Intelligent early warning method for financial risks of tourism enterprises based on deep learning
By constructing a multi-source fusion feature set and training an LSTM-attention model to generate dynamic thresholds, the problem of lagging early warning of financial risks for tourism enterprises in existing technologies is solved, and real-time adaptation and accurate early warning of seasonal fluctuations are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TOURISM COLLEGE OF ZHEJIANG
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-16
AI Technical Summary
Existing financial risk early warning technologies cannot effectively integrate multi-source data from tourism enterprises, resulting in delayed early warning responses, inability to adapt to seasonal fluctuations, low accuracy in cash flow forecasting, and ineffective early warnings.
Based on deep learning methods, we collect structured financial data, business operation data, and external dynamic data from tourism enterprises, construct a multi-source fusion feature set, train an LSTM-attention financial risk early warning model, generate dynamic thresholds for peak and off-peak seasons, and conduct real-time risk warnings.
It enables accurate early warning of financial risks for tourism enterprises, enhances the credibility of early warning results and the initiative of enterprises in resisting risks, and improves the efficiency of refined management and control of financial risks.
Smart Images
Figure CN122222754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and more specifically, to a method for intelligent early warning of financial risks for tourism enterprises based on deep learning. Background Technology
[0002] The tourism industry is characterized by strong seasonality, sensitivity to external shocks, and significant differentiation in business formats. Its financial risks exhibit unique characteristics such as "drastic fluctuations between peak and off-peak seasons, hidden off-balance-sheet risks, and strong cross-entity transmission." This makes it difficult for general financial risk early warning technologies to meet the precise early warning needs of tourism enterprises.
[0003] Financial risk early warning is a core part of enterprise risk management. Existing financial risk early warning technologies mainly originate from traditional industries such as manufacturing and finance. They construct linear early warning models by selecting core financial indicators such as debt-to-equity ratio, current ratio, and net profit margin to determine the risk level.
[0004] However, it still has some shortcomings in practical use, such as insufficient multi-source data fusion capability: the existing data processing methods of tourism enterprises cannot effectively integrate the tourism industry's unique feature system, resulting in a lag in early warning response; Existing financial risk early warning technologies for tourism enterprises use fixed thresholds and weights, which cannot adapt to the fluctuations in tourism seasons, resulting in low accuracy of cash flow forecasting and failure of early warning. Summary of the Invention
[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a deep learning-based intelligent early warning method for financial risks of tourism enterprises, which addresses the problems raised in the background section.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based intelligent early warning method for financial risks in tourism enterprises, comprising the following steps: Step S01: Data collection for tourism enterprises: This step involves collecting structured financial data, business operation data, and external dynamic data from target tourism enterprises, and then normalizing and removing outliers. Step S02: Construction of tourism enterprise suitability features: This step is used to extract the first feature set representing the financial health status of the target tourism enterprise from the time series dataset and to construct the risk suitability feature set. Step S03: Deep learning model training: Based on the time series dataset corresponding to the final multi-source fusion feature set, the model is trained and optimized to obtain a financial risk early warning model that adapts to the characteristics of the target tourism enterprise. Step S04: Peak and Off-Peak Season Dynamic Threshold Generation Mechanism: Based on the peak and off-peak season fluctuation coefficients of the target tourism enterprise, obtain the peak and off-peak season dynamic thresholds of the target tourism enterprise; Step S05: Financial Risk Early Warning: This step involves inputting the first feature set collected in real time into the trained improved LSTM-attention financial risk early warning model to obtain the real-time risk probability, which is then compared with the dynamic thresholds for peak and off-peak seasons for risk processing.
[0007] Preferably, step S01: data collection from tourism enterprises specifically includes: S21: Collect daily structured financial data of target tourism enterprises, including debt-to-equity ratio, current ratio, net operating cash flow, and gross profit margin; collect daily business operation data from scenic spot ticketing systems and secondary consumption systems, including total visitor flow, instantaneous visitor flow, ticket revenue, sales of cultural and creative products, and catering revenue; collect external dynamic data, including environmental data, public opinion data, and market data, by accessing third-party platforms through API interfaces. S22: Use box plots to remove outliers from financial structured data and use interpolation to fill in missing values in business operation data and external data; S23: Standardize the financial structured data, business operation data, and external dynamic data of the target tourism enterprise and map the data to the [0,1] interval; S24: Using "day" as the time granularity, align the financial structured data, business operation data, and external dynamic data along the time axis to generate a unified time series dataset. Each record in the time series dataset contains a timestamp and multi-dimensional feature values for the corresponding time point.
[0008] Preferably, step S02: constructing the adaptability features of tourism enterprises specifically involves: S31: Peak and Off-Peak Season Feature Construction: A clustering algorithm is used to divide the peak and off-peak seasons. The clustering features are average monthly revenue and average monthly customer traffic. The number of clusters is set to 3 (peak season, shoulder season, off-peak season). Based on the average monthly customer traffic / average monthly revenue of the cluster centers, the month with the highest output is marked as the peak season, the month with the lowest output is marked as the off-peak season, and the rest are shoulder seasons. S32: Extract the average monthly revenue during peak season, the average monthly revenue during off-season, and the average monthly revenue throughout the year, and calculate the peak-off-season fluctuation coefficient of the target tourism enterprise; S33: Add seasonal labels and seasonal fluctuation coefficients to each record in the time series dataset.
[0009] Preferably, the construction of the tourism enterprise adaptability features further includes: Supplier association feature construction: S41: Obtain the annual cooperation duration with suppliers, the total annual cooperation duration with the target tourism enterprise, the annual order volume with suppliers, and the total annual order volume of the target tourism enterprise from the time series dataset, and calculate the supplier association strength coefficient of the target tourism enterprise; S42: Add a supplier association strength coefficient to each record in the time series dataset, and finally output a multi-source fusion feature set.
[0010] Preferably, the model training specifically involves: S51: Divide the time series dataset corresponding to the final multi-source fusion feature set into a training set, a validation set, and a test set; the training set is used for fitting and learning the model parameters, the validation set is used for debugging and optimizing the model hyperparameters, and the test set is used for evaluating and verifying the final warning performance of the model; the temporal continuity of the time series data is maintained during the partitioning process to avoid data leakage affecting the model training effect; S52: Constructing an improved deep learning model: including an input layer, a seasonal adaptive adjustment layer, an LSTM feature learning layer, an attention layer, and a fully connected output layer. The parameters of each layer are initialized as follows: Input layer initialization: The input dimension is the number of features n in the final multi-source fusion feature set, used to receive temporal feature sequences. ,in This is represented as a multi-dimensional feature vector at time i, where the vector dimension is the same as the input dimension. Initialization of the seasonal adaptive adjustment layer: Introducing seasonal adjustment factors , Based on the peak and off-peak season labels, the specific assignment rules are as follows: Peak Season , shoulder season Off-season The feature vectors received by the input layer are weighted and adjusted using this adjustment factor, as shown in the formula: ,in, This represents the dynamic weight at time i; LSTM feature learning layer initialization: A two-layer stacked LSTM network is set as the core of feature learning. The number of neurons in the first LSTM hidden layer is 64, and the number of neurons in the second LSTM hidden layer is 32. The initial values of the parameters of the forget gate, input gate and output gate of the LSTM network are set using the Xavier initialization method to ensure that the model training process is stable and converges after parameter initialization. Attention layer initialization: Construct an attention layer to adaptively extract key risk information from time-series features. By calculating the attention weights of the feature vectors at each time step, a weighted aggregation of financial risk features is achieved. Fully connected output layer initialization: Set up two fully connected layers. The first fully connected layer has 16 neurons, and the second fully connected layer has 1 neuron. The second fully connected layer uses the Sigmoid activation function to output the risk probability at time i.
[0011] Preferably, the financial risk early warning model is as follows: The time-series feature sequences of the training set are input into the initialized improved deep learning model. The loss function is minimized through backpropagation, driving the model parameters to update iteratively. The model is iteratively trained until the loss on the validation set converges, and the current optimal model parameters are saved to obtain the trained improved LSTM-attention financial risk warning model. The improved deep learning model also includes an optimizer to improve the model's convergence speed and generalization ability. The loss function is used to calculate the deviation between the risk probability output by the model and the true risk label; The validation set loss converges as follows: during training, the change in the validation set loss value is monitored in real time, and when the fluctuation range of the validation set loss value is less than 10 consecutive rounds... At that time, the model training is considered to have converged.
[0012] Preferably, step S04: the dynamic threshold generation mechanism for peak and off-peak seasons specifically includes: Based on the peak and off-peak season fluctuation coefficients and the peak season adjustment factor, the peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated.
[0013] Preferably, step S05: financial risk warning specifically includes: S81: Extract the real-time first feature set representing the financial health status of the target tourism enterprise from the time series dataset, input it into the trained improved LSTM-attention financial risk early warning model, and obtain the real-time risk probability; S82: Combining real-time risk probability with dynamic thresholds for peak and off-peak seasons: If the real-time risk probability is less than the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold, a low-risk warning signal will be output to remind relevant management personnel of tourism enterprises to conduct routine monitoring. If the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold ≤ real-time risk probability < preset high-risk threshold, then a risk warning signal will be output and a risk cause analysis will be pushed to relevant management personnel of tourism enterprises. If the real-time risk probability is greater than or equal to the preset high-risk threshold, a high-risk warning signal will be output, and the emergency response mechanism will be triggered immediately. S53: Output early warning results, generate visual early warning reports, and push them to the enterprise financial terminal and management mobile terminal.
[0014] The technical effects and advantages of this invention are as follows: 1. This invention provides a deep learning-based intelligent early warning method for financial risks of tourism enterprises. By collecting structured financial data, business operation data, and external dynamic data of target tourism enterprises, a unified time-series dataset is generated. Peak and off-peak season features and supplier-related features are constructed, and finally, a multi-source fusion feature set is output. Through full coverage of multi-source data and comprehensive data collection, full-dimensional coverage of tourism enterprise financial data is achieved. By unifying the time-series dataset, the relevance and timeliness of data are ensured, laying the foundation for accurate early warning. The targeted construction of tourism enterprise features enhances the correlation between features and tourism financial risks, providing high-quality data and feature support for subsequent accurate early warning by deep learning models. 2. This invention provides a deep learning-based intelligent early warning method for financial risks in tourism enterprises. Based on the time-series dataset corresponding to the final multi-source fusion feature set, the method trains and optimizes the model to obtain a financial risk early warning model adapted to the characteristics of the target tourism enterprise. Based on the peak and off-peak season fluctuation coefficients and peak and off-peak season adjustment factors of the target tourism enterprise, the method calculates the peak season dynamic threshold, off-peak season dynamic threshold, and flat season dynamic threshold of the target tourism enterprise, outputs the early warning results, generates a visual early warning report, and pushes it to the enterprise's financial terminal and the management's mobile terminal. Through dynamic weight adjustment and dynamic threshold generation, the early warning model achieves real-time adaptation to seasonal fluctuations, improving the credibility of the early warning results. Based on the daily time-series dataset, it has the ability to quickly process and predict real-time data, enhances the enterprise's initiative in risk resistance, and improves the efficiency of refined management of financial risks in tourism enterprises. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the intelligent early warning method for financial risks of tourism enterprises based on deep learning, as described in this invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] Please see Figure 1 As shown, this invention provides a deep learning-based intelligent early warning method for financial risks in tourism enterprises, comprising the following steps: Step S01: Tourism Enterprise Data Collection: This step involves collecting the target tourism enterprise's structured financial data, business operation data, and external dynamic data, and then normalizing the data and removing outliers.
[0018] In one possible design, step S01: data collection from tourism enterprises specifically includes: S11: Collect daily structured financial data of target tourism enterprises, including debt-to-equity ratio, current ratio, net operating cash flow, and gross profit margin; collect daily business operation data from scenic spot ticketing systems and secondary consumption systems, including total visitor flow, instantaneous visitor flow, ticket revenue, sales of cultural and creative products, and catering revenue; collect external dynamic data, including environmental data, public opinion data, and market data, by accessing third-party platforms through API interfaces. S12: Use box plots to remove outliers from structured financial data, such as sudden increases / decreases in revenue due to accounting adjustments, and use interpolation to fill in missing values in business operation data and external data. S13: Standardize the financial structured data, business operation data, and external dynamic data of the target tourism enterprise and map the data to the [0,1] range; S14: Using "day" as the time granularity, align the financial structured data, business operation data, and external dynamic data along the time axis to generate a unified time series dataset. Each record in the time series dataset contains a timestamp and multi-dimensional feature values for the corresponding time point.
[0019] Step S02: Construction of tourism enterprise adaptability features: This step is used to extract the first feature set representing the financial health status of the target tourism enterprise from the time series dataset and to construct the risk adaptability feature set.
[0020] In one possible design, step S02: constructing the adaptability features of tourism enterprises specifically involves: S21: Extract a first feature set representing the financial health status of the target tourism enterprise from the time series dataset. The first feature set includes financial features, business features, and external features. S22: Peak and Off-Peak Season Feature Construction: A clustering algorithm is used to divide the peak and off-peak seasons. The clustering features are average monthly revenue and average monthly customer traffic. The number of clusters is set to 3 (peak season, shoulder season, off-peak season). Based on the average monthly customer traffic / average monthly revenue of the cluster centers, the month with the highest output is marked as the peak season, the month with the lowest output is marked as the off-peak season, and the rest are shoulder seasons. S23: Extract the average monthly revenue during peak season, the average monthly revenue during off-season, and the average monthly revenue throughout the year, and calculate the peak-off-season fluctuation coefficient of the target tourism enterprise; S24: Add seasonal labels and seasonal fluctuation coefficients to each record in the time series dataset.
[0021] In one possible design, the construction of the tourism enterprise adaptability features also includes: Supplier association feature construction: S25: Obtain the annual cooperation duration with suppliers, the total annual cooperation duration with the target tourism enterprise, the annual order volume with suppliers, and the total annual order volume of the target tourism enterprise from the time series dataset, and calculate the supplier association strength coefficient of the target tourism enterprise; S26: Add a supplier association strength coefficient to each record in the time series dataset, and finally output a multi-source fusion feature set.
[0022] In this embodiment, it should be specifically noted that the calculation formula for the peak and off-peak season fluctuation coefficient is as follows:
[0023] in, This is expressed as the seasonal fluctuation coefficient. This represents the average monthly income during peak season. This represents the average monthly income during the off-season. This represents the average monthly income throughout the year.
[0024] The formula for calculating the supplier association strength coefficient is as follows:
[0025] in, This is represented as the association strength coefficient with the x-th supplier. This represents the annual cooperation duration with the xth supplier. This represents the total annual cooperation duration with the target tourism enterprise. This is expressed as the annual order volume of the xth supplier. This represents the total number of orders placed by the target tourism enterprise in a given year.
[0026] Step S03: Deep learning model training: Based on the time series dataset corresponding to the final multi-source fusion feature set, the model is trained and optimized to obtain a financial risk early warning model that adapts to the characteristics of the target tourism enterprise.
[0027] In one possible design, step S03: deep learning model training specifically includes: S31: Divide the time series dataset corresponding to the final multi-source fusion feature set into a training set, a validation set, and a test set; the training set is used for fitting and learning the model parameters, the validation set is used for debugging and optimizing the model hyperparameters, and the test set is used for evaluating and verifying the final warning performance of the model; the temporal continuity of the time series data is maintained during the partitioning process to avoid data leakage affecting the model training effect; S32: Constructing an improved deep learning model: including an input layer, a seasonal adaptive adjustment layer, an LSTM feature learning layer, an attention layer, and a fully connected output layer. The parameters of each layer are initialized as follows: Input layer initialization: The input dimension is the number of features n in the final multi-source fusion feature set, used to receive temporal feature sequences. ,in This is represented as a multi-dimensional feature vector at time i, where the vector dimension is the same as the input dimension. Initialization of the seasonal adaptive adjustment layer: Introducing seasonal adjustment factors , Based on the peak and off-peak season labels, the specific assignment rules are as follows: Peak Season , shoulder season Off-season The feature vectors received by the input layer are weighted and adjusted using this adjustment factor, as shown in the formula: ,in, The dynamic weight at time i is used to achieve differentiated adaptation of the weights for peak and off-peak season characteristics, thereby strengthening the ability to represent core risk characteristics such as off-peak cash flow and peak-peak channel settlement. LSTM feature learning layer initialization: A two-layer stacked LSTM network is set as the core of feature learning. The number of neurons in the first LSTM hidden layer is 64, and the number of neurons in the second LSTM hidden layer is 32. The initial values of the parameters of the forget gate, input gate and output gate of the LSTM network are set using the Xavier initialization method to ensure that the model training process is stable and converges after parameter initialization. Attention Layer Initialization: An attention layer is constructed to adaptively extract key risk information from time-series features. By calculating the attention weights of the feature vectors at each time step, a weighted aggregation of financial risk features is achieved. The formula for calculating the attention weights is as follows: ,in, The attention score at time i is represented by the inner product of the hidden state output by the LSTM feature learning layer and the attention weight matrix. Fully connected output layer initialization: Two fully connected layers are set up. The first fully connected layer has 16 neurons and the second fully connected layer has 1 neuron. The second fully connected layer uses the Sigmoid activation function to output the risk probability at time i. The risk probability ranges from [0,1], representing the likelihood of financial risk faced by the tourism enterprise at time i. S33: Input the time-series feature sequence of the training set into the initialized improved deep learning model, minimize the loss function through backpropagation, and drive the model parameters to update iteratively; iterate the training until the loss on the validation set converges, save the current optimal model parameters, and obtain the trained improved LSTM-attention financial risk warning model; the improved deep learning model also includes an optimizer to improve the model's convergence speed and generalization ability; Specifically: The loss function is used to calculate the deviation between the risk probability output by the model and the true risk label. The formula for the loss function is:
[0028] in, Represented as a loss function, This represents the true risk label at time i. Let i be the risk probability at time i; Specifically: The validation set loss converges as follows: During training, the change in the validation set loss value is monitored in real time. When the fluctuation range of the validation set loss value is less than 10 consecutive rounds, the convergence occurs. At that time, the model training is considered to have converged.
[0029] Step S04: Peak and Off-Peak Season Dynamic Threshold Generation Mechanism: Based on the peak and off-peak season fluctuation coefficients of the target tourism enterprise, the peak and off-peak season dynamic thresholds of the target tourism enterprise are obtained.
[0030] In one possible design, step S04: the dynamic threshold generation mechanism for peak and off-peak seasons specifically includes: Based on the peak and off-peak season fluctuation coefficients and the peak season adjustment factor, the peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated.
[0031] In this embodiment, it should be specifically noted that the calculation formula for the peak season dynamic threshold is as follows:
[0032] in, This is represented as the peak season dynamic threshold. This is expressed as the seasonal fluctuation coefficient. This is represented as a peak season adjustment factor. This is represented as the baseline threshold and determined using the quantile method; The formula for calculating the off-season dynamic threshold is as follows:
[0033] in, This is represented as the off-season dynamic threshold. This is represented as a seasonal adjustment factor. The formula for calculating the off-season dynamic threshold is as follows:
[0034] in, This is represented as the off-season dynamic threshold. It is represented as the seasonal adjustment factor.
[0035] Step S05: Financial Risk Early Warning: This step involves inputting the first feature set collected in real time into the trained improved LSTM-attention financial risk early warning model to obtain the real-time risk probability, which is then compared with the dynamic threshold for peak and off-peak seasons for risk processing.
[0036] In one possible design, step S05: financial risk warning specifically includes: S51: Extract the real-time first feature set representing the financial health status of the target tourism enterprise from the time series dataset, input it into the trained improved LSTM-attention financial risk early warning model, and obtain the real-time risk probability; S52: Combining real-time risk probability with dynamic thresholds for peak and off-peak seasons: If the real-time risk probability is less than the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold, a low-risk warning signal will be output to remind relevant management personnel of tourism enterprises to conduct routine monitoring. If the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold ≤ real-time risk probability < preset high-risk threshold, then a risk warning signal will be output and a risk cause analysis will be pushed to relevant management personnel of tourism enterprises. If the real-time risk probability is greater than or equal to the preset high-risk threshold, a high-risk warning signal will be output, and the emergency response mechanism will be triggered immediately. S53: Output early warning results, generate visual early warning reports, and push them to the enterprise financial terminal and management mobile terminal.
[0037] In this embodiment, it should be specifically explained that the present invention generates a unified time-series dataset by collecting the target tourism enterprise's structured financial data, business operation data, and external dynamic data, constructs peak and off-peak season features and supplier association features, and finally outputs a multi-source fusion feature set. Through full coverage of multi-source data and comprehensive data collection, it achieves full-dimensional coverage of tourism enterprise financial data. By unifying the time-series dataset, it ensures data relevance and timeliness, lays the foundation for accurate early warning, and constructs targeted tourism enterprise features, thereby improving the correlation between features and tourism financial risks. This provides high-quality data and feature support for subsequent accurate early warning by deep learning models. This invention trains and optimizes a model based on the time-series dataset corresponding to the final multi-source fusion feature set, obtaining a financial risk early warning model adapted to the characteristics of the target tourism enterprise. Based on the peak and off-peak season fluctuation coefficients and peak and off-peak season adjustment factors of the target tourism enterprise, it calculates the peak season dynamic threshold, off-peak season dynamic threshold, and shoulder season dynamic threshold, outputs early warning results, generates a visual early warning report, and pushes it to the enterprise's financial terminal and management's mobile terminal. Through dynamic weight adjustment and dynamic threshold generation, the early warning model achieves real-time adaptation to seasonal fluctuations, improving the credibility of the early warning results. Based on the daily time-series dataset, it has the ability to quickly process and predict real-time data, enhancing the enterprise's initiative in risk resistance and improving the efficiency of refined management of financial risks for tourism enterprises.
[0038] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A deep learning-based intelligent early warning method for financial risks in tourism enterprises, characterized in that, Includes the following steps: Step S01: Data collection for tourism enterprises: This step involves collecting structured financial data, business operation data, and external dynamic data from target tourism enterprises, and then normalizing and removing outliers. Step S02: Construction of tourism enterprise suitability features: This step is used to extract the first feature set representing the financial health status of the target tourism enterprise from the time series dataset and to construct the risk suitability feature set. Step S03: Deep learning model training: Based on the time series dataset corresponding to the final multi-source fusion feature set, the model is trained and optimized to obtain a financial risk early warning model that adapts to the characteristics of the target tourism enterprise. Step S04: Peak and Off-Peak Season Dynamic Threshold Generation Mechanism: Based on the peak and off-peak season fluctuation coefficients of the target tourism enterprise, obtain the peak and off-peak season dynamic thresholds of the target tourism enterprise; Step S05: Financial Risk Early Warning: This step involves inputting the first feature set collected in real time into the trained improved LSTM-attention financial risk early warning model to obtain the real-time risk probability, which is then compared with the dynamic thresholds for peak and off-peak seasons for risk processing.
2. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: Step S01: Data collection from tourism enterprises specifically involves: S21: Collect daily structured financial data of target tourism enterprises, including debt-to-equity ratio, current ratio, net operating cash flow, and gross profit margin; collect daily business operation data from scenic spot ticketing systems and secondary consumption systems, including total visitor flow, instantaneous visitor flow, ticket revenue, sales of cultural and creative products, and catering revenue; collect external dynamic data, including environmental data, public opinion data, and market data, by accessing third-party platforms through API interfaces. S22: Use box plots to remove outliers from financial structured data and use interpolation to fill in missing values in business operation data and external data; S23: Standardize the financial structured data, business operation data, and external dynamic data of the target tourism enterprise and map the data to the [0,1] interval; S24: Using "day" as the time granularity, align the financial structured data, business operation data, and external dynamic data along the time axis to generate a unified time series dataset. Each record in the time series dataset contains a timestamp and multi-dimensional feature values for the corresponding time point.
3. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: Step S02: Construction of tourism enterprise adaptability features specifically involves: S31: Peak and Off-Peak Season Feature Construction: A clustering algorithm is used to divide the peak and off-peak seasons. The clustering features are average monthly revenue and average monthly customer traffic. The number of clusters is set to 3 (peak season, shoulder season, off-peak season). Based on the average monthly customer traffic / average monthly revenue of the cluster centers, the month with the highest output is marked as the peak season, the month with the lowest output is marked as the off-peak season, and the rest are shoulder seasons. S32: Extract the average monthly revenue during peak season, the average monthly revenue during off-season, and the average monthly revenue throughout the year, and calculate the peak-off-season fluctuation coefficient of the target tourism enterprise; S33: Add seasonal labels and seasonal fluctuation coefficients to each record in the time series dataset.
4. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 3, characterized in that: The construction of the tourism enterprise adaptability features also includes: Supplier association feature construction: S41: Obtain the annual cooperation duration with suppliers, the total annual cooperation duration with the target tourism enterprise, the annual order volume with suppliers, and the total annual order volume of the target tourism enterprise from the time series dataset, and calculate the supplier association strength coefficient of the target tourism enterprise; S42: Add a supplier association strength coefficient to each record in the time series dataset, and finally output a multi-source fusion feature set.
5. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: The model training specifically involves: S51: Divide the time series dataset corresponding to the final multi-source fusion feature set into a training set, a validation set, and a test set; the training set is used for fitting and learning the model parameters, the validation set is used for debugging and optimizing the model hyperparameters, and the test set is used for evaluating and verifying the final warning performance of the model; the temporal continuity of the time series data is maintained during the partitioning process to avoid data leakage affecting the model training effect; S52: Constructing an improved deep learning model: including an input layer, a seasonal adaptive adjustment layer, an LSTM feature learning layer, an attention layer, and a fully connected output layer. The parameters of each layer are initialized as follows: Input layer initialization: The input dimension is the number of features n in the final multi-source fusion feature set, used to receive temporal feature sequences. ,in This is represented as a multi-dimensional feature vector at time i, where the vector dimension is the same as the input dimension. Initialization of the seasonal adaptive adjustment layer: Introducing seasonal adjustment factors , Based on the peak and off-peak season labels, the specific assignment rules are as follows: Peak Season , shoulder season Off-season The feature vectors received by the input layer are weighted and adjusted using this adjustment factor, as shown in the formula: ,in, This represents the dynamic weight at time i; LSTM feature learning layer initialization: A two-layer stacked LSTM network is set as the core of feature learning. The number of neurons in the first LSTM hidden layer is 64, and the number of neurons in the second LSTM hidden layer is 32. The initial values of the parameters of the forget gate, input gate and output gate of the LSTM network are set using the Xavier initialization method to ensure that the model training process is stable and converges after parameter initialization. Attention layer initialization: Construct an attention layer to adaptively extract key risk information from time-series features. By calculating the attention weights of the feature vectors at each time step, a weighted aggregation of financial risk features is achieved. Fully connected output layer initialization: Set up two fully connected layers. The first fully connected layer has 16 neurons, and the second fully connected layer has 1 neuron. The second fully connected layer uses the Sigmoid activation function to output the risk probability at time i.
6. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: The financial risk early warning model is specifically as follows: The time-series feature sequences of the training set are input into the initialized improved deep learning model. The loss function is minimized through backpropagation, driving the model parameters to update iteratively. The model is iteratively trained until the loss on the validation set converges, and the current optimal model parameters are saved to obtain the trained improved LSTM-attention financial risk warning model. The improved deep learning model also includes an optimizer to improve the model's convergence speed and generalization ability. The loss function is used to calculate the deviation between the risk probability output by the model and the true risk label; The validation set loss converges as follows: during training, the change in the validation set loss value is monitored in real time, and when the fluctuation range of the validation set loss value is less than 10 consecutive rounds... At that time, the model training is considered to have converged.
7. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: The specific mechanism for generating dynamic thresholds for peak and off-peak seasons in step S04 is as follows: Based on the peak and off-peak season fluctuation coefficients and the peak season adjustment factor, the peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated. Based on the peak and off-peak season fluctuation coefficients and the off-peak season adjustment factor, the off-peak season dynamic threshold of the target tourism enterprise is calculated.
8. The intelligent early warning method for financial risks of tourism enterprises based on deep learning according to claim 1, characterized in that: Step S05: Financial Risk Warning specifically includes: S81: Extract the real-time first feature set representing the financial health status of the target tourism enterprise from the time series dataset, input it into the trained improved LSTM-attention financial risk early warning model, and obtain the real-time risk probability; S82: Combining real-time risk probability with dynamic thresholds for peak and off-peak seasons: If the real-time risk probability is less than the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold, a low-risk warning signal will be output to remind relevant management personnel of tourism enterprises to conduct routine monitoring. If the peak season dynamic threshold / off-season dynamic threshold / shoulder season dynamic threshold ≤ real-time risk probability < preset high-risk threshold, then a risk warning signal will be output and a risk cause analysis will be pushed to relevant management personnel of tourism enterprises. If the real-time risk probability is greater than or equal to the preset high-risk threshold, a high-risk warning signal will be output, and the emergency response mechanism will be triggered immediately. S53: Output early warning results, generate visual early warning reports, and push them to the enterprise financial terminal and management mobile terminal.