Carrying object safety detection management system and method based on Internet of Things

Abstract

The invention discloses a carried object safety detection management system and method based on the Internet of Things, and relates to the technical field of carried object safety detection.The method comprises the steps that a sensor combination is selected based on carried object type labels, and multi-mode sensing data of carried objects, vehicles and drivers are collected; applying a uniform timestamp and a space coordinate to the multi-modal data to generate a time-space associated data set; screening the time-space correlation data set to form a core feature set, analyzing a fusion weight in combination with the operation feature of the vehicle, and generating an optimization feature based on the fusion weight; training an SVM classifier based on the optimization features and the corresponding category labels, and training an LSTM model by using local data of each vehicle; analyzing the MMD distance between the source domain and the target domain; and multi-source evidence analysis is combined with the trust degree, and abnormal early warning is triggered based on dynamic safety evaluation. Multi-source features are adaptively fused, a threshold value is dynamically adjusted, risk fluctuations in different time periods are adapted, and the early warning sensitivity is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transport safety detection technology, and in particular to a transport safety detection management system and method based on the Internet of Things. Background Art

[0002] The logistics and transportation industry faces challenges, particularly in driver monitoring and transportation safety. Smart logistics offers advantages such as strong connectivity, broad integration, low economic costs, and high operational efficiency. Transportation is a crucial foundation for the development of a logistics system, and promoting the development of smart logistics has multiple application scenarios, including internet-connected road freight. Logistics companies need to maintain timely access to vehicle information and conduct real-time monitoring and dispatch.

[0003] In the existing real-time detection of cargo safety, traditional sensor networks use fixed combinations such as temperature, humidity and acceleration sensors, which cannot be dynamically adjusted according to cargo characteristics such as liquid sloshing and vibration of fragile items. This leads to missed detection or redundant collection of key parameters such as pressure distribution and vibration spectrum. Traditional analysis methods rely on a single signal or fixed threshold to judge risks, ignore the synergy of multiple evidence, and have a high false alarm rate.

[0004] Therefore, the present invention discloses a transport safety detection and management system and method based on the Internet of Things to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a transport safety detection and management system and method based on the Internet of Things to solve the problems raised in the prior art.

[0006] To achieve the above-mentioned object, the present invention provides the following technical solution: a method for safety detection and management of transported goods based on the Internet of Things, the method comprising the following steps: S1: Select a sensor combination based on the vehicle type label to collect multimodal sensor data from the vehicle, vehicle, and driver; apply unified timestamps and spatial coordinates to the multimodal data to generate a spatiotemporal correlation dataset; S2: Filter the spatiotemporal correlation dataset to form a core feature set, analyze the fusion weights based on the vehicle's operating characteristics, and generate optimized features based on the fusion weights; S3: Train an SVM classifier based on the optimized features and corresponding class labels, and train an LSTM model using the local data of each vehicle; analyze the MMD distance between the source domain and the target domain; S4: Integrate multi-source evidence to analyze joint trust and trigger abnormal warnings based on dynamic security assessment.

[0007] According to the above scheme, S1 includes the following contents: S101: Label L according to the type of cargo type Select the sensor combination; if L type=Liquid, enable pressure sensor array S p Collect pressure distribution ; where p n Represents the normalized measurement value of the nth pressure sensor, n∈[1,N], N represents the total number of pressure sensors, n is a positive integer; if L type =Fragile, activate vibration sensor D h Collect vibration spectrum D h (t); S102: Analyze the signal-to-noise ratio of the i-th sensor. If the signal-to-noise ratio of the i-th sensor is less than the signal-to-noise ratio threshold, switch to the backup sensor. After switching, compare the spectrum correlation between the backup sensor and the main sensor in the overlapping frequency band. Apply a unified timestamp t to the multimodal data. j and spatial coordinates (x, y, z) j ; Generate spatiotemporal correlation dataset X ST ={(t j , x j ,y j , z j , data j )};data j represents the set of sensor measurement values ​​corresponding to the jth spatiotemporal point.

[0008] This application intelligently selects pressure or vibration sensors based on the type of cargo (liquids / fragiles), avoiding redundant data collection and improving the monitoring accuracy of key parameters. It also calculates the signal-to-noise ratio in real time and switches to backup sensors, dynamically eliminating noise interference data and improving the reliability of subsequent feature extraction. It also unifies timestamps and spatial coordinates to eliminate spatiotemporal misalignment, constructing a spatiotemporal correlation dataset, and providing data support for multi-sensor collaborative analysis. According to the above scheme, S2 includes the following contents: S201: Spatiotemporal correlation dataset X ST Computing Random Forest Feature Importances for Physical Data in I RF d and mutual information I MI d ; Where d represents the feature dimension number; retain the RF d > α and I MI d The features of >β constitute the core feature set F core ={f d}, where α represents the random forest feature importance threshold, α=1 / d 1 / 2 , β represents the mutual information threshold, which is equal to the median of the mutual information in the data set; S202: Vehicle acceleration function a c(t) Convolution is performed with window length τ equal to τ1, τ2 and τ3, where τ1, τ2 and τ3 are preset constants, to obtain the local mutation amount M local =Conv1D(a c ,τ=S1), trend quantity M trend =Conv1D(a c ,τ=S2) and the event quantity M global =Conv1D(a c ,τ=S3); S203: Fusion feature X=[F core , M local , M trend , M global ] Projected to query matrix Q, key matrix K and value matrix V; Q=W q X,K=W K X,V=W V X; where W q 、W K and W V Represents the trainable parameter matrix; split the query matrix Q, key matrix K and value matrix V into A1, A2, ..., A according to the total number of attention heads M M Equal sub-vectors; calculate the matching score Score of the m-th head and all keys m =A m ⊺ K; where ⊺ represents the transpose of the vector; analyze the fusion weight w m =exp(A m ⊺ K) / exp(A r ⊺ K); where exp(∙) represents an exponential function with natural numbers as the base, r∈[1, M], r is a positive integer; based on the fusion weight, the optimized feature F is generated opt = w m ⋅[F core , M local , M trend , M global ].

[0009] This application combines random forest importance and mutual information to screen core features, taking into account both feature prediction ability and independence to suppress overfitting; extracts local mutations, trends, and global events of acceleration signals through different convolution windows to comprehensively characterize vehicle motion patterns; and adaptively fuses multi-source features to enhance the contribution of key features (such as sudden vibrations) and improve feature space discrimination.

[0010] According to the above solution, S3 contains the following: S301: Use optimization feature F opt And the corresponding category label b∈{-1,+1} trains the SVM classifier, where the radial basis kernel function is: ; Among them, x s represents the sth support vector, x represents the feature vector of the new sample to be classified, exp(∙) represents the exponential function with natural numbers as the base, γ represents the kernel function parameter, and D represents the optimized feature F opt The characteristic dimension of Var(F opt ) represents the sum of the variances of all features; Decision function: ; Among them, λ s represents the support vector x s The weight of b s represents the support vector x s The category label, c represents the bias term, sign (∙) represents the sign function, and the output +1 or -1 represents the classification result; When the new sample F new When it arrives, the support vector is filtered by KKT conditions, the original support vector and the newly added boundary samples are retained, and the reduced set S is constructed. ’ =S old ∪S new , in the reduced set S ’ Solve the dual problem and update λ s and c; S302: Use the optimized feature matrix stored in the onboard edge device of each vehicle to train the LSTM model and update the parameter set W local The parameter set includes the input gate, forget gate, output gate, and weight matrix of the candidate memory unit in the LSTM model. The locally trained parameters are encrypted and uploaded to the cloud server. The cloud adds Gaussian noise to the aggregation result. The standard deviation of Gaussian noise σ satisfies relaxed differential privacy. The cloud aggregation parameter W global = W e local +N(0,σ 2 ); E represents the total number of vehicles; W e local Represents the parameters of the e-th vehicle; N(0,σ 2 ) means the center is 0 and the variance is σ 2 Gaussian normal distribution sampling; distribute the cloud-aggregated parameters to the corresponding vehicles, replacing the parameter set; S303: The data distribution used in the model training phase is recorded as the source domain, and the new data distribution generated when the model is actually deployed is recorded as the target domain. The MMD distance between the source domain and the target domain is calculated; if the MMD distance is greater than the corresponding threshold, the model retraining is triggered.

[0011] By filtering support vectors using KKT conditions, only parameters of boundary samples are updated, reducing computational complexity and enabling rapid model iteration. Gaussian noise is added to aggregated parameters in the cloud to balance model performance and data security. Data distribution shifts are promptly identified, triggering model retraining to maintain prediction accuracy. According to the above scheme, S4 includes the following contents: S401: For each evidence source g∈{L obj , P state , F drift , Eye}, define the basic probability distribution function; where L obj represents the classification label output by the SVM classifier; P state represents the state probability of the carrier predicted by the LSTM model; F drift Represents the detected distribution shift mark; Eye represents the attention index collected by the eye tracker; DS evidence theory is used to calculate the joint trust; S402: Obtain the time interval t and the corresponding accident rate R from the historical accident data, and construct an exponential decay curve: R(t)=εexp(−κt)+μ, where ε represents the initial risk amplitude, κ represents the risk decay rate, and μ represents the baseline risk lower limit; use the nonlinear least squares method to fit the parameters ε, κ, and μ; calculate the real-time risk threshold Risk th (t)=R(t)+η∙std(R); η represents the system's preset sensitivity coefficient, and std(R) represents the standard deviation of the historical accident rate. If the real-time joint trust value exceeds the real-time risk threshold, an early warning is triggered and sent to the driver and system administrator.

[0012] Calculate the joint trust level to reduce the risk of misjudgment of a single piece of evidence; combine with the real-time standard deviation to dynamically adjust the threshold to adapt to risk fluctuations in different time periods (such as increased risk of nighttime transportation) and improve warning sensitivity.

[0013] Another aspect of the present application provides a transport safety detection and management system based on the Internet of Things, which is applied to the above-mentioned transport safety detection and management method based on the Internet of Things. The system includes a data acquisition and processing module, a feature extraction module, a model training module, and a dynamic warning module. The data acquisition and processing module is used to select a sensor combination based on the type label of the transported object, collect multimodal sensor data of the transported object, vehicle and driver; apply a unified time stamp and spatial coordinate to the multimodal data to generate a spatiotemporal correlation data set; The feature extraction module is used to filter the spatiotemporal correlation data set to form a core feature set, analyze the fusion weights in combination with the vehicle's operating characteristics, and generate optimized features based on the fusion weights; The model training module trains an SVM classifier based on optimized features and corresponding category labels, and trains an LSTM model using local data from each vehicle; analyzes the MMD distance between the source domain and the target domain; The dynamic warning module is used to fuse multi-source evidence to analyze joint trust and trigger abnormal warnings based on dynamic security assessments.

[0014] According to the above solution, the data acquisition and processing module includes a sensor selection unit and a data spatiotemporal association unit; The sensor selection unit is used to select a sensor combination according to the type label of the carried object. If the type label is liquid, the pressure sensor array is activated to collect pressure distribution; if the type label is fragile, the vibration sensor is activated to collect vibration spectrum; The data spatiotemporal association unit applies a unified timestamp and spatial coordinates to the multimodal data to generate a spatiotemporal association data set.

[0015] According to the above solution, the feature extraction module includes a feature analysis unit and a feature fusion unit; The feature analysis unit is used to calculate the random forest feature importance and mutual information of the physical data in the spatiotemporal correlation dataset, retaining the features that meet the random forest feature importance and mutual information greater than the corresponding threshold to form a core feature set. The vehicle acceleration function is convolved with different window lengths to generate acceleration feature quantities; The feature fusion unit analyzes fusion weights based on fusion features and generates optimized features based on the fusion weights.

[0016] According to the above solution, the model training module includes a data training unit and an offset analysis unit; The data training unit is used to train the SVM classifier using the optimized features and the corresponding category labels; and train the LSTM model using the local data of each vehicle; The offset analysis unit is used to record the data distribution used in the model training phase as the source domain, record the new data distribution generated when the model is actually deployed as the target domain, and calculate the MMD distance between the source domain and the target domain.

[0017] According to the above solution, the dynamic warning module includes a risk analysis unit and a dynamic threshold warning unit; The risk analysis unit is used to define a basic probability distribution function for each evidence source and calculate the joint trust level using DS evidence theory; The dynamic threshold warning unit is used to obtain the time intervals and corresponding accident rates in historical accident data, construct an exponential decay curve, and calculate the real-time risk threshold based on the exponential decay curve. If the real-time joint trust is greater than the real-time risk threshold, a warning is triggered and the warning information is sent to the driver and system administrator.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: the present application intelligently selects pressure or vibration sensors based on the type of carrier, avoids redundant data collection, and improves the monitoring accuracy of key parameters; calculates the signal-to-noise ratio in real time and switches to the backup sensor, dynamically eliminates noise interference data, and improves the reliability of subsequent feature extraction; unifies timestamps and spatial coordinates to eliminate spatiotemporal misalignment, constructs spatiotemporal correlation data sets, and provides data support for multi-sensor collaborative analysis; the present application combines random forest importance and mutual information to screen core features, taking into account feature prediction ability and independence, and suppressing overfitting; extracts acceleration signals through different convolution windows The local mutation volume, trend volume and global event volume comprehensively characterize the vehicle movement pattern; adaptively integrate multi-source features to enhance the contribution of key features and improve the feature space discrimination; filter support vectors through KKT conditions and only update boundary sample parameters to reduce computational complexity and achieve rapid model iteration; add Gaussian noise to cloud-based aggregation parameters to balance model performance and data security; timely identify data distribution offsets and trigger model retraining to maintain prediction accuracy; calculate joint trust to reduce the risk of misjudgment of single evidence; combine real-time standard deviation to dynamically adjust thresholds to adapt to risk fluctuations in different time periods and improve warning sensitivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a method for safety detection and management of transported goods based on the Internet of Things according to the present invention; Figure 2 This is a structural diagram of a transport safety detection and management system based on the Internet of Things of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] See also Figure 1The present invention provides a technical solution: a method for detecting and managing the safety of transported goods based on the Internet of Things, the method comprising the following steps: S1: Select a sensor combination based on the vehicle type label to collect multimodal sensor data from the vehicle, vehicle, and driver; apply unified timestamps and spatial coordinates to the multimodal data to generate a spatiotemporal correlation dataset; In S1, the following are included: S101: Label L according to the type of cargo type Select the sensor combination; if L type =Liquid, enable pressure sensor array S p Acquisition pressure distribution {p n} EQ \o(\s\up 6(N ),\s\do 2(n=1)); where p n Represents the normalized measurement value of the nth pressure sensor, n∈[1,N], N represents the total number of pressure sensors, n is a positive integer; if L type =Fragile, activate vibration sensor D h Collect vibration spectrum D h (t); S102: Analyze the signal-to-noise ratio of the i-th sensor. If the signal-to-noise ratio of the i-th sensor is less than the signal-to-noise ratio threshold, switch to the backup sensor. After switching, compare the spectrum correlation between the backup sensor and the main sensor in the overlapping frequency band. Apply a unified timestamp t to the multimodal data. j and spatial coordinates (x, y, z) j ; Generate spatiotemporal correlation dataset X ST ={(t j , x j ,y j , z j , data j )};data j represents the set of sensor measurement values ​​corresponding to the jth spatiotemporal point.

[0022] S2: Filter the spatiotemporal correlation dataset to form a core feature set, analyze the fusion weights based on the vehicle's operating characteristics, and generate optimized features based on the fusion weights; S2 contains the following: S201: Spatiotemporal correlation dataset X ST Computing Random Forest Feature Importances for Physical Data in I RF d and mutual information I MI d ; Where d represents the feature dimension number; retain the RF d > α and I MI dThe features of >β constitute the core feature set F core ={f d}, where α represents the random forest feature importance threshold, α=1 / d 1 / 2 , β represents the mutual information threshold, which is equal to the median of the mutual information in the data set; S202: Vehicle acceleration function a c (t) Convolution is performed with window length τ equal to τ1, τ2 and τ3, where τ1, τ2 and τ3 are preset constants, to obtain the local mutation amount M local =Conv1D(a c ,τ=S1), trend quantity M trend =Conv1D(a c ,τ=S2) and the event quantity M global =Conv1D(a c ,τ=S3); S203: Fusion feature X=[F core , M local , M trend , M global ] Projected to query matrix Q, key matrix K and value matrix V; Q=W q X,K=W K X,V=W V X; where W q 、W K and W V Represents the trainable parameter matrix; split the query matrix Q, key matrix K and value matrix V into A1, A2, ..., A according to the total number of attention heads M M Equal sub-vectors; calculate the matching score Score of the m-th head and all keys m =A m ⊺ K; where ⊺ represents the transpose of the vector; analyze the fusion weight w m =exp(A m ⊺ K) / exp(A r ⊺ K); where exp(∙) represents an exponential function with natural numbers as the base, r∈[1, M], r is a positive integer; based on the fusion weight, the optimized feature F is generated opt = w m ⋅[F core , M local , M trend , M global ].

[0023] S3: Train an SVM classifier based on the optimized features and corresponding class labels, and train an LSTM model using the local data of each vehicle; analyze the MMD distance between the source domain and the target domain; In S3, it contains the following: S301: Use optimization feature F opt And the corresponding category label b∈{-1,+1} trains the SVM classifier, where the radial basis kernel function is: ; Among them, x s represents the sth support vector, x represents the feature vector of the new sample to be classified, exp(∙) represents the exponential function with natural numbers as the base, γ represents the kernel function parameter, and D represents the optimized feature F opt The characteristic dimension of Var(F opt ) represents the sum of the variances of all features; Decision function: ; Among them, λ s represents the support vector x s The weight of b s represents the support vector x s The category label, c represents the bias term, sign (∙) represents the sign function, and the output +1 or -1 represents the classification result; When the new sample F new When it arrives, the support vector is filtered by KKT conditions, the original support vector and the newly added boundary samples are retained, and the reduced set S is constructed. ’ =S old ∪S new , in the reduced set S ’ Solve the dual problem and update λ s and c; S302: Use the optimized feature matrix stored in the onboard edge device of each vehicle to train the LSTM model and update the parameter set W local The parameter set includes the input gate, forget gate, output gate, and weight matrix of the candidate memory unit in the LSTM model. The locally trained parameters are encrypted and uploaded to the cloud server. The cloud adds Gaussian noise to the aggregation result. The standard deviation of Gaussian noise σ satisfies relaxed differential privacy. The cloud aggregation parameter W global = W e local +N(0,σ 2 ); E represents the total number of vehicles; W e local Represents the parameters of the e-th vehicle; N(0,σ 2 ) means the center is 0 and the variance is σ 2Gaussian normal distribution sampling; distribute the cloud-aggregated parameters to the corresponding vehicles, replacing the parameter set; S303: The data distribution used in the model training phase is recorded as the source domain, and the new data distribution generated when the model is actually deployed is recorded as the target domain. The MMD distance between the source domain and the target domain is calculated; if the MMD distance is greater than the corresponding threshold, the model retraining is triggered.

[0024] S4: Integrate multi-source evidence to analyze joint trust and trigger abnormal warnings based on dynamic security assessment.

[0025] In S4, the following contents are included: S401: For each evidence source g∈{L obj , P state , F drift , Eye}, define the basic probability distribution function; where L obj represents the classification label output by the SVM classifier; P state represents the state probability of the carrier predicted by the LSTM model; F drift Represents the detected distribution shift mark; Eye represents the attention index collected by the eye tracker; DS evidence theory is used to calculate the joint trust; Example 1: In this example, the classification label L is defined obj Basic probability distribution function: ; Among them, proposition H is high risk; Distribution offset marker F drift The generation rules are: ;D MMD represents the MMD distance between the source domain and the target domain, ξ represents the MMD distance threshold; The attention index is equal to the weighted average of the normalized gaze duration and blink frequency; S402: Obtain the time interval t and the corresponding accident rate R from the historical accident data, and construct an exponential decay curve: R(t)=εexp(−κt)+μ, where ε represents the initial risk amplitude, κ represents the risk decay rate, and μ represents the baseline risk lower limit; use the nonlinear least squares method to fit the parameters ε, κ, and μ; calculate the real-time risk threshold Risk th (t)=R(t)+η∙std(R); η represents the system's preset sensitivity coefficient, and std(R) represents the standard deviation of the historical accident rate. If the real-time joint trust value exceeds the real-time risk threshold, an early warning is triggered and sent to the driver and system administrator.

[0026] Example 2: Based on historical data fitting, ε=0.5, κ=0.1, μ=0.05; then R(t)=0.5exp(−0.1t)+0.05; when t=10 days, R(10)=0.5exp(−0.1×10)+0.05≈0.234; In this embodiment, the standard deviation of the historical accident rate std(R)=0.05; η=2; then Risk th (t)=0.5exp(−0.1t)+0.05+2×0.05; When t=5 days, Risk th (5)=0.5exp(−0.1×5)+0.05+0.1≈0.453; See also Figure 2 , the present invention provides a technical solution: a transport safety detection and management system based on the Internet of Things, the system includes a data acquisition and processing module, a feature extraction module, a model training module and a dynamic early warning module; The data acquisition and processing module is used to select a sensor combination based on the vehicle type label to collect multimodal sensor data of the vehicle, vehicle, and driver; and to apply a unified time stamp and spatial coordinate to the multimodal data to generate a spatiotemporal correlation data set; The feature extraction module is used to filter the spatiotemporal correlation data set to form a core feature set, analyze the fusion weights based on the vehicle's operating characteristics, and generate optimized features based on the fusion weights; The model training module trains an SVM classifier based on optimized features and corresponding category labels, and trains an LSTM model using local data from each vehicle; it also analyzes the MMD distance between the source and target domains; The dynamic warning module is used to integrate multi-source evidence to analyze joint trust and trigger abnormal warnings based on dynamic security assessments.

[0027] The data acquisition and processing module includes a sensor selection unit and a data spatiotemporal association unit; The sensor selection unit is used to select a sensor combination according to the type label of the carried object. If the type label is liquid, the pressure sensor array is activated to collect the pressure distribution; if the type label is fragile, the vibration sensor is activated to collect the vibration spectrum; The data spatiotemporal association unit imposes unified timestamps and spatial coordinates on multimodal data to generate spatiotemporal association datasets.

[0028] The feature extraction module includes a feature analysis unit and a feature fusion unit; The feature analysis unit calculates the random forest feature importance and mutual information for the physical data in the spatiotemporal correlation dataset, retaining the features that satisfy the random forest feature importance and mutual information greater than the corresponding threshold to form the core feature set. The vehicle acceleration function is convolved with different window lengths to generate acceleration features. The feature fusion unit analyzes the fusion weight based on the fusion feature and generates the optimized feature based on the fusion weight.

[0029] The model training module includes a data training unit and an offset analysis unit; The data training unit is used to train the SVM classifier using optimized features and corresponding category labels; and to train the LSTM model using the local data of each vehicle; The offset analysis unit is used to record the data distribution used in the model training phase as the source domain, record the new data distribution generated when the model is actually deployed as the target domain, and calculate the MMD distance between the source domain and the target domain.

[0030] The dynamic warning module includes a risk analysis unit and a dynamic threshold warning unit; The risk analysis unit is used to define the basic probability distribution function for each evidence source and calculate the joint trust level using DS evidence theory; The dynamic threshold warning unit is used to obtain the time intervals and corresponding accident rates in historical accident data, construct an exponential decay curve, and calculate the real-time risk threshold based on the exponential decay curve. If the real-time joint trust is greater than the real-time risk threshold, a warning is triggered and the warning information is sent to the driver and system administrator.

[0031] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0032] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for safety detection and management of transported goods based on the Internet of Things, characterized in that: The method comprises the following steps: S1: Select a sensor combination based on the vehicle type label to collect multimodal sensor data from the vehicle, vehicle, and driver; apply unified timestamps and spatial coordinates to the multimodal data to generate a spatiotemporal correlation dataset; S2: Filter the spatiotemporal correlation dataset to form a core feature set, analyze the fusion weights based on the vehicle's operating characteristics, and generate optimized features based on the fusion weights; S3: Train an SVM classifier based on the optimized features and corresponding class labels, and train an LSTM model using the local data of each vehicle; analyze the MMD distance between the source domain and the target domain; S4: Fusion of multi-source evidence to analyze joint trustworthiness and trigger anomaly warnings based on dynamic security assessments; In S1, the following are included: S101: Label L according to the type of cargo type Select the sensor combination; if L type =Liquid, enable pressure sensor array S p Collect pressure distribution ; where p n Represents the normalized measurement value of the nth pressure sensor, n∈[1,N], N represents the total number of pressure sensors, n is a positive integer; if L type =Fragile, activate vibration sensor D h Collect vibration spectrum D h (t); S102: Analyze the signal-to-noise ratio of the i-th sensor. If the signal-to-noise ratio of the i-th sensor is less than the signal-to-noise ratio threshold, switch to the backup sensor. After switching, compare the spectrum correlation between the backup sensor and the main sensor in the overlapping frequency band. Apply a unified timestamp t to the multimodal data. j and spatial coordinates (x, y, z) j ; Generate spatiotemporal correlation dataset X ST ={(t j , x j ,y j , z j , data j )};data j represents the set of sensor measurement values ​​corresponding to the jth spatiotemporal point; In S4, the following contents are included: S401: For each evidence source g∈{L obj , P state , F drift , Eye}, define the basic probability distribution function; where L obj represents the classification label output by the SVM classifier; P state represents the state probability of the carrier predicted by the LSTM model; F drift Represents the detected distribution shift mark; Eye represents the attention index collected by the eye tracker; DS evidence theory is used to calculate the joint trust; S402: Obtain the time interval t and the corresponding accident rate R from the historical accident data, and construct an exponential decay curve: R(t)=εexp(−κt)+μ, where ε represents the initial risk amplitude, κ represents the risk decay rate, and μ represents the baseline risk lower limit; use the nonlinear least squares method to fit the parameters ε, κ, and μ; calculate the real-time risk threshold Risk th (t)=R(t)+η∙std(R); η represents the system's preset sensitivity coefficient, and std(R) represents the standard deviation of the historical accident rate. If the real-time joint trust value exceeds the real-time risk threshold, an early warning is triggered and sent to the driver and system administrator.

2. The method for safety detection and management of transported goods based on the Internet of Things according to claim 1, characterized in that: S2 contains the following: S201: Spatiotemporal correlation dataset X ST Computing Random Forest Feature Importances for Physical Data in I RF d and mutual information I MI d ; Where d represents the feature dimension number; Retention Satisfaction I RF d > α and I MI d The features of >β constitute the core feature set F core ={f d }, where α represents the random forest feature importance threshold, α=1 / d 1 / 2 , β represents the mutual information threshold, which is equal to the median of the mutual information in the data set; S202: Vehicle acceleration function a c (t) Convolution is performed with window length τ equal to τ1, τ2 and τ3, where τ1, τ2 and τ3 are preset constants, to obtain the local mutation amount M local =Conv1D(a c ,τ=S1), trend quantity M trend =Conv1D(a c ,τ=S2) and the event quantity M global =Conv1D(a c ,τ=S3); S203: Fusion feature X=[F core , M local , M trend , M global ] Projected to query matrix Q, key matrix K and value matrix V; Q=W q X,K=W K X,V=W V X; where W q 、W K and W V represents the trainable parameter matrix; Split the query matrix Q, key matrix K and value matrix V into A1, A2, ..., A according to the total number of attention heads M. M Isovector; Calculate the matching score of the mth head and all keys m =A m ⊺ K; where ⊺ represents the transpose of the vector; analyze the fusion weight w m =exp(A m ⊺ K) / exp(A r ⊺ K); where exp(∙) represents an exponential function with natural numbers as the base, r∈[1, M], r is a positive integer; based on the fusion weight, the optimized feature F is generated opt = w m ⋅[F core , M local , M trend , M global ].

3. The method for safety detection and management of transported goods based on the Internet of Things according to claim 2, characterized in that: In S3, it contains the following: S301: Use optimization feature F opt And the corresponding category label b∈{-1,+1} trains the SVM classifier; When the new sample F new When it arrives, the support vector is filtered by the KKT condition, the original support vector and the newly added boundary samples are retained, and the reduced set S'=S is constructed. old ∪S new , solve the dual problem on the reduced set S' and update λ s and c; S302: Use the optimized feature matrix stored in the onboard edge device of each vehicle to train the LSTM model and update the parameter set W local The parameter set includes the input gate, forget gate, output gate, and weight matrix of the candidate memory unit in the LSTM model. The locally trained parameters are encrypted and uploaded to the cloud server. The cloud adds Gaussian noise to the aggregation result. The standard deviation of Gaussian noise σ satisfies relaxed differential privacy. The cloud aggregation parameter W global = W e local +N(0,σ 2 ); E represents the total number of vehicles; W e local Represents the parameters of the e-th vehicle; N(0,σ 2 ) means the center is 0 and the variance is σ 2 Gaussian normal distribution sampling; distribute the cloud-aggregated parameters to the corresponding vehicles, replacing the parameter set; S303: The data distribution used in the model training phase is recorded as the source domain, and the new data distribution generated when the model is actually deployed is recorded as the target domain. The MMD distance between the source domain and the target domain is calculated; if the MMD distance is greater than the corresponding threshold, the model retraining is triggered.

4. A transport safety detection and management system based on the Internet of Things, wherein the system is applied to the transport safety detection and management method based on the Internet of Things according to any one of claims 1 to 3, and is characterized in that: The system includes a data acquisition and processing module, a feature extraction module, a model training module and a dynamic warning module; The data acquisition and processing module is used to select a sensor combination based on the type label of the transported object and collect multimodal sensor data of the transported object, vehicle and driver; Applying unified timestamps and spatial coordinates to multimodal data to generate spatiotemporal correlation datasets; The feature extraction module is used to filter the spatiotemporal correlation data set to form a core feature set, analyze the fusion weights in combination with the vehicle's operating characteristics, and generate optimized features based on the fusion weights; The model training module trains an SVM classifier based on optimized features and corresponding category labels, and trains an LSTM model using local data from each vehicle; analyzes the MMD distance between the source domain and the target domain; The dynamic warning module is used to fuse multi-source evidence to analyze joint trust and trigger abnormal warnings based on dynamic security assessments.

5. The IoT-based transport safety detection and management system according to claim 4, characterized in that: The data acquisition and processing module includes a sensor selection unit and a data spatiotemporal association unit; The sensor selection unit is used to select a sensor combination according to the type label of the carried object. If the type label is liquid, the pressure sensor array is activated to collect pressure distribution; if the type label is fragile, the vibration sensor is activated to collect vibration spectrum; The data spatiotemporal association unit applies a unified timestamp and spatial coordinates to the multimodal data to generate a spatiotemporal association data set.

6. The IoT-based transport safety detection and management system according to claim 4, characterized in that: The feature extraction module includes a feature analysis unit and a feature fusion unit; The feature analysis unit is used to calculate the random forest feature importance and mutual information of the physical data in the spatiotemporal correlation dataset, retaining the features that meet the random forest feature importance and mutual information greater than the corresponding threshold to form a core feature set. The vehicle acceleration function is convolved with different window lengths to generate acceleration feature quantities; The feature fusion unit analyzes fusion weights based on fusion features and generates optimized features based on the fusion weights.

7. The IoT-based transport safety detection and management system according to claim 4, characterized in that: The model training module includes a data training unit and an offset analysis unit; The data training unit is used to train the SVM classifier using the optimized features and the corresponding category labels; and train the LSTM model using the local data of each vehicle; The offset analysis unit is used to record the data distribution used in the model training phase as the source domain, record the new data distribution generated when the model is actually deployed as the target domain, and calculate the MMD distance between the source domain and the target domain.

8. The IoT-based transport safety detection and management system according to claim 4, characterized in that: The dynamic warning module includes a risk analysis unit and a dynamic threshold warning unit; The risk analysis unit is used to define a basic probability distribution function for each evidence source and calculate the joint trust level using DS evidence theory; The dynamic threshold warning unit is used to obtain the time intervals and corresponding accident rates in historical accident data, construct an exponential decay curve, and calculate the real-time risk threshold based on the exponential decay curve. If the real-time joint trust is greater than the real-time risk threshold, a warning is triggered and the warning information is sent to the driver and system administrator.

Images