Electronic lead seal automatic detection method applied to logistics tracking

Abstract

The invention discloses an electronic lead seal automatic detection method applied to logistics tracking, and relates to the technical field of Internet of Things safety monitoring. The problems of packaging data loss, tampering risk increase and real-time monitoring failure caused by wireless communication signal interference, transmission delay and state updating lag of an existing electronic lead seal are solved. The method comprises the following steps: fusing multi-physical field sensing data through a space-time correlation sampling algorithm, and constructing an anti-interference characteristic matrix; evaluating communication quality based on the dynamic probability network model, and triggering a multi-path fragmentation concurrent transmission strategy to avoid signal attenuation; a hybrid reasoning model is deployed at an edge end to screen key event data, and the transmission efficiency is optimized in combination with differential coding and an IEEE 1588 clock synchronization mechanism; the cloud end adopts a space-time diagram fusion analysis model to carry out cross-modal abnormal association scoring, corrects misjudgment and updates an edge model; according to the invention, the communication reliability, the real-time transmission efficiency and the anomaly detection precision of the lead sealing state data in a complex environment are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things security monitoring, and more specifically to an automatic detection method for electronic seals applied to logistics tracking. Background Art

[0002] With the rapid development of globalization and information technology, the safe management of asset circulation has become the focus of enterprises. Against this background, electronic seal technology has emerged, providing strong support for the safety of asset circulation with its real-time monitoring, data tracking and intelligent early warning. Especially in the fields of logistics, warehousing, energy and finance, electronic seals have become a key technical means to ensure the safety of important materials.

[0003] The current automated detection of electronic seals mainly relies on the coordination of radio frequency identification (RFID), sensor networks and automation equipment. For example, Chinese patent CN117392680A proposes to automatically locate the seal position by carrying a camera on a robotic arm, and complete the seal number inspection in combination with image recognition technology to reduce manual intervention; patent CN118968204B uses blockchain technology to store the seal operation log to ensure that the data cannot be tampered with. At the hardware level, the existing technology generally integrates RFID tags (such as EPC Class 1Gen 2 protocol) and temperature and humidity sensors, and transmits status data to the cloud platform in real time through NB-IoT or 4G modules (patent CN119648102A). In the automated sealing process, patent CN119692890A uses load sensors and path optimization algorithms to prevent illegal loading and unloading of goods. These technologies not only improve the detection efficiency of seals, but also enhance the security and reliability of the system. However, the existing methods have the following defects in practical applications, which affect the stability and security of the logistics tracking system.

[0004] First, low-power wireless communication technologies (LoRa, NB-IoT) are vulnerable to external signal interference, resulting in data loss or transmission failure. For example, in a high-density metal container environment, LoRa signals are easily attenuated or reflected, causing the electronic lead seal to be unable to stably upload the packaging status data. In addition, when NB-IoT switches base stations or in areas with weak signal coverage (such as ocean transportation, mountain warehouses), there may be interruptions or delays in the transmission of packaging status data, affecting the real-time monitoring of packaging integrity. Second, there is a lag problem in the update of the packaging status, which affects the timeliness and accuracy of the data. Since LoRa adopts an Adaptive Data Rate (ADR) mechanism, when the network load is high, the data rate will decrease, resulting in an increased time interval for uploading packaging status data. Similarly, when NB-IoT devices enter the Power Saving Mode (PSM) or Extended Discontinuous Reception Mode (eDRX), it may take several minutes or even longer to update the packaging status data, increasing the risk of tampering. In addition, in a large-scale logistics environment, network congestion may cause the failure or delay of electronic lead seal data upload, making the logistics management system unable to obtain the packaging status information in a timely manner, affecting the accuracy of logistics tracking and the safety of goods. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention discloses an automated detection method for electronic lead seals applied to logistics tracking, aiming to solve the problems raised in the background technology.

[0006] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] An automated detection method for electronic lead seals applied to logistics tracking, comprising the following steps:

[0008] Step 1: Obtain lead seal status data through a spatio-temporal correlation sampling algorithm, and perform block preprocessing using a redundant coding mechanism. The lead seal status includes vibration signals, light interruption status, ambient temperature and humidity, and three-axis acceleration vectors.

[0009] Step 2: Perform wavelet packet decomposition on the vibration signal, extract the spectral energy ratio of the high-frequency subband and the low-frequency subband, and combine the pulse width of the light interruption status, the temperature and humidity change gradient, and the three-axis acceleration vector to construct an anti-interference feature matrix.

[0010] Step 3: Based on the anti-interference feature matrix, the real-time collected LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density, calculate the comprehensive channel quality index through a dynamic probability network model, and dynamically adjust the transmission strategy.

[0011] Step 4: Input the anti-interference feature matrix into the hybrid inference model deployed at the edge side to output the confidence score of the seal status; if the confidence is higher than the preset threshold or meets the vibration and light sensor combined trigger condition, compress and encode the current window data through the streaming compression algorithm, and upload the event-driven compressed package and hash value to the cloud; otherwise, upload the hash value.

[0012] Step 5: Apply IEEE 1588PTP clock synchronization to the spatio-temporal associated sampling data in Step 1 based on the GPS second pulse signal; and monitor the LoRa and NB-IoT network load rates in real time. If the load rate is higher than the preset threshold, enable differential encoding for the data stream output in Step 4, and only transmit the incremental data of adjacent windows to the cloud, and dynamically adjust the sleep cycle of the electronic seal communication module in combination with the confidence score of the seal status.

[0013] Step 6: Input the uploaded data in Step 4 into the spatio-temporal graph fusion analysis model, fuse the vibration spectrum, light sensor timing, temperature and humidity curve, and acceleration trajectory to generate a cross-modal anomaly association score, and correct the misjudgment events at the edge side; and update the edge model based on the verified abnormal samples using the distributed parameter aggregation framework.

[0014] As a further technical solution of the present invention, the working method of Step 2 is as follows:

[0015] S1: Perform 5-layer wavelet packet decomposition on the vibration signal through the D wavelet basis function to generate 32 frequency band nodes, and reorder the nodes in the 5th layer in ascending order of frequency through the node reordering algorithm.

[0016] S2: Based on the reconstructed sub-band signals, calculate the band energy ratios of the high-frequency band nodes from node 16 to node 31 and the low-frequency band nodes from node 0 to node 15 using the generalized relative wavelet energy ratio. The calculation formula is:

[0017]

[0018] In formula (1), Hilbert represents the square of the norm, which is used to capture the instantaneous energy of non-stationary signals. δ is an anti-zero division perturbation term, and d 5,i represents the multi-resolution characterization unit of the signal in time-frequency analysis;

[0019] S3: Parse the light sensor occlusion status using the pulse width modulation decoding mechanism, and extract the timestamp and pulse width parameters of the occlusion event.

[0020] S4: Calculate the gradient value of the temperature and humidity sensor through the sliding window difference method, and construct a four-dimensional auxiliary feature vector in combination with the Euler angle change amount of the three-axis acceleration vector.

[0021] S5. Align the band energy ratio, pulse width parameter, temperature and humidity gradient, and the change in Euler angles of the three-axis acceleration vector according to a time window. Eliminate the dimensional difference through Z-score standardization, and perform orthogonal dimensionality reduction on the multi-dimensional features using principal component analysis to generate a low-dimensional feature matrix with anti-interference. During the standardization process, if the high-frequency energy ratio exceeds the preset threshold, weight and enhance the three-axis acceleration vector to suppress mechanical noise.

[0022] As a further technical solution of the present invention, the dynamic probability network model includes a channel state evaluation layer, an interference perception layer, a link quality prediction layer, and a multi-objective fragmentation scheduling layer. The channel state evaluation layer is used to calculate the channel attenuation factor based on the LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density using statistical analysis methods, and predict the change trend of the channel state through time series modeling, and output the current channel stability parameter. The interference perception layer is used to identify interference sources through frequency domain analysis methods according to the signal spectrum distribution, and extract signal components through filtering algorithms, and output the interference intensity and interference influence range. The link quality prediction layer is used to calculate the future link signal change trend based on historical link data using an autoregressive model, and correct it in combination with real-time measurement values, and output the link reliability evaluation index. The multi-objective fragmentation scheduling layer is used to perform data fragmentation based on the data block size and channel availability, and use a multi-path fragmentation concurrent transmission strategy to allocate data blocks to different paths for transmission, and achieve concurrent transmission of Bluetooth Mesh ad hoc network and LoRa and NB-IoT links through a double-buffer mechanism.

[0023] As a further technical solution of the present invention, the working method of the multi-path fragmentation concurrent transmission strategy is as follows:

[0024] U1. Divide the data block into N original sub-blocks and M redundant sub-blocks through the RS erasure code fragmentation algorithm to generate an anti-loss fragmentation set, where N:M = 4:1.

[0025] U2. Based on the Bluetooth Mesh node topology structure, use a graph attention network to analyze the signal strength, hop count, and historical packet loss rate of each Bluetooth node, and select the optimal relay path set.

[0026] U3. Allocate the fragmented sub-blocks to the Bluetooth Mesh ad hoc network and LoRa and NB-IoT communication links for parallel transmission through a multi-threaded concurrent control protocol. If the RSSI of the Bluetooth Mesh path is lower than the preset threshold, migrate the sub-block to the sub-optimal path through a dynamic path switching mechanism.

[0027] U4. The receiving end verifies the integrity of the sub-blocks through a reverse hash check tree. If the number of lost sub-blocks exceeds the preset redundancy threshold, reconstruct the original data block based on the erasure code; at the same time, align the multi-path transmission timing through an end-to-end timestamp synchronization protocol.

[0028] As a further technical solution of the present invention, the vibration and light sensing combined trigger condition defines the time-domain integral value E of the vibration energy based on the mechanical deformation mechanical model of the lead seal v and the dynamic correlation function of the light sensing interruption pulse width T d The calculation formula of E v is as follows:

[0029]

[0030] In formula (2), E v represents the integral of the square of the modulus of the vibration acceleration vector a(t) collected by the triaxial acceleration sensor within the time window [t0, t1], which is used to quantify the impact energy intensity of the illegal opening event; when and T d > ργθ, it is determined as an illegal opening event, where a is the reference energy threshold, β is the interruption attenuation coefficient, ρ is the real-time metal density, γ is the metal density compensation factor, and θ is the environmental temperature and humidity compensation function. Among them, a, β, and γ are dynamically adjusted according to the type of transportation tool and the environmental temperature and humidity

[0031] As a further technical solution of the present invention, the working principle of step 5 is as follows: First, through the IEEE 1588v2 two-step synchronization mechanism, the Sync message is used to carry the hardware timestamp to record the sending moment u1 of the GPS second pulse source. The gateway side uses the transparent clock mode to calculate and correct the residence time c1 of the transmission path in real time, and superimposes the residence time c1 on the correction domain of the follow-up message. Combining the reception time u2 of the slave port, multi-sensor timestamp alignment is achieved based on formula (3):

[0032] R = u2 - u1 - D - c1 (3)

[0033] In formula (3), D is the preset path delay reference value; then, through the link layer probe of the LoRaWAN and NB-IoT dual-mode communication module, the MAC frame queue length and channel occupancy rate are statistically calculated in real time. If the load rate exceeds the preset threshold, the differential coding algorithm is triggered to split the sensor data stream according to the time window, and only the incremental data with the change amount exceeding the quantization threshold within adjacent windows is retained, and the Huffman compression coding table index is embedded; at the same time, based on the anomaly detection confidence feedback from the cloud, the sleep cycle of the communication module is dynamically adjusted. When the confidence is lower than 0.7, the exponential backoff algorithm is used to shorten the sleep interval to 1 / 2 of the reference value n , where n is the number of consecutive anomalies, otherwise it is extended to 2 n times the reference value

[0034] As a further technical solution of the present invention, the spatio-temporal graph fusion analysis model includes a spatio-temporal embedding alignment layer, a multi-modal graph convolutional layer, a cross-modal attention fusion layer, a spatio-temporal residual encoding layer, an anomaly correlation scoring layer, and a federated incremental learning layer; the spatio-temporal embedding alignment layer is used to align multi-modal time series according to the original timestamps of vibration spectra, light sensor timings, temperature and humidity curves, and acceleration trajectories through the dynamic time warping algorithm, and map them to a unified spatio-temporal coordinate system by combining the self-attention mechanism, and output a spatio-temporal alignment tensor; the multi-modal graph convolutional layer is used to fuse vibration frequency domain energy, light sensor pulse density, temperature and humidity gradient, and acceleration covariance matrix through a dynamic graph attention network, and model the spatio-temporal propagation delay using a gated recurrent unit, and output a spatio-temporal feature matrix; the cross-modal attention fusion layer is used to calculate the cross-modal correlation weights of vibration, light sensor, temperature and humidity, and acceleration through a multi-head cross-attention mechanism, and optimize the fusion weights through cosine similarity and KL divergence, and output a cross-modal fusion tensor; the spatio-temporal residual encoding layer is used to extract multi-scale time series features through dilated temporal convolution according to the cross-modal fusion tensor, and capture spatial dependence relationships through a graph residual network, and suppress noise using a gated linear unit, and output residual encoding features; the anomaly correlation scoring layer is used to calculate the node anomaly propagation probability through a graph random walk algorithm, and detect cross-modal joint outliers through an isolation forest, and output an anomaly correlation score; the federated incremental learning layer is used to aggregate feature distributions and fine-tune model parameters through a differential privacy federated average algorithm according to the anomaly correlation score and misjudged samples at the edge, and output an incrementally updated spatio-temporal graph model weight matrix.

[0035] As a further technical solution of the present invention, the distributed parameter aggregation framework performs homomorphic encryption on the edge model gradient parameters through a secure aggregation protocol to generate a set of ciphertext gradients; and based on a dynamic weight allocation mechanism, calculates weight coefficients according to the historical misjudgment rates, metal densities, and temperature and humidity of each node, and weighted aggregates the set of ciphertext gradients; when the Mahalanobis distance of the node parameters is detected to be higher than or equal to a preset threshold, the distributed parameter aggregation framework eliminates outlier parameters through an anomaly gradient filtering algorithm; after aggregation, a gradient sparsification compression method is used to retain the top 10% of the maximum amplitude gradients, and the rest are set to zero and compressed through Huffman coding to reduce the transmission volume; finally, Gaussian noise is injected into the aggregated gradients through a differential privacy mechanism to generate privacy-protected global model parameters, which are sent to the edge.

[0036] Based on the above technical solutions, the positive and beneficial effects of the present invention are as follows:

[0037] 1. The quality of the multi-mode communication channel is evaluated in real time through a dynamic probability network model, and combined with the adaptive switching strategy of Bluetooth Mesh self-organizing network and the main link, the electromagnetic shielding effect of high-density metal containers and the transmission interruption problems in weak signal areas such as the open ocean and mountainous areas are effectively avoided, ensuring the stable upload of the lead seal status data in a strong interference environment.

[0038] 2. The hardware-level clock synchronization mechanism based on the GPS second pulse signal and the IEEE 1588 protocol, combined with differential coding technology to transmit only incremental data, greatly reduces the transmission load in the network congestion scenario; through confidence-driven dynamic adjustment of the sleep cycle, it optimizes the real-time response ability of the low-power communication module, significantly shortens the state update delay, and reduces the risk of illegal tampering.

[0039] 3. Utilize wavelet packet decomposition and cross-modal feature fusion technology to extract the anti-interference feature matrix, combined with the collaborative verification mechanism of the edge-side hybrid inference model and the cloud spatio-temporal graph fusion analysis model, effectively distinguish real illegal opening events from complex environmental noises (such as mechanical vibration, electromagnetic interference, sudden changes in temperature and humidity), and greatly reduce misjudgment and missed detection.

[0040] 4. Optimize the edge model through the distributed parameter aggregation framework for federated learning, continuously use the abnormal samples verified by the cloud to iteratively update the model parameters, enabling the system to adapt to different logistics scenarios (such as cross-climate transportation, multi-vehicle switching), and breaking through the generalization bottleneck of traditional static models in dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings, where:

[0042] Figure 1 is the architecture diagram of an automatic detection method for electronic seals applied to logistics tracking according to the present invention;

[0043] Figure 2 is the working principle architecture diagram of step 1 of the present invention;

[0044] Figure 3 is the working principle step diagram of step 2 of the present invention;

[0045] Figure 4 is the architecture diagram of the dynamic probability network model of the present invention;

[0046] Figure 5 is the working method step diagram of the multi-path sharding concurrent transmission strategy of the present invention;

[0047] Figure 6 is the structural framework diagram of the hybrid inference model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] In an embodiment, an automated detection method for an electronic seal applied to logistics tracking is as Figure 1 shown, including;

[0050] Step 1: Obtain seal status data through a spatio-temporal correlation sampling algorithm, and perform block preprocessing using a redundant coding mechanism. The seal status includes vibration signals, light interruption status, ambient temperature and humidity, and three-axis acceleration vectors; as Figure 2 shown, the specific working principle is as follows: The multi-physical field perception module uses vibration sensors, an optoelectronic switch array, temperature and humidity sensors, and a MEMS three-axis accelerometer for data acquisition, and uses a clock synchronization protocol to eliminate the time reference drift between multiple sensors, and outputs a multi-source data stream; uses a spatio-temporal correlation sampling algorithm, and fuses the observation errors of each sensor in the spatio-temporal dimension through an adaptive weighted Kalman filter to generate a composite state matrix including vibration spectrum, light intensity mutation events, environmental parameters, and acceleration vectors; for the light interruption signal, the spatio-temporal correlation sampling algorithm uses pulse width modulation coding to convert the light intensity threshold mutation into binary timing data, and at the same time combines the Euler angle transformation matrix of the three-axis acceleration vector for spatial grid coding to realize the spatio-temporal coupling representation of optical-mechanical events and motion states; then, based on the redundant coding mechanism of the Reed-Solomon error correction algorithm in Step 1, the composite data stream is divided into fixed-length data blocks, and a CRC-32 cyclic redundancy check code and forward error correction redundant bits are embedded in each data block to construct a transmission unit; at the same time, when the temperature and humidity exceed the preset safety threshold in Step 1, a non-uniform block strategy is adopted, with small-grained blocks for high-frequency vibration signals and large-grained blocks for low-frequency environmental parameters; finally, a preprocessed data packet including spatio-temporal tags, physical field feature vectors, and redundant check bits is output.

[0051] Step 2: Perform wavelet packet decomposition on the vibration signal, extract the spectral energy ratio of the high-frequency subband and the low-frequency subband, and combine the pulse width of the light interruption status, the temperature and humidity change gradient, and the three-axis acceleration vector to construct an anti-interference feature matrix; as Figure 3 shown, the specific working method is as follows:

[0052] S1: Perform 5-layer wavelet packet decomposition on the vibration signal through the D wavelet basis function to generate 32 frequency band nodes, and reorder the nodes of the 5th layer in ascending order of frequency through a node reordering algorithm;

[0053] S2. Based on the reconstructed sub-band signals, calculate the band energy ratios of the high-frequency band from node 16 to node 31 and the low-frequency band from node 0 to node 15 using the generalized relative wavelet energy ratio. The calculation formula is as follows:

[0054]

[0055] In formula (1), Hilbert represents the square of the norm, which is used to capture the instantaneous energy of non-stationary signals. δ is an anti-zero-division perturbation term, and d 5,i represents the multi-resolution characterization unit of the signal in time-frequency analysis;

[0056] S3. Use the pulse-width modulation decoding mechanism to analyze the light-sensing blocking state, and extract the timestamp and pulse-width parameters of the blocking event;

[0057] S4. Calculate the gradient value of the temperature and humidity sensor by the sliding window difference method, and construct a four-dimensional auxiliary feature vector in combination with the Euler angle change amount of the three-axis acceleration vector;

[0058] S5. Align the band energy ratios, pulse-width parameters, temperature and humidity gradients, and Euler angle change amounts of the three-axis acceleration vector according to the time window. Eliminate the dimension difference through Z-score standardization, and use principal component analysis to orthogonalize and reduce the dimension of the multi-dimensional features to generate an anti-interference low-dimensional feature matrix. During the standardization process, if the high-frequency energy ratio exceeds the preset threshold, then weight and enhance the three-axis acceleration vector to suppress mechanical noise.

[0059] Step 3. Based on the anti-interference feature matrix, the LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density collected in real time, calculate the comprehensive channel quality index through a dynamic probability network model, and dynamically adjust the transmission strategy; among them, such as Figure 4As shown in the figure, the dynamic probability network model includes a channel state evaluation layer, an interference perception layer, a link quality prediction layer, and a multi-objective fragmentation scheduling layer. The channel state evaluation layer is used to calculate the channel attenuation factor by using statistical analysis methods based on the LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density, and predict the change trend of the channel state through time series modeling, and output the current channel stability parameter. The interference perception layer is used to identify interference sources by using frequency domain analysis methods according to the signal spectrum distribution, and extract signal components through filtering algorithms, and output the interference intensity and interference range. The link quality prediction layer is used to calculate the future link signal change trend by using an autoregressive model based on historical link data, and correct it in combination with real-time measurement values, and output the link reliability evaluation index. The multi-objective fragmentation scheduling layer is used to perform data fragmentation based on the data block size and channel availability, and adopt a multi-path fragmentation concurrent transmission strategy to allocate data blocks to different paths for transmission, and realize the concurrent transmission of Bluetooth Mesh self-organizing network and LoRa and NB-IoT links through a double buffering mechanism. As Figure 5 shown, the working method of the multi-path fragmentation concurrent transmission strategy is as follows:

[0060] U1. Divide the data block into N original sub-blocks and M redundant sub-blocks through the RS erasure code fragmentation algorithm to generate an anti-loss fragmentation set, where N:M = 4:1;

[0061] U2. Based on the Bluetooth Mesh node topology structure, use a graph attention network to analyze the signal strength, hop count, and historical packet loss rate of each Bluetooth node, and select the optimal relay path set;

[0062] U3. Allocate the fragmented sub-blocks to the Bluetooth Mesh self-organizing network and the LoRa and NB-IoT communication links for parallel transmission through a multi-threaded concurrent control protocol. If the RSSI of the Bluetooth Mesh path is lower than the preset threshold, migrate the sub-block to the sub-optimal path through a dynamic path switching mechanism;

[0063] U4. The receiving end verifies the integrity of the sub-blocks through a reverse hash check tree. If the number of lost sub-blocks exceeds the preset redundancy threshold, reconstruct the original data block based on the erasure code; at the same time, align the multi-path transmission timing through an end-to-end timestamp synchronization protocol.

[0064] Step 4. Input the anti-interference feature matrix into the hybrid inference model deployed at the edge side, and output the confidence score of the seal state; if the confidence is higher than the preset threshold or meets the vibration and light sensor joint trigger condition, compress and encode the current window data through a streaming compression algorithm, and upload the event-driven compression packet and hash value to the cloud; otherwise, upload the hash value; As Figure 6As shown in the figure, the hybrid inference model includes a feature preprocessing layer, a physical rule embedding layer, a multi-modal feature fusion layer, a lightweight inference layer, a confidence calibration layer, and an event trigger layer; the feature preprocessing layer is used to smooth the input anti-interference feature matrix through sliding window mean filtering, and normalize each feature dimension to zero mean unit variance using the Z-score normalization algorithm to generate a normalized feature vector; the physical rule embedding layer is used to define the joint determination rule of vibration energy and blocking duration based on the lead seal mechanical structure and light sensing blocking logic, and encode the rule matching result as a binary identifier, which is concatenated with the normalized feature vector; the multi-modal feature fusion layer is used to model the cross-modal correlation of vibration spectrum, temperature and humidity gradient, and acceleration vector through the multi-head attention mechanism, calculate the attention weight of each feature channel, and output a high-dimensional feature tensor; the lightweight inference layer is used to input the fused feature tensor into the knowledge distillation model, perform convolution and depthwise separable convolution operations, extract the spatio-temporal joint features, and output the original confidence score; the confidence calibration layer is used to dynamically correct the original confidence using the Bayesian probability calibration algorithm, combining the historical abnormal sample distribution and the current environmental parameters; the event trigger layer is used to compress and encode the original data window through the LZ77 streaming compression algorithm when the confidence is higher than the preset threshold or the rule is triggered, generate an event-driven compressed package and extract the data block hash value to upload to the cloud; when the confidence is lower than or equal to the preset threshold and the rule is not triggered, extract the data block hash value to upload to the cloud. The vibration and light sensing joint trigger condition is based on the lead seal mechanical deformation mechanics model to define the time-domain integral value E of the vibration energy. v And the light sensing blocking pulse width T d The dynamic correlation function of, E v The calculation formula of is:

[0065]

[0066] In formula (2), E v Represents the integral of the square of the modulus of the vibration acceleration vector a(t) collected by the triaxial acceleration sensor within the time window [t0, t1], which is used to quantify the impact energy intensity of the illegal opening event; when And T d > ργθ, it is determined as an illegal opening event, a is the reference energy threshold, β is the blocking attenuation coefficient, ρ is the real-time metal density, γ is the metal density compensation factor, θ is the environmental temperature and humidity compensation function, where a, β, γ are dynamically adjusted according to the transportation tool type and environmental temperature and humidity.

[0067] Step 5: Apply IEEE 1588 PTP clock synchronization to the spatio-temporal correlation sampling data in Step 1 based on the GPS second pulse signal; and monitor the LoRa and NB-IoT network load rates in real time. If the load rate is higher than the preset threshold, enable differential coding for the data stream output in Step 4, and only transmit the incremental data of adjacent windows to the cloud, and dynamically adjust the sleep cycle of the electronic seal communication module in combination with the confidence score of the seal status. The specific working principle is as follows: First, through the IEEE1588v2 two-step synchronization mechanism, use the Sync message to carry the hardware timestamp to record the sending time u1 of the GPS second pulse source. In the gateway side, use the transparent clock mode to calculate and correct the residence time C1 of the transmission path in real time, and superimpose the residence time c1 on the correction field of the follow-up message. Combine the reception time u2 of the slave port, and achieve multi-sensor timestamp alignment based on formula (3):

[0068] R = u2 - u1 - D - c1 (3)

[0069] In formula (3), D is the preset path delay reference value; then, through the link layer probe of the LoRaWAN and NB-IoT dual-mode communication module, statistically calculate the MAC frame queue length and channel occupancy rate in real time. If the load rate exceeds the preset threshold, trigger the differential coding algorithm, divide the sensor data stream by time window, and only retain the incremental data whose change amount in adjacent windows exceeds the quantization threshold, and embed the Huffman compression coding table index; at the same time, based on the anomaly detection confidence feedback from the cloud, dynamically adjust the sleep cycle of the communication module. When the confidence is lower than 0.7, use the exponential backoff algorithm to shorten the sleep interval to 1 / 2 of the reference value n , where n is the number of consecutive anomalies, otherwise extend it to 2 n times the reference value.

[0070] Step 6: Input the uploaded data in Step 4 into the spatio-temporal graph fusion analysis model, fuse the vibration spectrum, light sensor time series, temperature and humidity curve, and acceleration trajectory to generate a cross-modal anomaly correlation score, and correct the misjudgment events at the edge; and based on the verified anomaly samples, update the edge model using the distributed parameter aggregation framework. The spatio-temporal graph fusion analysis model includes a spatio-temporal embedding alignment layer, a multi-modal graph convolution layer, a cross-modal attention fusion layer, a spatio-temporal residual encoding layer, an anomaly correlation scoring layer, and a federated incremental learning layer; the spatio-temporal embedding alignment layer is used to align the multi-modal time series according to the original timestamps of the vibration spectrum, light sensor time series, temperature and humidity curve, and acceleration trajectory through the dynamic time warping algorithm, and map it to a unified spatio-temporal coordinate system in combination with the self-attention mechanism, and output a spatio-temporal alignment tensor; the multi-modal graph convolution layer is used to fuse the vibration frequency domain energy, light sensor pulse density, temperature and humidity gradient, and acceleration covariance matrix through the dynamic graph attention network, and model the spatio-temporal propagation delay using the gated recurrent unit, and output a spatio-temporal feature matrix; the cross-modal attention fusion layer is used to calculate the cross-modal correlation weights of vibration, light sensor, temperature and humidity, and acceleration through the multi-head cross-attention mechanism, and optimize the fusion weights through the cosine similarity and KL divergence, and output a cross-modal fusion tensor; the spatio-temporal residual encoding layer is used to extract multi-scale time series features through dilated temporal convolution according to the cross-modal fusion tensor, and capture the spatial dependence relationship through the graph residual network, and suppress the noise using the gated linear unit, and output a residual encoding feature; the anomaly correlation scoring layer is used to calculate the node anomaly propagation probability through the graph random walk algorithm, and detect the cross-modal joint outliers through the isolation forest, and output an anomaly correlation score; the federated incremental learning layer is used to aggregate the feature distribution and fine-tune the model parameters through the differential privacy federated average algorithm according to the anomaly correlation score and the misjudgment samples at the edge, and output the weight matrix of the spatio-temporal graph model after incremental update.

[0071] The distributed parameter aggregation framework performs homomorphic encryption on the edge model gradient parameters through the secure aggregation protocol to generate a set of ciphertext gradients; and based on the dynamic weight allocation mechanism, calculates the weight coefficients according to the historical misjudgment rate, metal density, and temperature and humidity of each node, and weighted aggregates the set of ciphertext gradients; when the Mahalanobis distance of the node parameters is detected to be higher than or equal to the preset threshold, the distributed parameter aggregation framework eliminates the outlier parameters through the anomaly gradient filtering algorithm; after aggregation, the gradient sparsification compression method is used to retain the top 10% of the maximum amplitude gradients, and the rest are set to zero and the transmission volume is compressed through Huffman coding; finally, Gaussian noise is injected into the aggregated gradients through the differential privacy mechanism to generate privacy-protected global model parameters and send them to the edge.

[0072] In step 1 of the above embodiments, the multi-physical field perception module captures high-frequency vibration signals of the mechanical deformation of the seal through vibration sensors (piezoelectric or MEMS type), the optoelectronic switch array monitors the cut-off state of the seal based on the design of reflective or transmissive optical paths, the temperature and humidity sensor (such as the SHT series) collects environmental parameters, and the MEMS triaxial accelerometer (such as ADXL345) quantifies the motion state of the vehicle; each sensor achieves hardware-level clock synchronization through the IEEE 1588PTP protocol, and uses the master-slave clock correction mechanism to eliminate the microsecond-level time reference deviation caused by crystal oscillator drift, ensuring the spatio-temporal consistency of multi-source data. The spatio-temporal correlation sampling algorithm is designed differently for the characteristics of heterogeneous sensors: for the optical cut-off signal, pulse width modulation coding (PWM) is used to convert the light intensity mutation event into a binary time sequence, and the cut-off event is spatially grid-coded through the Euler angle transformation matrix (pitch angle, roll angle, yaw angle based on the output of the accelerometer), realizing the spatio-temporal coupling representation of the opto-mechanical event and the vehicle attitude; for the vibration signal, the observation noise statistical characteristics of multiple sensors (such as the white noise variance of the vibration sensor and the quantization error of the accelerometer) are fused through adaptive weighted Kalman filtering (AWKF), and the weight coefficients of state prediction and measurement update are dynamically adjusted to generate a composite state matrix containing the vibration spectrum (0-2kHz), the light intensity mutation time sequence, the temperature and humidity gradient, and the acceleration vector. The redundant coding mechanism adopts the nested design of Reed-Solomon error correction code (RS code) and CRC-32 cyclic redundancy check: first, the composite data stream is divided into data blocks according to a fixed length (such as 512 bytes), and the CRC-32 check code is embedded to detect bit errors; then, the block division strategy is dynamically switched according to the environmental temperature and humidity threshold. In high-temperature and high-humidity environments, small-grain blocks of 64 bytes are used for high-frequency vibration data (to improve the error correction sensitivity), and large-grain blocks of 1024 bytes are used for low-frequency environmental parameters (to reduce the redundancy overhead); finally, RS code redundant bits (such as 4 redundant symbols can correct 2-byte errors) are added to each data block to construct a transmission unit with self-repair ability, ensuring data integrity in strong interference scenarios such as metal containers.

[0073] During implementation, the sensor module works in cooperation with the main control unit through a precision electrical interface: the vibration sensor (ADI ADXL357) is connected to the TI MSP432 microcontroller through the SPI interface at a clock rate of 8MHz, and the power supply uses 3.3V LDO voltage stabilization and is configured with an RC filter circuit to suppress high-frequency noise; the output end of the optoelectronic switch array (Vishay TCRT5000) is connected to the GPIO interrupt pin after being shaped by the LM393 voltage comparator, and a threshold voltage of 1.2V is set to achieve double-edge detection of light intensity mutation; the temperature and humidity sensor (Sensirion SHT45) passes through I 2The I²C bus communicates at a rate of 400 kHz, and 4.7 kΩ pull-up resistors on the SCL / SDA lines ensure signal integrity; the MEMS accelerometer (ST LIS3DH) transmits data via the I²S interface, is configured with a range of ±16g and a sampling rate of 1.6 kHz, and a bead filter is connected in series in the power path to eliminate switching noise. Clock synchronization is achieved by the Marvell 88E1111 Gigabit PHY chip implementing the IEEE 1588 PTP protocol, connecting to the MCU via the RMII interface, with a 25 MHz crystal oscillator providing the reference clock, and the FPGA global clock network distributing signals to make the sampling interval error of multiple sensors < 50 ns. The power supply system uses a TPS54331 module to output a 3.3V main power supply, and an LC filter (10 μH + 100 μF) is added to the analog branch to suppress the ripple to 20 mVpp, and the AD8605 operational amplifier is configured in the vibration signal channel to construct a second-order Butterworth filter (cut-off frequency of 5 kHz). 2

[0074] During the actual working process, the spatio-temporal correlation sampling algorithm runs on an embedded platform (such as ARM Cortex-M7), fusing vibration and acceleration data through adaptive weighted Kalman filtering (Q = diag(1e-4, 1e-5, 1e-3), R = diag(0.01, 0.1)), setting the rising / falling edge detection threshold for the PWM encoding of the light sensor interruption as the light intensity change rate > 5% / ms, measuring the pulse width of the light sensor interruption signal through the Timer capture mode and encoding it into an 8-bit binary sequence, and mapping the acceleration trajectory to a 1 cm 3 spatial grid; redundant encoding is implemented by the MCU calling an open-source RS library (such as libfec), adopting a joint mechanism of RS(255, 223) code and CRC-32 check, with the data block length fixed at 512 bytes, the CRC-32 polynomial being 0x04C11DB7, and the RS code configured as (255, 223) capable of correcting 16-byte errors; the non-uniform block strategy is triggered by temperature and humidity thresholds (such as temperature > 60°C or humidity > 90%RH), the high-frequency vibration data block is adjusted to 64 bytes (RS code configured as (63, 55)), and the low-frequency parameter block is extended to 1024 bytes (RS code configured as (1023, 975)). The preprocessed data packet is transmitted to the communication module via the SPI interface, and the data packet header contains a 4-byte timestamp (GPS synchronization), 2-byte sensor ID, and 2-byte check identification; the data transmission is encapsulated into a LoRa PHY frame (SF = 7, BW = 125 kHz) containing a 12-byte header, variable payload, and 4-byte checksum.

[0075] ​When configuring the present invention, in order to verify the beneficial effects brought by the spatio-temporal correlation sampling algorithm to the electronic seal for logistics, a simulation comparison experiment was carried out. The experiment built a multi-sensor data acquisition simulation platform for electronic seals through MATLAB / Simulink. The hardware configuration was TI MSP432 microcontroller, ADXL357 vibration sensor, TCRT5000 optoelectronic switch and LIS3DH accelerometer. The experimental conditions were set as a high-metal interference environment (simulating a container scenario), injecting Gaussian white noise (SNR = -5dB) and random pulse interference (duty cycle 10%). In the experiment, Group A applied the spatio-temporal correlation sampling algorithm, including clock synchronization protocol, adaptive Kalman filtering and spatio-temporal grid coding; Group B adopted the existing time-sharing sampling method, collected sensor data through time-sharing polling, and used moving average filtering. Each group carried out 5 independent experiments, each experiment lasted for 10 minutes, and the following parameters were recorded: timestamp alignment error (ns), vibration spectrum signal-to-noise ratio (dB), optical sensor event missed detection rate (%), acceleration trajectory root mean square error (mm), data packet CRC check failure rate (%) and transmission delay (ms). The experimental data record form is shown in Table 1:

[0076] Table 1 Comparison experimental record form of the application of spatio-temporal correlation sampling algorithm

[0077]

[0078] The experiment shows that the spatio-temporal correlation sampling algorithm of Group A is significantly better than the existing method of Group B in terms of time synchronization accuracy (error 12.5ns vs. 90.1ns), anti-interference ability (vibration SNR 18.7dB vs. 6.5dB), event capture reliability (optical sensor missed detection rate 0.8% vs. 13.3%) and data integrity (CRC failure rate 0.2% vs. 8.5%). Through the fusion of spatio-temporal grid coding and Kalman filtering, it effectively suppresses the multipath interference and noise coupling in the metal environment, reduces the acceleration trajectory error to the millimeter level (0.55mm vs. 3.95mm), and compresses the transmission delay to 12.7ms (45.6ms for Group B). While ensuring real-time performance, the CPU utilization rate of this algorithm is reduced by about 10%, verifying its technical advantages in complex logistics scenarios.

[0079] In step 2 of the above embodiment, the wavelet packet decomposition uses the Daubechies wavelet basis function to decompose the vibration signal into 5 layers, generating 32 frequency band nodes. The principle is that the wavelet packet decomposition realizes the balanced processing of high frequency and low frequency by recursively dividing the frequency band, overcoming the defect of insufficient high-frequency resolution of traditional wavelet decomposition. The node reordering algorithm is based on the frequency distribution characteristics of the wavelet packet tree, and through the mapping relationship between the frequency band node number and the Nyquist frequency (such as the frequency interval of the 5th layer nodes is the sampling frequency / 2 5), combined with the wpfrqord function in Matlab, it is sorted in ascending order of frequency to ensure that the energy calculation of the sub-band signals conforms to the physical frequency band distribution; the generalized relative wavelet energy ratio is calculated by calculating the energy ratio of the high-frequency band (nodes 16 - 31) to the low-frequency band (nodes 0 - 15), and the L 2 norm square (‖·‖2 2 ) characterizes the non-stationary energy accumulation characteristics of the sub-band signals. The mathematical essence of this method is to decouple the transient shock (high frequency) and steady-state ambient noise (low frequency) in mechanical vibrations through the time-frequency localization ability of the orthogonal wavelet basis, where the anti-zero-division perturbation term ε (such as 10 -6 ) is used to avoid the numerical instability problem of the denominator being zero while retaining the characteristics of weak signals. Pulse-width modulation decoding is based on the PWM signal output by the optoelectronic switch. The rising edge / falling edge is used to trigger an interrupt to record the timestamp, and the clock counter is combined to quantify the pulse width. Its principle lies in that the change in the PWM pulse width caused by the interruption event is linearly related to the physical seal opening and closing state, and random electromagnetic interference is suppressed through threshold filtering (such as a pulse width > 1 ms being determined as a valid event). The sliding window difference method quantifies the dynamic perturbation of environmental parameters by calculating the change rate of the average temperature and humidity within adjacent time windows; the change in the Euler angles of the three-axis acceleration vector is used to calculate the attitude angle offset through the quaternion transformation matrix, and its physical meaning lies in correlating the mechanical vibrations during transportation with the sudden changes in spatial attitude events. Z-score eliminates the dimensional difference through the mean-standard deviation normalization of the feature dimension, and its mathematical expression is z = (x - μ) / σ z = (x - μ) / σ, ensuring the comparability of multi-dimensional features in the orthogonal space; PCA realizes feature dimensionality reduction based on the eigenvalue decomposition of the covariance matrix, suppresses redundant noise by retaining the direction of the largest variance (such as the contribution rate of the first 3 principal components > 85%), and at the same time, the weighted enhancement strategy (such as assigning a 1.5-fold weight to the acceleration feature when the high-frequency energy ratio exceeds the threshold) can dynamically increase the representation weight of mechanical shock events.

[0080] During implementation, the vibration signal is collected by the ADI ADXL1002 broadband accelerometer (range ±50g, bandwidth 5 kHz), digitized by the ADS1278 at a sampling rate of 51.2 kHz, and then input into the STM32H743 microcontroller; the optical interruption signal is detected by the Vishay VCNL4200 optoelectronic sensor, and its I 2The C interface is configured with an interrupt response cycle of 1 ms; the temperature and humidity sensor uses Sensirion SHT35 (accuracy ±0.5% RH, sampling rate 4 Hz). The software algorithm is implemented on the DSP. The wavelet packet decomposition calls an open-source library (such as Wavelet Toolbox for C), sets the D wavelet basis function as db4, the decomposition level is 5 layers, and the node reordering uses Gray code index mapping; in the GRWER calculation, the norm square integral window is 10 ms, and the anti-zero division perturbation term is set to 1e-6. The light sensing interruption decoding sets the debounce filtering window to 5 ms, and the pulse width detection accuracy is 0.1 ms. The temperature and humidity gradient sliding window length is 10 sampling points (1-second window), and the Euler angle differential calculation uses the fourth-order Runge-Kutta method for iteration. In feature standardization, the mean and standard deviation of Z-score are dynamically updated based on historical data. PCA dimensionality reduction retains the first 3 principal components (cumulative variance > 95%), and the high-frequency energy ratio threshold is set to 3.0 (when exceeding the threshold, the acceleration feature weight is increased by 50%). The final anti-interference feature matrix is transmitted to the edge computing unit (such as NVIDIA Jetson Nano) through the CAN bus for use in the subsequent hybrid inference model.

[0081] In step 3 of the above embodiment, the channel state evaluation layer realizes the comprehensive quantization of the channel quality by fusing the multi-dimensional signal characteristics of LoRa, NB-IoT, and Bluetooth Mesh. The LoRa signal strength reflects the link attenuation characteristics of long-distance communication. Its spread spectrum modulation technology distinguishes channels through the orthogonal spreading factor (SF) to reduce co-channel interference. The NB-IoT signal-to-noise ratio (SNR) is based on narrowband modulation technology, and optimizes the anti-interference ability through the dynamic allocation of frequency-domain resource blocks. Its link budget model calculates the theoretical capacity boundary in combination with the Shannon formula. The Bluetooth Mesh node density realizes the network connectivity evaluation through topological graph analysis. Utilizing the multi-hop relay characteristics in the flooding communication mode, combined with the graph attention network (GAT) to quantify the correlation of signal strength, hop count, and historical packet loss rate between nodes, a path weight matrix is constructed. The channel attenuation factor uses statistical analysis methods, combines the time series model (ARIMA) to predict the channel state change trend, and dynamically updates the stability parameters through the sliding window difference method to avoid misjudgment caused by sudden environmental changes.

[0082] The interference perception layer uses frequency-domain decomposition and signal reconstruction technologies to enhance anti-interference capabilities. Spectrum analysis based on the Fast Fourier Transform (FFT) identifies the interference source frequency band. Combining with the wavelet threshold filtering algorithm, it separates signal components, suppresses narrowband noise and impulse interference. The interference intensity calculates the power spectral density (PSD) within the interference frequency band through the energy integration method, and the influence range estimates the spatial correlation of the interference signal through the covariance matrix. The adaptive notch filter dynamically adjusts the stopband bandwidth to achieve targeted suppression for common interference sources such as Wi-Fi and microwave ovens, ensuring the integrity of effective signal components. The link quality prediction layer is based on a hybrid prediction framework of the autoregressive model (AR) and Kalman filtering to achieve dynamic correction of link reliability. Historical link data extracts features such as the RSSI mean and SNR variance through a sliding time window. The AR model predicts the future signal trend, and the real-time measurement value is error-corrected through the Kalman gain coefficient to form a closed-loop feedback mechanism. The link reliability evaluation index combines the weighted entropy values of the packet loss rate (PLR) and delay jitter (Jitter) to quantify the link health and provide a basis for scheduling decisions. The core of the multi-objective fragmentation scheduling layer lies in multi-path concurrent transmission and redundant fault-tolerant design. The Reed-Solomon (RS) erasure code uses linear block coding in the Galois field (GF(2 8 )) to divide the data block into k original sub-blocks and m redundant sub-blocks, generating a loss-resistant fragmentation set. The fragmentation strategy is based on the data block size and channel availability. It uses a greedy algorithm to select the optimal fragmentation granularity and combines a double-buffer mechanism to achieve parallel transmission of Bluetooth Mesh and LoRa / NB-IoT. The dynamic path switching mechanism triggers sub-optimal path migration based on the RSSI threshold and verifies the integrity of the sub-blocks through a hash check tree. If the number of lost sub-blocks exceeds the redundancy threshold (m ≥ k / 3), the RS decoder is called to reconstruct the original data. The timing synchronization protocol uses an improved version of NTP (Network Time Protocol) to eliminate clock drift in multi-path transmission through end-to-end timestamp compensation.

[0083] During implementation, the main control unit uses the NXP i.MX RT1060 cross-border processor, which integrates a dual-core Cortex-M7 (600 MHz) and a hardware encryption engine; the LoRa communication module selects the Semtech SX1276, configured with a center frequency of 868 MHz, a bandwidth of 125 kHz, and a spreading factor SF = 7; the NB-IoT module is the Quectel BC66, supporting Band 8 (900 MHz) and the 3GPP Release 14 protocol; the Bluetooth Mesh network uses the Nordic nRF52840 chipset, supporting BLE 5.0 and dynamic relay routing. The channel state evaluation layer runs the GLMM model in the DSP core embedded in the processor. The parameter β is pre-trained to β = [0.32, -0.15, 0.28] through an offline training set (scenario data such as metal containers, open fields, mountains, etc.). The order of the ARIMA model (p, d, q) = (2, 1, 1), and the prediction window Δt = 5 seconds. The interference perception layer uses the FFT accelerator of the processor to perform 4096-point spectrum analysis, with a sampling rate of 51.2 ksps, 100 iterations of the ICA algorithm, and the stopband attenuation of the notch filter > 40 dB. The parameters of the ARMA model in the link quality prediction layer (p, q) = (3, 2), the covariance Q of the Kalman process noise = 1e-4, and the covariance R of the observation noise = 0.01. In the multi-objective fragmentation scheduling layer, the RS coding realizes the Galois field GF(2^8) operation through a hardware accelerator, and the fragmentation size is 256 bytes / block; the input feature dimension of the GAT network is 32, the number of attention heads is 4, the number of nodes in the hidden layer is 64, and the Adam optimizer (lr = 0.001) is used during training; the double buffer queue is configured as a 2×8KB circular buffer, and the DMA channel priority is LoRa > NB-IoT > Bluetooth Mesh; the hash check tree is constructed using the hardware SHA-256 engine, and the parallelism of the leaf node hash calculation is 4. The update period of the Bluetooth Mesh routing table is 10 seconds, and the ETX metric is calculated by statistically counting the periodic Hello packets. The path switching threshold RSSI < -90 dBm. The timing synchronization is marked by the time stamp of the PHY layer, the calibration period of the master-slave clock deviation is 1 second, and the clock source uses a TCXO (accuracy of ±0.5 ppm). The software protocol stack is based on the FreeRTOS real-time kernel, and the multi-threaded scheduling strategy is fixed-priority preemptive. The communication task priorities are: emergency events > data fragmentation > signaling control. In the actual measurement, the RS(16, 12) coding can recover 100% of the data at a 25% packet loss rate. The double buffer mechanism increases the concurrent transmission throughput to 2.8 times that of a single path, and the standard deviation of the end-to-end delay is controlled within ±1.2 ms.

[0084] In the design of this step, to verify the effectiveness of the hybrid inference model, a test platform was built based on the NXP i.MX RT1060 processor, Semtech SX1276 LoRa module, Quectel BC66 NB-IoT module, and Nordic nRF52840 Bluetooth Mesh node to simulate a high-interference logistics scenario (a metal container environment with urban multipath effects). Gaussian white noise (SNR = -10 dB) and random pulse interference (duty cycle 15%) were injected into the experiment. Group A applied the dynamic probability network model, including channel state assessment, interference awareness, and multipath sharding concurrent transmission strategy; Group B adopted the existing static multipath transmission protocol, polling transmission based on the LoRa / NB-IoT dual-link, using RS(12,8) erasure code and without dynamic path switching. Each group conducted 5 independent experiments, each lasting 10 minutes, and the data recording form is shown in Table 2:

[0085] Table 2 Comparison Experiment Record Table of Hybrid Inference Model Application

[0086]

[0087] Experiments show that this model has achieved high-reliability and low-latency concurrent transmission in complex interference environments, systematically solving the problems of data loss and excessive energy consumption caused by rigid paths and inefficient redundancy mechanisms in traditional methods.

[0088] In step 4 of the above embodiment, the feature preprocessing layer performs time-domain smoothing processing on the input anti-interference feature matrix through sliding window mean filtering (window length 10 sampling points) to suppress high-frequency jitter introduced by sensor noise or instantaneous interference (such as electromagnetic pulses); subsequently, the Z-score normalization algorithm is used to normalize heterogeneous features such as vibration spectrum energy ratio, temperature and humidity gradient, and acceleration Euler angle to the zero-mean unit variance interval based on the mean and standard deviation statistically calculated from historical data, eliminating the interference of dimension differences on model inference and generating standardized feature vectors.

[0089] The physical rule embedding layer defines the time-domain integral value E of vibration energy v and the optical sense interruption pulse width T dNonlinear dynamic correlation function: Ev > α·e-β·Td and Td > γ·ρ·η(T,H), where α is the benchmark energy threshold adapted to the type of transportation vehicle (truck / ship), β is the cut-off attenuation coefficient (representing the negative exponential relationship between energy and time), γ is the metal density compensation factor, ρ is the real-time metal density (inverted through the reflection characteristics of millimeter-wave radar), and η(T,H) is the temperature-humidity coupling function (compensating for the drift of light sensor sensitivity). This rule encodes the physical mechanism into a binary trigger identifier (set to 1 if the condition is met, otherwise set to 0), which is concatenated with the standardized feature vector to enhance the logical interpretability of the model for illegal opening events.

[0090] The multi-modal feature fusion layer adopts the multi-head attention mechanism (4 attention heads) to calculate the cross-modal correlation weights of the vibration spectrum (frequency-domain energy distribution), temperature-humidity gradient (time-domain differential slope), and acceleration vector (spatial Euler angle change) respectively. By normalizing the attention scores through the Softmax function, the weight ratio of the vibration spectrum is 60%, and the temperature-humidity and acceleration each account for 20%, strengthening the dominant features of impact events; after weighted fusion, a high-dimensional feature tensor is generated to suppress the interference of redundant noise on model inference.

[0091] The lightweight inference layer compresses the pre-trained ResNet-50 model into a MobileNet architecture through knowledge distillation technology, retaining the sensitivity to high-frequency vibration features. Specifically, depthwise separable convolutions are used to replace standard convolutions, reducing the number of parameters to 1 / 5 of the original model; during the training phase, the KL divergence loss function is used to align the feature distributions of the teacher model (ResNet-50) and the student model (MobileNet), ensuring that the edge device can still extract spatio-temporal joint features (such as the frequency-domain aggregation of vibration energy and the trajectory continuity of the acceleration vector) under limited computing power, and output the original confidence score.

[0092] The confidence calibration layer introduces the Bayesian probability calibration algorithm, combines the long-tail distribution of historical abnormal samples (abnormal proportion < 5%) and the current environmental parameters (metal density ρ, temperature-humidity η(T,H)), and dynamically corrects the original confidence. Specifically, a prior probability P(abnormal|ρ,η) is constructed. If the metal density ρ > 500 kg / m3, the confidence weight of the low-frequency vibration feature is reduced; through the posterior probability update formula Pcalibration = Poriginal·P(abnormal|ρ,η), the physical environment adaptive optimization of the confidence is realized.

[0093] The event trigger layer designs a dual-mode compression strategy: If the calibrated confidence level > 90% or a physical rule is triggered (binary identifier is 1), the original data window is compressed and encoded through the LZ77 stream compression algorithm (sliding window dictionary size 32KB, maximum match length 258 bytes) to generate an event-driven compressed packet; otherwise, only the SHA-256 hash value (32 bytes) of the data block is extracted and uploaded. This strategy reduces the network load by more than 70% while ensuring the integrity of critical event data.

[0094] During implementation, in terms of hardware deployment, the edge computing unit uses NVIDIA Jetson Nano (4GB memory, 128-core CUDA), and sensor data is transmitted to the processing unit in real time through the CAN bus (baud rate 500kbps). The vibration signal is collected by an ADI ADXL357 accelerometer (sampling rate 4kHz, range ±40g), the optical sensing interruption module uses an Omron EE-SX671 optoelectronic switch (response time 0.5ms), and the temperature and humidity sensor uses a Sensirion SHT45 (accuracy ±1.0%RH, sampling interval 1 second).

[0095] In terms of software implementation, the moving window mean filtering in the feature preprocessing layer is implemented in embedded C language, with a window length of 10 points, and the mean and standard deviation of Z-score normalization are updated every 10 minutes; the parameter settings of the dynamic correlation function in the physical rule embedding layer are: for the truck scenario, αtruck = 0.8J, for the ship scenario, αship = 1.2J, β = 0.05, γ = 0.2s / (kg / m3), and the temperature and humidity compensation function η(T,H) = 1 + 0.01(T - 25) + 0.005(H - 50). The multi-head attention mechanism in the multi-modal feature fusion layer is implemented based on TensorFlow Lite, with an input feature dimension of 64, 4 attention heads, and an output feature dimension of 256.

[0096] For the knowledge distillation model training in the lightweight inference layer, the PyTorch framework is used. The teacher model (ResNet-50) is trained on the server side, and the student model (MobileNet) is optimized through the distillation loss function (λ = 0.7); after the model weights are solidified in the inference stage, they are deployed to Jetson Nano, and the single-inference latency < 50ms. The Bayesian calibration module in the confidence calibration layer integrates the metal density sensor data (read through the SPI interface), and the prior probability table is pre-trained and generated based on historical data. The LZ77 compression algorithm in the event trigger layer is implemented using the zlib library, and the compression level is set to 6 (balancing speed and compression ratio), and the SHA-256 function of the OpenSSL library is called for hash value calculation.

[0097] Similarly, a test platform was built based on the STM32H743 microcontroller, Kistler 8704B accelerometer, Omron EE-SX671 optoelectronic switch, and Amphenol T9602 temperature and humidity sensor to simulate a high-interference metal container environment, and Gaussian white noise (SNR = -8 dB) and random electromagnetic pulse interference (frequency 1 - 100 MHz, field strength 10 V / m) were injected. Group A applied the hybrid inference model in step 4, including physical rule embedding, multimodal fusion, and Bayesian calibration, and used LZ77 stream compression and event-driven transmission; Group B used the existing fixed-threshold detection method, triggered upload only based on a single vibration energy threshold, and used the gzip compression algorithm. Each group conducted 5 independent experiments, each lasting 15 minutes, and the experimental data are recorded in Table 3 as follows:

[0098] Table 3 Comparison Experiment Record Table

[0099]

[0100] Experiments show that the proposed method achieves accurate and efficient seal status monitoring and data transmission in high-interference scenarios, systematically solving the performance bottleneck caused by the rigid thresholds and redundant data processing of traditional static rule engines. Therefore, the hybrid inference model can more accurately identify abnormal events and effectively reduce the amount of data transmitted through an intelligent compression strategy, thereby reducing power consumption and storage pressure, making the system more suitable for low-power and low-bandwidth logistics tracking application environments.

[0101] In step 5 of the above embodiment, the core principle of IEEE 1588 PTP clock synchronization is to spread high-precision time information in the network through the Precision Time Protocol (PTP), so that multiple sensor nodes are aligned on the same time basis. Specifically, the IEEE 1588v2 two-step synchronization mechanism is adopted, where the Sync message carries the timestamp of the GPS second pulse signal, and the gateway in transparent clock mode calculates and corrects the path dwell time when forwarding the message. Finally, an accurate global timestamp is calculated on the slave device through the correction field and the reception time, thus ensuring the time consistency of multi-sensor data.

[0102] In terms of network load monitoring, the LoRaWAN and NB-IoT dual-mode communication module uses the link layer probe mechanism to collect the MAC frame queue length and channel occupancy rate in real time. The probe module analyzes the channel load trend in different time windows. When the load rate exceeds the set threshold, it triggers the differential coding algorithm for data stream optimization. The basic idea of differential coding is to only store the incremental data that exceeds the quantization threshold between adjacent time windows to reduce the transmission of redundant data. This method combines Huffman coding to efficiently compress data blocks with small changes, thereby reducing the consumption of uplink communication bandwidth.

[0103] In addition, the communication sleep strategy of the electronic seal dynamically adjusts the sleep cycle based on the anomaly detection confidence feedback from the cloud. When the confidence is lower than 0.7, the exponential backoff algorithm is adopted to shorten the sleep interval to ensure higher-frequency data updates and improve the real-time performance of the system. When the confidence is high and no anomaly is detected, the system extends the sleep time to reduce energy consumption. The adjustment formula of the exponential backoff depends on the continuous anomaly count, and the maximum multiple is set to limit the variation range of the sleep interval, so as to maximize the energy-saving effect while ensuring the response speed.

[0104] During implementation, the IEEE 1588v2 Precision Time Protocol (PTP) based on the GPS second pulse (PPS) signal achieves the unification of the time reference for multi-sensor data through a two-step synchronization mechanism: the master clock (GPS module) periodically broadcasts Sync messages and records the hardware timestamp of the message sending moment t1; after receiving the Sync message, the slave port (electronic seal) records the arrival moment t2 and transmits the t1 value through the Follow_Up message. Combining the calculation of the transmission path residence time Δt_residence of the transparent clock (border gateway) (Δt_residence = t2 - t1 - t_delay_base, where t_delay_base is the preset reference path delay), the Δt_residence value is superimposed in the correction field. Finally, the slave port calculates the clock offset Δt_offset = [(t2 - t1) - (t4 - t3)] / 2 through the interaction of Delay_Req / Delay_Resp messages to achieve sub-microsecond timestamp alignment. Network load monitoring real-time statistics the channel occupancy rate (CBR = current number of frames / maximum frame capacity) and queue length (QL) through the MAC layer probes of LoRaWAN and NB-IoT. If CBR > 70% or QL > 15 frames, the differential coding algorithm is triggered: the sensor data stream is segmented by a time window (T_window = 500ms), and the difference ΔDCT of the discrete cosine transform (DCT) coefficients of adjacent window data is calculated. Only the incremental data blocks with ΔDCT ≥ quantization threshold Q_th (Q_th = 0.1·max(DCT)) are retained, and the Huffman coding table index (predefined codebook ID) is embedded. The compressed data packet further reduces redundancy through entropy coding. The dynamic sleep cycle adjustment is based on the confidence score (from step 4) and the continuous anomaly count N_abnormal: when the confidence score < 0.7, the sleep cycle T_sleep = T_base·exp(-0.5·N_abnormal), where T_base = 10s; if the confidence score ≥ 0.7 and N_abnormal = 0, then T_sleep = 1.5·T_base, balancing the real-time performance and energy consumption through the exponential backoff algorithm.

[0105] In implementation, the GPS module uses ublox ZED-F9P (PPS output accuracy ±30 ns), connects to the STM32H743 main control chip through the RS-422 interface, and the PTP protocol stack is implemented based on the LWIP library; the LoRaWAN module is Semtech SX1302, the NB-IoT module is Quectel BC95-G, and the MAC layer probe reads the PHY register status (CBR and QL) through the SPI interface; the power management module uses the TPS62840 buck converter to dynamically adjust the core voltage (0.9V - 1.2V) in the sleep mode. In software implementation, the PTP clock synchronization configures the Sync message period to 1 s, the transparent clock dwell time measurement accuracy is ±100 ns, and the clock offset compensation is implemented through the hardware timer TIM2 (32-bit); in the differential coding algorithm, the DCT coefficient calculation uses the FFT acceleration library (CMSIS-DSP), the quantization threshold Q(th) = 0.1·max(DCT), the Huffman codebook pre-stores 4 dynamic tables (compression ratio 45% - 65%), and the index field occupies 2 bits; the sleep cycle control is triggered by the RTC wake-up interrupt, the exponential backoff coefficient α = 0.5, and the confidence threshold S_c = 0.7 is determined by training with historical data. In actual measurement, the GPS-PTP synchronization error ≤200 ns, the differential coding reduces the data volume by 58%, the dynamic sleep mechanism shortens the response time to 5 s (benchmark value 10 s) in the low-confidence scenario, and the energy consumption is reduced by 33% in the high-confidence scenario.

[0106] In implementation, compared with the traditional method of uploading data at fixed time intervals, this method achieves more accurate time synchronization, dynamic data optimization, and intelligent sleep management. Through the IEEE 1588 PTP synchronization mechanism, the time deviation of multi-sensor data is eliminated, and the reliability of data fusion is improved; the differential coding technology effectively reduces the communication bandwidth occupancy, reduces the data upload volume, and alleviates the network load pressure of LoRa and NB-IoT; the dynamic sleep strategy adjusts the communication cycle based on the abnormal confidence, optimizes the power management, enables the system to significantly reduce the energy consumption and equipment operation and maintenance costs while ensuring the detection accuracy, and improves the overall efficiency and stability of the logistics tracking system.

[0107] In step 6 of the above embodiment, the core objective of the spatio-temporal embedding alignment layer is to uniformly align the time axes of different modality data (vibration, light sense, temperature and humidity, acceleration) to eliminate the errors caused by the asynchronous sampling time of sensors. This process relies on the dynamic time warping algorithm, whose basic principle is to calculate the optimal matching path between different time series through dynamic programming to align them in the time dimension. Specifically, DTW calculates the optimal matching path W=(w1, w2,..., wk) between two time series X=(x1, x2,..., xm) and Y=(y1, y2,..., yn) based on the recursive formula to minimize the cumulative distance:

[0108] D(i,j) = d(xi,yj)+min(D(i - 1,j),D(i,j - 1),D(i - 1,j - 1))

[0109] where d(xi,yj) is the distance between time points i and j, and D(i,j) is the cumulative distance of the first i and j points.

[0110] In addition, to further optimize the data alignment effect, a self-attention mechanism is used to map multi-modal data to a unified spatio-temporal coordinate system. The self-attention mechanism calculates the correlation weights between data based on Scaled Dot-Product Attention, and the core calculation formula is as follows:

[0111] Attention(Q,K,V) = softmax(QKTdk)V

[0112] where Q, K, and V represent the query, key value, and value matrices respectively, and dk is the dimension of the key value. This mechanism can adaptively adjust the distribution of multi-modal data on the time axis to align it to the same time reference. After spatio-temporal alignment, the multi-modal data is constructed into a graph structure and input into a dynamic graph convolutional network for feature extraction. Its core principle is to calculate the relationship between nodes (data points) using the adjacency matrix A and adaptively adjust the weights of different modality features through a graph attention network. The basic propagation rule of GCN is:

[0113] H(l + 1) = σ(D~-1 / 2A~D~-1 / 2H(l)W(l))

[0114] where A~ = A + I is the adjacency matrix with self-connections, D~ is the degree matrix, H(l) is the feature matrix of the l-th layer, W(l) is the trainable weight, and σ is the activation function. To model the propagation delay of different physical sensing data on the time axis, a gated recurrent unit (GRU) is used to process the time series information. GRU controls the storage and forgetting of information through a reset gate and an update gate, and the update formula is:

[0115] rt = σ(Wrxt + Urht-1)

[0116] zt = σ(Wzxt + Uzht-1)

[0117] ht = (1 - zt)ht-1 + zttanh(Whxt + Uh(rt·ht-1))

[0118] This can effectively model the spatio-temporal dependence relationship between the vibration frequency-domain energy, light-sensing pulse density, temperature and humidity gradient, and acceleration covariance matrix, and improve the accuracy of anomaly detection. The cross-modal attention fusion layer optimizes the feature fusion effect by calculating the mutual correlation of multi-modal data. This layer uses the multi-head cross-attention mechanism to calculate the correlation of different modal features, and combines the cosine similarity and Kullback-Leibler (KL) divergence for feature weighting. Its core formula is as follows:

[0119] αij = QiKjTdk

[0120] KL(P||Q) = ∑iP(i)logP(i) / Q(i)

[0121] The cosine similarity is used to calculate the similarity of different modal features, while the KL divergence is used to measure the difference between different feature distributions, so as to optimize the fusion weight and improve the anomaly detection ability of the model. The anomaly correlation scoring layer uses graph random walk to calculate the propagation probability of abnormal data on the spatio-temporal graph, and combines the isolation forest to identify cross-modal joint outliers. The transition probability matrix of graph random walk is calculated as follows:

[0122] P = D-1A

[0123] Among them, P is the transition probability matrix, A is the adjacency matrix, and D is the degree matrix. This method can identify the propagation path of abnormal data between different sensor modalities and improve the detection accuracy. The federated incremental learning layer uses the differential privacy federated average algorithm for edge model update. Its core idea is to use the model parameters locally trained by multiple edge devices for encrypted aggregation to ensure that data privacy is not leaked. The basic update formula is as follows:

[0124] wt+1 = ∑i=1Nninwit + N(0,σ2)

[0125] Among them, wit is the model parameter of the i-th device, ni is the data volume, and N(0,σ2) is Gaussian noise. In addition, the distributed parameter aggregation framework uses the anomaly gradient filtering algorithm to eliminate abnormal nodes, and uses gradient sparsification compression to only retain the top 10% of the maximum amplitude gradients to reduce communication overhead. Finally, the aggregated gradient injects noise through the differential privacy mechanism to improve the privacy protection ability while ensuring the accuracy of model update.

[0126] During implementation, the cloud server uses an NVIDIA A100 GPU cluster (80GB of video memory), and the edge device is a Jetson AGX Xavier (32GB of memory), which communicates with the cloud via the MQTT protocol (QoS = 1). The dynamic time warping (DTW) algorithm of the spatiotemporal embedding alignment layer sets the window constraint to 100ms, and the feature dimension of the self-attention mechanism is 512; the DGAT network node features of the multimodal graph convolution layer include vibration energy (normalized to [0,1]), light pulse density (counts / second), temperature and humidity gradient (℃ / s) and acceleration covariance matrix (3x3), the graph edge weights are dynamically updated based on the Pearson correlation coefficient (threshold>0.6), and the number of GRU hidden layer units is 128; the cosine similarity temperature coefficient τ of the cross-modal attention fusion layer is 0.07, and the KL divergence regularization weight λ is 0.1; the dilated convolution kernel size of the spatiotemporal residual coding layer is 3, the number of GraphResNet residual blocks is 3, and the GLU gated activation function is Sigmoid; the graph random walk step size of the abnormal association scoring layer is 5, and the number of isolated forest trees is 100; the DP-FedAvg noise scale σ of the federated incremental learning layer is 0.1, the clipping threshold C is 1.0, and the number of fine-tuning iterations is 10.

[0127] The homomorphic encryption of the distributed parameter aggregation framework uses the Paillier library (key length 2048 bits), the historical misjudgment rate weight accounts for 60% of the dynamic weight allocation, and the environmental interference accounts for 40%; the Mahalanobis distance threshold is set to 3σ (based on the global gradient distribution), and the number of K-means clustering centers is 3; the Top-K selection of gradient sparsification is based on absolute value sorting, and the Huffman coding code table is updated every 24 hours; the Gaussian noise of the differential privacy mechanism is injected through CUDA acceleration, and the single aggregation takes <200ms. The edge model update is achieved through OTA wireless upgrade, the weight matrix differential transmission (only the change amount is sent), and the bandwidth occupancy is reduced by 65%.

[0128] During implementation, this step uses a spatiotemporal graph fusion analysis model to achieve deep correlation verification of multimodal data, effectively correcting misjudgment events caused by environmental interference at the edge, and improving the accuracy and robustness of anomaly detection; the federated incremental learning and differential privacy aggregation mechanism achieve dynamic optimization of model parameters under the premise of protecting data privacy, so that the system can adapt to the complex changes of different logistics scenarios; the distributed parameter aggregation framework uses gradient sparsification and homomorphic encryption to balance communication efficiency and security, significantly reducing the collaborative overhead between the cloud and the edge. Compared with traditional independent verification or centralized training solutions, this solution achieves a comprehensive improvement in detection accuracy and system generalization capabilities while ensuring privacy and security.

[0129] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Without departing from the principles and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps so as to perform substantially the same function in a substantially the same way to achieve substantially the same result falls within the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. An automated detection method for electronic seals applied to logistics tracking; characterized in that: The method includes the following steps: Step 1: Obtain the seal status data through a spatio-temporal correlation sampling algorithm, and perform block preprocessing using a redundant coding mechanism. The seal status includes vibration signals, light sensing interruption status, ambient temperature and humidity, and three-axis acceleration vector; Step 2: Perform wavelet packet decomposition on the vibration signals, extract the spectral energy ratio of the high-frequency subband and the low-frequency subband, and combine the pulse width of the light sensing interruption status, the temperature and humidity change gradient, and the three-axis acceleration vector to construct an anti-interference feature matrix; Step 3: Based on the anti-interference feature matrix, the LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density collected in real time, calculate the comprehensive channel quality index through a dynamic probability network model, and dynamically adjust the transmission strategy; Step 4: Input the anti-interference feature matrix into the hybrid inference model deployed at the edge side, and output the seal status confidence score; if the confidence is higher than the preset threshold or meets the vibration and light sensing joint trigger condition, compress and encode the current window data through a streaming compression algorithm, and upload the event-driven compressed package and hash value to the cloud; otherwise, upload the hash value; Step 5: Apply IEEE 1588PTP clock synchronization to the spatio-temporal correlation sampling data in Step 1 based on the GPS second pulse signal; and monitor the LoRa and NB-IoT network load rates in real time. When the load rate is higher than the preset threshold, transmit the incremental data of adjacent windows in the data output in Step 4 using differential coding, and dynamically adjust the sleep cycle of the electronic seal communication module in combination with the seal status confidence score; Step 6: Input the uploaded data in Step 4 into the spatio-temporal graph fusion analysis model, fuse the vibration spectrum, light sensing time series, temperature and humidity curve, and acceleration trajectory, generate a cross-modal anomaly correlation score, and correct the misjudgment events at the edge side; and based on the verified abnormal samples, update the edge model using a distributed parameter aggregation framework.

2. An automated detection method for electronic seals applied to logistics tracking according to claim 1, characterized in that: The working principle of Step 1 is: The multi-physical field perception module uses a vibration sensor, a photoelectric switch array, a temperature and humidity sensor, and a MEMS three-axis accelerometer for data collection, and uses a clock synchronization protocol to eliminate the time reference drift between multiple sensors, and outputs a multi-source data stream; Using a spatio-temporal correlation sampling algorithm, the observation errors of each sensor in the spatio-temporal dimension are fused through adaptive weighted Kalman filtering to generate a composite state matrix containing vibration spectra, light intensity mutation events, environmental parameters, and acceleration vectors; for the light sensing interruption signal, the spatio-temporal correlation sampling algorithm uses pulse width modulation coding to convert the light intensity threshold mutation into binary timing data, and at the same time combines the Euler angle transformation matrix of the three-axis acceleration vector for spatial grid coding to realize the spatio-temporal coupling characterization of opto-mechanical events and motion states; then, based on the redundancy coding mechanism of the Reed-Solomon error correction algorithm in step 1, the composite data stream is segmented into fixed-length data blocks, and CRC-32 cyclic redundancy check codes and forward error correction redundancy bits are embedded in each data block to construct a transmission unit; at the same time, in step 1, when the temperature and humidity exceed the preset safety threshold, a non-uniform block strategy is adopted, with small-grain blocks for high-frequency vibration signals and large-grain blocks for low-frequency environmental parameters; finally, a preprocessed data packet containing spatio-temporal tags, physical field feature vectors, and redundancy check bits is output.

3. An automated detection method for electronic seals applied to logistics tracking according to claim 1, characterized in that: The working method of step 2 is as follows: S1. Perform 5-layer wavelet packet decomposition on the vibration signal through D wavelet basis functions to generate 32 frequency band nodes, and reorder the nodes of the 5th layer in ascending order of frequency through the node reordering algorithm; S2. Based on the reconstructed sub-band signals, calculate the frequency band energy ratios of the high-frequency band nodes from node 16 to node 31 and the low-frequency band nodes from node 0 to node 15 using the generalized relative wavelet energy ratio, and the calculation formula is: In Equation (1), Hilbert represents the squared norm, which is used to capture the instantaneous energy of non-stationary signals. δ is a perturbation term to prevent division by zero, and d 5,i represents the multi-resolution characterization unit of the signal in time-frequency analysis; S3. Use a pulse width modulation decoding mechanism to analyze the light sensing interruption state and extract the timestamp and pulse width parameters of the interruption event; S4. Calculate the gradient value of the temperature and humidity sensor through the sliding window difference method, and construct a four-dimensional auxiliary feature vector in combination with the Euler angle change amount of the three-axis acceleration vector; S5. Align the frequency band energy ratio, pulse width parameter, temperature and humidity gradient, and Euler angle change amount of the three-axis acceleration vector according to the time window, eliminate the dimension difference through Z-score standardization, and perform orthogonal dimensionality reduction on the multi-dimensional features using principal component analysis to generate an anti-interference low-dimensional feature matrix; During the standardization process, if the high-frequency energy ratio exceeds the preset threshold, the three-axis acceleration vector is weighted and enhanced to suppress mechanical noise.

4. An automated detection method for electronic seals applied to logistics tracking according to claim 1, characterized in that: The dynamic probability network model includes a channel state evaluation layer, an interference perception layer, a link quality prediction layer, and a multi-target fragmentation scheduling layer; the channel state evaluation layer is used to calculate the channel attenuation factor using statistical analysis methods based on LoRa signal strength, NB-IoT signal-to-noise ratio, and Bluetooth Mesh node density, and predict the channel state change trend through time series modeling, and output the current channel stability parameters; the interference perception layer is used to identify interference sources through frequency domain analysis methods according to the signal spectrum distribution, and extract signal components through filtering algorithms, and output the interference intensity and interference range; the link quality prediction layer is used to calculate the future link signal change trend using an autoregressive model based on historical link data, and correct it in combination with real-time measurement values, and output the link reliability evaluation index; The multi-objective sharding scheduling layer is used to perform data sharding based on the data block size and channel availability, and adopt a multi-path sharding concurrent transmission strategy to allocate data blocks to different paths for transmission, and achieve concurrent transmission of Bluetooth Mesh ad hoc network and LoRa and NB-IoT links through a double-buffer mechanism.

5. An automated detection method for electronic seals applied to logistics tracking according to claim 4, characterized in that: The working method of the multi-path sharding concurrent transmission strategy is as follows: U1. Divide the data block into N original sub-blocks and M redundant sub-blocks through the RS erasure code sharding algorithm to generate a loss-resistant sharding set, where N:M = 4:1; U2. Based on the Bluetooth Mesh node topology structure, use a graph attention network to analyze the signal strength, hop count, and historical packet loss rate of each Bluetooth node, and select the optimal relay path set; U3. Allocate the sharded sub-blocks to the Bluetooth Mesh ad hoc network and LoRa and NB-IoT communication links for parallel transmission through a multi-threaded concurrent control protocol. If the RSSI of the Bluetooth Mesh path is lower than the preset threshold, migrate the sub-block to the sub-optimal path through a dynamic path switching mechanism; U4. The receiving end verifies the integrity of the sub-blocks through a reverse hash check tree. If the number of lost sub-blocks exceeds the preset redundancy threshold, reconstruct the original data block based on the erasure code; at the same time, align the multi-path transmission timing through an end-to-end timestamp synchronization protocol.

6. An automated detection method for an electronic seal applied to logistics tracking according to claim 1, characterized in that: The hybrid inference model includes a feature preprocessing layer, a physical rule embedding layer, a multi-modal feature fusion layer, a lightweight inference layer, a confidence calibration layer, and an event trigger layer; the feature preprocessing layer is used to smooth the input anti-interference feature matrix through sliding window mean filtering, and use the Z-score normalization algorithm to normalize each feature dimension to zero mean and unit variance to generate a normalized feature vector; the physical rule embedding layer is used to define the joint determination rule of vibration energy and blocking duration based on the lead seal mechanical structure and light sensing blocking logic, and encode the rule matching result as a binary identifier, which is concatenated with the normalized feature vector; the multi-modal feature fusion layer is used to model the cross-modal correlation of vibration spectrum, temperature and humidity gradient, and acceleration vector through a multi-head attention mechanism, calculate the attention weights of each feature channel, and output a high-dimensional feature tensor; the lightweight inference layer is used to input the fused feature tensor into a knowledge distillation model, perform convolution and depthwise separable convolution operations, extract spatio-temporal joint features, and output an original confidence score; the confidence calibration layer is used to adopt a Bayesian probability calibration algorithm, combine the historical abnormal sample distribution and the current environmental parameters, and dynamically correct the original confidence; the event trigger layer is used to compress and encode the original data window through the LZ77 streaming compression algorithm when the confidence is higher than the preset threshold or the rule is triggered, generate an event-driven compressed packet and extract the data block hash value and upload it to the cloud; when the confidence is lower than or equal to the preset threshold and the rule is not triggered, extract the data block hash value and upload it to the cloud.

7. An automated detection method for electronic seals applied to logistics tracking according to claim 1, characterized in that: The vibration and light-sensing combined trigger condition defines the time-domain integral value E of vibration energy based on the mechanical deformation mechanics model of the lead seal v and the light-sensing interruption pulse width T d of the dynamic correlation function. The calculation formula of E v is as follows: In formula (2), E v represents the integral of the square of the magnitude of the vibration acceleration vector a(t) collected by the triaxial acceleration sensor within the time window [t0, t1], and is used to quantify the impact energy intensity of an illegal opening event; when is satisfied and T d > ργθ, it is determined as an illegal opening event, where a is the reference energy threshold, β is the cut-off attenuation coefficient, ρ is the real-time metal density, γ is the metal density compensation factor, and θ is the environmental temperature and humidity compensation function. Among them, a, β, and γ are dynamically adjusted according to the type of transportation vehicle and the environmental temperature and humidity.

8. An automated detection method for electronic seals applied to logistics tracking according to claim 1, characterized in that: The working principle of step 5 is as follows: First, through the IEEE 1588v2 two-step synchronization mechanism, the Sync message is used to carry the hardware timestamp to record the transmission moment u1 of the GPS second pulse source. In the gateway side, the transparent clock mode is adopted to calculate and correct the residence time c1 of the transmission path in real time, and the residence time c1 is superimposed on the correction field of the follow-up message. Combining with the reception time u2 of the slave port, multi-sensor timestamp alignment is achieved based on formula (3): R = u2 - u1 - D - c1 (3) In formula (3), D is a preset path delay reference value; then, through the link layer probe of the LoRaWAN and NB-IoT dual-mode communication module, the MAC frame queue length and channel occupancy rate are statistically monitored in real time. If the load rate exceeds the preset threshold, the differential coding algorithm is triggered to split the sensor data stream according to a time window, and only the incremental data whose change amount exceeds the quantization threshold within adjacent windows is retained, and the Huffman compression coding table index is embedded; at the same time, based on the anomaly detection confidence feedback from the cloud, the sleep cycle of the communication module is dynamically adjusted. When the confidence is lower than 0.7, the exponential backoff algorithm is used to shorten the sleep interval to 1 / 2 of the reference value n , where n is the number of consecutive anomalies, otherwise it is extended to 2 n times the reference value.

9. An automated detection method for an electronic seal applied to logistics tracking according to claim 1, characterized in that: The spatio-temporal graph fusion analysis model includes a spatio-temporal embedding alignment layer, a multi-modal graph convolution layer, a cross-modal attention fusion layer, a spatio-temporal residual coding layer, an anomaly correlation scoring layer, and a federated incremental learning layer; the spatio-temporal embedding alignment layer is used to align the multi-modal time series according to the original timestamps of the vibration spectrum, light sensor time series, temperature and humidity curve, and acceleration trajectory through the dynamic time warping algorithm, and map them to a unified spatio-temporal coordinate system by combining the self-attention mechanism, and output a spatio-temporal alignment tensor; the multi-modal graph convolution layer is used to fuse the vibration frequency domain energy, light sensor pulse density, temperature and humidity gradient, and acceleration covariance matrix through a dynamic graph attention network, and use a gated recurrent unit to model the spatio-temporal propagation delay, and output a spatio-temporal feature matrix; The cross-modal attention fusion layer is used to calculate the cross-modal correlation weights of vibration, light sensor, temperature and humidity, and acceleration through a multi-head cross-attention mechanism, and optimize the fusion weights through cosine similarity and KL divergence, and output a cross-modal fusion tensor; the spatio-temporal residual coding layer is used to extract multi-scale time series features through dilated time convolution according to the cross-modal fusion tensor, and capture the spatial dependence relationship through a graph residual network, and use a gated linear unit to suppress noise, and output a residual coding feature; the anomaly correlation scoring layer is used to calculate the node anomaly propagation probability through a graph random walk algorithm, and detect cross-modal joint outliers through an isolation forest, and output an anomaly correlation score; the federated incremental learning layer is used to aggregate the feature distribution and fine-tune the model parameters through the differential privacy federated average algorithm according to the anomaly correlation score and the misjudgment samples at the edge, and output the spatio-temporal graph model weight matrix after incremental update.

10. An automated detection method for an electronic seal applied to logistics tracking according to claim 1, characterized in that: The distributed parameter aggregation framework homomorphically encrypts the edge model gradient parameters through a secure aggregation protocol to generate a ciphertext gradient set; and based on a dynamic weight allocation mechanism, calculates the weight coefficient according to the historical misjudgment rate, metal density, and temperature and humidity of each node, and weighted aggregates the ciphertext gradient set; when the Mahalanobis distance of the node parameters is detected to be higher than or equal to a preset threshold, the outlier parameters are removed through an abnormal gradient filtering algorithm; after aggregation, the gradient sparsification compression method is used to retain the top 10% of the maximum amplitude gradients, and the rest are set to zero and compressed through Huffman coding; finally, Gaussian noise is injected into the aggregated gradient through the differential privacy mechanism to generate privacy-protected global model parameters and send them to the edge.

Images