A store risk analysis method and system based on Internet of Things roller shutter door data

Abstract

The application discloses a shop risk analysis method and system based on Internet of Things roll-up door data, which comprises the following steps: obtaining a shop roll-up door opening and closing event sequence, constructing a multi-dimensional state feature vector containing a first opening time and a total business duration, extracting a multi-scale residual feature tensor after preprocessing and time series decomposition; adopting a hybrid deep neural network for unsupervised training to learn a normal operation mode; calculating a standardized abnormal score based on adaptive weighted fusion of multi-scale reconstruction errors, combining a number of opening and closing events with a quantity decay factor and a trend decay amplitude, and outputting a shop closure risk probability; and generating a regional industry state index through spatial distance weighted aggregation. The application realizes an Internet of Things roll-up door data analysis method and system for presenting shop risks through roll-up door data in a data-driven manner.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) data processing technology, and in particular to a method and system for store risk analysis based on IoT roller shutter door data. Background Technology

[0002] Assessing the operational status and closure risk of physical stores is crucial for business management, financial credit risk control, and regional economic monitoring. Traditional risk assessment methods primarily rely on publicly available business registration information, corporate financial statements, or third-party business survey data.

[0003] However, the above-mentioned existing technologies have the following obvious technical defects: (1) Information lag: The administrative process of deregistration or inclusion in the list of abnormal business operations has a significant delay, usually lagging behind the actual closure or bankruptcy of the store by several months, making it impossible to achieve early warning of risk events. (2) Difficulty in data acquisition: The financial data of small and micro stores are usually not disclosed to the public, and there is a lack of effective digital operation record methods, resulting in a lack of reliable quantitative analysis basis for risk assessment. (3) Reliance on subjective rules: Some analysis methods based on IoT data usually require the pre-setting of fixed time thresholds or business rules, such as "not opening the door for N consecutive days is judged as abnormal". These rules cannot adapt to the personalized operation mode of different stores and different industries, and the false alarm rate and false alarm rate are high. The data of roller shutters connected to the Internet of Things is a kind of data that is obviously related to the daily operation of the store. When the store is equipped with roller shutters, it will inevitably use roller shutters when opening or closing the door. The application of smart roller shutters in the industry is mostly still at the level of basic remote control, status monitoring and simple data statistics, without processing and analyzing IoT roller shutters.

[0004] Therefore, how to utilize the objective data generated by IoT roller shutters widely deployed in physical stores to construct a fully data-driven IoT roller shutter data analysis method and system that presents store risks through IoT roller shutter data, and present the current status of the store through IoT roller shutter data, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for store risk analysis based on IoT roller shutter door data. By acquiring roller shutter door data, processing and analyzing the data, and mining feature vectors from the roller shutter door opening and closing event sequences, the system analyzes and calculates the data to intuitively reflect the store's risk status. Based on a specified time interval, a specified geographical area, and a defined industry category, the system integrates all data to intuitively reflect the status of the same industry in the region. This achieves a fully data-driven IoT roller shutter door data analysis method and system that presents store risks through IoT roller shutter door data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for store risk analysis based on IoT roller shutter door data includes: S1. Obtain the sequence of roller shutter door opening and closing events for all stores within a specified time interval, a specified geographical area, and a specified industry category through the Internet of Things platform. Each opening and closing event is recorded as a tuple containing store identifier, timestamp, action type, and geographical coordinates. S2. For each store, extract complete switch event pairs from the roller shutter door switch event sequence, and construct a state feature vector in natural days. The state feature vector includes at least the first opening time, the last closing time, the total business hours, the number of switch event pairs, and the average time interval between switch events. S3. Preprocess the state feature vector, including missing value processing, outlier truncation and standardization, to obtain a standardized feature vector; S4. Construct a state feature time series from the standardized feature vectors of each store over multiple consecutive days, and decompose it into trend components, periodic components, and residual components using a decomposition method; based on the residual components, extract multi-scale time series feature tensors through multiple sliding windows of different time lengths. S5. Construct and train a hybrid deep neural network model based on variational autoencoder and long short-term memory network. The hybrid deep neural network model uses the multi-scale temporal feature tensor of stores in the same industry during the historical normal operation period as training data and adopts unsupervised learning to learn the feature distribution of the normal operation mode. S6. For the store and date to be analyzed, input the residual components of the corresponding time window into the trained hybrid deep neural network model, calculate the multi-scale reconstruction error, and obtain the standardized anomaly score of the store on that date through adaptive weighted fusion and standardization. S7. Based on the standardized anomaly score, combined with the decay factor of the number of switch events and the decay magnitude of the trend component, calculate the probability of store closure risk. S8. For stores in a specified geographical area and industry category, the probability of store closure risk of each store in the specified geographical area is aggregated by spatial distance and combined with the store size weight to calculate the regional industry status index, so as to represent the overall status of the specified geographical area and industry.

[0007] Furthermore, in step S2, the state feature vector constructed in calendar days includes seven feature components, namely the first opening time, the last closing time, the total operating time, the number of opening and closing event pairs, the average interval between opening and closing events, the fluctuation of opening time, and the fluctuation of closing time.

[0008] Further, in step S4, the time series decomposition method is a locally weighted seasonal trend decomposition method to decompose the state feature time series, and the relationship is: ,in, Let be the standardized feature vector of store s on day t. As a trend component, For periodic components, These are the residual components; the residual components are used in subsequent steps. Perform model training and anomaly detection.

[0009] Furthermore, in step S4, the multi-scale temporal feature tensor is constructed in the following way: Set three sliding windows with different time scales. , and These correspond to the weekly, bi-weekly, and monthly scales, respectively. For each timescale sliding window, a subsequence is extracted from the residual sequence with a fixed step size to form a feature window.

[0010] Furthermore, the hybrid deep neural network model in step S5 includes: The encoder, consisting of a multi-layer LSTM network, encodes the input feature window into a fixed-dimensional latent state vector; the variational inference layer is based on the mean vector and log-variance vector of the output latent variable distribution. The latent space sampling layer employs reparameterization techniques to sample latent variables from a normal distribution; The decoder, composed of an LSTM network, is used to reconstruct the input sequence from the latent variables to obtain the reconstruction window; The conformance loss function minimized during model training is: ,in, To reconstruct the mean squared error loss, The KL divergence regularization term is used to represent the potential distribution between the standard normal distribution. To reconstruct the temporal smoothing regularization term of the sequence, and To weigh the parameters.

[0011] Furthermore, in step S6, the relationship between the multi-scale reconstruction error and the adaptive weighted fusion is as follows: ,in, For reconstruction error, Let be the original reconstruction error of store s at the i-th time scale within the current date D to be analyzed. and These are the mean and standard deviation of the historical reconstruction error of store s at this time scale, respectively. The weights for adaptive weighted fusion are determined by the following formula: ,in, Let be the coefficient of variation at the i-th time scale, and j be the variable to be summed. Let j be the sum of all exponential terms across j time scales. This is the temperature coefficient.

[0012] Furthermore, in step S6, the standardized anomaly score specifically includes: ,in, and These represent the mean and standard deviation of the fusion and reconstruction error of store s during its historical normal operating period, respectively. Standardized anomaly scoring.

[0013] Furthermore, in step S7, the attenuation factor for the number of switching event pairs is specifically as follows: ,in, For store s, the decay factor for the number of on / off events on the current date D to be analyzed. For store s, a recent sliding window on the current date D to be analyzed The average number of daily on / off events Preset a fixed time scale for store s during its normal historical operating period. The average daily number of on / off events; The specific probability of store S going bankrupt is as follows: ,in, Let S be the probability of store S closing down on the current date D to be analyzed. Here, C is the sigmoid function, and C is the cumulative anomaly index. , and These are the weighting coefficients, and n is the bias term. Let L2 norm be the trend component of store s within natural day D. Let L2 norm be the trend component of store s on day t. C represents the magnitude of the decline in the trend component. ,in, This is a pre-defined sliding window that traces to the left from the current date to be analyzed. Heaven, from The continuous natural day interval up to D is used to calculate the average intensity of significant anomalous events during this period. Let be the standardized anomaly score for store s on day t. For indicator functions, This is the threshold for anomaly detection.

[0014] Further, in step S8, stores in a specified geographical area and a specified industry category are obtained and grouped into a store set S. The regional industry status index is specifically as follows: Among them, RSI is the regional industry status index. and For the index stores of the store collection, For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The scale weighting coefficient, K, is the spatial decay kernel function, specifically: Where L represents two stores and The geographical straight-line distance between them, H is a preset spatial bandwidth parameter used to control the distance attenuation frequency.

[0015] A store risk analysis system based on IoT roller shutter door data includes: The data acquisition and annotation module is used to acquire roller shutter door opening and closing event data from the Internet of Things platform for a specified time, region and industry, and to assign spatiotemporal semantic annotations to each record. The feature engineering module is used to extract valid event pairs from switch event data, construct daily multidimensional state feature vectors, and perform preprocessing and time-series decomposition. The model training module is used to build and train a hybrid deep neural network model based on LSTM-VAE to learn the feature distribution of the normal operating mode. The anomaly score calculation module is used to calculate multi-scale reconstruction error based on the trained model, and obtain the standardized anomaly score of individual stores through weighted fusion and standardization. The risk probability assessment module is used to calculate the probability of store closure risk by combining the persistence of abnormal scores, event-to-decay factor, and trend decline magnitude. The regional status aggregation module is used to spatially weight and aggregate the risk probabilities of each store within a region to generate a regional industry status index. The results output and visualization module is used to output the risk probability of individual stores and the regional industry status index in a visual form.

[0016] By adopting the above technical solution, the present invention has the following advantages compared with the prior art: This invention collects and deeply processes IoT roller shutter door opening and closing event data strongly correlated with store risk status through an IoT platform. This enables fully data-driven, subjective threshold-free intelligent analysis of physical store operations. First, the IoT platform acquires roller shutter door opening and closing event sequences, allowing for the filtering of specific data. Discrete opening and closing events are transformed into state feature vectors, elevating roller shutter door data from simple remote control and monitoring to a deeper level of analysis that quantifies the rhythm and activity of store risk status. Then, by training a hybrid deep neural network model, it automatically learns personalized normal operating patterns for different stores and industries, eliminating the need for manually setting fixed rules and effectively overcoming the errors of traditional threshold methods. It identifies and addresses both reported and missed defects; utilizing multi-scale reconstruction error adaptive fusion and standardization processing, it can sensitively and robustly capture subtle anomalies deviating from their historical norms, transforming roller shutter data into objective, standardized anomaly scores; finally, it integrates multi-dimensional signals of anomaly persistence, frequency decay, and trend decline to output data with early warning and intuitive store risk status, spatially weighting and aggregating individual risks to generate a regional industry status index, providing a group situational awareness capability not available in traditional roller shutter applications; it elevates IoT roller shutter data from a basic remote control tool to a fully data-driven, rule-free, and adaptive core intelligent analysis basis for different store operation models. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] It should be noted that in this invention, the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are all based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element of this invention must have a specific orientation, and therefore should not be construed as a limitation of this invention.

[0020] Example 1 refer to Figure 1As shown, this invention discloses a store risk analysis method based on IoT roller shutter door data, including: A method for store risk analysis based on IoT roller shutter door data includes: S1. Obtain the sequence of roller shutter door opening and closing events for all stores within a specified time interval, a specified geographical area, and a specified industry category through the Internet of Things platform. Each opening and closing event is recorded as a tuple containing store identifier, timestamp, action type, and geographical coordinates. S2. For each store, extract complete switch event pairs from the roller shutter door switch event sequence, and construct a state feature vector in natural days. The state feature vector includes at least the first opening time, the last closing time, the total business hours, the number of switch event pairs, and the average time interval between switch events. S3. Preprocess the state feature vector, including missing value processing, outlier truncation and standardization, to obtain a standardized feature vector; S4. Construct a state feature time series from the standardized feature vectors of each store over multiple consecutive days, and decompose it into trend components, periodic components, and residual components using a decomposition method; based on the residual components, extract multi-scale time series feature tensors through multiple sliding windows of different time lengths. S5. Construct and train a hybrid deep neural network model based on variational autoencoder and long short-term memory network. The hybrid deep neural network model uses the multi-scale temporal feature tensor of stores in the same industry during the historical normal operation period as training data and adopts unsupervised learning to learn the feature distribution of the normal operation mode. S6. For the store and date to be analyzed, input the residual components of the corresponding time window into the trained hybrid deep neural network model, calculate the multi-scale reconstruction error, and obtain the standardized anomaly score of the store on that date through adaptive weighted fusion and standardization. S7. Based on the standardized anomaly score, combined with the decay factor of the number of switch events and the decay magnitude of the trend component, calculate the probability of store closure risk. S8. For stores in a specified geographical area and industry category, the probability of store closure risk of each store in the specified geographical area is aggregated by spatial distance and combined with the store size weight to calculate the regional industry status index, so as to represent the overall status of the specified geographical area and industry.

[0021] Step S1: Obtain roller shutter door opening and closing event data within a specified time and space range. This step obtains raw data through a data interface with the IoT platform, such as an MQTT message queue or a RESTful API. Specifically, set the time range for the analysis task, such as a start time from January 1, 2024 to an end time from June 30, 2024; specify the geographical area as the central business district of a certain city, a circular area with a radius of 3 kilometers centered on the specified latitude and longitude N; and specify the industry category as "catering services". Obtain the opening and closing event records reported by the intelligent controllers of the roller shutter doors of all catering establishments within this time and space range from the IoT platform.

[0022] Each switch event record is a 5-tuple e i This includes a unique identifier for each store, a timestamp of the event, the action type, device status information (such as battery level and signal strength, used for data quality assessment), and the store's fixed geographical coordinates. It also retrieves static metadata for each store, including the store's building area and opening date, used for subsequent calculations of scale weighting coefficients.

[0023] In step S2, the state feature vector constructed in calendar days includes seven feature components: the first opening time, the last closing time, the total operating time, the number of opening and closing event pairs, the average interval between opening and closing events, the fluctuation of opening time, and the fluctuation of closing time.

[0024] The first time the store opens, the timestamp of the first "open" event of the day is taken and converted into minutes from 00:00:00 of that day. If there is no event, it is recorded as a missing value and processed in step S3. This feature reflects the regularity of the store's opening time. The last time the store closes, the timestamp of the last "close" event of the day is taken and converted into minutes. This feature reflects the regularity of the store's closing time. Total operating time. The summation of operating hours for all valid event pairs on a given day, expressed in minutes, is a core indicator for measuring store operational activity. The number of on / off event pairs counts the total number of valid event pairs on a given day; for restaurants, this reflects the frequency of different operating hours such as lunch and dinner. The average time interval between on / off events calculates the average time difference between adjacent on / off events (i.e., the interval between closing and opening), reflecting the frequency of business interruptions within a day. The fluctuation in closing time is calculated as the absolute value of the difference between the last closing time of the day and the average last closing time of the store over the past 30 days; this feature is used to detect abnormal deviations in closing time.

[0025] The feature preprocessing in step S3 is as follows: First, missing value handling: For dates with no on / off event records throughout the day (such as store holidays or closed days), the status feature vector is set to a zero vector to distinguish them from operating days with events but lower feature values.

[0026] Secondly, outlier truncation: For each feature dimension, based on historical data of all stores in the same industry, calculate the 1% quantile Q1 and 99% quantile Q1 of that feature. 99 For values ​​outside the interval [Q1, Q...] 99 The extreme values ​​of [] are forcibly assigned the corresponding quantile values ​​to eliminate noise interference caused by sensor failures or occasional extreme events (such as overnight operation).

[0027] Finally, Min-Max standardization: The truncated feature values ​​are linearly transformed to map them to the [0, 1] interval, resulting in a standardized feature vector. The linear transformation formula is a current technique. The standardized vector eliminates the impact of differences in the units of different features (such as business hours of several hundred minutes or the number of event pairs of single digits) on model training.

[0028] In step S4, the time series decomposition method is a locally weighted seasonal trend decomposition method to decompose the state feature time series, and the relationship is: ,in, Let be the standardized feature vector of store s on day t. As a trend component, For periodic components, These are the residual components; the residual components are used in subsequent steps. Perform model training and anomaly detection.

[0029] The core point is that trend changes and cyclical fluctuations are part of the normal operating rhythm of a store and should not be seen as risk signals. Therefore, only the residual component is retained. This data serves as input for subsequent anomaly detection models, thereby significantly reducing the interference of normal business cycles on risk warnings.

[0030] In step S4, the multi-scale temporal feature tensor is constructed in the following way: Set three sliding windows with different time scales. , and These correspond to the weekly, bi-weekly, and monthly scales, respectively. For each timescale sliding window, a subsequence is extracted from the residual sequence with a fixed step size to form a feature window. The sliding window is operated with a step size of 1 day, truncating continuous w-sequences. k The residual vectors of each day form a feature window, which ends with a set window end date.

[0031] The hybrid deep neural network model in step S5 includes: The encoder, composed of multiple LSTM networks, encodes the input feature window into a fixed-dimensional hidden state vector. The variational inference layer outputs the mean vector and log-variance vector of the latent variable distribution. It consists of two stacked LSTM layers, each with a hidden state dimension of 64. The input is a fixed-length feature window; the encoder processes the sequence data progressively, ultimately outputting a fixed-length context vector. Subsequently, two independent fully connected layers take the context vector as input and output the mean vector and log-variance vector of the latent variable distribution, respectively. In this embodiment, the latent space dimension is set to 16.

[0032] The latent space sampling layer employs reparameterization techniques to sample latent variables from a normal distribution; this ensures the model has generative capabilities and a smooth latent space.

[0033] The decoder, composed of an LSTM network, is used to reconstruct the input sequence from the latent variables to obtain the reconstruction window. Its structure is symmetrical to that of the encoder and is composed of an LSTM network. The decoder uses the sampled latent variables as the initial state and gradually reconstructs a reconstruction window with the same size as the input window.

[0034] The model training in this embodiment adopts an unsupervised learning approach, using only multi-scale feature windows extracted from stores in the same industry during their historical normal operating period, such as the 31st to 120th day after opening, avoiding the unstable stage at the beginning of opening, as the training set.

[0035] The conformance loss function minimized during model training is: ,in, To reconstruct the mean squared error loss, The KL divergence regularization term is used to represent the potential distribution between the standard normal distribution. To reconstruct the temporal smoothing regularization term of the sequence, and To weigh the parameters.

[0036] Specifically, the formula for calculating the reconstruction mean squared error loss is as follows: Where F is the dimension of the state feature vector. The sliding window length used in the current training. Let y be the actual value of the y-th feature in the input feature window at the p-th time step. This is the corresponding value reconstructed by the decoder, that is, the reconstructed value of the y-th feature at the p-th time step in the reconstruction window; The formula for calculating the KL divergence regularization term between the latent distribution and the standard normal distribution is: ,in, Let be the dimension of the latent space, which is 16, and and are the variance values ​​corresponding to the homogeneity and log-variance of the o-th dimension of the latent variable distribution output by the encoder, respectively. The formula for calculating the temporal smoothing regularization term of the reconstructed sequence is: ,in, Using the L1 norm, this term penalizes drastic changes between adjacent time steps in the reconstructed sequence, resulting in a smoother reconstruction that better reflects the continuous and gradual changes in store operation data. Let y be the reconstructed value of the y-th feature at time step p+1 in the reconstruction window.

[0037] By using deep neural networks to automatically and unsupervisedly learn the complex temporal dependencies and probability distribution boundaries of residual signals of store groups under normal operating conditions, without the need for manual definition of "normal" rules, adaptive modeling of personalized operating modes for different industries and stores is achieved.

[0038] In step S6, the relationship between multi-scale reconstruction error and adaptive weighted fusion is as follows: ,in, For reconstruction error, Let be the original reconstruction error of store s at the i-th time scale within the current date D to be analyzed. and These represent the mean and standard deviation of the historical reconstruction error of store s at this time scale; the original reconstruction error. The calculation was performed by reconstructing the formula for the mean squared error loss; )and This is a personalized baseline parameter for store s at the i-th time scale. It needs to be statistically analyzed offline when the store is first connected to the analytics system and can be updated periodically based on the accumulation of subsequent normal operational data. The specific calculation method is as follows: Select all dates within the historical normal operating period of store s, such as the 31st to 120th day after opening, or any operating period of more than 60 consecutive days without any abnormalities as confirmed by manual verification. For each date in the historical normal operating period, calculate the original reconstruction error at the i-th scale using the above method to obtain the reconstruction error sequence, and then calculate the mean and standard deviation based on the reconstruction error sequence.

[0039] The weights for adaptive weighted fusion are determined by the following formula: ,in, Let be the coefficient of variation at the i-th time scale, and j be the variable to be summed. Let j be the sum of all exponential terms across j time scales. This is the temperature coefficient.

[0040] The coefficient of variation eliminates the influence of the difference in the absolute magnitude of reconstruction errors at different scales, and can fairly measure the relative stability of historical errors at each scale. The smaller the value, the smoother the reconstruction error fluctuation at that scale, and the more reliable the output signal.

[0041] In step S6, the standardized anomaly scoring specifically involves: ,in, and These represent the mean and standard deviation of the fusion and reconstruction error of store s during its historical normal operating period, respectively. Standardized anomaly scoring.

[0042] In step S7, the attenuation factor for the number of switch event pairs is specifically as follows: ,in, For store s, the decay factor for the number of on / off events on the current date D to be analyzed. For store s, a recent sliding window on the current date D to be analyzed The average number of daily on / off events Preset a fixed time scale for store s during its normal historical operating period. The average daily number of on / off events; The specific probability of store S going bankrupt is as follows: ,in, Let S be the probability of store S closing down on the current date D to be analyzed. Here, C is the sigmoid function, and C is the cumulative anomaly index. , and These are the weighting coefficients, and n is the bias term. Let L2 norm be the trend component of store s within natural day D. Let L2 norm be the trend component of store s on day t. C represents the magnitude of the decline in the trend component. ,in, This is a pre-defined sliding window that traces to the left from the current date to be analyzed. Heaven, from The continuous natural day interval up to D is used to calculate the average intensity of significant anomalous events during this period. Let be the standardized anomaly score for store s on day t. For indicator functions, This is the threshold for anomaly detection.

[0043] In step S8, stores within a specified geographical region and industry category are obtained and grouped into a store set S. The specific regional industry status index is as follows: Among them, RSI is the regional industry status index. and For the index stores of the store collection, For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The scale weighting coefficient, K, is the spatial decay kernel function, specifically: Where L represents two stores and The geographical straight-line distance between them, H is a preset spatial bandwidth parameter used to control the distance attenuation frequency.

[0044] This embodiment collects and deeply processes IoT roller shutter door opening and closing event data that is strongly correlated with the store's risk status through an IoT platform. This enables fully data-driven, subjective threshold-free intelligent analysis of the physical store's operational status. First, the IoT platform acquires the roller shutter door opening and closing event sequence, allowing for the filtering of specific data. Discrete opening and closing events are transformed into state feature vectors, elevating roller shutter door data from simple remote control and monitoring to a deeper level of analysis that quantifies the rhythm and activity of the store's risk status. Then, by training a hybrid deep neural network model, it automatically learns the personalized normal operating modes of different stores and industries, eliminating the need for manually setting fixed rules and effectively overcoming the errors of traditional threshold methods. It identifies and addresses both reported and missed defects; utilizing multi-scale reconstruction error adaptive fusion and standardization processing, it can sensitively and robustly capture subtle anomalies deviating from their historical norms, transforming roller shutter data into objective, standardized anomaly scores; finally, it integrates multi-dimensional signals of anomaly persistence, frequency decay, and trend decline to output data with early warning and intuitive store risk status, spatially weighting and aggregating individual risks to generate a regional industry status index, providing a group situational awareness capability not available in traditional roller shutter applications; it elevates IoT roller shutter data from a basic remote control tool to a fully data-driven, rule-free, and adaptive core intelligent analysis basis for different store operation models.

[0045] Example 2 A store risk analysis system based on IoT roller shutter door data, implemented according to the method of Embodiment 1, includes: The data acquisition and labeling module is used to acquire roller shutter door opening and closing event data from stores in specified time, region and industry from the IoT platform, and to assign spatiotemporal semantic labels to each record; it is deployed in the cloud and acquires opening and closing event data in real time by subscribing to the IoT platform message queue, completes data cleaning and deduplication, and labels each record with spatiotemporal semantic tags according to the store registration information; The feature engineering module is used to extract valid event pairs from switch event data, construct daily multidimensional state feature vectors, and perform preprocessing and time series decomposition. Based on the Spark distributed computing framework, it implements daily offline batch processing tasks, performs all feature extraction, preprocessing and time series decomposition work in steps S2 to S4, and stores the results in the ClickHouse time series database for online service query. The model training module is used to build and train a hybrid deep neural network model based on LSTM-VAE, learning the feature distribution of normal operation mode; it implements the definition, training, and version management of LSTM-VAE models based on the PyTorch framework and GPU cluster. The module performs incremental training on the model monthly using newly added historical data and updates the model parameters. The anomaly score calculation module is used to calculate multi-scale reconstruction error based on the trained model, and obtain the standardized anomaly score of each store through weighted fusion and standardization. Deployed in the form of microservice, it receives store ID and date requests, reads residual feature windows from the time series database, calls the trained model file to calculate and return the standardized anomaly score. The risk probability assessment module is used to calculate the probability of store closure risk by combining the persistence of abnormal scores, event-pair decay factor and trend decline magnitude; it retrieves historical abnormal score sequences and trend component data from the database, performs the fusion calculation in step S7, updates the risk probability of all stores daily and stores it in the results database. The regional status aggregation module is used to spatially weight and aggregate the risk probabilities of each store within a region to generate a regional industry status index; in response to front-end analysis requests, it filters stores according to the region and industry specified by the user, queries the risk probability from the result database, executes the spatial aggregation algorithm, and returns the RSI index and heat distribution data in GeoJSON format; The results output and visualization module is used to output the risk probability of individual stores and the regional industry status index in a visual format. The web front-end interface is built based on Vue.js and ECharts, providing functions such as single-store risk probability trend line charts, regional risk heatmaps, and alert list subscription and push notifications.

[0046] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for store risk analysis based on IoT roller shutter door data, characterized in that, include: S1. Obtain the sequence of roller shutter door opening and closing events for all stores within a specified time interval, a specified geographical area, and a specified industry category through the Internet of Things platform. Each opening and closing event is recorded as a tuple containing store identifier, timestamp, action type, and geographical coordinates. S2. For each store, extract complete switch event pairs from the roller shutter door switch event sequence, and construct a state feature vector in natural days. The state feature vector includes at least the first opening time, the last closing time, the total business hours, the number of switch event pairs, and the average time interval between switch events. S3. Preprocess the state feature vector, including missing value processing, outlier truncation and standardization, to obtain a standardized feature vector; S4. Construct a state feature time series from the standardized feature vectors of each store over multiple consecutive days, and decompose it into trend components, periodic components, and residual components using a decomposition method; based on the residual components, extract multi-scale time series feature tensors through multiple sliding windows of different time lengths. S5. Construct and train a hybrid deep neural network model based on variational autoencoder and long short-term memory network. The hybrid deep neural network model uses the multi-scale temporal feature tensor of stores in the same industry during the historical normal operation period as training data and adopts unsupervised learning to learn the feature distribution of the normal operation mode. S6. For the store and date to be analyzed, input the residual components of the corresponding time window into the trained hybrid deep neural network model, calculate the multi-scale reconstruction error, and obtain the standardized anomaly score of the store on that date through adaptive weighted fusion and standardization. S7. Based on the standardized anomaly score, combined with the decay factor of the number of switch events and the decay magnitude of the trend component, calculate the probability of store closure risk. S8. For stores in a specified geographical area and industry category, the probability of store closure risk of each store in the specified geographical area is aggregated by spatial distance and combined with the store size weight to calculate the regional industry status index, so as to represent the overall status of the specified geographical area and industry.

2. The store risk analysis method based on IoT roller shutter door data as described in claim 1, characterized in that: In step S2, the state feature vector constructed in calendar days includes seven feature components: the first opening time, the last closing time, the total operating time, the number of opening and closing event pairs, the average interval between opening and closing events, the fluctuation of opening time, and the fluctuation of closing time.

3. The store risk analysis method based on IoT roller shutter door data as described in claim 1, characterized in that: In step S4, the time series decomposition method is a locally weighted seasonal trend decomposition method to decompose the state feature time series, and the relationship is: ,in, Let be the standardized feature vector of store s on day t. As a trend component, For periodic components, These are the residual components; the residual components are used in subsequent steps. Perform model training and anomaly detection.

4. The store risk analysis method based on IoT roller shutter door data as described in claim 3, characterized in that, In step S4, the multi-scale temporal feature tensor is constructed in the following way: Set three sliding windows with different time scales. , and These correspond to the weekly, bi-weekly, and monthly scales, respectively. For each timescale sliding window, a subsequence is extracted from the residual sequence with a fixed step size to form a feature window.

5. The store risk analysis method based on IoT roller shutter door data as described in claim 4, characterized in that, The hybrid deep neural network model in step S5 includes: The encoder, consisting of a multi-layer LSTM network, encodes the input feature window into a fixed-dimensional latent state vector; the variational inference layer is based on the mean vector and log-variance vector of the output latent variable distribution. The latent space sampling layer employs reparameterization techniques to sample latent variables from a normal distribution; The decoder, composed of an LSTM network, is used to reconstruct the input sequence from the latent variables to obtain the reconstruction window; The conformance loss function minimized during model training is: ,in, To reconstruct the mean squared error loss, The KL divergence regularization term is used to represent the potential distribution between the standard normal distribution. To reconstruct the temporal smoothing regularization term of the sequence, and To weigh the parameters.

6. The store risk analysis method based on IoT roller shutter door data as described in claim 5, characterized in that: In step S6, the relationship between multi-scale reconstruction error and adaptive weighted fusion is as follows: ,in, For reconstruction error, Let be the original reconstruction error of store s at the i-th time scale within the current date D to be analyzed. and These are the mean and standard deviation of the historical reconstruction error of store s at this time scale, respectively. The weights for adaptive weighted fusion are determined by the following formula: ,in, Let be the coefficient of variation at the i-th time scale, and j be the variable to be summed. Let j be the sum of all exponential terms across j time scales. This is the temperature coefficient.

7. The store risk analysis method based on IoT roller shutter door data as described in claim 6, characterized in that: In step S6, the standardized anomaly scoring specifically involves: ,in, and These represent the mean and standard deviation of the fusion and reconstruction error of store s during its historical normal operating period, respectively. Standardized anomaly scoring.

8. The store risk analysis method based on IoT roller shutter door data as described in claim 7, characterized in that: In step S7, the attenuation factor for the number of switch event pairs is specifically as follows: ,in, For store s, the decay factor for the number of on / off events on the current date D to be analyzed. For store s, a recent sliding window on the current date D to be analyzed The average number of daily on / off events Preset a fixed time scale for store s during its normal historical operating period. The average daily number of on / off events; The specific probability of store S going bankrupt is as follows: ,in, Let S be the probability of store S closing down on the current date D to be analyzed. Here, C is the sigmoid function, and C is the cumulative anomaly index. , and These are the weighting coefficients, and n is the bias term. Let L2 norm be the trend component of store s within natural day D. Let L2 norm be the trend component of store s on day t. C represents the magnitude of the decline in the trend component. ,in, This is a pre-defined sliding window that traces to the left from the current date to be analyzed. Heaven, from The continuous natural day interval up to D is used to calculate the average intensity of significant anomalous events during this period. Let be the standardized anomaly score for store s on day t. For indicator functions, This is the threshold for anomaly detection.

9. A store risk analysis method based on IoT roller shutter door data as described in claim 8, characterized in that: In step S8, stores within a specified geographical region and industry category are obtained and grouped into a store set S. The specific regional industry status index is as follows: Among them, RSI is the regional industry status index. and For the index stores of the store collection, For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The probability of bankruptcy risk on the current date D to be analyzed. For the store The scale weighting coefficient, K, is the spatial decay kernel function, specifically: Where L represents two stores and The geographical straight-line distance between them, H is a preset spatial bandwidth parameter used to control the distance attenuation frequency.

10. A store risk analysis system based on IoT roller shutter door data, characterized in that, include: The data acquisition and annotation module is used to acquire roller shutter door opening and closing event data from the Internet of Things platform for a specified time, region and industry, and to assign spatiotemporal semantic annotations to each record. The feature engineering module is used to extract valid event pairs from switch event data, construct daily multidimensional state feature vectors, and perform preprocessing and time-series decomposition. The model training module is used to build and train a hybrid deep neural network model based on LSTM-VAE to learn the feature distribution of the normal operating mode. The anomaly score calculation module is used to calculate multi-scale reconstruction error based on the trained model, and obtain the standardized anomaly score of individual stores through weighted fusion and standardization. The risk probability assessment module is used to calculate the probability of store closure risk by combining the persistence of abnormal scores, event-to-decay factor, and trend decline magnitude. The regional status aggregation module is used to spatially weight and aggregate the risk probabilities of each store within a region to generate a regional industry status index. The results output and visualization module is used to output the risk probability of individual stores and the regional industry status index in a visual form.

Images