Shared pallet data acquisition management system and method based on machine learning model

Abstract

The application discloses a shared tray data acquisition management system and method based on a machine learning model, and relates to the technical field of data management. The shared tray data acquisition management system and method based on the machine learning model comprises the following steps: S1, preprocessing shared tray operation data, environment perception positioning data and historical impact sample data; S2, completing logistics working condition marking according to impact logarithmic likelihood ratio analysis results; S3, outputting impact type discrimination conclusions according to impact logarithmic probability ratio discrimination results; and S4, analyzing disposal priority value of the impact type discrimination conclusions, and performing intelligent early warning triggering, communication strategy scheduling and archiving management. The application solves the problem that existing systems rely on inertial measurement units three-axis acceleration peak values and fixed thresholds for impact discrimination, but normal loading and unloading impacts and drop collision impacts have similar characteristics under multiple working conditions such as warehouse loading and unloading and transportation, so that traditional methods are difficult to accurately distinguish, and false positives or false negatives are prone to occur.

Description

Technical Field

[0001] This invention relates to the field of data management technology, specifically to a shared tray data acquisition and management system and method based on machine learning models. Background Technology

[0002] With the continuous development of modern logistics, pallets, as an important carrier for unitized transportation and loading / unloading of goods, are widely used in warehousing, transportation, distribution, and delivery. Standardized loading and mechanized handling of goods using pallets can effectively improve logistics efficiency, reduce manual handling costs, and promote collaborative operations among various links in the supply chain. In large-scale warehousing systems, port logistics, and manufacturing supply chains, shared pallets are gradually becoming an important logistics infrastructure, achieving resource sharing and efficient utilization through cyclical use among different enterprises and logistics nodes. Simultaneously, with the development of IoT technology and information management methods, logistics companies are beginning to record and manage the circulation status, usage, and logistics node information of pallets through information systems to achieve pallet asset tracking and scheduling, and improve the transparency and management efficiency of logistics operations. In actual logistics operations, pallets typically undergo multiple stages, including warehouse loading / unloading, vehicle transportation, distribution and transshipment, and last-mile delivery. Their operational status is affected by various factors such as transportation conditions, loading / unloading operations, and environmental factors. Therefore, collecting and managing pallet status information during the logistics process is gradually becoming an important component of intelligent logistics management systems and is receiving continuous attention and application in the development of modern logistics informatization and intelligence.

[0003] For example, the invention patent with publication number CN121329564A discloses a method and system for optimizing the rental scheme of shared pallets. The method includes acquiring historical rental data, real-time inventory data, and predicted demand data of each node in the supply chain; collecting pallet full life cycle trajectory data based on RFID technology and establishing a digital twin model of pallet circulation to map the entire process status of pallets from warehousing, rental, transportation, recycling to repair or scrapping; optimizing pallet scheduling paths, inventory distribution, and rental pricing through an improved genetic algorithm; and combining edge computing and 5G communication to achieve data collection and cloud processing. The system also predicts the remaining service life of pallets through time series analysis and realizes functions such as dynamic adjustment of rental pricing, intelligent order matching, pallet recycling optimization, and abnormal order processing. By constructing an evaluation index system, the optimization effect is continuously evaluated and iterated, thereby improving the operational efficiency and resource utilization of shared pallet rental management.

[0004] For example, the invention patent with the publication number CN120562480A discloses an automatic loading method for freight trucks based on deep learning and stereo vision. This method includes obtaining the pixel coordinates of the cargo space and converting them to the coordinate system of the AGV forklift, planning the driving trajectory of the AGV forklift based on the converted cargo space coordinates and moving it to the front station, collecting point cloud data during the movement and performing denoising processing, extracting features and matching through an improved point cloud registration network based on Transformer, calculating the three-dimensional feature points of the cargo space and the pallet and converting them to the coordinate system of the AGV forklift; calculating the rotation adjustment amount and displacement compensation parameters of the AGV forklift in combination with the three-dimensional feature point information, correcting the driving trajectory and achieving docking; further achieving precise matching and pose correction of the pallet slots through an adaptive local attention weighting strategy, aligning the forklift forks with the pallet slots and completing the cargo grasping; optimizing the cargo stacking posture and loading angle according to the point cloud feedback information after grasping, controlling the AGV forklift to complete automatic loading, and recording the data during the loading process to continuously optimize the loading strategy and achieve the automatic loading operation of the freight truck.

[0005] However, most of the existing shared pallet-related technologies focus on the functional realization of a single link, lacking the comprehensive perception and unified management of the state of the pallet throughout the whole process of transportation, loading and unloading, and warehousing. At the same time, some systems mainly rely on the peak values of the three-axis accelerations of the inertial measurement unit and fixed thresholds for discrimination in impact event recognition. In multi-condition environments such as warehouse loading and unloading and transportation, the characteristics of normal loading and unloading impacts and drop collision impacts are relatively similar, and it is difficult for traditional discrimination methods to accurately distinguish them, easily resulting in false alarms or missed alarms. Moreover, there is a lack of an effective fusion analysis mechanism between different sources of data, making it difficult to achieve accurate identification and risk assessment of the pallet operation state.

[0006] Therefore, in view of the above problems, there is an urgent need for a shared pallet data acquisition and management system and method based on a machine learning model. Summary of the Invention

[0007] Technical Problems to be Solved

[0008] In view of the deficiencies of the prior art, the present invention provides a shared pallet data acquisition and management system and method based on a machine learning model, which solves the problem that existing systems mostly rely on the peak values of the three-axis accelerations of the inertial measurement unit and fixed thresholds for impact discrimination, but in multi-condition environments such as warehouse loading and unloading and transportation, the characteristics of normal loading and unloading impacts and drop collision impacts are similar, resulting in difficulty in accurately distinguishing by traditional methods and easy generation of false alarms or missed alarms.

[0009] Technical Solutions

[0010] To achieve the above objectives, the present invention provides the following technical solution: a shared pallet data acquisition and management method based on a machine learning model, comprising: S1, acquiring shared pallet operation data and environmental perception positioning data, obtaining historical impact sample data, and preprocessing the shared pallet operation data, environmental perception positioning data, and historical impact sample data; S2, generating an event analysis window based on the shared pallet operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis, and completing logistics condition labeling based on the impact log-likelihood ratio analysis results; S3, performing impact log-probability ratio discrimination based on the event analysis window data, and outputting impact type discrimination conclusions based on the impact log-probability ratio discrimination results; S4, performing disposal priority value analysis on the impact type discrimination conclusions, and establishing a risk priority assessment and early warning control mechanism, and performing intelligent early warning triggering, communication strategy scheduling, and archiving management of impact events.

[0011] Further, the specific process of collecting shared pallet operation data and environmental perception positioning data to obtain historical impact sample data is as follows: Collecting shared pallet operation data includes: pallet identification data, pallet-end collection timestamp data, gateway reception timestamp data, communication protocol identification data, triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, temperature data, humidity data, light intensity data, battery voltage data, and remaining power data; collecting environmental perception positioning data includes: positioning longitude data, positioning latitude data, positioning speed data, positioning status marker data, loading and unloading area identification result data, pallet stack type tilt angle data, RFID read / write event timestamp data, and read / write location identifier data; obtaining historical impact sample data and establishing a historical impact time database.

[0012] Furthermore, the specific preprocessing steps for shared pallet operation data, environmental perception positioning data, and historical impact sample data are as follows: The pallet-end collected timestamp data, gateway received timestamp data, and RFID read / write event timestamp data are aligned to a unified time reference; the read / write location identifier data and loading / unloading area identification result data are used as node anchors and bound to a unified time axis; outlier removal and noise smoothing are performed on the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data; zero-point calibration and missing data completion are performed on the outrigger weighing channel data. Background baseline updates and abrupt segment markings are performed on the numerical data of illumination intensity; continuity verification and unusable segment markings are performed on the positioning longitude data, positioning latitude data, positioning speed data, and positioning status marker data; the shared pallet operation data and environmental perception positioning data are standardized using a zero-mean unit variance standardization algorithm; and the range normalization algorithm is used to perform range normalization on the triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, illumination intensity numerical data, pallet stack type skew angle data, and positioning speed data.

[0013] Furthermore, the specific process of generating an event analysis window based on shared pallet operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis is as follows: The acceleration magnitude at the sampling time is calculated for the triaxial acceleration sequence data; values ​​greater than the previous time interval are considered as... The sampling time of the statistical mean of the acceleration magnitude of the time segment is determined as the start time of the event, and time segments of duration before and after the start time are respectively... The data fragments are used to construct an event analysis window; the outrigger weighing channel data are summed to obtain the total pallet weight sequence, and the median is taken within the event analysis window to obtain the load conditions; the working condition markers are obtained by performing working condition rule recognition calculations on the positioning speed data, positioning status marker data, RFID read / write event timestamp data, and loading / unloading area identification results data; the peak value, duration, directional component ratio, and frequency band energy distribution are extracted from the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window, and the probability density is evaluated based on historical impact sample data to obtain the drop collision observation likelihood and the loading / unloading impact observation likelihood. However, the consistency of the numerical data of light intensity, data of outrigger weighing channels, data of loading and unloading area identification, and data of positioning status markers within the event analysis window is compared to obtain the degree of evidence divergence; the ratio of the likelihood of drop collision observation plus a zero-protection constant to the likelihood of loading and unloading impact observation plus a zero-protection constant is calculated and the natural logarithm is removed to obtain the impact log-likelihood ratio; the compression function mapping result of the impact log-likelihood ratio is calculated to obtain the log-likelihood probability mapping term; the exponential function value of the opposite of the degree of evidence divergence is calculated to obtain the evidence divergence suppression term; and the product of the log-likelihood probability mapping term and the evidence divergence suppression term is calculated to obtain the impact separability confidence value.

[0014] Furthermore, the specific process of completing the logistics condition labeling based on the impact log-likelihood ratio analysis results is as follows: Real-time comparison of the impact separable confidence value and the impact separable confidence threshold: When the impact separable confidence value is less than the impact separable confidence threshold, output candidate labels for loading and unloading impacts and disturbances, and trigger data sampling enhancement, event recording, and priority communication reporting strategies. Record the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data corresponding to this event analysis window as loading and unloading impact and disturbance sample data, and write them together with the original sensor data fragments and node anchor indexes corresponding to the event analysis window into the impact event count. According to the database; when the impact separability confidence value is greater than or equal to the impact separability confidence threshold, the drop collision output candidate mark is output and the data sampling enhancement, event recording and priority communication reporting strategy are triggered. The sampling frequency of the nine-axis inertial measurement unit and vibration sensor is increased, and the sampling density of pressure and weighing sensors is increased. The event analysis window, load conditions, working condition mark and evidence divergence degree are transmitted as constraint information. At the same time, the data priority transmission mechanism is triggered to increase the transmission priority of the data corresponding to the current event analysis window in the data transmission queue of the communication module, and the data reporting is completed according to the communication link corresponding to the data identified by the communication protocol.

[0015] Furthermore, the specific process of impact log probability ratio discrimination based on event analysis window data is as follows: The impact peak value and duration are extracted from the triaxial acceleration sequence data; the attitude change amplitude is extracted from the triaxial angular velocity sequence data; the frequency band energy distribution is extracted from the vibration acceleration sequence data; the transient drift amplitude and the time required to recover to the average load range before the event are extracted from the outrigger weighing channel data; the abrupt change amplitude and duration are extracted from the light intensity numerical data; and the change within the event analysis window is extracted from the pallet-type tilt angle data. The extracted features are sequentially concatenated to form a feature vector, which is then constructed into a fusion feature set through vectorization encoding. Historical impact sample data is input, and the peak value, duration, directional component ratio, and frequency band energy distribution extracted from the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window constitute the sample features. These sample features are then vectorized to form a unified-dimensional impact feature vector, and kernel density estimation is used to determine the vector vector. An algorithm and Bayesian discriminant computation are used to construct an impact classification machine learning model. The fusion feature set corresponding to the event analysis window is input into the impact classification machine learning model, and the output is the posterior probability of drop collision and the posterior probability of loading and unloading impact. Uncertainty encoding is performed on the update discontinuity between unusable segments marked in the positioning status label data and adjacent timestamps in the positioning speed data to obtain the positioning uncertainty. The ratio of the sum of the drop collision posterior probability plus the division by zero protection constant to the sum of the loading and unloading impact posterior probability plus the division by zero protection constant is calculated and the natural logarithm is removed to obtain the impact discriminant log probability ratio. The compression function mapping result of the impact discriminant log probability ratio is calculated to obtain the impact discriminant probability mapping term. The negative exponential function value of the evidence divergence degree is calculated to obtain the evidence divergence suppression term. The negative exponential function value of the positioning uncertainty is calculated to obtain the positioning uncertainty suppression term. The product of the impact discriminant probability mapping term, the evidence divergence suppression term, and the positioning uncertainty suppression term is calculated to obtain the impact discriminant confidence value.

[0016] Furthermore, the specific process of outputting the impact type judgment conclusion based on the impact log probability ratio judgment result is as follows: Real-time comparison of the impact judgment confidence value and the impact judgment confidence threshold: When the impact judgment confidence value is less than the impact judgment confidence threshold, the loading and unloading impact and disturbance markers are output, and the current event analysis window is written into the loading and unloading impact and disturbance samples according to the working condition markers and recorded in the impact event database; When the impact judgment confidence value is greater than or equal to the impact judgment confidence threshold, the drop collision establishment marker is output, and an event conclusion log containing pallet identification data, event timestamp anchor point, positioning longitude data, positioning latitude data and read / write position marker data is generated; At the same time, the sampling strategy of the nine-axis inertial measurement unit, vibration sensor and pressure and weighing sensor is maintained for n hours after the event conclusion log is generated.

[0017] Furthermore, the specific process for analyzing the priority value of the impact type discrimination conclusion is as follows: Normalize the abrupt change amplitude and duration of the light intensity numerical data within the event analysis window to obtain the unpacking-related intensity term; normalize the transient drift amplitude and the time required to recover to the average load range before the event from the outrigger weighing channel data to obtain the weight-related intensity term; normalize the change in pallet stack type skew angle data within the event analysis window to obtain the stack type-related intensity term; calculate the ratio of the sum of the impact discrimination confidence value plus the zero-division protection constant and the sum of the complement of the impact discrimination confidence value plus the zero-division protection constant, and take the natural logarithm to obtain the logarithmic probability transformation value; calculate the product of the logarithmic probability transformation value minus the unpacking-related intensity term, the weight-related intensity term minus the weight-related intensity term, and the stack type-related intensity term minus the weight-related intensity term to obtain the evidence joint suppression term; calculate the difference between the evidence joint suppression term and the evidence joint suppression term to obtain the disposal priority value.

[0018] Furthermore, a risk priority assessment and early warning control mechanism is established. The specific process for intelligent early warning triggering, communication strategy scheduling, and archiving management of impact events is as follows: Real-time comparison of the handling priority value and the handling priority threshold: When the handling priority value is less than the handling priority threshold, recording is performed and the event entry is written to the impact event database; when the handling priority value is greater than or equal to the handling priority threshold, an early warning signal is output and pushed to the management platform; at the same time, the sending priority in the data sending queue of the communication module is increased, and the data reporting initiation timestamp data and gateway reception timestamp data are recorded; when the remaining power data is higher than the lower power limit threshold and the battery voltage data is higher than the lower voltage threshold, post-event sampling is maintained, and the corresponding event sampling frequency range is selected according to the size of the handling priority value; if the remaining power data is lower than the lower power limit threshold or the battery voltage data is lower than the lower voltage threshold, the sampling frequency of the nine-axis inertial measurement unit, vibration sensor, and positioning module is reduced, and the sampling strategy and data reporting mode are switched according to the scheduling level corresponding to the handling priority value.

[0019] The second aspect of this invention provides a shared pallet data acquisition and management system based on a machine learning model, comprising: an acquisition and preprocessing module for acquiring shared pallet operation data and environmental perception positioning data, obtaining historical impact sample data, and preprocessing the shared pallet operation data, environmental perception positioning data, and historical impact sample data; a working condition identification and event analysis window generation module for generating an event analysis window based on the shared pallet operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis, and completing logistics working condition labeling based on the impact log-likelihood ratio analysis results; a feature fusion and impact type discrimination module for performing impact log-probability ratio discrimination based on the event analysis window data, and outputting impact type discrimination conclusions based on the impact log-probability ratio discrimination results; and an intelligent early warning control and evidence package archiving module for performing disposal priority value analysis on the impact type discrimination conclusions, establishing a risk priority assessment and early warning control mechanism, and performing intelligent early warning triggering, communication strategy scheduling, and archiving management of impact events.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention collects and preprocesses shared pallet operation data and environmental perception positioning data in a unified manner, and establishes a unified time benchmark and status expression system for multi-source data, so that the pallet operation status can be continuously recorded and analyzed throughout the logistics process, thereby achieving the effect of full-process visualization monitoring of pallet operation status, effectively solving the problem of scattered pallet status information and difficulty in unified management in the prior art.

[0023] (2) This invention constructs an event analysis window and combines it with the impact log-likelihood ratio analysis mechanism to perform probabilistic discrimination processing on impact events, so that different types of impacts can be effectively distinguished in complex logistics environments, thereby achieving the effect of accurate identification of abnormal impact events, and effectively solving the problem of false alarms or missed alarms caused by relying only on the peak acceleration and fixed threshold for impact discrimination in the prior art.

[0024] (3) This invention integrates load conditions, logistics condition markings and multi-source perception evidence for comprehensive analysis, and imposes contextual constraints on the background of the impact event, thereby achieving the effect of reliably explaining the impact behavior in multiple working conditions, effectively solving the problem of difficulty in distinguishing between loading and unloading impact and drop collision in the prior art.

[0025] (4) This invention integrates and models the multi-dimensional features in the event analysis window and combines them with a machine learning model to determine the impact type, thereby transforming the impact event recognition from a single rule judgment to a multi-feature comprehensive decision-making, thus achieving the effect of improving the accuracy of impact type recognition and effectively solving the problem of insufficient single sensor data discrimination capability in the prior art.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a flowchart of the shared tray data acquisition and management method based on a machine learning model according to the present invention;

[0028] Figure 2 This is a structural diagram of the shared tray data acquisition and management system based on a machine learning model according to the present invention;

[0029] Figure 3 This is a sample similarity bubble matrix diagram of the present invention;

[0030] Figure 4 This is a three-dimensional surface diagram illustrating the probability relationship of the present invention;

[0031] Figure 5 This is a flowchart of the shared tray early warning and handling process of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figures 1-5 This invention provides a technical solution: a shared pallet data collection and management method based on a machine learning model, comprising the following steps: S1, collecting shared pallet operation data and environmental perception positioning data, obtaining historical impact sample data, and preprocessing the shared pallet operation data, environmental perception positioning data, and historical impact sample data; S2, generating an event analysis window based on the shared pallet operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis, and completing logistics condition labeling based on the impact log-likelihood ratio analysis results; S3, performing impact log-probability ratio discrimination based on the event analysis window data, and outputting impact type discrimination conclusions based on the impact log-probability ratio discrimination results; S4, performing disposal priority value analysis on the impact type discrimination conclusions, and establishing a risk priority assessment and early warning control mechanism, performing intelligent early warning triggering, communication strategy scheduling, and archiving management of impact events.

[0034] Specifically, the process of collecting shared pallet operation data and environmental perception positioning data, and obtaining historical impact sample data is as follows: Collect shared pallet operation data, which includes: pallet identification data, pallet end collection timestamp data, gateway reception timestamp data, communication protocol identification data, triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, temperature data, humidity data, light intensity data, battery voltage data, and remaining power data.

[0035] Collect environmental perception and positioning data, which includes: positioning longitude data, positioning latitude data, positioning speed data, positioning status marker data, loading and unloading area identification result data, pallet stack type tilt angle data, RFID read and write event timestamp data and read and write location identifier data.

[0036] Acquire historical impact sample data and establish a historical impact time database; the historical impact sample data was generated by pallet operation data collection during the initial deployment of the system and includes drop impact records and loading and unloading impact records.

[0037] Data is directly acquired by a nine-axis inertial measurement unit, vibration sensor, pressure and load cell, temperature and humidity sensor, photosensitive sensor, energy management unit, positioning module, 3D vision camera, and UHF RFID reader. The energy management unit includes a rechargeable lithium battery or energy harvesting module to power the multimodal sensing unit. The communication module uses Bluetooth Low Energy or Huawei NearLink communication protocol for data upload, and the gateway can transmit data back to the platform via 4G or 5G mobile communication.

[0038] This implementation plan unifies the collection and aggregation of shared pallet operation data and environmental perception positioning data to construct a multi-source observation data system covering pallet operation status, spatial location status, load change status, and environmental condition status. This enables continuous and structured recording of pallet operation behavior during loading, unloading, transportation, and warehousing, forming a comprehensive data foundation that reflects impact response characteristics, attitude change characteristics, load change characteristics, and environmental change characteristics. This provides a stable and reliable data input source for subsequent impact event analysis, working condition identification calculation, feature fusion modeling, and impact type discrimination, improving the shared pallet's status perception capability and event identification reliability in complex logistics environments, and providing a complete data support foundation for impact event tracing and regulatory analysis.

[0039] Specifically, the preprocessing process for shared pallet operation data, environmental perception positioning data, and historical impact sample data is as follows: The pallet-end acquisition timestamp data, gateway reception timestamp data, and RFID read / write event timestamp data are aligned to a unified time reference, and the read / write location identification data and loading / unloading area identification results are used as node anchors bound to a unified time axis; Outlier removal and noise smoothing are performed on the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data. Outlier removal uses a median deviation discrimination method based on a sliding time window to identify and remove abnormal sampled values ​​exceeding the local statistical range. Noise smoothing uses a sliding window mean filter or a first-order Kalman filter to smooth the sequence and performs range validity verification; Zero-point calibration and missing data completion are performed on the outrigger weighing channel data. Zero-point calibration corrects the offset using the baseline value of the weighing channel in a static state. Missing data completion restores the continuous sequence through linear interpolation of adjacent time segments, forming a continuous sequence usable for load condition calculation; Temperature numerical data... The system performs effective range verification on humidity and light intensity data. Background baseline updates and abrupt segment marking are applied to the light intensity data. The background baseline is updated using a sliding time window to obtain the ambient light reference value. Abrupt segments are marked based on the time interval where the light intensity change exceeds the baseline change threshold, used for subsequent evidence discrepancy constraints. Continuity verification and unusable segment marking are applied to positioning longitude, latitude, speed, and status marker data, used for subsequent operational condition segmentation and positioning uncertainty constraints. The shared pallet operation data and environmental perception positioning data are standardized using a zero-mean unit variance standardization algorithm to unify data scale. Range normalization is applied to triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, light intensity data, pallet stack skew angle data, and positioning speed data, mapping various continuous numerical features to a unified numerical range to form dimensionless data for subsequent feature calculations and priority formula calculations.

[0040] In this implementation plan, a unified timeline expression structure is formed across devices and sensor sources by aligning with a unified time reference and binding node anchors. A data sequence foundation with stable physical meaning is constructed through sequence anomaly verification, noise suppression, and range validity checks. A continuous observation sequence capable of stably characterizing changes in pallet load-bearing status is obtained through load-related data calibration and continuous completion. A stable observation benchmark capable of characterizing environmental state changes is formed through environmental perception data validity screening and background baseline updates. A state expression reflecting spatial location reliability is formed through location data continuity identification and unusable segment marking. A dimensionless feature expression structure with consistent numerical range and comparability is formed by uniformly scaling the shared pallet operation data and environmental perception positioning data using zero-mean unit variance standardization and range normalization algorithms. This provides a stable, unified, and computable data foundation for subsequent impact feature extraction, impact discrimination confidence value calculation, and disposal priority assessment.

[0041] Specifically, the process of generating an event analysis window based on shared pallet operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis is as follows: Calculate the acceleration magnitude at the sampling time for the triaxial acceleration sequence data, and select the acceleration magnitude greater than the previous time interval. The sampling time of the statistical mean of the acceleration magnitude of the time segment is determined as the start time of the event, and time segments of duration before and after the start time are respectively... Data fragments are used to construct an event analysis window. The event analysis window half-width duration represents the impact duration distribution range obtained from the statistical analysis of historical pallet loading / unloading impacts and drop impacts, ranging from 20 milliseconds to 200 milliseconds. The total pallet weight sequence is obtained by summing the outrigger weighing channel data, and the median is taken within the event analysis window to obtain the load condition. Working condition rules are calculated based on positioning speed data, positioning status marker data, RFID read / write event timestamp data, and loading / unloading area identification results to obtain the working condition marker. Specifically, a working condition is determined when the positioning speed data is consistently greater than the transportation speed threshold, and when the positioning speed data remains within a continuous time segment and the speed change amplitude is less than the historical threshold, a working condition is determined. The lower quantile interval of the transport velocity distribution is used to determine the loading and unloading condition when the loading and unloading area identification result data indicates that the area is in the loading and unloading area. When the positioning status marker data indicates that the positioning is unavailable and the RFID read / write event timestamp data and read / write location identifier data match the loading and unloading node anchor point in time neighborhood on the time axis, the loading and unloading operation stage condition is determined. Peak values, durations, directional component proportions, and frequency band energy distributions are extracted from the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window. Based on historical impact sample data, probability density assessment is performed to obtain the drop collision observation likelihood and the loading and unloading impact observation likelihood. By analyzing different impact modes... Statistical distribution modeling is used to create distinguishable probabilistic representations of loading / unloading impacts and drop impacts in the feature space, thereby improving the stability of subsequent likelihood judgments. Consistency comparisons are performed on numerical data of light intensity, outrigger weighing channels, loading / unloading area identification results, and positioning status marker data within the event analysis window to calculate the degree of evidence divergence. The changes in light intensity, total pallet weight, and loading / unloading area identification results within the event analysis window are calculated, and their consistency with the positioning status marker data is assessed. When discrepancies exist in the judgments of pallet status changes given by multiple data sources, statistical analysis is used to determine the differences in status judgments between different data sources. The proportion of inconsistencies yields the evidence divergence degree, which quantifies the degree of difference between the interpretation results of impact events from different sensor sources. The evidence divergence degree is used to identify normal impact behavior caused by manual handling or pallet movement in loading and unloading scenarios. When there is a significant inconsistency between inertial impact characteristics and information on illumination, load, or area status, the interpretation of drop collisions is suppressed, thereby reducing the possibility of loading and unloading impacts being misjudged as drop events. The condition identification and event analysis window generation are used to form comparable impact observation segments under multiple conditions such as loading and unloading platforms and transportation, supporting subsequent dead reckoning using inertial measurement unit data in scenarios where positioning signals are lost and maintaining the continuous expression of the event analysis window.

[0042] The ratio of the likelihood of drop collision observations plus a zero-protection constant to the likelihood of loading / unloading impact observations plus a zero-protection constant is calculated and then the natural logarithm is removed to obtain the impact log-likelihood ratio. By performing a logarithmic ratio transformation on the likelihoods of the two types of impact observations, the original probability scale is transformed into a log-likelihood space, enabling stable comparisons of observation probabilities of different orders of magnitude under a unified scale, thereby enhancing the sensitivity to differences in impact patterns. The compression function mapping result of the impact log-likelihood ratio is calculated to obtain the log-likelihood probability mapping term. The compression function restricts the log-likelihood result to within the standard probability interval, giving the impact event discrimination result a stable probabilistic expression form, while suppressing the over-amplification effect caused by extreme observation values. The phase of evidence divergence is calculated. The exponential function value of the inverse number yields the evidence discrepancy suppression term. This exponential decay mechanism is used to penalize inconsistencies among multi-source evidence, automatically reducing the weight of events with significant evidence discrepancies in subsequent calculations, thereby improving the credibility of the impact identification results. The product of the log-likelihood probability mapping term and the evidence discrepancy suppression term is calculated to obtain the impact separability credibility value. By jointly modeling the probability mapping result with evidence consistency constraints, the impact separability credibility value simultaneously reflects the degree of difference in impact patterns and the degree of consistency among multi-source evidence. This reduces the risk of false alarms and improves the reliability of drop event identification in logistics operations where loading and unloading impacts are frequent, thus enhancing the discrimination stability in complex logistics environments. The specific calculation formula is as follows:

[0043] ;

[0044] In the formula, This represents the impact separability confidence value, which measures the degree of distinguishability between two types of impact modes under given load conditions and logistics constraints. The event analysis window represents a local time analysis segment constructed around the moment of impact, used to carry out the impact characteristic calculation process; This indicates the load conditions, used to characterize the pallet load-bearing state within the event analysis window; This indicates the current stage of the pallet's logistics operation and is subject to impact mode constraints. This represents the observational likelihood of the drop collision, characterizing the degree of matching between the impact characteristics and the drop collision characteristic distribution within the event analysis window; This represents the observed likelihood of loading and unloading impacts, characterizing the degree of matching between the impact characteristics within the event analysis window and the distribution of loading and unloading impact characteristics. Indicates the degree of divergence in evidence, quantifying the level of consistency between different sensor sources' interpretations of the impact event; This represents a compression function that maps the input logarithmic ratio to a numerical range between 0 and 1 to form a probabilistic expression; This represents the division-to-zero protection constant, which prevents the numerator and denominator from reaching zero values ​​when calculating the logarithmic ratio, thus making the calculation impossible. Its value ranges from 0.0005 to 0.001.

[0045] In this implementation scheme, an event analysis window is constructed from triaxial acceleration sequence data, and a load condition expression structure is formed by combining outrigger weighing channel data, thereby achieving a localized observation expression of pallet impact behavior at a unified time scale. Through rule-based calculations using positioning velocity data, positioning status marker data, RFID read / write event timestamp data, and loading / unloading area identification results, a condition marker expression capable of characterizing the stage of logistics operations is formed. Furthermore, by statistically extracting multi-source inertial response characteristics within the event analysis window and establishing drop collision observation likelihood and loading / unloading impact observation likelihood expression structures based on historical impact sample data, a comprehensive observation expression structure is formed. This study establishes an observational probability basis capable of characterizing the probability matching degree of different impact modes; it forms an evidence divergence expression structure through multi-source state change consistency assessment, establishing quantitative constraints on the consistency of interpretation of multi-sensor information; and it constructs an impact separability confidence value expression through log-likelihood ratio transformation, compression function mapping, and evidence divergence suppression calculation, enabling impact mode difference information and multi-source evidence consistency information to form a stable expression structure in a unified probability space. This achieves a quantitative assessment of the distinguishability between loading and unloading impacts and drop collision impacts, providing a reliable basis with stable discrimination capability for subsequent impact type discrimination and abnormal event identification.

[0046] Specifically, the process of labeling logistics conditions based on the results of the impact log-likelihood ratio analysis is as follows: real-time comparison of the impact separability confidence value and the impact separability confidence threshold.

[0047] When the impact separability confidence value is less than the impact separability confidence threshold, output the analysis window marker for loading and unloading impact and disturbance candidate events. Record the triaxial acceleration sequence data, triaxial angular velocity sequence data and vibration acceleration sequence data corresponding to this event analysis window as loading and unloading impact and disturbance sample data, and write them together with the original sensor data fragments and node anchor indexes corresponding to the event analysis window into the impact event database for subsequent classification machine learning model training and sample statistical updates.

[0048] When the impact separability confidence value is greater than or equal to the impact separability confidence threshold, a drop collision candidate event analysis window marker is output, and the sampling frequency of the nine-axis inertial measurement unit and vibration sensor is increased. The sampling frequency of the nine-axis inertial measurement unit and vibration sensor is switched to the event sampling frequency, with a value range of 50 Hz to 800 Hz, to obtain denser impact response data. At the same time, the sampling density of the pressure and load sensors is increased to strengthen the load evidence within the event analysis window, and the sampling frequency of the pressure and load sensors is switched to the load sampling frequency, with a value range of 10 Hz to 200 Hz. The event analysis window, load conditions, operating condition markers, and evidence divergence are transmitted as constraint information. At the same time, a data priority transmission mechanism is triggered to increase the transmission priority of the data corresponding to the current event analysis window in the data transmission queue of the communication module, and data reporting is completed according to the communication link corresponding to the data identified by the communication protocol.

[0049] In this implementation plan, a graded response mechanism for the distinguishability of impact events is formed by real-time calculation of the impact separability confidence value and impact separability confidence threshold, enabling differentiated processing strategies for loading / unloading impacts and drop collision impacts. By accumulating samples and recording impact event data in the analysis window for events with low separability confidence values, a continuously expanding foundation of impact sample data is established to support the training and statistical updating of classification machine learning models. A high-sampling-density observation mechanism is triggered for candidate events with high separability confidence values, enhancing the response recording capabilities of the nine-axis inertial measurement unit, vibration sensors, and pressure and weighing sensors during the impact phase, thus forming more complete evidence of dynamic impact response. By uniformly transmitting event analysis windows, load conditions, operating condition markers, and evidence divergence as constraint information, combined with a data-priority transmission mechanism, a rapid reporting channel and complete evidence recording mechanism for abnormal impact events are constructed, thereby improving the data acquisition accuracy, abnormal event response efficiency, and impact behavior tracing capabilities during the impact event identification process.

[0050] Specifically, the process of determining the impact logarithmic probability ratio based on event analysis window data is as follows: The impact peak value and duration are extracted from the triaxial acceleration sequence data. The extreme value changes and duration intervals of the acceleration modulus within the event analysis window are analyzed to characterize the intensity and duration of the impact. The attitude change amplitude is extracted from the triaxial angular velocity sequence data. The range of angular velocity changes within the event analysis window is statistically analyzed to characterize the degree of attitude disturbance generated by the pallet under impact. The frequency band energy distribution is extracted from the vibration acceleration sequence data. Frequency domain analysis of the vibration acceleration sequence is performed to calculate the energy proportion of each frequency band, which characterizes the impact vibration response. The system extracts the corresponding spectral characteristics; extracts the transient drift amplitude and the time required to recover to the average load range before the event from the outrigger weighing channel data, and analyzes the changing trend of the load sequence before and after the impact to characterize the load disturbance degree and recovery process characteristics; extracts the abrupt change amplitude and duration from the light intensity numerical data to reflect the pallet's possible opening or ambient light changes during the impact; extracts the change within the event analysis window from the pallet stack skew angle data to characterize the degree of attitude displacement of the cargo stacking structure under impact; and concatenates the extracted features in sequence to form a feature vector, which is then constructed into a fusion feature set through vectorized encoding. This process forms a unified-dimensional multi-source feature representation structure to support subsequent probability discrimination calculations. Inputting historical impact sample data includes historical drop-collision impact samples and historical loading / unloading impact samples. Sample features are derived from peak values, durations, directional component proportions, and frequency band energy distribution extracted from triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window. Each sample feature is vectorized to form a unified-dimensional impact feature vector. An impact classification machine learning model is then constructed based on kernel density estimation and Bayesian discriminant calculations to perform probability density fitting and discrimination of impact behavior. The fusion feature set corresponding to the event analysis window is input into the impact classification machine learning model, and the output is the posterior probability of drop collision and the posterior probability of loading and unloading impact. Uncertainty encoding is performed on the update discontinuity between the unusable segments marked in the positioning status marker data and the adjacent timestamps in the positioning velocity data to obtain the positioning uncertainty. The positioning uncertainty is a continuous numerical index, which is calculated by normalizing the update discontinuity duration between adjacent timestamps in the positioning velocity data. When the positioning status marker data indicates that the positioning is unusable or the update discontinuity duration increases, the positioning uncertainty increases accordingly. It is used to quantify the reliability of the positioning information and participate in the subsequent impact discrimination confidence value calculation.

[0051] The ratio of the sum of the posterior probabilities of drop collisions plus a protection constant divided by zero to the sum of the posterior probabilities of loading / unloading impacts plus a protection constant divided by zero is calculated and then the natural logarithm is taken to obtain the impact discrimination logarithmic probability ratio. By transforming the posterior probabilities into logarithmic ratios, the probability results output by the classification machine learning model can express the relative dominance of the two types of impact events in logarithmic probability form. The compression function mapping result of the impact discrimination logarithmic probability ratio is calculated to obtain the impact discrimination probability mapping term. A nonlinear mapping is applied to the logarithmic probability result through the compression function to maintain a stable probability expression and enhance the numerical robustness of the impact discrimination result. The inverse exponential function value of the degree of evidence divergence is calculated to obtain the evidence divergence suppression term. The consistency of multi-source evidence is used as... Constraints are introduced into the impact discrimination calculation process to avoid excessive influence of single feature anomalies on the discrimination results. The negative exponential function value of the positioning uncertainty is calculated to obtain the positioning uncertainty suppression term. Exponential decay processing is applied to the reliability of positioning information, automatically reducing the confidence level of the impact discrimination results when the positioning state is unstable, thereby improving the overall recognition reliability. The product of the impact discrimination probability mapping term, the evidence divergence suppression term, and the positioning uncertainty suppression term is calculated to obtain the impact discrimination confidence value. The joint product structure achieves comprehensive constraints on three factors: feature matching probability, multi-source evidence consistency, and positioning reliability, enabling the impact discrimination results to remain stable and reliable under multi-source information conditions. The specific calculation formula is as follows:

[0052] ;

[0053] In the formula, This represents the impact discrimination confidence value, indicating the overall confidence level of an impact event within the event analysis window belonging to the drop collision type. This represents the fused feature set, and the impact behavior representation space formed by the combination of feature vectors from multiple sensors. This represents the posterior probability of a fall collision, and the probability that the fused feature is classified as a fall collision event in the classification machine learning model. represents the posterior probability of loading and unloading impact, and represents the probability result of the fused feature being judged as a loading and unloading impact event in the classification machine learning model; Indicates the degree of divergence in evidence, quantifying the level of consistency between different sensor sources' interpretations of the impact event; This indicates the uncertainty in positioning, describing the degree of change in the reliability of positioning information within the event analysis window; This indicates the event analysis window, which limits the time frame for impact characteristic calculation and probability assessment; This indicates the load conditions, reflecting the influence of the pallet's load-bearing state on the impact response characteristics; This indicates the operational status markers and the logistics operation context in which the impact event is determined. This represents a compression function that maps the input logarithmic ratio to a numerical range between 0 and 1 to form a probabilistic expression; This represents the division-to-zero protection constant, which prevents the numerator and denominator from reaching zero values ​​when calculating the logarithmic ratio, thus making the calculation impossible. Its value ranges from 0.0005 to 0.001.

[0054] In this embodiment, Table 1 is a data table of parameters for calculating the impact discrimination confidence value. It records in detail the posterior probability of drop collision, posterior probability of loading and unloading impact, degree of evidence divergence, degree of positioning uncertainty, and the finally calculated impact discrimination confidence value for different event examples in the process of calculating the impact discrimination confidence value. It is used to quantify the degree of confidence that the impact event belongs to the drop collision type under different logistics conditions and multi-source evidence conditions. Specifically: For Example 1, the posterior probability of drop collision is 0.20, the posterior probability of loading / unloading impact is 0.80, the degree of evidence divergence is 0.10, the location uncertainty is 0.10, and the impact judgment confidence value is 0.164; for Example 2, the posterior probability of drop collision is 0.45, the posterior probability of loading / unloading impact is 0.55, the degree of evidence divergence is 0.40, the location uncertainty is 0.20, and the impact judgment confidence value is 0.247; for Example 3, the posterior probability of drop collision is 0.80, the posterior probability of loading / unloading impact is 0.20, the degree of evidence divergence is 0.10, the location uncertainty is 0.10, and the impact judgment confidence value is 0.65. 4; The posterior probability of drop collision for example 4 is 0.80, the posterior probability of loading and unloading impact is 0.20, the degree of evidence divergence is 0.80, the location uncertainty is 0.10, and the confidence value of impact judgment is 0.325; The posterior probability of drop collision for example 5 is 0.80, the posterior probability of loading and unloading impact is 0.20, the degree of evidence divergence is 0.10, the location uncertainty is 0.80, and the confidence value of impact judgment is 0.325; The posterior probability of drop collision for example 6 is 0.52, the posterior probability of loading and unloading impact is 0.48, the degree of evidence divergence is 0.10, the location uncertainty is 0.10, and the confidence value of impact judgment is 0.426.

[0055] Table 1 Data table of parameters for calculating the confidence value of impact discrimination

[0056]

[0057] like Figure 3The bubble matrix diagram of sample similarity is shown. Combined with Table 1, it can be seen that there are significant differences in the similarity of different event samples in the key feature parameter space. Specifically, Sample 3 and Sample 6 show high consistency in parameters such as drop collision posterior probability and evidence divergence, with larger bubble sizes and higher average confidence values ​​corresponding to their colors, indicating high similarity between them in the impact event discrimination feature space. Sample 1 and Sample 3 show significantly reduced similarity due to a large difference in drop collision posterior probability, with noticeably smaller bubble sizes. Although Sample 4 and Sample 5 maintain consistency in drop collision posterior probability, their similarity is moderate due to increased differences in evidence divergence or location uncertainty. Overall, the bubble matrix diagram of impact discrimination confidence value sample similarity intuitively reflects the distribution relationship and similarity changes of different impact event samples in the feature space, and can be used to analyze the differences in discrimination features between different impact patterns.

[0058] like Figure 4 The diagram shows a three-dimensional surface plot of the probability relationship. Combined with Table 1, it can be seen that the impact discrimination confidence value exhibits a clear non-linear distribution trend with the changes in the posterior probability of the drop collision and the degree of evidence divergence. Specifically, when the degree of evidence divergence remains low, the impact discrimination confidence value shows a significant upward trend as the posterior probability of the drop collision increases. For example, in Example 3, when the posterior probability of the drop collision is 0.80 and the degree of evidence divergence is 0.10, its impact discrimination confidence value reaches 0.654. ​​However, when the degree of evidence divergence increases significantly, even if the posterior probability of the drop collision remains high, the impact discrimination confidence value is still significantly suppressed. For example, in Example 4, when the degree of evidence divergence increases to 0.80, its impact discrimination confidence value decreases to 0.325. Furthermore, when the two types of impact posterior probabilities are close, such as in Example 6, its impact discrimination confidence value is in the intermediate range. Overall, the three-dimensional surface plot of the probability relationship of the impact discrimination confidence value intuitively demonstrates the coupling relationship between the probability discrimination result and the consistency of multi-source evidence, providing a visual representation of the changing pattern of the confidence value of impact event identification.

[0059] In this implementation scheme, multi-source feature extraction and fusion are performed on triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, illumination intensity numerical data, and pallet stacking tilt angle data to construct a unified-dimensional impact behavior feature representation space. This enables a comprehensive characterization of the pallet's dynamic response characteristics, attitude disturbance characteristics, load change characteristics, and environmental change characteristics during impact events. An impact classification machine learning model is constructed based on historical impact sample data, outputting the posterior probability of drop collisions and the posterior probability of loading / unloading impacts, allowing impact behavior to be distinguished probabilistically. Furthermore, by introducing positioning uncertainty and evidence analysis... Dissimilarity is used as a constraint in the impact discrimination calculation, enabling the impact event discrimination process to simultaneously consider the degree of feature matching, consistency of multi-source evidence, and reliability of location information, thereby reducing the impact of single sensor feature anomalies on the discrimination results. By performing logarithmic probability transformation on the posterior probability and combining it with compression function mapping to form a stable probability expression, and then using an exponential decay mechanism to suppress and adjust evidence discrepancies and location uncertainties, a multi-factor joint constraint impact discrimination credibility value calculation mechanism is formed. This enables impact events to maintain stable and reliable identification results under complex logistics environments and multiple working conditions, thereby improving the separability between drop collisions and loading / unloading impacts and improving the overall impact identification accuracy.

[0060] Specifically, the process of outputting the impact type judgment conclusion based on the impact log probability ratio judgment result is as follows: Real-time comparison of the impact judgment confidence value and the impact judgment confidence threshold:

[0061] When the impact discrimination confidence value is less than the impact discrimination confidence threshold, the loading and unloading impact and disturbance markers are output. The event analysis window is written into the loading and unloading impact and disturbance samples according to the working condition markers and recorded in the impact event database. It is then transmitted to the intelligent early warning, adaptive control and evidence package archiving modules for sample library accumulation and model version iteration training.

[0062] When the impact discrimination confidence value is greater than or equal to the impact discrimination confidence threshold, a drop collision establishment marker is output, and an event conclusion log containing pallet identification data, event timestamp anchor points, positioning longitude data, positioning latitude data, and read / write location identifier data is generated for risk tracing. Simultaneously, the sampling strategy of the nine-axis inertial measurement unit, vibration sensor, and pressure and weighing sensor is maintained for n time duration after the event conclusion log is generated. n represents the sampling duration after the event, which is derived from the statistical results of the subsequent vibration attenuation process in historical drop impact and loading / unloading impact events. This is to ensure that the subsequent evolution stage of the impact can be completely recorded. The value of n ranges from 0.5 seconds to 5 seconds to ensure that the evidence package fragments cover the subsequent evolution of the event. The fused feature index and discrimination conclusion are then transmitted to the intelligent early warning, adaptive control, and evidence package archiving modules for priority calculation and early warning output.

[0063] In this implementation plan, by performing real-time calculations of the impact discrimination confidence value and impact discrimination confidence threshold, a hierarchical identification result expression structure for impact event types is formed, realizing a differentiated discrimination and handling mechanism for loading and unloading impact disturbance behavior and drop collision anomaly events. By recording samples of loading and unloading impact and disturbance events and accumulating impact event databases, a continuously expanding impact sample data foundation is formed to support the version iteration training of classification machine learning models. By generating event conclusion logs for drop collision events and maintaining a multi-sensor continuous sampling strategy, a complete observation evidence fragment covering the impact occurrence stage and the post-impact evolution stage is formed, improving the continuity and completeness of abnormal impact behavior records. By transmitting the fused feature index and discrimination conclusions to the intelligent early warning, adaptive control, and evidence package archiving modules, a closed-loop data support structure is constructed to transmit impact event identification results to the disposal priority assessment and early warning response process, thereby improving the overall reliability and regulatory capability of shared pallet impact event identification, risk tracing, and anomaly early warning handling.

[0064] Specifically, the process of prioritizing the impact type determination is as follows: The amplitude and duration of abrupt changes in light intensity data within the event analysis window are normalized to obtain the unpacking-related intensity term; the transient drift amplitude and the time required to recover to the average load range before the event are extracted from the outrigger weighing channel data and normalized to obtain the weight-related intensity term; the change in pallet stack skew angle data within the event analysis window is normalized to obtain the stack-related intensity term. Normalization eliminates differences in the dimensions of different sensors, mapping various features to a unified numerical range. The upper and lower bounds of normalization are jointly determined by the corresponding sensor range and the feature distribution range obtained from historical impact sample data statistics, thus forming comparable evidence indicators. A logarithmic probability transformation is performed on the impact determination confidence value to obtain a logarithmic probability transformation value, which maps probabilistic indicators to evidence strength quantities, enabling different pieces of evidence to express their contribution to the priority of handling in a non-linear form during fusion calculation.

[0065] The logarithmic probability transformation value is obtained by calculating the ratio of the sum of the impact discrimination confidence value plus the zero-protection constant to the sum of the complement of the impact discrimination confidence value plus the zero-protection constant, and taking the natural logarithm. This logarithmic probability transformation converts the probabilistic index into an unbounded expression of evidence strength, facilitating its participation in multi-source evidence fusion calculations. The joint suppression term is obtained by multiplying the logarithmic probability transformation value, the unpacking-related strength term, the weight-related strength term, and the stacking-related strength term. This product structure is used to simulate the joint suppression relationship between multi-source risk evidence. When multiple risk factors are simultaneously weak, the overall risk is significantly suppressed, thus improving the stability of risk assessment. The difference between the joint suppression term and the evidence joint suppression term is calculated to obtain the disposal priority value. By performing a complement transformation on the joint suppression result, the risk strength can be expressed in the form of cumulative probability, thereby achieving a comprehensive assessment of multi-source risk signals. The specific calculation formula is as follows:

[0066] ;

[0067] ;

[0068] In the formula, This indicates the priority value, representing the overall urgency of handling the impact event under the combined effect of multiple sources of risk evidence; This represents the logarithmic probability transformation value, used to map the impact discrimination confidence value from a probability interval to an intensity characterization quantity that can participate in multi-source evidence fusion calculations; This indicates the intensity of the unboxing-related factors, characterizing the degree of abnormal unboxing risk reflected by the sudden changes in lighting behavior within the event analysis window; This indicates a weight-related strength item, describing the risk of abnormal movement or loss of goods as reflected by changes in pallet load; This represents the strength terms related to the pallet type, reflecting the stability risk of the stacking structure corresponding to changes in the pallet stack's posture; This represents the division-to-zero protection constant, which prevents the numerator and denominator from reaching zero values ​​when calculating the logarithmic ratio, thus making the calculation impossible. Its value ranges from 0.0005 to 0.001.

[0069] In this implementation plan, by normalizing the numerical data of light intensity, the data of the support leg weighing channel, and the data of the pallet stack tilt angle, and constructing the unpacking-related intensity item, the weight-related intensity item, and the stack type-related intensity item, a unified intensity expression for changes in environmental state, load state, and stack structure state is achieved. By performing a logarithmic probability transformation on the impact discrimination confidence value to form a logarithmic probability transformation value, the probabilistic discrimination result can be transformed into an evidence strength quantity that can participate in evidence fusion calculation. By constructing a multi-source evidence joint suppression structure and calculating the disposal priority value, the impact discrimination result, the unpacking risk, the load anomaly risk, and the stack stability risk are fused and evaluated under a unified mathematical expression framework, thereby forming a quantitative indicator that can comprehensively reflect the degree of impact event risk and the urgency of disposal, providing a unified risk assessment basis for subsequent early warning output, anomaly disposal decision-making, and evidence package archiving.

[0070] Specifically, the process of establishing a risk priority assessment and early warning control mechanism, and carrying out intelligent early warning triggering, communication strategy scheduling, and archiving management for impact events is as follows: Figure 5 The diagram shows the shared tray early warning and handling process, which compares the handling priority value and the handling priority threshold in real time.

[0071] When the handling priority value is less than the handling priority threshold, the event entry is recorded and written to the impact event database, and the impact event data update and model version iteration are initiated to form a continuous convergence closed loop of false alarms and false negatives.

[0072] When the handling priority value is greater than or equal to the handling priority threshold, an early warning signal is output and pushed to the management platform. The early warning information includes at least the tray identification data, event timestamp anchor point, positioning longitude data, positioning latitude data, read / write location identification data, impact type conclusion, and handling priority value. Simultaneously, the transmission priority in the communication module's data transmission queue is increased, and the data reporting initiation timestamp data and gateway reception timestamp data are recorded for link tracing. When the remaining battery power data is higher than the lower battery power threshold and the battery voltage data is higher than the lower voltage threshold, post-event sampling is maintained. The corresponding event sampling frequency range is selected according to the handling priority value to obtain impact response data with higher time resolution, used to supplement evidence of the recovery status after the event. The sampling frequency of the nine-axis inertial measurement unit and vibration sensor is determined according to the handling priority value. Dynamic mapping adjustments are implemented to continuously change the sampling frequency within the event sampling frequency range according to the handling priority value. If the remaining power data is lower than the lower power threshold or the battery voltage data is lower than the lower voltage threshold, the sampling frequency of the nine-axis inertial measurement unit, vibration sensor, and positioning module is reduced. The sampling strategy and data reporting mode are switched according to the scheduling level corresponding to the handling priority value. The sampling frequency and data reporting cycle are dynamically adjusted through the handling priority value to balance the integrity of event evidence and terminal energy consumption control under power-limited conditions. The scheduling mechanism links the impact event risk level with the sampling strategy, communication priority, and reporting behavior, so that high-risk events receive higher sampling density and higher transmission priority, while reducing the sampling and communication load in low-risk or low power states, thereby achieving terminal energy consumption control and communication link congestion relief. Early warning information and event entry archiving are used to form traceable regulatory evidence for the entire logistics chain, supporting intelligent and precise supervision of the flow process of palletized products such as petrochemicals, food, and pharmaceuticals. Multimodal collaborative perception and multi-level data fusion reduce false alarm rates and improve the integrity of status perception.

[0073] This implementation plan triggers an impact event early warning and data reporting mechanism by real-time determination of the disposal priority value and disposal priority threshold, enabling rapid identification and timely notification of high-risk impact events. It records key information such as pallet identification data, event timestamp anchor points, location longitude data, location latitude data, read / write location identifier data, impact type conclusions, and disposal priority values, while simultaneously recording the data reporting initiation timestamp data and gateway reception timestamp data, forming a traceable event chain record. Through dynamic mapping of sampling frequency and communication priority scheduling based on disposal priority values, it establishes a linkage between the impact event risk level and data collection density, communication transmission priority, and data reporting cycle. This allows for the acquisition of higher time-resolution impact response data and ensures timely transmission of key evidence during high-risk events. Under power-constrained conditions, it achieves terminal energy consumption control by reducing sampling frequency and adjusting data reporting cycle. Finally, through the archiving of early warning information and event entries, it forms a traceable regulatory evidence system covering pallet transportation, loading / unloading, and warehousing processes, improving the impact risk monitoring capability, event response efficiency, and the completeness of status perception in the logistics chain.

[0074] like Figure 2 As shown, the second aspect of this invention provides a shared pallet data acquisition and management system based on a machine learning model, comprising: an acquisition and preprocessing module, used to acquire shared pallet operation data and environmental perception positioning data, obtain historical impact sample data, preprocess the shared pallet operation data, environmental perception positioning data, and historical impact sample data to construct a multi-source observation data foundation covering pallet motion state, attitude response state, load change state, and environmental perception state, and form a unified data expression structure that can participate in subsequent impact event analysis and calculation through time base alignment and data quality verification; and a working condition identification and event analysis window generation module, used to generate event analysis windows based on shared pallet operation data and environmental perception positioning data and perform impact log-likelihood ratio analysis, and complete logistics working condition labeling based on the impact log-likelihood ratio analysis results to realize the identification of loading, unloading, transportation, and operation stages. The system identifies the status of different logistics operation scenarios and provides impact observation segments with unified time range constraints for impact behavior feature calculation; the feature fusion and impact type discrimination module is used to perform impact log probability ratio discrimination based on event analysis window data, and outputs impact type discrimination conclusions based on the impact log probability ratio discrimination results. It achieves reliable differentiation between drop collision impact and loading and unloading impact behavior through the fusion expression and probability discrimination calculation of feature vectors from multiple sources; the intelligent early warning control and evidence package archiving module is used to analyze the handling priority value of the impact type discrimination conclusion, and establish a risk priority assessment and early warning control mechanism. It performs intelligent early warning triggering, communication strategy scheduling and archiving management of impact events to form adaptive sampling and communication scheduling control based on the degree of impact risk, and builds an impact event evidence recording and traceable supervision data system for the entire logistics process.

[0075] This implementation plan unifies the collection, processing, and analysis of shared pallet operation data, environmental perception and positioning data, and historical impact sample data to construct a multi-source data processing and decision-making system covering pallet operation status, logistics operation condition identification, impact behavior characteristic expression, and impact type discrimination. This enables automatic identification and type differentiation of impact events during logistics transportation and loading / unloading. Through event analysis window generation, impact probability discrimination, and disposal priority value evaluation mechanisms, an intelligent early warning control and communication scheduling strategy based on the degree of impact risk is formed. This ensures that impact events can be promptly identified, recorded, and reported after they occur, forming a complete event evidence record structure. This improves the shared pallet's impact risk perception capability, event response efficiency, and full-process traceability and supervision capability in complex logistics environments.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A shared tray data collection and management method based on a machine learning model, characterized in that, Includes the following steps: S1, collect shared pallet operation data and environmental perception positioning data, obtain historical impact sample data, and preprocess the shared pallet operation data, environmental perception positioning data and historical impact sample data; S2, based on shared pallet operation data and environmental perception positioning data, generate an event analysis window and perform impact log-likelihood ratio analysis, and complete logistics condition marking based on the impact log-likelihood ratio analysis results; S3, based on the event analysis window data, performs impact log probability ratio discrimination, and outputs the impact type discrimination conclusion based on the impact log probability ratio discrimination result; S4 analyzes the priority values ​​of the impact type identification conclusions, establishes a risk priority assessment and early warning control mechanism, and performs intelligent early warning triggering, communication strategy scheduling and archiving management for impact events.

2. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for collecting shared pallet operation data and environmental perception positioning data to obtain historical impact sample data is as follows: Collect shared pallet operation data, which includes: pallet identification data, pallet end collection timestamp data, gateway received timestamp data, communication protocol identification data, triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, temperature data, humidity data, light intensity data, battery voltage data, and remaining power data. Collect environmental perception and positioning data, which includes: positioning longitude data, positioning latitude data, positioning speed data, positioning status marker data, loading and unloading area identification result data, pallet stack type tilt angle data, RFID read and write event timestamp data and read and write location identifier data; Acquire historical impact sample data and establish a historical impact time database.

3. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for preprocessing shared pallet operation data, environmental perception and positioning data, and historical impact sample data is as follows: The pallet-side timestamp data, gateway-received timestamp data, and RFID read / write event timestamp data are aligned to a unified time reference, and the read / write location identification data and loading / unloading area identification results are used as node anchors and bound to a unified time axis; outlier removal and noise smoothing are performed on the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data; zero-point calibration and missing data completion are performed on the outrigger weighing channel data. Background baseline updates and abrupt segment markings are performed on the numerical data of illumination intensity; continuity verification and unusable segment markings are performed on the positioning longitude data, positioning latitude data, positioning speed data, and positioning status marker data; the shared pallet operation data and environmental perception positioning data are standardized using a zero-mean unit variance standardization algorithm; and the range normalization algorithm is used to perform range normalization on the triaxial acceleration sequence data, triaxial angular velocity sequence data, vibration acceleration sequence data, outrigger weighing channel data, illumination intensity numerical data, pallet stack type skew angle data, and positioning speed data.

4. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process of generating an event analysis window based on shared tray operation data and environmental perception positioning data and performing impact log-likelihood ratio analysis is as follows: Calculate the acceleration magnitude at each sampling time point for the triaxial acceleration sequence data, and assign the acceleration magnitude greater than the previous time point by a certain value. The sampling time of the statistical mean of the acceleration magnitude of the time segment is determined as the start time of the event, and time segments of duration before and after the start time are respectively... The data fragments are used to construct an event analysis window; the data of the outrigger weighing channel are summed to obtain the total weight sequence of the pallet, and the median is taken within the event analysis window to obtain the load condition; The working condition marker is obtained by performing working condition rule recognition calculations on the positioning speed data, positioning status marker data, RFID read / write event timestamp data, and loading / unloading area identification result data. Peak values, durations, directional component proportions, and frequency band energy distributions were extracted from the triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window. Probability density assessments were performed based on historical impact sample data to obtain the observed likelihoods of drop collisions and loading / unloading impacts. Consistency comparisons were performed on the numerical data of illumination intensity, outrigger weighing channels, loading / unloading area identification results, and positioning status marker data within the event analysis window to calculate the degree of evidence divergence. The ratio of the drop collision observation likelihood plus a zero-protection constant to the loading and unloading impact observation likelihood plus a zero-protection constant is calculated and then the natural logarithm is removed to obtain the impact log-likelihood ratio. The compression function mapping result of the impact log-likelihood ratio is calculated to obtain the log-likelihood probability mapping term. The exponential function value of the inverse of the degree of evidence divergence is calculated to obtain the evidence divergence suppression term. The product of the log-likelihood probability mapping term and the evidence divergence suppression term is calculated to obtain the impact separability confidence value.

5. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for completing the logistics condition labeling based on the impact log-likelihood ratio analysis results is as follows: Real-time comparison of shock separability confidence value and shock separability confidence threshold: When the impact separability confidence value is less than the impact separability confidence threshold, output the loading and unloading impact and disturbance output candidate flags and trigger the data sampling enhancement, event recording and priority communication reporting strategy. Record the triaxial acceleration sequence data, triaxial angular velocity sequence data and vibration acceleration sequence data corresponding to the event analysis window as loading and unloading impact and disturbance sample data, and write them together with the original sensor data fragments and node anchor point indexes corresponding to the event analysis window into the impact event database. When the impact separability confidence value is greater than or equal to the impact separability confidence threshold, a drop collision output candidate marker is output, triggering data sampling enhancement, event logging, and priority communication reporting strategies. The sampling frequency of the nine-axis inertial measurement unit and vibration sensor is increased, while the sampling density of the pressure and weighing sensors is also increased. The event analysis window, load conditions, working condition markers, and evidence disagreement are transmitted as constraint information, and a data priority transmission mechanism is triggered to increase the transmission priority of the data corresponding to the current event analysis window in the data transmission queue of the communication module. Data reporting is completed according to the communication link corresponding to the data identified by the communication protocol.

6. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for determining the log-probability ratio of impacts based on event analysis window data is as follows: The impact peak value and duration are extracted from triaxial acceleration sequence data; attitude change amplitude is extracted from triaxial angular velocity sequence data; frequency band energy distribution is extracted from vibration acceleration sequence data; transient drift amplitude and time required to recover to the average load range before the event are extracted from outrigger weighing channel data; abrupt change amplitude and duration are extracted from light intensity numerical data; and changes within the event analysis window are extracted from pallet stack type tilt angle data. The extracted features are sequentially concatenated to form a feature vector, which is then constructed into a fusion feature set through vectorization encoding. Historical impact sample data is input, and the peak value, duration, directional component ratio, and frequency band energy distribution extracted from triaxial acceleration sequence data, triaxial angular velocity sequence data, and vibration acceleration sequence data within the event analysis window constitute sample features. The sample features are vectorized to form a unified dimension impact feature vector, and an impact classification machine learning model is constructed based on kernel density estimation algorithm and Bayesian discriminant calculation. The fusion feature set corresponding to the event analysis window is input into the impact classification machine learning model, which outputs the posterior probability of drop collision and the posterior probability of loading and unloading impact. The positioning uncertainty is obtained by encoding the update discontinuities between unavailable segments marked in the positioning status marker data and adjacent timestamps in the positioning velocity data. The ratio of the sum of the posterior probability of drop collision plus a zero protection constant to the sum of the posterior probability of loading / unloading impact plus a zero protection constant is calculated and then the natural logarithm is removed to obtain the impact discrimination log probability ratio. The compression function mapping result of the impact discrimination log probability ratio is calculated to obtain the impact discrimination probability mapping term. The inverse exponential function value of the degree of evidence divergence is calculated to obtain the evidence divergence suppression term. The inverse exponential function value of the positioning uncertainty is calculated to obtain the positioning uncertainty suppression term. The impact discrimination confidence value is obtained by multiplying the impact discrimination probability mapping term, the evidence divergence suppression term, and the location uncertainty suppression term.

7. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for outputting the impact type judgment conclusion based on the impact log probability ratio judgment result is as follows: Real-time comparison of impact discrimination confidence value and impact discrimination confidence threshold: When the impact discrimination confidence value is less than the impact discrimination confidence threshold, output the loading and unloading impact and disturbance flags, write the current event analysis window into the loading and unloading impact and disturbance samples according to the working condition flags and record it to the impact event database; When the impact judgment confidence value is greater than or equal to the impact judgment confidence threshold, a drop collision establishment mark is output, and an event conclusion log containing pallet identification data, event timestamp anchor point, positioning longitude data, positioning latitude data and read / write location identification data is generated; at the same time, the sampling strategy of the nine-axis inertial measurement unit, vibration sensor and pressure and weighing sensor is maintained for n time periods after the event conclusion log is generated.

8. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process for analyzing the priority value of the impact type discrimination conclusion is as follows: Normalize the abrupt change amplitude and duration of the light intensity numerical data within the event analysis window to obtain the unpacking-related intensity term; extract the transient drift amplitude and the time required to recover to the average load range before the event from the outrigger weighing channel data and normalize them to obtain the weight-related intensity term; normalize the change in pallet stacking skew angle data within the event analysis window to obtain the stacking-related intensity term. The logarithmic probability transformation value is obtained by calculating the ratio of the sum of the impact discrimination confidence value plus the zero-divide-off protection constant to the sum of the complement of the impact discrimination confidence value plus the zero-divide-off protection constant, and taking the natural logarithm. The joint evidence suppression term is obtained by calculating the product of the logarithmic probability transformation value minus the unpacking related strength term, the weight related strength term minus the weight related strength term, and the stacking type related strength term. The disposal priority value is obtained by calculating the difference between the joint evidence suppression term and the evidence suppression term.

9. The shared tray data collection and management method based on a machine learning model according to claim 1, characterized in that: The specific process of establishing a risk priority assessment and early warning control mechanism, and carrying out intelligent early warning triggering, communication strategy scheduling, and archiving management for impact events is as follows: Real-time comparison of processing priority values ​​and processing priority thresholds: When the handling priority value is less than the handling priority threshold, the record is executed and the event entry is written to the impact event database; When the handling priority value is greater than or equal to the handling priority threshold, an early warning signal is output and pushed to the management platform; at the same time, the sending priority in the data sending queue of the communication module is increased, and the data reporting initiation timestamp data and gateway reception timestamp data are recorded. When the remaining power data is higher than the lower power limit threshold and the battery voltage data is higher than the lower voltage limit threshold, the sampling continues after the event, and the corresponding event sampling frequency range is selected according to the processing priority value. If the remaining power data is lower than the lower power limit threshold or the battery voltage data is lower than the lower voltage limit threshold, the sampling frequency of the nine-axis inertial measurement unit, vibration sensor and positioning module will be reduced, and the sampling strategy and data reporting mode will be switched according to the scheduling level corresponding to the handling priority value.

10. A shared tray data acquisition and management system based on a machine learning model, employing the shared tray data acquisition and management method based on a machine learning model as described in any one of claims 1-9, comprising: The data acquisition and preprocessing module is used to acquire shared pallet operation data and environmental perception and positioning data, obtain historical impact sample data, and preprocess the shared pallet operation data, environmental perception and positioning data, and historical impact sample data. The working condition identification and event analysis window generation module is used to generate event analysis windows based on shared pallet operation data and environmental perception positioning data, and perform impact log-likelihood ratio analysis. Based on the impact log-likelihood ratio analysis results, the logistics working condition is marked. The feature fusion and impact type discrimination module is used to discriminate the impact log probability ratio based on event analysis window data, and output the impact type discrimination conclusion based on the impact log probability ratio discrimination result; The intelligent early warning control and evidence archiving module is used to analyze the priority value of the impact type judgment conclusion, establish a risk priority assessment and early warning control mechanism, and carry out intelligent early warning triggering, communication strategy scheduling and archiving management of impact events.

Images