Iot-based factory real-time digital monitoring method and system and storage medium

Abstract

The application relates to the technical field of industrial monitoring, and discloses a factory real-time digital monitoring method and system based on an Internet of Things and a storage medium. The method comprises the following steps: performing encrypted access to heterogeneous Internet of Things equipment through double identity authentication to obtain a secure authentication equipment identifier pool; performing multi-dimensional data collection on the equipment identifier pool based on hash verification to obtain encrypted time sequence data chains; performing dynamic identification on factory visual images by using a confidence adaptive threshold to obtain multi-task fusion feature codes; performing heterogeneous fusion on the time sequence data chains and the feature codes by using a sliding window weight distribution algorithm to obtain factory digital state fingerprints; and performing knowledge graph reasoning on the state fingerprints by using a three-level early warning threshold to obtain a real-time risk level early warning code. The application solves the technical problems that heterogeneous Internet of Things equipment cannot be uniformly accessed and authenticated, multi-source data lacks secure fusion processing, and a factory state cannot be intelligently reasoned and warned.

Description

Technical Field

[0001] This application relates to the field of industrial monitoring technology, and in particular to a method, system and storage medium for real-time digital monitoring of factories based on the Internet of Things. Background Technology

[0002] Current factory monitoring technologies primarily rely on a combination of traditional single-sensor data acquisition and manual inspection. This involves installing sensors such as temperature, pressure, and vibration sensors on production equipment to acquire operating parameters, with operators periodically inspecting the equipment's status and safety. This sensor data is transmitted to a central monitoring system via wired or wireless networks, where monitoring personnel determine whether the equipment is operating normally based on preset thresholds, issuing alarm signals when monitored parameters exceed normal ranges. Simultaneously, computer vision technology is also beginning to be applied to factory monitoring, mainly for single-task target detection, such as identifying whether workers are wearing safety helmets or detecting obvious defects on equipment surfaces.

[0003] However, existing technologies have significant shortcomings, primarily manifested in the severe problem of data silos. Different manufacturers' devices employ different communication protocols and data formats, leading to ineffective integration of various sensors and monitoring equipment, resulting in independent monitoring systems. Network security capabilities are weak; a large number of IoT devices connect to the network but lack unified identity authentication and data encryption mechanisms, posing risks of data leakage and system attacks. Existing visual recognition technologies are mainly optimized for single tasks and cannot simultaneously handle multi-dimensional monitoring needs such as device detection, personnel behavior recognition, and environmental safety assessment, and lack effective integration with sensor data. Summary of the Invention

[0004] This application provides a method, system, and storage medium for real-time digital monitoring of factories based on the Internet of Things (IoT), which solves the problems of heterogeneous IoT devices being unable to access and authenticate in a unified manner, lack of secure fusion processing of multi-source data, and inability to intelligently infer and warn about factory status, and establishes a method for real-time digital monitoring of factories based on the IoT.

[0005] Firstly, this application provides a method for real-time digital monitoring of factories based on the Internet of Things (IoT). The method includes: encrypting heterogeneous IoT devices through a dual authentication mechanism to obtain a secure authentication device identifier pool; performing multi-dimensional data acquisition and processing on the secure authentication device identifier pool based on hash verification integrity verification to obtain an encrypted time-series data chain; dynamically recognizing factory visual images using a confidence-adaptive threshold mechanism to obtain a multi-task fusion feature code; performing heterogeneous fusion processing on the encrypted time-series data chain and the multi-task fusion feature code based on a sliding window weight allocation algorithm to obtain a factory digital status fingerprint; and performing knowledge graph reasoning processing on the factory digital status fingerprint through a three-level early warning threshold determination to obtain a real-time risk level early warning code.

[0006] Secondly, this application provides an IoT-based real-time digital monitoring system for factories, the IoT-based real-time digital monitoring system for factories comprising:

[0007] The encryption module is used to perform encrypted access processing on heterogeneous IoT devices through a dual authentication mechanism to obtain a secure authentication device identifier pool.

[0008] The data acquisition module is used to perform multi-dimensional data acquisition and processing on the security authentication device identifier pool based on hash verification integrity verification to obtain an encrypted time-series data chain.

[0009] The recognition module is used to dynamically recognize factory visual images using a confidence-adaptive threshold mechanism to obtain multi-task fusion feature codes.

[0010] The fusion module is used to perform heterogeneous fusion processing on the encrypted time-series data chain and the multi-task fusion feature code according to the sliding window weight allocation algorithm to obtain the factory digital status fingerprint.

[0011] The reasoning module is used to perform knowledge graph reasoning processing on the digital status fingerprint of the factory through a three-level early warning threshold to obtain a real-time risk level early warning code.

[0012] Thirdly, an IoT-based real-time digital monitoring device for factories is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the IoT-based real-time digital monitoring device for factories to execute the aforementioned IoT-based real-time digital monitoring method for factories.

[0013] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described IoT-based real-time digital monitoring method for factories.

[0014] The technical solution provided in this application employs a dual authentication mechanism to encrypt and process heterogeneous IoT devices, resulting in a secure authentication device identifier pool. This addresses the issues of poor device compatibility and weak security protection in traditional factory monitoring systems, establishing a unified device access standard and a multi-level security verification system to ensure the reliability and security of data collection. Based on hash verification integrity checks, the secure authentication device identifier pool undergoes multi-dimensional data collection and processing to obtain an encrypted time-series data chain, effectively solving the problems of integrity protection and anti-tampering during data transmission. The combined application of the SHA-256 hash algorithm and AES-256 encryption technology establishes an end-to-end data security mechanism. A confidence-adaptive threshold mechanism is used to dynamically recognize and process factory visual images, obtaining multi-task fusion feature codes. This overcomes the limitation of existing visual recognition technologies that can only handle single tasks. By improving the YOLOv8 network architecture, parallel processing of device detection, personnel segmentation, and posture recognition is achieved. Simultaneously, the confidence-adaptive adjustment mechanism dynamically optimizes the detection threshold based on historical recognition accuracy, significantly improving the recognition accuracy and stability in complex factory environments.

[0015] The factory's digital status fingerprint is obtained by heterogeneously fusing encrypted temporal data chains and multi-task fusion feature codes using a sliding window weight allocation algorithm. This innovatively solves the problem of spatiotemporal alignment and fusion of sensor and visual data. An intelligent fusion system for heterogeneous data is established through a 30-second sliding window mechanism and dynamic weight allocation of sensor reliability scores. An improved Kalman filter algorithm with an adaptive noise estimation mechanism further enhances the accuracy of status prediction. Deep autoencoder feature dimensionality reduction effectively extracts the core features of factory operation and removes data redundancy. Real-time risk level warning codes are obtained by knowledge graph reasoning based on a three-level warning threshold determination of the factory's digital status fingerprint. An intelligent reasoning system based on RDF triple knowledge graphs and graph convolutional networks is constructed, capable of accurately identifying the propagation path and impact range of equipment faults. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of one embodiment of the IoT-based real-time digital monitoring method for factories in this application.

[0018] Figure 2This is a schematic diagram of one embodiment of the IoT-based real-time digital monitoring system for factories in this application.

[0019] Figure 3 This is a schematic block diagram of the structure of a factory real-time digital monitoring device based on the Internet of Things in an embodiment of the present invention. Detailed Implementation

[0020] This application provides a method, system, and storage medium for real-time digital monitoring of factories based on the Internet of Things (IoT). The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the IoT-based real-time digital monitoring method for factories in this application includes:

[0022] Step S101: Encrypt the access of heterogeneous IoT devices through a dual authentication mechanism to obtain a secure authentication device identifier pool;

[0023] Step S102: Perform multi-dimensional data acquisition and processing on the security authentication device identifier pool based on hash verification integrity verification to obtain an encrypted time-series data chain;

[0024] Step S103: Apply the confidence-adaptive threshold mechanism to the factory visual image for dynamic recognition processing to obtain the multi-task fusion feature code;

[0025] Step S104: Perform heterogeneous fusion processing on the encrypted time-series data chain and the multi-task fusion feature code according to the sliding window weight allocation algorithm to obtain the factory digital status fingerprint.

[0026] Step S105: The factory's digital status fingerprint is processed by knowledge graph reasoning based on the three-level early warning threshold to obtain a real-time risk level early warning code.

[0027] It is understood that the implementing entity of this application can be an IoT-based real-time digital monitoring system for factories, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0028] Specifically, the dual authentication mechanism first verifies the MAC addresses of heterogeneous IoT devices such as temperature sensors, vibration sensors, and pressure sensors within the factory, extracting the physical identifier code for each device. Then, it uses an elliptic curve cryptography algorithm to perform digital certificate authentication on the physical identifier code. Elliptic curve cryptography is an asymmetric encryption technology based on the mathematical structure of elliptic curves. It generates public-private key pairs by performing point group operations on elliptic curves defined over a finite field. It features short key length but high security strength. During the authentication process, the device sends an encrypted request containing its identifier code to the gateway. The gateway matches and verifies the device's security credentials against a pre-stored device whitelist. After successful verification, the system assigns a unique communication key pair to the device according to a temporary session key allocation mechanism. Finally, a secure authentication device identifier pool is generated through unified identifier encoding processing.

[0029] Based on the established secure authentication device identifier pool, the system performs multi-dimensional data acquisition. It synchronously collects temperature values ​​from temperature sensors, acceleration frequency data from vibration sensors, and Pascal pressure values ​​from pressure sensors at a fixed 100-millisecond cycle, forming raw equipment operation data. Then, it performs SHA-256 hash calculation on each data packet. SHA-256 is a family of secure hash algorithms that can convert input data of arbitrary length into a fixed-length 256-bit hash value. The calculation process includes message preprocessing, message segmentation, and compression function iteration. The generated data packet integrity fingerprint is compared with the device's historical data fingerprint for difference verification. When data anomalies are detected, a data retransmission trigger mechanism automatically initiates a re-acquisition process. The filtered trusted device monitoring data is then time-series encoded according to the factory production cycle timestamp, arranging data from different time points in the production process sequence to form an encrypted time-series data chain.

[0030] An adaptive confidence threshold mechanism is used for dynamic recognition of factory visual images. First, adaptive enhancement preprocessing is performed on the acquired factory images, automatically adjusting contrast and saturation parameters according to the image brightness distribution to generate illumination-normalized image data. Then, the processed image is input into an improved YOLOv8 network. YOLOv8 is the eighth generation of the You Only Look Once object detection algorithm, which adopts a single-stage detection architecture and can simultaneously complete object detection, instance segmentation, and key point detection tasks in a single forward propagation. The network outputs the coordinates of the device detection box, the personnel segmentation mask region, and the coordinates of the operator's pose key points through multi-task parallel detection processing. The system dynamically adjusts the confidence threshold for different types of targets based on historical recognition accuracy statistics. When the historical recognition accuracy of a certain type of target is low, its confidence threshold is increased accordingly to reduce false detections. TensorRT acceleration technology controls the single-frame image processing time to within 50 milliseconds through model quantization, layer fusion, and memory optimization. Finally, the factory scene recognition feature vector is processed through multi-task fusion encoding to generate multi-task fusion feature codes.

[0031] A sliding window weight allocation algorithm is used to heterogeneously fuse encrypted temporal data chains and multi-task fusion feature codes. The sliding window mechanism sets a 30-second time window, sliding once every 10 seconds. Spatiotemporal alignment processing is performed on sensor data and visual recognition results within the window. Spatiotemporal alignment unifies data with different sampling frequencies to the same time granularity through interpolation algorithms. At the same time, a spatial coordinate mapping relationship is established based on the physical location of sensors and cameras to form a synchronous heterogeneous data matrix. The system dynamically assigns weights to the data based on the historical reliability scores of each sensor. The reliability scores are calculated based on a comprehensive calculation of indicators such as sensor failure rate, data consistency, and environmental adaptability. An improved Kalman filter algorithm adds an adaptive noise estimation mechanism to the standard Kalman filter, which can dynamically adjust the process noise and observation noise covariance matrix according to the actual observation data. The equipment state prediction vector is obtained through iterative calculation in two stages: state prediction and update. The deep autoencoder compresses the high-dimensional state vector into a low-dimensional representation through the encoder and then reconstructs it through the decoder. During the training process, the reconstruction error is minimized to learn the essential features of the data. The RSA encryption algorithm is based on the mathematical problem of large integer factorization. A 1024-bit key is used to encrypt the extracted core features of factory operation, generating a unique digital state fingerprint of the factory.

[0032] The system employs a three-tiered early warning threshold to determine and execute knowledge graph reasoning. The factory's digital status fingerprint is input into an RDF triple knowledge graph containing 50,000 entity nodes and 200,000 relational edges. The RDF triples describe the relationships between entities using a subject-verb-object structure, such as "Equipment A - Operating Status - Abnormal". The system matches entity relationships to find nodes related to the current status, including equipment, operating procedures, and safety standards, forming an equipment status association graph. A graph convolutional network performs convolution operations on the graph data, updating node representations by aggregating neighbor node information. After iterative computation, the multi-layer graph convolutional network outputs a risk propagation path vector, which describes the propagation path and impact of faults or abnormal states within the equipment network. The Analytic Hierarchy Process (AHP) decomposes the complex risk assessment problem into multiple levels. The weights of each risk factor are determined through expert scoring and consistency checks. The risk propagation path vector is weighted and summed to obtain a numerical risk level. The system sets three warning thresholds: 0.6, 0.8, and 0.9. When the risk value exceeds 0.6, a Level 1 warning is triggered to alert operators; when it exceeds 0.8, a Level 2 warning is triggered, requiring immediate inspection of equipment status; and when it exceeds 0.9, a Level 3 warning is triggered, initiating an emergency shutdown procedure. Digital signature technology uses a private key to sign the risk level classification result. The recipient uses the corresponding public key to verify the integrity and authenticity of the signature, ultimately generating a real-time risk level warning code containing the risk level, timestamp, and digital signature.

[0033] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0034] Perform MAC address verification on heterogeneous IoT devices to obtain the device physical identifier code;

[0035] The device physical identifier is digitally authenticated using an elliptic curve cryptography algorithm to obtain a device security credential.

[0036] Input the device security credentials into the pre-stored device whitelist for matching and verification to obtain the device access permission identifier;

[0037] The device access permission identifier is encrypted and authorized according to the temporary session key allocation mechanism to obtain the device communication key pair;

[0038] A secure authentication device identifier pool is obtained by performing unified identifier encoding processing on the device communication key pairs.

[0039] Specifically, the Media Access Control (MAC) address is extracted from the network interface cards of various IoT devices within the factory. The MAC address is a unique identifier for a network device at the data link layer, consisting of 48 binary bits, typically represented as 12 hexadecimal digits. The first 24 bits represent the organization's unique identifier, and the last 24 bits represent the device identifier assigned by the manufacturer. The verification process obtains the raw MAC address data by reading the physical address register of the device's network interface. Then, the read MAC address is compared with the standard MAC address format preset at the device's factory to check whether the address conforms to the IEEE standard specification. At the same time, the unicast identifier and the globally unique identifier of the address are verified to confirm the legality of the device's physical identifier. After successful verification, the MAC address is converted into a unified 64-bit device physical identifier code format. The identifier code includes a device type field, a manufacturer code field, and a serial number field. When using elliptic curve cryptography (ECC) to authenticate a device's physical identifier (PID) with a digital certificate, the algorithm first selects appropriate elliptic curve parameters, including the modulus p of the finite field, the coefficients a and b in the elliptic curve equation, the coordinates of the base point G, and the order n of the base point. An elliptic curve in a finite field is defined as the set of points satisfying the equation y²≡x³+ax+b. Based on the mathematical difficulty of the elliptic curve discrete logarithm problem, the algorithm generates a public-private key pair through scalar multiplication. The private key is a randomly selected integer d, and the public key is a point Q = d × G on the elliptic curve, where G is the base point. During authentication, the device's PID is used as the message content, combined with the device's private key to generate a digital signature. The signature algorithm first calculates the hash value of the message, then uses the elliptic curve digital signature algorithm to generate a signature pair. The verifier verifies the validity of the signature using the corresponding public key. Upon successful authentication, a device security credential containing the device identifier, public key information, validity period, and signature is generated.

[0040] The matching and verification process between device security credentials and the pre-stored device whitelist is completed by searching the authorized device records in the whitelist database. The device whitelist is stored in a hash table structure, using the hash value of the device identifier code as an index. The record content includes device model, permission level, access time limit, and functional scope. The matching process first calculates the hash value of the device identifier code in the input credentials, then searches for the corresponding record in the whitelist hash table, comparing the consistency of key fields such as device model, manufacturer information, and serial number in the credentials with the whitelist record. At the same time, it verifies the device's digital certificate chain, checking the certificate's validity period, issuing authority, and revocation status. The matching verification also includes checking the device's functional permissions, determining its allowed operating range based on the device type. For example, a temperature sensor can only perform data acquisition operations and cannot execute control commands. After successful verification, a corresponding device access permission identifier is generated based on the device's permission level in the whitelist. The identifier includes the device's unique ID, permission level code, allowed network segment, and data type marker. The temporary session key allocation mechanism performs encrypted authorization processing based on the device access permission identifier. The session key generation uses a cryptographically secure pseudo-random number generator, which combines the current timestamp, device identifier, and system random seed to generate a 128-bit or 256-bit symmetric key. The key allocation process uses a variant of the Diffie-Hellman key exchange protocol. The device and gateway each generate a temporary public-private key pair, and calculate the shared session key by exchanging the public key and combining it with their respective private keys. The key transmission process uses the device's long-term public key for encryption protection to prevent the key from being intercepted during transmission. The generated device communication key pair contains an encryption key and a message authentication key. The encryption key is used to protect the confidentiality of data transmission, and the message authentication key is used to verify the integrity and source authentication of the data.

[0041] The unified identifier encoding process converts device communication key pairs into a standardized device identifier format. The encoding process first combines the key pair information with the device's physical characteristics, network parameters, and permission information. Then, the Base64 encoding algorithm is used to convert the binary data into a printable ASCII string. Base64 encoding maps every 3 bytes of binary data to 4 characters. The character set includes AZ, az, 0-9, and two additional characters. The encoded identifier string contains fields such as version number, device type, network address, key fingerprint, and checksum. The version number identifies the version of the encoding format. The device type field distinguishes between sensor, actuator, and gateway devices. The network address field records the device's IP address and port information. The key fingerprint is a key digest calculated using a hash algorithm. The checksum is calculated using a cyclic redundancy check algorithm to detect transmission errors in the identifier string. After encoding, all authenticated device identifiers are classified and stored according to device type, access time, and permission level, forming a hierarchical secure authentication device identifier pool structure.

[0042] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0043] Based on the safety certification equipment identification pool, temperature sensors, vibration sensors, and pressure sensors are synchronously acquired and processed in 100-millisecond cycles to obtain the raw data of equipment operation;

[0044] The raw data of device operation is processed by SHA-256 hash value calculation to obtain the data packet integrity fingerprint;

[0045] The data packet integrity fingerprint is compared with the device's historical data fingerprint to obtain the data quality assessment result.

[0046] Based on the data retransmission triggering mechanism, abnormal data is filtered out from the data quality assessment results to obtain reliable device monitoring data.

[0047] Based on the factory production cycle timestamp, the monitoring data of trusted equipment is processed by time-series chain encoding to obtain an encrypted time-series data chain.

[0048] Specifically, the 100-millisecond cycle synchronous acquisition and processing based on the safety certification equipment identification pool coordinates the data acquisition of multiple types of sensors within the factory through a precise clock synchronization mechanism. The acquisition controller sends a data request command to the temperature sensor to obtain the current Celsius value according to the equipment communication parameters and access permissions stored in the safety certification equipment identification pool. At the same time, it sends a sampling command to the vibration sensor to obtain triaxial acceleration and frequency data, and sends a query command to the pressure sensor to obtain Pascal pressure readings. The synchronous acquisition mechanism uses the Network Time Protocol to ensure that all sensors read data at the same point in time, with timestamp accuracy reaching the millisecond level. The temperature sensor returns a data packet containing the sensor ID, timestamp, temperature value, and unit. The vibration sensor returns multi-dimensional data containing X-axis, Y-axis, and Z-axis acceleration values ​​and vibration frequency. The pressure sensor returns complete information including pressure value, measurement accuracy, and calibration status. The acquired raw data is organized according to a predefined data structure, including equipment identification field, data type field, numerical field, timestamp field, and status flag field, forming a standardized raw data format for equipment operation.

[0049] The SHA-256 hash value calculation process performs a cryptographic digest operation on the raw data processed by the device. SHA-256 belongs to the family of secure hash algorithms and can convert input data of arbitrary length into a fixed-length 256-bit output digest. The calculation process first preprocesses the raw data by adding padding bits to the end of the data to ensure that the total length is a multiple of 512 bits. Then, the data is divided into 512-bit message blocks. Each message block is processed by a compression function in 64 rounds to generate an intermediate hash value. The compression function uses 8 32-bit working variables and 64 constants to perform non-linear transformations, including logical function operations, modular addition operations, and circular shift operations. Finally, a 256-bit hash digest is output as the data packet integrity fingerprint. The input to the hash calculation contains all fields of the original data, ensuring that any small change in the data will result in a significant difference in the hash value. The generated integrity fingerprint is stored in hexadecimal string format, containing 64 characters, with each character representing 4 bits of binary data.

[0050] The difference verification process between data packet integrity fingerprints and historical device data fingerprints detects data anomalies by comparing the current data packet fingerprint with the device fingerprint patterns stored in the historical database. The historical data fingerprint database is indexed according to device ID and time window, storing normal data fingerprint samples from a certain period of time in the past. The verification program first looks up the corresponding historical fingerprint record based on the current device ID, and then calculates the Hamming distance between the current fingerprint and the historical fingerprint. The Hamming distance represents the number of different characters at corresponding positions in two strings of equal length. The distance value reflects the degree of data change. The verification algorithm also checks the distribution characteristics of the fingerprint, including character frequency analysis and pattern matching. When the Hamming distance exceeds a preset threshold or an abnormal pattern appears, it is marked as a data quality anomaly. Under normal circumstances, small fluctuations in sensor data will cause slight changes in the fingerprint, while device failure or data transmission errors will cause drastic changes in the fingerprint. The evaluation result includes data quality level, anomaly type identifier, and confidence score, forming a structured data quality evaluation result.

[0051] The data retransmission trigger mechanism performs abnormal data filtering based on the data quality assessment results. When an abnormal data quality is detected, the trigger automatically starts the re-acquisition process. The retransmission mechanism first analyzes the specific type of abnormality, including transmission errors, sensor failures, and environmental interference. For transmission error abnormalities, the trigger sends a resampling request to the corresponding device, requiring the device to reacquire data in the next acquisition cycle. For sensor failure abnormalities, the trigger starts the device self-test program, checking the sensor's working status and calibration parameters by sending diagnostic commands. For environmental interference abnormalities, the trigger adjusts the acquisition parameters, including increasing the number of samplings and extending the stabilization time. The filtering process uses a majority voting mechanism. When most of the data from multiple acquisitions of the same data point is consistent, the data is accepted. When the number of retransmissions exceeds the preset limit or abnormalities continue to occur, the device is marked as faulty and an alarm message is generated. After filtering, the data undergoes final verification to confirm its integrity and accuracy, forming reliable device monitoring data containing valid data records, quality identifiers, and verification timestamps.

[0052] The time-series chained encoding process arranges and encrypts the monitoring data of trusted equipment according to the factory production cycle timestamp. The factory production cycle timestamp reflects the actual operating rhythm of the production line, including the start time, duration, and end time of each process step. The encoding program first associates the monitoring data with the corresponding production cycle based on the production plan and equipment location, establishing a time correspondence between the data and the production process. Then, the data is organized into a chain structure according to the chronological order. Each data block contains the current data content, the hash fingerprint of the previous data block, and its own timestamp. The chain structure is connected by hash pointers to ensure the temporal integrity and tamper-proof characteristics of the data. The encoding process uses the AES-256 symmetric encryption algorithm to encrypt and protect sensitive data. The encryption key comes from the data protection key in the device communication key pair. The encrypted data blocks are stored in a distributed database according to the production process order, forming a complete encrypted time-series data chain. The head of the data chain contains the start timestamp and production batch information, and the tail contains the end timestamp and data integrity check code.

[0053] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0054] Adaptive enhancement preprocessing is performed on the factory visual images to obtain illumination-normalized image data;

[0055] The illumination-normalized image data is input into the improved YOLOv8 network for multi-task parallel detection processing to obtain the device detection box, personnel segmentation mask and pose key point coordinates.

[0056] Based on historical recognition accuracy, the confidence threshold of the device detection frame, personnel segmentation mask, and posture key point coordinates are dynamically adjusted to obtain adaptive detection results.

[0057] Based on TensorRT acceleration technology, the adaptive detection results are processed in real time for 50 milliseconds to obtain the factory scene recognition feature vector.

[0058] The factory scene recognition feature vector is subjected to multi-task fusion encoding processing to obtain the multi-task fusion feature code.

[0059] Specifically, the adaptive enhancement preprocessing of factory visual images addresses uneven lighting by analyzing the image's brightness distribution characteristics. The preprocessing algorithm first calculates the image's grayscale histogram, statistically analyzing the pixel distribution at each brightness level. Then, it determines the overall brightness characteristics of the image based on the peak position and dispersion of the histogram. When the image is detected to be too dark, the enhancement algorithm automatically increases the contrast parameter and brightness offset; when the image is detected to be too bright, the algorithm reduces exposure compensation and lowers the intensity of highlight areas. Adaptive enhancement also includes saturation adjustment, determining the vibrancy of colors by analyzing the numerical distribution of the RGB three color channels. The color space is converted to the HSV color space, where H represents hue, S represents saturation, and V represents brightness. The adjustment process mainly targets the S and V channels for numerical correction. The adjustment of the V channel is based on a local adaptive histogram equalization algorithm. This algorithm divides the image into multiple small regions, performs histogram equalization processing on each region independently, and then smoothly connects the processing results of each region through bilinear interpolation to generate image data with overall illumination standardization. The standardized image has a uniform brightness range and contrast level, and the pixel values ​​are normalized to the standard range of 0 to 255, eliminating the influence of different lighting conditions on subsequent recognition algorithms.

[0060] This improved YOLOv8 network performs multi-task parallel detection after receiving illumination-normalized image data. The YOLOv8 network architecture comprises three main components: a backbone network, a neck network, and a detection head. The backbone network extracts image features using a CSPDarknet structure, employing a combination of convolutional layers, batch normalization layers, and activation function layers to learn hierarchical feature representations of the image. The neck network uses a feature pyramid network structure to fuse feature information at different scales. The detection head is improved to include a multi-task structure comprising an object detection head, an instance segmentation head, and a keypoint detection head. The object detection head outputs the class probability, bounding box coordinates, and confidence score for each candidate box. The segmentation head generates pixel-level segmentation masks to identify the precise contours of each detected target. The keypoint detection head outputs the two-dimensional coordinates of human keypoints, including key positions such as the head, shoulders, elbows, wrists, hips, knees, and ankles. During the forward propagation of the network, image data undergoes multi-layer convolution operations to extract features. Each convolutional layer uses a learnable filter to perform convolution operations on the input feature map. Local feature responses are calculated through a sliding window mechanism, and nonlinear transformations are introduced by the activation function to enhance the network's expressive power. Finally, the output is a multi-dimensional detection result containing the coordinates of the device bounding box, the personnel segmentation mask region, and the coordinates of the operator's pose keypoints.

[0061] The dynamic adjustment of the confidence threshold optimizes detection performance based on the historical recognition accuracy of various target types. The dynamic adjustment mechanism maintains a historical statistical database, recording the detection accuracy, false positive rate, and false negative rate of each target category within a certain time window. The adjustment algorithm calculates the performance evaluation score for each category based on these statistical indicators. When the historical accuracy of a certain target category is low, the algorithm automatically increases the confidence threshold for that category, reducing the output of low-quality detection results. When the false negative rate of a certain target category is high, the algorithm appropriately lowers the confidence threshold, increasing the detection sensitivity. The threshold adjustment uses an exponential smoothing algorithm. The new threshold equals the historical threshold multiplied by the smoothing coefficient plus the current performance index multiplied by one minus the smoothing coefficient. The smoothing coefficient controls the weight ratio of historical data and current data. The adjusted detection results are then processed by a non-maximum suppression algorithm to remove duplicate detections. This algorithm selects the optimal detection result based on the overlap between detection boxes and the confidence score. The overlap is measured by calculating the intersection-union ratio (IU) of two detection boxes. When the IU exceeds a preset threshold, the detection box with higher confidence is retained, forming a deduplicated adaptive detection result.

[0062] TensorRT acceleration technology achieves real-time processing in 50 milliseconds through model optimization and inference engine acceleration. TensorRT is a deep learning inference optimization library developed by NVIDIA. The accelerated processing includes three key steps: model quantization, layer fusion, and memory optimization. Model quantization converts network weights from 32-bit floating-point numbers to 8-bit integers, reducing computational complexity and memory usage. The quantization process uses a calibration dataset to statistically analyze the distribution range of activation values ​​in each layer to determine the optimal quantization parameters. Layer fusion combines consecutive convolutional layers, batch normalization layers, and activation layers into a single operation, reducing memory accesses and computational overhead. Memory optimization reduces memory requirements by reusing the storage space of intermediate results. The TensorRT engine selects the optimal algorithm implementation based on the characteristics of the target hardware platform, including convolutional algorithms, activation function implementations, and memory allocation strategies. The optimized inference engine adaptively converts the detection results into a standardized feature vector format. The feature vector contains the category encoding, location information, size parameters, and confidence values ​​of the detected target, forming a factory scene recognition feature vector describing the factory scene content.

[0063] Multi-task fusion coding process unifies the results of different tasks in the factory scene recognition feature vector. The fusion coding adopts a hierarchical coding structure, encoding the results of target detection, instance segmentation, and keypoint detection into different feature subspaces. The target detection result coding includes one-hot coding of the target category, normalized values ​​of bounding box coordinates, and detection confidence. The instance segmentation result coding converts pixel-level masks into compact shape descriptors, generating shape codes by calculating geometric features of the mask such as area, perimeter, aspect ratio, and centroid position. The keypoint detection result coding converts human keypoint coordinates into pose description vectors, describing the human's action state by calculating the relative positions and angular relationships between keypoints. The fusion coding algorithm uses a concatenation method to connect the three sub-codes into a unified feature representation, and then uses principal component analysis to reduce the feature dimensionality, retaining the most important feature information, and finally generating a fixed-length multi-task fusion feature code, which contains a compact representation of all relevant information in the factory scene.

[0064] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0065] Spatiotemporal alignment of encrypted time-series data chains and multi-task fusion feature codes is performed based on a 30-second sliding window to obtain a synchronous heterogeneous data matrix.

[0066] Based on the sensor reliability score, a dynamic weight allocation process is performed on the synchronous heterogeneous data matrix to obtain a weighted fused dataset.

[0067] The weighted fusion dataset is input into the improved Kalman filter algorithm for state prediction processing to obtain the device state prediction vector;

[0068] Deep autoencoder feature extraction is performed on the equipment status prediction vector to obtain the core features of factory operation.

[0069] The factory's digital status fingerprint is obtained by generating digital fingerprints based on the core characteristics of the factory's operation using the RSA encryption algorithm.

[0070] Specifically, the 30-second sliding window spatiotemporal alignment process synchronizes the encrypted time-series data chain and the multi-task fusion feature code by establishing a unified time benchmark. The sliding window mechanism sets a fixed 30-second time span, moving forward once every 10 seconds. The window contains time-series data from different data sources. The spatiotemporal alignment algorithm first extracts the timestamps of each data block in the encrypted time-series data chain. These timestamps originate from the precise records during sensor data acquisition. Then, it extracts the generation timestamp of the multi-task fusion feature code, which corresponds to the moment when visual recognition is completed. The alignment process uses a linear interpolation algorithm to unify data with different sampling frequencies to the same time granularity. When the sensor data acquisition frequency is once every 100 milliseconds and the visual data update frequency is once every 200 milliseconds, the interpolation algorithm generates intermediate values ​​within the time interval of the visual data. The interpolation calculation is based on the time weight allocation of adjacent data points, with data closer to the target time point having a greater weight. Spatial alignment establishes a mapping relationship based on the physical coordinates of the sensor and camera, associating data from the same production area monitored by different devices. After alignment, a synchronous heterogeneous data matrix containing time and spatial dimension information is formed. The rows of the matrix represent time series, and the columns represent different data characteristics. Each matrix element contains the monitoring value of the corresponding spatiotemporal location.

[0071] The dynamic weighting process for sensor reliability scoring determines the contribution to data fusion based on the historical performance of the devices. The reliability score calculation includes three dimensions: accuracy, stability, and responsiveness. Accuracy is assessed by comparing the deviation of sensor readings from standard reference values; the smaller the deviation, the higher the accuracy score. Stability is measured by analyzing the degree of data fluctuation of the sensors in a stable environment; sensors with smaller fluctuation amplitudes receive higher stability scores. Responsiveness is determined by measuring the sensor's reaction speed to environmental changes; sensors with faster response speeds receive higher responsiveness scores. The comprehensive score is obtained by weighting and summing the scores of the three dimensions according to preset weights. The weighting algorithm assigns corresponding weight coefficients to each column of data in the synchronous heterogeneous data matrix according to the reliability score. Data from high-reliability sensors receives greater weights, while data from low-reliability sensors receives correspondingly lower weights. The weighting process multiplies the original data values ​​by the corresponding weight coefficients and then performs row-wise normalization to ensure that the numerical ranges of each feature dimension are consistent, forming a weighted fusion dataset that has been reliability corrected.

[0072] An improved Kalman filter algorithm is used for state prediction on a weighted fusion dataset. Kalman filtering is a recursive digital filter that estimates the state of a dynamic system through two steps: state prediction and state update. The improved version adds an adaptive noise estimation mechanism to the standard Kalman filter, dynamically adjusting the covariance matrix of process noise and observation noise based on actual observation data. The state prediction step uses the state transition matrix and the state estimate from the previous time step to predict the system state at the current time step. The state transition matrix describes the change of the system state over time and is derived by analyzing the physical model of the equipment operation and statistically analyzing historical data. The prediction process also considers the influence of process noise, which reflects the uncertainty of the system model and external disturbances. The state update step uses the current observation data to correct the prediction results. The update process calculates the Kalman gain, which determines the weight ratio of the predicted value and the observed value in the final estimate. When the quality of the observation data is good, the gain is biased towards the observed value; when the confidence of the prediction model is high, the gain is biased towards the predicted value. The filtering algorithm iteratively executes the prediction and update steps, ultimately outputting a predicted vector of the equipment state containing predicted values ​​of state parameters such as equipment temperature, vibration, and pressure, as well as prediction uncertainty.

[0073] Deep autoencoders extract features from equipment state prediction vectors. An autoencoder is an unsupervised learning neural network structure consisting of an encoder and a decoder. The encoder maps the high-dimensional input state vector to a low-dimensional hidden space through a multi-layer neural network. Each layer learns an abstract representation of the data through linear transformations and non-linear activation functions. The encoding process progressively reduces the data dimensionality, compressing the multi-dimensional state parameters from the input layer to the core features of the hidden layer. The decoder performs the opposite process, restoring the low-dimensional hidden representation to the original dimensional data. The training process optimizes the network parameters by minimizing the mean square error between the input and the reconstructed output. After training, the encoder can extract the essential features of the data, removing redundant information and noise interference. The feature extraction process inputs the equipment state prediction vector into the trained encoder network and calculates the compressed feature representation through forward propagation. This feature representation contains key information about the equipment's operating status, such as abnormal patterns, trend changes, and correlations, forming the core features of the factory operation that describe the overall operating status of the factory.

[0074] The RSA encryption algorithm generates digital fingerprints for the core characteristics of factory operation. RSA is an asymmetric encryption algorithm based on the mathematical problem of factoring large integers. The digital fingerprint generation first uses a hash function to calculate the digest of the core characteristics of factory operation. The hash function converts feature data of arbitrary length into a digest value of fixed length. The digest calculation process includes data preprocessing, block processing, and compression function iteration to ensure that small changes in the input data will result in significant differences in the digest value. Then, the hash digest is digitally signed using the RSA private key. The signing process uses the digest value as plaintext and uses the private key to perform modular exponentiation to generate the signature value. The security of the RSA algorithm is based on the computational complexity of factoring large composite numbers. The private key contains the product of two large prime numbers as the modulus and the corresponding exponent parameter. The signature verification uses the corresponding public key to perform the inverse operation to verify the validity of the signature and the integrity of the data. The generated digital fingerprint contains the hash digest of the original features, the RSA signature value, the timestamp, and the version identifier, forming a unique digital status fingerprint of the factory that identifies the current operating status of the factory.

[0075] In one specific embodiment, the process of performing deep autoencoder feature extraction on the device state prediction vector may specifically include the following steps:

[0076] The device state prediction vector is input into the encoder network for dimensionality compression to obtain a low-dimensional state code.

[0077] The low-dimensional state code is transformed by a nonlinear activation function to obtain the hidden layer feature representation;

[0078] The hidden layer feature representation is input into the decoder network to calculate the reconstruction error and obtain the feature reconstruction loss value.

[0079] The feature reconstruction loss value is optimized using the backpropagation algorithm to obtain the convergent encoder.

[0080] Based on the convergent state encoder, feature dimensionality reduction processing is performed on multidimensional equipment parameters to obtain the core features of factory operation.

[0081] Specifically, the encoder network performs dimensionality compression on the equipment state prediction vector. It extracts the essential features of the data through progressive dimensionality reduction using a multi-layer neural network. The encoder employs a fully connected layer structure, with each layer containing a weight matrix and a bias vector. The input equipment state prediction vector contains multi-dimensional parameters such as temperature, vibration, pressure, and current, typically ranging from tens to hundreds of dimensions. The first layer of the encoder performs matrix multiplication on the input vector and the weight matrix, calculating the sum of the products of each input feature and its corresponding weight. Then, the bias value is added to obtain the output of that layer. The second layer continues to perform the same linear transformation on the output of the first layer. The number of neurons in each layer gradually decreases, thus achieving progressive dimensionality compression. During the compression process, the network learns to map the high-dimensional original data to a low-dimensional feature space, removing redundant information and noise interference while retaining the most important state information. The final encoding layer outputs a low-dimensional state code containing the core features of the original data, typically compressed to one-tenth or less of the original dimension. The encoder's weight parameters are learned through a training process, with training data derived from historical equipment state data from the factory.

[0082] Nonlinear activation function transformation processes perform nonlinear mapping on low-dimensional state codes. The role of activation functions is to introduce nonlinear factors into neural networks, enabling the network to learn and represent complex nonlinear relationships. Commonly used activation functions include ReLU, Sigmoid, and Tanh. The ReLU function performs thresholding on the input value; when the input is greater than zero, the output equals the input, and when the input is less than or equal to zero, the output is zero. The ReLU function is simple to compute and can alleviate the gradient vanishing problem. The Sigmoid function maps the input to continuous values ​​between 0 and 1, and the output value undergoes an S-shaped curve transformation. The Tanh function maps the input to between -1 and 1, exhibiting symmetry. Activation function transformation processes each element in the low-dimensional state code one by one, converting the linear transformation result into a nonlinear hidden layer feature representation. The transformation process maintains the data dimension unchanged but changes the distribution characteristics of the values. Nonlinear transformation enhances the expressive power of the feature representation, enabling the encoding to capture complex relationships between device states. The hidden layer feature representation contains abstract features that have undergone nonlinear processing. These features have stronger discriminative power and generalization performance than the original linear encoding.

[0083] The decoder network calculates the reconstruction error of the hidden layer feature representation. The decoder's structure is symmetrical to the encoder's. It restores low-dimensional features to high-dimensional data by increasing the number of neurons layer by layer. Each layer of the decoder also uses a weight matrix and bias vector for linear transformation, and then applies an activation function for non-linear processing. The reconstruction process starts from the hidden layer features and gradually restores the dimension and numerical range of the original data. Ideally, the reconstructed output should be as close as possible to the original input. The reconstruction error is measured by calculating the difference between the original input and the reconstructed output. The error is calculated using the mean squared error function, which calculates the sum of the squares of the differences between corresponding elements and then divides it by the total number of elements to obtain the average reconstruction error. The reconstruction error reflects the quality of the features extracted by the encoder and the ability of the decoder to restore the data. The smaller the error value, the more accurate the feature extraction. The training goal of the autoencoder is to minimize the reconstruction error. By adjusting the network parameters, the reconstructed output is made as close as possible to the original input. The feature reconstruction loss value includes not only the mean squared error but also a regularization term. The regularization term prevents the network from overfitting and encourages the learning of concise feature representations.

[0084] Backpropagation is a core algorithm for training neural networks in deep learning. It involves two phases: forward propagation and backpropagation. Forward propagation calculates the network output and loss function value, while backpropagation calculates the gradient of the loss function with respect to the network parameters. The gradient calculation uses the chain rule, propagating the error signal layer by layer from the output layer to the input layer. The gradient of each layer is equal to the product of the gradient of the next layer and the derivative of the activation function of that layer. Parameter updates use the gradient descent algorithm, adjusting the parameters along the negative direction of the gradient. The update step size is controlled by the learning rate, which determines the magnitude of parameter adjustment. An excessively large learning rate leads to training instability, while an excessively small learning rate results in slow convergence. The optimization process iterates for multiple epochs, each traversing all training data. As training progresses, the loss value gradually decreases, and the network parameters gradually converge to their optimal values. Training stops when the magnitude of the loss value change is less than a preset threshold or the maximum number of iterations is reached, resulting in a converged state encoder. This converged encoder possesses stable feature extraction capabilities.

[0085] The convergent state encoder performs feature dimensionality reduction on multi-dimensional equipment parameters. The trained encoder can compress high-dimensional equipment monitoring data into low-dimensional core features. The dimensionality reduction process first standardizes the input multi-dimensional equipment parameters, unifying parameters with different dimensions to the same numerical range to prevent some parameters from dominating the feature extraction process due to excessively large values. Then, the standardized parameters are input into the convergent state encoder, which performs forward propagation calculations according to the trained weight parameters, extracting and compressing feature information layer by layer. The final output is a low-dimensional feature vector containing key information about equipment operation. This feature vector includes core information such as equipment health status, operating mode, and abnormal trends, removing noise and redundancy from the original data. Feature dimensionality reduction solves the computational complexity and storage burden brought by high-dimensional data, while retaining important information for status analysis and fault prediction. The dimensionality-reduced core features of factory operation have good interpretability and discriminative ability, effectively supporting subsequent decision analysis and early warning processing.

[0086] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0087] The digital status fingerprint of the factory is input into the RDF triple knowledge graph for entity relationship matching to obtain the equipment status association graph.

[0088] Based on graph convolutional networks, reasoning and computation are performed on the device state association graph to obtain the risk propagation path vector;

[0089] The risk propagation path vector is quantitatively assessed using the analytic hierarchy process to obtain a numerical risk level.

[0090] The numerical risk level is compared with the preset three-level thresholds of 0.6, 0.8, and 0.9 to determine the risk level classification result.

[0091] The risk level classification results are processed using digital signature technology to obtain a real-time risk level warning code through identity authentication encoding.

[0092] Specifically, the entity relationship matching process of the RDF triple knowledge graph uses the factory's digital status fingerprint as query input to retrieve relevant equipment entities and relationship connections. RDF triples employ a standard subject-verb-object structure to describe knowledge, where the subject represents the entity object, the verb represents the relationship type, and the object represents the attribute value or associated entity. The knowledge graph stores structured knowledge such as attribute information, operating procedures, failure modes, and safety standards for all equipment within the factory. The matching process first parses the equipment identifier, status characteristics, and time information contained in the factory's digital status fingerprint, and then searches for entity nodes related to this information in the knowledge graph. The search process uses SPA (Specific Application Programming Language). RQL query language constructs complex graph pattern matching expressions. The expressions define the entity types, relationship paths, and constraints to be retrieved. The matching algorithm traverses the graph structure to find all triple combinations that satisfy the query conditions. The scattered triples are then reorganized into a subgraph structure centered on the current equipment status. The subgraph contains equipment entity nodes, related other equipment nodes, operating procedure nodes, safety standard nodes, and relational edges connecting these nodes, forming an equipment status association graph that describes the current factory status and its relationships. This graph not only contains direct information about the currently monitored equipment but also information about other equipment and processes that have dependencies, influences, or collaborations with it.

[0093] Graph convolutional networks (GCNNs) perform inference computation on device state association graphs. GCNNs are deep learning models specifically designed for graph-structured data, enabling feature learning and information propagation within the graph's topology. The inference computation process initializes each node in the graph as a feature vector, containing node attribute information such as device type, current state, and historical fault records. The graph convolution operation updates the node's feature representation by aggregating information from its neighboring nodes. The aggregation process calculates a weighted average of the neighboring node features, with weights determined by the edge type and importance. The updated node features not only contain their own information but also incorporate information from surrounding related nodes. This process is iteratively executed across multiple layers of GCNNs, allowing information to propagate and merge over a wider range. Each layer expands the scope of information propagation, ultimately ensuring that each node's feature representation includes information from its multi-hop neighbors in the graph. Inference computation pays particular attention to identifying risk propagation paths, predicting the direction and scope of fault or abnormal state propagation in the device network by analyzing the dependencies and influence relationships between nodes. The computation results output a risk propagation path vector, which encodes the probability distribution and degree of impact on other devices that may be affected starting from the currently abnormal device.

[0094] The Analytic Hierarchy Process (AHP) is used to quantify and assess the risk propagation path vector. AHP is a multi-criteria decision analysis method that systematically assesses complex problems by decomposing them into multiple levels. In this risk quantification assessment, factory safety risks are decomposed into four main dimensions: equipment failure risk, personnel safety risk, production interruption risk, and environmental impact risk. Each dimension is further subdivided into specific risk factors. Equipment failure risk includes sub-factors such as mechanical damage, electrical failure, and control system anomalies; personnel safety risk includes sub-factors such as operational injuries, chemical exposure, and fire / explosion. The assessment process first constructs a hierarchical model to determine the subordinate relationships between factors at each level. Then, the relative importance weights of each factor are determined through expert evaluation or historical data statistics. Weight calculation uses a pairwise comparison matrix, where matrix elements represent the importance ratio between two factors. Normalized weight coefficients are obtained by calculating the maximum eigenvalue and corresponding eigenvector of the matrix. The risk quantification calculation multiplies the risk probability of each device in the risk propagation path vector by its corresponding weight coefficient and then sums the results to obtain a comprehensive numerical risk level. The numerical range is typically between 0 and 1, where 0 represents no risk and 1 represents extremely high risk.

[0095] The three-tiered threshold comparison and judgment process compares the numerical risk level with three preset thresholds of 0.6, 0.8, and 0.9 step by step. The judgment process adopts a ladder-like comparison logic. First, it judges whether the risk level exceeds the first-level warning threshold of 0.6. When the risk value is greater than 0.6, a first-level warning state is triggered, reminding operators to pay attention to changes in equipment status. Next, it judges whether it exceeds the second-level warning threshold of 0.8. When the risk value is greater than 0.8, a second-level warning state is triggered, requiring immediate inspection of relevant equipment and implementation of preventive measures. Finally, it judges whether it exceeds the third-level warning threshold of 0.9. When the risk value is greater than 0.9, a third-level warning state is triggered, initiating emergency response procedures including equipment shutdown, personnel evacuation, and emergency handling. The comparison and judgment also consider the duration and trend of risk. When the risk level rises rapidly in a short period of time, the trigger threshold is lowered; when the risk level changes slowly, the tolerance is appropriately increased. The judgment result includes the current risk level identifier, warning level code, trigger timestamp, and recommended measures, forming a structured risk level classification result.

[0096] Digital signature technology uses risk level classification results for identity authentication encoding. Digital signatures are based on asymmetric encryption algorithms to verify the integrity and authenticity of the data source. The signature processing first uses a hash function to calculate the digest of the risk level classification results. The hash function converts input data of arbitrary length into a fixed-length digest value. The digest calculation process includes three stages: data preprocessing, block compression, and output generation. In the preprocessing stage, padding bits are added to the end of the original data to ensure the total length meets the algorithm requirements. In the block compression stage, the data is divided into fixed-size blocks and compressed block by block. In the output generation stage, the final compressed result is used as the data digest. Yes, then the digest value is encrypted and signed using the private key of the issuing party. The signature algorithm adopts RSA or elliptic curve digital signature algorithm. The digest is taken as plaintext input, and the signature value is generated by mathematical transformation using the private key. The recipient uses the corresponding public key to verify the validity of the signature. The verification process decrypts the signature value with the public key to obtain the digest. At the same time, the digest of the received data is recalculated, and the consistency of the two digest values ​​is compared to determine whether the data has been tampered with. The encoding process combines the original risk level classification result, digital signature value, timestamp and issuing party identifier into a complete authentication data packet to generate a real-time risk level warning code containing identity verification information.

[0097] The above describes the IoT-based real-time digital monitoring method for factories in the embodiments of this application. The following describes the IoT-based real-time digital monitoring system for factories in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the IoT-based real-time digital monitoring system for factories in this application includes:

[0098] The encryption module is used to perform encrypted access processing on heterogeneous IoT devices through a dual authentication mechanism to obtain a secure authentication device identifier pool.

[0099] The data acquisition module is used to perform multi-dimensional data acquisition and processing on the security authentication device identifier pool based on hash verification integrity verification to obtain an encrypted time-series data chain.

[0100] The recognition module is used to dynamically recognize factory visual images using a confidence-adaptive threshold mechanism to obtain multi-task fusion feature codes.

[0101] The fusion module is used to perform heterogeneous fusion processing on the encrypted time-series data chain and the multi-task fusion feature code according to the sliding window weight allocation algorithm to obtain the factory digital status fingerprint.

[0102] The reasoning module is used to perform knowledge graph reasoning processing on the digital status fingerprint of the factory through a three-level early warning threshold to obtain a real-time risk level early warning code.

[0103] above Figure 2The IoT-based real-time digital monitoring system for factories in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The IoT-based real-time digital monitoring equipment for factories in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0104] Reference Figure 3 This invention also provides an IoT-based real-time digital monitoring device for factories. This IoT-based real-time digital monitoring device can be a server, and its internal structure can be as follows: Figure 3 As shown, this IoT-based real-time digital monitoring device for factories includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of this IoT-based real-time digital monitoring device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of this IoT-based real-time digital monitoring device stores the data corresponding to this embodiment. The network interface of this IoT-based real-time digital monitoring device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0105] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the IoT-based real-time digital monitoring equipment for factories to which the present invention is applied.

[0106] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the IoT-based factory real-time digital monitoring method.

[0107] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0108] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an IoT-based real-time digital monitoring device for factories (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time digital monitoring of factories based on the Internet of Things, characterized in that, The method includes: A secure authentication device identifier pool is obtained by encrypting the access of heterogeneous IoT devices through a dual authentication mechanism. The secure authentication device identifier pool is subjected to multi-dimensional data acquisition and processing based on hash verification integrity verification to obtain an encrypted time-series data chain. This includes: synchronously acquiring data from temperature sensors, vibration sensors, and pressure sensors at 100-millisecond intervals based on the secure authentication device identifier pool to obtain raw device operation data; performing SHA-256 hash value calculation on the raw device operation data to obtain a data packet integrity fingerprint; performing difference verification processing between the data packet integrity fingerprint and the device's historical data fingerprint to obtain a data quality assessment result; filtering abnormal data from the data quality assessment result based on a data retransmission trigger mechanism to obtain trusted device monitoring data; and performing time-series chain encoding processing on the trusted device monitoring data based on the factory production cycle timestamp to obtain the encrypted time-series data chain. The factory visual images are dynamically recognized based on the confidence-adaptive threshold mechanism to obtain multi-task fusion feature codes. Heterogeneous fusion processing is performed on the encrypted time-series data chain and the multi-task fusion feature code according to the sliding window weight allocation algorithm to obtain the factory digital status fingerprint, including: Spatiotemporal alignment processing is performed on the encrypted time-series data chain and the multi-task fusion feature code based on a 30-second sliding window to obtain a synchronous heterogeneous data matrix; dynamic weight allocation processing is performed on the synchronous heterogeneous data matrix according to the sensor reliability score to obtain a weighted fusion dataset; the weighted fusion dataset is input into an improved Kalman filter algorithm for state prediction processing to obtain an equipment state prediction vector; deep autoencoder feature extraction processing is performed on the equipment state prediction vector to obtain the core features of factory operation; digital fingerprint generation processing is performed on the core features of factory operation based on the RSA encryption algorithm to obtain the digital state fingerprint of the factory. The factory's digital status fingerprint is subjected to knowledge graph reasoning processing based on a three-level early warning threshold to obtain a real-time risk level early warning code.

2. The IoT-based real-time digital monitoring method for factories according to claim 1, characterized in that, The encrypted access process for heterogeneous IoT devices through a dual authentication mechanism yields a secure authentication device identifier pool, including: Perform MAC address verification on heterogeneous IoT devices to obtain the device physical identifier code; The device physical identifier is digitally authenticated using an elliptic curve cryptography algorithm to obtain a device security credential. The device security credentials are entered into a pre-stored device whitelist for matching and verification to obtain the device access permission identifier. The device access permission identifier is encrypted and authorized according to the temporary session key allocation mechanism to obtain the device communication key pair; The security authentication device identifier pool is obtained by performing unified identifier encoding processing on the device communication key pair.

3. The IoT-based real-time digital monitoring method for factories according to claim 1, characterized in that, The process of dynamically recognizing factory visual images based on a confidence-adaptive threshold mechanism to obtain multi-task fusion feature codes includes: Adaptive enhancement preprocessing is performed on the factory visual images to obtain illumination-normalized image data; The illumination-normalized image data is input into an improved YOLOv8 network for multi-task parallel detection processing to obtain the device detection box, personnel segmentation mask, and pose key point coordinates. Based on the historical recognition accuracy, the confidence threshold of the device detection frame, personnel segmentation mask, and posture key point coordinates is dynamically adjusted to obtain adaptive detection results. The adaptive detection results are processed in real time for 50 milliseconds using TensorRT acceleration technology to obtain the factory scene recognition feature vector. The factory scene recognition feature vector is subjected to multi-task fusion encoding processing to obtain the multi-task fusion feature code.

4. The IoT-based real-time digital monitoring method for factories according to claim 1, characterized in that, The deep autoencoder feature extraction process is performed on the equipment state prediction vector to obtain the core features of factory operation, including: The device state prediction vector is input into the encoder network for dimensionality compression to obtain a low-dimensional state code. The low-dimensional state code is subjected to nonlinear activation function transformation to obtain the hidden layer feature representation; The hidden layer feature representation is input into the decoder network for reconstruction error calculation to obtain the feature reconstruction loss value; The feature reconstruction loss value is optimized using the backpropagation algorithm to obtain a convergent encoder. Based on the convergent state encoder, feature dimensionality reduction processing is performed on the multidimensional equipment parameters to obtain the core features of the factory operation.

5. The IoT-based real-time digital monitoring method for factories according to claim 1, characterized in that, The process of using a knowledge graph to reason about the factory's digital status fingerprint through a three-level early warning threshold determination to obtain a real-time risk level early warning code includes: The digital status fingerprint of the factory is input into the RDF triple knowledge graph for entity relationship matching to obtain the equipment status association graph. The device state association graph is inferred and processed using a graph convolutional network to obtain the risk propagation path vector. The risk propagation path vector is subjected to risk quantification assessment using the analytic hierarchy process to obtain a numerical risk level. The numerical risk level is compared with the preset three-level thresholds of 0.6, 0.8, and 0.9 to determine the risk level classification result. The risk level classification results are processed using digital signature technology to obtain the real-time risk level warning code through identity authentication encoding.

6. A real-time digital monitoring system for factories based on the Internet of Things, characterized in that, For implementing the IoT-based real-time digital monitoring method for factories as described in any one of claims 1-5, the IoT-based real-time digital monitoring system for factories comprises: The encryption module is used to perform encrypted access processing on heterogeneous IoT devices through a dual authentication mechanism to obtain a secure authentication device identifier pool. The data acquisition module is used to perform multi-dimensional data acquisition and processing on the security authentication device identifier pool based on hash verification integrity verification to obtain an encrypted time-series data chain. This includes: synchronously acquiring data from temperature sensors, vibration sensors, and pressure sensors at 100-millisecond intervals based on the security authentication device identifier pool to obtain raw device operation data; performing SHA-256 hash value calculation on the raw device operation data to obtain a data packet integrity fingerprint; performing difference verification processing between the data packet integrity fingerprint and the device's historical data fingerprint to obtain a data quality assessment result; performing abnormal data filtering processing on the data quality assessment result according to a data retransmission triggering mechanism to obtain trusted device monitoring data; and performing time-series chain encoding processing on the trusted device monitoring data based on the factory production cycle timestamp to obtain the encrypted time-series data chain. The recognition module is used to dynamically recognize and process factory visual images based on a confidence-adaptive threshold mechanism to obtain multi-task fusion feature codes. The fusion module is used to perform heterogeneous fusion processing on the encrypted time-series data chain and the multi-task fusion feature code according to a sliding window weight allocation algorithm to obtain a factory digital status fingerprint. This includes: performing spatiotemporal alignment processing on the encrypted time-series data chain and the multi-task fusion feature code based on a 30-second sliding window to obtain a synchronous heterogeneous data matrix; performing dynamic weight allocation processing on the synchronous heterogeneous data matrix according to sensor reliability scores to obtain a weighted fusion dataset; inputting the weighted fusion dataset into an improved Kalman filter algorithm for state prediction processing to obtain an equipment state prediction vector; performing deep autoencoder feature extraction processing on the equipment state prediction vector to obtain core factory operation features; and performing digital fingerprint generation processing on the core factory operation features based on an RSA encryption algorithm to obtain the factory digital status fingerprint. The reasoning module is used to perform knowledge graph reasoning processing on the digital status fingerprint of the factory through a three-level early warning threshold to obtain a real-time risk level early warning code.

7. A factory real-time digital monitoring device based on the Internet of Things, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the IoT-based real-time digital monitoring method for factories as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the IoT-based real-time digital monitoring method for factories as described in any one of claims 1 to 5.

Images