An enterprise intelligent early warning method based on an artificial intelligence model

Abstract

The application provides an enterprise intelligent early warning method based on an artificial intelligence model, comprising: extracting spatio-temporal trajectory features of a target product in combination with a circulation data set, identifying circulation time intervals, mapping geographical position information and time intervals by using a geographic information system, and generating a spatio-temporal feature matrix of product circulation; combining abnormal geographical position deviation values, dispersion indexes corresponding to dispersion degrees, and product batch sizes, arrival times and network sizes to grade product circulation abnormalities; sending early warning signals to a management platform for different grades of abnormal risks, determining early warning triggering times and objects; generating abnormal intervention adjustment schemes according to early warning signal triggering conditions, monitoring product circulation trajectories, and obtaining trajectory change data after intervention; updating the spatio-temporal feature matrix according to the trajectory change data, re-inputting the data into a deviation detection and dispersion degree analysis process through a cyclic feedback mechanism, and obtaining new abnormal identification results.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to an enterprise intelligent early warning method based on an artificial intelligence model. BACKGROUND

[0002] The product traceability system is an important pillar for safeguarding market order and consumer rights and interests, and its core lies in monitoring the circulation path of products through technical means to ensure that the source of goods is traceable and the destination is traceable. With the continuous expansion of the market size and the complication of the circulation link, how to effectively identify abnormal circulation and timely realize enterprise intelligent early warning through artificial intelligence model has become a key field for maintaining market fairness and cracking down on counterfeiting and shoddy. However, many current traceability methods still have obvious deficiencies, and generally rely on static data records, lack of deep analysis ability of dynamic circulation track, especially in the face of complex market environment, it is difficult to quickly capture abnormal phenomena and respond. This makes some potential risks, such as counterfeit products flowing into the market or the circulation path deviating from the expected, often ignored. In-depth, the core challenge faced by this field is first how to extract meaningful spatio-temporal trajectory features from massive historical circulation data. Products involve multiple geographic locations and time nodes in the circulation link, and the data is huge and unevenly distributed, and it is difficult to accurately describe its regularity by relying solely on traditional analysis methods. This problem further leads to another technical difficulty, that is, when identifying abnormal circulation path, the system often cannot accurately judge the deviation of geographic location from the expected sales area, or the abnormal change of the dispersion degree of the same batch of products. This progressive difficulty from data feature extraction to abnormality identification makes the traceability system appear helpless when facing sudden or hidden problems. Therefore, how to build an intelligent abnormality identification mechanism based on historical circulation data and realize hierarchical early warning for key indicators such as geographic location deviation and dispersion degree anomaly has become a key problem that needs to be overcome. SUMMARY

[0003] The present application provides an enterprise intelligent early warning method based on an artificial intelligence model, mainly including:

[0004] By collecting and cleaning the historical circulation data, the product flow record of each network point is identified, and the time stamp, geographic location information, product batch size, arrival time and network point size involved are extracted, and the structured circulation data set is arranged;

[0005] The spatio-temporal trajectory features of the target product are extracted in combination with the circulation data set, the circulation time interval is identified, the geographic location information and time interval are mapped by using the geographic information system, and the spatio-temporal feature matrix of product circulation is generated;

[0006] The product flow path nodes and node distances are obtained by clustering analysis on the space-time feature matrix, and the deviation percentage is obtained by comparing the nodes and node distances of the preset standard sales area path one by one, and the abnormal geographic position deviation value is determined according to the deviation percentage;

[0007] The distribution proportion of the same batch of products in different sales areas, different level network points and different time windows is obtained by statistical analysis on the space-time feature matrix, the Gini coefficient is calculated to obtain the dispersion index of each dimension, and the dispersion degree is determined by the maximum dispersion index;

[0008] The abnormal product circulation is classified according to the abnormal geographic position deviation value, the dispersion index corresponding to the dispersion degree, and the product batch size, arrival time and network point size;

[0009] According to the abnormal risk of different grades, an early warning signal is sent to the management platform, and the early warning trigger time and object are determined;

[0010] According to the triggering condition of the early warning signal, an abnormal intervention adjustment scheme is generated, the product circulation track is monitored, and the track change data after intervention is obtained;

[0011] The space-time feature matrix is updated according to the track change data, the data is re-input into the deviation detection and dispersion degree analysis process through the cyclic feedback mechanism, and the new abnormal identification result is obtained.

[0012] Further, by collecting and cleaning historical circulation data, identifying product circulation records of each network point, extracting time stamp, geographic location information, product batch size, arrival time and network point size, and arranging structured circulation data sets, including: according to historical circulation records, data cleaning is carried out, regular expression matching is used to extract timestamp information, and standard format network product circulation time series data set is obtained through data standardization processing. For network product circulation time series data set, sliding time window method is used to segment time series data according to 24 hours as window length, from which product circulation interval matrix and batch size matrix between network points are extracted. According to product circulation interval matrix and batch size matrix, combined with regional distribution route distance data, dynamic programming algorithm is used to calculate network storage capacity threshold vector with the objective function of minimizing the total distribution cost. For network storage capacity threshold vector and real-time logistics tracking data, the inventory turnover rate index of each network point is obtained through the inventory turnover rate calculation formula, so as to construct the network inventory capacity saturation curve. According to the network inventory capacity saturation curve and the congestion degree data of peak period, the network is clustered by using the density clustering algorithm with the inventory turnover efficiency as the feature, and the clustering distance threshold is set as the standard deviation of the average inventory turnover rate, so as to obtain the network clustering result. For network clustering result and product circulation interval matrix, network product circulation relationship graph is generated by graph network construction algorithm, node weight is determined by inventory turnover rate, and edge weight is determined by circulation interval.

[0013] Further, combined with the circulation data set, the spatio-temporal trajectory features of the target product are extracted, the circulation time interval is identified, the geographic information system is used to map the geographic location information and the time interval, and the spatio-temporal feature matrix of product circulation is generated, including: according to the product location update interval and the geographic coordinate collection frequency, the space is divided into grid units with a side length of 500 meters by using the quadtree index, and the product trajectory point sequence data is obtained by time stamp marking. For product trajectory point sequence data, the distance and time difference between trajectory points are extracted by using geographic information encoder, from which product circulation time interval matrix and spatial displacement matrix are calculated. According to product circulation time interval matrix and spatial displacement matrix, the optimal path set between stations is calculated by using the shortest path algorithm, and the path weight is determined by distance and traffic period limit. For the optimal path set between stations, a three-dimensional path graph is constructed by using the space-time cube method, the horizontal coordinate is the longitude, the vertical coordinate is the latitude, and the height axis is the timestamp, and the path space-time distribution data is obtained. According to the path space-time distribution data and the path congestion delay coefficient, the traffic speed distribution map of each region at different time periods is obtained by density calculation, and then the regional distribution network density heat map is identified. For the regional distribution network density heat map, the spectral clustering method is used to divide the region according to the traffic speed and the residence time, and the spatio-temporal feature mapping matrix of product circulation is obtained.

[0014] Further, by clustering analysis on the space-time feature matrix, the product flow path node and node distance are obtained, and the deviation percentage is obtained by comparing the node and node distance of the preset standard sales area path one by one, and the abnormal geographic position deviation value is determined according to the deviation percentage, including: according to the space-time feature matrix, using hierarchical clustering algorithm with node distance and time interval as features to group the distribution path, and obtaining the node sequence data of the actual distribution path. According to the actual distribution path node sequence data, the standard distribution area is divided according to the fixed side length by the space gridding method, and the standard distribution road network topology graph is constructed from the regional distribution point density data. According to the standard distribution road network topology graph, the minimum spanning tree algorithm is used to calculate the standard distribution path node sequence, and the node weight is determined by the residence time and the traffic restriction degree. According to the actual distribution path node sequence and the standard distribution path node sequence, the spatial distance difference value between the corresponding nodes is calculated by the coordinate projection method. According to the spatial distance difference value between the nodes, the normalized processing is used to obtain the node position deviation proportion value, and then the node deviation degree distribution matrix is generated. According to the node deviation degree distribution matrix, the abnormal judgment is made by setting the reference deviation threshold, and the abnormal geographic position deviation value sequence is identified.

[0015] Further, by statistical analysis on the space-time feature matrix, the distribution proportion of products in the same batch in different sales areas, different level nodes and different time windows is obtained, and the Gini coefficient is used to calculate the dispersion index of each dimension, and the dispersion degree is determined by the maximum dispersion index, including: according to the space-time feature matrix, the hierarchical sampling method is used to divide the node level according to the storage capacity of the node, and the sampling proportion is determined based on the number of nodes at each level, to obtain the node hierarchical structure data. According to the node hierarchical structure data, the hierarchical statistical calculator is used to calculate the product batch distribution value of each level node, and the node level proportion data set is obtained by combining the inventory turnover rate. According to the node level proportion data set, the region division tool is used to divide the space according to the sales area boundary, and the sales area proportion data set is obtained from the regional commercial activity data. According to the sales area proportion data set, the time window division tool is used to divide the time period according to the sales peak law, and the time dimension proportion data set is obtained. According to the time dimension proportion data set, the data perspective tool is used to construct a three-dimensional joint distribution table, and a multi-dimensional distribution matrix is obtained from the regional distribution resource data. According to the multi-dimensional distribution matrix, the Gini coefficient calculator is used to calculate the node level dimension dispersion index value, the sales area dimension dispersion index value and the time dimension dispersion index value respectively. According to the dispersion index values of the three dimensions, the maximum dispersion index is extracted by the extreme value selector, and the product dispersion degree identification value is obtained from the sales area overlap degree data.

[0016] Further, combined with the abnormal geographical position deviation value, the dispersion index corresponding to the dispersion degree, and the product batch size, the arrival time and the network point size, the product circulation anomaly is classified, including: according to the abnormal geographical position deviation value, the spatial feature extractor is adopted to construct the path anomaly index according to the offset direction and distance, and the position anomaly feature set is obtained from the node-to-node transportation time data. For the position anomaly feature set, the product batch size and the network storage capacity are mapped and calculated by the multi-dimensional data integrator to obtain the batch size conformity, and the scale anomaly index set is obtained. According to the scale anomaly index set, the time window analyzer is adopted to calculate the batch arrival delay time and the network turnover speed, and the time anomaly index set is obtained from the network inventory saturation data. For the time anomaly index set, the network feature extractor is used to calculate the correlation degree of the network size and the distribution efficiency, and the network anomaly index set is obtained from the regional circulation anomaly rate data. According to the position anomaly feature set, the scale anomaly index set, the time anomaly index set and the network anomaly index set, the k-means clustering method is adopted to construct a four-dimensional feature space. For the four-dimensional feature space data, the abnormal degree quantifier is used to calculate the abnormal degree value of each dimension, and the abnormal feature weight matrix is obtained. According to the abnormal feature weight matrix, the hierarchical standard generator is used to divide the abnormal level according to the abnormal degree value, and the product circulation anomaly classification identifier is obtained.

[0017] Further, for different levels of abnormal risk, a warning signal is sent to the management platform to determine the warning trigger time and object, including: according to different levels of abnormal risk data, the risk quantization calculator is used to generate a risk level matrix, and the risk quantization index includes three dimensions of abnormal degree, duration and influence range. For the risk level matrix, the warning rule generator is used to set the warning trigger threshold of each level of risk, and the threshold parameter includes risk accumulation value, propagation speed and correlation strength. According to the warning trigger threshold, the warning classification mapper is used to classify the network warning objects, and the hierarchical warning object list is obtained from the network response time limit data. For the hierarchical warning object list, the decision tree algorithm is used to calculate the warning priority index, and the decision node includes the risk level, the warning timeliness and the network level. According to the warning priority index, the time window divider is used to set the warning information push frequency, and the warning time trigger table is obtained from the warning duration data. For the warning time trigger table, the message queue distributor is used to build a warning information package, and the warning information includes the risk level identifier, the warning object number and the trigger timestamp. According to the warning information package, the multi-channel pusher is used to send a hierarchical warning signal to the management platform to realize the differentiated push of different levels of warning information.

[0018] Further, according to the early warning signal triggering condition, an abnormal intervention adjustment scheme is generated, the product circulation track is monitored, and track change data after intervention is obtained, including: according to the early warning signal triggering record, an inspection node priority sequence is generated using an association rule extractor, and the node priority is calculated based on the abnormal occurrence frequency and the disposal timeliness. For the inspection node priority sequence, the resource scheduler is used to calculate the inspection resource configuration scheme of each node, and the sampling frequency reference value is obtained from the inspection point distribution data. According to the sampling frequency reference value, a random forest algorithm is used to calculate the sampling time window, and the decision features include abnormal level, disposal speed and resource density. For the sampling time window, a real-time stream processor is used to build a track monitoring scheme, and the monitoring indicators include position deviation, time delay and path change. According to the track monitoring scheme, a data stream collector is used to record product position data at a preset sampling interval to obtain real-time track data stream. For the real-time track data stream, a track comparator is used to calculate the track change value before and after intervention to obtain the track offset metric value. According to the track offset metric value, a feedback regulator is used to dynamically update the inspection node configuration and sampling frequency, and the closed-loop optimization of the intervention scheme is realized.

[0019] Further, the severity and triggering frequency of the early warning signal are identified, the frequently occurring abnormal distributors are identified according to the triggering frequency of the early warning signal, the product circulation abnormal type library is matched to generate a targeted intervention scheme, the suspicious distributors are set with temporary sales limits and circulation time window restrictions in specific areas, the distributors with serious violations are implemented with temporary qualification freezing, and temporary traceability identifiers are added to suspicious product batches for key monitoring, forming an abnormal intervention execution list for closed-loop control, including: according to the early warning signal density and triggering time interval, an abnormality detector is used to calculate a distributor abnormality index, and the abnormality index is calculated by weighting the early warning frequency and severity to obtain a high-frequency abnormal distributor list. For the high-frequency abnormal distributor list, a rule matcher is used to classify against the abnormal behavior feature library, and a violation behavior level distribution table is obtained from the distributor credit level data. According to the violation behavior level distribution table, a sales limit calculator is used to set regional limits and time period restrictions, and limit constraint parameter sets are obtained from regional transaction records. For the limit constraint parameter set, a qualification evaluator is used to generate a qualification freeze period, and a hierarchical control measure set is obtained from the violation behavior type. According to the hierarchical control measure set, a traceability identifier generator is used to add temporary traceability codes to suspicious product batches, and a key monitoring batch list is obtained from batch circulation records. For the key monitoring batch list, an execution tracker is used to record the implementation of the control measures, and a measure implementation progress table is obtained from the execution feedback data. According to the measure implementation progress table, a closed-loop monitor is used to calculate the control effect indicators, and an abnormal intervention execution list is obtained from the abnormal rectification records.

[0020] Further, the space-time feature matrix is updated according to the trajectory change data, the data is re-input into the deviation detection and dispersion degree analysis process through a loop feedback mechanism, and a new abnormality recognition result is obtained, including: according to the trajectory change data, a trajectory feature point is extracted according to a preset sampling interval by using a data synchronizer, and a trajectory feature sequence is obtained from trajectory sampling density data. According to the trajectory feature sequence, a feature matrix first version is constructed by a space-time feature extractor, and the matrix dimension includes position coordinates, time stamps, speed and direction. According to the feature matrix first version, a position deviation value is calculated by using a recurrent neural network, and a position and time double-dimensional feature is set in the network input layer. According to the position deviation value, a regional dispersion index is generated by using a dispersion calculator, and the calculation index includes spatial dispersion and temporal dispersion. According to the regional dispersion index, a feature matrix second version is generated by using a feature updater, and the update content includes new deviation values and dispersion indexes. According to the feature matrix second version, the updated data is input into the abnormality detection process by using a data backflow device, and a new abnormality feature vector is generated. According to the new abnormality feature vector, an abnormality recognition parameter is updated by using a feedback optimizer, and the dynamic update of the abnormality recognition result is completed.

[0021] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0022] The present application discloses an enterprise intelligent early warning method based on an artificial intelligence model, which collects historical circulation data, extracts product space-time trajectory features, constructs a space-time feature matrix, compares a preset standard sales area path, calculates abnormal geographic position deviation values and dispersion degrees, and combines multi-dimensional features such as batch size and arrival time to divide abnormality levels. For different levels of abnormal risks, the present application automatically pushes an early warning signal, generates an intervention adjustment scheme, and monitors trajectory changes using real-time data stream processing technology. Through a loop feedback mechanism, the present application continuously updates the feature matrix, realizes dynamic identification and early warning of product circulation abnormalities, effectively improves the accuracy and timeliness of product circulation supervision, and provides strong support for preventing illegal circulation and market risks. BRIEF DESCRIPTION OF DRAWINGS

[0023] Fig. 1 A flowchart of an enterprise intelligent early warning method based on an artificial intelligence model of the present application.

[0024] Fig. 2 A schematic diagram of an enterprise intelligent early warning method based on an artificial intelligence model of the present application. DETAILED DESCRIPTION

[0025] The technical scheme in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application.

[0026] As Figs. 1-2The embodiment of the enterprise intelligent early warning method based on an artificial intelligence model can specifically include the following steps:

[0027] S101 Collect and process historical circulation data, identify product circulation records of each network point, extract time stamp, geographic location, product batch size, arrival time and network point size, and generate a structured circulation data set.

[0028] S1011 extracts time stamp information from historical circulation records, uses a regular expression to standardize the format, and generates a uniform format of network point circulation time sequence data set. In the embodiment of the application, the time stamp information covers the year, month, day, hour, minute and second format, and is standardized to the format of 2024-05-21 14:30:00. For abnormal time stamps, such as future time or records exceeding a reasonable range, a preset time interval is used for filtering to ensure data validity.

[0029] S1012 uses a sliding time window method to segment the time sequence data with a window length of 24 hours for the network point circulation time sequence data set, extracts the product circulation interval matrix and batch size matrix between network points. In the embodiment of the application, the window length is set to 24 hours, which conforms to the daily cycle planning of product distribution. For example, a network point distributes 300 products to another network point at 8:00, and 200 products to another network point at 12:00, generating a matrix with a circulation interval of 4 hours and batch sizes of 300 and 200.

[0030] S1013 calculates a network storage capacity threshold vector based on the circulation interval matrix and batch size matrix, and in combination with the regional distribution route distance, using a dynamic programming algorithm to minimize the distribution cost as the target. In the embodiment of the application, the peak period traffic congestion coefficient is considered, such as 1.5 for 8:00-10:00 and 1.0 for the flat peak period, to optimize the inventory allocation of each network point and avoid overstocking or shortages.

[0031] S1014 calculates the inventory turnover rate of each network point based on the network storage capacity threshold vector and real-time logistics data, and generates an inventory capacity saturation curve. In the embodiment of the application, the turnover rate is calculated by dividing the daily average outbound quantity by the average inventory. For example, a network point has a daily average outbound quantity of 500 pieces and an average inventory of 1000 pieces, with a turnover rate of 0.5, reflecting a 2-day inventory turnover.

[0032] S1015 uses a density clustering algorithm to cluster network points based on the inventory capacity saturation curve and peak period congestion data, with turnover efficiency as the feature, and the clustering distance threshold is set to the standard deviation of the turnover rate. In the embodiment of the application, assuming that the average turnover rate of 10 network points is 0.6 and the standard deviation is 0.1, the clustering is divided into high-efficiency, medium-efficiency and low-efficiency turnover groups.

[0033] S1016 generates a network product flow relationship graph based on the network point clustering result and the flow interval matrix through a graph network algorithm, the node weight is determined by the turnover rate, and the edge weight is determined by the flow interval. In the embodiment of the application, the graph intuitively displays the business connection between the network points, for example, the network point node with a turnover rate of 0.8 is greater than the node with a turnover rate of 0.4, and the edge with a flow interval of 2 hours is thicker than the edge with a flow interval of 6 hours.

[0034] In the embodiment of the application, the standardization processing of the historical circulation data lays a foundation for analysis. The uniform time stamp format avoids data inconsistency problems, and the sliding time window method captures dynamic flow rules. The dynamic programming algorithm optimizes the storage capacity allocation, and the density clustering and graph network algorithm further reveals the business correlation between the network points. Through these steps, the embodiment of the application generates a high-quality circulation data set, which provides reliable support for abnormal identification. The network flow relationship graph can identify key logistics nodes and distribution paths, optimize resource allocation, and significantly improve the traceability efficiency.

[0035] S102 extracts the spatio-temporal trajectory features of the target product based on the structured circulation data set, maps the flow time interval and geographic location information through a geographic information system, and generates a spatio-temporal feature matrix of product circulation.

[0036] S1021 divides the distribution area into grids according to the product position update frequency and geographic coordinate collection interval, generates trajectory point sequence data by using spatial indexing technology. In the embodiment of the application, the spatial area is divided into grids with a grid length of 1000 meters, the product position in each grid is marked by a time stamp, and a trajectory point sequence is formed. For example, a distribution vehicle records geographic coordinates every 20 seconds, generates about 4320 trajectory points in a single day, and the time stamp format is unified as 2024-05-21 09:15:00. Abnormal data, such as coordinate points outside the distribution area, are filtered by a preset geographic range to ensure the accuracy of the trajectory data.

[0037] S1022 uses a geocoding algorithm to extract the distance and time difference between adjacent trajectory points for the trajectory point sequence data, and generates a time interval matrix and a spatial displacement matrix of product flow. In the embodiment of the application, the geocoding algorithm calculates the Euclidean distance and actual driving time between two points. For example, the straight-line distance between two network points on a distribution path is 1500 meters, the actual driving distance is 2000 meters, and the time consumed is 12 minutes, indicating that the traffic speed on this section is 10 kilometers per hour, which is lower than the expected 15 kilometers per hour, reflecting potential congestion or path restrictions.

[0038] S1023 Based on the time interval matrix and the space displacement matrix, an improved Dijkstra algorithm is used to calculate the optimal path set between the network points, and the path weight comprehensively considers the distance, the traffic restriction and the traffic period. In the embodiment of the present application, the algorithm preferentially selects a path avoiding the traffic restricted area during the peak period. For example, a certain path increases the detour distance by 1000 meters during the traffic restriction period of 7:00-9:00, but the total time consumption is reduced by 15 minutes, thereby optimizing the distribution efficiency. The path set covers all the feasible routes between the network points, thereby providing a basis for analysis.

[0039] S1024 For the optimal path set, a three-dimensional path graph of product circulation is generated through a space-time three-dimensional modeling technology, the horizontal axis represents the longitude, the vertical axis represents the latitude, and the height axis represents the time. In the embodiment of the present application, the three-dimensional path graph directly displays the space-time distribution of the distribution vehicles. For example, the trajectories of 15 distribution vehicles in a certain area are concentrated on the urban trunk roads during 10:00-12:00 in the morning, and the branch trajectories are sparse, which reflects the regional preference of the distribution path.

[0040] S1025 Based on the traffic speed and the stay time data in the three-dimensional path graph, a spectral clustering algorithm is used to divide the regions with the distribution efficiency as the feature, thereby generating a space-time feature matrix of product circulation. In the embodiment of the present application, the spectral clustering divides the high-efficiency distribution area and the low-efficiency distribution area with the traffic speed and the stay time as the features. For example, the average stay time in the core commercial area is 20 minutes, and the traffic speed is 8 km / h, while the stay time in the suburban area is 5 minutes, and the traffic speed is 30 km / h. The clustering distance threshold is set to 1.2 times the standard deviation of the traffic speed, thereby ensuring the rationality of the regional division.

[0041] In the embodiment of the present application, the generation of the space-time feature matrix fully considers the dynamic characteristics of product circulation. The space indexing technology improves the trajectory point retrieval efficiency, and the geocoding algorithm accurately describes the space-time relationship between the network points. The improved Dijkstra algorithm optimizes the path selection by comprehensively considering the weight, thereby overcoming the shortcomings of the traditional method of ignoring the traffic restriction. The space-time three-dimensional modeling technology integrates the time dimension into the path analysis, thereby revealing the regional distribution rule of the distribution peak period. The spectral clustering algorithm further refines the regional features, thereby identifying the distribution efficiency difference of different functional areas, for example, the high stay time in the commercial area reflects the complex distribution demand. Through these steps, the space-time feature matrix generated by the embodiment of the present application provides high-precision data support for abnormal identification, thereby significantly improving the analysis capability, and laying a foundation for optimizing the distribution path and resource allocation.

[0042] S103 Cluster analysis is performed on the space-time feature matrix, the nodes and the node distances of the product circulation path are extracted, and the extracted nodes and node distances are compared with the preset standard sales area path, thereby calculating the deviation percentage to determine the abnormal geographical position deviation value.

[0043] S1031Based on the distance between nodes and the time interval in the space-time feature matrix, a hierarchical clustering algorithm is used to group the distribution path to generate node sequence data of the actual distribution path. In the embodiment of the application, the hierarchical clustering algorithm takes the Euclidean distance and the time interval as the characteristics to calculate the space-time similarity between nodes. For example, a certain distribution route contains 12 nodes, the average node distance is 700 meters, the time interval is 12 minutes, and after clustering, it is divided into 3 distribution areas, respectively corresponding to the core business circle, residential area and industrial area, reflecting the difference of regional distribution characteristics.

[0044] S1032For the node sequence data of the actual distribution path, the space grid division technology is used to grid the distribution area with a side length of 800 meters, and the standard distribution road network topology graph is constructed combined with the distribution point density data. In the embodiment of the application, the grid division divides the area into regular grids, and each grid records the number of distribution points and the traffic attribute. For example, a certain commercial area of 4 square kilometers is divided into 5x5 grids, and each grid contains an average of 4 distribution points, while the same area grid in the suburb contains an average of only 1 distribution point, highlighting the difference in regional distribution density.

[0045] S1033Based on the standard distribution road network topology graph, the Kruskal algorithm is used to calculate the standard distribution path node sequence, and the node weight is determined by the stay time and the traffic restriction coefficient. In the embodiment of the application, the Kruskal algorithm takes the minimization of the total traffic time as the goal, and comprehensively considers the distance between nodes and traffic control. For example, the total weight of the minimum spanning tree of 10 distribution points in a certain area is 150 minutes, the average stay time is 8 minutes, and the traffic restriction coefficient during peak period is 1.4. The generated standard path covers the main distribution route.

[0046] S1034For the node sequence of the actual distribution path and the node sequence of the standard distribution path, the spatial distance difference between nodes is calculated by the geographic coordinate projection method. In the embodiment of the application, the Web Mercator projection is used to convert latitude and longitude into plane coordinates. For example, the actual path node coordinates are (450, 620), the standard path corresponding node is (470, 600), the distance difference is calculated as 28.28 meters, reflecting the deviation degree of node position.

[0047] S1035Based on the spatial distance difference between nodes, a node deviation degree distribution matrix is generated by normalization processing, and an abnormal geographic position deviation value sequence is determined by a preset threshold. In the embodiment of the application, the normalization processing maps the distance difference to the interval of 0 to 1, and sets the deviation threshold to 25% of the average node distance. For example, the average distance of a certain path is 900 meters, the threshold is 225 meters, 80% of the node deviation is less than 0.2, showing high path consistency, and 15% of the node deviation is more than 0.3, which is judged as abnormal.

[0048] In the embodiment of the present application, the hierarchical clustering algorithm groups the space-time features, clearly distinguifies the path characteristics of different distribution areas, and overcomes the shortcomings of traditional methods in dynamic path analysis. The grid processing provides high-precision area division, and the standard road network topology map generated in combination with the distribution point density data lays a foundation for path comparison. The Kruskal algorithm optimizes the standard path calculation, fully considers the traffic restrictions in actual operation, and improves the practicability of path planning. The coordinate projection and normalization processing ensure the accuracy and consistency of the deviation calculation. For example, the deviation threshold of the city trunk road is set to 250 meters, and the branch road is relaxed to 400 meters. The analysis finds that the deviation of the core business circle is mostly caused by traffic control, with an average deviation of 320 meters, while the deviation of the suburb is mostly caused by road expansion, with an average deviation of 510 meters. Through these steps, the embodiment of the present application effectively identifies the abnormal geographical position deviation, provides reliable data support for hierarchical early warning, and significantly improves the accuracy and efficiency of path anomaly detection.

[0049] S104 calculates the distribution proportion of the same batch of products in different sales areas, network levels and time windows by conducting in-depth statistical analysis on the space-time feature matrix, quantifies the dispersion degree of each dimension by using Gini coefficient, and determines the dispersion degree of product circulation according to the maximum dispersion index.

[0050] S1041 generates network hierarchical structure data based on the space-time feature matrix and using hierarchical sampling technology to classify the distribution network according to the storage capacity of the network. In the embodiment of the present application, the network is divided into three categories according to the daily inventory: large warehouse network with inventory exceeding 1200 pieces, accounting for 12%; medium distribution network with inventory between 400 and 1200 pieces, accounting for 50%; and small retail network with inventory less than 400 pieces, accounting for 38%. The hierarchical sampling is set according to the proportion of network quantity to ensure the representativeness of the sample. For example, among 100 network points in a certain area, 12 large network points, 50 medium network points and 38 small network points are sampled to form the hierarchical structure data set.

[0051] S1042 calculates the product batch distribution of each level network by using statistical analysis tool for network hierarchical structure data, and generates network level proportion data set in combination with inventory turnover rate. In the embodiment of the present application, the turnover rate is calculated by dividing the average daily delivery quantity by the average inventory. For example, the turnover rate of large network is 0.7, and the daily processing quantity is 5 batches; the turnover rate of medium network is 1.1, and the daily processing quantity is 3 batches; the turnover rate of small network is 1.8, and the daily processing quantity is 1 batch. This analysis reveals the differences in operation efficiency of different level networks, and provides a basis for distribution analysis.

[0052] S1043, based on the grid level proportion data set, using regional division algorithm, combined with business circle boundary and regional commercial activity data, generate sales area proportion data set. In the embodiment of the application, the regional division is based on administrative division and commercial density, for example, the core business circle grid density reaches 10 per square kilometer, and the suburb is only 3 per square kilometer. The commercial activity data shows that the sales in the core business circle during weekdays is 3.8 times that in the suburb, and during weekends is 4.5 times, reflecting the significant difference in regional sales capacity.

[0053] S1044, for the sales area proportion data set, through time window division technology, according to the sales peak law, the time period is divided, and the time dimension proportion data set is generated. In the embodiment of the application, the business hours are divided into early peak 7:00-9:30, daytime flat peak 9:30-16:30 and late peak 16:30-19:00. For example, the sales of large grid in the core business circle in the early peak accounts for 40% of the whole day, the medium grid accounts for 30%, and the small grid accounts for 20%, which shows the interaction effect of time period and grid type.

[0054] S1045, based on the time dimension proportion data set, using data perspective analysis technology, a three-dimensional joint distribution table containing grid level, sales area and time dimension is constructed, and the discrete index of multi-dimensional distribution matrix is calculated. In the embodiment of the application, Gini coefficient is used to quantify the distribution imbalance. For example, the Gini coefficient of grid level dimension is 0.42, indicating that the distribution is relatively balanced; the Gini coefficient of sales area dimension is 0.70, reflecting the significant imbalance of core business circle and suburb distribution; the Gini coefficient of time dimension is 0.55, showing the concentration of sales peak. The maximum discrete index 0.70 comes from the sales area dimension, indicating that the spatial distribution imbalance is the main feature.

[0055] In the embodiment of the application, the scientificity of grid classification is ensured by stratified sampling, and the difference of grid operation efficiency is revealed by combining turnover rate analysis, which lays a foundation for accurate evaluation of product distribution. The regional division algorithm clearly describes the difference of sales area by integrating commercial activity data, especially in the comparison between core business circle and suburb, providing more fine-grained spatial distribution information. The time window division captures the dynamic law of sales peak, and the data perspective analysis integrates multi-dimensional data through three-dimensional joint distribution table, and the application of Gini coefficient further quantifies the imbalance of distribution. For example, the sales proportion of a batch of products in the early peak of large grid in the core business circle is 28%, while in the late peak of small grid in the suburb is only 12%, highlighting the asymmetry of space-time distribution. Through the extraction of the maximum discrete index, the key influencing factors of product dispersion degree are effectively identified in the embodiment of the application, which provides high-precision data support for abnormal classification and early warning response, and significantly improves the analysis ability of the system to complex circulation scenarios.

[0056] S105 combines the abnormal geographical location deviation value, the dispersion index corresponding to the dispersion degree, the product batch size, the arrival time and the network point size to construct a multi-dimensional feature matrix to perform hierarchical evaluation on the product circulation abnormality.

[0057] S1051 uses a spatial feature extraction algorithm to calculate the path deviation direction and distance according to the abnormal geographical location deviation value and the transportation time length between nodes, and generates a location abnormality feature set. In the embodiment of the present application, the spatial feature extraction algorithm quantifies the deviation degree by analyzing the vector characteristics of the path deviation. For example, the actual track of a certain distribution route deviates from the standard path by 600 meters to the northeast, the angle is 135 degrees, the transportation time length is extended from the expected 40 minutes to 62 minutes, and the location abnormality feature value is calculated as 0.82. This method accurately captures the dynamic characteristics of the path abnormality by fusing the spatial deviation and the time delay, and provides a reliable basis for multi-dimensional analysis.

[0058] S1052 uses a multi-dimensional data integration algorithm to calculate the matching degree between the product batch size and the network point storage capacity for the location abnormality feature set, and generates a size abnormality index set. In the embodiment of the present application, the integration algorithm evaluates the size matching by comparing the batch size and the remaining capacity of the network point. For example, a certain network point is designed to have a storage capacity of 1200 pieces, the current remaining capacity is 400 pieces, but the single batch arrival quantity is 900 pieces, which exceeds the capacity by 125%, and combined with the previous backlog of 200 pieces, the size abnormality index value is calculated as 0.70. This analysis reveals the impact of the mismatch between the batch size and the network point capacity on the circulation efficiency.

[0059] S1053 calculates the batch arrival delay time length from the network point inventory saturation and turnover rate data based on the size abnormality index set by using a time window analysis algorithm, and generates a time abnormality index set. In the embodiment of the present application, the time window analysis evaluates the arrival delay and the turnover efficiency in units of 6 hours. For example, a certain batch is originally scheduled to arrive at 9:00, but is actually delayed to 12:15, the delay time length is 195 minutes, the network point turnover rate decreases from 150 pieces / hour to 90 pieces / hour during the same period, and the inventory saturation is 0.90, the time abnormality index value is calculated as 0.80, which reflects the significant impact of time deviation on operational efficiency.

[0060] S1054 calculates the regional circulation abnormality rate by using a network point feature analysis algorithm in combination with the network point size and the distribution efficiency for the time abnormality index set, and generates a network point abnormality index set. In the embodiment of the present application, the analysis algorithm evaluates the deviation between the network point processing capacity and the actual distribution quantity. For example, a certain large network point is designed to have a daily processing capacity of 5000 pieces, but actually only processes 3000 pieces, the efficiency is reduced by 40%, and the proportion of similar abnormal network points in the region is 30%, the network point abnormality index value is calculated as 0.68. This analysis highlights the mismatch between the network point size and the actual operation capacity.

[0061] In the embodiment of the present application, by k-means clustering algorithm, the abnormal indexes of four dimensions of location, size, time and network are integrated into a four-dimensional feature space, and the weight of each dimension is calculated by the abnormal degree quantization algorithm to generate an abnormal feature weight matrix. For example, the weight distribution is 0.35 for location abnormality, 0.25 for size abnormality, 0.25 for time abnormality, and 0.15 for network abnormality. The comprehensive abnormal value of a certain route is 0.75, which exceeds the warning threshold 0.65. Based on this, the system uses a hierarchical standard generation algorithm to divide the abnormality into three levels of mild 0.4, moderate 0.6 and severe 0.8 according to the abnormal degree. For example, the mild abnormality of the core business circle accounts for 30%, the moderate abnormality accounts for 15%, and the severe abnormality accounts for 8%. The suburb is 20%, 25% and 15%, respectively, showing significant regional differences. The above steps clearly depict the complexity of the abnormality through multi-dimensional feature integration and weighted analysis, significantly improve the accuracy and operability of the abnormality classification, and provide solid data support for early warning response.

[0062] S106 according to the different levels of product circulation abnormality, an automatic push mechanism is used to send a hierarchical early warning signal to the management platform to accurately determine the early warning trigger time and target object.

[0063] S1061 utilizes a risk quantization algorithm to generate a risk level matrix from three dimensions of abnormal degree, duration and influence range. In the embodiment of the present application, the risk quantization algorithm calculates the abnormal risk by weighted calculation. For example, the abnormal degree of a certain network batch backlog is 0.85, the duration is 4.5 hours, and the influence on 4 surrounding networks is calculated as a risk index of 0.78, which is classified as a high risk level. Another network has a short-term delay abnormality degree of 0.25, a duration of 0.8 hours, and only affects itself, with a risk index of 0.20, which is classified as a low risk level. This multi-dimensional quantization method ensures the comprehensiveness and accuracy of risk assessment.

[0064] S1062 sets the early warning trigger threshold value for the risk level matrix through a rule generation algorithm based on risk accumulation value, propagation speed and correlation strength parameters. In the embodiment of the present application, the high risk threshold is set to risk accumulation value 0.75, propagation speed 3 network points per hour, and correlation strength 0.65. The medium risk is 0.50, 1.5 network points per hour, and 0.40. The low risk is 0.30, 0.5 network points per hour, and 0.20. For example, the cumulative value of a certain high-risk abnormality reaches 0.80, the propagation speed is 2.5 network points per hour, and the high-risk early warning is triggered. The rule generation algorithm dynamically adjusts the threshold to adapt to different scenarios.

[0065] S1063, based on the early warning trigger threshold, adopts a hierarchical mapping algorithm to extract early warning object features from the network response time limit and service attributes, and generates a hierarchical early warning object list. In the embodiment of the application, the core distribution center response time limit is 10 minutes, which is classified as a first-level object; the regional distribution center response time limit is 25 minutes, which is classified as a second-level object; and the end network point response time limit is 50 minutes, which is classified as a third-level object. For example, a high-risk abnormality triggers a first-level object warning, and the list contains 3 core distribution centers to ensure fast response.

[0066] S1064, for the hierarchical early warning object list, calculates the early warning priority index by a decision tree algorithm, considering the risk level, early warning timeliness and network level. In the embodiment of the application, the decision tree takes the risk level as the root node, and the timeliness and network level as branch nodes. For example, a high-risk abnormality of a first-level network point requires a response within 5 minutes in terms of timeliness, and the priority index is 0.92; the priority index of a medium-risk abnormality of a second-level network point is 0.60. The decision tree algorithm optimizes the priority allocation through multi-layer judgment.

[0067] S1065, based on the early warning priority index, sets the push frequency by a time window scheduling algorithm, and generates an early warning time trigger table from the early warning duration data. In the embodiment of the application, high-priority early warnings are pushed every 4 minutes during the morning peak period 7:00-9:30, and every 12 minutes during the flat peak period; medium-priority early warnings are pushed every 10 minutes during the peak period, and every 25 minutes during the flat peak period; and low-priority early warnings are fixedly pushed every 50 minutes. For example, a high-risk early warning is pushed 8 times during 8:00-8:30, ensuring timely delivery.

[0068] In the embodiment of the application, a message queue distribution mechanism is used to generate an early warning information package containing a risk level identifier, an early warning object number and a trigger timestamp. For example, an information package {H-DC002-20250521104500} represents a high-risk early warning, aimed at the second distribution center, triggered at 10:45:00 on May 21, 2025. The system selects a push mode according to the early warning level through a multi-channel push mechanism: high-risk early warnings are pushed through the management platform pop-up window, SMS and email simultaneously; medium-risk early warnings are pushed through the pop-up window and SMS; and low-risk early warnings are only pushed through the pop-up window. For example, a certain region continuously receives 4 high-risk early warnings, and the system automatically expands the push range to the entire distribution network, with early warning information covering 5 network points and the response time being shortened to within 15 minutes. This hierarchical push mechanism significantly improves the pertinence and timeliness of early warnings, providing efficient support for the management platform to quickly respond to potential risks.

[0069] S107 generates an abnormal intervention adjustment scheme according to the trigger condition of the early warning signal, including suggestions for checking nodes and adjusting the frequency of sampling, uses real-time data stream processing technology to monitor changes in product circulation track, and obtains track deviation data after intervention.

[0070] S1071Based on the early warning signal trigger record, a correlation rule mining algorithm is used to generate a check node priority sequence, and the priority is jointly determined by the abnormality occurrence frequency and the treatment timeliness. In the embodiment of the application, the correlation rule mining algorithm calculates the node priority by analyzing the frequency and response time of the early warning signal. For example, a certain distribution center triggered 10 early warnings in the past week, with an average treatment time of 50 minutes, which is 30 minutes higher than the average value, and the priority index reaches 0.88, ranking as the primary check node. This method optimizes the allocation of inspection resources by quantifying the frequency of abnormality and treatment efficiency.

[0071] S1072For the check node priority sequence, a resource scheduling algorithm is used to develop an inspection resource allocation scheme, and the sampling frequency reference value is extracted from the inspection point distribution data. The resource scheduling algorithm allocates manpower and equipment according to the node priority, with 3 full-time inspectors and 100% equipment coverage for high-priority nodes, 2 inspectors and 75% equipment coverage for medium-priority nodes, and 1 mobile inspector and 50% equipment coverage for low-priority nodes. For example, the sampling frequency of a certain high-priority node is set to once every 1.5 hours to ensure rapid detection of abnormalities.

[0072] S1073Based on the sampling frequency reference value, a random forest algorithm is used to calculate the sampling time window, taking into account the abnormality level, treatment speed and resource density characteristics. The random forest algorithm evaluates feature weights through multiple decision trees to generate an optimized time window. For example, the treatment speed of a high-risk node is 30% lower than the standard, and the resource density is insufficient, so the sampling interval is shortened from 2 hours to 1 hour; the low-risk node has sufficient resources, and the interval is extended to 5 hours. This method dynamically balances inspection efficiency and resource utilization.

[0073] S1074For the sampling time window, a real-time stream processing technology is used to build a trajectory monitoring scheme, record position deviation, time delay and path change, and use a data stream collector to obtain real-time trajectory data at a dynamic sampling interval. The normal transportation sampling interval is 4 minutes, and the abnormal trigger shortens it to 1 minute. For example, a batch of products generated 120 trajectory points in 6 hours of transportation, with a position deviation of 600 meters, a detention time exceeding 100 minutes, and a path change of 3 times, triggering real-time intervention.

[0074] S1075For the real-time trajectory data stream, a trajectory comparison algorithm is used to calculate the trajectory change before and after intervention, generate a trajectory deviation metric value, and use a feedback adjustment mechanism to optimize the inspection configuration. The trajectory comparison algorithm compares the paths before and after intervention. For example, the average deviation of a certain route before intervention is 500 meters, and the maximum deviation is 900 meters; after intervention, the average deviation is reduced to 120 meters, and the detention time is reduced from 80 minutes to 30 minutes. The feedback adjustment mechanism adjusts the resources according to the deviation improvement rate. When the improvement rate exceeds 50%, the sampling frequency is reduced, and when it is lower than 15%, the inspection resources are increased.

[0075] S1076 For high-frequency abnormal dealers, calculate the early warning signal density and trigger interval through the anomaly detection algorithm, generate a high-frequency abnormal dealer list, and classify it against the abnormal behavior feature library. The anomaly detection algorithm weights the warning frequency and severity, for example, a certain dealer triggers 18 warnings in 30 days, 10 high-risk warnings, and an anomaly index of 0.90, which is included in the high-frequency abnormal list. The classification shows that its behavior matches the "cross-region string goods" feature, with a matching degree of 0.95, combined with a D-level credit rating, and is determined to be a high-risk violation.

[0076] S1077 Based on the distribution of violation behavior levels, set regional limits and time limits through the sales limit algorithm, and determine the freeze period through the qualification assessment algorithm to generate graded control measures. The daily sales limit of a high-risk dealer is reduced to 40% of the historical average, with a constraint period of 8:00-17:00; multiple violations trigger a 30-day qualification freeze. For example, a certain dealer's daily sales decreased from 1200 to 480, orders during the freeze period need to be audited, and abnormal behavior has significantly decreased.

[0077] Generate temporary traceability codes for suspicious batches of products, use the execution tracking algorithm to record the implementation of control measures, and evaluate the effect through the closed-loop monitoring algorithm. For example, batch "BT20250521" is attached with code "TMP-BT20250521-H2", and monitoring shows that its cross-region sales behavior has decreased from an average of 2 times a day to 0 times, with an execution rate of 97%. Closed-loop monitoring statistics show that a certain dealer's warning frequency has decreased from an average of 0.6 times a day to 0.08 times a day after control, transaction data has returned to the compliance range, and logistics records show that path deviation rate has decreased by 70%, providing data basis for lifting control.

[0078] S108 Update the spatio-temporal feature matrix according to the trajectory change data after intervention, and re-execute the deviation detection and dispersion degree analysis through the loop feedback mechanism to generate new abnormal recognition results.

[0079] S1081 Use the data synchronization algorithm to extract feature points from the trajectory change data at a preset sampling interval to form a trajectory feature sequence. The data synchronization algorithm dynamically adjusts the sampling frequency according to the transportation scenario, with an interval of 20 seconds for urban distribution routes, generating 1080 trajectory points within 6 hours; an interval of 90 seconds for rural trunk transportation, generating 480 trajectory points within 12 hours. Abnormal points are filtered out by geographical range to ensure high-quality sequence data and provide a reliable basis for feature extraction.

[0080] S1082, for the trajectory feature sequence, a preliminary version of the feature matrix is constructed through a spatio-temporal feature extraction algorithm, including four dimensions of position coordinates, time stamps, instantaneous speed and motion direction. The preliminary version of the feature matrix records the latitude and longitude to 0.0001 degrees, the time stamp to the second, the speed in kilometers per hour, and the direction in degrees. For example, a delivery vehicle has an average speed of 12 kilometers per hour during peak hours, which is lower than the standard of 20 kilometers per hour, and the direction changes 5 times per hour, indicating that there is detouring or congestion.

[0081] S1083, based on the preliminary version of the feature matrix, a long short-term memory network is used to calculate the position deviation value, with position and time features as input. The long short-term memory network contains two layers of hidden layers, with 128 neurons in each layer, which analyzes the trajectory data in a continuous time window and outputs the deviation value. For example, the actual trajectory of a certain route deviates from the planned path by 600 meters, lasts for 50 minutes, and the deviation value is 0.78, which exceeds the normal fluctuation threshold of 0.5, indicating potential abnormalities.

[0082] S1084, for the position deviation value, a dispersion analysis algorithm is used to calculate the regional dispersion index, covering spatial dispersion and temporal dispersion. Spatial dispersion is calculated based on the distribution density of delivery points, with a core business district of 0.30 and a suburban area of 0.68; temporal dispersion reflects the uniformity of delivery time, with a peak period of 0.70 and a flat period of 0.28. For example, a certain area has a deviation value of 450 meters and a time dispersion of 0.72, indicating low delivery efficiency.

[0083] S1085, based on the regional dispersion index, a feature update algorithm is used to generate the second version of the feature matrix, integrating the new deviation value and dispersion index, and the updated data is returned to the anomaly detection process. The second version of the matrix shows that the mean deviation of a certain area has risen to 550 meters, the spatial dispersion is 0.80, and the temporal dispersion is 0.78, and the abnormal feature vector is significantly enhanced. The backflow mechanism re-enters the updated data into the deviation detection through the data backflow device, generating a new abnormal feature vector.

[0084] A parameter optimization algorithm is used to dynamically adjust the identification parameters according to the abnormal feature vector, and the update of the abnormal identification result is completed. For example, the initial position deviation threshold is adjusted from 350 meters to 280 meters, the time dispersion threshold is reduced from 0.65 to 0.48, and the spatial dispersion threshold remains at 0.68. After optimization, the accuracy of abnormal identification is improved from 82% to 95%, and the false negative rate is reduced from 10% to 3%. Continuous 90-day data analysis shows that the abnormal frequency in a certain area has decreased from 0.4 times per day to 0.1 times per day, and the mean trajectory deviation has decreased from 500 meters to 200 meters, reflecting the significant effect of intervention measures, providing precise data support for early warning and optimization.

[0085] The above embodiment is only one of the preferred embodiments of the present application, and should not be used to limit the protection scope of the present application, but any modification or polishing without substantial meaning made in the main design idea and spirit of the present application, and the technical problems solved are still consistent with the present application, and should be included in the protection scope of the present application.

Claims

1. An enterprise intelligent early warning method based on an artificial intelligence model, characterized in that, The method comprises: By collecting and cleaning historical circulation data, identifying product flow records of each network point, extracting time stamps, geographic location information, product batch size, arrival time and network point size, and arranging into a structured circulation data set; Combined with the circulation data set, the spatio-temporal trajectory features of the target product are extracted, the flow time interval is identified, the geographic information system is used to map the geographic location information and the time interval, and the spatio-temporal feature matrix of product circulation is generated; By clustering analysis on the spatio-temporal feature matrix, the product flow path nodes and node distances are obtained, which are compared with the nodes and node distances of the preset standard sales area path to obtain the deviation percentage, and the abnormal geographic location deviation value is determined according to the deviation percentage; By statistical analysis on the spatio-temporal feature matrix, the distribution proportion of the same batch of products in different sales areas, different level network points and different time windows is obtained, the Gini coefficient is calculated to obtain the dispersion index of each dimension, and the dispersion degree is determined by the maximum dispersion index; Combined with the abnormal geographic location deviation value, the dispersion index corresponding to the dispersion degree, and the product batch size, arrival time and network point size, the product circulation anomaly is graded, including: According to the node-to-node transportation time data, a spatial feature extractor is used, which constructs a path anomaly index according to the offset direction and distance to obtain a location anomaly feature set; For the location anomaly feature set, the product batch size is mapped and operated with the network point storage capacity to obtain a scale anomaly index set; According to the scale anomaly index set, the batch arrival delay time and the network turnover speed are calculated from the network inventory saturation data to obtain a time anomaly index set; For the location anomaly feature set, the scale anomaly index set and the time anomaly index set, the abnormal degree value of each dimension is calculated, and the abnormal level is divided according to the abnormal degree value; For different levels of abnormal risk, an early warning signal is sent to the management platform to determine the early warning trigger time and object; According to the early warning signal trigger condition, an abnormal intervention adjustment scheme is generated, the product circulation trajectory is monitored, and the trajectory change data after intervention is obtained; For the trajectory change data, the spatio-temporal feature matrix is updated, the data is re-input into the deviation detection and dispersion degree analysis process through a cyclic feedback mechanism, and a new abnormality recognition result is obtained. 2.The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The method comprises: According to historical circulation records, regular expression matching is used to extract timestamp information, and network product flow time series data set is obtained through data standardization processing; For the network product flow time series data set, a sliding time window method is used to segment the time series data, and a product flow interval matrix and a batch size matrix between network points are obtained through time window segmentation; According to the product flow interval matrix and the batch size matrix, a dynamic programming algorithm is used to calculate a network storage capacity threshold vector; The density clustering algorithm is used for network point clustering according to the network point storage capacity threshold vector, a network product flow relationship graph is generated through a graph network construction algorithm, the node weight of the relationship graph is determined by the inventory turnover rate, and the edge weight of the relationship graph is determined by the flow interval. 3.The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The spatial and temporal trajectory features of the target product are extracted by combining the circulation data set, the flow time interval is identified, the geographic location information and the time interval are mapped by using a geographic information system, and a spatial and temporal feature matrix of product circulation is generated, including: The spatial region is divided into grids by using a quadtree index, and product trajectory point sequence data is obtained by time stamp marking; According to the product trajectory point sequence data, a geographic information encoder is used to extract the distance and time difference between trajectory points, and a product flow time interval matrix and a spatial displacement matrix are obtained; For the time interval matrix and the spatial displacement matrix, a shortest path algorithm is used to calculate a path set between stations, and a three-dimensional path graph is constructed by using a space-time cube method; According to the passing speed and residence time data in the three-dimensional path graph, a spectral clustering method is used to divide the region, and the spatial and temporal feature matrix of product circulation is obtained. 4.The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The deviation percentage is obtained by comparing the product flow path nodes and the node distance with the nodes and the node distance of the preset standard sales area path one by one, and the abnormal geographic position deviation value is determined according to the deviation percentage, including: According to the distance and time interval between nodes in the space-time feature matrix, a hierarchical clustering algorithm is used to group the distribution path, and actual distribution path node sequence data is obtained; The actual distribution path node sequence data is divided by a fixed edge length by using a spatial gridding method, and a standard distribution road network topology graph is constructed from the regional distribution point density data; For the standard distribution road network topology graph, a minimum spanning tree algorithm is used to calculate the standard distribution path node sequence, and the node weight is determined by the residence time and the passing limit degree; The spatial distance difference between the actual distribution path node sequence and the standard distribution path node sequence is calculated by a coordinate projection method, and if the spatial distance difference exceeds the preset deviation threshold, it is determined as an abnormal geographic position deviation value sequence. 5.The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The distribution proportion of the same batch of products in different sales areas, different levels of network points and different time windows is obtained by statistical analysis of the space-time feature matrix, the Gini coefficient is calculated to obtain the dispersion index of each dimension, and the dispersion degree is determined by the maximum dispersion index, including: The network point storage capacity is processed by using a hierarchical sampling method, and network point hierarchical structure data is obtained according to the network point storage capacity division; The batch distribution of the network point hierarchical structure data is calculated by using a hierarchical statistical calculator, and a network point level proportion data set is obtained from the batch distribution calculation; The network point level proportion data set is divided according to the region boundary, and a sales area proportion data set is obtained according to the region boundary division; The three-dimensional joint distribution table is constructed for the sales area proportion dataset by using a data pivot tool, a dimension dispersion index value is calculated from the three-dimensional joint distribution table, and if the dimension dispersion index value is greater than a preset threshold value, a product dispersion degree identification value is obtained.

6. The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The management platform is sent a warning signal according to different levels of abnormal risks, and a warning trigger time and object are determined, including: A risk level matrix is generated according to an abnormal degree parameter, a duration parameter and an influence range parameter; A warning trigger threshold is obtained according to a risk accumulation value parameter, a propagation speed parameter and a correlation strength parameter; According to the warning trigger threshold, a warning classification mapper is used to extract warning object features from the branch response time limit data, and the warning classification mapper obtains a classified warning object list according to the warning object features; If the classified warning object list triggers a warning condition, the classified warning object list is processed by a decision tree algorithm, and a warning priority index is obtained by the decision tree algorithm according to a risk level node, a warning timeliness node and a branch level node.

7. The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, According to the warning signal trigger condition, an abnormal intervention adjustment scheme is generated, the product circulation track is monitored, and track change data after intervention is obtained, including: A check node priority sequence is generated by using an association rule extractor according to a warning signal trigger record, and the check node priority sequence obtains a node priority value based on abnormal occurrence frequency and disposal timeliness; A node check resource configuration scheme is calculated for the check node priority sequence by a resource scheduler, and an inspection frequency benchmark value is obtained from the inspection point distribution data by the resource scheduler; The inspection frequency benchmark value is processed by using a random forest algorithm, and an inspection time window is obtained based on abnormal level, disposal speed and resource density characteristics by the random forest algorithm; According to the inspection time window, a data stream collector is used to record the product position according to a preset sampling interval, and the data stream collector obtains a real-time track data stream and performs track comparison to obtain a track deviation metric value.

8. The method of claim 7, wherein, Further comprising: The severity and trigger frequency of the warning signal are identified, the dealers with frequent abnormal occurrences are identified according to the trigger frequency of the warning signal, a targeted intervention scheme is generated by matching a product circulation abnormal type library, a temporary sales limit and a circulation time window restriction in a specific area are set for suspicious dealers, a temporary qualification freeze is implemented for dealers with serious irregularities, a temporary traceability identifier is added to suspicious product batches for key monitoring, and an abnormal intervention execution list for closed-loop control is formed, specifically including: The dealer warning signal density and trigger time interval data are obtained, an abnormal index corresponding to the warning signal is calculated, and a high-frequency abnormal dealer list is obtained according to the abnormal index; According to the high-frequency abnormal dealer list, classification is performed by comparing an abnormal behavior characteristic library, and a violation behavior level distribution table is obtained from dealer credit level data; According to the violation behavior level distribution table, a regional limit and a time period restriction are set, and a limit constraint parameter set is obtained from regional transaction records; According to the limit constraint parameter set, a qualification freeze period is generated, and a hierarchical control measure set is obtained from the violation behavior type. 9.The enterprise intelligent early warning method based on an artificial intelligence model according to claim 1, characterized in that, The trajectory change data is used to update the space-time feature matrix, the data is re-input into the deviation detection and dispersion analysis process through a loop feedback mechanism, and a new abnormality recognition result is obtained, including: According to the trajectory change data, trajectory feature points are extracted from the trajectory change data according to a preset sampling interval, and a trajectory feature sequence is obtained. For the trajectory feature sequence, a feature matrix first version is constructed through a space-time feature extractor, and the feature matrix first version contains position coordinates, time stamps, and speed and direction data. According to the feature matrix first version, a position deviation value is calculated using a recurrent neural network, and the recurrent neural network obtains position and time two-dimensional features from the feature matrix first version as input data. For the position deviation value, a regional dispersion index is generated by a dispersion calculator, and the regional dispersion index contains spatial dispersion and temporal dispersion data. If the regional dispersion index exceeds a preset threshold, an abnormality recognition parameter is updated using a feature updater, and an abnormality recognition result is obtained.

Images