Order exception response method based on logistics status

Abstract

This invention relates to the field of logistics distribution network monitoring technology, and discloses a method for order anomaly response based on logistics status. The method includes: acquiring parcel and vehicle data from network points and stitching them together to obtain first node load characteristics; obtaining second node load characteristics through random forest classification; adjusting the RFID tag acquisition frequency to obtain a first acquisition frequency and acquiring a first set of abnormal orders; spatially clustering the first set of abnormal orders to obtain initial clusters; calculating the stacking density based on their boundary contours and area; analyzing the average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density to obtain the RFID wake-up period; comparing with order tracking data to obtain tracking delay results; reducing the dimensionality of the temperature and humidity sensor distribution density and acquired data to obtain a second resource allocation matrix; matching a preset frequency mapping interval to obtain a second acquisition frequency and acquiring a second set of abnormal orders. This method can solve the problem of delayed anomaly response caused by mismatch between fixed acquisition frequency and node load.

Description

Technical Field

[0001] This invention relates to the field of logistics and distribution network monitoring technology, and in particular to a method for responding to order anomalies based on logistics status. Background Technology

[0002] Currently, in the field of logistics and distribution network monitoring technology, the logistics and distribution network serves as a key infrastructure supporting the operation of the e-commerce economy. It consists of a network of sensors distributed across various nodes that collaboratively collect data on the status of packages and vehicles. Its operational efficiency directly affects the timely delivery of hundreds of millions of orders.

[0003] In one existing technology, a logistics monitoring system collects data on parcel accumulation and vehicle entry / exit at fixed intervals. Regardless of node load, RFID devices and temperature / humidity sensors periodically go into sleep mode and wake up at preset frequencies. The dispatch center responds to parcel flow and capacity allocation based on the collected status data. When a large number of parcels suddenly flood into a transit warehouse due to promotional activities or weather conditions, the fixed collection frequency cannot capture the parcel accumulation status and flow anomalies in time. RFID devices miss key statuses during peak periods due to low-frequency cycles. The dispatch center struggles to grasp the true load of nodes due to information lag, and the physical stagnation of parcels on the sorting line is delayed in detection. Environmentally sensitive orders also experience quality deviations due to insufficient temperature and humidity monitoring frequency. Congestion that could have been resolved by adding vehicles and adjusting sorting resources in certain areas evolves into systemic anomalies affecting the entire delivery chain due to delayed detection.

[0004] Existing technologies suffer from a mismatch between the fixed data collection frequency of IoT devices and the actual load of nodes, leading to delayed abnormal order responses. Summary of the Invention

[0005] This invention provides a method for responding to order anomalies based on logistics status, in order to solve the problem of delayed anomaly response caused by mismatch between fixed collection frequency and node load.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for responding to order anomalies based on logistics status, comprising:

[0007] Obtain parcel accumulation data and vehicle entry and exit data of the target logistics network point, concatenate the parcel accumulation data and vehicle entry and exit data to obtain the first node load feature, and use a pre-constructed random forest to classify the first node load feature to obtain the second node load feature.

[0008] The sensory resource distribution data of the RFID tag is obtained based on the load characteristics of the second node as a first resource allocation matrix. The acquisition frequency of the RFID tag is adjusted according to the first resource allocation matrix to obtain a first acquisition frequency. The status data of the unsorted package node is collected according to the first acquisition frequency to classify and obtain a first abnormal order set.

[0009] Spatial clustering is performed on the first abnormal order set to obtain initial clusters. The stacking density is calculated based on the boundary contour and area of ​​the initial clusters. The average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density are analyzed to obtain the radio frequency wake-up cycle.

[0010] The order tracking data of the first abnormal order set is obtained according to the radio frequency wake-up cycle. The actual flow time difference is calculated based on the order tracking data. The time difference deviation value is obtained by comparing it with the preset standard flow time difference. The tracking delay result is obtained based on the time difference deviation value.

[0011] Based on the tracking delay result, obtain the spatial coordinates of the temperature and humidity sensor network and calculate the sensor distribution density. Reduce the dimensionality of the sensor distribution density and the temperature and humidity sensor data to obtain a second resource allocation matrix. Match the second resource allocation matrix with a preset frequency mapping interval to obtain a second acquisition frequency. Collect the environmental condition data of the first abnormal order set according to the second acquisition frequency to obtain a second abnormal order set.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] (1) The present invention performs feature splicing operation on the parcel accumulation data and vehicle entry and exit data of the target logistics network point, and uses a pre-constructed random forest to classify the spliced ​​first node load features to obtain the second node load features that can characterize the true load level of the node. Through this load feature classification, the system can dynamically trigger subsequent resource allocation based on the storage capacity saturation and transportation scheduling tension of the node, avoid scheduling misjudgment caused by a single indicator, and provide a high-precision basis for the dynamic adjustment of the collection frequency.

[0014] (2) The present invention adjusts the acquisition frequency of the radio frequency identification tag based on the first resource allocation matrix, and compares the time difference of the order tracking data by combining the radio frequency wake-up period obtained by spatial clustering; by dynamically allocating the distribution of the radio frequency identification tag sensing resources, the system can increase the acquisition frequency when the node load increases to capture the status changes of unsorted packages in a timely manner, and adjust the tag communication timing in areas with high stacking density, thereby reducing the probability of reading collisions and the impact of scanning blind spots, and significantly advancing the discovery of abnormal orders.

[0015] (3) The present invention obtains the spatial coordinates of the temperature and humidity sensor network points and calculates the sensor distribution density based on the tracking delay result. The sensor distribution density and the temperature and humidity sensor data are reduced in dimensionality to obtain the second resource allocation matrix. The second acquisition frequency is obtained by matching the preset frequency mapping interval. By dynamically allocating the temperature and humidity sensor acquisition resources, the system can perform high-frequency environmental data acquisition on environmentally sensitive orders when the tracking delay exceeds the limit, and promptly identify quality abnormalities caused by improper temperature control, thereby reducing the risk of order delays and cargo damage, and improving the real-time performance and accuracy of abnormal handling in the logistics and distribution network. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an order anomaly response method based on logistics status provided in the first embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Reference Figure 1 The first embodiment of the present invention provides a method for responding to order anomalies based on logistics status, including the following steps:

[0019] S1, acquire parcel accumulation data and vehicle entry and exit data of the target logistics network point, concatenate the parcel accumulation data and vehicle entry and exit data to obtain the first node load feature, and use a pre-constructed random forest to classify the first node load feature to obtain the second node load feature.

[0020] S2, based on the load characteristics of the second node, obtain the sensing resource distribution data of the RFID tag as the first resource allocation matrix, adjust the acquisition frequency of the RFID tag according to the first resource allocation matrix to obtain the first acquisition frequency, and collect the status data of the unsorted package node according to the first acquisition frequency to classify and obtain the first abnormal order set.

[0021] S3, perform spatial clustering on the first abnormal order set to obtain initial clusters, calculate the stacking density based on the boundary contour and area of ​​the initial clusters, and analyze the average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density to obtain the radio frequency wake-up cycle.

[0022] S4, obtain the order tracking data of the first abnormal order set according to the radio frequency wake-up cycle, calculate the actual flow time difference based on the order tracking data, compare it with the preset standard flow time difference to obtain the time difference deviation value, and obtain the tracking delay result based on the time difference deviation value;

[0023] S5. Based on the tracking delay result, obtain the spatial coordinates of the temperature and humidity sensor network points and calculate the sensor distribution density. Reduce the dimensionality of the sensor distribution density and the temperature and humidity sensor data to obtain a second resource allocation matrix. Match the second resource allocation matrix with a preset frequency mapping interval to obtain a second acquisition frequency. Collect the environmental condition data of the first abnormal order set according to the second acquisition frequency to obtain a second abnormal order set.

[0024] In step S1, parcel accumulation data and vehicle entry / exit data of the target logistics network point are obtained. The parcel accumulation data and vehicle entry / exit data are concatenated to obtain the first node load feature. The first node load feature is then classified using a pre-constructed random forest to obtain the second node load feature, including:

[0025] Obtain the original parcel entry time record and vehicle arrival time record, and perform timestamp alignment processing according to the time window to obtain aligned parcel stacking data and aligned vehicle entry and exit data;

[0026] The warehouse capacity saturation of the target logistics network is calculated based on the aligned parcel stacking data and the maximum design capacity of the target logistics network. The network capacity scheduling tension is determined based on the truck queuing waiting time and the number of currently available trucks in the aligned vehicle entry and exit data.

[0027] The saturation of the storage capacity of the network points and the tension of the transportation capacity scheduling of the network points are concatenated into a two-dimensional vector to obtain the load characteristics of the first node.

[0028] The load features of the second node are obtained by classifying the load features of the first node using a pre-built random forest.

[0029] First, the original parcel entry time records and vehicle arrival time records are obtained, and timestamp alignment is performed according to the time window to obtain aligned parcel stacking data and aligned vehicle entry and exit data. There is usually a time difference between parcel scanning records and vehicle gate records at logistics network points; timestamp alignment eliminates analysis errors caused by asynchronous data. The parcel stacking volume and vehicle entry and exit frequency within the same hour are matched, with a time window length of 1 hour and a sliding step of 10 minutes. The window overlap rate is no less than 80% to ensure temporal continuity. The time window length and sliding step are determined based on the fluctuation cycle (15 to 30 minutes) of parcel arrival volume in the historical entry records of the target logistics network point. The window length covers at least two complete fluctuation cycles to smooth random fluctuations, and the sliding step is no greater than the minimum fluctuation cycle to avoid missing sudden load changes. The window overlap rate is verified through historical data playback; when it is no less than 80%, the load characteristic jump between adjacent windows is less than 5%, meeting the temporal continuity requirement. The aligned parcel stacking data (unit: pieces) and aligned vehicle entry and exit data (unit: vehicles / hour) within each window are output.

[0030] Secondly, the saturation of the network's storage capacity is calculated based on the aligned parcel stacking data and the maximum design capacity of the target logistics network. The network's capacity scheduling tension is determined based on the truck queuing time and the current number of available trucks in the aligned vehicle entry and exit data. The network's storage capacity saturation is expressed as a percentage, calculated as the ratio of the current number of stacked parcels to the network's maximum design capacity. For example, if there are 5000 stacked parcels and the network's maximum design capacity is 8000, the saturation is 62.5%. The network's capacity scheduling tension is assessed using a weighted summation based on preset weights. The ratio of truck queuing time to the preset maximum tolerable waiting time is used as the time urgency component, and the ratio of the available truck shortage to the total number of rated trucks is used as the resource shortage component. These components are weighted and summed using preset weights and mapped to a tension score from 0 to 100. The preset maximum tolerable waiting time is set at 60 minutes, based on the 90th percentile statistical value of tolerable waiting time in the historical queuing records of the target logistics network and the network scheduling rules; the total number of rated trucks is set at 5, which is the rated number of vehicles in the network scheduling configuration table; the weight of the time urgency component is set at 0.6, and the weight of the resource gap component is set at 0.4, determined by least squares fitting from historical congestion samples—using the two components of the same historical period as independent variables and whether an actual congestion alarm occurred as the dependent variable in the regression, the regression coefficients are normalized to 0.6 and 0.4, and the classification consistency rate is verified to be no less than 85% after leaving a sample. For example, when the queuing waiting time is 40 minutes and there are 2 available trucks, the time urgency component is 40 / 60≈0.67, the resource gap component is (5-2) / 5=0.6, and the tension score is (0.6×0.67+0.4×0.6)×100=64.2 points.

[0031] Then, the saturation of the storage capacity at the network points and the tension of the network point transportation capacity scheduling are concatenated into a two-dimensional vector to obtain the load feature of the first node. The saturation value representing storage pressure and the tension value representing transportation capacity status are concatenated into a vector to output the load feature of the first node (for example, when the saturation is 62.5% and the tension is 64.2%, the load feature of the first node is a two-dimensional vector [62.5, 64.2]), which is used for subsequent random forest classification.

[0032] Finally, a pre-constructed random forest was used to classify the load features of the first node to obtain the load features of the second node. The pre-constructed random forest is a supervised classification model. The training samples are the load features of the first node (two components: network storage capacity saturation and network transportation capacity scheduling tension) for each time window in the operation logs of the past 3 months. The sample labels are manually labeled according to the formation rules: if there are additional vehicle dispatches or congestion alarm records in the scheduling logs within the following 1 hour of the window, it is labeled as a high load warning; otherwise, it is labeled as a low load and stable. Sample preprocessing includes removing missing windows caused by sensor offline and performing z-score standardization on the two components respectively. The sample set was randomly divided into training and validation sets in an 8:2 ratio. Offline training was performed using Gini impurity as the node splitting criterion. The number of decision trees and the maximum depth were determined through grid search. Among candidate combinations of 50 / 100 / 200 trees and 5 / 10 / 15 maximum depths, 100 trees and a maximum depth of 10 resulted in the highest validation set accuracy (not less than 90%) and manageable overfitting risk; therefore, these were selected as the model structure parameters. A validation set accuracy of not less than 90% was used as the criterion for deploying the model online. Once this criterion was met, online inference was deployed. The load characteristics of the first node were input into the random forest, which output the load characteristics of the second node. The second node load characteristics consisted of the load level label (high load warning or low load stability) and the corresponding classification confidence score output by the random forest, used to trigger subsequent resource allocation for IoT devices. When the classification result was low load stability, resource allocation was not triggered, the RFID tag maintained its current baseline acquisition frequency, and the load assessment was re-executed in the next monitoring cycle. The two types of outputs formed a closed loop.

[0033] In step S2, the sensing resource distribution data of the RFID tags is obtained based on the load characteristics of the second node as a first resource allocation matrix. The acquisition frequency of the RFID tags is adjusted according to the first resource allocation matrix to obtain a first acquisition frequency. Unsorted package node status data is collected according to the first acquisition frequency and classified to obtain a first abnormal order set, including:

[0034] The frequency band occupancy distribution of the RFID tags is extracted based on the first resource allocation matrix to obtain the initial RFID coverage network;

[0035] Based on the initial radio frequency coverage network and the time-series position coordinates of the RFID tag, smoothing and trajectory tracking are performed on the time-series position coordinates to extract the movement trajectory of the RFID tag. The reading collision probability is obtained by calculating the ratio of the number of time slots that collide within the reading cycle to the total number of time slots based on the movement trajectory.

[0036] The frequency adjustment step size is determined based on the mapping relationship between the read collision probability and the preset adjustment amount. The larger the read collision probability, the larger the frequency adjustment step size.

[0037] The current reference acquisition frequency of the RFID tag is updated according to the frequency adjustment step size to obtain the first acquisition frequency;

[0038] The RF reader antenna is controlled to transmit a continuous wave signal according to the first acquisition frequency to obtain the RF echo signal reflected by the unsorted package;

[0039] The received signal strength indication of the radio frequency echo signal is extracted as an attenuation feature, and the arrival time difference between signal transmission and reception is extracted. The attenuation feature and the arrival time difference are smoothed using a Kalman filter algorithm to obtain smoothed position coordinates. The actual moving speed of the unsorted package is calculated based on the difference between the smoothed position coordinates of adjacent sampling times and the sampling time interval. Timing starts when the actual moving speed approaches zero to obtain the dwell time sequence of the unsorted package.

[0040] If the dwell time sequence exceeds the preset timeout warning threshold, an abnormal status identifier is generated;

[0041] Node status data is generated based on the status anomaly identifier and the logistics waybill number and volume information of the unsorted packages. The node status data is then filtered and classified based on the status anomaly identifier to obtain the first set of abnormal orders.

[0042] First, after the high-load warning value in the second node's load characteristics is triggered, the sensor resource distribution data of RFID tags within the target logistics network point is obtained through the warehouse IoT platform. This includes the deployment density, operating frequency band, and transmission power of tags in each area, and is organized into a first resource allocation matrix. The matrix rows represent network point areas, and the columns represent resource dimensions (number of tags, occupied frequency band, and transmission power). Based on the first resource allocation matrix, the number of RFID tags and their occupied frequency bands in each area are counted according to the network point area, resulting in the frequency band occupancy distribution of each area. For example, in the sorting exit area, 42 out of 50 tags occupy the ultra-high frequency band (920MHz to 925MHz). The frequency band occupancy distribution of each area is correlated with the spatial location of the area, and an initial RFID coverage network characterizing the signal coverage range and frequency band overlap of each area is output.

[0043] Secondly, based on the initial RF coverage network and the temporal position coordinates of the RFID tags, smoothing and trajectory tracking are performed on the temporal position coordinates to extract the movement trajectory of the RFID tags. The RFID reader reports the temporal position coordinates of the tags at a fixed sampling period. The original coordinate sequence has jitter due to signal reflection and multipath effects. The jitter is eliminated by performing a moving average smoothing process on the coordinate sequences of adjacent sampling points. Then, the smoothed coordinate points are connected in chronological order to obtain the dynamic movement trajectory of the RFID tags on the conveyor belt. Based on the movement trajectory, the set of tags within the same reading range in the same reading period is determined. The RFID reader initiates reading one by one using the slotted ALOHA anti-collision mechanism. The ratio of the number of time slots with response collisions in the reading period to the total number of reading time slots is counted, and the reading collision probability is output. For example, if 13 out of 20 reading time slots collide, the reading collision probability is 65%.

[0044] Then, the frequency adjustment step size is determined based on the mapping relationship between the read collision probability and the preset adjustment amount. The preset adjustment amount mapping relationship is a pre-configured positive correlation mapping table between the collision probability and the frequency adjustment step size. This mapping table is established by historical test statistics: the read collision probability, channel occupancy rate and tag read success rate are recorded at different load periods of the target network. The collision probability corresponding to a read success rate of less than 95% (20%) is used as the starting adjustment boundary, and the collision probability corresponding to a success rate of less than 85% (50%) is used as the increased adjustment boundary. The adjustment step size is the minimum channel spacing (2MHz) and the cross-channel group spacing (5MHz) required to reduce the collision probability to below 20%. The higher the read collision probability, the larger the frequency adjustment step size. For example, the adjustment step size is 0 when the collision probability is less than 20%, 2MHz when it is between 20% and 50%, and 5MHz when it is greater than 50%. In this embodiment, the read collision probability is 65%, and the frequency adjustment step size is 5MHz according to the mapping table. The current reference acquisition frequency (e.g., 920MHz) of the RFID tag is updated according to the frequency adjustment step size. The reference acquisition frequency is then increased by the adjustment step size, and the tag is switched to a nearby channel with a lighter load to obtain the first acquisition frequency (e.g., 925MHz). It should be noted that in this embodiment, the first acquisition frequency refers to the RFID tag's RF operating frequency (carrier frequency, unit: MHz), which is used to reduce read collisions by switching channels. The RFID reader performs reads according to a fixed acquisition period (unit: s, which can be converted to acquisition frequency, unit: Hz). In this step, the acquisition period remains unchanged, and only the operating channel is switched.

[0045] Furthermore, the RFID reader antenna is controlled to transmit a continuous wave signal according to the first acquisition frequency. When an unsorted package on the conveyor belt passes through the scanning area, the RFID tag on the package surface is activated and reflects the signal. The receiving hardware captures the reflected wave and outputs an RFID echo signal containing amplitude and phase information. The Received Signal Strength Indication (RSSI, dBm) of the RFID echo signal is extracted as an attenuation feature, and the time difference of arrival (ns) between signal transmission and reception is extracted. The above features are smoothed using a Kalman filter algorithm. The state vector of the Kalman filter represents the position and velocity of the package along the conveyor belt. The state transition matrix is ​​discretized according to the uniform motion model of the conveyor belt. The observation vector consists of RSSI and time difference of arrival. The observation equation uses the RSSI path loss model (RSSI decreases logarithmically with transmission distance, and the attenuation coefficient is calibrated by the RF reader at the factory) and the time difference of arrival ranging formula (speed of light multiplied by half of the time difference of arrival after deducting the fixed system delay) to convert the two observations into the distance observation between the package and the reader antenna. The fixed system delay includes the reader link delay, tag response delay, and hardware synchronization error, which is determined in advance through calibration experiments: the reference tag is placed at a known distance (e.g., 1m) directly in front of the reader antenna, and the time difference of arrival is measured multiple times. The average of the difference between the measured time difference of arrival and the theoretical propagation delay corresponding to the known distance is the fixed system delay. Each reader is calibrated once after installation and the configuration is written. The package position observation (unit: m) is then calculated by combining the known installation coordinates of the reader antenna. That is, the observation vector is a 1-dimensional position observation. The process noise covariance matrix Q is a 2×2 matrix, with the diagonal elements representing the variances of the position components (taken as 1×10⁻). 4The variance of the velocity component (0.01 m / s², corresponding to the conveyor belt positioning jitter) and the variance of the speed component (0.01 m / s², corresponding to the random disturbance variance of the conveyor belt speed) are used, with off-diagonal elements set to 0. The observation noise covariance R is the scalar position observation variance (0.1 m², determined by the comprehensive measurement variance of distance converted from RSSI and arrival time difference through reader factory calibration and on-site calibration tests). The system outputs smoothed position coordinates within a specific physical grid of the sorting machine through a two-step iterative prediction and update process. The actual moving speed of the package (unit: mm / s) is calculated based on the difference in smoothed position coordinates between adjacent sampling times and the sampling time interval. For example, if the difference in smoothed position coordinates between adjacent sampling times is 12 mm and the sampling time interval is 100 ms, then the actual moving speed is 120 mm / s. Timing starts when the actual moving speed approaches 0 (less than 5 mm / s). The criteria for setting this threshold are: the lower limit of speed noise corresponding to the comprehensive error of position calculation (approximately 2 mm / s), 10% of the minimum normal operating speed of the conveyor belt (approximately 50 mm / s), and an integer value of 5 mm / s between the two values ​​is obtained by combining the speed distribution statistics of historical stationary samples. This value can be adjusted within the range of 3 mm / s to 10 mm / s according to the positioning accuracy. The output is a dwell time sequence representing the duration of the package's stay at the current node, for example, 120 s.

[0046] Then, if the dwell time series exceeds the preset timeout warning threshold, a status anomaly flag is generated. The preset timeout warning threshold is 90 seconds, which physically represents the maximum dwell time of a package during normal flow at a single node of the sorting line. It is set based on the empirical value of normal flow time calculated from the average conveyor belt speed of 0.2 m / s and the single-segment flow path length of 16 m (16 m ÷ 0.2 m / s = 80 s), with a 10-second tolerance allowed. This tolerance can be adjusted according to the conveyor belt speed within the range of 60 to 120 seconds. The dwell time series of 120 seconds is compared with the preset timeout warning threshold of 90 seconds. If it exceeds the normal flow range, a Boolean logic judgment rule triggers the status anomaly flag (anomaly code 1), indicating that the package has physically stopped on the sorting line. If the dwell time series does not exceed the preset timeout warning threshold, the status anomaly flag is normal code 0, and the package is not included in the subsequent abnormal order set.

[0047] Finally, node status data is generated based on the status anomaly identifier and the tracking number and volume information of the unsorted packages. This node status data is then filtered and categorized according to the status anomaly identifier to obtain the first set of abnormal orders. The node status data (tracking number, volume, dwell time, location coordinates) of packages with a status anomaly identifier of 1 are filtered out from the full set of node status data and categorized into anomaly categories. The first set of abnormal orders belonging to these categories is then output, for example, containing 50 abnormal tracking numbers requiring manual intervention.

[0048] In step S3, spatial clustering is performed on the first abnormal order set to obtain initial clusters. The stacking density is calculated based on the boundary contours and area of ​​the initial clusters. The average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density are analyzed to obtain the radio frequency wake-up cycle, including:

[0049] Obtain the indoor spatial coordinates and shelf level of each package in the first abnormal order set, and construct a three-dimensional location matrix;

[0050] Based on the three-dimensional position matrix, the density peak clustering algorithm is used to calculate the local density of each package node and the relative distance to higher density nodes. Nodes whose local density and relative distance both satisfy the preset clustering parameters are taken as cluster centers, and the remaining package nodes are assigned to the nearest cluster center to obtain multiple initial clusters.

[0051] Based on the node coordinates of the initial clusters, the convex hull algorithm is used to extract the line connecting the outermost wrapping nodes of each initial cluster as the boundary contour, and the area of ​​the region enclosed by the boundary contour is calculated. The total number of wrapping nodes in the initial clusters is divided by the area of ​​the region to obtain the stacking density.

[0052] The initial clusters whose stacking density meets the preset congestion conditions are taken as high-density congestion areas. The average physical distance between adjacent packages in the high-density congestion areas is calculated, and the distribution characteristics of scanning blind spots that cannot be covered by the signal are extracted by combining the radiation angle model of the radio frequency reader.

[0053] When the proportion of the blind area of ​​the scanning blind area distribution feature is not lower than the preset blind area proportion threshold, the communication timing of the radio frequency identification tag is adjusted according to the scanning blind area distribution feature to obtain the radio frequency wake-up period.

[0054] First, obtain the indoor spatial coordinates and shelf level of each package in the first abnormal order set to construct a three-dimensional position matrix. Input the indoor positioning data of 50 abnormal packages in the first abnormal order set, extract the horizontal and vertical spatial coordinates (unit: m) of each package in the sorting workshop, and obtain its shelf level (e.g., the second level of the temporary storage rack). Convert the shelf level to the actual height coordinates (e.g., the second level corresponds to a height of 3.6m) according to the preset shelf height (1.8m per level, determined according to the temporary storage rack design drawings). Obtain the unified horizontal, vertical, and height three-dimensional spatial coordinates in meters. Vectorize the above three-dimensional spatial data and output a three-dimensional position matrix containing the coordinate information of 50 nodes to represent the physical spatial distribution of abnormal packages.

[0055] Secondly, based on the three-dimensional position matrix, a density peak clustering algorithm is used to calculate the local density of each package node and its relative distance to higher-density nodes. The local density ρ_i is the number of nodes in a spherical neighborhood centered on node i with a preset cutoff distance d_c as its radius. The relative distance δ_i to higher-density nodes is the Euclidean distance from node i to the nearest node among all nodes with a local density greater than ρ_i. The preset cutoff distance d_c is set to 2.0m, which physically represents the spatial threshold for determining whether two packages belong to the same cluster. This threshold is set based on the physical size of a single cell on the sorting line's temporary storage rack (approximately 1.5m) and considers package overflow margins; it can be adjusted within the range of 1.0m to 3.0m. The preset clustering parameters include two components: a lower limit for local density and a lower limit for relative distance. The lower limit for local density is set to 5 nodes (below this value is considered sparse noise), and the lower limit for relative distance is set to 3.0m (above this value indicates that the node is independent of other high-density areas). Both components jointly determine the cluster center, avoiding misjudgments caused by comparisons of a single dimension. Nodes whose local density and relative distance both meet the preset clustering parameters are established as cluster centers, and the remaining package nodes are assigned to the nearest cluster center, outputting 3 initial clusters to indicate 3 independent package retention and accumulation areas existing next to the sorting line.

[0056] Then, based on the node coordinates of the initial clusters, the convex hull algorithm is used to extract the line connecting the outermost nodes of each initial cluster as the boundary contour, and the area enclosed by the boundary contour is calculated. The total number of packages in the initial cluster is divided by the area to obtain the stacking density. Before the convex hull calculation, the nodes in the initial clusters are grouped according to shelf levels based on height coordinates, and nodes at the same level are projected onto the horizontal plane of that level. The convex hull algorithm uses the Graham scan method, inputting the two-dimensional planar coordinates of the nodes at the same level in the cluster, and outputting the clockwise line connecting the outermost nodes as the boundary contour. The area enclosed by the contour (unit: m²) is then calculated using the shoelace formula. The stacking density is calculated for each level, and the maximum value is taken as the stacking density of the initial cluster. For example, if the area enclosed by the boundary contour of the second level of an initial cluster is 2 m², and the total number of packages in that level is 30, then the stacking density of that level is 30 ÷ 2 = 15 packages / m².

[0057] Furthermore, initial clusters whose stacking density meets the preset congestion conditions are designated as high-density congestion areas. The preset congestion conditions are defined as stacking densities within a pre-configured high-load range (e.g., stacking density greater than 10 pieces / m²). Physically, this means the critical density at which package stacking affects RFID signal reading. The setting is based on the sorting line's designed throughput density (approximately 8 pieces / m²) with a 20% safety margin (8 × 1.2 = 9.6 pieces / m², rounded to 10 pieces / m² for engineering purposes). The range boundaries can be adjusted according to the sorting line width. In this embodiment, an initial cluster has a stacking density of 15 pieces / m², meeting the preset congestion conditions, and is thus designated as a high-density congestion area for subsequent blind zone analysis. If the stacking density does not meet the preset congestion conditions, this cluster is not designated as a high-density congestion area, and the corresponding package maintains its existing RF wake-up cycle.

[0058] Then, the average physical distance between adjacent packages in the high-density congestion area is calculated, and the distribution characteristics of the scanning blind zone that the signal cannot cover are extracted by combining the radiation angle model of the radio frequency reader. The average physical distance is the mean of the Euclidean distances (in meters) between all adjacent packages within the area, used to characterize the density of package stacking. The radiation angle model of the RF reader is pre-constructed as follows: with the known installation coordinates of the reader antenna as the vertex and the antenna orientation (the beam pointing angle issued during installation) as the axis, a cone-shaped coverage area is formed by opening up according to the main lobe half-power angle (e.g., 60°), and the effective range (e.g., 3m) is truncated along the axis. Packages within the cone-shaped area are determined to be reachable by the signal. The blocking determination rule is as follows: for any package A within the cone-shaped area, if there is another package B located on the line segment connecting the antenna vertex and package A, and the perpendicular distance between package B and the line segment is less than the cross-sectional radius of package B (calculated based on package volume), then package A is determined to be blocked by package B. When the average physical distance between adjacent packages is less than the effective range of the reader and there is the above mutual blocking between packages, the area where the blocked package is located is determined to be a scanning blind zone. The area that the output signal cannot cover is used as the distribution characteristic of the scanning blind zone, and the ratio of the scanning blind zone area to the total area of ​​the cluster is calculated as the blind zone area ratio.

[0059] Finally, when the proportion of the blind zone area in the blind zone distribution characteristics is not lower than the preset blind zone proportion threshold, the communication timing of the RFID tags is adjusted according to the blind zone distribution characteristics to obtain the RFID wake-up cycle. The preset blind zone proportion threshold is 30%, which means that the area of ​​the blind zone seriously affects the collection of package status. The setting is based on the historical blind zone miss reading test statistics of the target network: several batches (e.g., 100 batches) of blind zone proportion and corresponding tag miss reading rate samples are collected at different times on the sorting line. When the blind zone proportion exceeds 30%, the miss reading rate increases from the average of 2% to more than 8% (the increase exceeds 3 times the standard deviation, which is judged as a significant increase). It can be adjusted within the range of 20% to 40% according to the reader sensitivity. If the blind zone area proportion is lower than the preset blind zone proportion threshold, the default wake-up cycle (e.g., 1 second) is maintained. The correspondence between the blind zone percentage and the wake-up cycle is provided by a pre-configured mapping table, which was established through historical calibration experiments: the wake-up cycle was gradually shortened under different blind zone percentages until the missed read rate fell back to within 2%, and the corresponding cycle values ​​were recorded. For example, a blind zone percentage of 30% to 40% corresponds to 500ms, and above 40% corresponds to 250ms. In this embodiment, the blind zone area percentage is 35%, which is not lower than the preset blind zone percentage threshold of 30%, indicating severe physical obstruction. The mapping table shows that the radio frequency wake-up cycle is 500ms, thereby increasing the activation frequency of the tag to compensate for the impact of the blind zone.

[0060] In step S4, order tracking data of the first abnormal order set is acquired according to the radio frequency wake-up cycle. The actual flow time difference is calculated based on the order tracking data, and compared with the preset standard flow time difference to obtain the time difference deviation value. The tracking delay result is obtained based on the time difference deviation value, including:

[0061] According to the radio frequency wake-up cycle, a radio frequency activation signal is sent to the radio frequency read and write control device to collect real-time dynamic information of packages in the first abnormal order set and obtain the order tracking data;

[0062] By analyzing the order tracking data, the key physical nodes that the package passed through and their corresponding arrival times were separated, resulting in node flow records and status update timestamps.

[0063] The actual flow time difference between adjacent physical nodes is calculated based on the state update timestamp, and the actual flow time difference is compared with the preset standard flow time difference to obtain the time difference deviation value.

[0064] If the time difference deviation exceeds the upper limit of the preset flow deviation range, the starting node of the package is extracted as the stagnant node, and the tracking delay result is generated based on the stagnant node and the time difference deviation.

[0065] First, control parameters are input to the RF read / write control device according to the RF wake-up cycle, with the RF wake-up cycle of 500ms as the control cycle, and an RF activation signal is sent. Real-time dynamic information of packages within the workshop from the first set of abnormal orders is collected, and order tracking data containing the package's unique identifier and historical movement trajectory is output to reconstruct the physical flow process of the abnormal packages.

[0066] Secondly, the order tracking data is analyzed to separate the key physical nodes the package passes through on the sorting line and their corresponding arrival times, resulting in node flow records and status update timestamps. For example, the node flow record of an abnormal package includes the unloading storage area and the main sorting chute. The status update timestamp for arriving at the unloading storage area is 10:05:20, and the status update timestamp for arriving at the main sorting chute is 10:15:30.

[0067] Then, the actual time difference in parcel flow between adjacent physical nodes is calculated based on the status update timestamp, and compared with the preset standard time difference to obtain the time difference deviation value. The actual time difference is the difference in status update timestamps between adjacent physical nodes. The preset standard time difference is 300s, which physically represents the standard time for a parcel to flow normally between adjacent nodes. It is set based on the sorting line design cycle time (an average flow of 5 minutes per node) and takes into account peak load fluctuations, and can be adjusted within the range of 180s to 420s according to the node distance. For example, if the actual time difference is 610s, comparing it with the preset standard time difference of 300s yields a time difference deviation value of 310s.

[0068] Finally, if the time difference deviation exceeds the upper limit of the preset flow deviation range, the starting node of the package is extracted as the stagnant node, and a tracking delay result is generated based on the stagnant node and the time difference deviation. The preset flow deviation range is from 0s to 180s, which physically represents the deviation range of package flow delay within the tolerable range. It is set based on the maximum delay margin tolerated by the logistics scheduling system, and the upper limit of the range can be adjusted according to the business rhythm. The time difference deviation of 310s is compared with the upper limit of the preset flow deviation range of 180s. If it exceeds the upper limit, the starting node (unloading temporary storage area) in the flow record is extracted as the stagnant node. It should be noted that in this embodiment, the stagnant node is defined by the system as the preceding node of the timed flow segment, which is used for scheduling attribution and resource location, rather than directly measuring the physical stop of the package at that node. The output includes the package identification code, the name of the stagnant node, and the deviation duration. If the time difference deviation falls within the preset flow deviation range, no tracking delay result is generated, and the corresponding package is maintained under normal monitoring.

[0069] In step S5, the spatial coordinates of the temperature and humidity sensor network points are obtained based on the tracking delay result, and the sensor distribution density is calculated. The sensor distribution density and the temperature and humidity sensor data are then dimensionality-reduced to obtain a second resource allocation matrix. A preset frequency mapping interval is matched to the second resource allocation matrix to obtain a second acquisition frequency. Environmental condition data of the first abnormal order set is collected according to the second acquisition frequency to obtain a second abnormal order set, including:

[0070] If the time difference deviation value in the tracking delay result exceeds the upper limit of the preset flow deviation range, the spatial coordinates of the temperature and humidity sensor network points within the target logistics network point are obtained through the warehouse IoT platform.

[0071] Using a kernel density estimation algorithm, the number of sensors per unit volume is calculated based on the spatial coordinates of the temperature and humidity sensor network as input, and the sensor distribution density is obtained by calculating the number of sensors per unit volume according to a preset search radius.

[0072] The temperature range and humidity range of the temperature and humidity sensors are collected within a preset time period to obtain the environmental fluctuation amplitude.

[0073] Principal component analysis is used to perform linear transformation and dimensionality reduction on the sensor distribution density and the environmental fluctuation amplitude to obtain the second resource allocation matrix.

[0074] The second resource allocation matrix is ​​processed using a singular value decomposition algorithm to extract the maximum singular value. The maximum singular value is then compared with the preset frequency mapping interval to obtain the second acquisition frequency.

[0075] The temperature and humidity sensor data of the logistics storage area where the packages in the first abnormal order set are located are continuously read at the second collection frequency to obtain an environmental condition data sequence.

[0076] Extract the temperature range and humidity range of the environmental condition data sequence within a preset collection period to form a temperature and humidity fluctuation feature vector.

[0077] Based on the temperature and humidity fluctuation feature vector, calculate the environmental deviation value of each order, divide the environmental deviation value into a preset level range, and assign an environmental abnormality status label to the order that falls into the abnormal level of the preset level range.

[0078] The first abnormal order set is filtered based on the environmental abnormality status identifier to obtain the second abnormal order set.

[0079] First, when the time difference deviation of the tracking delay result of the package in the unloading temporary storage area is determined to be greater than the upper limit of the preset flow deviation range of 180 seconds, the environmental monitoring mechanism is triggered. The spatial coordinates of the temperature and humidity sensor network deployed in and around the unloading temporary storage area within the target logistics point are obtained through the warehouse IoT platform. For example, the position data of the five temperature and humidity sensors numbered 1 to 5 in the three-dimensional coordinate system (sensor number 1 is 17m horizontally, 20m vertically, and 3m high; sensor number 2 is 25m horizontally, 20m vertically, and 3m high; sensor number 3 is 27m horizontally, 20m vertically, and 3m high). The Euclidean distances between the five sensors (20m longitudinally, 3m high, number 4 is 29m horizontally, 20m longitudinally, 3m high, and number 5 is 32m horizontally, 20m longitudinally, 3m high) and the center grid point of the unloading temporary storage area (20m horizontally, 20m vertically, 3m high) are 3m, 5m, 7m, 9m, and 12m, respectively. When the time difference deviation value in the tracking delay result falls within the preset flow deviation range, it indicates that although the parcel flow fluctuates, it does not affect subsequent scheduling and does not trigger the environmental monitoring mechanism. The corresponding parcel can maintain the regular radio frequency monitoring of the first abnormal order set.

[0080] Secondly, a kernel density estimation algorithm is used to perform spatial analysis on the spatial coordinates of the five sensor points. The kernel density estimation algorithm uses a uniform kernel function (weight 1 within the search radius, weight 0 outside the radius). It counts the number of sensors falling within the area covered by the preset search radius, centered on the grid point to be evaluated. This number is then divided by the volume of the area to obtain the number of sensors per unit volume, which is used as the sensor distribution density. The preset search radius is 10m, which physically represents the spatial radius for determining whether sensor coverage is sufficient. This radius is set based on the empirical value of the effective sensing radius of a single temperature and humidity sensor (approximately 8m to 12m), and can be adjusted within the range of 8m to 15m. For example, using the center grid point of the unloading storage area as the evaluation point, the area covered by the preset search radius of 10m is determined according to the measured spatial dimensions of the unloading storage area (10m horizontal × 20m vertical × 3m height, totaling 600m³). All five sensors fall within the search radius (each with a weight of 1), resulting in a sensor distribution density of 5 ÷ 600m³ ≈ 0.0083 sensors / m³, or approximately 0.83 sensors / 100m³.

[0081] Then, the temperature and humidity ranges from the temperature and humidity sensors are collected within a preset time period to obtain the environmental fluctuation amplitude. The preset time period is 1 hour, which physically represents a statistical window for assessing environmental stability. This setting is based on the typical adjustment cycle of the temperature control system and can be adjusted within the range of 30 minutes to 2 hours. The difference between the maximum and minimum temperature values ​​over the past hour is collected as the temperature range (e.g., 2°C), and the difference between the maximum and minimum humidity values ​​is collected as the humidity range (e.g., 5%). These values ​​are combined to obtain the environmental fluctuation amplitude.

[0082] Furthermore, principal component analysis (PCA) is used to linearly transform and reduce the dimensionality of sensor distribution density and environmental fluctuation amplitude to obtain the second resource allocation matrix. To ensure consistency in the direction of each input dimension (larger values ​​indicate greater resource scarcity), a resource gap transformation is performed on the sensor distribution density before input: the difference between the preset baseline coverage density and the actual sensor distribution density is taken as the resource gap component; the preset baseline coverage density is taken as the reciprocal of the effective sensing volume of a single temperature and humidity sensor, calculated based on the spherical volume of approximately 4189 m³ corresponding to an effective sensing radius of 10 m (taking the empirical median of 8 m to 12 m). The baseline coverage density is approximately 1 ÷ 4189 m³ ≈ 0.024 sensors / m³, or approximately 2.4 sensors / 100 m³; when the actual sensor distribution density is lower than the baseline coverage density, the resource gap component is positive, indicating insufficient coverage, and the larger the difference, the more insufficient the coverage and the more scarce the resources; when the actual distribution density reaches or exceeds the baseline coverage density, the resource gap component is 0, indicating sufficient coverage. Principal component analysis (PCA) uses resource gap components, temperature range, and humidity range as input feature vectors (3D). Feature vectors from multiple sampling times are combined to form a sample matrix. The eigenvalues ​​and eigenvectors of the covariance matrix (3×3) of the sample matrix are calculated. The eigenvectors corresponding to the first two principal components are selected to construct a 3×2 transformation matrix. The criterion for retaining two principal components is that the cumulative variance contribution rate is not less than 85%. This criterion is verified by historical samples—in the past three months of normal operation samples at the target site, the cumulative variance contribution rate of the first two principal components is stable between 85% and 92%. Retaining the third principal component improves the resource status discrimination by less than 2%, so two dimensions are fixed. The input feature vectors are first standardized using z-scores (mean and standard deviation are statistically derived from the same historical normal sample set). The transformation matrix is ​​also pre-calculated using this historical normal sample set and re-estimated and updated quarterly using the latest samples. The standardized input feature vectors are projected through the transformation matrix to output a 1×2 second resource allocation matrix, which characterizes the comprehensive status of environmental monitoring resources in the area. Principal component analysis is used here instead of direct weighting because the sensor distribution density, temperature range, and humidity range have different dimensions, and their influence on the sampling frequency also varies with the type of goods. Direct weighting would lead to the results being dominated by the temperature range, which has a larger dimension. Principal component analysis automatically determines the weight of each dimension based on the variance contribution, avoiding a single dimension dominating the dimensionality reduction results and ensuring that the resource stress assessment is adaptable to different types of goods.

[0083] Then, the 1×2 second resource allocation matrix is ​​processed using a singular value decomposition (SVD) algorithm, decomposing it into the product of a left singular matrix, a singular value matrix, and a right singular matrix. The maximum singular value in the singular value matrix is ​​extracted. The maximum singular value is then input into a preset frequency mapping interval for comparison. The preset frequency mapping interval includes a low-frequency interval (sampling period 10s to 30s, corresponding to a sampling frequency of 0.033Hz to 0.1Hz), a mid-frequency interval (sampling period 5s to 9s), and a high-frequency interval (sampling period 1s to 4s). It should be noted that although the first sampling frequency and the second sampling frequency in this embodiment are both named "sampling frequency," they belong to different sampling devices: the first sampling frequency refers to the radio frequency operating frequency (carrier frequency, unit: MHz, e.g., 920MHz to 925MHz) of the RFID tag, used to reduce read collisions by switching channels; the second sampling frequency refers to the sampling frequency of the temperature and humidity sensor, expressed in sampling period (unit: s) and can be converted to Hz, for example, a sampling period of 15s corresponds to 0.067Hz, 5s corresponds to 0.067Hz, and 1s corresponds to 0.033Hz. The sampling frequency should be 0.2Hz and 2s corresponds to 0.5Hz. The shorter the sampling period, the higher the sampling frequency. The physical meaning, unit and equipment of the two types of sampling frequencies are different and should not be mixed. The setting of each interval is based on the historical temperature and humidity monitoring data of the network point, the sensor response period and the statistical calibration of environmental abnormal samples: the 95th percentile of the temperature and humidity fluctuation range under historical normal working conditions corresponds to the sampling period (10s to 30s) as the normal range, the sampling period (5s to 9s) corresponds to the historical events of concern (temperature exceeding the limit by less than 2℃) as the attention range, and the sampling period (1s to 4s) corresponds to the historical emergency events (temperature exceeding the limit by more than 2℃ or environmental abnormal work orders) as the emergency range. The boundary of the three sampling periods is determined by the quantile of the historical sample distribution and can be readjusted according to the type of goods. The mapping relationship between the maximum singular value and frequency is achieved through a lookup table. The mapping boundary is statistically determined by the distribution of the maximum singular values ​​of historical environmental anomaly samples: the upper quartile of the singular values ​​of historical normal operating condition samples (0.5) is used as the low-to-mid frequency boundary, and the lower quartile of the singular values ​​of historical anomaly samples (1.0) is used as the mid-to-high frequency boundary. When the maximum singular value is less than 0.5 (corresponding to low distribution density and small fluctuation amplitude), it matches the low-frequency range, and the output acquisition period is 15s (approximately 0.067Hz); when the maximum singular value is between 0.5 and 1.0 (corresponding to medium distribution density or small fluctuation amplitude), it matches the low-frequency range, and the output acquisition period is 15s (approximately 0.067Hz). When the amplitude of fluctuation is moderate, the mid-frequency range is matched, and the output acquisition period is 5s (0.2Hz); when the maximum singular value is greater than 1.0 (corresponding to high distribution density and large fluctuation amplitude), the high-frequency range is matched, and the output acquisition period is 2s (0.5Hz). That is, the larger the maximum singular value, the greater the resource gap and fluctuation, the more scarce the resources, and the shorter the acquisition period. This monotonic mapping relationship is obtained by calibration of historical samples (on historical normal and abnormal working condition samples, the maximum singular value is significantly positively correlated with the manually set monitoring level, and the correlation coefficient is not less than 0.8).In this embodiment, the maximum singular value of the unloading temporary storage area is 0.8 (this value is obtained by projecting the PCA transformation matrix and decomposing it by SVD; since the input feature vector is first standardized by z-score, the maximum singular value reflects the relative position of the resource shortage at this sampling moment in the historical sample distribution, rather than the absolute magnitude of the input value - in this embodiment, the resource gap of 1.57 is in the middle of the historical samples, and the corresponding maximum singular value of 0.8 falls into the mid-frequency range), matching the mid-frequency range, and the output acquisition period is 5s.

[0084] Then, data reading operations are performed at the second collection frequency. For fresh produce packages involved in the first abnormal order set, the temperature and humidity sensor data of the unloading temporary storage area where the package is located is continuously read at a cycle of 5 seconds. By continuously collecting environmental data in the past 1 minute, an environmental condition data sequence containing 12 discrete data points is obtained, such as the temperature sequence fluctuating between 4℃ and 6℃ and the humidity sequence fluctuating between 60% and 65%.

[0085] Furthermore, the temperature and humidity ranges of the environmental condition data sequence within the preset collection period are extracted to form a temperature and humidity fluctuation feature vector. The preset collection period is 1 minute, which is set based on the following: the response cycle of the temperature and humidity sensor (approximately 5 to 10 seconds) ensures that more than 6 effective sampling points can be obtained within 1 minute; it is more than twice the minimum time scale (approximately 30 seconds) for significant changes in the ambient temperature of the target fresh goods; and it has been verified by statistical analysis of historical environmental anomaly samples—the complete capture rate of historical anomaly events within the 1-minute window is not less than 95%, and it can be adjusted according to the type of goods within the range of 30 seconds to 5 minutes; the temperature range is the difference between the maximum and minimum temperature values ​​within this period (e.g., 6℃ minus 4℃ equals 2℃), and the humidity range is the difference between the maximum and minimum humidity values ​​(e.g., 65% minus 60% equals 5%). The above two statistics are combined into a temperature and humidity fluctuation feature vector, for example, [2℃, 5%].

[0086] Then, the environmental deviation value for each order is calculated based on the temperature and humidity fluctuation characteristic vector. The environmental deviation value is the ratio of the temperature range to the preset temperature baseline range, plus the ratio of the humidity range to the preset humidity baseline range, and then weighted and summed according to weighting coefficients. The weighting coefficients α are 0.6 (temperature term) and 0.4 (humidity term), and are set based on historical temperature and humidity anomaly sample statistics: a retrospective analysis of environmental anomaly work orders from the target network over the past 6 months shows that anomalies caused by exceeding temperature limits accounted for approximately 60%, and those caused by exceeding humidity limits accounted for approximately 40%. The weights are determined based on these proportions and can be readjusted according to the type of goods. The preset temperature baseline range is 1℃, which physically means the allowable range of temperature fluctuation within 1 minute under normal operating conditions in cold chain warehousing, and is set based on the measured average of steady-state fluctuations in the cold chain warehousing temperature control system; the preset humidity baseline range is 2%, which physically means the allowable range of humidity fluctuations under normal operating conditions, and is set based on the measured value of the humidifier response characteristics; both baselines can be adjusted within ±20% of the baseline value according to the type of goods. For example, when the temperature range is 3℃ and the humidity range is 5%, the environmental deviation value is 0.6×(3℃ / 1℃)+0.4×(5% / 2%) =0.6×3+0.4×2.5=1.8+1.0=2.8 (dimensionless). The environmental deviation value is divided into three levels according to a preset level range, which includes normal level (deviation value less than 1.0), attention level (1.0 to 2.0), and abnormal level (greater than 2.0). The level boundary is determined by the deviation distribution of historical abnormal samples: the 95th quantile (1.0) of the deviation value of historical normal working condition samples is used as the boundary between normal and attention, and the 5th quantile (2.0) of the deviation value corresponding to historical abnormal work orders is used as the boundary between attention and abnormal, so that the two levels correspond to statistically significant deviation and significant abnormality, respectively. Both benchmarks can be adjusted within ±20% of the benchmark value according to the type of goods. The adjustment range of 20% corresponds to the dispersion range of benchmark values ​​for different types of goods in historical statistics (coefficient of variation of about 15% to 20%). In this embodiment, if the environmental deviation value of 2.8 is greater than 2.0, it falls into the abnormal level within the preset level range, and the order is given an environmental abnormality status mark; if the environmental deviation value falls into the normal level or the attention level, no environmental abnormality status mark is given.

[0087] Finally, the first set of abnormal orders is filtered based on the environmental anomaly status identifier to obtain the second set of abnormal orders. Packages with assigned environmental anomaly status identifiers (e.g., 8 abnormal fresh food packages) are filtered out from the first set of abnormal orders, and the second set of abnormal orders is output to guide the dispatch center to prioritize handling environmentally sensitive anomalies and initiate temperature control remedial measures.

[0088] In summary, this invention discloses a method for order anomaly response based on logistics status. The method includes: acquiring parcel and vehicle data from network points and stitching them together to obtain a first node load characteristic; obtaining a second node load characteristic through random forest classification; adjusting the RFID tag acquisition frequency accordingly to obtain a first acquisition frequency and acquiring a first set of abnormal orders; spatially clustering the first set of abnormal orders to obtain initial clusters; calculating the stacking density based on their boundary contours and area; analyzing the average physical distance and scanning blind zone characteristics of the initial clusters with high stacking density to obtain the RFID wake-up period; comparing this with order tracking data to obtain the tracking delay result; and using this to reduce the dimensionality of the temperature and humidity sensor distribution density and acquired data to obtain a second resource allocation matrix; matching a preset frequency mapping interval to obtain a second acquisition frequency and acquiring a second set of abnormal orders. This invention can solve the problem of delayed anomaly response caused by a mismatch between fixed acquisition frequency and node load.

[0089] The second embodiment of the present invention provides an order anomaly response system based on logistics status, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described above.

[0090] It should be noted that the order anomaly response system based on logistics status provided in this embodiment of the invention is used to execute all the process steps of the order anomaly response method based on logistics status in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0091] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for responding to order anomalies based on logistics status, characterized in that, include: Obtain parcel accumulation data and vehicle entry and exit data of the target logistics network point, concatenate the parcel accumulation data and vehicle entry and exit data to obtain the first node load feature, and use a pre-constructed random forest to classify the first node load feature to obtain the second node load feature. The sensory resource distribution data of the RFID tag is obtained based on the load characteristics of the second node as a first resource allocation matrix. The acquisition frequency of the RFID tag is adjusted according to the first resource allocation matrix to obtain a first acquisition frequency. The status data of the unsorted package node is collected according to the first acquisition frequency to classify and obtain a first abnormal order set. Spatial clustering is performed on the first abnormal order set to obtain initial clusters. The stacking density is calculated based on the boundary contour and area of ​​the initial clusters. The average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density are analyzed to obtain the radio frequency wake-up cycle. The order tracking data of the first abnormal order set is obtained according to the radio frequency wake-up cycle. The actual flow time difference is calculated based on the order tracking data. The time difference deviation value is obtained by comparing it with the preset standard flow time difference. The tracking delay result is obtained based on the time difference deviation value. Based on the tracking delay result, obtain the spatial coordinates of the temperature and humidity sensor network and calculate the sensor distribution density. Reduce the dimensionality of the sensor distribution density and the temperature and humidity sensor data to obtain a second resource allocation matrix. Match the second resource allocation matrix with a preset frequency mapping interval to obtain a second acquisition frequency. Collect the environmental condition data of the first abnormal order set according to the second acquisition frequency to obtain a second abnormal order set.

2. The order anomaly response method based on logistics status according to claim 1, characterized in that, The process of acquiring parcel accumulation data and vehicle entry / exit data of the target logistics network point, and concatenating the parcel accumulation data and vehicle entry / exit data to obtain the first node load characteristics includes: Obtain the original parcel entry time record and vehicle arrival time record, and perform timestamp alignment processing according to the time window to obtain aligned parcel stacking data and aligned vehicle entry and exit data; The warehouse capacity saturation of the target logistics network is calculated based on the aligned parcel stacking data and the maximum design capacity of the target logistics network. The network capacity scheduling tension is determined based on the truck queuing waiting time and the number of currently available trucks in the aligned vehicle entry and exit data. The saturation of the storage capacity of the network points and the tension of the transportation capacity scheduling of the network points are concatenated into a two-dimensional vector to obtain the load characteristics of the first node.

3. The order anomaly response method based on logistics status according to claim 1, characterized in that, The step of adjusting the acquisition frequency of the RFID tag according to the first resource allocation matrix to obtain the first acquisition frequency includes: The frequency band occupancy distribution of the RFID tags is extracted based on the first resource allocation matrix to obtain the initial RFID coverage network; Based on the initial radio frequency coverage network and the time-series position coordinates of the RFID tag, smoothing and trajectory tracking are performed on the time-series position coordinates to extract the movement trajectory of the RFID tag. The reading collision probability is obtained by calculating the ratio of the number of time slots that collide within the reading cycle to the total number of time slots based on the movement trajectory. The frequency adjustment step size is determined based on the mapping relationship between the read collision probability and the preset adjustment amount. The larger the read collision probability, the larger the frequency adjustment step size. The current reference acquisition frequency of the RFID tag is updated according to the frequency adjustment step size to obtain the first acquisition frequency.

4. The order anomaly response method based on logistics status according to claim 1, characterized in that, The step involves collecting node status data of unsorted packages at the first collection frequency, and classifying the node status data to obtain a first abnormal order set, including: The RF reader antenna is controlled to transmit a continuous wave signal according to the first acquisition frequency to obtain the RF echo signal reflected by the unsorted package; The received signal strength indication of the radio frequency echo signal is extracted as an attenuation feature, and the arrival time difference between signal transmission and reception is extracted. The attenuation feature and the arrival time difference are smoothed using a Kalman filter algorithm to obtain smoothed position coordinates. The actual moving speed of the unsorted package is calculated based on the difference between the smoothed position coordinates of adjacent sampling times and the sampling time interval. Timing starts when the actual moving speed approaches zero to obtain the dwell time sequence of the unsorted package. If the dwell time sequence exceeds the preset timeout warning threshold, an abnormal status identifier is generated; Node status data is generated based on the status anomaly identifier and the logistics waybill number and volume information of the unsorted packages. The node status data is then filtered and classified based on the status anomaly identifier to obtain the first set of abnormal orders.

5. The order anomaly response method based on logistics status according to claim 1, characterized in that, The step of spatially clustering the first abnormal order set to obtain initial clusters, and calculating the stacking density based on the boundary contours and area of ​​the initial clusters, includes: Obtain the indoor spatial coordinates and shelf level of each package in the first abnormal order set, and construct a three-dimensional location matrix; Based on the three-dimensional position matrix, the density peak clustering algorithm is used to calculate the local density of each package node and the relative distance to higher density nodes. Nodes whose local density and relative distance both satisfy the preset clustering parameters are taken as cluster centers, and the remaining package nodes are assigned to the nearest cluster center to obtain multiple initial clusters. Based on the node coordinates of the initial clusters, the convex hull algorithm is used to extract the line connecting the outermost wrapping nodes of each initial cluster as the boundary contour, and the area of ​​the region enclosed by the boundary contour is calculated. The total number of wrapping nodes in the initial clusters is divided by the area of ​​the region to obtain the stacking density.

6. The order anomaly response method based on logistics status according to claim 1, characterized in that, The process of analyzing the average physical distance and scanning blind zone distribution characteristics of the initial clusters with high stacking density to obtain the radio frequency wake-up cycle includes: The initial clusters whose stacking density meets the preset congestion conditions are taken as high-density congestion areas. The average physical distance between adjacent packages in the high-density congestion areas is calculated, and the distribution characteristics of scanning blind spots that cannot be covered by the signal are extracted by combining the radiation angle model of the radio frequency reader. When the proportion of the blind area of ​​the scanning blind area distribution feature is not lower than the preset blind area proportion threshold, the communication timing of the radio frequency identification tag is adjusted according to the scanning blind area distribution feature to obtain the radio frequency wake-up period.

7. The order anomaly response method based on logistics status according to claim 1, characterized in that, The step of calculating the actual turnaround time difference based on the order tracking data, comparing it with a preset standard turnaround time difference to obtain a time difference deviation value, and obtaining a tracking delay result based on the time difference deviation value includes: According to the radio frequency wake-up cycle, a radio frequency activation signal is sent to the radio frequency read and write control device to collect real-time dynamic information of packages in the first abnormal order set and obtain the order tracking data; By analyzing the order tracking data, the key physical nodes that the package passed through and their corresponding arrival times were separated, resulting in node flow records and status update timestamps. The actual flow time difference between adjacent physical nodes is calculated based on the state update timestamp, and the actual flow time difference is compared with the preset standard flow time difference to obtain the time difference deviation value. If the time difference deviation exceeds the upper limit of the preset flow deviation range, the starting node of the package is extracted as the stagnant node, and the tracking delay result is generated based on the stagnant node and the time difference deviation.

8. The order anomaly response method based on logistics status according to claim 1, characterized in that, The process of obtaining the spatial coordinates of the temperature and humidity sensor network points based on the tracking delay result and calculating the sensor distribution density, reducing the dimensionality of the sensor distribution density and the temperature and humidity sensor data to obtain a second resource allocation matrix, and matching the second resource allocation matrix with a preset frequency mapping interval to obtain a second acquisition frequency includes: If the time difference deviation value in the tracking delay result exceeds the upper limit of the preset flow deviation range, the spatial coordinates of the temperature and humidity sensor network points within the target logistics network point are obtained through the warehouse IoT platform. Using a kernel density estimation algorithm, the number of sensors per unit volume is calculated based on the spatial coordinates of the temperature and humidity sensor network as input, and the sensor distribution density is obtained by calculating the number of sensors per unit volume according to a preset search radius. The temperature and humidity ranges of the temperature and humidity sensors are collected within a preset time period to obtain the environmental fluctuation range. Principal component analysis is used to perform linear transformation and dimensionality reduction on the sensor distribution density and the environmental fluctuation amplitude to obtain the second resource allocation matrix. The second resource allocation matrix is ​​processed using a singular value decomposition algorithm to extract the maximum singular value. The maximum singular value is then compared with the preset frequency mapping interval to obtain the second acquisition frequency.

9. The order anomaly response method based on logistics status according to claim 1, characterized in that, The step of collecting environmental condition data of the first abnormal order set at the second collection frequency to obtain the second abnormal order set includes: The temperature and humidity sensor data of the logistics storage area where the packages in the first abnormal order set are located are continuously read at the second collection frequency to obtain an environmental condition data sequence. Extract the temperature range and humidity range of the environmental condition data sequence within a preset collection period to form a temperature and humidity fluctuation feature vector. Based on the temperature and humidity fluctuation feature vector, calculate the environmental deviation value of each order, divide the environmental deviation value into a preset level range, and assign an environmental abnormality status label to the order that falls into the abnormal level of the preset level range. The first abnormal order set is filtered based on the environmental abnormality status identifier to obtain the second abnormal order set.

10. An order anomaly response system based on logistics status, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 9.

Images