermore, the parameter descriptions are as follows: X k X is the optimal estimation vector of the true internal state of the semiconductor process system at time k, representing the ideal true values ​​of the chamber temperature and pressure without disturbance; k-1 Z represents the historical backtracking state vector from the previous moment; A is the fixed coefficient matrix for one-step transition of the system state, which is calibrated to conform to the physical characteristics of semiconductor thermal inertia and pressure hysteresis; k H is the noisy time-series vector of the actual high-frequency sampling observation at the edge at time k; H is the observation mapping coefficient matrix, which is matched to the unified calibration of the physical dimensions of the ADC sampling; W k V represents the inherent weak Gaussian white noise sequence of the process system, with Q being its fixed covariance matrix; k The noise sequence is the superimposed observation noise sequence of electromagnetic coupling and sensor temperature drift in the clean room. R is its fixed covariance matrix. The Q and R matrices are calibrated offline and then fixedly written into the firmware, eliminating the need for on-site debugging. w is the original noise coefficient of the high-frequency layer of wavelet decomposition. w' is the effective wavelet coefficient reconstructed after noise reduction. λ is the multi-layer adaptive soft threshold noise reduction threshold coefficient, which is adaptively adjusted according to the electromagnetic environment intensity of the production line. sign(·) is the sign discrimination function, which retains the effective abrupt polarity and suppresses irregular impulse noise. All parameters are adapted to low-computing-power embedded iterative calculations and do not occupy additional edge computing resources.

[0049] The core technology implementation process is as follows: First, a lightweight Kalman recursive filter link is built, abandoning the complex calculation logic of dynamically updating the covariance matrix in traditional Kalman filtering. Instead, an offline calibration method is used to fix the process noise covariance matrix Q and the observation noise covariance matrix R, reducing the iterative overhead of embedded hardware computing power. Then, state equations and observation equations adapted to the inertial characteristics of semiconductor processes are constructed, where the state equation is X. k =AX k-1 +W k The observation equation is Z k =HX k +V kBy recursively iteratively calculating, the optimal estimate of the process timing signal is output, thereby achieving accurate filtering of slow baseline drift noise caused by sensor temperature drift. Simultaneously, a multi-layer wavelet soft threshold denoising branch is deployed in parallel to perform wavelet multi-scale decomposition on the acquired signal, extract wavelet coefficients containing high-frequency electromagnetic coupling pulse noise, and use the soft threshold calculation formula w'=sign(w)(|w|-λ) to suppress and reconstruct the high-frequency noise coefficients, where λ is calibrated offline according to the electromagnetic environment intensity of the production line and can be adaptively fine-tuned to achieve efficient filtering of instantaneous strong pulse interference.

[0050] After the dual-branch noise reduction link works in concert, an innovative process timing gradient mutation correlation discrimination mechanism is introduced. Based on the timing differential gradient calculation results of adjacent sampling points and combined with the prior characteristics of semiconductor equipment start-up and shutdown timing ledger, a gradient mutation correlation criterion is constructed. By quantifying the matching degree between the gradient mutation amplitude, duration and process start-up and shutdown timing, a binary accurate distinction is achieved between the inherent impact pseudo-anomaly of equipment start-up and shutdown and the real process parameter drift anomaly.

[0051] The specific calculation steps for the temporal difference gradient between adjacent sampling points are as follows: First, data preprocessing: After composite noise reduction and anomaly removal by the improved 3σ criterion, the effective process time series data sequence obtained is denoted as {y1,y2,...,y...}. m}, where m is the number of valid sampling points within the current sliding window, and y j The first step is to calculate the effective value of the process parameter at the j-th sampling time (j=1,2,...,m), with a sampling time interval of Δt0 (determined by the sampling frequency; 1kHz sampling corresponds to Δt0=1ms, and 10kHz sampling corresponds to Δt0=0.1ms), ensuring continuous and complete data timing. The second step is to calculate the point-by-point differential gradient using the first-order forward differencing method, performing differential operations on the process parameter values ​​of adjacent sampling points one by one, and calculating the timing differential gradient g between the j-th sampling point and the (j+1)-th sampling point. j The calculation formula is g j =(y j+1 -y j ) / Δt0, where j=1,2,...,m-1, this gradient value intuitively reflects the rate and direction of change of the process parameters at adjacent sampling times. Positive values ​​indicate that the parameters are increasing, negative values ​​indicate that the parameters are decreasing, and the absolute value indicates the degree of change. The third step is gradient smoothing preprocessing. To avoid gradient misjudgment caused by small fluctuations at a single sampling point, the calculated differential gradient sequence {g1,g2,...,g...} is processed. m-1 Perform sliding window smoothing, selecting a sliding smoothing window of length k (k ranges from 3 to 5, adaptively adjusted according to the process type), and calculate the window mean value corresponding to each gradient value. (i is the offset within the window), resulting in the smoothed gradient sequence { The first step is to reduce the interference of small fluctuations and retain the core features of gradient mutations; the fourth step is to extract gradient mutation features and set a gradient mutation judgment threshold G. th (Based on the parameter fluctuation characteristics of different processes such as photolithography, etching, and CVD, offline calibration and solidification are performed), and the smoothed gradient values ​​are checked one by one. Make a judgment, if If a gradient mutation occurs at that moment, the mutation time, gradient value, and direction are recorded to form a gradient mutation feature set. The fifth step is gradient correlation verification. Combining the prior features of the semiconductor equipment start-up and shutdown timing log (written in advance to the firmware, including the preset start-up and shutdown timing intervals and the gradient change range during the start-up and shutdown phases), the extracted gradient mutation features are correlated and verified. If the gradient mutation time falls within the preset start-up and shutdown timing intervals, and the mutation gradient value and direction match the inherent gradient features of the start-up and shutdown phases, it is determined to be a pseudo-anomaly of inherent start-up and shutdown impacts. If the gradient mutation time does not fall within the start-up and shutdown timing intervals, or the mutation features do not match the inherent start-up and shutdown features, it is determined to be a true process parameter drift anomaly.

[0052] Furthermore, an improved 3σ criterion algorithm is introduced, combined with the timing continuity of semiconductor process data to optimize statistical logic, thereby achieving accurate removal of outlier data in the buffer. This embodiment targets the process timing data in the edge buffer, introducing an improved 3σ criterion algorithm to accurately remove outliers. It overcomes the limitations of the traditional 3σ criterion, which relies solely on a single global mean and variance and is susceptible to misjudgments due to extreme values. The algorithm combines the timing continuity characteristics of semiconductor process data with optimized iterative logic. The specific algorithm steps are as follows: First, data selection and window division: Extract the process timing data sequence {x1, x2, ..., x...} within the current sliding analysis window from the edge buffer. n}, where n is the number of sampling points within the window, and the window size is dynamically adapted according to the sampling frequency (1kHz sampling corresponds to a window of n=1000, 10kHz sampling corresponds to a window of n=5000), ensuring that both steady-state characteristics of the process are covered and instantaneous anomalies are captured in a timely manner; the second step is robust mean and variance calculation. Abandoning the traditional global mean calculation method, a robust statistical method is adopted after removing initial extreme values. First, the initial mean of the original sequence, μ0=∑x, is calculated. i / n and preliminary variance σ0²=∑(x iThe first step involves setting a temporary threshold [μ0-2σ0, μ0+2σ0] based on the initial mean and variance. Extreme outliers exceeding this temporary threshold are then removed. The robust mean μ and robust variance σ² are recalculated on the remaining valid data to avoid interference from single extreme values. The second step is outlier threshold calibration. Based on the robust statistically obtained μ and σ², the outlier discrimination threshold interval is set to [μ-3σ, μ+3σ], where 3σ is the deviation range corresponding to the 99.73% confidence interval of a normal distribution, adapting to the normal distribution characteristics of semiconductor process data. The third step is outlier discrimination and removal. Each sampled data point x within the window is analyzed one by one. i Perform a judgment, if x i If x [μ-3σ,μ+3σ], then the data is considered an outlier (caused by electromagnetic interference, sensor malfunction, etc.) and is removed from the buffer data; if x i Data ∈[μ-3σ,μ+3σ] is considered valid and retained for subsequent feature extraction. The fifth step involves iterative updates and adaptive adjustments. After each sliding window's anomaly removal, the robust mean μ and robust variance σ² of the current window are automatically used as initial reference values ​​for the next window, reducing redundant calculations. Simultaneously, the window size and 3σ threshold coefficient are adaptively adjusted based on the process type (lithography / etching / CVD) to ensure the algorithm adapts to the data fluctuation characteristics of different processes. The entire algorithm is implemented in a lightweight embedded manner, with a single-window computation latency ≤0.3ms. It seamlessly integrates with the aforementioned composite denoising algorithm, further improving the purity of the buffered process data and providing high-quality data support for subsequent feature extraction and transmission.

[0053] Step 4: Edge-cloud layered collaborative transmission and timing fault-tolerant fallback closed-loop technical solution; After full algorithm preprocessing is completed at the edge, the data is prioritized in three levels: abnormal event data, key closed-loop process data, and routine steady-state characteristic data. Hierarchical scheduling and transmission are implemented using dual-redundant physical communication channels. Under normal operating conditions, low-latency wired Ethernet transmission is prioritized, and lightweight MQTT protocol is used to encapsulate structured data packets, reducing cloud-based parsing overhead. The hardware link layer samples the link error rate and latency jitter quantification indicators every millisecond in real time. When the error rate continuously exceeds the limit, triggering a link failure judgment, the hardware automatically and seamlessly switches to the wireless backup channel without interrupting the front-end acquisition and algorithm processing. In the event of a network anomaly and disconnection, the main controller immediately generates a time-series index timestamp, writes the currently incomplete process time-series data closed-loop into a local independent Flash hardware partition, and archives it in a three-dimensional structured archive according to batch number, device ID, and time-series stamp. After the link is restored, the background time-series alignment program automatically compares the time-series breakpoints in the cloud ledger, and performs batch differential retransmission according to production time-series priority, synchronously verifying data packet integrity to prevent time-series out-of-order issues, data loss, and duplicate uploads, forming a complete closed-loop technology link of acquisition-processing-transmission-fault tolerance.

[0054] Based on the aforementioned dedicated lightweight data acquisition device hardware, coupled with an adapted edge-layered architecture and a dedicated algorithm engine, a standardized, mass-producible, and traceable closed-loop data acquisition and control method is formed. The specific steps are coherent and closed-loop, with strong adaptability to various operating conditions. Step 1: Rapid on-site adaptation of network configuration and initial deployment of edge computing power; Maintenance personnel can quickly connect the corresponding pluggable acquisition sub-modules to different process equipment based on the on-site lithography, etching, CVD, and testing processes without tools. The main control module automatically identifies the peripheral model and matches the native communication protocol in seconds via the I2C high-speed bus. Process ledger parameters, alarm threshold ranges, and sampling baseline configurations can be sent remotely via cloud or local touch control with one click. The device automatically completes the initialization of the edge lightweight algorithm engine computing power, local Flash storage timing partition formatting, and dual communication link heartbeat networking. The power supply module defaults to energy-saving mode and is on standby 24 / 7, seamlessly connecting to the existing industrial control network of the production line without downtime or modification of the existing industrial control architecture.

[0055] Step 2: Parallel acquisition of raw data across the entire domain via high-frequency synchronous acquisition at the edge; The device is deployed close to the PLC, process sensors, and chamber controller of the machine, using in-situ data sources. It collects all raw data in parallel and synchronously according to a preset high-frequency sampling strategy of 1kHz to 10kHz. This includes core process parameters such as real-time chamber temperature, vacuum closed-loop pressure, RF output power, instantaneous flow rate of special gases, and machine axis motion status. Simultaneously, it collects cleanroom environmental temperature and humidity, electromagnetic interference background values, and auxiliary linkage parameters such as equipment power supply voltage and current. All raw high-frequency data is written to the local high-speed cache in milliseconds through a DMA hardware pass-through channel, without consuming the main control core computing power, ensuring no frame loss or timing offset throughout the acquisition process.

[0056] Step 3: Edge-side lightweight engine performs full-domain preprocessing to achieve a high proportion of redundancy reduction and accuracy improvement; The dedicated edge lightweight processing engine is activated to complete intelligent pre-processing of the entire chain on-site at the edge, without uploading any invalid raw redundant data: First, adaptive threshold linkage differentiated downsampling + multi-dimensional feature extraction, real-time comparison of real-time process operating parameters with preset safety fluctuation thresholds, maintaining 10kHz full raw sampling retention in critical fluctuations, anomaly initiation, and high-risk intervals of process switching, and automatically and intelligently downsampling in stable and steady-state production intervals, simultaneously batch extracting time series mean, fluctuation variance, peak deviation, step change and other multi-dimensional effective process features, with an overall lossless data compression ratio of ≥80%; Second, multi-level composite noise reduction + process The system employs a dual algorithm that combines lightweight Kalman real-time filtering with wavelet hierarchical denoising to precisely eliminate process-specific anomalies. This algorithm filters out high-frequency electromagnetic coupling interference in cleanrooms, long-term temperature drift and baseline drift of sensors, and spurious disturbance noise from momentary impacts during equipment start-up and shutdown. A dedicated semiconductor process timing anomaly detection model is then used for secondary precise identification. Two levels of noise reduction differentiate between genuine process anomalies and environmental interference false trigger signals, ensuring robust control over the accuracy error of all effective data acquisition to ≤±0.1%. Thirdly, the system uses structured and lightweight packaging and labeling to assign high-priority timing tags to anomaly event fragments and key closed-loop control data, generating standardized and lightweight structured data packages.

[0057] Step 4: Layered collaboration between edge and cloud, with critical data being sent up reliably and with low latency; Strictly adhering to the principles of lightweight real-time edge control and layered division of labor for in-depth backtracking modeling in the cloud, only high-priority key process data packets, equipment anomaly alarm feature data, and batch traceability core ledger data that have been pre-processed at the edge are selectively uploaded to the cloud-based industrial IoT platform with low latency through dual redundant communication modules. The cloud backend does not need to repeatedly use computing power to clean up the original redundant data, and can directly and quickly carry out batch yield correlation analysis, full-domain equipment status assessment, full-process process backtracking modeling, and compliance audit archiving. Lightweight feature backup data is locally retained at the edge and can be remotely retrieved for backtracking as needed, which significantly reduces uplink bandwidth load and cloud computing power consumption, and reduces end-to-end linkage latency across the entire domain.

[0058] Step 5: Real-time link awareness + local caching as a backup + orderly resumption of transmission after breakpoints; The communication module performs millisecond-level real-time self-checks of wired and wireless dual-link communication quality, and synchronously links with a local 8GB industrial-grade Flash fault-tolerant cache unit for routine standby: Under healthy network conditions, lightweight structured data is uploaded in an orderly and prioritized manner, with local silent backup and retention; in the event of network jitter, factory network segment maintenance, wireless signal blockage or interruption, etc., the local closed-loop timing cache mechanism is automatically triggered instantly to completely lock all lithography and etching key process data for the current period without interrupting the on-site acquisition and algorithm operation process; after the network link quality recovers to the standard, the unuploaded cached data is automatically re-uploaded in batches according to the production sequence and priority level, eliminating data disorder, duplication, and loss problems throughout the process, and ensuring full-chain traceability of mass production data.

[0059] Step 6: Full-condition adaptive linkage parameter tuning + terminal health self-inspection closed-loop operation and maintenance; The main control module monitors three core operating states in real time around the clock: real-time process production status of the machine, health and quality of the on-site industrial control network link, and the operating temperature and real-time power consumption of the device itself. It dynamically and adaptively matches the optimal operating strategy: during peak and intensive process phases, it automatically switches to full computing power operation mode, fully activating high-frequency sampling and high-precision algorithm computing power to ensure real-time closed-loop control of the process; during equipment standby and idle periods, and network channel congestion periods, it automatically reduces the sampling frequency, optimizes algorithm computing power allocation, and enhances data compression efficiency to reduce overall power consumption; during the equipment shutdown and maintenance phase, it switches to ultra-low power sleep mode, retaining only link monitoring and log caching functions; it synchronously and in real-time inspects the temperature rise, power supply voltage, and interface connection status of each module, providing local audible and visual alarms for abnormalities and remotely pushing alarm logs to the cloud, achieving long-term stable unattended operation.

[0060] Working principle: First, this embodiment optimizes computing resource allocation at the topology level through a layered edge computing architecture, significantly reducing the uplink bandwidth consumption of industrial control networks and the concurrent computing load on the cloud. End-to-end measurement and control latency is controllable at the millisecond level, meeting the timing requirements of closed-loop linkage in semiconductor precision processes. Second, the self-developed lightweight composite algorithm cluster has strong hardware adaptability, achieving wideband noise reduction, adaptive sampling, and accurate anomaly detection without additional computing modules. The acquisition and measurement accuracy is significantly better than that of general acquisition equipment, effectively avoiding process misjudgments and false alarms caused by environmental interference. Third, the integrated isolated cabling and compact hardware... The device topology features strong electromagnetic compatibility protection, making it suitable for confined, embedded installations in cleanrooms. It requires no on-site wiring modifications, resulting in low deployment and construction costs and no secondary interference. Fourthly, the hardware-level time-series fault-tolerant storage coupled differential transmission logic ensures the complete and traceable timing of critical process data at all times, fully meeting the technical specifications for wafer mass production quality control and compliance auditing across the entire chain. Fifthly, the partitioned dynamic power consumption control hardware link adapts to process conditions and adaptively schedules the equipment, ensuring high stability during long-term continuous operation and significantly reducing the frequency of manual intervention in maintenance, making it suitable for 24 / 7 unattended mass production.

[0061] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0062] Example 4: This embodiment is based on any one of embodiments 1-3 above. 1. Lightweight Industrial IoT Data Acquisition Device The device includes a compact hardware module and a lightweight software system. Step 1: The hardware module adopts a layered integration design. The main control computing core layer, signal acquisition isolation layer, power supply protection layer, and peripheral interface expansion layer are integrated using a four-layer isolation cabling process. The analog signal routing area and the digital computing power routing area are strictly separated, and ground plane isolation and shielding are used at the intersection. The industrial-grade main control chip, SECS / GEM hard decoding protocol core, 16-bit synchronous ADC array, dual-channel communication physical module, 8GB independent wide-temperature Flash memory chip, and multi-level electromagnetic protection power supply device are soldered and cured. After the hardware assembly is completed, it is connected to a standard 9~24V wide voltage tooling power supply. The port impedance, communication signal-to-noise ratio, and timing synchronization accuracy are tested point by point. A 72-hour high and low temperature full-load electrical aging test is carried out to select hardware units that meet the electrical parameter consistency standards for backup.

[0063] Step 2: Lightweight Edge Algorithm Cluster Embedded Firmware Fixing and Parameter Calibration In an embedded real-time operating system environment, the modular compilation of adaptive multi-threshold downsampling program, lightweight Kalman filter iterative program, wavelet hierarchical soft threshold denoising program, process timing gradient anomaly discrimination program, and timing index cache alignment and continuation program is uniformly packaged into a low-computing-power firmware image. According to the on-site process logs of three types of processes, namely photolithography, etching, and CVD, the steady-state threshold ranges of four core parameters, namely temperature, vacuum degree, RF power, and gas flow rate, are calibrated offline. The basic parameters of Q / R noise covariance, wavelet hierarchical threshold coefficients, and timing gradient criterion thresholds of the algorithm are solidified. After the firmware is burned, a standard timing signal source is connected to carry out closed-loop calibration of algorithm accuracy. Only after verifying that the compression ratio, noise reduction signal-to-noise ratio, and steady-state error meet the full technical indicators can the product be shipped.

[0064] Step 3: Embedded installation and bus timing network synchronization at the nearest site In the confined installation area near the semiconductor etching machine, lithography machine chamber controller, and front-end process sensors, a lightweight acquisition device is fixedly and securely deployed. It interfaces with the equipment's native SECS / GEM process interface, industrial bus interface, and analog sensor interface in designated areas. It is connected to the original 24V DC industrial power supply of the production line, with a closed multi-level protection power supply circuit. Through the local maintenance serial port, the device discovers the field equipment ID, batch log, and sampling bandwidth configuration parameters of 1kHz~10kHz. The device automatically completes bus clock synchronization, dual communication link heartbeat handshake, and local Flash timing storage partition formatting. It seamlessly connects to the existing factory industrial control network without altering the original production line measurement and control wiring topology.

[0065] Step 4: In-situ parallel high-frequency acquisition and edge-side full-process algorithm synchronous iterative calculation The device's hardware DMA channel synchronously acquires multiple raw process timing signals and equipment status messages in parallel. The data is directly written to the on-chip cache without consuming core computing power. The background real-time task scheduling mechanism polls and calls the solidified algorithm cluster at millisecond intervals, performs adaptive threshold comparison frame by frame to dynamically adjust the sampling density, performs two-level composite noise reduction operations in parallel, substitutes the process-specific gradient model to complete the binary discrimination of anomalies, and generates lightweight structured effective feature data packets and abnormal event marker frames in real time. All mathematical processing is completed in a closed loop at the edge throughout the process, and the original redundant data is not transparent to the external network.

[0066] Step 5: Hierarchical priority dual-link scheduling transmission and real-time link status inspection After edge-side preprocessing, lightweight encrypted messages are encapsulated according to data priority sequence. Under normal circumstances, they are uploaded to the cloud platform with low latency via wired Ethernet channel, while hardware collects link error rate and latency jitter quantification parameters in real time. Once the link communication indicators are detected to deteriorate beyond the limit, the hardware link seamless switching mechanism is immediately triggered to switch to the wireless backup channel to continue transmission. The data collection and algorithm calculation process is not interrupted throughout the process. The cloud directly receives structured and effective data to carry out deep modeling and backtracking without the need for secondary computing power to analyze and reduce noise.

[0067] Step 6: Closed-loop verification of link failure timing caching and network recovery differential timing retransmission The hardware link sensing unit monitors the communication link connectivity in real time. In the event of a momentary network outage, the main controller automatically overlays high-precision time stamps and batch index tags, and writes all unuploaded process data to an independent non-volatile Flash partition for archiving storage. After the link communication quality recovers to the compliance threshold, it automatically reads the cloud breakpoint time sequence ledger, sorts it by production time and classifies it by process priority, and re-uploads the cached data in batches in a differential and orderly manner. It verifies the data integrity and time sequence alignment packet by packet. After the re-upload is completed, it automatically cleans up redundant backup cache entries. After closed-loop verification that there is no loss, no errors, and no duplicates, it enters the normal data collection and operation state.

[0068] Step 7: Full-condition adaptive power consumption and computing power linkage regulation and long-term steady-state monitoring The device's backend collects four types of quantitative parameters in real time: hardware junction temperature, real-time power consumption, front-end process operating conditions, and network load. It dynamically adjusts the main control computing power and frequency, front-end sampling rate, and communication transmission power. During periods of high process load, it operates at full computing power in a steady state, while during periods of low standby load, it reduces computing power and sampling frequency, automatically optimizing the overall power consumption distribution. It also periodically inspects the electrical conditions of each hardware unit, generates standardized operation and maintenance logs, and stores them locally, achieving 24 / 7 unattended continuous steady-state operation and adapting to the stringent operating conditions of the entire semiconductor mass production cycle.

[0069] 2. Data Acquisition Methods Based on the above device, the following steps are performed: Step 1: Interface Adaptation and Initialization 1.1 The user inserts the corresponding pluggable acquisition module according to the interface type of the target semiconductor device, and the control module automatically identifies the module type through the I2C bus; 1.2 The control module loads the preset protocol configuration file (built-in drivers for protocols such as SECS / GEM, EtherNet / IP, etc.), establishes a communication connection with the device, and configures the acquisition parameters (sampling frequency 1Hz-100Hz, data type filtering). 1.3 The power supply module starts up, enters energy-saving mode by default, initializes the communication module and connects to the edge gateway.

[0070] Step 2: Multi-source data acquisition 2.1 The control module triggers the acquisition module according to the sampling frequency to synchronously acquire digital (process parameters, equipment status codes) and analog (ambient temperature, operating voltage) data of the semiconductor equipment; 2.2 The collected data is transmitted in real time to the control module buffer (capacity ≥16KB) via the DMA channel to avoid excessive CPU usage.

[0071] Step 3: Edge Preprocessing 3.1 Outlier Removal: An improved 3σ criterion algorithm is applied to the cached data to remove outlier data that deviates from the mean by three times the standard deviation (such as jump values ​​caused by sudden equipment failure). 3.2 Lightweight noise reduction: Simplified Kalman filtering is performed on the data after outlier removal to reduce noise caused by electromagnetic interference; 3.3 Data Compression: The noise-reduced data is compressed in stages according to type. Key process data (such as wafer etching accuracy) retains the original accuracy, while ordinary state data is compressed at a ratio of 5:1.

[0072] Step 4: Dual-link data transmission 4.1 The control module sorts data according to priority (process alarm data > critical process data > normal status data) and prioritizes the transmission of high-priority data; 4.2 The communication module prioritizes data transmission via the Wi-Fi 6 link using the MQTT protocol, and monitors the link quality in real time (triggering switching when the packet loss rate is >5%). 4.3 When the wireless link fails, it automatically switches to the wired Ethernet link and caches the untransmitted data to the local eMMC storage (capacity 8GB). The data will be retransmitted after the link is restored.

[0073] Step 5: Adaptive parameter adjustment 5.1 The control module monitors three types of statuses in real time: equipment operating status (whether it is in the process execution period), network link quality (packet loss rate, latency), and device power consumption; 5.2 Dynamic Adjustment Strategy: During process execution: switch to full load mode, increase the sampling frequency to 50-100Hz, and turn off data compression; When the network packet loss rate is less than 5% and the device is in standby mode: switch to power saving mode, reduce the sampling frequency to 1-10Hz, and increase the compression ratio to 8:1; No data acquisition required (device shutdown): Switch to sleep mode, only keep the communication module listening, power consumption ≤3W.

[0074] Step 6: Status monitoring and fault feedback; 6.1 The control module monitors the temperature (threshold ≤ 85℃), power supply voltage (9-24V), and communication link status of each module in real time; 6.2 When a module overheats, experiences abnormal voltage, or has a link interruption for more than 5 minutes, a fault alarm is sent to the edge gateway via the communication module, and a fault log is recorded to local storage.

[0075] 3. Software and hardware system deployment and integration (1) Hardware selection and assembly Control module: Uses STM32H743VIT6 microcontroller (480MHz clock frequency, built-in 2MB Flash, 1MB RAM); Plug-in data acquisition module: SECS / GEM submodule: ADLINKMXE-5400 chip, supports SECS-II protocol, board-to-board connector model: HRSDF30FC-20DS-0.4V; General interface sub-modules: MAX485 (RS485), W5500 (EtherNet / IP), PNcoupler (Profinet) integrated board; Analog quantum module: ADS1256 chip, sampling rate 100 SPS, accuracy ±0.0015%; Communication module: ESP32-C3-WROOM-02 (supports 802.11ax, BLE5.0), RJ45 interface with integrated network transformer HR601680; Power supply module: TPS5430 step-down chip, input 9-24VDC, output 3.3V / 5V, maximum output current 3A; Structural components: Aluminum housing (9cm×7cm×4cm), with built-in 1mm thick aluminum heat sink; housing grounding terminal model: M3 grounding post; Assembly method: Each module is integrated in layers through a PCB board. Pluggable modules are connected to the motherboard through board-to-board connectors. The shell is fixed by clips and has an IP65 protection rating.

[0076] (2) Software implementation Control module firmware: Based on the FreeRTOS operating system and developed in C language, it is divided into acquisition tasks, preprocessing tasks, transmission tasks, and status monitoring tasks, and task synchronization is achieved through semaphores; Preprocessing algorithms: Improved 3σ criterion algorithm (adding a sliding window mechanism with configurable window size), lightweight Kalman filter (simplifying the Q / R matrix to a constant), and adding a semiconductor data feature dictionary during data compression; Communication protocols: The MQTT protocol uses the EMQX client, and the CoAP protocol uses the libcoap library, supporting QoS level 1 transmission; Configuration and monitoring: Provides local button configuration (sampling frequency, interface selection) and cloud remote configuration (commands are sent via MQTT topics), and supports remote export of fault logs.

[0077] (3) Application verification This device was applied to a 12-inch wafer etching production line, connecting a certain brand of etching machine (interface: SECS / GEM) to an auxiliary temperature control device (interface: EtherNet / IP): Installation: The device is fixed to the side of the etching machine (space size 8cm×8cm), and the power supply is the production line's 24VDC power supply; Data acquisition parameters: sampling frequency of 50Hz during process execution and 10Hz during standby; key data (etching depth, chamber pressure) have the highest priority. Verification results: Data acquisition delay ≤8ms, error ±0.25%, transmission packet loss rate 0.08%, average power consumption 6.2W, continuous operation for 30 days without failure, meeting production line requirements.

[0078] 4. Implementation Examples (1) Example of hardware integration and standardized assembly of the whole machine In the actual mass production assembly stage, a four-layer high-density PCB layered integrated surface mount technology is adopted to strictly control the overall size and fit the narrow installation requirements: the main control unit uses an industrial-grade STM32H743VIT6 main control chip, with onboard firmware partitioning for a lightweight edge computing algorithm engine; the pluggable acquisition interface partitions are independently surface mounted according to SECS / GEM semiconductor dedicated protocol channels, industrial bus channels, and high-precision analog channels, and are equipped with high-temperature resistant industrial-grade board-to-board connectors; the edge computing power scheduling circuit has independent routing to avoid electromagnetic crosstalk of process signals; and local fault-tolerant buffering is used. The unit features industrial-grade wide-temperature 8GB Flash memory chips, with separate isolated partitions for storing process-related confidential data; the dual-link communication unit integrates an anti-interference network transformer and an industrial-grade Wi-Fi shielded antenna; the power supply unit is equipped with a high-efficiency step-down and voltage-regulating circuit and triple protection circuits against overcurrent, overvoltage, and surge; all core modules are integrated and encapsulated in an integrated die-cast aluminum shielded shell with built-in ultra-thin micro heat sinks, achieving an IP65 protection rating. Before leaving the factory, the unit undergoes 72 hours of full-load high and low temperature aging testing and EMC class 3 electromagnetic compatibility testing, and can only be deployed after passing the tests.

[0079] (2) Embedded Deployment Implementation Example of Edge Lightweight Algorithm Engine Firmware The firmware runtime environment is built based on the FreeRTOS real-time lightweight embedded operating system. It adopts standard C language modular development and compilation, adapts to the low computing power hardware resources of the main control unit, and has no redundant background programs consuming computing power. A dedicated real-time task thread for the algorithm is created, with a higher priority than the regular operation and maintenance thread. The adaptive threshold linkage downsampling function, multi-dimensional process feature extraction function, lightweight Kalman filter function, wavelet hierarchical denoising simplification function, semiconductor process timing anomaly discrimination model, high-ratio lossless compression function, and timing cache index arrangement function are modularly packaged and solidified into the main control firmware. On-site, only the corresponding process threshold ledger needs to be remotely issued to automatically match the algorithm running parameters for different scenarios such as lithography, etching, CVD, and testing. The entire process on the edge side is automatically and seamlessly run without manual secondary algorithm debugging and adaptation.

[0080] (3) On-site verification of 12-inch wafer etching mass production line A full-condition field test and verification was conducted in the 12-inch advanced dry etching mass production workshop of a leading wafer fab. The site was equipped with multiple heterogeneous etching machines, chamber vacuum closed-loop controllers, and high-precision temperature and pressure sensors. The site had strong electromagnetic interference, limited installation space, and required 24 / 7 uninterrupted high-frequency production. The lightweight acquisition device was embedded in the narrow mounting slot on the side of the etching machine, connected to the production line's standard 24V industrial DC power supply, and quickly plugged into the SECS / GEM dedicated acquisition submodule to interface with the machine's native process interface. The sampling range was configured to cover the entire range of 1kHz to 10kHz, and the etching chamber temperature, vacuum level, and RF power safety fluctuation threshold were recorded. The device automatically completed network initialization, edge algorithm engine startup, and local time-series partition mounting. Actual test results: The compression ratio of edge in-situ preprocessing data is stable at ≥82%, the accuracy error of key process parameter acquisition across the entire link is ≤±0.08%, and the latency of edge single batch data processing across the entire process is stable at ≤0.8ms; the dual-link switching is seamless, the core etching batch data is completely cached during network outages, and the automatic timing retransmission is error-free after reconnection; the uplink bandwidth usage is reduced by more than 80% year-on-year, and the computing power load of the cloud backend is significantly reduced; the device has been running continuously at full load for 30 days without crashes, disconnections, or data anomalies, perfectly adapting to the harsh working conditions of semiconductor mass production and fully meeting the requirements for industrial deployment and patent mass production applications.

[0081] The other parts of this embodiment are the same as any one of the embodiments 1-3 above, so they will not be described again.

[0082] Example 5: Based on any one of Embodiments 1-4 above, this embodiment proposes a lightweight industrial IoT data acquisition system for semiconductor production lines, which is implemented based on the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines; it includes an acquisition module, a lightweight noise reduction and anomaly identification module, and an edge-cloud layered collaborative transmission module; The acquisition module is used to acquire semiconductor device data according to a set sampling frequency and store it in a data buffer. The lightweight noise reduction anomaly identification module is used to reduce semiconductor equipment data in parallel based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, and input it into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain process data packets. The edge-cloud layered collaborative transmission module is used to upload process data packets to the cloud-based industrial IoT platform with low latency through dual-redundant communication, based on the set time sequence tags.

[0083] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described lightweight industrial IoT data acquisition method for semiconductor production lines.

[0084] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned lightweight industrial IoT data acquisition method for semiconductor production lines is implemented.

[0085] The other parts of this embodiment are the same as any one of the embodiments 1-4 above, so they will not be described again.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A lightweight industrial IoT data acquisition method for semiconductor production lines, characterized in that, Specifically, the following steps are included: Step S1: Collect semiconductor device data according to the set sampling frequency and store it in the data buffer area; Step S2: Based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, the semiconductor device data is denoised in parallel and input into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain the process data package; Step S3: Based on the set time sequence tags, upload the process data package to the cloud-based industrial IoT platform with low latency through dual redundant communication.

2. The lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Based on the 3σ criterion algorithm, construct a semiconductor process timing gradient correlation anomaly discrimination model to remove abnormal data from the semiconductor equipment data that needs noise reduction; Step S22: Based on the constructed lightweight Kalman recursive filter link, the slow-varying baseline drift noise of the semiconductor data is reduced to obtain the first denoised semiconductor data; Step S23: Based on the deployed multi-layer wavelet soft threshold denoising branch, filter out the instantaneous strong pulse interference of the semiconductor data to obtain the second denoised semiconductor data.

3. The lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 2, characterized in that, Step S21 specifically includes the following steps: Step S211: According to the 3σ criterion, abnormal data in the process timing data in the buffer area are removed. Outlier detection and removal are performed to obtain the effective semiconductor process timing data sequence. Step S212: Use the first-order forward difference method to perform differential operations on the semiconductor effective process timing data of two adjacent sampling points one by one to obtain the timing differential gradient; Step S213: Call the sliding window to smooth the temporal difference gradient, calculate the window mean corresponding to the temporal difference gradient, and obtain the smoothed gradient sequence; Step S214: Determine the smoothed gradient sequence according to the set gradient mutation determination threshold, and obtain the gradient mutation feature set; Step S215: Based on the prior features of the semiconductor device start-up and shutdown timing log, verify the gradient mutation feature set with correlation to obtain the gradient mutation feature set after binary precise differentiation.

4. The lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 2, characterized in that, Step S22 specifically includes the following steps: Step S221: Construct the state equation based on the obtained optimal estimation vector of the true internal state of the semiconductor process system, the fixed coefficient matrix of the one-step transition of the system state, and the Gaussian white noise sequence of the inherent weak process of the process system; Step S221: Construct the observation equation based on the obtained observation mapping coefficient matrix, the obtained optimal estimation vector of the true internal state of the semiconductor process system, and the sensor observation noise sequence; Step S223: Simplify the fixed process noise covariance matrix according to the state equation, and simplify the observation noise covariance matrix according to the observation equation to obtain the semiconductor data of the noise-reduced slow-varying baseline drift noise.

5. A lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 2, characterized in that, Step S23 specifically includes the following steps: Step S231: Based on the original noise coefficients of the high-frequency layer of wavelet decomposition and the threshold coefficients of the multi-layer adaptive soft threshold denoising, call the symbol discrimination function to calculate the soft threshold; Step S232: Suppress and reconstruct the high-frequency noise figure based on the soft threshold to obtain semiconductor data with transient strong pulse interference filtered out.

6. The lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 5, characterized in that, The specific operation of step S235 is as follows: based on the prior characteristics of the semiconductor equipment start-up and shutdown timing ledger, the gradient mutation feature set is verified by correlation. If the gradient mutation time falls within the preset start-up and shutdown timing interval of the equipment, and the mutation gradient value and change direction match the inherent gradient characteristics of the start-up and shutdown stage, it is determined to be a pseudo-anomaly of the inherent impact of the equipment start-up and shutdown. If the gradient mutation time does not fall within the start-up and shutdown timing interval, or the mutation characteristics do not match the inherent characteristics of start-up and shutdown, it is determined to be an anomaly of the actual process parameter drift.

7. A lightweight industrial IoT data acquisition method for semiconductor production lines according to claim 5, characterized in that, Step S231 specifically includes the following steps: Step S2311: Extract the process timing data sequence within the current sliding analysis window from the edge buffer based on the number of sampling points within the window; Step S2312: Based on the preliminary mean and preliminary variance of the calculated process time series data sequence, obtain a temporary threshold, remove extreme outliers that exceed the temporary threshold, and calculate the robust mean and robust variance. Step S2313: Obtain the anomaly detection threshold range based on the robust mean and robust variance; Step S2314: Based on the anomaly discrimination threshold range, process timing data within the window are discriminated one by one to obtain the semiconductor valid process timing data sequence.

8. A lightweight industrial IoT data acquisition system for semiconductor production lines, implemented based on the lightweight industrial IoT data acquisition method for semiconductor production lines as described in claim 1; characterized in that, This includes a data acquisition module, a lightweight noise reduction and anomaly detection module, and an edge-cloud layered collaborative transmission module. The acquisition module is used to acquire semiconductor device data according to a set sampling frequency and store it in a data buffer. The lightweight noise reduction anomaly identification module is used to reduce semiconductor equipment data in parallel based on the constructed lightweight Kalman recursive filter link and the deployed multi-layer wavelet soft threshold denoising branch, and input it into the constructed semiconductor process timing gradient correlation anomaly discrimination model to obtain process data packets. The edge-cloud layered collaborative transmission module is used to upload process data packets to the cloud-based industrial IoT platform with low latency through dual-redundant communication, based on the set time sequence tags.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements a lightweight industrial IoT data acquisition method for semiconductor production lines as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 9, they implement a lightweight industrial IoT data acquisition method for semiconductor production lines as described in any one of claims 1-7.

Images