A fruit skin box overflow early warning management method and system

Abstract

The application relates to the technical field of fruit peel box management, and discloses a fruit peel box overflow early warning management method and system, which comprises the following steps: acquiring, in real time, the in-box garbage images of a plurality of target management objects in a target management area, surrounding instantaneous passenger flow data and position information; identifying the preprocessed in-box garbage images based on a U-Net semantic segmentation network and a bidirectional LSTM neural network, and analyzing the in-box garbage images in combination with the surrounding instantaneous passenger flow data; generating the space-time management load entropy of each target management object; strengthening the correlation between the target management objects through the generated dynamic mutual construction weight management matrix, avoiding management deviation caused by isolated analysis; calculating the management index of the driving analysis unit; clearly defining the load emergency degree of the target management object; facilitating the judgment of the management priority in the later period; and issuing the collaborative early warning management instruction of the early warning management unit, integrating the scattered early warning information, and collaboratively and efficiently managing the target management object, so that the management efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of peel box management, in particular to a peel box overflow early warning management method and system. BACKGROUND

[0002] At present, the peel box overflow early warning management still mainly adopts manual inspection as the main mode, and relies on the on-site troubleshooting of operation and maintenance personnel to realize overflow state recognition and basic early warning. However, with the application of intelligent technology, many management methods are through image acquisition and neural network algorithm. Among them, the real-time image of the peel box is obtained through a high-definition camera, and the neural network algorithm is used to extract and intelligently determine the garbage accumulation and the overflow degree of the box body, so as to realize the automatic early warning of the single box overflow state.

[0003] However, in the grandstand area of large-scale events, the existing early warning management method still has the following defects, specifically: the instantaneous passenger flow of the grandstand area is large during the event, the grandstand area is wide, the peel box points are relatively scattered and the total quantity is large, and under the independent early warning of the peel box, multiple peel boxes may appear to be simultaneously overflowed and early warned, which will cause the early warning information to be out of control and cannot be effectively managed. SUMMARY

[0004] In view of the defects of the prior art, the present application provides a peel box overflow early warning management method and system, which solves the above problems.

[0005] The above technical purpose of the present application is realized by the following technical scheme:

[0006] A peel box overflow early warning management method, comprising:

[0007] Step S1, real-time acquisition of the in-box garbage image, surrounding instantaneous passenger flow data and position information of a plurality of target management objects in a target management area, identification of the preprocessed in-box garbage image based on a U-Net semantic segmentation network and a bidirectional LSTM neural network, and analysis combined with the surrounding instantaneous passenger flow data to generate the space-time management load entropy of each target management object, including: identification of the preprocessed in-box garbage image based on the U-Net semantic segmentation network to generate a capacity fluctuation spectrum, including: processing a single frame of in-box garbage image by using a semantic segmentation network with an encoder-decoder structure to obtain a binary mask of the garbage area, and then calculating the capacity rate based on the binary mask to form a capacity rate time sequence, and performing bidirectional time sequence analysis on the capacity rate time sequence according to the bidirectional long short-term memory network to capture the change trend and mutual influence of the capacity rate, and generating the capacity fluctuation spectrum; the target management area is a grandstand area of a stadium, and the target management object is a peel box;

[0008] Step S2, analyze the space-time management load entropy of all target management objects, generate a regional load situation tensor, and calculate the regional load situation tensor based on a neural network algorithm of a self-attention mechanism to generate a dynamic mutual construction weight management matrix of the target management objects in the target management region;

[0009] Step S3, dynamically calculate the dynamic mutual construction weight management matrix and the space-time management load entropy to obtain a management index of each target management object;

[0010] Step S4, analyze the management index and position information of all target management objects to generate an optimal sequence of the target management objects;

[0011] Step S5, perform global optimization sorting management and regional cluster division management on the optimal sequence to generate a collaborative early warning management instruction of the target management objects.

[0012] Further, based on the U-Net semantic segmentation network and the bidirectional LSTM neural network, the preprocessed garbage image in the box is identified and analyzed in combination with the surrounding instantaneous passenger flow data to generate the space-time management load entropy of each target management object, which further includes:

[0013] The U-Net semantic segmentation network is used to identify the preprocessed garbage image in the box to generate a container fluctuation spectrum;

[0014] The target management region is divided into four sub-regions;

[0015] Based on the position information of the target management object, the path distance of the target management object to each sub-region is calculated, and the surrounding instantaneous passenger flow of the target management object is fused and calculated to generate a spatial attraction mode value representing the probability of the target management object being preferentially selected for deployment;

[0016] The stage of the stadium is obtained, and a time sequence pressure factor is assigned to different stages.

[0017] Further, based on the U-Net semantic segmentation network and the bidirectional LSTM neural network, the preprocessed garbage image in the box is identified and analyzed in combination with the surrounding instantaneous passenger flow data to generate the space-time management load entropy of each target management object, which further includes:

[0018] The time sequence analysis is performed on the container fluctuation spectrum to generate an instantaneous change gradient, and the instantaneous change gradient is combined with the spatial attraction mode value and the time sequence pressure factor to generate a load surge expectation vector;

[0019] The physical volume upper limit of the target management object is obtained;

[0020] Based on the container fluctuation spectrum, the load surge expectation vector and the physical volume upper limit are fused to generate the space-time management load entropy.

[0021] Further, analyze the spatio-temporal management load entropy of all target management objects, generate the regional load situation tensor, including:

[0022] Spatial gridding mapping of the spatio-temporal management load entropy of all target management objects, combined with the spatial attributes of the sub-regions to which each target management object belongs, calculate the load superposition coefficient of target management objects in each grid cell;

[0023] Based on the time sequence division of the stadium event stage, the load superposition coefficient of each grid cell is dynamically correlated and analyzed across time periods, and the regional load time sequence linkage index is generated;

[0024] Fusion of load superposition coefficient and regional load time sequence linkage index, combined with the distribution density characteristics of target management objects in grid cells, generate regional load situation tensor.

[0025] Further, based on the neural network algorithm of self-attention mechanism, the regional load situation tensor is calculated to generate the dynamic mutual construction weight management matrix of target management objects in the target management region, including:

[0026] Extract the load dynamic characteristics of each target management object from the regional load situation tensor, analyze the influence probability of the load change of different target management objects on the surrounding target management objects, and generate the load induced potential;

[0027] Based on the neural network of self-attention mechanism, the load induced potential is taken as the core basis of attention allocation, the asymmetric correlation weight formed between different target management objects due to the difference in load induced potential is calculated, and the induced correlation matrix is generated;

[0028] Based on the position information of the target management object and the sub-region to which it belongs, the spatial steric hindrance degree between each target management object is calculated;

[0029] According to the spatial steric hindrance, the induced correlation matrix is corrected to reduce false correlation, and the spatial tuning correlation matrix is generated;

[0030] Analyze the dynamic change rate of the time sequence pressure factor and the load induced potential to obtain the time sequence sensitivity coefficient.

[0031] Further, based on the neural network algorithm of self-attention mechanism, the regional load situation tensor is calculated to generate the dynamic mutual construction weight management matrix of target management objects in the target management region, also including:

[0032] Based on the time sequence sensitivity coefficient, adjust the time decay characteristics of each correlation weight in the spatial tuning correlation matrix, so that the correlation relationship changes dynamically with the progress of the event, and generate the time-space covariant correlation matrix;

[0033] The multi-dimensional feature enhancement is performed on the time-space covariant correlation matrix through the multi-head attention of the self-attention mechanism, the correlation characteristics between the high-induced potential target management object and the high-load target management object are analyzed, and a feature highlighted correlation matrix is generated;

[0034] The load superposition coefficient of each sub-region and the surrounding instantaneous passenger flow data distribution are analyzed to determine the correlation adaptation threshold of each sub-region;

[0035] Based on the correlation adaptation threshold, the feature highlighted correlation matrix is corrected to generate a dynamic mutual construction weight management matrix of the target management objects in the target management area.

[0036] Further, the dynamic mutual construction weight management matrix and the space-time management load entropy are dynamically calculated to obtain the management index of each target management object, including:

[0037] Based on the dynamic mutual construction weight management matrix, the space-time management load entropy of each target management object is calculated by bidirectional mapping to generate a coupling interaction strength;

[0038] The residual physical volume of the target management object is obtained in real time;

[0039] The coupling interaction strength and the residual physical volume are calculated to generate a load criticality ratio;

[0040] The time sequence change of the load criticality ratio is analyzed to generate a load criticality increase rate;

[0041] The load criticality ratio and the load criticality increase rate are fused to generate the management index of each target management object.

[0042] Further, the management index and the position information of all target management objects are analyzed to generate an optimal sequence driven management sequence of the target management objects, including:

[0043] The management index and the position information of each target management object are analyzed, and the potential level-topological coupling value is obtained in combination with the sub-region;

[0044] The distribution characteristics of the potential level-topological coupling value are analyzed to generate a scheduling conduction dependency between adjacent objects;

[0045] Based on the scheduling conduction dependency, the potential level-topological coupling value is corrected to generate a dynamic scheduling potential;

[0046] Based on the dynamic scheduling potential, the convenience of the unloading path of the sub-region division is analyzed to obtain a resource reachability adaptation degree of each target management object;

[0047] The dynamic scheduling potential and the resource reachability adaptation degree are fused and sorted to generate the optimal sequence driven management sequence.

[0048] Further, the optimal sequence driving management sequence is globally optimized and sorted and regionally clustered, and a cooperative early warning management instruction of the target management object is generated, including:

[0049] The dynamic scheduling potential of each target management object in the optimal sequence driving management sequence and the resource reach adaptation degree are analyzed, and the global scheduling influence entropy of each object is obtained.

[0050] Based on the global scheduling influence entropy and the sub-region, the optimal sequence driving management sequence is clustered, and the cluster internal cooperative efficiency of each cluster is calculated, and a load cooperative cluster is generated.

[0051] The cluster internal cooperative efficiency of each load cooperative cluster and the load critical speed of the target management object in the cluster are fused and calculated, and the cluster level early warning priority is obtained.

[0052] According to the cluster level early warning priority, the load cooperative cluster is globally sorted, and the spatial position of each cluster is combined to generate a cooperative early warning management instruction of the target management object.

[0053] Further, a fruit peel box overflow early warning management system is applied to the fruit peel box overflow early warning management method, and includes:

[0054] The recognition unit is used for real-time acquisition of the in-box garbage image of a plurality of target management objects in a target management region, the surrounding instantaneous passenger flow data and the position information, and the in-box garbage image after preprocessing is recognized based on a U-Net semantic segmentation network and a bidirectional LSTM neural network and analyzed in combination with the surrounding instantaneous passenger flow data to generate a spatiotemporal management load entropy of each target management object. The target management region is a stand area of a stadium, and the target management object is a fruit peel box.

[0055] The calculation unit is used for analyzing the spatiotemporal management load entropy of all target management objects, generating a regional load situation tensor, and calculating the regional load situation tensor based on a neural network algorithm of a self-attention mechanism to generate a dynamic mutual construction weight management matrix of the target management object in the target management region.

[0056] The driving analysis unit is used for dynamically calculating the dynamic mutual construction weight management matrix and the spatiotemporal management load entropy to obtain a management index of each target management object.

[0057] The management optimization unit is used for analyzing the management index and the position information of all target management objects to generate an optimal sequence driving management sequence of the target management object.

[0058] The early warning management unit is used for globally optimizing and sorting the optimal sequence driving management sequence and regionally clustering to generate a cooperative early warning management instruction of the target management object.

[0059] In summary, the present application has the following advantages:

[0060] By acquiring the in-box garbage image of the target management object, the surrounding instantaneous passenger flow data and the position information in real time, the spatial management load entropy is generated by combining the neural network algorithm to generate the capacity fluctuation spectrum, the spatial gravity modal value and the time series pressure factor, further fusing the load surge expectation vector and the physical volume upper limit, and further can avoid the management decision mistakes caused by data deviation, so that the load state judgment and early warning of the target management object is more accurate; at the same time, the regional load situation tensor containing the load superposition coefficient, the regional load time series linkage index and the distribution density is generated based on the spatial management load entropy, and then the dynamic mutual construction weight management matrix is generated through the optimization of multiple steps such as load induced potential, induced correlation matrix, spatial tuning correlation matrix and time series-space covariant correlation matrix through the self-attention mechanism, so as to strengthen the associated management between the target management objects, change the target management object management from single to cooperation, and make the early warning information more complete.

[0061] The management index is calculated by coupling the interaction strength, the load critical ratio and the load critical speed increase, so that the load emergency degree of each target management object can be clearly determined, the management priority can be sorted in the later period, the management focus can be quickly focused on the target management object with high emergency degree, and the management response efficiency is improved; and through the generated optimal sequence driving management sequence, the target management object management can have clear priority, the processing confusion when multiple target management objects are simultaneously warned is avoided, and the order of the management process is improved; and the final collaborative early warning management instruction can realize the cluster collaborative management of the target management objects, cooperates the scattered early warning information, solves the early warning information blowout problem, and improves the overall management efficiency of the target management objects in the grand event stand area. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a fruit peel box overflow early warning management method flow chart of the present application;

[0063] Figure 2 is a fruit peel box overflow early warning management system schematic diagram of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] Reference Figure 1 and Figure 2 A fruit peel box overflow early warning management method, comprising:

[0066] Step S1, real-time acquisition of the in-box garbage image of a plurality of target management objects in the target management area, the surrounding instantaneous passenger flow data and the position information, identification of the preprocessed in-box garbage image based on the U-Net semantic segmentation network and the bidirectional LSTM neural network, analysis combined with the surrounding instantaneous passenger flow data, generation of the space-time management load entropy of each target management object, the target management area is the stand area of the stadium, and the target management object is the fruit box;

[0067] The surrounding instantaneous passenger flow data includes the number of moving people, the moving track of people, the staying time of people, the density distribution of people flow, the number of audience, the number of staff and the like.

[0068] The position information includes the coordinate, the spatial level attribute and the like; the spatial level attribute mainly refers to the stand spatial level to which the fruit box belongs, such as the first stand, the second stand, the stand under the podium and the like.

[0069] Step S2, analysis of the space-time management load entropy of all target management objects, generation of the regional load situation tensor, calculation of the regional load situation tensor based on the neural network algorithm of the self-attention mechanism, generation of the dynamic mutual construction weight management matrix of the target management objects in the target management area.

[0070] Step S3, dynamic calculation of the dynamic mutual construction weight management matrix and the space-time management load entropy, obtaining of the management index of each target management object.

[0071] Step S4, analysis of the management index and the position information of all target management objects, generation of the optimal sequence driven management sequence of the target management objects.

[0072] Step S5, global optimization sorting management and regional cluster division management of the optimal sequence driven management sequence, generation of the collaborative early warning management instruction of the target management objects.

[0073] Through real-time acquisition of the in-box garbage image, the surrounding instantaneous passenger flow and the position information, generation of the space-time management load entropy of each target management object combined with the neural network algorithm, it is convenient for accurate management in the later period, analysis of the space-time management load entropy to generate the regional load situation tensor, generation of the dynamic mutual construction weight management matrix relying on the self-attention mechanism, optimization of the management logic related between the target management objects, subsequent calculation of the management index, generation of the optimal sequence driven management sequence combined with the position information, avoidance of the disordered management of independent early warning, global optimization sorting and regional cluster division of the optimal sequence driven management sequence, generation of the collaborative early warning management instruction, which can upgrade the independent early warning of a single target management object to the collaborative management of a plurality of target management objects, effectively avoiding the blowout of early warning information.

[0074] In one case of the embodiment, the pre-processed waste image in the box is recognized based on a U-Net semantic segmentation network and a bidirectional LSTM neural network, and combined with the surrounding instantaneous passenger flow data for analysis to generate the space-time management load entropy of each target management object, and further comprising:

[0075] The pre-processed waste image in the box is recognized based on a U-Net semantic segmentation network to generate a container load fluctuation spectrum, specifically including: for the gray image of the pre-processed continuous waste image in the box, a single frame image is processed using a U-Net semantic segmentation network: wherein the encoder is composed of 4 layers of module convolution, each layer containing a 3x3 convolution kernel (step 1, padding 1) to extract edge and texture features, cooperating with a 2x2 max pooling layer for down-sampling; the decoder is symmetrically provided with 4 layers of deconvolution modules, which recover the image size through up-sampling, while fusing the features of the corresponding layers of the encoder to enhance the recognition accuracy of the garbage and the box boundary; the output layer uses sigmoid operation to map the output value to the range of 0 to 1, and then generate a binary mask: the initial empty box image of the target management object is taken as the effective area in the box of the target management object, the garbage pixels in the effective area in the mask are marked as 1, and the pixels not occupied by garbage are marked as 0; the total number of pixels with a value of 1 in the single frame mask is divided by the total number of effective area pixels to obtain the container load rate, which ranges from 0 to 1, with 0 indicating an empty box and 1 indicating a complete overflow; images are collected at a frequency of 1 frame per 5 minutes, and 12 frames of data are obtained continuously for 60 minutes, corresponding to 12 container load rates, which are sorted by time to form a container load rate time sequence, and each data is time-stamped;

[0076] The container load rate time sequence is input into a bidirectional LSTM network: the forward LSTM processes from the 1st frame to the 12th frame in turn to capture the trend of garbage accumulation; the backward LSTM processes from the 12th frame to the 1st frame in reverse to find out the influence of the later state on the earlier state; the bidirectional LSTM network dynamically retains key changes through the forget gate, input gate and output gate; the container load rate difference between adjacent two frames is calculated: the container load rate of the latter frame is subtracted from the container load rate of the former frame to obtain the container load rate difference between adjacent two frames, if the difference is positive, it means that the garbage is increasing, if the difference is negative, it means that the garbage is decreasing after cleaning, and the larger the absolute value of the difference, the faster the change; taking 12 time points as the horizontal axis and the container load rate as the vertical axis, if the difference is positive, it is marked with red, if the difference is negative, it is marked with blue, and the larger the absolute value of the difference, the darker the color, and then a container load fluctuation spectrum is drawn;

[0077] The target management area is divided into four sub-areas, which are the front row, the back row, the channel and the middle area of the stand;

[0078] Based on the position information of the target management object, the path distance of the target management object and each sub-region is calculated, and the path distance is fused with the surrounding instantaneous passenger flow of the target management object to generate a spatial attraction mode value representing the probability of the target management object being preferentially selected for delivery, specifically including: taking the center of each sub-region as the starting point and the geographic coordinates of the target management object as the ending point, the shortest path distance is calculated by avoiding seat arrangement obstacles and only along passable routes such as corridors and aisles; the number of moving people in each sub-region is multiplied by 0.6 plus the average value of the passenger flow density to obtain the passenger flow weight; for each sub-region, the passenger flow weight is divided by the corresponding shortest path distance to obtain a contribution value, and the four contribution values are added to generate a spatial attraction mode value representing the probability of the target management object being preferentially selected for delivery.

[0079] Among them, due to the garbage disposal demand of the stadium stand, it is mainly generated by the moving crowd, because the audience moves during the match interval and the halftime break, and may discard garbage, so the number of moving people is the core driving factor of the selection probability of the target management object, so the weight is high. 0.6; and the passenger flow density means that the crowd is more concentrated, but needs to be combined with the number of moving people to be effective, and its role is to correct the deviation of only focusing on the number of moving people, and to reflect the feature that the crowd in the crowded area is more inclined to nearby disposal, so as a supplement to the core factor, the weight is low. 0.4;

[0080] The stage of the stadium event is obtained, and a time pressure factor is assigned to different stages, specifically including: the stadium event stage is divided into three stages, namely the event is in progress, the event is in halftime, and the event is over; the time pressure factor is set to 0.6 in the event in progress stage, because at this time most of the audience are sitting to watch the game, the number of moving people is small, and the increment of garbage is slow; the time pressure factor is set to 1.2 in the event halftime stage, because at this time most of the audience get up and move, and may discard garbage when moving, so the increment of garbage is fast; the time pressure factor is set to 1.5 in the event over stage, because at this time the audience is concentrated and leaves, and the garbage carried needs to be discarded, so the increment of garbage will increase sharply.

[0081] In one case of the embodiment, the pre-processed garbage image in the box is recognized based on the U-Net semantic segmentation network and the bidirectional LSTM neural network, and the surrounding instantaneous passenger flow data is analyzed to generate the spatiotemporal management load entropy of each target management object, which also includes:

[0082] The time series analysis is performed on the container load fluctuation spectrum to generate an instantaneous change gradient. The instantaneous change gradient is combined with the spatial gravity mode value and the time series pressure factor to generate a load surge expectation vector. Specifically, the time series sequence of 12 container load rates in the container load fluctuation spectrum is analyzed by using a 3-frame sliding window method. The container load rates of frames 1-3 are summed and divided by 3 to obtain the mean value of window 1. The container load rates of frames 2-4 are summed and divided by 3 to obtain the mean value of window 2. The process is sequentially slid to frames 10-12 to obtain 10 window mean values. The instantaneous change gradient is obtained by subtracting the mean value of the previous window from the mean value of the subsequent window and dividing the result by the window interval (the window interval is 5 minutes). Each instantaneous change gradient corresponds to a time stamp. The instantaneous change gradient is multiplied by the event stage time series pressure factor corresponding to the time stamp and then multiplied by the spatial gravity mode value to obtain the load surge value at a single time point. The 10 load surge values are arranged in the order of the time stamps to generate a load surge expectation vector.

[0083] The upper limit of the physical volume of the target management object is obtained.

[0084] Based on the container load fluctuation spectrum, the load surge expectation vector and the physical volume upper limit are fused to generate a space-time management load entropy. Specifically, the mean value of the load surge values at the current time point and the subsequent two time points in the load surge expectation vector is calculated. The container load rate at the current time point in the container load fluctuation spectrum is added to the mean value of the load surge values to obtain an expected container load rate. The expected container load rate is divided by the normalized physical volume upper limit to obtain the space-time management load entropy. The space-time management load entropy ranges from 0 to 1. The closer the value of the space-time management load entropy to 1, the higher the risk of overflow.

[0085] The U-Net semantic segmentation network is used to accurately extract the features of the garbage and the box boundary in the target management object, combined with the calculation of the container load rate, and the generated container load fluctuation spectrum clearly presents the dynamic changes of the container load of the target management object, which is convenient for dynamic management. The spatial gravity mode value is calculated by fusing the surrounding instantaneous passenger flow and the path distance, which accurately judges the probability of the target management object being put, optimizes the demand management of the target management object, and improves the efficiency of the daily management of the target management object.

[0086] The 3-frame sliding window method is used to generate the instantaneous change gradient, which is combined with the time series pressure factor and the spatial gravity mode value to generate the load surge expectation vector, which can predict the trend of the load change of the target management object in advance. The space-time management load entropy is generated by fusing the physical volume upper limit of the target management object, which directly reflects the overflow risk and avoids the confusion of warning and management caused by the simultaneous overflow of multiple target management objects, thereby improving the overall management level of the target management object in the grandstand area of the large-scale event.

[0087] In one case of the present embodiment, the space-time management load entropy of all target management objects is analyzed to generate a regional load situation tensor, including:

[0088] The space grid mapping of the space-time management load entropy of all target management objects is performed, and the load superposition coefficient of the target management objects in each grid unit is calculated in combination with the space attribute of the sub-region to which each target management object belongs, specifically including: dividing the stand area of the gymnasium into grid units according to 5m*5m, comparing the coordinates of each target management object with the grid coordinate range, and entering the unique grid in which the target management object is located, to complete the space grid mapping of all target management objects; the sub-regions are weighted according to the space attribute of the sub-regions, wherein if the sub-region belongs to a passageway, the weight is 0.3, if the sub-region belongs to a middle region, the weight is 0.3, if the sub-region belongs to a front row, the weight is 0.2, and if the sub-region belongs to a back row, the weight is 0.2; for each grid, the space-time management load entropy of all target management objects in the grid is counted, each space-time management load entropy is multiplied by the weight of the sub-region to which the target management object belongs, and then all the product results are added, and the sum is the load superposition coefficient of the grid;

[0089] Wherein, because the passageway is the only dynamic line of the audience to and from the seats, the selling point and the toilet, the number of moving people is the most, and the probability of garbage disposal when the personnel stay is extremely high, which is the most intensive area of garbage generation, and contributes most to the grid load, so the weight is high at 0.3; the middle region of the stand is the core seating area of the audience, with the highest proportion of people, although most people sit during the game, but the demand for garbage disposal is large when standing up during the middle of the game or when leaving, so the weight is high at 0.3; the number of front rows of the stand is less than the middle region, and is mostly close to the edge of the stand, far from the exit and the passageway, with a lower moving frequency than the middle region, and less garbage generation, so the weight is low at 0.2; the audience density of the back row of the stand is the lowest, although it may be close to the exit, but the degree of concentration of people flow is not high, the number of moving people and the demand for disposal are the least, and the contribution to the grid load is the least, so the weight is low at 0.2;

[0090] Based on the time sequence division of the gymnasium event stage, the load superposition coefficients of each grid unit are dynamically correlated and analyzed across time periods to generate a regional load time sequence linkage index, specifically including: dividing the time period of 30 minutes according to the event stage, each time period may contain 1-2 event stages, such as containing the event in progress or the event halftime, or containing the event halftime and the end of the event, which needs to be matched according to the actual event length; for each grid, the load superposition coefficients of adjacent two time periods are selected, and the difference between the load superposition coefficient of the latter time period and the load superposition coefficient of the former time period is calculated; the difference value is multiplied by the average value of the time sequence pressure factors of the two time periods to obtain the time period linkage value of a single grid; the average value of the time period linkage values of all grids is calculated, and the average value is the regional load time sequence linkage index;

[0091] The load superposition coefficient and the regional load time sequence linkage index are fused, and the distribution density characteristics of the target management object in the grid unit are combined to generate a regional load situation tensor, specifically including: dividing the number of target management objects in each grid by the grid area to obtain the distribution density of the target management object; multiplying the grid load superposition coefficient by 0.5, multiplying the regional load time sequence linkage index by 0.3, and multiplying the distribution density by 0.2 to obtain the load comprehensive value of the grid, and arranging the load comprehensive values of all grids in order according to the three dimensions of spatial x coordinate-spatial y coordinate-time stamp to generate a regional load situation tensor; wherein the grid load superposition coefficient is calculated based on the spatiotemporal management load entropy of the target management object in the grid and the sub-region weight, which reflects the actual garbage load of the current grid, so the weight is high, that is, 0.5; and the regional load time sequence linkage index mainly reflects the cross-period load change trend, realizes early warning, but the essence is a trend index rather than the current actual load, so the weight is low, that is, 0.3; the distribution density of the target management object indicates that there are more target management objects in the grid, which can share the garbage pressure, but it does not directly reflect the garbage amount, and the influence on early warning is relatively secondary, so the weight is low, that is, 0.2.

[0092] By dividing the stand into 5m*5m grids, the target management object is accurately assigned to the corresponding grid, the grid load superposition coefficient is calculated combined with the sub-region spatial attribute weight, and then the load contribution of the target management object in different regions is understood, so that the regional management is more targeted. Then, the period is divided according to the event stage, the cross-period load superposition coefficient is analyzed, and then the regional load time sequence linkage index is generated, which can capture the dynamic change trend of the load. Finally, the load superposition coefficient, the time sequence linkage index and the distribution density of the target management object are fused to generate a three-dimensional regional load situation tensor of space-time-load, realizing the transformation of the target management object management from single point isolation to collaborative management, and improving the overall management efficiency.

[0093] In one case of the embodiment, a neural network algorithm based on a self-attention mechanism calculates the regional load situation tensor to generate a dynamic mutual construction weight management matrix of the target management object in the target management region, including:

[0094] extracting load dynamic characteristics of each target management object from the regional load situation tensor, analyzing the influence probability of load changes of different target management objects on surrounding target management objects, and generating load-induced potential, specifically including: finding the load comprehensive value of the grid where each target management object is located from the regional load situation tensor, arranging it in time stamp order to form the load dynamic characteristic sequence of the target management object; taking a single target management object as the center, setting 10 meters as the surrounding threshold, screening other target management objects within a distance of ≤10 meters, and the other target management objects are the surrounding objects; assigning influence probability according to distance: 0-5 meters influence probability is 0.8, 5-10 meters influence probability is 0.4; calculate the difference value of adjacent time stamps in the load dynamic characteristic sequence, multiply the difference value by the influence probability of the target management object on the surrounding objects, and obtain the load-induced potential of the target management object on the surrounding objects;

[0095] Based on the neural network of self-attention mechanism, taking the load-induced potential as the core basis of attention allocation, calculating the asymmetric correlation weight formed between different target management objects due to the difference in load-induced potential, generating an induced correlation matrix, specifically including: setting each target management object as a query object, and setting the surrounding objects within 10 meters of the query object as key objects, and the load-induced potential of the query object and the key object as the attention value of the self-attention mechanism; for the query object, add the load-induced potential of all key objects to obtain the total induced potential; divide the load-induced potential of the query object to a single key object by the total induced potential of the query object to obtain the normalized weight; because the load-induced potential of the query object to each key object is different, the normalized weight at this time reflects the asymmetry; taking the query object as the row and the key object as the column to construct the matrix, and filling each weight into the corresponding position, and if it is a non-surrounding target management object, the weight is filled with 0, to generate the induced correlation matrix;

[0096] Based on the position information of the target management object and the sub-region to which it belongs, the spatial steric hindrance degree between each target management object is calculated, specifically including: taking the coordinates of two target management objects as the starting point and the end point, the path needs to avoid seat arrangement obstacles, and only along the passable routes such as corridors and aisles, the shortest path distance is calculated; set the hindrance coefficient according to the cross-domain situation of the sub-region, if the two target management objects belong to one sub-region, the hindrance coefficient is 1, if the two target management objects belong to different sub-regions, the hindrance coefficient is 2; multiply the shortest path distance by the hindrance coefficient to obtain the spatial steric hindrance degree of the two target management objects;

[0097] According to the degree of steric hindrance, the induced correlation matrix is corrected to reduce false correlation, and a space tuning correlation matrix is generated, which specifically includes: taking the maximum value in all degrees of steric hindrance as a threshold; for the weight of the induced correlation matrix, for two target management objects, if the degree of steric hindrance is less than or equal to the threshold, then the attenuation coefficient is obtained by subtracting the degree of steric hindrance from 1 and dividing by the threshold; if the degree of steric hindrance is greater than the threshold, the attenuation coefficient is 0; the weight in the induced correlation matrix is multiplied by the corresponding attenuation coefficient to obtain the correlation weight; all correlation weights are arranged according to the original matrix row and column to generate a space tuning correlation matrix;

[0098] The time sequence sensitivity coefficient is obtained by analyzing the dynamic change rate of the time sequence pressure factor and the load induced potential, specifically including: calculating the difference value of the time sequence pressure factor of adjacent event periods, and dividing the difference value by the time interval to obtain the time sequence change rate; calculating the difference value of the load induced potential of adjacent time stamps, and dividing the difference value by the time interval to obtain the load change rate; multiplying the time sequence change rate by 0.6 and the load change rate by 0.4 to obtain the time sequence sensitivity coefficient; wherein the time sequence change rate is directly related to the event stage, and the event stage determines the audience moving frequency, which is the root cause of the change of the amount of garbage, and is the core driving the dynamic change of the load, so the weight is high, which is 0.6; and the load change rate reflects the change of the load influence between the target management objects, which is secondary and needs the initial load change driven by the time sequence pressure factor to produce the induced correlation between the target management objects, which belongs to local response, so the weight is low, which is 0.4.

[0099] In one case of the embodiment, the neural network algorithm based on the self-attention mechanism calculates the regional load situation tensor to generate a dynamic mutual construction weight management matrix of the target management objects in the target management region, and further includes:

[0100] Based on the time sequence sensitivity coefficient, the time decay characteristics of each correlation weight in the space tuning correlation matrix are adjusted to make the correlation relationship change dynamically with the progress of the event, and a time-space covariant correlation matrix is generated, specifically including: multiplying the time sequence sensitivity coefficient by the time interval of the current event to obtain a time decay coefficient; then multiplying each correlation weight in the space tuning correlation matrix by (1-time decay coefficient) to obtain the weight after time adjustment; finally, the weights are arranged according to the original matrix row and column to generate a time-space covariant correlation matrix;

[0101] The multi-dimensional feature enhancement is performed on the time-space covariant correlation matrix through the multi-head attention of the self-attention mechanism, the correlation characteristics between the high-induced potential target management object and the high-load target management object are analyzed, and a characteristic highlighted correlation matrix is generated, specifically including: three independent attention heads are set, which focus on three dimensions of high-induced potential, high-load and time-space cooperation: the first attention head selects the objects with the load-induced potential of the top 30% as the query objects, and the rest as the key objects; the second attention head selects the objects with the top 30% of the space management load entropy as the query objects, and the rest as the key objects; the third attention head takes the target management object with high time-space cooperation, that is, low spatial steric hindrance and high time sequence sensitivity coefficient, as the query object, and the rest as the key object, wherein the average of the spatial steric hindrance between all pairs of target management objects in the target management area is calculated, and the average is taken as the threshold value, if the spatial steric hindrance is lower than the threshold value, it is low spatial steric hindrance; the sum of the time sequence sensitivity coefficients of all pairs of target management objects in the target management area is calculated, and the sum is taken as the threshold value, if the time sequence sensitivity coefficient is higher than the threshold value, it indicates that the time sequence sensitivity coefficient is high;

[0102] For the first attention head, the load-induced potential is taken as the characteristic value, for the second attention head, the space management load entropy is taken as the characteristic value, and for the third attention head, the spatial steric hindrance is divided by the time sequence sensitivity coefficient, and the characteristic value is obtained; for the query object and the key object of the three attention heads, the matching degree of each attention head is obtained according to 1 minus the absolute value of the characteristic value of the query object minus the characteristic value of the key object, and then divided by max (the characteristic value of the query object, the characteristic value of the key object);

[0103] For the query object and the key object of the three attention heads, the sum of the matching degrees of the query object to all key objects is obtained, and then the single matching degree is divided by the sum to obtain the attention weight of each attention head; the attention weight is multiplied by the weight of the time-space covariant correlation matrix to obtain the enhanced weight, and the enhanced weight is arranged according to the original matrix row and column to generate the enhanced sub-matrix of each attention head;

[0104] Among them, the row and column numbers of the three enhanced sub-matrices are consistent, and each same row-column position corresponds to an element of the three enhanced sub-matrices; for each row-column position, the elements of the three enhanced sub-matrices at this position are packaged into a three-dimensional vector; the three elements of the three-dimensional vector are added to obtain the final value of the position; all final values (elements) of all positions are arranged according to the original matrix row and column to generate the characteristic highlighted correlation matrix;

[0105] The load superposition coefficient of each sub-region is analyzed and the distribution of the surrounding instantaneous passenger flow data is analyzed to determine the correlation adaptation threshold of each sub-region, specifically including: for each sub-region, the load superposition coefficient of the sub-region in the past 10 events and the moving number in the corresponding period are found respectively; find all the load superposition coefficients corresponding to the period with the most moving number, sort these load superposition coefficients from high to low according to the value, and take the load superposition coefficient with the highest frequency as the correlation adaptation threshold of the sub-region;

[0106] Based on the correlation adaptation threshold, the feature highlighting correlation matrix is corrected to generate a dynamic mutual construction weight management matrix of the target management object in the target management area, specifically including: determining the sub-region corresponding to each element in the feature highlighting correlation matrix; compare the correlation adaptation threshold of the corresponding sub-region with the element, if the element > correlation adaptation threshold, multiply the element by (threshold ÷ element) to get the corrected element; if the element ≤ correlation adaptation threshold, the element value is retained; then arrange all elements according to the rows and columns of the original matrix, and the dynamic mutual construction weight management matrix can be generated.

[0107] By extracting the load dynamic characteristics of the target management object from the regional load state tensor, combining the 10-meter surrounding threshold and the segmented influence probability to generate the load induction potential, the load influence relationship between the target management objects is determined, and then the asymmetric induction correlation matrix is generated with the load induction potential as the core., combined with the space steric degree to correct false correlation, cooperate with the time sequence sensitivity coefficient to adjust the weight time decay characteristics, let the target management object correlation management dynamically adapt with the progress of the event, and through the three independent attention heads to strengthen the features, generate the dynamic mutual construction weight management matrix, avoid the accumulation of early warning information, and significantly improve the overall management efficiency and accuracy of the target management object of the large-scale event stand.

[0108] In one case of the embodiment, the dynamic mutual construction weight management matrix and the space-time management load entropy are dynamically calculated to obtain the management index of each target management object, including:

[0109] Based on the dynamic mutual construction weight management matrix, the space-time management load entropy of each target management object is calculated by bidirectional mapping to generate the coupling interaction strength, specifically including: for a single target management object, determine its row and column positions in the dynamic mutual construction weight management matrix; multiply all element values in the row where the target management object is located by the space-time management load entropy of the corresponding column target management object, and sum all the products to obtain the action party strength; multiply all element values in the column where the target management object is located by the space-time management load entropy of the corresponding row object, and sum all the products to obtain the acted party strength; add the action party strength and the acted party strength to obtain the coupling interaction strength of the target management object;

[0110] The residual physical volume of the target management object is obtained in real time;

[0111] The coupling interaction strength and the residual physical volume are calculated to generate a load criticality ratio, specifically including: dividing the residual physical volume of the target management object by the total physical volume of the target management object to obtain a normalized value, ranging from 0 to 1; dividing the coupling interaction strength of the target management object by the normalized value to obtain the load criticality ratio;

[0112] The time sequence change of the load criticality ratio is analyzed to generate a load criticality increase rate, specifically including: recording the load criticality ratio of the target management object at an interval of 5 minutes and forming a time sequence; calculating the difference of the load criticality ratio at adjacent two time points; dividing the difference by the time interval to obtain an increase rate per unit time, which is the load criticality increase rate;

[0113] The load criticality ratio and the load criticality increase rate are fused to calculate a management index of each target management object, specifically including: setting the weight of the load criticality ratio as 0.6 and the weight of the load criticality increase rate as 0.4; adding the load criticality ratio and the load criticality increase rate multiplied by the corresponding weights to obtain the management index of each target management object; wherein the load criticality ratio reflects the current load state of the target management object and is the core basis for determining whether immediate treatment is needed, so the weight is high at 0.6; while the load criticality increase rate only reflects the change trend of the load of the target management object, which is a potential risk, so the weight is low at 0.4.

[0114] Through dynamic inter-constructing weight management matrix, bidirectional mapping calculation is performed on each target management object to generate coupling interaction strength, and the load correlation between target management objects is determined, and the load criticality ratio is calculated by combining the real-time acquired residual physical volume of the target management object, reflecting the current load state of the target management object, and the load criticality ratio is recorded at an interval of 5 minutes to generate a time sequence, and the load criticality increase rate is calculated, and then the load change trend can be captured, and the management index is generated to realize the change of the target management object from isolated management to hierarchical precise management, avoid the accumulation of early warning information, and improve the overall management efficiency and pertinence of the target management object of the grand event stand.

[0115] In one case of the embodiment, the management index and the position information of all target management objects are analyzed to generate an optimal sequence of target management object driving management, including:

[0116] The management index and the position information of each target management object are analyzed, and the potential level-topological coupling value is obtained by combining the sub-regions, specifically including: taking the reciprocal of the shortest path distance from the target management object to the center of the sub-region as the topological correlation coefficient; multiplying the management index of the target management object by the weight of the sub-region and then by the topological correlation coefficient to obtain the potential level-topological coupling value;

[0117] The distribution characteristics of the potential level-topological coupling value are analyzed to generate a scheduling conduction dependency between adjacent objects, specifically including: for the potential level-topological coupling value of the target management object and the adjacent peripheral object, the minimum potential level-topological coupling value is divided by the maximum potential level-topological coupling value to obtain a coupling ratio; 1 is subtracted from (the steric hindrance degree of the target management object and the adjacent peripheral object divided by the maximum value of all steric hindrance degrees) to obtain an attenuation coefficient; the attenuation coefficient is multiplied by the coupling ratio to obtain the scheduling conduction dependency between adjacent objects.

[0118] Based on the scheduling conduction dependency, the potential level-topological coupling value is corrected to generate a dynamic scheduling potential, specifically including: calculating the mean of all scheduling conduction dependencies, and taking the mean as a neighboring influence comprehensive coefficient; the potential level-topological coupling value of the target management object is multiplied by (1+the neighboring influence comprehensive coefficient) to obtain the dynamic scheduling potential of the target management object.

[0119] Based on the dynamic scheduling potential, the convenience of the cleaning and transporting path of the sub-region division is analyzed to obtain the resource reachability adaptation degree of each target management object, specifically including: calculating the shortest path distance from the target management object to the entrance of the sub-region, avoiding seat obstacles, and calculating along the channel, and taking the reciprocal of the shortest path distance as the path convenience coefficient; the dynamic scheduling potential of the target management object is multiplied by the path convenience coefficient, and then multiplied by the weight of the sub-region to obtain the resource reachability adaptation degree of the target management object.

[0120] The dynamic scheduling potential and the resource reachability adaptation degree are fused and sorted to generate an optimal sequence driven management sequence, specifically including: multiplying the dynamic scheduling potential by 0.6 and the resource reachability adaptation degree by 0.4 to obtain a comprehensive scheduling value; sorting the comprehensive scheduling values of all target management objects from large to small to obtain the optimal sequence driven management sequence; wherein the dynamic scheduling potential mainly reflects whether the target management object needs to be processed, which is the primary goal of the target management object management, so a high weight of 0.6 is given; and the resource reachability adaptation degree only reflects whether the cleaning and transporting path is convenient, which belongs to efficiency optimization, so a low weight of 0.4 is given.

[0121] The potential level-topological coupling value is calculated through the management index of the target management object, the weight of the sub-region and the topological correlation coefficient, so that the management of the target management object is more targeted, and the dynamic scheduling potential is used to make the management fully consider the synergistic effect of adjacent target management objects, and then the resource reachability adaptation degree is combined to calculate the comprehensive scheduling value from the dynamic scheduling potential and the resource reachability adaptation degree, and the optimal sequence driven management sequence generated by sorting the comprehensive scheduling value, and then the ordered management of the target management object can be realized.

[0122] In one case of the embodiment, the global optimization sequence of the priority-driven management sequence is managed and the regional cluster division management is managed, and the cooperative early warning management instruction of the target management object is generated, including:

[0123] The dynamic scheduling potential and resource reach adaptation degree of each target management object in the priority-driven management sequence are analyzed to obtain the global scheduling influence entropy of each object, specifically including: the mean value of the dynamic scheduling potential and the mean value of the resource reach adaptation degree of all target management objects are calculated respectively; for each target management object, the absolute deviation of the dynamic scheduling potential and the mean value, and the absolute deviation of the resource reach adaptation degree and the mean value are calculated respectively; the absolute deviation of the dynamic scheduling potential×0.6+the absolute deviation of the resource reach adaptation degree×0.4 is obtained as the fusion deviation; the global scheduling influence entropy of the target management object is obtained by dividing the fusion deviation by the maximum fusion deviation of all target management objects;

[0124] Based on the global scheduling influence entropy and the sub-region, the priority-driven management sequence is clustered, and the intra-cluster cooperative efficiency of each cluster is calculated to generate a load cooperative cluster, specifically including: the target management objects are grouped according to the sub-region, wherein the channel and the intermediate region are a group, and the front and rear regions are a group; the standard deviation of all global scheduling influence entropies in each group is calculated, and the standard deviation is multiplied by 0.5 to obtain a similar threshold; the target management objects with a global scheduling influence entropy less than the similar threshold are grouped into a cluster to form an initial sub-region cluster, and the target management objects with a global scheduling influence entropy greater than or equal to the similar threshold are not grouped; for each initial sub-region cluster, the mean value of the dynamic scheduling potential and the mean value of the resource reach adaptation degree of all target management objects in the cluster are calculated, and the mean value of the dynamic scheduling potential is multiplied by the mean value of the resource reach adaptation degree to obtain the intra-cluster cooperative efficiency; the clusters with an intra-cluster cooperative efficiency≥0.5 are retained as the load cooperative cluster, and the clusters with an intra-cluster cooperative efficiency<0.5 are not retained;

[0125] The intra-cluster cooperative efficiency of each load cooperative cluster and the load critical speed increment of the intra-cluster target management object are fused to obtain the cluster-level early warning priority, specifically including: the minimum value of the intra-cluster cooperative efficiency of all clusters and the minimum value of the corresponding load critical speed increment in the cluster are taken as the cooperative efficiency reference and the critical speed increment reference respectively; for a single cluster, the intra-cluster cooperative efficiency is subtracted from the cooperative efficiency reference to obtain the cooperative deviation; the load critical speed is subtracted from the critical speed increment reference to obtain the speed deviation; the cooperative deviation and the speed deviation are added to obtain the cluster-level early warning score; the cluster-level early warning scores of all clusters are sorted from large to small, the higher the score, the higher the cluster-level early warning priority, so that the cluster-level early warning priority can be obtained;

[0126] The cluster-level early warning priority is used for globally sorting the load cooperative clusters, and a cooperative early warning management instruction of the target management object is generated in combination with the spatial position of each cluster, specifically including: sorting the load cooperative clusters in descending order of cluster-level early warning priority; calculating the shortest path of the adjacent cluster center for the sorted clusters to generate a clean-up route; combining the cluster-level early warning priority, the clean-up route and the position of the target management object to generate the cooperative early warning management instruction of the target management object, which includes the position of the target management object in the cluster, the clean-up route and the early warning priority;

[0127] The global scheduling influence entropy of the target management object is used to provide a management basis for cluster division, and then the initial sub-regional clusters are divided according to the grouping of the sub-regions and the similar threshold value, the load cooperative clusters with a cooperative efficiency greater than or equal to 0.5 are retained, then the clusters are sorted according to the cluster-level early warning priority, and the cooperative early warning management instruction containing the position of the target management object in the cluster, the clean-up route and the early warning priority is generated in combination with the spatial position, so as to avoid disorderly accumulation or blowout of early warning information, and then resources can be efficiently deployed to improve the overall management efficiency of the target management objects in the stand area of the large-scale event.

[0128] In one case of the embodiment, a fruit peel box overflow early warning management system is applied to the fruit peel box overflow early warning management method described above, and includes:

[0129] The recognition unit is configured to acquire the in-box garbage image, the surrounding instantaneous passenger flow data and the position information of the target management objects in the target management area in real time, identify the preprocessed in-box garbage image based on the U-Net semantic segmentation network and the bidirectional LSTM neural network, analyze the surrounding instantaneous passenger flow data, and generate the spatio-temporal management load entropy of each target management object. The target management area is the stand area of a stadium, and the target management object is a fruit peel box.

[0130] The calculation unit is configured to analyze the spatio-temporal management load entropy of all target management objects, generate a regional load situation tensor, and calculate the regional load situation tensor based on a neural network algorithm of a self-attention mechanism to generate a dynamic mutual construction weight management matrix of the target management objects in the target management area.

[0131] The driving analysis unit is configured to dynamically calculate the dynamic mutual construction weight management matrix and the spatio-temporal management load entropy to obtain the management index of each target management object.

[0132] The management optimization unit is configured to analyze the management index and the position information of all target management objects to generate an optimal sequence driving management sequence of the target management objects.

[0133] The early warning management unit is configured to globally optimize and sort manage the optimal sequence driving management sequence and divide the regional clusters to generate a cooperative early warning management instruction of the target management objects.

[0134] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since numerous changes, modifications, substitutions and variations can be made thereto without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A method for early warning management of overflowing fruit peel bins, characterized in that, include: Step S1: In real time, acquire images of garbage inside bins, instantaneous passenger flow data and location information of multiple target management objects within the target management area. Based on the U-Net semantic segmentation network and bidirectional LSTM neural network, identify the preprocessed garbage images inside bins and analyze them in combination with the instantaneous passenger flow data of the surrounding area to generate the spatiotemporal management load entropy of each target management object. The target management area is the grandstand area of ​​the stadium, and the target management object is the garbage bin. The generation of the spatiotemporal management load entropy includes: identifying the preprocessed garbage image inside the container based on the U-Net semantic segmentation network to generate the load fluctuation spectrum; The generation of the capacity fluctuation spectrum includes: processing a single frame of garbage image in a bin using a semantic segmentation network with an encoder-decoder structure to obtain a binarized mask of the garbage area, calculating the capacity rate based on the binarized mask to form a capacity rate time series, performing bidirectional time series analysis on the capacity rate time series using a bidirectional long short-term memory network to capture the changing trend and mutual influence of the capacity rate, and generating the capacity fluctuation spectrum. Step S2: Analyze the spatiotemporal management load entropy of all target management objects, generate the regional load status tensor, and calculate the regional load status tensor based on the self-attention mechanism neural network algorithm to generate the dynamic mutual construction weight management matrix of target management objects within the target management area. The generation of the regional load situation tensor includes: spatially gridding the spatiotemporal management load entropy of all target management objects, and calculating the load superposition coefficient of the target management objects in each grid cell by combining the spatial attributes of the sub-region to which each target management object belongs; Based on the time sequence division of the stadium's event phases, a cross-time dynamic correlation analysis is performed on the load superposition coefficient of each grid unit to generate a regional load time sequence linkage index. The load superposition coefficient and the regional load time-series linkage index are fused together, and the distribution density characteristics of the target management object in the grid cell are combined to generate the regional load situation tensor. Step S3: Dynamically calculate the dynamic mutual construction weight management matrix and the spatiotemporal management load entropy to obtain the management index of each target management object; The calculation process of the management index includes: based on the dynamic mutual construction weight management matrix, performing bidirectional mapping calculation on the spatiotemporal management load entropy of each target management object to generate the coupling interaction strength; Real-time acquisition of the remaining physical volume of the target managed object; The coupling interaction strength and the remaining physical volume are calculated to generate the critical load ratio; Analyze the time-series changes of the critical load ratio to generate the critical load growth rate; The critical load ratio and critical load growth rate are combined and calculated to generate a management index for each target management object. Step S4: Analyze the management index and location information of all target management objects to generate a priority-driven management sequence for the target management objects; Step S5: Perform global optimization sorting management and regional cluster division management on the priority-driven management sequence to generate collaborative early warning management instructions for the target management object.

2. The method for early warning management of overflowing fruit peel bins according to claim 1, characterized in that, Based on the U-Net semantic segmentation network and bidirectional LSTM neural network, the preprocessed images of waste inside the bins are identified and analyzed in conjunction with surrounding instantaneous passenger flow data to generate the spatiotemporal management load entropy for each target management object. This also includes: The target management area is divided into four sub-areas; Based on the location information of the target management object, the path distance between the target management object and each sub-region is calculated, and then fused with the instantaneous passenger flow around the target management object to generate a spatial gravity mode value representing the probability that the target management object is preferentially selected for deployment. The generation of the spatial gravity mode value includes: taking the center of each sub-region as the starting point and the geographical coordinates of the target management object as the ending point, calculating the shortest path distance; combining the number of people moving in each sub-region with the passenger flow density to obtain the passenger flow weight; dividing the passenger flow weight by the corresponding shortest path distance to obtain the contribution value; and adding the four contribution values ​​to obtain the spatial gravity mode value. Acquire the event phases of the stadium and assign time-series pressure factors to different phases.

3. The method for early warning management of overflowing fruit peel bins according to claim 2, characterized in that, Based on the U-Net semantic segmentation network and bidirectional LSTM neural network, the preprocessed images of waste inside the bins are identified and analyzed in conjunction with surrounding instantaneous passenger flow data to generate the spatiotemporal management load entropy for each target management object. This also includes: A time series analysis of the load fluctuation spectrum is performed to generate an instantaneous gradient. The instantaneous gradient is then combined with the spatial gravitational mode value and the time series pressure factor to generate a load surge prediction vector. Get the maximum physical volume of the target managed object; Based on the load fluctuation spectrum, the spatiotemporal management load entropy is generated by integrating the load surge expectation vector and the physical volume limit.

4. The method for early warning management of overflowing fruit peel bins according to claim 1, characterized in that, A neural network algorithm based on a self-attention mechanism calculates the regional load situation tensor to generate a dynamic mutual weight management matrix for target management objects within the target management area, including: Extract the load dynamic characteristics of each target management object from the regional load situation tensor, analyze the probability of the impact of load changes of different target management objects on surrounding target management objects, and generate load induced potential; The neural network based on the self-attention mechanism uses the load-induced potential as the core basis for attention allocation, calculates the asymmetric correlation weights formed by the difference in load-induced potential between different target management objects, and generates the induced correlation matrix. Based on the location information of the target management object and its sub-region, calculate the spatial resistivity between each target management object; The induced correlation matrix is ​​corrected based on the spatial potential resistance to reduce spurious correlations and generate a spatially tuned correlation matrix. The time-series pressure factor and the dynamic rate of change of the load-induced potential are analyzed to obtain the time-series sensitivity coefficient.

5. The method for early warning management of overflowing fruit peel bins according to claim 4, characterized in that, A neural network algorithm based on a self-attention mechanism calculates the regional load situation tensor, generates a dynamic mutual weight management matrix for target management objects within the target management area, and also includes: Based on the time sensitivity coefficient, the time decay characteristics of each correlation weight in the spatially tuned correlation matrix are adjusted so that the correlation changes dynamically with the progress of the event, thereby generating a time-space covariant correlation matrix. Multi-head attention using a self-attention mechanism is used to enhance the multi-dimensional features of the temporal-spatial covariant correlation matrix, analyze the correlation features between high-inducibility target management objects and high-load target management objects, and generate a feature-highlighting correlation matrix. Analyze the load superposition coefficient of each sub-region and the distribution of instantaneous passenger flow data in the surrounding area to determine the correlation adaptation threshold of each sub-region; Based on the association adaptation threshold, the feature salience association matrix is ​​modified to generate a dynamic mutual construction weight management matrix for target management objects within the target management area.

6. The method for early warning management of overflowing fruit peel bins according to claim 1, characterized in that, Analyze the management index and location information of all target management objects to generate a priority-driven management sequence for the target management objects, including: The management index and location information of each target management object are analyzed, and combined with the sub-regions, the potential-topology coupling value is obtained; Analyze the distribution characteristics of potential-topology coupling values ​​to generate scheduling transit dependencies between adjacent objects; Based on the scheduling propagation dependency, the potential-topology coupling value is corrected to generate a dynamic scheduling potential; Based on dynamic scheduling potential, the convenience of the waste collection path in the sub-region division is analyzed to obtain the resource accessibility of each target management object; The dynamic scheduling potential and resource accessibility are integrated and sorted to generate a priority-driven management sequence.

7. The fruit peel bin overflow early warning management method according to claim 6, characterized in that, The priority-driven management sequence is subjected to global optimization sorting management and regional cluster partitioning management to generate collaborative early warning management instructions for the target management object, including: Analyze the dynamic scheduling potential and resource reachability of each target management object in the priority-driven management sequence to obtain the global scheduling influence entropy of each object; Based on the global scheduling influence entropy and sub-regions, the priority-driven management sequence is divided into clusters, and the intra-cluster synergy validity of each cluster is calculated to generate load synergy clusters. The cluster-level early warning priority is obtained by integrating the intra-cluster synergy effectiveness of each load synergy cluster with the critical load growth rate of the target management object within the cluster. The load coordination clusters are globally sorted according to the cluster-level early warning priority, and the coordination early warning management instructions for the target management objects are generated based on the spatial location of each cluster.

8. A waste bin overflow warning and management system, applied to the waste bin overflow warning and management method described in any one of claims 1-7, characterized in that, include: The identification unit is used to acquire in real time images of garbage inside bins, instantaneous passenger flow data and location information of multiple target management objects within the target management area. Based on the U-Net semantic segmentation network and bidirectional LSTM neural network, the preprocessed garbage images inside bins are identified and analyzed in combination with the instantaneous passenger flow data of the surrounding area to generate the spatiotemporal management load entropy of each target management object. The target management area is the grandstand area of ​​the stadium, and the target management object is the garbage bin. The calculation unit is used to analyze the spatiotemporal management load entropy of all target management objects, generate the regional load status tensor, and calculate the regional load status tensor based on the self-attention mechanism neural network algorithm to generate the dynamic mutual construction weight management matrix of target management objects within the target management area. The driving analysis unit is used to dynamically calculate the dynamic mutual construction weight management matrix and the spatiotemporal management load entropy to obtain the management index of each target management object; The management optimization unit is used to analyze the management index and location information of all target management objects and generate a priority-driven management sequence for the target management objects. The early warning management unit is used to perform global optimization sorting management and regional cluster division management of the priority-driven management sequence, and generate collaborative early warning management instructions for the target management object.

Images