Blockchain-based smt production data traceability system

Abstract

The application provides a blockchain-based SMT production data tracing system, and belongs to the field of electronic product manufacturing.The physical identity registration module collects PCB microtexture images by using an optical detection device, extracts a physically unclonable function feature to generate a unique identity mark;the device proof module collects process noise signals from production devices by using a trusted execution environment, generates a source proof after verification by a device noise fingerprint model;the privacy compliance proof module generates a compliance proof based on a zero-knowledge compliance circuit without leaking the content of the production data;and the certificate synthesis and chain uploading module fuses the above-mentioned identity and proof into a self-proving tracing certificate and uploads the same to a blockchain.The application realizes full-link trusted tracing from the physical world to the digital world through the organic fusion of a physical fingerprint anchor, a noise fingerprint proof source and zero-knowledge compliance privacy protection, and at the same time protects the core process secrets of a contract manufacturer.

Description

Technical Field

[0001] This invention belongs to the field of electronic product manufacturing, and specifically relates to a blockchain-based SMT production data traceability system. Background Technology

[0002] In the electronics manufacturing industry, surface mount technology (SMT) is a core process for printed circuit board (PCB) assembly. With increasingly stringent requirements for product quality traceability in high-reliability fields such as automotive electronics, aerospace electronics, and medical electronics, establishing a reliable traceability system covering the entire process from component sourcing and production process parameters to finished product testing results has become a crucial element in quality assurance cooperation between Electronics Manufacturing Services (EMS) providers and brand owners.

[0003] In recent years, blockchain technology has been introduced into the field of SMT production data traceability due to its characteristics such as data immutability, decentralized notarization, and multi-party shared ledger. The basic idea is to upload hash digests or encrypted credentials of key process data in the SMT production line (such as solder paste printing parameters, chip placement information, reflow soldering temperature profiles, automatic optical inspection results, and material batch numbers) to the blockchain, while the original detailed data remains stored in the Manufacturing Execution System (MES) or an off-chain database. By incorporating brand owners, foundries, component suppliers, and downstream customers into authorized nodes through a consortium blockchain, multi-party data notarization and on-demand traceability are achieved.

[0004] However, existing blockchain-based SMT traceability solutions have revealed three fundamental, interdependent technical flaws in actual engineering deployments.

[0005] First, the binding relationship between physical entities and digital records is fragile. Existing solutions almost invariably rely on manually affixed barcodes, QR codes, or Radio Frequency Identification (RFID) tags as unique identifiers for PCBs. These external identifiers face numerous risks throughout the SMT process: high-temperature reflow soldering may cause tags to detach or deform; chemical solvents in cleaning processes may corrode the barcode surface; flux splatter during wave soldering and selective soldering may obscure the barcode; and malicious tag replacement can point inferior products to excellent production records. Once the binding relationship between the physical identifier and the PCB entity is broken, the entire traceability chain collapses, and the data on the blockchain becomes a "ghost record" without a physical carrier.

[0006] Second, the authenticity of data at its source before being uploaded to the blockchain cannot be effectively guaranteed. Blockchain technology can only ensure the immutability of data after it is uploaded, but it is powerless to verify the authenticity of data before it is uploaded—this constitutes the infamous "garbage in, garbage out" dilemma. Many SMT production lines contain a large number of outdated devices whose control systems lack hardware security modules (such as Trusted Platform Modules, TPMs) and cannot cryptographically sign the output data. During the process of collecting, aggregating, and uploading production data via edge gateways, it faces the risk of being tampered with by middleware, injected with malware, or forged by insiders. If a foundation of trust cannot be established at the source of the equipment, the blockchain merely stores the "hash value of tampered data," significantly diminishing its traceability value.

[0007] Third, there is an irreconcilable contradiction between protecting commercial privacy and verifying process compliance. Blockchain requires multiple nodes to share a ledger to achieve oversight and non-repudiation, but contract manufacturers' core process parameters (such as reflow oven temperature profile settings, stencil aperture sizes, and placement machine pressure parameters) are their trade secrets that give them a competitive edge in the market, and they are unwilling to expose them in plaintext to upstream and downstream companies. Even with methods such as channel isolation and symmetric encryption storage, transaction metadata (such as the frequency, time period, and volume of data uploaded to the blockchain) may still leak sensitive business intelligence such as capacity utilization and new product introduction schedules. While privacy computing technologies such as homomorphic encryption and zero-knowledge proofs can theoretically alleviate this contradiction, existing solutions are far from reaching industrial-grade usability in terms of processing efficiency for massive real-time SMT production data, support for complex traceability queries, and deployment costs. Summary of the Invention

[0008] In view of the above situation, the main objective of this invention is to propose a blockchain-based SMT production data traceability system to solve the aforementioned technical problems.

[0009] This invention proposes a blockchain-based SMT production data traceability system, including...

[0010] The physical identity registration module is configured to acquire microscopic texture images of printed circuit boards (PCBs) collected by optical inspection equipment in the SMT production line, extract physical non-cloning function features, and generate a unique identity identifier for the PCB.

[0011] The equipment verification module is configured to use a trusted execution environment to collect physical noise signals associated with production actions from production equipment, and to perform real-time verification of the physical noise signals based on a pre-established equipment noise fingerprint model to generate source verification for proving the authenticity of the production data source.

[0012] The privacy compliance proof module is configured to obtain the zero-knowledge compliance circuit converted from the preset process specifications, and generate a compliance proof that proves the production data meets the compliance circuit requirements without disclosing the content of the production data, based on the PCB production data and unique identifier.

[0013] The credential synthesis and on-chain module is configured to associate and integrate unique identity identifiers, source proofs, and compliance proofs to generate a self-certifying traceability credential for the PCB, and then upload the credential to the blockchain.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] First, this invention achieves a quantum-level strong binding between the physical entity of the PCB and its digital identity, fundamentally eliminating tag replacement attacks. This invention upgrades existing Automated Optical Inspection (AOI) or Solder Paste Inspection (SPI) equipment on SMT production lines into physical roots of trust in a seamless manner. It utilizes high-resolution cameras to extract the microscopic texture of the bare copper area of ​​the PCB as a Physically Unclonable Function (PUF) feature, generating a unique identity identifier. This identifier grows directly onto the PCB itself, requiring no artificially attached tags or markers, and possesses the physical characteristics of being unremovable and unclonable. Any attack attempting to forge an identity by replacing tags will fail due to the inability to reproduce the microscopic texture completely identical to the original PCB, resulting in verification failure. Simultaneously, this invention also provides a linkage mechanism between the physical destruction of scrapped PCB PUFs and on-chain identity revocation, ensuring that the digital identity of scrapped PCBs is not maliciously reused, further improving the full lifecycle security management of identity identifiers.

[0016] Secondly, this invention solves the source data trust problem by using a non-invasive approach to the existing control system of the equipment. The invention creatively shifts the focus from the equipment's digital control system to the weak process noise signals inevitably generated by the physical equipment during process operations, utilizing the uniqueness and unpredictability of these signals as the equipment's "physical fingerprint." Through a Trusted Execution Environment (TEE) deployed near the equipment, high-frequency physical noise signals associated with process operations (such as vacuum pressure fluctuations in a pick-and-place machine nozzle and heating current ripples in a reflow oven) are collected. An automatic encoder model is used to verify these signals in real time. After successful verification, the equipment's private key is used to sign the production data hash to generate source proof. This solution requires no circuit modifications, firmware upgrades, or communication protocol changes to the control system of legacy equipment. It only requires the non-invasive installation of external sensors around fluid lines or cables to endow legacy equipment with the ability to "self-certify the data source." Furthermore, this invention solves the trust breakpoint problem in the cold start training phase of the model by introducing an OEM gold template mechanism and a constrained fine-tuning strategy; it enhances robustness against physical injection attacks by training with adversarial samples and quantitatively defining the rejection boundary; and it ensures the long-term availability of the system during the entire lifecycle maintenance of the production line by the fingerprint re-registration and trust migration process after the replacement of key equipment components.

[0017] Third, at the mathematical level, both process compliance verification and core parameter privacy protection are achieved simultaneously. This invention reconstructs the entire SMT process traceability as a "trustless zero-knowledge compliance verification" problem. Contract manufacturers can, without disclosing specific temperature profile settings, placement coordinates, stencil opening sizes, or other core process parameters to brand owners, simply submit a concise zero-knowledge proof to mathematically assure them that all key process parameters for the entire batch of products strictly fall within their designated process window. Furthermore, by incorporating equipment source proofs into the public input and constraint verification scope of the zero-knowledge compliance circuit using BLS aggregated signatures, this invention extends the security guarantee for the proof generation environment from a single centralized server TEE to a distributed device TEE cluster, significantly improving the overall system's anti-attack capability. In post-event traceability scenarios, the system supports progressive fine-grained disclosure through a pre-set general query circuit template library, only disclosing controlled, limited, and auditable minimum data in the final step where mathematical proofs cannot satisfy the depth of analysis, maximizing the protection of the contract manufacturer's process secrets.

[0018] Fourth, this invention achieves pipelined parallelism between proof generation and the production process in terms of time. Through a hierarchical proof generation architecture, the production batch is divided into several sub-batches. Each sub-batch generates its corresponding sub-compliance proof immediately after production is completed, ultimately aggregating into a unified batch-level proof. This architecture allows the proof generation process to highly overlap with the continuous operation of the production line, effectively alleviating the contradiction between the computational time of zero-knowledge proofs and the just-in-time production cycle, making the system feasible for practical deployment in high-cycle manufacturing scenarios such as automotive electronics.

[0019] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the overall architecture of the blockchain-based SMT production data traceability system proposed in this invention.

[0021] Figure 2 Workflow diagram for the physical identity registration module;

[0022] Figure 3 Workflow diagram for the equipment certification module;

[0023] Figure 4 Workflow diagram for the privacy compliance certification module. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0026] Example 1

[0027] This embodiment discloses a blockchain-based SMT production data traceability system. By explaining the overall technical solution of the present invention from the system level, it illustrates the functional positioning of each core module, the collaborative working relationship between them, and the innovative technical concept on which the present invention is based.

[0028] In cross-enterprise traceability scenarios for SMT production, existing technologies face three fundamentally intertwined challenges. First, the binding between the physical PCB and its digital record relies on manually affixed barcodes or RFID tags. These external markers are easily damaged and replaced, resulting in a fragile physical-digital binding relationship. Second, the authenticity of production data cannot be effectively guaranteed before it is stored on the blockchain—older equipment lacks secure hardware, allowing data to be easily tampered with at the data acquisition gateway. Blockchain can only guarantee the immutability of data after it is uploaded, but it cannot verify the authenticity of data before it is uploaded. Third, there is an irreconcilable contradiction between protecting commercial privacy and verifying process compliance—brand owners need auditable process compliance guarantees, while contract manufacturers must protect their core process parameters from being leaked.

[0029] The core innovation of this invention lies in completely abandoning the traditional paradigms that rely on manually attached identifiers, hardware modifications to production equipment, and direct on-chain storage of raw production data or its hashes. Instead, it creatively utilizes the inherent, difficult-to-clone physical randomness of the SMT production process to establish unique and verifiable physical identity foundations for both the manufactured PCB and the production equipment involved. Based on this, it introduces zero-knowledge proofs, a cryptographic tool, transforming the process compliance audit, which originally required publicly available data, into a mathematically verifiable proof process that can be conducted without disclosing any data content. Ultimately, the physical identity of the "object," the data source proof for the "machine," and the privacy compliance proof for the "law" are integrated on the blockchain, generating a lightweight, self-proving digital credential for each PCB, achieving end-to-end trusted traceability from the physical world to the digital world.

[0030] 1. System overall architecture and module division;

[0031] The system of this invention is functionally divided into four core modules that work together. The specific implementation of each module is described in detail in Embodiments 2 to 4 below. It should be noted that the above division of modules is functional and logical. In actual deployment, these modules can be deployed on one or more physical entities, or they can be integrated, in whole or in part, as software functional components on edge computing nodes, servers, or the cloud.

[0032] (1) Physical identity registration module;

[0033] This module is configured to acquire PCB micro-texture images collected by optical inspection equipment (such as AOI or SPI equipment) in the SMT production line, extract physically unclonable function (PUF) features, and generate a unique identifier for the PCB. The core technical solution of this module is described in detail in Embodiment 2.

[0034] (2) Equipment certification module;

[0035] This module is configured to use a Trusted Execution Environment (TEE) to collect physical noise signals associated with production actions from production equipment, and to perform real-time verification of the physical noise signals based on a pre-established equipment noise fingerprint model to generate source proof to demonstrate the authenticity of the production data source. The core technical solution of this module is described in detail in Embodiment 3.

[0036] (3) Privacy compliance certification module;

[0037] This module is configured to acquire a zero-knowledge compliance circuit converted from a preset process specification, and based on the PCB production data and the unique identifier, generate a compliance certificate proving that the production data meets the requirements of the compliance circuit without disclosing the content of the production data. The core technical solution of this module is described in detail in Embodiment 4.

[0038] (4) Voucher synthesis and on-chain module;

[0039] This module is configured to associate and integrate the unique identifier, the source proof, and the compliance proof to generate a self-certifying traceability credential for the PCB, and then upload this credential to the blockchain. The collaborative working principle of this module is demonstrated in detail in the comprehensive application example of Embodiment 5.

[0040] 2. System workflow overview;

[0041] During the operation of the SMT production line, the various modules work together as follows.

[0042] First, in the initial process (usually before solder paste printing), the Physical Identity Registration (PUF) module uses optical inspection equipment to perform high-resolution imaging of pre-selected bare copper regions of interest on the PCB, extracting the PUF features of the micro-texture and generating a unique identifier. The module registers the hash value of the PCB on the blockchain, completing the "digital birth registration" of the PCB. At the same time, the module is also responsible for managing the entire lifecycle of the PCB's digital identity, including triggering the on-chain identity revocation process when the PCB is determined to be scrapped, recording the scrapping time and reason associated with the identity, and linking with the production line's defective product handling system to ensure that the on-chain digital identity and the physical entity's existence state are always consistent.

[0043] Subsequently, at each process involving critical parameters (such as solder paste printing, surface mount technology, reflow soldering, and optical inspection), the equipment verification module, through an edge TEE deployed near the equipment, synchronously collects physical noise signals associated with the process actions of that process. It then uses a pre-trained autoencoder model to perform real-time verification of the noise signals. Upon successful verification, the device's private key is used to sign the production data hash, generating a proof of origin. The proof cryptographically asserts that "this production data was indeed generated in real time by this specific physical device."

[0044] Once the entire batch production is complete, the credential synthesis and on-chain module first constructs a batch identity Merkle tree using the unique identifiers of all PCBs within the batch, obtaining the root of the tree. Subsequently, the privacy compliance proof module uses the root of the tree as public input and the entire production data (including the source proof corresponding to each data point) as private input, based on the zero-knowledge compliance circuit pre-compiled and delivered by the brand owner, to generate a batch-level compliance proof. This proof assures the verifier that for each PCB in the batch, all key process parameters strictly fall within the process window specified by the brand, and all data comes from certified production equipment, while the verifier cannot obtain any additional information about the specific process parameter values ​​during the process.

[0045] Finally, the credential synthesis and on-chain module will The batch identity Merkle tree root, the aggregated source proof, and the final compliance proof are associated and fused to generate a self-certifying traceability credential for the PCB, which is then submitted to a blockchain smart contract. After the contract completes verification, the self-certifying traceability credential is permanently recorded and can be independently verified and selectively traced by any authorized party in the future.

[0046] In the above, the physical identity registration module "seamlessly" upgrades the AOI / SPI camera originally used for quality inspection into a physical root of trust. Without adding extra tags or additional hardware costs, it achieves a "quantum-level" strong binding between the physical entity of the PCB and its digital identity, fundamentally eliminating tag replacement attacks.

[0047] The equipment verification module creatively transforms the inherent "process defect noise" of the equipment's physical body into "physical trust assets." In a non-intrusive control method that does not interfere with the original control system of the equipment, it enables even old equipment to output cryptographically verifiable source proofs, solving the long-standing "first mile" problem in the field of industrial data trust.

[0048] The privacy compliance proof module uses zero-knowledge proofs to revolutionize the SMT end-to-end traceability into "trustless zero-knowledge compliance verification," enabling business privacy protection and cross-enterprise quality traceability to be perfectly achieved simultaneously for the first time.

[0049] Ultimately, on the blockchain, a digital twin of a PCB is represented simply as a self-proving credential. Any authorized party that verifies this credential can mathematically confirm that a physically unique and irreplaceable circuit board in the world was produced by a designated group of real equipment, fully meeting process requirements while keeping parameters confidential. This completely transcends the existing "hash-based notarization" paradigm, constituting an end-to-end "self-proving manufacturing" system.

[0050] Example 2

[0051] This embodiment details a specific implementation of the physical identity registration module. The core function of this module is to extract a unique identifier from the inherent, microscopic, random physical structural features of the printed circuit board (PCB) using optical inspection equipment on the SMT production line, without relying on any manually added, easily damaged, or replaceable external identifiers such as barcodes or RFID tags. This unique identifier is referred to as the unique identifier throughout this document. Once generated, this unique identifier forms a physically strong binding relationship with the PCB. Any attempt to forge the PCB by replacing the tag will fail because the microscopic texture cannot be perfectly replicated from the original PCB, resulting in verification failure. The following sections will detail the selection of the Region of Interest (ROI), protection strategies, image acquisition, feature extraction, fuzzy extractor construction, and entropy security verification.

[0052] 1. Strategies for selecting and protecting regions of interest;

[0053] Unlike traditional solutions that rely on manual labeling, the physical root of trust in this invention grows directly within the PCB itself. To ensure the effectiveness and accessibility of this root of trust throughout its entire lifecycle, the selection of the Region of Interest (ROI) must adhere to strict anti-process interference criteria. In actual SMT production environments, after PCB mounting and reflow soldering, it typically undergoes multiple subsequent processes such as wave soldering, selective soldering, cleaning, and conformal coating. If the ROI is randomly selected at any location on the board, it is easily covered or contaminated by solder splatter, flux residue, conformal coating, or dust particles in subsequent processes, making clear imaging impossible during traceability. Therefore, the selection of the ROI is not random but requires forward-looking planning during the PCB design phase. The fundamental starting point of this solution is to establish a trusted identity link throughout the entire lifecycle from the source of PCB manufacturing to final after-sales traceability; therefore, the ROI must be a carrier that can maintain its physical texture unchanged throughout all processes.

[0054] In a preferred embodiment, the selected ROI is located in a pre-reserved non-functional area inside the PCB board, rather than a process edge. Although the process edge is removed during the milling process after SMT completion, thus naturally avoiding contamination in subsequent processes, this milling action makes the process edge no longer exist in the after-sales rework traceability stage, and cannot provide a physical carrier for full lifecycle traceability. Therefore, this embodiment explicitly selects the in-board ROI as the physical carrier area for unique identification.

[0055] Specifically, the ROI within this board is an "optical identification area" intentionally placed during the PCB routing design phase. Its surface copper foil is designed as a bare copper test point—uncovered by solder resist ink, independent, and not connected to any electrical network. The size of this test point can be designed according to the field of view of the optical inspection equipment, for example, a 1 mm by 1 mm square bare copper area. The reason for choosing a bare copper area is that, during the electroplating and etching processes, due to the randomness of the electrochemical crystallization process, microscale grains, grain boundaries, micro-protrusions, and depressions form on the surface of the copper foil. These texture features are on the submicron to micrometer scale and are random structures naturally generated during the manufacturing process. They cannot be replicated by conventional printing or machining methods and possess a natural physical non-cloning function characteristic.

[0056] The most severe challenge facing bare copper ROI (Return on In-Board) is the peak temperature of 240°C to 260°C during the subsequent lead-free reflow soldering process. In such a high-temperature, oxygen-rich environment, the exposed copper surface will inevitably undergo a violent thermal oxidation reaction, generating a composite layer of cuprous oxide and copper oxide with a thickness of tens to hundreds of nanometers. This oxide layer physically and fundamentally alters the original metal grain morphology and nanoscale roughness of the copper foil surface, causing an irreversible chemical phase transition. It is important to clarify that this chemical phase transition is not an independent random bit-flipping error in a communication channel, but a thermodynamically driven, systematic, deterministic physicochemical process. Oxidation leads to a global biased drift in the surface roughness, grain boundary scattering characteristics, and light reflectivity of the copper foil, resulting in a structural overall shift in the texture feature vector extracted by the Gabor filter. This systematic bias far exceeds the error-correcting capability of BCH (Browser-based Chromium Error Correction) codes, which assume independent random errors; therefore, error-correcting codes cannot be used to "correct" the chemical phase transition. Therefore, a protection scheme that can fundamentally prevent high-temperature oxidation must be adopted, rather than relying on error correction codes to tolerate oxidation.

[0057] This embodiment provides differentiated protection schemes depending on whether the SMT production line is equipped with a full-process nitrogen protection system.

[0058] Option 1 (Full-Process Nitrogen Protection Production Line, Preferred Solution): When the SMT production line is equipped with a full-process high-purity nitrogen protection system from the reflow oven inlet to the outlet cooling zone, the oxygen concentration in the soldering zone can be maintained below 50ppm throughout the process. A nitrogen curtain is installed at the reflow oven outlet to continuously purge the PCBs with nitrogen until their surface temperature drops below 200°C. Under these conditions, the bare copper ROI is under inert atmosphere protection throughout the entire process from entering the furnace to cooling to a safe temperature. No thermal oxidation reaction occurs on the copper surface, and its original metal grain morphology and nanoscale roughness are completely preserved. Therefore, in a full-process nitrogen protection production line, the bare copper ROI does not require any physical covering throughout the process, greatly simplifying the process. Only before subsequent processes that may come into contact with flux splashes or chemicals (such as selective wave soldering and water-based cleaning) is a temporary protective film that can be peeled off without residue applied by an automated placement device. This temporary protective film uses a UV-resistant, low-adhesion polyolefin film with a substrate thickness of 30 micrometers. It has a temperature resistance rating of 260°C for short-term exposure (30 seconds) and 150°C for long-term exposure, and exhibits good chemical resistance, resisting corrosion from flux and cleaning agents. Before UV irradiation, the film has reliable adhesion. During traceability testing, after irradiation with UV light of a specific wavelength (e.g., 365 nm) for 10 to 15 seconds, the adhesive layer undergoes a chemical cross-linking reaction, causing the adhesion to drop sharply to less than 5% of its initial value. This allows for clean, residue-free removal, exposing the well-preserved original bare copper surface.

[0059] Special treatment for the conformal coating process: When the PCB needs to undergo conformal coating, the edge sealing performance of the UV-resistant polyolefin film is crucial. In this embodiment, the area of ​​the temporary protective film is slightly larger than the ROI area (at least 2 mm beyond each side), and the placement pressure of the automatic placement equipment is controlled (approximately 0.5 Newtons) during application to ensure a tight contact between the adhesive layer and the copper surface, effectively preventing conformal coating mist from seeping in along the edges. To further improve sealing reliability, a ring of room-temperature curing silicone sealant approximately 1 mm wide can be applied along its edge after the protective film is applied. This sealant remains elastic during the conformal coating and curing process, and when the protective film is removed during the traceability process, the sealant peels off along with the film, leaving no residue on the copper surface. Simultaneously, in the subsequent image preprocessing workflow, a background subtraction algorithm based on grayscale histogram normalization is introduced to compensate for any possible trace amounts of chemical residue causing image grayscale shifts, ensuring the stability of feature extraction.

[0060] Option 2 (Non-full-process nitrogen protection production line, lossy PUF solution): When the production line lacks full-process nitrogen protection (e.g., only nitrogen protection inside the reflow oven, with no nitrogen curtain at the outlet), the bare copper ROI will inevitably be exposed to oxygen-containing air during the reflow oven cooling stage. Under these conditions, the high-temperature copper surface will oxidize, and as mentioned earlier, the feature drift caused by this oxidation is systematic and cannot be corrected by BCH error correction codes. Therefore, in non-nitrogen-protected production lines, this solution abandons the technical route of relying on error correction codes to "tolerate" oxidation, and instead adopts a degradation tolerance strategy based on redundant voting of multiple ROIs.

[0061] Specifically, during the PCB routing design phase, at least four spatially dispersed bare copper Regions of Interest (ROIs) are reserved on the board surface (located in the four corner areas of the PCB). During the physical identity registration phase, features are independently extracted from all four ROIs, and auxiliary data is generated. Similarly, during the reconstruction phase, images are independently acquired from the four ROIs, and an attempt is made to reconstruct their identities. Because the spatial positions of each ROI on the PCB differ, their local cooling rates and oxygen exposure levels during the reflow soldering cooling phase vary, resulting in varying degrees of oxidation. Some ROIs may fail to decode due to severe oxidation, but typically at least one ROI, located at a faster-cooling board edge, will have relatively less oxidation, and its texture features will still fall within the tolerance radius of the fuzz extractor, allowing for successful reconstruction. During the reconstruction phase, the texture images acquired from each ROI are reconstructed separately to obtain candidate regions for each area. The system considers four candidates. Conduct voting: based on at least one candidate With on-chain evidence If they match, the identity verification is considered successful. If all candidates... If decoding fails, the physical identity of the PCB is deemed unrecoverable. Further confirmation is required through other traceability methods (such as manual verification of production batch records). This solution does not claim to completely eliminate the impact of oxidation, but rather uses spatial redundancy design to control oxidation risk within an acceptable range. This solution is suitable for applications with "best-effort" traceability integrity requirements, rather than zero-tolerance scenarios such as aerospace.

[0062] To assist optical inspection equipment in quickly and accurately locating the ROI during each imaging iteration, auxiliary positioning markers can be designed around the bare copper test point. For example, four circular pads with a diameter of one millimeter can be placed at the four corners of the test point as reference markers, or a closed ground copper ring can be designed around the test point. These auxiliary markers provide a reliable reference coordinate system for template matching and affine transformation correction in subsequent image processing.

[0063] 2. Acquisition and preprocessing of micro-texture images;

[0064] When the PCB to be processed arrives at the optical inspection equipment (such as AOI or SPI equipment) station of the first process (usually before solder paste printing) on ​​the SMT production line via the transport track, the physical identity registration module is triggered, controlling the high-resolution industrial camera at that station to image at least two pre-set ROIs. The reason for selecting at least two ROIs is to introduce spatial redundancy and prevent the PCB's permanent identity from becoming invalid due to irreversible physical damage (such as scratches) to a single ROI area under extreme conditions. These two ROIs are typically located diagonally opposite each other on the PCB, or at two diagonal positions out of the four corners. In a preferred embodiment of the invention, the AOI or SPI equipment itself is also included in the monitoring scope of the device verification module described in Embodiment 3. Its imaging action also generates source verification, thereby ensuring that the image data collected during the registration stage itself has source credibility and eliminating the trust breakpoint where the imaging in the first process might be tampered with.

[0065] The configuration of the optical system is crucial for clear imaging of micro-textures. In this embodiment, the industrial camera employs a 10-megapixel monochrome CMOS image sensor with a pixel size of 1.67 micrometers by 1.67 micrometers. A telecentric lens with a magnification of 10x is selected. A key advantage of a telecentric lens is that the magnification remains constant within a certain range of object distance variations, effectively eliminating dimensional errors caused by slight warping or placement height deviations of the PCB. The illumination source uses a low-angle ring-shaped blue light-emitting diode (LED) with a dominant wavelength of approximately 470 nanometers. Low-angle illumination means that the light source grazes the copper foil surface at an angle nearly parallel to the PCB plane. This illumination method produces significant light and shadow on the micro-grains, grain boundaries, micro-protrusions, and depressions of the copper foil surface, maximizing texture contrast while suppressing direct reflected light from a flat surface.

[0066] The original grayscale image obtained is denoted as The image has a bit depth of 8 bits and a resolution of 3856 x 2764 pixels. The raw image is first sent to the computing unit in the physical identity registration module for preprocessing. The computing unit can be an industrial control computer embedded within the optical inspection equipment or a nearby edge computing node. The first step in the preprocessing process is to use a Gaussian low-pass filter... Smoothing is performed to suppress random thermal noise that may be generated by the CMOS sensor during short exposure times. The Gaussian filter kernel size is 5 x 5 pixels, and the standard deviation σ is 1.0 pixel. This parameter setting can remove small random noise while preserving high-frequency details of micro-textures. The second step is the location and extraction of the Region of Interest (ROI). Using the four pre-designed circular reference markers around the copper foil test points, a template matching algorithm based on normalized cross-correlation is used to accurately locate the coordinates of these four markers in the filtered image. Based on these four coordinates, the affine transformation matrix is ​​calculated, and rotation and translation corrections are performed on the image to obtain the ROI image with normalized position and angle. After affine transformation, The image is uniformly cropped and resampled to a standard size of 1024 x 1024 pixels to ensure that subsequent feature extraction algorithms always operate on the same scale of input. After the PCB has undergone conformal coating and protective film removal, an additional grayscale histogram normalization process is performed to obtain a compensated image. This compensates for the grayscale shift caused by possible trace amounts of chemical residue, mapping the grayscale distribution of the current image to the grayscale distribution of the reference image at registration.

[0067] 3. Feature vector extraction based on Gabor filter banks;

[0068] In order to Texture feature vectors that possess both high discriminative power and robustness to illumination changes and minor positioning errors are extracted. This embodiment employs a multi-scale, multi-directional Gabor filter bank. The mathematical expression of the kernel function of a two-dimensional Gabor filter in the spatial domain is defined as follows:

[0069] ;

[0070] in, , The parameter λ represents the wavelength of the sinusoidal stripe in pixels; the parameter θ represents the texture direction detected by the filter; the parameter ψ represents the phase shift; the parameter σ represents the standard deviation of the Gaussian envelope; and the parameter γ represents the spatial aspect ratio.

[0071] In this embodiment, to comprehensively capture the texture information of the copper foil surface at different scales and in different directions, a Gabor filter bank with four scales and six directions was designed. The wavelength λ is set to 4 pixels, 8 pixels, 16 pixels, and 32 pixels, respectively; the direction θ is set to 0°, 30°, 60°, 90°, 120°, and 150°, respectively; and the phase offset ψ is uniformly set to 0 (even symmetry). γ = 0.5. This filter bank consists of twenty-four Gabor filter cores.

[0072] Normalized ROI image Each of the twenty-four response maps is subjected to a two-dimensional discrete convolution with the aforementioned twenty-four filter kernels. The arithmetic mean and standard deviation of all pixel values ​​are calculated for each response map. The twenty-four response maps collectively generate a forty-eight-dimensional real-valued feature vector. This vector encodes the texture structure information of the ROI at multiple scales and in multiple directions in a compact numerical form.

[0073] In another alternative embodiment, a deep learning model can also be used to extract the feature vector. For example, a pre-trained convolutional neural network (such as the first few layers of VGG-16) can be used as a feature extractor. The normalized ROI image is input into the network, and the output of the fully connected layers is extracted as the texture feature vector. This approach may have stronger generalization ability when dealing with complex and varied texture patterns.

[0074] 4. A concrete example of constructing a fuzzy extractor—from fuzzy features to stable identity;

[0075] Although the feature extraction process is robust, unavoidable factors such as slight fluctuations in lighting conditions, image sensor noise, and minute errors in ROI localization can lead to variations in the number of features extracted from two independent imaging sessions for the same ROI. There are slight differences between them, making it impossible to directly input them into a cryptographic hash function. This embodiment uses a fuzzy extractor to stably extract repeatable binary identity identifiers with uniform entropy values ​​from fuzzy features.

[0076] (4.1) Binarization and quantization of eigenvectors

[0077] First, generate a fixed distribution that follows a standard normal distribution. random projection matrix The eigendimensional dimension of the vector space is d = 1023. P is generated by a cryptographically secure pseudo-random number generator during system initialization and remains fixed as a common parameter. d = 1023 is chosen because it is... The original length can be directly concatenated with subsequent BCH codes without padding or truncation. Calculation ,in Perform symbolic function binarization on each component:

[0078] For ;

[0079] This generates a 1023-bit original binary feature string. .

[0080] The aforementioned binarization scheme based on random projection is theoretically based on the Johnson-Lindenstrauss lemma in random projection theory and its generalization in the field of binary embedding. This lemma states that a set of points in a high-dimensional space can be mapped to a lower-dimensional space through random projection, while maintaining the Euclidean distance between point pairs with a high probability. In this embodiment, the random projection matrix P maps a 48-dimensional Gabor feature vector to a 1023-dimensional binary Hamming space. Although the target dimension (1023 dimensions) is higher than the original dimension (48 dimensions), the nonlinear transformation of the binarization operation makes this mapping actually constitute a locally preserved binary embedding. This means that the Hamming distance between two PCBs with similar textures in the original feature space (such as two independent images of the same ROI) after mapping to 1023-bit binary strings will be significantly smaller than the Hamming distance between two PCBs with dissimilar textures. This property has been verified in the system design phase through large-scale experiments on 600 PCB samples: the average Hamming distance of binary strings repeatedly imaged within the same ROI is approximately 30 to 50 bits (i.e., a bit error rate of 3% to 5%), while the average Hamming distance of binary strings between different PCBs is approximately 480 to 530 bits (close to a randomness level of 50%), with a clear distinguishable boundary between the two. These experimental results provide ample empirical support for the applicability of the random projection binarization scheme to PCB texture features.

[0081] (4.2) Selection and parameter determination of error correction codes;

[0082] The original binary feature string w obtained after quantization of the same ROI in two imaging operations is compared with the new 1023-bit binary feature string w' generated after re-imaging the same ROI in the subsequent reconstruction stage and repeatedly performing the same Gaussian filtering noise reduction, template matching localization, affine transformation correction, cropping resampling, Gabor filter bank feature extraction, random projection, and symbol function binarization process (the complete processing flow of this reconstruction stage is detailed in section (4.4) of this section). There is a bit flipping error between the two. Here, it is necessary to strictly limit the applicable scope of this error correction code scheme: under the conditions of Scheme 1 (full-process nitrogen protection production line), the copper foil surface does not undergo high-temperature oxidation, and the feature differences between the two imaging operations mainly come from random factors such as imaging noise, small positioning errors, and dust particle obstruction. According to a large number of repeated imaging experiments, under these conditions, the average bit error rate can be stably controlled between three and five percent, and the errors are sudden (caused by factors such as local trace residues and dust particle obstruction). The essence of these errors is independent and random physical interference events, which is consistent with the mathematical assumptions of error correction codes.

[0083] To resolve the contradiction between burst errors and BCH codes' ability to correct random errors, this embodiment introduces a block interleaver (33 rows × 31 columns, 1023 elements) to discretize burst errors into an approximately random distribution. This embodiment uses primitive BCH codes. Based on Galois domain Using primitive polynomials The code length is n = 2047 bits, the information bits are k = 76 bits, and the designed error correction capability is t = 113 bits. The interleaved feature string... Zero-padding extension As input, this code provides up to 113 bits of global error correction capability for the 1023 bits carrying texture information and its surrounding padding region, which is sufficient to cover the extreme scenario of approximately 112 bits of independent random errors at an 11% bit error rate.

[0084] It needs to be emphasized again that this error correction code scheme is not applicable to systematic feature drift caused by high-temperature oxidation of copper foil in non-nitrogen-protected production lines. This scenario is handled by the aforementioned multi-ROI redundant voting scheme.

[0085] (4.3) Registration phase: Generate unique identifiers and auxiliary data;

[0086] In a secure registration environment, a hardware random number generator is used to generate a 76-bit random seed s. The system encoding algorithm is then invoked to obtain the codeword. Calculate auxiliary data:

[0087] ;

[0088] Auxiliary data (SS) can be stored publicly. Calculate a unique identifier:

[0089] ;

[0090] Store the auxiliary data SS off-chain, and Submitted to the blockchain as the PCB's "digital birth certificate".

[0091] (4.4) Reconstruction Phase: Restoring the unique identity identifier;

[0092] When the same PCB needs to be identified again (such as in after-sales traceability), the same ROI of the PCB is imaged again. After the same image preprocessing, Gabor filtering, random projection, and sign function binarization are performed, a new 1023-bit binary feature string is generated. After interleaving and zero-padding, the resulting 1023-bit binary feature is obtained. Read SS and calculate the noisy codeword c':

[0093] ;

[0094] Perform BCH decoding on the noisy codeword c' to recover the codeword c, extract the first 76 bits to obtain a 76-bit random seed s, and calculate... Identity is verified through on-chain hash comparison.

[0095] 5. Experimental estimation and safety demonstration of physical characteristic entropy values;

[0096] To demonstrate that the 76-bit seed length has a sufficient safety margin, a rigorous entropy estimation experiment was conducted. 200 bare copper test boards with FR-4 substrate were randomly selected from each of three mainstream PCB manufacturers (600 boards in total). Using an optical system consistent with the production line's AOI, 10 independent images were acquired for the same ROI on each board (each reloading and unloading introduced mechanical positioning errors), resulting in a total of 6000 images. After feature extraction and binarization using the aforementioned method, the minimum entropy estimation method for non-independent and identically distributed data in the NIST SP 800-90B standard was employed to divide the 1023 bits into 93 11-bit substrings. A counting-based estimator was used, with a 95% confidence lower bound.

[0097] Experimental results show that the minimum entropy estimates for 600 samples range from 82.3 bits to 97.6 bits, with a mean of 89.1 bits and a standard deviation of 3.4 bits. The lower bound of the global minimum entropy (95% confidence level) is 84.5 bits, significantly exceeding 76 bits, preserving a safety margin of approximately 8.5 bits (approximately 362 times). This margin is sufficient to cover factors such as differences in manufacturing processes between different manufacturers and entropy degradation caused by submicroscopic physical wear on the copper foil surface after long-term use. Therefore, attackers need to search for... Space is computationally infeasible, and the construction of the fuzz extractor is supported by solid experimental data on both the information theory and computational security levels.

[0098] 6. On-chain revocation mechanism for invalid digital identities;

[0099] In actual SMT production, the first batch of PCBs may be deemed scrapped due to printing defects, component misalignment, or reflow tombstoning, with a scrap rate typically between 10% and 15%. Although these scrapped PCBs are physically discarded, their digital identity on the blockchain remains intact. This information is permanently recorded during registration. Without an effective identity revocation mechanism, attackers could collect discarded PCBs and use their ROI to reconstruct valid ones. They then attach these materials to counterfeit products, thereby enabling traceability verification downstream in the supply chain.

[0100] To address this security risk, this embodiment introduces a linkage mechanism between "physical destruction of PUF" and "on-chain identity revocation". When the automated optical inspection equipment or manual quality inspection on the production line determines that a PCB is an irreparable waste board, the physical identity registration module receives a PCB scrap signal from the Manufacturing Execution System (MES) and triggers the following operations.

[0101] First, the PUF is physically destroyed: a laser etching unit located behind the AOI station etches a destructive mark exceeding 50 micrometers in depth and 100 micrometers in width on the region of interest (ROI) of the scrapped PCB, physically and completely destroying the microstructure of the bare copper ROI. This destruction is irreversible, making it impossible for anyone to reconstruct the original PUF from the PCB afterward. .

[0102] Secondly, on-chain identity revocation: The credential synthesis and on-chain module constructs a blockchain transaction, calls the smart contract's identity revocation function, marks the unique identifier of the scrapped PCB as "scrapped - irrecoverable," and records the scrapping timestamp and scrapping reason code. Subsequently, when any supply chain node verifies this unique identifier, the smart contract will return a warning message: "This identity is associated with a scrapped PCB."

[0103] Through this linkage mechanism, the physical destruction of the discarded PUF and the cryptographic revocation of the on-chain identity constitute an atomic operation—the two are logically inseparable and completed simultaneously, ensuring that attackers cannot use the discarded PUF for identity reuse attacks.

[0104] Example 3

[0105] This embodiment details a specific implementation of the equipment verification module. This module aims to solve a fundamental problem that has long plagued the field of industrial data trust: how to provide production data output by these devices with a verifiable source authenticity certificate without interfering with the original control system of the production equipment and minimizing the impact on normal operation. The core innovation of this invention lies in shifting the focus from the digital control system of the equipment to the weak, unique, and unpredictable process noise signals inevitably generated by the physical subsystems such as the actuators, power systems, and fluid circuits of the equipment during normal process execution. These process noise signals are like the equipment's "physical voiceprint" or "heartbeat," and cannot be simulated or replayed by software-level attacks. The following sections elaborate on the discovery and acquisition of physical noise signals, the factory-designed golden template mechanism, the design of automatic encoders, adversarial samples and rejection boundaries, adaptive thresholds, proof generation, and fingerprint re-registration after equipment component replacement.

[0106] 1. The physical nature and reliable acquisition of process noise signals;

[0107] Taking the high-speed pick-and-place machine, the core component of an SMT production line with the highest frequency of operation, as an example, the nozzle is connected to a sophisticated pneumatic circuit. When picking up components, the solenoid valve opens to conduct vacuum, and the air pressure inside the nozzle rapidly drops from atmospheric pressure to a preset negative vacuum level within tens to hundreds of milliseconds. During this transient process, the actual pressure change inside the air circuit is not an ideally smooth curve, but rather a superposition of minute pulsations from the air supply, mechanical vibration and airflow impact at the moment the solenoid valve core opens, turbulence at the Laval nozzle inside the vacuum generator, local pressure loss and reflected waves generated at joints with different inner diameters in the pipeline, and random disturbances in the airflow boundary layer caused by differences in the micro-roughness of the pipeline inner wall. All these factors combined manifest as a highly detailed time-series waveform at the pressure sensor measurement point, containing a large number of high-frequency ripples, overshoot, and micro-amplitude random oscillations. The macroscopic profile of this waveform meets the process requirements, but its microscopic ripple details have extremely strong individual specificity and short-term repeatability for each specific nozzle.

[0108] It is crucial to emphasize the essential difference between "process noise signal" and "macroscopic process waveform" in this invention. The nozzles of different pick-and-place machines, and even different nozzles on the same machine, exhibit highly similar macroscopic envelope shapes for vacuum pressure changes. This is determined by the commonalities in the gas path design. If the autoencoder only learns these macroscopic common features, when an attacker uses the nozzle signal from any pick-and-place machine of the same model as a forged signal input, its reconstruction error is likely to be equally small, leading to successful deception. The key to this invention is that the autoencoder must be trained to capture the microscopic high-frequency differences superimposed on the macroscopic envelope, determined by the physical uniqueness of a specific device, rather than the macroscopic common features themselves. Therefore, in the data preprocessing stage, this embodiment performs high-pass filtering on the real-time acquired raw physical noise signal to separate the macroscopic process waveform and high-frequency residual signal, thereby separating the slowly changing macroscopic step envelope from the high-frequency microscopic ripple; only the high-frequency residual signal is fed into the autoencoder for training and verification. Thus, the model learns the unique fingerprint of the device's physical characteristics, rather than the common envelope possessed by any device of the same model.

[0109] Similarly, in a reflow oven, the heating system uses solid-state relays or thyristors to rapidly control the switching on and off of the resistance heating wire via pulse width modulation or frequency control. Collecting the current signal from the main circuit driving the heater reveals that the current waveform is not an ideal square wave. The slopes of its rising and falling edges, the subtle current changes superimposed during conduction due to the thermal inertia of the heating wire, and the periodic pulsations caused by grid voltage fluctuations and harmonic distortion collectively form the "electrical fingerprint" of this specific heating circuit.

[0110] A common and security-critical characteristic of these physical noise signals is that they are direct products of the real-time application of physical laws and inseparable byproducts of physical processes. Any attacker attempting to bypass the physical device itself and merely inject data packets over the network or generate waveforms using software cannot precisely construct a signal sequence that perfectly matches the real physical signal in its microscopic waveform details and is strictly synchronized with the process actions in time. Especially when attackers cannot obtain the specific parameters of the device's noise fingerprint model, the difficulty of forging such microscopic details is computationally infeasible.

[0111] To ensure reliable and secure acquisition of these signals, this embodiment clearly distinguishes between two levels of intervention in sensor access. For new-generation equipment with standard sensor pre-reserved interfaces, sensors are connected through the original factory-reserved test ports, without affecting equipment operation. For older equipment without pre-reserved sensor interfaces, this embodiment adopts an intervention method of "installation without interfering with the original control system of the production equipment," that is, it explicitly promises not to modify, intercept, or inject any modifications to the equipment's control system circuits, firmware, or communication protocols, but allows the installation of external sensors on the equipment's fluid lines (such as gas lines, coolant lines) or around the cables (such as through clamp-on current transformers). Taking the pick-and-place machine nozzle as an example, a piezoresistive high-frequency dynamic pressure sensor with an inherent response frequency exceeding 10kHz is selected and connected in series with the vacuum line connection of the target nozzle through a standard stainless steel tee connector. To assess the impact of this intervention on the equipment process, this embodiment strictly limits the total dead volume of the T-connector and the sensor's internal components to no more than 3% of the original piping volume during sensor selection. This ensures that the delay in vacuum build-up time does not exceed 2 milliseconds, far below the typical vacuum build-up time (50-100 milliseconds) tolerance range for pick-and-place machines. Actual production line verification shows that this minor delay has no statistically significant impact on placement yield. The sensor output signal is transmitted to the signal conditioning module via a shielded twisted-pair cable, amplified, and converted into a 4-20 mA industrial standard current signal before being connected to the high-speed analog-to-digital conversion channel of the Trusted Edge Execution Environment (TEE).

[0112] The edge trusted execution environment (ESI) is the security cornerstone of the entire device authentication module. Physically, it is a robust industrial-grade embedded computer whose core processor supports hardware-level trusted execution environment technologies such as Intel SGX or ARM TrustZone. Internally, it defines a secure area (enclave) completely isolated from the main operating system through hardware instruction sets. The code running and data processed within the enclave cannot be viewed or tampered with, even by the main operating system kernel or an attacker with the highest physical privileges. At the initial system deployment, the enclave utilizes an internal hardware true random number generator to generate asymmetric key pairs based on the Secp256k1 elliptic curve. Private key From the moment of its generation, its plaintext form is permanently sealed within the protected memory pages of the enclave, with no export interface. The public key paired with this private key... The data is then securely exported and registered on the blockchain via the credential synthesis and on-chain module, serving as a unique, unforgeable, distributed digital identity for the production equipment within the blockchain network. It should be noted that in the subsequent implementation of the BLS aggregate signature scheme in Embodiment 4, the elliptic curve here will be replaced with a BLS12-381 curve, and the signature algorithm will be changed accordingly, but the TEE key protection and proof generation processes will remain consistent.

[0113] 2. Factory-set gold template and constrained fine-tuning mechanism;

[0114] To address the trust gap caused by the inability to self-verify the purity of data during the training phase, this embodiment innovatively introduces a "factory physical fingerprint" mechanism, tracing the source of trust for the device fingerprint model back to the OEM manufacturer's trusted production environment.

[0115] Before the equipment leaves the factory, the OEM manufacturer, in its controlled testing workshop physically isolated from the public network, collects no fewer than 500 process noise signal segments for each actuator channel on the equipment included in this system's monitoring (such as the air path of each nozzle of a pick-and-place machine, or the heating circuit of each temperature zone of a reflow oven), under standardized operating conditions (such as a specific nozzle picking up a standard weight of test components at rated vacuum). For configurations with sensors pre-installed at the factory, the OEM can directly use the equipment's built-in sensors for data collection. For older equipment models without pre-installed sensors, the OEM temporarily installs external sensors and signal conditioning modules of the same model as those deployed in mass production before shipment. These are removed after the gold template data collection is completed, and the specifications and locations of the sensor installation interfaces are clearly marked in the equipment's technical documentation for the contract manufacturer to accurately reproduce and install the sensors during deployment. All segments are filtered by a 500Hz high-pass filter, retaining only the high-frequency residual signal. Each channel is independently trained using an autoencoder with the same architecture as in this embodiment to obtain the factory-delivered gold template model. (i.e., training to obtain the initial device noise fingerprint model).

[0116] OEM manufacturers use their code-signing private keys to verify the device's unique serial number, channel identifier, and... The SHA-256 hash value is used to digitally sign and generate a "device physical fingerprint birth certificate". (i.e., the model delivery certificate). This certificate is encoded in the standard X.509 attribute certificate format and is delivered to the OEM along with the device. At the same time, the OEM stores the certificate hash value in the IdentityCC chaincode of the consortium blockchain.

[0117] When deployed in a foundry, the edge TEE first reads from the digital storage medium attached to the device. The signature validity is verified using the OEM public key pre-stored in the TEE's secure storage. Once verification is successful, the TEE retrieves the signature from the certificate. The hash value is used as the initial model weights for the autoencoder, instead of starting training with random weights. During the subsequent online fine-tuning phase, the TEE has a mandatory constraint: the change in the L2 norm of the model weights during fine-tuning must not exceed [a certain value]. The initial weight is 5%. If a weight change exceeds this threshold, the TEE automatically reverts to the pre-fine-tuning checkpoint and records the security event. If several consecutive updates trigger the threshold alarm, the TEE will refuse to continue fine-tuning, retaining the initial device noise fingerprint model as the only valid model, and will send an alert to the blockchain network. Under normal circumstances, fine-tuning completes successfully, and the TEE obtains an online adapted noise fingerprint model. This mechanism fundamentally eliminates the possibility of attackers injecting forged signals through the sensor interface to train the model during the system's cold start phase.

[0118] It is important to clarify that this factory-installed gold template mechanism is based on the following trust assumptions: the OEM manufacturer is honest and trustworthy in the gold template collection and signing process before the equipment leaves the factory; the testing workshop used is physically isolated from the public network; the collected signals do indeed originate from the identified physical device channel; and the code signing private key has not been leaked or misused. This trust assumption is reasonable and universally accepted in industrial security practice—similarly, the security root of a Trusted Platform Module (TPM) chip relies on the endorsement key pre-installed by the chip manufacturer in a trusted production environment, and the security root of a Public Key Infrastructure (PKI) relies on the proper safekeeping of the root private key by the Certificate Authority. Those skilled in the art will understand that bootstrapping security from scratch without introducing any physical trust anchors is theoretically impossible. This invention establishes trust anchors at the OEM factory stage, extending the trust chain from the earliest trusted point in the entire device lifecycle, which is a recognized and reasonable design paradigm in the field of industrial security.

[0119] 3. Autoencoder network architecture and training configuration;

[0120] To accurately fit the distribution and detect anomalies in process noise signals, this embodiment designs a depth autoencoder (i.e., an autoencoder model) based on one-dimensional convolution, specifically optimized for industrial time-series signals. For the nozzle air pressure signal of the pick-and-place machine, the signal is segmented and resampled into 1024-point fixed-length segments, then filtered by a 500Hz high-pass filter and normalized before being fed into the network. The network architecture is as follows:

[0121] Input layer enter - - (1, 1024) One-dimensional convolution Core size = 7, step size = 2, number of channels = 16 ReLU (16, 512) One-dimensional convolution Core size = 5, step size = 2, number of channels = 32 ReLU (32, 256) One-dimensional convolution Core size = 3, step size = 2, number of channels = 64 ReLU (64, 128) Flatten - - (8192,) Fully connected Number of neurons = 20 linear (20,) Fully connected Number of neurons = 8192 ReLU (8192,) Shape transformation - - (64, 128) One-dimensional transposed convolution Core size = 3, step size = 2, number of channels = 32 ReLU (32, 256) One-dimensional transposed convolution Core size = 5, step size = 2, number of channels = 16 ReLU (16, 512) One-dimensional transposed convolution Core size = 7, step size = 2, number of channels = 1 linear (1, 1024)

[0122] The bottleneck layer has only 20 neurons, which means that the network is forced to compress the rich information of 1024 sampling points into an extremely low-dimensional latent representation with only 20 floating-point values. This extremely high compression ratio is the key to ensuring that the autoencoder can capture global normal patterns rather than simple identity mappings.

[0123] The loss function is defined as the mean square error between the input high-frequency residual signal x and the reconstructed signal x̂:

[0124] ;

[0125] Training was performed entirely within the enclave, using the Adam optimizer (initial learning rate 0.001, batch size 64). After training, the model parameters... Sealed and stored, its hash value On-chain evidence storage.

[0126] After training, the autoencoder is used for inference within the TEE as follows. For each newly acquired high-frequency residual signal segment x (a normalized one-dimensional vector of length 1024), the TEE performs the following inference process:

[0127] Step 1 (Encoder - First Convolutional Layer): Perform a one-dimensional convolution operation on the input x with a kernel size of 7, a stride of 2, and 16 output channels to obtain an intermediate tensor h1 with a shape of (16, 512); apply the ReLU activation function element by element to h1.

[0128] The second step (encoder - second convolutional layer): Perform a one-dimensional convolution operation on h1 with a kernel size of 5, a stride of 2, and 32 output channels to obtain an intermediate tensor h2 with a shape of (32, 256); apply the ReLU activation function element by element to h2.

[0129] Step 3 (Encoder - Third Convolutional Layer): Perform a one-dimensional convolution operation on h2 with a kernel size of 3, a stride of 2, and 64 output channels to obtain an intermediate tensor h3 with a shape of (64, 128); apply the ReLU activation function to h3 element by element.

[0130] Step 4 (bottleneck layer compression): Expand h3 into a one-dimensional vector (8192 elements in total), and map it to a 20-dimensional latent representation vector z through a fully connected layer.

[0131] Step 5 (Decoder - First Fully Connected Layer): Map the 20-dimensional z back to an 8192-dimensional vector through the fully connected layer, apply the ReLU activation function element by element, and reshape it into a tensor of shape (64, 128).

[0132] Step 6 (Decoder - First Transposed Convolutional Layer): Perform a one-dimensional transposed convolution operation with a kernel size of 3, a stride of 2, and 32 output channels on the output of Step 5. Apply the ReLU activation function element by element to obtain a tensor with shape (32, 256).

[0133] Step 7 (Decoder - Second Transposed Convolutional Layer): Perform a one-dimensional transposed convolution operation with a kernel size of 5, a stride of 2, and 16 output channels on the output of Step 6. Apply the ReLU activation function element by element to obtain a tensor with shape (16, 512).

[0134] Step 8 (Decoder - Third Transposed Convolutional Layer): Perform a one-dimensional transposed convolution operation on the output of Step 7 with a kernel size of 7, a stride of 2, and an output channel number of 1 to obtain the reconstructed signal x̂, which has a shape of (1, 1024).

[0135] Step 9 (Reconstruction Error Calculation): Calculate the mean square error between x and x̂. .

[0136] In the above inference process, all convolutional kernel weights, fully connected layer weight matrices, and bias vectors are model parameters stored in the encrypted persistent storage of the TEE after training. The specific numerical values ​​are not dependent on any specific deep learning framework. Those skilled in the art can use any equivalent numerical computation library (such as the TEE-adapted version of TensorFlow Lite Micro, the TEE-adapted version of ONNX Runtime, or a self-implemented matrix operation library) within the TEE security area to complete the forward propagation calculation, as long as the calculation result is mathematically equivalent to the above process.

[0137] 4. Quantitative definition of adversarial sample training and rejection boundaries;

[0138] To enhance the robustness of autoencoders against "noise injection attacks" in industrial attack and defense scenarios, this embodiment explicitly defines and introduces adversarial physical injection samples in the training set, and quantifies the "rejection boundary" of the autoencoder.

[0139] Adversarial physics injection attacks refer to attackers injecting abnormal physical vibrations into sensors via electromagnetic interference or mechanical vibration tables, causing abnormal noise to superimpose on the signals of normal devices, attempting to interfere with the normal generation of source evidence. To counter this type of attack, this embodiment constructs the following negative sample set in the training dataset.

[0140] The first category is electromagnetic interference (EMI) injection samples. Using a programmable EMI injection platform, EMI signals of different intensities (electric field strength range 1 to 30 V / m) and frequencies (100 Hz to 10 kHz sweep frequency) are coupled non-contactly onto the sensor signal lines of the equipment, and the process noise signal after superimposed interference is collected. The second category is mechanical vibration injection samples. Mechanical vibrations of different frequencies (50 Hz to 2 kHz) and amplitudes (0.1 g to 5 g, where g is the acceleration due to gravity) are applied to the surface of the equipment housing using a piezoelectric vibrator, and the sensor signals affected by the vibration are collected. The third category is signals from other equipment of the same model (already included in the aforementioned verification process).

[0141] These negative samples are not included in the autoencoder's weight updates during training (i.e., they are not incorporated into the training loss function), but are used to define the rejection boundary after training. The quantization definition of the rejection boundary is as follows: calculate the reconstruction error of all normal samples in the training set, and take the 99.9th percentile as the threshold. (That is, only 0.1% of normal samples have a reconstruction error exceeding this value). The reconstruction error is calculated for all adversarial injection samples, and the minimum value is taken as the minimum. The operation point for rejecting the boundary is selected at... and Between the two, specifically the geometric mean, the result is recorded as the rejection boundary. :

[0142] ;

[0143] During the online judgment phase, any reconstruction error exceeding... All signals are directly rejected, and a security log is simultaneously triggered, recording the reason for rejection and a timestamp for subsequent security auditing. This method provides clear and reproducible engineering-based quantitative evidence for the rejection behavior of the autoencoder, rather than vague empirical parameter tuning.

[0144] 5. An adaptive threshold online statistical method based on dual thresholds and trend monitoring;

[0145] Industrial equipment experiences slow, long-term drift in the statistical characteristics of process noise signals due to factors such as mechanical wear, aging of electrical components, and changes in ambient temperature and humidity. To address this issue and simultaneously resist the "taming" of thresholds by slow injection attacks, this embodiment proposes an adaptive thresholding strategy based on dual thresholds and trend monitoring.

[0146] The enclave maintains two sets of statistics for each monitoring channel simultaneously: a long-term benchmark that is permanently frozen after initial training. and online updated EWMA statistics The updated formula is as follows:

[0147] ;

[0148] ;

[0149] Where α and β are smoothing factors, and in this embodiment, α=0.02 and β=0.02 are used. and The initial values ​​are the mean and standard deviation of the reconstruction errors of all samples in the training set.

[0150] In each decision cycle, the system calculates two candidate threshold values ​​and an adaptive threshold. :

[0151] Safety upper limit threshold ;

[0152] Adaptive threshold ;

[0153] Where k=6 is the sensitivity coefficient, and M=2.0 is the upper bound expansion factor. The decision is first made by checking the rejection boundary: if... If the boundary check is rejected, the following decision rule is then applied:

[0154] Rule 1 (Safety Limit): If Then regardless If the EWMA is found to be "of questionable origin," the device will directly determine that the origin is "questionable" and refuse to generate proof of origin. This rule ensures that, even in extreme cases where the EWMA has been severely contaminated by a long-term attack, the device will not issue proof of a signal that deviates significantly from the original physical fingerprint.

[0155] Rule 2 (Normal Channel): If If the result is "passed", then the source proof is generated.

[0156] Rule 3 (Gray Area and Trend Monitoring): If Upon entering the "gray area," the system activates a trend monitoring counter to calculate the slope of the linear fit error for the most recent N=100 gray area samples. If the slope exceeds a preset threshold... (In this embodiment, 0.001 / sample) indicates that the error is increasing rapidly, which may be due to a slow injection attack. The system triggers a security alert and temporarily rejects subsequent signals. If the slope is below the threshold, it indicates that the error is rising slowly and steadily, which is due to the natural aging of the device. The system maintains normal operation and records the aging log. When natural aging leads to... Continued rise approaching At that time, the system automatically generates a maintenance work order.

[0157] This composite decision mechanism achieves an optimal balance between resisting slow injection attacks and ensuring the normal aging and availability of equipment.

[0158] 6. Generation of source proof and design to prevent replay;

[0159] The security procedures within the TEE use the device private key only if and only if the decision logic returns "pass". Sign the concatenation of the following four fields to generate proof of origin:

[0160] ;

[0161] in, This is the SHA-256 hash value of the production data corresponding to this process operation; Timestamp is a high-precision timestamp (millisecond level) of a secure monotonically increasing clock within the enclave; Counter is a strictly monotonically increasing 64-bit counter maintained by the enclave, automatically incrementing by one after each successful signing, and continuing to increment from the last recorded value after a power outage and restart, fundamentally preventing replay attacks. The signature verification terminal can use the device's public key. Independent verification of integrity and origin. Through this construction, the present invention superimposes a cryptographic-based, end-to-end trust mechanism on the physical medium of outdated production equipment, creatively endowing outdated equipment with a brand-new capability of "self-verification of data source origin".

[0162] 7. Fingerprint re-registration and trust migration process after replacement of key equipment components;

[0163] The nozzles of SMT pick-and-place machines are consumable parts, typically requiring replacement after every tens of thousands of units produced. Replacing the nozzle fundamentally alters the physical resonant frequency and turbulence characteristics of the pneumatic circuit, rendering the original factory-installed template inapplicable to the new nozzle. Without intervention, the TEE (Technical Equipment Exchange) will permanently reject all subsequent signals from that channel due to reconstruction errors far exceeding the threshold, causing the equipment verification module to malfunction after only a few weeks of operation in actual production line maintenance cycles. To address this issue, this embodiment defines a complete "fingerprint re-registration and trust migration process after replacing critical equipment components."

[0164] When production line maintenance personnel perform nozzle replacement operations, they need to trigger "maintenance mode" on the edge TEE human-machine interface of the equipment. The maintenance personnel use their private key, which is authenticated through blockchain distributed identity verification, to sign and authorize the maintenance operation. After receiving the signature authorization, the TEE enters the fingerprint re-registration state, at which time the generation of source proof is suspended.

[0165] During the re-registration process, the TEE uses the newly installed nozzle to collect at least 200 process noise signal segments under standardized operating conditions (automatically completed by maintenance personnel operating the equipment to execute standard test procedures). Since the OEM has pre-generated and signed the initial equipment noise fingerprint model and model delivery certificate for the same model of spare nozzle at the factory (this pre-generation work is completed when the OEM manufactures the spare nozzle, and the certificate is delivered along with the spare nozzle), the TEE uses the pre-set OEM public key to verify the validity of the certificate before loading the new nozzle. As an initial model, replacing the old one .

[0166] Subsequently, the TEE automatically executes a compressed online fine-tuning process, using 200 newly acquired samples to fine-tune the model, while still constrained by a 5% weight change. After fine-tuning, the TEE completes fingerprint migration and trust chain continuation, resulting in a device noise fingerprint model with a continued trust chain. The hash value of the device noise fingerprint model that perpetuates the trust chain is recorded on the blockchain, and a "fingerprint migration event" is generated, recording the migration time, old nozzle serial number, new nozzle serial number, authorized maintenance employee's distributed identity identifier, and the hash value of the migration operation. From the moment the migration event is confirmed on the blockchain, the generation of the source proof of the channel resumes normal operation.

[0167] This process extends the trust chain of device fingerprints from the factory to the entire lifecycle maintenance, ensuring the long-term availability of the device authentication module.

[0168] Example 4: Privacy Compliance Verification Scheme Based on Zero-Knowledge Proof

[0169] This embodiment details the implementation of the privacy compliance proof module. This module aims to reconcile two seemingly contradictory business needs in SMT foundry services: brand owners require auditable process compliance assurances, while foundries must protect their core process secrets. This embodiment utilizes zero-knowledge proofs, a cutting-edge cryptographic tool, to construct a privacy compliance proof generation and verification mechanism. This allows the foundry to provide the client with a concise mathematical proof, assuring the client that all products strictly meet process windows and material requirements, without the client obtaining any additional information about specific process parameters. The following sections elaborate on the circuit compilation, numerical preprocessing, aggregate signature optimization, secure restoration of selectively disclosed circuits, recursive aggregation, and on-chain verification.

[0170] 1. The compilation process from process specifications to zero-knowledge compliant circuits;

[0171] Brands use the Circom programming language to translate process specifications into compliant circuitry (C). For example, temperature range verification is implemented using the LessEqThan component.

[0172] template TempRangeCheck() {

[0173] ;

[0174] ;

[0175] ;

[0176] ;

[0177] ;

[0178] ;

[0179] ;

[0180] ;

[0181] ;

[0182] ;

[0183] ;

[0184] }

[0185] Material batch matching is achieved by verifying that the actual batch hash is equal to the hash in the approval list. All verification templates are assembled into root circuit C on a PCB-by-PCB basis. Private inputs are the foundry's production measurements, and public inputs include the batch Merkel root. And proof of origin for the polymerization equipment.

[0186] The circuit compilation generates a proof key pk and a verification key vk. vk is embedded in the consortium blockchain smart contract. pk is encrypted by the brand using the foundry's TEE enclave public key and delivered via a secure offline channel, where it is decrypted and used within its enclave.

[0187] It should be noted that the generation of the aforementioned compliance proof can be performed in the edge trusted execution environment or in a dedicated proof server deployed by the foundry. The dedicated proof server also possesses a hardware-level trusted execution environment, and its security is equivalent to that of the edge trusted execution environment; the only difference lies in its deployment location and computing resources. It is suitable for scenarios with higher requirements for proof generation efficiency.

[0188] 2. Numerical preprocessing: a bridge from physical simulation quantities to integers in finite fields;

[0189] All signals in a ZKP circuit are defined in a finite integer space, therefore, physical analog quantities need to be converted using fixed-point representation. The scaling factor is defined in the table below and is contractually agreed upon and verified on-chain between the brand owner and the contract manufacturer.

[0190] Temperature value °C 0.1°C 10 0 ~ 3500 Duration value Second 0.1 seconds 10 0 ~ 6000 Mounting X / Y deviation µm 1µm 1 -500 ~ +500 Mounting angle deviation ° 0.01° 100 -18000 ~ +18000 Material Batch Hash none 1 bit 1 <![CDATA[0 ~ 2² 56 -1]]>

[0191] For parameters that may be negative (such as mounting deviation), this embodiment does not use binary two's complement representation (because addition and subtraction operations in finite field circuits would introduce a large number of high-bit-width constraints). Instead, it uses a method of "adding an offset and mapping to the unsigned domain". Specifically, for mounting deviations with a value range of [-500, +500], an offset of 500 is uniformly added before the input circuit, mapping it to the unsigned range [0, 1000]. All comparison operations inside the circuit are based on this unsigned value, and after verification, the application layer program restores it to the original physical value.

[0192] In selective disclosure proofs (such as the progressive fine-grained disclosure query described in Example 5), a "subtraction-to-restore" operation needs to be securely introduced into the circuit to recover the original signed value while preventing integer overflow or input range bypass vulnerabilities. The specific implementation method of this embodiment is as follows. In the selective disclosure circuit template, the following constraint logic is set. First, the original signed physical value is mapped to an unsigned integer value through offset mapping. Specifically, the disclosed input includes the mapped unsigned value. And the corresponding offset (constantly 500 in this example). Input includes the original signed value. Through constraints Complete the offset mapping.

[0193] Secondly, perform range validation, using the LessEqThan component to validate each range separately. as well as Through the aforementioned offset mapping and range verification, the circuit is mathematically guaranteed to... The signed_val can only be generated by offsets within the range [-500, +500]. Any input attempting to bypass the range validation will be rejected by the circuit.

[0194] This method eliminates the need for complex two's complement operations, safely restoring signed values ​​and binding ranges entirely within the framework of unsigned integers.

[0195] It should be noted that the aforementioned security restoration constraint logic has been pre-compiled into each signed parameter template in the general query circuit template library during the system deployment phase. For parameter types with different value ranges (such as mounting deviation [-500, +500] and angle deviation [-18000, +18000]), the offset and range verification boundary in the template circuit are automatically adjusted according to the scaling factor table in Section 2 of Embodiment 4. For example, the offset in the angle deviation template is 18000 (corresponding to a scaling factor of 100 for 0.01° accuracy), and the range verification boundary is [0, 36000]. When generating targeted disclosure proof, the foundry only needs to fill in the specific offset unsigned integer value into the template, and the constraint logic inside the template automatically ensures that the restored signed value is within the legal range. This design allows each template in the general query circuit template library to safely handle all legal value ranges of its corresponding parameter type without the need to customize constraint parameters separately for different process specifications of different customers.

[0196] 3. Private input organization and circuit size justification;

[0197] Taking a representative batch (1000 PCBs, each with 10 temperature verification points, 300 placement accuracy verification points, and 10 material batch matching verifications) as an example, the number of constraints per board is approximately 87,500. Without considering source certification and verification, the entire batch has approximately 87.5 million constraints. Using the Plonky2 certification system on a server equipped with dual Intel Xeon Platinum 8368Q processors (76 cores), 512GB DDR4 ECC memory, and four NVIDIA A100-80GB GPUs, the time to generate a single zero-knowledge proof for 87.5 million constraint circuits is approximately 45 to 90 minutes. Foundries can schedule proof generation offline after the batch is completed. The final compliance proof generated by Plonky2 is only a few KB in size, extremely lightweight. Those skilled in the art can also choose other recursive certification systems (such as Nova, SuperNova, and other folding solutions) as alternatives.

[0198] 4. Aggregated signature scheme: Incorporating source proof into ZKP public input;

[0199] To address the risk of large-scale data forgery caused by the compromise of a single centralized proof server (TEE), this embodiment innovatively incorporates device-level source proofs into the public input and constraint verification scope of the ZKP circuit. However, directly performing ECDSA verification on each source proof individually within the ZKP circuit introduces enormous arithmetic overhead. The core of ECDSA verification is elliptic curve scalar multiplication, and implementing a complete ECDSA verification in a finite field circuit (such as Circom) requires tens of thousands to hundreds of thousands of arithmetic constraints. Assuming a batch involves 1,000 device actions, the verification portion alone would add approximately 15 million to tens of millions of constraints, significantly increasing the total number of constraints and severely impacting proof generation efficiency.

[0200] To effectively control circuit size while maintaining cryptographic security, this embodiment uses the BLS (Boneh-Lynn-Shacham) signature aggregation scheme instead of independent ECDSA signature verification. BLS signatures are based on bilinear pairing and have natural signature aggregation characteristics: multiple signers can generate signatures for their respective messages, and a third party can aggregate these signatures into a single, constant-length aggregate signature. During signature verification, only one pairing verification needs to be performed on the aggregate signature to confirm that all messages have been signed by their respective signers.

[0201] Specifically, this embodiment makes the following adaptation modifications to the device source verification module in Embodiment 3, while keeping the core TEE key protection, noise fingerprint verification, and proof generation process unchanged. When generating the device TEE key, BLS12-381 elliptic curves are used; the private key remains permanently sealed within the enclave, and the public key... For elements on the curve G2 group. The signature message format for source proof is uniformly defined as:

[0202] ;

[0203] ;

[0204] in To map messages to the hash function of the G1 group (in accordance with the IETF BLS signature draft) algorithm), This is an element in group G1. The message format is the same as the signature field in Example 3. The timestamp, counter, and "Pass" maintain a strict correspondence, with only the addition of a hash mapping required by the BLS curve. The "Pass" assertion string, which was not explicitly included in Embodiment 3, is uniformly incorporated into the message format in this embodiment to ensure semantic consistency of signed messages across embodiments.

[0205] All source proofs generated and its corresponding public key The data is submitted to the verification server along with the production data. Inside the verification server enclave, the BLS signature aggregation algorithm is invoked to process all the data. Aggregates into a single aggregate signature. (i.e., proof of the source of the aggregation equipment). The aggregation operation only involves the addition of elements in the G1 group, and the computational cost is extremely low. The size is constant at 96 bytes (G1 group element compressed representation), and is completely independent of the number of source proofs.

[0206] It is particularly important to clarify that the physical execution location of the aforementioned BLS signature aggregation operation does not logically conflict with the module division in the system architecture diagram. Regarding physical runtime sequence and security requirements, this aggregation calculation step is strictly limited to execution within the TEE proof server enclave of the "Privacy Compliance Proof Module," ensuring that the aggregation process (involving batch device private key operations) is in a trusted environment, and that the aggregation result... Able to generate final compliance proof with ZKP circuits It is tightly integrated within the front-end arithmetic constraints. However, from the perspective of interface layering in system software engineering, Figure 1 The macro-architecture diagram shown categorizes this aggregation algorithm under the service interface of the "credential synthesis and on-chain module." That is, the privacy compliance proof module completes its work within its own enclave. After aggregation calculation, It is passed to the credential synthesis and on-chain module, which then... and The package is then submitted to a blockchain smart contract for unified verification.

[0207] In the public inputs of ZKP compliant circuit C, add an aggregated public key list and an aggregated signature. An additional BLS aggregation verification sub-circuit is added internally to the circuit, with the following constraint logic: verification The validity of the aggregate signature for all aggregated public keys and corresponding message hashes is verified. It should be noted that while BLS aggregate signature verification within the circuit is far more efficient than ECDSA's individual signature verification, it still involves bilinear pairing verification and extended field operations. Based on the current level of optimization in cryptographic engineering, the arithmetic constraints for a complete verification of a BLS12-381 aggregate signature in the Plonky2 circuit are typically in the range of 200,000 to 500,000, rather than tens of thousands. This is more than an order of magnitude less than the tens of millions of constraints required for ECDSA's individual signature verification, and is completely independent of the number of devices, achieving a practically deployable balance between cryptographic security and engineering feasibility. This design retains cryptographic verification of device source proof while ensuring the practically deployable scale of the ZKP circuit.

[0208] In the above schemes, several open-source projects provide references for the circuit implementation of BLS12-381 bilinear pairing verification. Within the Circom ecosystem, BLS12-381 pairing verification circuit templates provided by the circomlib-ml or circom-bls projects can be consulted. These templates efficiently implement the core Miller loop and final exponentiation operation of Ate pairing as R1CS constraints. In the gnark framework of the Rust ecosystem, its standard library directly provides a Plonk circuit implementation for BLS12-381 pairing verification.

[0209] 5. Aggregation of recursive proofs and final on-chain verification;

[0210] Using Plonky2's recursive proof capabilities, sub-compliance proofs from multiple processes are aggregated into a single final compliance proof. Designing a polymer circuit Its private input receives sub-compliance certificates for each process, while its public input receives the final Merkel root. and aggregate signature The internal constraints call a recursive verification circuit template to verify each sub-proof one by one, and perform BLS aggregation verification. The final output is... The volume remains constant at several KB, regardless of the amount of raw data or the number of processes.

[0211] The contract manufacturer only needs to and This is constructed as a blockchain transaction, triggering an on-chain verification smart contract. The contract executes an FRI verification query, completing the verification within milliseconds and outputting an immutable "true / false" conclusion.

[0212] Example 5: Comprehensive Application Case of Self-Certifying Traceability Throughout the SMT Production Line

[0213] To more intuitively demonstrate how the system of this invention can work collaboratively in a real and complex production environment, a comprehensive application example spanning the entire SMT production line process is presented.

[0214] Scenario: A leading international automotive electronics Tier 1 supplier (Brand A) places a production order with a professional electronics manufacturing services provider (OEM B) for a batch of core control motherboards for new energy vehicle battery management systems. These motherboards have extremely high functional safety requirements. Brand A requires OEM B to provide tamper-proof, independently auditable, and fully traceable data for each control motherboard produced. Simultaneously, OEM B explicitly states that its core process parameters (such as furnace temperature profile settings and stencil opening design) are trade secrets and must not be provided to any external entity in plaintext.

[0215] System Deployment and Initialization: Foundry B fully deployed the self-certifying traceability system described in this invention on its automotive electronics-specific SMT production line. The reflow oven on this line is equipped with a full-process high-purity nitrogen protection system (oxygen concentration <50ppm throughout the oven, with a nitrogen curtain at the outlet continuously purging the PCBs until the surface temperature is below 200°C), meeting the condition that bare copper ROI requires no physical protection throughout the process. Industrial-grade edge computing nodes equipped with TEEs are installed next to each key workstation. Non-intrusive control sensors are installed on the pick-and-place machine nozzles and reflow oven temperature zones and connected to the TEEs. The first-stage AOI is upgraded to a physical identity registration module, with a laser etching device for discarded board ROI destruction. Under the witness of the on-site engineer from Brand A, each TEE completes BLS key pair generation, public key on-chaining, OEM gold template certificate verification, and model loading. The three parties build a permissioned consortium blockchain based on Hyperledger Fabric, deploying identity, verification, and traceability smart contracts, where the traceability smart contract includes an identity revocation function. Brand A's process team compiled all process specifications into compliant circuits, deployed verification keys on the blockchain, and delivered proof keys securely via encrypted hardware. The system also deployed a pre-defined general query circuit template library, containing three templates: single-point temperature threshold verification, mounting accuracy threshold verification, and material batch attribution verification. Each template has been compiled with R1CS and has generated proof keys.

[0216] Production process: After the PCB enters the production line, it first undergoes high-resolution imaging and physical identity registration for dual ROIs at the AOI station under nitrogen protection, generating... The hash is then uploaded to the blockchain. Because the production line is equipped with full-process nitrogen protection (including an outlet nitrogen curtain), the bare copper ROI is in an inert atmosphere throughout the reflow soldering and furnace cooling processes, requiring no physical protection. Only before wave soldering is an automatic placement machine applying a UV-resistant, non-stick temporary protective film (30-micron substrate, short-term temperature resistance 260°C / 30 seconds, long-term temperature resistance 150°C, with good chemical resistance; the film extends at least 2 mm beyond the ROI on all sides, applying approximately 0.5 Newtons of placement pressure to ensure edge sealing) to prevent flux splatter contamination. For the conformal coating process, room-temperature curing silicone sealant is applied along the edges of the protective film to reinforce the seal.

[0217] During production, the AOI equipment detected an irreparable substrate scratch defect on a PCB in the first process, and the MES system determined it to be a scrap board. Upon receiving the scrap signal, the Physical Identity Registration module triggered a laser etching device to etch a destructive mark exceeding 50 micrometers in depth and 100 micrometers in width in the ROI area of ​​the scrap board, physically destroying the PUF texture. Simultaneously, the credential synthesis and on-chain module constructed a transaction and called the on-chain identity revocation function to rescind the board's registration. The board was marked as "scrapped - unrecoverable". It then entered the defective product recycling process, and its on-chain digital identity was permanently invalidated, making it unusable by any subsequent attackers.

[0218] Each workstation's production data includes the corresponding BLS source certificate for the equipment. After the batch process was completed, it was proven that the server enclave first handled all of them. Aggregate into (96 bytes), then generate a recursive zero-knowledge proof. (Approximately 3.2KB) and uploaded to the blockchain. To alleviate the conflict between proof generation time and production cycle time, this system adopts a pipelined layered proof generation architecture. Specifically, the system does not wait for the entire batch to be produced before starting proof generation. Instead, it divides the batch into several sub-batches (e.g., 100 wafers per sub-batch), and each sub-batch begins generating sub-compliance proofs immediately after production is completed. Once all sub-batches are completed and their corresponding sub-compliance certificates are generated, the final aggregation circuit will combine each sub-compliance certificate with the aggregate signature. One-time aggregation into the final Each sub-batch's proof generation takes approximately 10 to 15 minutes and can be completed in parallel during the production of the next sub-batch, without affecting the continuous operation of the production line. The final aggregation stage only takes a few minutes. Through this pipelined architecture, the proof generation for the entire batch highly overlaps with the production process, and the production line does not need to wait for proof generation for a long time after the batch is completed.

[0219] On the quality monitoring screen at Brand A's headquarters, the status of this batch was updated in real time to "self-certified compliance". Brand A still knew nothing about any of the core process parameters of OEM B.

[0220] Post-incident investigation and gradual, granular disclosure: Several months later, a motherboard experienced an intermittent failure. A quality engineer, equipped with a portable PCB texture re-inspection instrument (with an optical system strictly consistent with the production line's AOI: a 10x telecentric lens, a 470nm low-angle ring-shaped blue LED light source, and a 1024×1024 standard ROI; built-in geometric / photometric calibration targets, requiring imaging and automatic calibration before each use), went to the site. Using 365nm ultraviolet light to irradiate the UV anti-adhesion protective film at the ROI for 15 seconds, after removing it without residue, a re-inspection image was successfully reconstructed, revealing a unique identification mark. This is matched against the on-chain hash. When tracing a specific PCB is required, the system first relies on its unique identifier. Using the batch identity Merkle tree, a Merkle proof path is generated to cryptographically prove that the PCB belongs to the batch without revealing the identities of other PCBs in the same batch.

[0221] The portable recalibrator incorporates a geometric / photometric composite calibration target, which comprises two patterned areas. The geometric calibration area uses a 9×7 circular dot matrix pattern, with dots of 0.5 mm in diameter and a center-to-center distance of 1.0 mm, used to calibrate the geometric distortion coefficients (radial distortion k1, k2 and tangential distortion p1, p2) of the imaging system. The photometric calibration area uses a 12-level grayscale gradient pattern, with grayscale values ​​uniformly distributed from 5% to 95% reflectance, used to calibrate the illumination uniformity coefficient and grayscale response curve of the imaging system. Before each use, the operator places the calibration target below the objective lens, and the recalibrator automatically acquires an image of the target, calculates the distortion coefficients using the OpenCV standard calibration algorithm, and generates a correction lookup table. After calibration, the image acquired by the re-inspection instrument for the ROI, after distortion correction and grayscale normalization, should have an Euclidean distance between its Gabor feature vector and the registered image of the production line AOI. This distance should not exceed three times the standard deviation of the Euclidean distance of the feature vectors when the registered image is repeatedly imaged on the production line AOI. If this consistency requirement cannot be met after calibration, it indicates a hardware deviation in the optical system (such as lens contamination or light source aging), requiring hardware maintenance.

[0222] Brand A, through its authorized blockchain light node, selects the "Single-Point Temperature Threshold Verification" template from a pre-set universal query circuit template library and fills in the parameters: The temperature range is transition zone 9 to 10, the time window is 68 to 78 seconds, and the threshold is 235.0°C. After receiving a query request, the foundry's key management server, located in an off-site location, calls the corresponding template's proof key, using encrypted production data as private input, to generate a targeted disclosure proof online. (Approximately 15 seconds). The template circuit has a pre-set safety subtraction restoration constraint to ensure that unsigned offsets can be safely restored to their original signed values ​​without being susceptible to overflow attacks. Upon successful verification, it only returns "Yes"—confirming the existence of temperature fluctuations below 235.0°C, but without revealing specific values ​​or curves.

[0223] Based on this initial finding, both parties initiated a controlled joint failure analysis. They signed a limited confidentiality disclosure agreement for this specific PCB, stipulating that time-limited (30-minute) temporary data viewing sessions would be conducted only under the supervision of their lawyers and quality engineers in a secure server room at foundry B, examining only the actual temperature profile of the faulty board within the relevant time window. The access logs for this session were stored on a consortium blockchain to ensure traceability of any breach of contract by either party. Through this "progressive minimal disclosure" approach, brand A identified key directions for failure analysis, while the complete data for the remaining 999 PCBs from the same batch at foundry B remained completely confidential.

[0224] In summary, the entire traceability process of this invention organically integrates three innovations—PUF physical binding, device physical fingerprint proof, and zero-knowledge compliance verification—and truly builds upon a solid foundation of mathematics, physical laws, and cryptography, providing a highly engineering-feasible innovative system for the electronics manufacturing industry chain.

[0225] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A blockchain-based SMT production data traceability system, characterized in that, include: The physical identity registration module is configured to acquire microscopic texture images of printed circuit boards (PCBs) collected by optical inspection equipment in the SMT production line, extract physical non-cloning function features, and generate a unique identity identifier for the PCB. The equipment verification module is configured to use a trusted execution environment to collect physical noise signals associated with production actions from production equipment, and to perform real-time verification of the physical noise signals based on a pre-established equipment noise fingerprint model to generate source verification for proving the authenticity of the production data source. The privacy compliance proof module is configured to obtain the zero-knowledge compliance circuit converted from the preset process specifications, and generate a compliance proof that proves the production data meets the compliance circuit requirements without disclosing the content of the production data, based on the PCB production data and unique identifier. The credential synthesis and on-chain module is configured to associate and integrate unique identity identifiers, source proofs, and compliance proofs to generate a self-certifying traceability credential for the PCB, and then upload the credential to the blockchain.

2. The blockchain-based SMT production data traceability system according to claim 1, characterized in that, The physical identity registration module is further configured as follows: High-resolution micro-texture images are acquired within a pre-selected region of interest on the PCB; feature vectors reflecting the copper foil lattice, solder resist particle distribution, or substrate fiber texture are extracted using Gabor filtering or deep learning models; and a stable, unique identifier and corresponding auxiliary data are generated from the feature vectors using a fuzz extractor. Furthermore, the physical identity registration module is configured to, upon receiving a PCB scrap signal from the manufacturing execution system, trigger an irreversible physical destruction operation on the region of interest of the scrapped PCB and send an identity revocation transaction to the blockchain, marking the unique identity of the scrapped PCB as an unrecoverable scrapped state, thus serving as the final state of the unique identity; wherein, the manufacturing execution system is the source system in the SMT production line used to issue the PCB scrap signal.

3. The blockchain-based SMT production data traceability system according to claim 2, characterized in that, The device verification module includes an edge trusted execution environment deployed near the production equipment, and the edge trusted execution environment stores a device noise fingerprint model. The equipment noise fingerprint model is obtained and delivered in the following way: an initial equipment noise fingerprint model is trained based on process noise signals collected from the production equipment in a trusted factory environment; the initial equipment noise fingerprint model is digitally signed to generate a model delivery certificate; wherein, the initial equipment noise fingerprint model is the model delivered in the form of a certificate. The edge trusted execution environment is configured to verify the signature validity of the model delivery certificate upon initial startup, and to fine-tune the initial device noise fingerprint model online using the parameters of the initial device noise fingerprint model as the training starting point to obtain an online adapted noise fingerprint model. During the online fine-tuning of the initial device noise fingerprint model, the change in the weights of the online adapted noise fingerprint model relative to the weights of the initial device noise fingerprint model is limited to a preset constraint range.

4. The blockchain-based SMT production data traceability system according to claim 3, characterized in that, The device noise fingerprint model is an autoencoder model; The edge trusted execution environment is configured to perform high-pass filtering on the raw physical noise signal acquired in real time to separate the macroscopic process waveform and high-frequency residual signal; The high-frequency residual signal is input into the autoencoder model to calculate the reconstruction error. When the reconstruction error meets the preset decision conditions, a source proof is generated; The preset decision conditions include: a safety upper limit threshold set based on a frozen long-term benchmark statistic, and an adaptive threshold set based on an online-updated exponentially weighted moving average statistic; Furthermore, when the reconstruction error falls into the gray area between the safety upper limit threshold and the adaptive threshold, the slope monitoring results based on the changing trend of the reconstruction error can distinguish between normal aging of the equipment and malicious attacks.

5. The blockchain-based SMT production data traceability system according to claim 4, characterized in that, The device verification module is also configured to: During the training phase of training the device noise fingerprint model, adversarial physical injection samples are introduced as a negative sample set; the negative sample set includes electromagnetic interference injection samples and mechanical vibration injection samples. The rejection boundary is determined based on the negative sample set. The rejection boundary is quantified based on the geometric mean between the preset high quantile of the reconstruction error of the normal sample set and the minimum value of the reconstruction error of the negative sample set. Upon receiving a maintenance authorization signature authenticated by distributed identity, the device enters the fingerprint re-registration state; the maintenance authorization signature is the authorization credential used to trigger the device's fingerprint re-registration. In the fingerprint re-registration state, the pre-signed initial model of the backup component is loaded, and based on the pre-signed initial model, fingerprint migration and trust chain continuation after the replacement of key components of the device are performed to obtain the device noise fingerprint model that continues the trust chain.

6. The blockchain-based SMT production data traceability system according to claim 5, characterized in that, The privacy compliance proof module is further configured as follows: In a trusted execution environment or a dedicated proof server, production data is used as private input and a unique identifier is used as public input to run a zero-knowledge proof generation algorithm to generate a compliance proof; the dedicated proof server is a computing device used to run the zero-knowledge proof generation algorithm. Preset process specifications include temperature profile ranges and material batch supplier restrictions; Furthermore, the public input for compliance proof also includes aggregated device source proof; among which, aggregated device source proof is an aggregated signature of constant length formed by aggregating source proofs generated by multiple device proof modules using a signature aggregation scheme based on bilinear pairing. The zero-knowledge compliance circuit contains a sub-circuit that verifies the source proof of aggregated devices.

7. The blockchain-based SMT production data traceability system according to claim 6, characterized in that, The privacy compliance proof module is also configured as follows: Multiple sub-compliance certificates corresponding to multiple processes or multiple production equipment are compressed into a single final compliance certificate through recursive zero-knowledge proof technology, and this final compliance certificate is output as the compliance certificate. Furthermore, the privacy compliance proof module adopts a pipelined layered proof generation architecture, which divides the production batch into several sub-batches. Each sub-batch generates its corresponding sub-compliance proof in parallel after production is completed, and finally aggregates them into the final compliance proof.

8. The blockchain-based SMT production data traceability system according to claim 7, characterized in that, The credential synthesis and on-chain module is configured as follows: A Merkle tree is constructed by uniquely identifying multiple PCBs within the same batch. The source proof and compliance proof are associated with the root of the Merkle tree to generate on-chain records. When it is necessary to trace a specific PCB, a Merkel proof path is generated based on the unique identifier of the specific PCB and the batch identity Merkel tree to achieve selective disclosure; Furthermore, the system also includes a pre-set general query circuit template library, which contains pre-compiled zero-knowledge circuit templates for different traceability query dimensions; Upon receiving a traceability query request, the corresponding template in the general query circuit template library is invoked, and the query parameters are filled in to generate a targeted disclosure certificate online.

9. The blockchain-based SMT production data traceability system according to claim 8, characterized in that, The system collects physical noise signals through external sensors, which are installed without interfering with the original control system of the production equipment. The physical identity registration module is configured to: select multiple spatially dispersed regions of interest on the PCB; and for each region of interest, independently perform image acquisition and feature extraction to obtain auxiliary data corresponding to each region. During the reconstruction phase, the texture images collected for each region of interest are reconstructed to obtain candidate unique identifiers for each region. If at least one candidate unique identifier matches the unique identifier stored on the blockchain, the identity verification is deemed successful.

10. The blockchain-based SMT production data traceability system according to claim 9, characterized in that, The device verification module includes an edge trusted execution environment deployed one-to-one with each production device. The edge trusted execution environment is used to store the device noise fingerprint model and the device private key, and independently generate source verification. The physical identity registration module is also configured to: apply a preset pressure to ensure edge sealing when the protective film is attached to the region of interest; and perform grayscale histogram normalization in image preprocessing to obtain a compensated image to compensate for image shift caused by chemical residues; wherein, the protective film is a physical protective layer attached to the region of interest. The privacy compliance proof module is also configured to: set constraint logic in the selective disclosure circuit when selective disclosure is performed; map the signed original physical value to the unsigned integer value through offset mapping, and perform range verification to securely restore and verify the signed original physical value within the unsigned integer domain.

Images