Data analysis method based on chain operation benefit supervision system

Abstract

The present application relates to the technical field of data analysis, and discloses a data analysis method based on a chain operation benefit supervision system, the method comprising: collecting and preprocessing multi-dimensional data of each store of a chain operation enterprise to obtain multi-dimensional time series data; performing multi-dimensional similarity calculation and fuzzy clustering to obtain store clustering results and fuzzy membership; constructing a decision tree model and performing store operation status evaluation to obtain store operation status evaluation results; performing time series feature extraction, constructing a semantic segmentation network, and obtaining a store business prediction model; performing in-depth prediction of multi-store business data to obtain store business prediction results; performing comprehensive abnormality prediction and resource optimization allocation based on store operation status evaluation results, store business prediction results and fuzzy membership to obtain abnormality prediction reports and optimization plans, thereby improving the accuracy of operation status evaluation.

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a data analysis method based on a chain operation benefit supervision system. Background Art

[0002] As the scale of chain enterprises continues to expand and the market environment becomes increasingly complex, traditional management methods can no longer meet the needs of enterprises for refined and intelligent management. Especially in the multi-store and multi-regional operating environment, enterprises are faced with problems such as data dispersion, information islands, and decision-making lags, which seriously affect the overall operational efficiency and competitiveness of enterprises.

[0003] At the same time, the intensification of market competition and the diversification of consumer demand require chain enterprises to respond to market changes more quickly and accurately. However, traditional data analysis methods often have difficulty processing massive, multi-dimensional, and real-time business data, and cannot provide timely and comprehensive decision-making support for enterprises. This leads to low decision-making efficiency in resource allocation, inventory management, marketing strategies, etc., and it is difficult to give full play to the scale advantages of chain operations. In addition, the traditional business efficiency supervision system lacks the ability to deeply mine multi-dimensional data, and cannot achieve accurate prediction and rapid response to abnormal situations, and it is difficult to provide enterprises with personalized and dynamic resource optimization allocation solutions. Summary of the invention

[0004] The present application provides a data analysis method based on a chain operation benefit supervision system, thereby improving the accuracy of business status assessment.

[0005] In a first aspect, the present application provides a data analysis method based on a chain operation benefit supervision system, the data analysis method based on the chain operation benefit supervision system comprising:

[0006] Collect and preprocess the multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data;

[0007] Performing multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain store clustering results and fuzzy membership;

[0008] Building a decision tree model based on the store clustering results and the fuzzy membership degree, and performing store operation status evaluation on the multidimensional time series data to obtain a store operation status evaluation result;

[0009] Extracting time series features from the multidimensional time series data, building a semantic segmentation network, and obtaining a store business prediction model;

[0010] Based on the store business prediction model and the store clustering result, deep prediction is performed on the business data of multiple stores to obtain store business prediction results;

[0011] Based on the store operation status evaluation results, the store business forecast results and the fuzzy membership, comprehensive abnormality prediction and resource optimization configuration are performed to obtain an abnormality prediction report and an optimization plan.

[0012] A second aspect of the present application provides a data analysis device based on a chain operation benefit supervision system, the data analysis device based on a chain operation benefit supervision system comprising:

[0013] The collection module is used to collect and preprocess the multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data;

[0014] A calculation module, used to perform multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain store clustering results and fuzzy membership;

[0015] An evaluation module, used to construct a decision tree model based on the store clustering results and the fuzzy membership, and to evaluate the store operation status of the multidimensional time series data to obtain a store operation status evaluation result;

[0016] A construction module is used to extract time series features from the multi-dimensional time series data, construct a semantic segmentation network, and obtain a store business prediction model;

[0017] A prediction module, used to perform in-depth prediction on multiple store business data based on the store business prediction model and the store clustering result to obtain store business prediction results;

[0018] A configuration module is used to perform comprehensive abnormality prediction and resource optimization configuration based on the store operation status evaluation results, the store business prediction results and the fuzzy membership, and obtain an abnormality prediction report and an optimization plan.

[0019] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned data analysis method based on the chain operation efficiency supervision system.

[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned data analysis method based on the chain operation benefit supervision system.

[0021] Compared with the prior art, the present application has the following beneficial effects: by collecting and preprocessing multi-dimensional data such as sales, inventory, and customer flow, the standardization and high quality of data are achieved. The use of technologies such as dynamic time warping and kernel function transformation improves the accuracy and flexibility of store clustering, which helps to more accurately identify different types of stores. Through entropy weight analysis and principal component analysis, the selection of evaluation features is optimized, and the accuracy and interpretability of business status evaluation are improved. Combining time convolutional networks, long short-term memory networks, and self-attention mechanisms, the feature extraction and prediction capabilities of time series data are improved, and more accurate business forecasts are achieved. Through a multi-task prediction model, sales, customer flow, and inventory levels are predicted at the same time, which improves the comprehensiveness and efficiency of the prediction. The use of autoencoders and dynamic threshold calculations improves the accuracy and sensitivity of anomaly detection, and at the same time, more reasonable resource allocation is achieved through decision rule libraries and constraint optimization. From time series decomposition to multi-scale time window division, and then to multi-task learning, multi-level and multi-scale analysis of data is achieved, and the scalability of the model and the adaptability to different store characteristics are enhanced through technologies such as conditional embedding and residual connection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0024] Figure 1 It is a flow chart of a data analysis method based on a chain operation benefit supervision system provided by an embodiment of the present invention;

[0025] Figure 2 It is a schematic block diagram of the structure of a data analysis device based on a chain operation benefit supervision system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0028] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0029] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 In the embodiment of the present application, an embodiment of the data analysis method based on the chain operation benefit supervision system includes:

[0030] Step 100: Collect and preprocess multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data;

[0031] It is understandable that the execution subject of the present application may be a data analysis device based on the chain operation benefit supervision system, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0032] Specifically, the sales data, inventory data and customer flow data of each store of the chain operation enterprise are collected to obtain the original multidimensional data. The original multidimensional data is processed for missing values ​​to fill the gaps in the data. Mean filling, interpolation or filling methods based on machine learning are used to ensure the integrity and continuity of the data and obtain the filled data. Outlier detection and correction are performed on the filled data to identify and process outliers in the data. Common methods include statistical methods such as box plots, Z-Score analysis and machine learning methods such as isolation forest algorithm to ensure the authenticity and reliability of the data and obtain the corrected data. Multi-dimensional data fusion is performed on the corrected data. Data fusion technology is used to integrate data of different dimensions, eliminate data redundancy, enhance the comprehensiveness and consistency of data, and obtain fused data. Time series decomposition is performed on the fused data. Decomposition techniques such as wavelet transform and empirical mode decomposition are used to decompose the time series data into different time scale components to capture the multi-scale characteristics of the data. Then the time series reconstruction is performed to recombine the decomposed components to obtain reconstructed data that better reflects the essence of the data. Perform multi-scale time window division on the reconstructed data. By setting different time window lengths, the data is divided into multiple time scales for multi-level analysis and multi-scale time window data is obtained. Perform time series correlation analysis on multi-scale time window data. Through correlation analysis techniques such as Pearson correlation coefficient and Granger causality test, identify the time series relationship and causal relationship in the data and explore the potential laws in the data. Perform feature selection on the results of time series correlation analysis. Through feature selection techniques such as principal component analysis and Lasso regression, screen out the features that have the greatest impact on the analysis target and obtain multi-dimensional time series data.

[0033] Step 200: Perform multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain store clustering results and fuzzy membership;

[0034] Specifically, the multidimensional time series data is segmented and dynamically time-warped to align data from different time periods to make them comparable. The dynamic time warping method is used to deal with the nonlinear time deformation of the time series. The key points of the time series are aligned through the dynamic programming algorithm to obtain the aligned time series. According to the aligned time series, the Euclidean distance matrix is ​​calculated. The Euclidean distance can measure the straight-line distance between time series to construct the initial similarity matrix. The initial similarity matrix is ​​transformed by kernel function. Commonly used kernel functions such as the Gaussian kernel function can map low-dimensional data to high-dimensional feature space, better reveal the potential structure and similarity characteristics of the data, and obtain the similarity matrix in the high-dimensional feature space. The k-nearest neighbor graph is constructed according to the similarity matrix in the high-dimensional feature space. The k-nearest neighbor graph is formed by connecting each data point with its most similar k neighbors to form a network structure reflecting the relationship between stores, and the store relationship network is obtained. Spectral clustering analysis is performed on the store relationship network. Spectral clustering is a clustering method based on graph theory. By calculating the Laplace matrix of the graph and finding its eigenvectors, the data is mapped to a new low-dimensional space, in which cluster analysis is performed to obtain the initial clustering results. After obtaining the initial clustering results, the fuzzy membership matrix is ​​calculated. The fuzzy membership matrix is ​​used to represent the probability that each store belongs to each cluster. By calculating the initial membership, the degree of membership of each store to different clusters is obtained. In order to optimize the clustering effect, the initial membership is iteratively optimized by fuzzy c-means. Fuzzy c-means iteratively updates the membership and cluster center to make the clustering results more accurate and stable, and obtains the optimized fuzzy membership. The number of clusters of each store is redistributed and dynamically adjusted according to the optimized fuzzy membership. Through dynamic adjustment, the number of clusters is adjusted according to the actual situation to better reflect the real relationship and business characteristics between stores, and the final store clustering results are obtained.

[0035] Step 300: construct a decision tree model based on the store clustering results and fuzzy membership, and evaluate the store operation status of the multidimensional time series data to obtain the store operation status evaluation results;

[0036] It should be noted that the entropy weight method is used to analyze the store clustering results. The entropy weight method is used to quantify the weight coefficients of each cluster, which effectively measures the importance of different clusters. Based on the weight coefficient and fuzzy membership, the comprehensive score is calculated to obtain the initial evaluation index of each store. The principal component analysis is performed on the initial evaluation index of the store, and the high-dimensional data is reduced in dimension to extract the main evaluation features. The feature importance ranking is generated based on the evaluation features after dimensionality reduction to determine which features have the greatest impact on the store's operating status. Based on the feature importance ranking, the optimized evaluation feature set is selected. By selecting those features that have the greatest impact on the evaluation results, the model is simplified and the model performance is improved. A decision tree model is constructed based on the optimized evaluation feature set. The decision tree model is a commonly used classification and regression model. By learning the relationship between the features in the data and the target variable, a series of decision rules are formed to perform prediction and evaluation. The multidimensional time series data is input into the constructed decision tree model to evaluate the store's operating status and obtain the initial operating status evaluation results. Through the calculation of the model, the operating status of each store is preliminarily determined. The confidence calculation of the initial operating status evaluation results is performed to obtain the evaluation confidence interval. Confidence calculation can determine the reliability range of the evaluation results through statistical analysis methods and provide an evaluation confidence interval. According to the evaluation confidence interval, the initial business status evaluation results are revised to eliminate some possible errors and uncertainties, and obtain more accurate and reliable store business status evaluation results.

[0037] Step 400: extract time series features from multi-dimensional time series data, build a semantic segmentation network, and obtain a store business prediction model;

[0038] Specifically, the multidimensional time series data is divided into three subsets: sales data, customer flow data, and inventory data. Wavelet transform is performed on these three subsets to obtain time-frequency features at different time scales and frequencies. Multi-scale time-frequency features can effectively capture periodic and non-stationary changes in data. A temporal convolutional network is constructed based on the multi-scale time-frequency features. The temporal convolutional network extracts features through three one-dimensional convolutional layers. Each layer uses 64 convolution kernels with a kernel size of 3 and a step size of 1. The ReLU activation function is used to extract the first local time series features. The first local time series features are processed by a long short-term memory network (LSTM). The long short-term memory network contains two LSTM layers, each of which contains 128 hidden units, which are used to capture long-term dependencies and global time series features in time series data to obtain the first global time series features. The first local time series features and the first global time series features are fused through the attention mechanism. The attention mechanism uses scaled dot product attention with 8 attention heads. Through this mechanism, the model can focus on the key parts of the temporal features and obtain the first mixed temporal features. The first mixed temporal features are processed by the self-attention mechanism. The self-attention mechanism contains 3 self-attention layers, and the hidden dimension of each layer is 256. The self-attention mechanism can strengthen the interdependence and importance between features and obtain the first attention weighted features. Based on the first attention weighted features, the encoder-decoder structure is constructed to obtain the initial semantic segmentation network. The encoder part contains 5 downsampling blocks, each of which contains two 3x3 convolutional layers and a maximum pooling layer to gradually reduce the size of the feature map while extracting high-level features. The decoder part contains 5 upsampling blocks, each of which contains a transposed convolutional layer and two 3x3 convolutional layers to gradually restore the size of the feature map while retaining important feature information. In order to optimize the initial semantic segmentation network, residual connections are added between each encoder and decoder block, and channel matching is performed through 1x1 convolution to obtain the optimized semantic segmentation network. Residual connections help alleviate the gradient vanishing problem and promote the training of deeper networks. The optimized semantic segmentation network is conditionally embedded with the store attribute information. After the store attribute information is processed by the embedding layer, it is connected with the output of each decoder block, so as to add more background information in the feature extraction process and improve the prediction performance of the model. Multi-task learning is performed on the conditional semantic segmentation network to simultaneously predict sales, customer flow and inventory levels to obtain a multi-task prediction model. The multi-task prediction model is divided into three branches at the end of the network. Each branch contains 2 fully connected layers. The output layer uses a linear activation function to output the prediction results of sales, customer flow and inventory levels respectively. The multi-task prediction model is integrated with the time series decomposition model, combining the advantages of the time series decomposition model to improve the prediction accuracy and robustness, and finally obtain a store business prediction model.

[0039] Step 500: Perform deep prediction on multiple store business data based on the store business prediction model and the store clustering results to obtain store business prediction results;

[0040] Specifically, the time series of multi-store business data is decomposed into trend, seasonal and random components. The trend, seasonal and random components are input into the feature extraction layer of the store business prediction model and processed through three one-dimensional convolutional layers. Each convolutional layer uses 64 convolution kernels with a kernel size of 3 and a step size of 1. The ReLU activation function is used to extract the second local time series feature. The local time series feature captures the local pattern and detail information in the data. The second local time series feature is input into the time series modeling layer of the store business prediction model and processed through two LSTM layers. Each LSTM layer contains 128 hidden units. The long short-term memory network can capture the long-term dependency in the time series data to obtain the second global time series feature. The second local time series feature and the second global time series feature are input into the attention fusion mechanism. The attention mechanism weightedly fuses the local features and the global features by scaling the dot product attention, so that the model can better focus on important features and obtain the second mixed time series feature. The second mixed time series feature is input into three self-attention layers for attention weighting processing. The hidden dimension of each self-attention layer is 256. The self-attention mechanism strengthens the interdependence and importance between features, and obtains the second attention weighted feature. The second attention weighted feature is input into the encoder-decoder structure of the semantic segmentation layer. The encoder part contains 5 downsampling blocks, each of which contains two 3x3 convolutional layers and a maximum pooling layer. The decoder part contains 5 upsampling blocks, each of which contains a transposed convolutional layer and two 3x3 convolutional layers. Through the processing of the encoder and decoder, the size of the feature map is gradually extracted and restored to obtain the initial semantic feature. Store attribute information is extracted according to the store clustering results, and the store attribute information is conditionally embedded with the initial semantic feature. After the store attribute information is processed by the embedding layer, it is connected with the output of each decoder block to enhance the adaptability of the model to different store characteristics and obtain the conditional semantic feature. The conditional semantic feature is input into the multi-task prediction layer, and sales, customer flow and inventory level are predicted respectively through 3 parallel fully connected layers. Each fully connected layer contains 2 hidden layers, and the output layer uses a linear activation function to obtain the preliminary prediction result. The preliminary forecast results are integrated, and the forecast results of different tasks are combined through the integration method to obtain the final store business forecast results. The store business forecast results include point forecast values ​​and forecast intervals. The point forecast value provides a specific forecast value, and the forecast interval gives the confidence range of the forecast result.

[0041] Step 600: Based on the store operation status evaluation results, store business forecast results and fuzzy membership, comprehensive abnormality prediction and resource optimization configuration are performed to obtain an abnormality prediction report and an optimization plan.

[0042] Specifically, the store operation status evaluation results and the store business forecast results are aligned in time series to obtain aligned time series data, eliminating the impact of time differences. An anomaly detection model is constructed based on the aligned time series data and the fuzzy membership. The anomaly detection model uses an autoencoder to calculate the reconstruction error. The autoencoder is an unsupervised learning model that encodes the input data into a low-dimensional space and reconstructs it back to the original space to calculate the reconstruction error to measure the degree of abnormality of the data. A dynamic threshold is calculated for the reconstruction error, and the threshold is dynamically adjusted to set the abnormality judgment standard according to the actual situation, so as to detect anomalies more flexibly. According to the abnormal judgment standard, the reconstruction error is binarized and converted into a binary result, that is, normal or abnormal, to obtain an anomaly detection result. Anomaly feature extraction is performed on the anomaly detection result, and a feature vector that can represent the anomaly is extracted by analyzing the features of the abnormal data. The abnormal feature vector is deeply analyzed to identify different types of anomalies and obtain anomaly type classification. The anomaly type classification helps to clarify the specific manifestations and causes of different anomalies. According to the anomaly type classification, a decision rule library is constructed, and the decision rule library contains processing rules and resource allocation suggestions for different types of anomalies. Constrained optimization is performed on the resource allocation suggestion. Constrained optimization optimizes resource allocation by considering actual constraints, such as limited resources, costs and benefits, and obtains the best resource allocation plan. Anomaly prediction reports are generated based on anomaly detection results and anomaly type classification. The anomaly prediction report records the detected anomalies, anomaly types and their characteristics, as well as the corresponding resource allocation and optimization plans.

[0043] In the embodiment of the present application, by collecting and preprocessing multi-dimensional data such as sales, inventory, and passenger flow, the standardization and high quality of data are achieved. The use of technologies such as dynamic time warping and kernel function transformation improves the accuracy and flexibility of store clustering, which helps to more accurately identify different types of stores. Through entropy weight analysis and principal component analysis, the selection of evaluation features is optimized, and the accuracy and interpretability of business status evaluation are improved. Combining time convolutional networks, long short-term memory networks, and self-attention mechanisms, the feature extraction and prediction capabilities of time series data are improved, and more accurate business forecasts are achieved. Through a multi-task prediction model, sales, passenger flow, and inventory levels are predicted at the same time, which improves the comprehensiveness and efficiency of the prediction. The use of autoencoders and dynamic threshold calculations improves the accuracy and sensitivity of anomaly detection, and at the same time, through decision rule bases and constraint optimization, more reasonable resource allocation is achieved. From time series decomposition to multi-scale time window division, and then to multi-task learning, multi-level and multi-scale analysis of data is achieved, and the scalability of the model and the adaptability to different store characteristics are enhanced through technologies such as conditional embedding and residual connection.

[0044] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0045] Collect sales data, inventory data and customer flow data of each store of the chain enterprise to obtain original multi-dimensional data;

[0046] Process the missing values ​​of the original multidimensional data to obtain the filled data, and perform outlier detection and correction on the filled data to obtain the corrected data;

[0047] Perform multi-dimensional data fusion on the corrected data to obtain fused data, and perform time series decomposition and time series reconstruction on the fused data to obtain reconstructed data;

[0048] The reconstructed data is divided into multi-scale time windows to obtain multi-scale time window data, and time series correlation analysis and feature selection are performed on the multi-scale time window data to obtain multi-dimensional time series data.

[0049] Specifically, through a unified data collection interface, the sales data, inventory data, and customer flow data of each store are collected in real time or regularly to obtain the original multidimensional data. Sales data includes information such as sales and sales volume of each store in each time period. Inventory data includes the inventory quantity of each commodity, inventory turnover rate, etc., and customer flow data includes the number of customers per day or hour, entry and exit time, etc. Missing values ​​are processed for the original multidimensional data. During the data collection process, missing values ​​may occur due to various reasons. In order to ensure the integrity of the data and the accuracy of the analysis, these missing values ​​are filled. Common filling methods include mean filling, interpolation, and filling methods based on machine learning. For example, when using mean filling, the missing value is replaced by the average value of the data column. Suppose there is an inventory data series , where there are missing values , can be filled in by the formula:

[0050] ;

[0051] The integrity of the filled data is restored. Perform outlier detection and correction on the filled data. Outliers are values ​​that deviate significantly from other data points, which may be caused by data entry errors, sensor failures, etc. The presence of outliers will affect the accuracy of data analysis, so they need to be detected and corrected. Common detection methods include box plot analysis, Z-Score analysis, etc. For example, when using Z-Score analysis, the Z value of each data point can be calculated by the formula:

[0052] ;

[0053] in represents the value of a data point. represents the mean of the data, Represents the standard deviation of the data. When the Z value exceeds a certain threshold, the data point can be considered an outlier. For the detected outliers, the mean substitution method, neighboring value substitution method, etc. are used for correction to obtain the corrected data. The corrected data is subjected to multi-dimensional data fusion. Multi-dimensional data fusion is the integration of data from different sources and different types to form a unified data set for comprehensive analysis. The data fusion process includes steps such as data standardization and data alignment. For example, the sales data, inventory data, and passenger flow data are standardized so that they have the same dimension, and then aligned according to the timestamp to form a complete data matrix. Assume that the sales data, inventory data, and passenger flow data are , and , the data matrix obtained after fusion can be expressed as . Perform time series decomposition and time series reconstruction on the fused data. Decompose the original time series into trend components, seasonal components and random components. Common decomposition methods include moving average method, exponential smoothing method, etc. Assume that the time series of the fused data is , can be obtained through the formula Decompose it into Represents the trend component, Indicates seasonal components, Represents random components. The decomposed components can be analyzed independently to extract useful information. Through time series reconstruction, the components are recombined to obtain reconstructed data. Reconstructed data can eliminate noise and improve the smoothness and predictability of data. Multi-scale time window division is performed on the reconstructed data, and the time series data is divided according to different time scales so as to be analyzed at different time granularities. For example, the data of a day can be divided into different time windows such as hour level, day level, week level, etc. to obtain multi-scale time window data. Assume that the reconstructed data is , can be divided according to hours, days, and weeks, and we get , and . Perform time series correlation analysis and feature selection on multi-scale time window data. Time series correlation analysis is to discover the potential patterns and laws in the data by analyzing the correlation between time series data. Commonly used methods include Pearson correlation coefficient, Granger causality test, etc. Assume that the two time series data are and , its correlation can be calculated by the Pearson correlation coefficient formula:

[0054] ;

[0055] According to the results of correlation analysis, the most influential features on the analysis target are selected, and feature selection is performed to obtain multidimensional time series data.

[0056] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0057] Segment and dynamically time warp multidimensional time series data to obtain aligned time series;

[0058] The Euclidean distance matrix is ​​calculated according to the aligned time series to obtain the initial similarity matrix, and the initial similarity matrix is ​​transformed by the kernel function to obtain the similarity matrix in the high-dimensional feature space;

[0059] A k-nearest neighbor graph is constructed based on the similarity matrix in the high-dimensional feature space to obtain a store relationship network, and spectral clustering is performed on the store relationship network to obtain an initial clustering result;

[0060] The fuzzy membership matrix is ​​calculated according to the initial clustering results to obtain the initial membership, and the fuzzy c-means iterative optimization is performed on the initial membership to obtain the fuzzy membership;

[0061] The number of clusters of each store is redistributed and dynamically adjusted according to the fuzzy membership to obtain the store clustering results.

[0062] Specifically, multidimensional time series data is segmented and dynamically time-warped. Time series segmentation refers to dividing time series data into fixed or variable time windows for subsequent analysis. Dynamic time warping is a technique for aligning time series data by allowing nonlinear time deformation to find the best match between two time series. Suppose there are two time series and , dynamic time warping calculates the alignment path between them by minimizing the cumulative distance. The cumulative distance can be expressed as:

[0063] ;

[0064] in, Indicates The distance between two points is usually calculated using the Euclidean distance. Through dynamic time warping, the aligned time series is obtained. The Euclidean distance matrix is ​​calculated based on the aligned time series. The Euclidean distance is used to measure the similarity between two time series. The smaller the distance, the higher the similarity. Assume that the aligned time series data is , where each is a time series. Euclidean distance matrix Elements Representing time series and The Euclidean distance between them can be expressed as:

[0065] ;

[0066] in, and Represents time series and In time The kernel function transformation is performed on the initial similarity matrix to obtain the similarity matrix in the high-dimensional feature space. Kernel function transformation is a method of mapping low-dimensional data to high-dimensional space, so that the intrinsic structure of data can be more easily found in high-dimensional space. Commonly used kernel functions include Gaussian kernel function (RBF), which is in the form of:

[0067] ;

[0068] in, is the bandwidth parameter of the kernel function. Through the kernel function transformation, the initial similarity matrix Transformed into a similarity matrix in a high-dimensional feature space According to the similarity matrix in the high-dimensional feature space Construct a k-nearest neighbor graph. The k-nearest neighbor graph is formed by connecting each data point with its k most similar neighbors. Assume that Data points , then the k-nearest-neighbor graph It can be represented as an adjacency matrix ,in Represents data points and Are they neighbors? After constructing the k-nearest neighbor graph, we get the store relationship network. We perform spectral clustering on the store relationship network to get the initial clustering results. Spectral clustering is a clustering method based on graph theory. By calculating the Laplacian matrix of the graph and finding its eigenvectors, we map the data to a new low-dimensional space and perform cluster analysis in this space. Laplacian matrix It can be expressed as:

[0069] ;

[0070] in, is the degree matrix, with diagonal elements Representation Node By Perform eigendecomposition to obtain the eigenvector matrix , then Perform k-means clustering to obtain the initial clustering results. Calculate the fuzzy membership matrix based on the initial clustering results. The fuzzy membership matrix is ​​used to represent the membership of each data point to each cluster. By calculating the initial membership, the membership degree of each data point to different clusters is obtained. Fuzzy c-means is a commonly used fuzzy clustering algorithm. By iteratively optimizing the membership and cluster center, the clustering results are more accurate and stable. Membership matrix Elements Represents data points Belongs to cluster The membership degree of , the optimization objective can be expressed as:

[0071] ;

[0072] in, Represents data points Belongs to the cluster center The membership degree of is a fuzzy index used to control the degree of fuzziness of membership. Through iterative optimization, the final fuzzy membership matrix is ​​obtained. The number of clusters of each store is redistributed and dynamically adjusted according to the fuzzy membership to obtain the final store clustering result. By analyzing the membership matrix, the number and allocation of clusters are dynamically adjusted to better reflect the actual structure of the data. For example, if the membership of a cluster is generally low, you can consider splitting the cluster into multiple sub-clusters or merging it into other clusters. Through dynamic adjustment, a more reasonable and accurate store clustering result is obtained.

[0073] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0074] The entropy weight method is used to analyze the store clustering results to obtain the weight coefficient of each cluster, and the comprehensive score is calculated based on the weight coefficient and fuzzy membership to obtain the initial evaluation index of the store;

[0075] Conduct principal component analysis on the initial store evaluation indicators to obtain the evaluation features after dimensionality reduction, and generate feature importance ranking based on the evaluation features after dimensionality reduction;

[0076] Select an optimized evaluation feature set based on feature importance ranking, and build a decision tree model based on the optimized evaluation feature set;

[0077] Input the multidimensional time series data into the decision tree model to evaluate the store operation status and obtain the initial operation status evaluation results;

[0078] The confidence level of the initial operating status evaluation result is calculated to obtain an evaluation confidence interval, and the initial operating status evaluation result is corrected based on the evaluation confidence interval to obtain the store operating status evaluation result.

[0079] Specifically, the entropy weight method is used to analyze the store clustering results. The entropy weight method is a weight determination method based on information entropy, which determines the weight coefficient by calculating the entropy value of each cluster. Assume that stores, the clustering result is , the membership matrix of each cluster is ,in Indicates store Belongs to cluster Calculate the entropy value of each cluster:

[0080] ;

[0081] in, Representation clustering The entropy value of . Calculate the weight coefficient based on the entropy value:

[0082] ;

[0083] in, Representation clustering The weight coefficient is calculated based on the weight coefficient and fuzzy membership to obtain the initial evaluation index for each store:

[0084] ;

[0085] in, Indicates store The comprehensive score of the store is obtained by principal component analysis. Principal component analysis is a dimensionality reduction technique that converts the original high-dimensional data into a new low-dimensional feature space through linear transformation. Assume that the initial evaluation index is , the evaluation features after dimensionality reduction are obtained through principal component analysis ,in The goal of principal component analysis is to find new eigenvectors that maximize the variance of the transformed data in the new feature space. By calculating the covariance matrix and performing eigenvalue decomposition, new eigenvectors and eigenvalues ​​are obtained. The features are sorted according to the size of the eigenvalues ​​to generate a feature importance ranking. Based on the feature importance ranking, the optimized evaluation feature set is selected. The features with larger eigenvalues ​​are selected to form the optimized evaluation feature set. ,in . Build a decision tree model based on the optimized evaluation feature set. The decision tree model is a commonly used classification and regression model that generates a tree structure by dividing the data into hierarchical decisions. Assume that the optimized evaluation feature set is , the decision tree model generates a series of decision rules by learning the relationship between these features and the store operating status. The multidimensional time series data is input into the decision tree model to evaluate the store operating status and obtain the initial operating status evaluation results. Assume that the multidimensional time series data is , the decision tree model can predict the operating status of each store at different time points. The initial operating status evaluation result is . The confidence calculation of the initial operating status assessment results is performed to obtain the assessment confidence interval. The confidence calculation can be achieved through statistical methods. Assuming that the mean of the initial assessment results is , the standard deviation is , according to the confidence interval formula:

[0086] ;

[0087] in, is the critical value of the standard normal distribution, is the number of samples. The reliability range of the evaluation results is determined by the confidence interval. According to the evaluation confidence interval, the initial business status evaluation results are corrected to obtain the store business status evaluation results.

[0088] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0089] The multidimensional time series data is divided into three subsets: sales data, customer flow data, and inventory data, and wavelet transform is performed on each of them to obtain multi-scale time-frequency features;

[0090] A temporal convolutional network is constructed based on the multi-scale time-frequency features to obtain the first local time series features. The temporal convolutional network contains three one-dimensional convolutional layers, each layer uses 64 convolution kernels, the kernel size is 3, the step size is 1, and the ReLU activation function is adopted;

[0091] The first local time series feature is processed by a long short-term memory network to obtain a first global time series feature. The long short-term memory network includes two LSTM layers, and each LSTM layer includes 128 hidden units.

[0092] The first local temporal feature and the first global temporal feature are fused through the attention mechanism to obtain the first mixed temporal feature. The attention mechanism adopts scaled dot product attention, and the number of attention heads is 8;

[0093] The first mixed time series feature is processed by the self-attention mechanism to obtain the first attention weighted feature. The self-attention mechanism includes 3 self-attention layers, and the hidden dimension of each layer is 256;

[0094] The encoder-decoder structure is constructed according to the first attention weighted features to obtain the initial semantic segmentation network. The encoder contains 5 downsampling blocks, each of which contains two 3x3 convolutional layers and a maximum pooling layer. The decoder contains 5 upsampling blocks, each of which contains a transposed convolutional layer and two 3x3 convolutional layers.

[0095] The initial semantic segmentation network is optimized by residual connection to obtain the optimized semantic segmentation network. A residual connection is added between each encoder and decoder block. The residual connection uses 1x1 convolution for channel matching.

[0096] The optimized semantic segmentation network is conditionally embedded with the store attribute information to obtain a conditional semantic segmentation network. The store attribute information is processed by the embedding layer and then connected with the output of each decoder block.

[0097] The conditional semantic segmentation network is subjected to multi-task learning to simultaneously predict sales, customer flow, and inventory levels, and a multi-task prediction model is obtained. The multi-task prediction model is divided into three branches at the end of the network, each branch contains two fully connected layers, and the output layer uses a linear activation function.

[0098] The multi-task prediction model is integrated with the time series decomposition model to obtain the store business prediction model.

[0099] Specifically, the multidimensional time series data is divided into three subsets: sales data, customer flow data, and inventory data, and wavelet transform is performed on each of them to obtain multi-scale time-frequency features. Wavelet transform is a signal processing technology that decomposes time series data into different frequency components and extracts multi-scale time-frequency features. Suppose there is a time series data , a set of wavelet coefficients is obtained through wavelet transform ,in is the scale parameter, is the translation parameter. A temporal convolutional network is constructed based on the multi-scale time-frequency features. The temporal convolutional network contains three one-dimensional convolutional layers, each layer uses 64 convolution kernels, the kernel size is 3, the step size is 1, and the ReLU activation function is adopted. Assume that the input multi-scale time-frequency features are , No. The output of the convolutional layer can be expressed as:

[0100] ;

[0101] in, It is convolution kernels, * indicates convolution operation, is the bias term. The first local time series feature is obtained by stacking three convolutional layers. The first local time series feature is processed by a long short-term memory network (LSTM) to capture the long-term dependencies in the time series data. The LSTM network contains 2 LSTM layers, each of which contains 128 hidden units. Assume that the first local time series feature is , No. The output of the LSTM layer can be expressed as:

[0102] ;

[0103] Through the processing of two layers of LSTM, the first global time series feature is obtained. The first local time series feature and the first global time series feature are fused through the attention mechanism. The attention mechanism uses scaled dot product attention, and the number of attention heads is 8. Assume that the first local time series feature is , the first global timing feature is , the output of the attention mechanism can be expressed as:

[0104] ;

[0105] in, , , is the dimension of the key vector. Through the attention mechanism, the first mixed temporal feature is obtained. The first mixed temporal feature is processed by the self-attention mechanism, which contains 3 self-attention layers, and the hidden dimension of each layer is 256. The self-attention mechanism can further enhance the interdependence and importance between features, and obtain the first attention weighted feature. The encoder-decoder structure is constructed according to the first attention weighted feature to obtain the initial semantic segmentation network. The encoder contains 5 downsampling blocks, each of which contains two 3x3 convolutional layers and a maximum pooling layer, and the decoder contains 5 upsampling blocks, each of which contains a transposed convolutional layer and two 3x3 convolutional layers. The encoder is used to extract high-level features, while the decoder is used to restore the size of the feature map. A residual connection is added between each encoder and decoder block. The residual connection uses 1x1 convolution for channel matching, which helps to alleviate the gradient vanishing problem and promotes the training of deeper networks, and the optimized semantic segmentation network is obtained. The optimized semantic segmentation network is conditionally embedded with the store attribute information. After the store attribute information is processed by the embedding layer, it is connected with the output of each decoder block to enhance the adaptability of the model to different store characteristics and obtain a conditional semantic segmentation network. The conditional semantic segmentation network is subjected to multi-task learning to simultaneously predict sales, customer flow, and inventory levels to obtain a multi-task prediction model. The multi-task prediction model is divided into three branches at the end of the network, each branch contains two fully connected layers, and the output layer uses a linear activation function to output the prediction results of sales, customer flow, and inventory levels respectively. The multi-task prediction model is integrated with the time series decomposition model to obtain a store business prediction model. The time series decomposition model can decompose time series data into trend, seasonality, and residual components, thereby improving the accuracy of the prediction. By combining the multi-task prediction model with the time series decomposition model, the multiple characteristics of the time series are considered at the same time to improve the prediction effect.

[0106] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0107] Decompose the time series of multi-store business data to obtain trend, seasonality and random components;

[0108] The trend, seasonality and random components are input into the feature extraction layer of the store business prediction model and processed through three one-dimensional convolutional layers to obtain the second local time series features;

[0109] The second local time series feature is input into the time series modeling layer of the store business prediction model and processed through two LSTM layers to obtain the second global time series feature.

[0110] Inputting the second local temporal feature and the second global temporal feature into the attention fusion mechanism to obtain a second mixed temporal feature;

[0111] The second mixed time series feature is input into three self-attention layers for attention weighting processing to obtain the second attention weighted feature;

[0112] The second attention weighted features are input into the encoder-decoder structure of the semantic segmentation layer, and processed by 5 downsampling blocks and 5 upsampling blocks to obtain the initial semantic features;

[0113] Extract store attribute information based on store clustering results, conditionally embed the store attribute information with the initial semantic features to obtain conditional semantic features;

[0114] The conditional semantic features are input into the multi-task prediction layer, and sales, customer flow, and inventory levels are predicted respectively through three parallel fully connected layers to obtain preliminary prediction results;

[0115] The preliminary forecast results are integrated to obtain the store business forecast results, which include point forecast values ​​and forecast intervals.

[0116] Specifically, we perform time series decomposition on the business data of multiple stores and decompose the data into trend, seasonality and random components. Assume that there is time series data , after decomposition, it can be expressed as:

[0117] ;

[0118] in, It represents the trend component, reflecting the long-term trend of data; Represents seasonal components, reflecting periodic changes, Denotes random components, representing unpredictable noise or short-term fluctuations. The trend, seasonality, and random components are input into the feature extraction layer of the store business forecasting model. The feature extraction layer consists of three one-dimensional convolutional layers, each of which uses 64 convolution kernels with a kernel size of 3, a step size of 1, and a ReLU activation function. Assume that the decomposed components of the input are , No. The output of the convolutional layer can be expressed as:

[0119] ;

[0120] in, It is convolution kernels, * indicates convolution operation, is the bias term. Through these three convolutional layers, the second local time series feature is obtained. The second local time series feature is subjected to time series modeling. Two LSTM layers are used to capture the long-term dependencies in the time series data. Each LSTM layer contains 128 hidden units. Assume that the second local time series feature is , No. The output of the LSTM layer can be expressed as:

[0121] ;

[0122] After two layers of LSTM processing, the second global temporal feature is obtained. The second local temporal feature and the second global temporal feature are input into the attention fusion mechanism. The attention mechanism uses scaled dot product attention, and the number of attention heads is 8. Assume that the second local temporal feature is , the second global timing characteristic is , the output of the attention mechanism can be expressed as:

[0123] ;

[0124] in, , , is the dimension of the key vector. Through the attention mechanism, the second mixed time series feature is obtained. The second mixed time series feature is processed by the self-attention mechanism. The self-attention mechanism contains 3 self-attention layers, each with a hidden dimension of 256. The self-attention mechanism can further enhance the interdependence and importance between features, and obtain the second attention weighted feature. The second attention weighted feature is input into the encoder-decoder structure of the semantic segmentation layer. The encoder contains 5 downsampling blocks, each of which contains two 3x3 convolutional layers and a maximum pooling layer. The decoder contains 5 upsampling blocks, each of which contains a transposed convolutional layer and two 3x3 convolutional layers. Through the processing of the encoder and decoder, the size of the feature map can be gradually extracted and restored to obtain the initial semantic feature. Store attribute information is extracted according to the store clustering results, and the store attribute information is conditionally embedded with the initial semantic feature. After the store attribute information is processed by the embedding layer, it is connected with the output of each decoder block to enhance the adaptability of the model to different store characteristics and obtain conditional semantic features. The conditional semantic features are input into the multi-task prediction layer, and sales, customer flow, and inventory levels are predicted respectively through three parallel fully connected layers. Each branch contains two fully connected layers, and the output layer uses a linear activation function to output preliminary prediction results for sales, customer flow, and inventory levels respectively. The preliminary prediction results are integrated to obtain the store business prediction results. The integration method improves the accuracy and robustness of the prediction by combining the various prediction results. The store business prediction results include point prediction values ​​and prediction intervals.

[0125] In a specific embodiment, the process of executing step 600 may specifically include the following steps:

[0126] Perform time series alignment on the store operation status evaluation results and the store business forecast results to obtain aligned time series data;

[0127] An anomaly detection model is built based on the aligned time series data and fuzzy membership, and the reconstruction error is calculated using the autoencoder;

[0128] Perform dynamic threshold calculation on the reconstruction error to obtain the abnormality judgment standard, and perform binarization processing on the reconstruction error according to the abnormality judgment standard to obtain the abnormality detection result;

[0129] Extract abnormal features from the abnormal detection results to obtain abnormal feature vectors, and analyze the abnormal feature vectors to obtain abnormal type classification;

[0130] Build a decision rule base based on the abnormal type classification to obtain resource allocation suggestions, and perform constraint optimization on the resource allocation suggestions to obtain an optimization plan;

[0131] Generate anomaly prediction reports based on anomaly detection results and anomaly type classification.

[0132] Specifically, the store operation status evaluation results and store business forecast results are aligned in time series to ensure the consistency of the two time series data in the time dimension. and , respectively represent the store operation status evaluation results and store business forecast results, which are aligned by interpolation or dynamic time warping to obtain the aligned time series data . An anomaly detection model is constructed based on the aligned time series data and fuzzy membership. The anomaly detection model uses an autoencoder to calculate the reconstruction error. An autoencoder is an unsupervised learning model that encodes the input data into a low-dimensional space and then reconstructs it back to the original space, calculating the reconstruction error to measure the degree of data anomaly. Assume that the aligned time series data is , the input of the autoencoder is , the reconstructed output is , the reconstruction error can be expressed as:

[0133] ;

[0134] in, is the difference between the input and the reconstructed output. By minimizing the reconstruction error, the autoencoder can learn the normal mode, thereby showing a large reconstruction error when anomalies occur. The dynamic threshold calculation is performed on the reconstruction error to obtain the abnormality judgment standard. The dynamic threshold calculation can be determined based on the statistical characteristics of the reconstruction error. Assume that the distribution of the reconstruction error is , its mean can be calculated by statistical methods and standard deviation The abnormality judgment criteria can be set as:

[0135] ;

[0136] in, It is a tuning parameter used to control the sensitivity of anomaly detection. When the reconstruction error exceeds the threshold, it is judged as abnormal. The reconstruction error is binarized according to the abnormality judgment standard to obtain the anomaly detection result. . Perform abnormal feature extraction on the abnormal detection result to obtain the abnormal feature vector. Extract the vector that can describe the abnormal characteristics from the detected anomaly. Assume that the abnormal detection result is , through clustering or classification methods, the anomalies are divided into different feature categories to obtain the anomaly feature vector . Analyze the abnormal feature vector to obtain the abnormal type classification. Classify the abnormal feature vector to identify different types of abnormalities. Use classification algorithms such as support vector machine (SVM) or decision tree to classify the abnormal feature vector to obtain the abnormal type . According to the classification of exception types, a decision rule library is constructed to obtain resource allocation suggestions. The decision rule library contains processing rules and resource allocation suggestions for different exception types. By analyzing historical data and expert experience, corresponding decision rules are formulated. For example, for inventory anomalies, it can be recommended to increase inventory; for sales anomalies, it can be recommended to adjust marketing strategies. Constrained optimization is performed on the resource allocation suggestions to obtain the optimization plan. The purpose of constrained optimization is to find the optimal resource allocation plan while satisfying resource constraints and other constraints. Assume that the resource allocation suggestion is , the constraints are , the optimization goal is , the optimization problem can be solved by linear programming:

[0137] ;

[0138] By solving the optimization problem, the optimal resource allocation plan is obtained. Anomaly prediction reports are generated based on anomaly detection results and anomaly type classification. The anomaly prediction report records the detected anomalies, anomaly types, processing suggestions and optimization plans to provide decision support for managers. The report content includes the anomaly type at the time of anomaly detection, possible causes and recommended solutions.

[0139] The above describes the data analysis method based on the chain operation benefit supervision system in the embodiment of the present application. The following describes the data analysis device 10 based on the chain operation benefit supervision system in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the data analysis device 10 based on the chain operation benefit supervision system includes:

[0140] The collection module 11 is used to collect and pre-process the multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data;

[0141] The calculation module 12 is used to perform multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain the store clustering results and fuzzy membership;

[0142] Evaluation module 13, used to build a decision tree model based on the store clustering results and fuzzy membership, and evaluate the store operation status of the multi-dimensional time series data to obtain the store operation status evaluation results;

[0143] Construction module 14 is used to extract time series features from multi-dimensional time series data, build a semantic segmentation network, and obtain a store business prediction model;

[0144] Prediction module 15, used to perform in-depth prediction on multiple store business data based on the store business prediction model and store clustering results to obtain store business prediction results;

[0145] Configuration module 16 is used to perform comprehensive abnormality prediction and resource optimization configuration based on the store operation status evaluation results, store business prediction results and fuzzy membership, and obtain an abnormality prediction report and optimization plan.

[0146] Through the collaboration of the above components, the data is standardized and high-quality through the collection and preprocessing of multi-dimensional data such as sales, inventory, and customer flow. The use of technologies such as dynamic time warping and kernel function transformation improves the accuracy and flexibility of store clustering, which helps to more accurately identify different types of stores. Through entropy weight analysis and principal component analysis, the selection of evaluation features is optimized, and the accuracy and interpretability of business status evaluation are improved. Combining time convolutional networks, long short-term memory networks, and self-attention mechanisms, the feature extraction and prediction capabilities of time series data are improved, and more accurate business forecasts are achieved. Through multi-task prediction models, sales, customer flow, and inventory levels are predicted at the same time, which improves the comprehensiveness and efficiency of predictions. The use of autoencoders and dynamic threshold calculations improves the accuracy and sensitivity of anomaly detection. At the same time, more reasonable resource allocation is achieved through decision rule bases and constraint optimization. From time series decomposition to multi-scale time window division, and then to multi-task learning, multi-level and multi-scale analysis of data is achieved. Through conditional embedding and residual connection technologies, the scalability of the model and its adaptability to different store characteristics are enhanced.

[0147] The present application also provides an electronic device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the data analysis method based on the chain operation efficiency supervision system in the above-mentioned embodiments.

[0148] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein instructions are stored in the computer, and when the instructions are executed on a computer, the computer executes the steps of the data analysis method based on the chain operation benefit supervision system.

[0149] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0151] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data analysis method based on a chain operation benefit supervision system, characterized in that: The method comprises: Collect and preprocess the multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data; Performing multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain store clustering results and fuzzy membership; Building a decision tree model based on the store clustering results and the fuzzy membership degree, and performing store operation status evaluation on the multidimensional time series data to obtain a store operation status evaluation result; Extracting time series features from the multidimensional time series data, constructing a semantic segmentation network, and obtaining a store business prediction model; specifically comprising: dividing the multidimensional time series data into three subsets: sales data, customer flow data, and inventory data, and performing wavelet transforms on each of them to obtain multi-scale time-frequency features; constructing a time convolution network based on the multi-scale time-frequency features to obtain a first local time series feature, wherein the time convolution network comprises three one-dimensional convolution layers, each layer uses 64 convolution kernels, the kernel size is 3, the step length is 1, and the ReLU activation function is adopted; extracting time series features from the multi-dimensional time series data, constructing a semantic segmentation network, and obtaining a store business prediction model; specifically comprising: dividing the multi- ... and constructing a semantic segmentation network; The first local time series feature and the first global time series feature are processed by a long short-term memory network to obtain a first global time series feature, wherein the long short-term memory network includes two LSTM layers, and each LSTM layer includes 128 hidden units; the first local time series feature and the first global time series feature are fused by an attention mechanism to obtain a first mixed time series feature, wherein the attention mechanism adopts scaled dot product attention, and the number of attention heads is 8; the first mixed time series feature is processed by a self-attention mechanism to obtain a first attention weighted feature, wherein the self-attention mechanism includes three self-attention layers, and the hidden dimension of each layer is 2 56; construct an encoder-decoder structure according to the first attention weighted feature to obtain an initial semantic segmentation network, wherein the encoder includes 5 downsampling blocks, each of which includes two 3x3 convolutional layers and one maximum pooling layer, and the decoder includes 5 upsampling blocks, each of which includes one transposed convolutional layer and two 3x3 convolutional layers; perform residual connection optimization on the initial semantic segmentation network to obtain an optimized semantic segmentation network, add residual connections between each encoder and decoder block, and use 1x1 convolution for channel matching in the residual connection; conditionally embed the optimized semantic segmentation network with store attribute information to obtain a conditional semantic segmentation network, and the store attribute information is processed by the embedding layer and then connected to the output of each decoder block; perform multi-task learning on the conditional semantic segmentation network, and simultaneously predict sales, customer flow and inventory levels to obtain a multi-task prediction model, which is divided into three branches at the end of the network, each of which includes 2 fully connected layers, and the output layer uses a linear activation function; integrate the multi-task prediction model with the time series decomposition model to obtain a store business prediction model; Based on the store business prediction model and the store clustering results, the multi-store business data is deeply predicted to obtain the store business prediction results; specifically including: performing time series decomposition on the multi-store business data to obtain trend, seasonality and random components; inputting the trend, seasonality and random components into the feature extraction layer of the store business prediction model, processing through three one-dimensional convolutional layers, and obtaining a second local time series feature; inputting the second local time series feature into the time series modeling layer of the store business prediction model, processing through two LSTM layers, and obtaining a second global time series feature; inputting the second local time series feature and the second global time series feature into the attention fusion mechanism to obtain a second mixed time series feature; The time series features are input into three self-attention layers for attention weighting processing to obtain the second attention weighted features; the second attention weighted features are input into the encoder-decoder structure of the semantic segmentation layer, and are processed through five downsampling blocks and five upsampling blocks to obtain the initial semantic features; store attribute information is extracted according to the store clustering results, and the store attribute information is conditionally embedded with the initial semantic features to obtain conditional semantic features; the conditional semantic features are input into the multi-task prediction layer, and sales, customer flow and inventory levels are predicted respectively through three parallel fully connected layers to obtain preliminary prediction results; the preliminary prediction results are integrated to obtain store business prediction results, and the store business prediction results include point prediction values ​​and prediction intervals; Based on the store operation status evaluation results, the store business forecast results and the fuzzy membership, comprehensive abnormality prediction and resource optimization configuration are performed to obtain an abnormality prediction report and an optimization plan.

2. The data analysis method based on the chain operation benefit supervision system according to claim 1 is characterized in that: The multi-dimensional data of each store of the chain operation enterprise is collected and preprocessed to obtain multi-dimensional time series data, including: Collect sales data, inventory data and customer flow data of each store of the chain enterprise to obtain original multi-dimensional data; Performing missing value processing on the original multidimensional data to obtain filled data, and performing outlier detection and correction on the filled data to obtain corrected data; Performing multi-dimensional data fusion on the corrected data to obtain fused data, and performing time series decomposition and time series reconstruction on the fused data to obtain reconstructed data; The reconstructed data is divided into multi-scale time windows to obtain multi-scale time window data, and time series correlation analysis and feature selection are performed on the multi-scale time window data to obtain multi-dimensional time series data.

3. The data analysis method based on the chain operation benefit supervision system according to claim 1 is characterized in that: The multi-dimensional similarity calculation and fuzzy clustering are performed on the multi-dimensional time series data to obtain store clustering results and fuzzy membership, including: Segmenting and dynamically time warping the multidimensional time series data to obtain an aligned time series; Calculating a Euclidean distance matrix according to the aligned time series to obtain an initial similarity matrix, and performing a kernel function transformation on the initial similarity matrix to obtain a similarity matrix in a high-dimensional feature space; Constructing a k-nearest neighbor graph according to the similarity matrix in the high-dimensional feature space to obtain a store relationship network, and performing spectral clustering on the store relationship network to obtain an initial clustering result; Calculating a fuzzy membership matrix according to the initial clustering result to obtain an initial membership, and performing fuzzy c-means iterative optimization on the initial membership to obtain a fuzzy membership; The number of clusters of each store is reallocated and dynamically adjusted according to the fuzzy membership degree to obtain a store clustering result.

4. The data analysis method based on the chain operation benefit supervision system according to claim 3 is characterized in that: The decision tree model is constructed based on the store clustering results and the fuzzy membership, and the store operation status evaluation is performed on the multi-dimensional time series data to obtain the store operation status evaluation result, including: Performing entropy weight analysis on the store clustering results to obtain weight coefficients of each cluster, and calculating a comprehensive score based on the weight coefficients and the fuzzy membership degree to obtain an initial store evaluation index; Performing principal component analysis on the initial store evaluation indicators to obtain evaluation features after dimensionality reduction, and generating feature importance rankings based on the evaluation features after dimensionality reduction; Selecting an optimized evaluation feature set based on the feature importance ranking, and constructing a decision tree model based on the optimized evaluation feature set; Inputting the multidimensional time series data into the decision tree model to evaluate the store operation status and obtain an initial operation status evaluation result; A confidence calculation is performed on the initial business status evaluation result to obtain an evaluation confidence interval, and the initial business status evaluation result is corrected according to the evaluation confidence interval to obtain a store business status evaluation result.

5. The data analysis method based on the chain operation benefit supervision system according to claim 1 is characterized in that: The comprehensive abnormality prediction and resource optimization configuration are performed based on the store operation status evaluation result, the store business prediction result and the fuzzy membership to obtain an abnormality prediction report and an optimization plan, including: Performing time series alignment on the store operation status evaluation result and the store business forecast result to obtain aligned time series data; Building an anomaly detection model based on the aligned time series data and the fuzzy membership, and using an autoencoder to calculate a reconstruction error; Performing dynamic threshold calculation on the reconstruction error to obtain an abnormality determination standard, and performing binarization processing on the reconstruction error according to the abnormality determination standard to obtain an abnormality detection result; Extracting abnormal features from the abnormal detection results to obtain abnormal feature vectors, and analyzing the abnormal feature vectors to obtain abnormal type classification; A decision rule base is constructed according to the abnormal type classification to obtain resource allocation suggestions, and constraint optimization is performed on the resource allocation suggestions to obtain an optimization solution; An anomaly prediction report is generated according to the anomaly detection result and the anomaly type classification.

6. A data analysis device based on a chain operation benefit supervision system, characterized in that: The device is used to execute the data analysis method based on the chain operation benefit supervision system according to any one of claims 1 to 5, and comprises: The collection module is used to collect and preprocess the multi-dimensional data of each store of the chain operation enterprise to obtain multi-dimensional time series data; A calculation module, used to perform multi-dimensional similarity calculation and fuzzy clustering on the multi-dimensional time series data to obtain store clustering results and fuzzy membership; An evaluation module, used to construct a decision tree model based on the store clustering results and the fuzzy membership, and to evaluate the store operation status of the multidimensional time series data to obtain a store operation status evaluation result; A construction module is used to extract time series features from the multi-dimensional time series data, construct a semantic segmentation network, and obtain a store business prediction model; A prediction module, used to perform in-depth prediction on multiple store business data based on the store business prediction model and the store clustering result to obtain store business prediction results; A configuration module is used to perform comprehensive abnormality prediction and resource optimization configuration based on the store operation status evaluation results, the store business prediction results and the fuzzy membership, and obtain an abnormality prediction report and an optimization plan.

7. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the electronic device executes the data analysis method based on the chain operation benefit supervision system as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the data analysis method based on the chain operation benefit supervision system as described in any one of claims 1 to 5 is implemented.

Images