Navigation equipment abnormity monitoring method based on inverted attention mechanism and diffusion model

Abstract

The invention relates to a navigation equipment abnormity monitoring method based on an inverted attention mechanism and a diffusion model. The method comprises the following steps: acquiring time sequence data related to navigation equipment from a navigation system; an inverted attention mechanism is adopted, and a Transform model is improved; training an improved Transform model based on a diffusion model, and in a diffusion stage, adopting an adaptive noise injection mechanism, dynamically adjusting a noise level, and selecting noise intensity to gradually add noise to multivariate time sequence data; in the denoising stage, using noisy navigation system time sequence sample data to train an improved Transform model; and using an improved Transform model to obtain predicted time sequence data according to the real-time navigation system data, comparing the predicted time sequence data with the healthy time sequence data, and detecting and identifying abnormity. According to the method, an inverted attention mechanism is adopted, and the defects of traditional Transform in multivariate time sequence coding are overcome; data distribution is learned through the diffusion-denoising process of the diffusion model, and the accuracy and prediction precision of abnormal detection of the navigation equipment are improved.

Description

Technical Field

[0001] The present invention belongs to the field of navigation control, and particularly relates to an abnormal monitoring method for navigation equipment based on an inverted attention mechanism and a diffusion model. Background Art

[0002] In the application scenarios of the Three Gorges navigation equipment, data collection and monitoring systems including network monitoring, CCTV systems, dispatching systems, etc. have been established. These systems collect data such as the operating status of navigation equipment and the status of the navigation environment. The existing data center performs basic data processing and analysis on these data, as well as an alarm system based on preset thresholds. Although these methods can provide certain monitoring capabilities, there are obvious limitations in terms of real-time performance, prediction accuracy, and in-depth data utilization. With the continuous growth of the national shipping demand, the navigation informatization system has become increasingly complex, and the data it needs to collect and process has become increasingly huge. The amount of time series data and data types required to be processed in the navigation scenario have reached an unprecedented level of complexity. Therefore, there is an urgent need for a more accurate time series-based abnormal detection method that can handle more complex data.

[0003] Traditional abnormal detection methods for navigation equipment mainly rely on statistical methods, often based on simple data processing, and simply judge abnormalities through fixed thresholds. When new abnormal patterns appear, this method is often difficult to respond in a timely manner. By simply judging whether it is abnormal through the warning threshold after simple data processing, this method can only linearly process data and cannot handle complex abnormal situations, resulting in missed potential risks. In addition, traditional data processing methods process single time series data alone and lack comprehensive analysis of multiple data, which may lead to incorrect abnormal alarms due to missing or incorrect isolated data sources, affecting subsequent maintenance and detection work.

[0004] In the work of abnormal detection of navigation equipment, multivariate time series data is a processing difficulty. In other fields, there are also some time series prediction technologies such as exponential smoothing models and ARIMA models. However, in the navigation information system, the data covers multiple aspects such as the above-mentioned dispatching system data, supervision system data, CCTV system data, and navigation environment data. The traditional time series prediction technologies mentioned above only have good prediction performance for single-variable time series data and cannot handle the multi-variable time series in the navigation information system.

[0005] In recent years, deep learning technologies have demonstrated powerful capabilities in processing time series data. Applying this ability to the abnormal detection task of navigation system equipment can significantly improve the accuracy of system abnormal detection. Common deep learning models for time series data are RNN and LSTM models. Although they can capture the dependencies in the time series, they often face bottlenecks in efficiency and performance when dealing with long sequences. Summary of the Invention

[0006] The object of the present invention is to address the above problems and provide a method for abnormal monitoring of navigation equipment based on an inverted attention mechanism and a diffusion model. The traditional Transformer model is improved using the inverted attention mechanism to solve the deficiencies of the traditional Transformer in multi-variate time series encoding. The diffusion model is integrated with the improved Transformer model, and the data distribution is learned through the diffusion-denoising process of the diffusion model, breaking through the dependence of traditional time series models on linear or stationarity assumptions, and improving the accuracy and prediction precision of abnormal detection of navigation equipment. The prediction stability under complex working conditions is improved through the adaptive noise mechanism of the diffusion model, avoiding false alarms caused by environmental noise.

[0007] To achieve the above object, the technical solution provided by the present invention is as follows: A method for abnormal monitoring of navigation equipment based on an inverted attention mechanism and a diffusion model, comprising the following steps: S1: Obtain time series data related to navigation equipment from the navigation system; S2: Preprocess the time series data of the navigation system obtained in step S1 to obtain multi-variate time series data. The preprocessing includes timestamp alignment, missing value filling, data normalization, and outlier filtering; S3: Use the inverted attention mechanism to improve the Transformer model to obtain an improved Transformer model; S4: Train the improved Transformer model based on the diffusion model; S401: In the diffusion stage, adopt an adaptive noise injection mechanism to dynamically adjust the noise level, and gradually add noise to the multi-variate time series data obtained in step S2 at a selected noise intensity to obtain noisy time series sample data of the navigation system; S402: In the denoising stage, use the noisy time series sample data of the navigation system to train the improved Transformer model; S5: After processing the time series data of the navigation system obtained in real time by the method of step S2, input it into the trained improved Transformer model to obtain predicted time series data. Adopt a detection method based on an error threshold to compare the predicted time series data with the healthy time series data and detect and identify abnormalities. If an abnormality is detected and identified, a warning signal is issued; S6: Perform adaptive adjustment and optimization on the model for abnormal monitoring of navigation equipment.

[0008] Further, in step S1, multi-source data is collected from the navigation system, and the multi-source data includes navigation equipment data, navigation equipment server logs, navigation monitoring data, navigation execution work orders, and navigation vessel information data.

[0009] Preferably, step S2 specifically includes the following sub-steps: S201: Unify the timestamps of different data sources to ensure the synchronization of time series data; S202: For missing data points, use interpolation, average value, or machine learning methods to fill them to ensure data continuity; S203: Scale the data of each dimension to a unified range to reduce the difference in different feature scales; S204: Identify and filter out outliers in the data to avoid negative impacts on model training and prediction; S205: Extract the volatility and trend characteristics of the navigation system time series data to provide a basis for subsequent dynamic adjustment of the noise level.

[0010] Preferably, step S401 specifically includes the following sub-steps: S401a: Generate a series of noise samples from the Gaussian distribution where represents the standard deviation of the Gaussian distribution, reflecting the noise intensity, dynamically adjust the intensity of the noise according to the following formula ; where k is the dynamic adjustment factor, represents the error of the navigation system time series feature data obtained by real-time monitoring; The calculation formula for the dynamic adjustment factor is: ; where is the i-th type of navigation system time series feature value, used to dynamically adjust the noise level, is the weight coefficient of the i-th type of navigation system time series feature data, used to balance the impacts of the volatility and trend characteristics of the navigation system time series data on the noise level; S401b: Gradually superimpose the generated noise samples onto the preprocessed navigation system time series data to generate noisy sample data with different noise levels, ; where represents the sample data before adding noise, represents the sample data after adding noise, represents the noise, calculated according to ; S401c: Generate a noisy navigation system data sample set for training an improved Transformer model.

[0011] Preferably, step S402 specifically includes the following sub-steps: S402a: Use the inverted self-attention mechanism of the improved Transformer model to predict the noise components at each step according to the long-range dependence relationship of the input data; encode the time series data into independent Tokens, and use inverted Tokens to invert the time dimension and spatial dimension of the time series data, and use the attention mechanism to model the correlation between different variables in the multivariate time series data; S402b: Starting from the highest noise level, gradually remove the noise through the improved Transformer model to generate time series data close to the original data distribution. S402c: Calculate the error between the denoised navigation system data and the original noise-free data, use the mean square error MSE as the loss function for optimization, and iteratively adjust the parameters of the improved Transformer model; S402d: Evaluate the performance of the improved Transformer model on an independent validation set to ensure its generalization ability and make necessary parameter adjustments.

[0012] Preferably, step S5 specifically includes the following sub-steps: S501: Use the trained improved Transformer model to denoise the real-time acquired navigation system time series data and predict the future time series trend; S502: Compare the predicted time series data obtained by the improved Transformer model with the actually observed healthy time series data, calculate the error, and use a detection method based on an error threshold to identify the time points with large errors and determine whether there is an abnormal state; S503: If an abnormal state is detected, send a warning signal to remind the operation and maintenance personnel to take corresponding measures in time to avoid equipment failures or other risks.

[0013] Preferably, step S6 specifically includes the following sub-steps: S601: Continuously update the parameters of the improved Transformer model according to the newly collected navigation system time series data to enhance the adaptability of the improved Transformer model to the new data distribution; S602: Based on the feedback information in the actual navigation system application scenario, automatically adjust the noise intensity of the diffusion model and the parameters of the attention mechanism in the improved Transformer model to cope with the data characteristics in different environments. S603: Ensure the robustness and prediction accuracy of the improved Transformer model in various different scenarios of the navigation system through periodic verification and fine-tuning.

[0014] As another object of the present invention, there is provided a navigation equipment anomaly monitoring system, including: Data processing module: Collect multi-source data from the navigation informatization business system, preprocess the time series data of the navigation system therein to obtain multivariate time series data; Improved Transformer module: Construct an improved Transformer model, and the improved Transformer model adopts an inverted attention mechanism; Model training module: Use a diffusion model to train the improved Transformer model; Anomaly detection module: Invoke the data preprocessing module and the improved Transformer module, obtain predicted time series data according to the real-time obtained time series of the navigation system; Adopt a detection method based on an error threshold, compare the predicted time series data with the healthy time series data and detect and identify anomalies, and if anomalies are detected and identified, an early warning signal is issued; Model optimization module: Used to optimize and adjust the trained improved Transformer model and diffusion model; Through continuous learning of the time series data of different navigation subsystems and continuously adjusting the model parameters according to different situations in the actual application of the navigation system, ensure the robustness and accuracy of the improved Transformer model in different scenarios and environments of the navigation system.

[0015] Compared with the prior art, the beneficial effects of the present invention include: 1) The present invention integrates a diffusion model with an improved Transformer model, introduces the diffusion model into the time series prediction task, and learns the data distribution through the diffusion-denoising process, breaking through the dependence of traditional time series models such as ARIMA and LSTM on linear or stationarity assumptions, and is especially suitable for modeling non-linear and high-noise scenarios of the navigation system, effectively improving the accuracy and prediction accuracy of navigation equipment anomaly detection, and improving the prediction stability under complex working conditions such as extreme weather and sudden load through the adaptive noise mechanism of the diffusion model, and can avoid false alarms caused by environmental noise by traditional methods.

[0016] 2) The improved Transformer model introduces an inverted attention mechanism on the basis of the traditional Transformer. Different from the traditional Transformer model that uses the attention mechanism to model the relationships between time steps and variables simultaneously, the inverted attention establishes the relationships between time steps through an MLP, and then further models the relationships between variables through the attention mechanism. It can solve the problems of the traditional model being sensitive to local noise or insufficient capture of long-range dependencies by inversely focusing on key features in time-series data such as low-frequency signals or long-term dependencies, and enhances the parsing ability for complex time-series patterns of navigation equipment.

[0017] 3) The present invention dynamically adjusts the noise intensity during the diffusion stage and optimizes the noise injection strategy based on data characteristics. Compared with the fixed noise scheme, it can significantly improve the adaptability of the improved Transformer model to data distribution shifts and avoid overfitting or underfitting.

[0018] 4) The present invention effectively integrates multi-source asynchronous sensor data in the navigation system by co-processing multi-modal time-series data and through timestamp alignment and adaptive missing value filling techniques, and solves the prediction deviation problem caused by data fragmentation in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the drawings and embodiments.

[0020] Figure 1 It is a schematic diagram of implementing abnormal detection of navigation equipment for multi-source data in the navigation system in the embodiment of the present invention.

[0021] Figure 2 It is a schematic diagram of the prediction process of the improved Transformer model in the embodiment of the present invention.

[0022] Figure 3 It is a comparison diagram between the improved Transformer model and the traditional Transformer model in the embodiment of the present invention.

[0023] Figure 4 It is a schematic diagram of the inverted attention mechanism of the improved Transformer model in the embodiment of the present invention.

[0024] Figure 5 It is a schematic diagram of the data flow of the navigation equipment abnormal system in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model in the embodiment includes: Step S1: Obtain time-series data related to navigation equipment from the navigation system; Step S101: Collect time series data from various subsystems in the general aviation informatization business system.

[0026] As Figure 1 shown, the multi-source data collected from the general aviation informatization business system includes real-time operation index data of equipment (temperature, pressure, speed, etc.), general aviation equipment server logs (equipment startup, stop, abnormal events, etc.), general aviation monitoring data (sensor data, CCTV system, environmental information, etc.), general aviation execution work orders (equipment ID, maintenance records, processing status, etc.), and general aviation ship data (ship number, declaration time, arrival at anchor time, etc.). The storage formats of some of the general aviation data are shown in Table 1.

[0027] Table 1 Data storage format

[0028] Step S102: Manage the objects to be analyzed in the management center to determine the analysis scope.

[0029] Step S103: Query the managed objects, generate learning object by associating with the learning component configuration table, and store it in the database.

[0030] Step S104: Prepare the learning component + learning component parameters, call the anomaly detection interface of the iFormerFusion model, and perform subsequent model processing. Step S105: Perform data format conversion. For some fields that need to be utilized but are not numerical values, perform format conversion. The types of fields such as the declaration unique identifier (ship declaration number), lockage method, and declaration method are varchar, and they are converted into numerical sequences through numerical mapping.

[0031] Step S2: Preprocess the general aviation system time series data to obtain multivariate time series data; Step S201: Timestamp alignment: Unify the timestamps of different data sources to ensure the synchronization of time series data; Step S202: Missing value filling: For missing data points, use interpolation, average value, or machine learning methods for filling to ensure the continuity of the data; Step S203: Data normalization: Scale the data of each dimension to a unified range such as 0 to 1 to reduce the impact of different feature scale differences on model training; The normalization formula is as follows: ; Step S204: Outlier filtering: Identify and filter out the outliers in the data to avoid negative impacts on model training.

[0032] In the embodiment, the outlier detection is performed through the following formula: ; Among them, Q1 and Q3 are the first quartile and the third quartile of the data respectively, and IQR is the interquartile range.

[0033] Step S205: Feature extraction: Extract useful device timing features such as data volatility, trends, etc., to provide a basis for subsequent dynamic adjustment of the noise level.

[0034] The data format after outlier filtering is shown in Table 2.

[0035] Table 2: Data format

[0036] Step S3: Adopt an inverted attention mechanism to improve the Transformer model to obtain an improved Transformer model, as Figure 3 and Figure 4 shown.

[0037] The inverted attention mechanism separately considers different variables in the time series data. Each variable is encoded into an independent token. The attention mechanism is used to model the correlation between different variables, and the time series correlation of the variables is modeled through a feed-forward network. This processing can solve the deficiencies of traditional Transformers in multivariate time series encoding. This inverted processing method can effectively capture the time dependence and the relationship between features, improving the accuracy and robustness of anomaly detection. The navigation informatization system has data in multiple aspects such as dispatching system data, supervision system data, CCTV system data, and navigation environment data. Conducting anomaly detection based on time series prediction in this system requires processing relatively complex multivariate time series data. Using this inverted attention mechanism can improve the deficiencies of the traditional Transformer model in multivariate time series encoding ability in the system.

[0038] Step S4: Train the improved Transformer model based on the diffusion model; Step S401 specifically includes the following sub-steps: S401a: Noise generation: Generate a series of noise samples from the Gaussian distribution , represents the standard deviation of the Gaussian distribution, reflecting the noise intensity; Dynamically adjust the intensity of the noise according to the following formula, ; Among them, k is the dynamic adjustment factor, represents the error of the navigation device timing feature data obtained by real-time monitoring; The calculation formula for the dynamic adjustment factor is: ; Among them, is the time-series characteristic value of the i-th navigation equipment, which is used to dynamically adjust the noise level, is the weight coefficient of the time-series characteristic data of the i-th navigation equipment, which is used to balance the influence of the volatility and trend of the time-series data of the navigation equipment on the noise level; S401b: Noise addition: Gradually superimpose the generated noise samples onto the preprocessed time-series data of the navigation equipment to generate noisy sample data with different noise levels, ; Among them represents the sample data before noise addition, represents the sample data after noise addition, represents the noise, which is calculated according to ; S401c: Generate a set of noisy time-series data samples of the navigation equipment for training the improved Transformer model; S402a: Noise prediction: Use the inverted self-attention mechanism of the improved Transformer model to predict the noise component at each step according to the long-range dependence relationship of the input data, as shown in Figure 2 ; First, obtain the sequence encoding of a single variable in the multi-variable time-series data of the navigation equipment, and obtain the time feature information of the noisy sample data through the linear MLP layer, ; ; In the formula, represents the time-series encoding output by the linear MLP layer, MLP( ) is the mapping function of the linear MLP layer, and are both parameters of the linear MLP layer; Encode the time-series data into independent tokens, and use the attention mechanism to model the correlation between different variables; In the traditional Transformer model, all feature data are normalized at fixed timestamps, which will introduce interactive noise, meaning that the model is learning useless relationships; in addition, it may cause the signal to be too smooth.

[0039] Therefore, in the improved Transformer model, the LayerNorm layer is used to replace the normalization process of the traditional Transformer model, ; where LayerNorm( ) represents the mapping function of the LayerNorm layer; Mean( ) is the mean function; Var( ) is the variance function; represents the time series encoding output by the LayerNorm layer; After that, it is input into the attention layer: ; where represents the output of the attention layer; Transformer( ) represents the mapping function of the attention layer; Q, K, and V respectively represent the query, key, and value of the attention layer; Finally, the correlation between different time steps in the time series data is obtained through a linear layer: ; where FFN( ) represents the mapping function of the linear layer; S402b: Gradual denoising: Starting from the highest noise level, gradually remove the noise through the improved Transformer model to generate time series data close to the original data distribution, ; S402c: Loss function optimization: Calculate the error between the denoised navigation equipment data and the original noise-free data, and use the mean squared error MSE as the loss function for optimization, and iteratively adjust the parameters of the improved Transformer model; ; where represents the loss function; represents the optimization parameter; N represents the number of samples; represents the true value of the sample, s represents the sample number, and t represents the time step; represents the predicted value; represents the noise term; S402d: Model verification: Evaluate the performance of the improved Transformer model on an independent validation set to ensure its generalization ability and make necessary parameter adjustments.

[0040] Step S5: Anomaly detection and warning; Step S501: Data restoration and prediction: Use the trained improved Transformer model to denoise the time series data of the navigation system obtained in real time and predict the future time series trend; This process restores the data step by step through reverse denoising,; ; Step S502: Anomaly Detection: Compare the predicted data obtained from the improved Transformer model with the actually observed healthy time series data, calculate the error, and use a detection method based on an error threshold to identify time points with large errors and determine whether there is an abnormal state; The anomaly detection formula is: ; where represents the error threshold, represents the abnormal state; The error threshold Threshold is dynamically adjusted according to the average prediction error calculated by the navigation system, ; In the formula represents the threshold reference; represents the error reference; represents the number of time steps.

[0041] Step S503: Issuing a Warning Signal: If an abnormal state is detected, a warning signal is issued to remind the operation and maintenance personnel to take corresponding measures in a timely manner to avoid equipment failures or other risks.

[0042] Step S6: Adaptive adjustment and optimization of the model for anomaly monitoring of navigation equipment; S601: Continuous Learning: Continuously update the parameters of the improved Transformer model according to the newly collected multi-source time series data of the navigation system to enhance the adaptability of the improved Transformer model to the new data distribution; S602: Parameter Adjustment: Automatically adjust the noise intensity of the diffusion model and the parameters of the attention mechanism in the improved Transformer model based on the feedback information in the actual navigation system application scenario to cope with the data characteristics in different environments; The parameter adjustment is achieved through the following optimization objective function: ; where, represents the regularization term, is the regularization strength coefficient; represents the total loss function; represents the optimization parameter; S603: Model Validation and Fine-Tuning: Ensure the robustness and prediction accuracy of the improved Transformer model in various different scenarios of the navigation system through periodic validation and fine-tuning.

[0043] As Figure 5 shown, in another embodiment of the present invention, a navigation equipment anomaly monitoring system based on the above method is provided, including the following modules: Data processing module: Collect multi-source data from the general aviation informatization business system, preprocess the general aviation system time series data therein, and obtain multivariate time series data; Improved Transformer module: Construct an improved Transformer model, and the improved Transformer model adopts an inverted attention mechanism; Model training module: Use a diffusion model to train the improved Transformer model; Anomaly detection module: Call the data preprocessing module and the improved Transformer module, obtain predicted time series data according to the real-time obtained general aviation system time series; adopt a detection method based on an error threshold, compare the predicted time series data with the healthy time series data and detect and identify anomalies, and if anomalies are detected and identified, issue a warning signal; Model optimization module: Used to optimize and adjust the trained improved Transformer model and diffusion model; continuously learn from the time series data of different general aviation subsystems, and continuously adjust the model parameters according to different situations in the actual application of the general aviation system to ensure the robustness and accuracy of the improved Transformer model in different scenarios and environments of the general aviation system.

Claims

1. An abnormal monitoring method for navigation equipment based on an inverted attention mechanism and a diffusion model, characterized in that It includes the following steps: S1: Obtain time series data related to navigation equipment from the navigation system; S2: Preprocess the time series data of the navigation system obtained in step S1 to obtain multivariate time series data. The preprocessing includes timestamp alignment, missing value filling, data normalization, and outlier filtering; S3: Improve the Transformer model using an inverted attention mechanism to obtain an improved Transformer model; S4: Train the improved Transformer model based on the diffusion model; S401: In the diffusion stage, adopt an adaptive noise injection mechanism to dynamically adjust the noise level, and gradually add noise to the multivariate time series data obtained in step S2 with a selected noise intensity to obtain noisy time series sample data of the navigation system; S402: In the denoising stage, use the noisy time series sample data of the navigation system to train the improved Transformer model; S5: After processing the time series data of the navigation system obtained in real time by the method of step S2, input it into the trained improved Transformer model to obtain predicted time series data; adopt a detection method based on an error threshold to compare the predicted time series data with the healthy time series data and detect and identify anomalies. If an anomaly is detected and identified, a warning signal is issued.

2. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 1, characterized in that In step S1, multi-source data is collected from the navigation system. The multi-source data includes navigation equipment data, navigation equipment server logs, navigation monitoring data, navigation execution work orders, and navigation vessel information data.

3. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S201: Unify the timestamps of different data sources to ensure the synchronization of time series data; S202: For missing data points, use interpolation, average value, or machine learning methods to fill them to ensure data continuity; S203: Scale the data of each dimension to a unified range to reduce the difference in different feature scales; S204: Identify and filter out outliers in the data; S205: Extract the volatility and trend characteristics of the time series data of the navigation system to provide a basis for dynamically adjusting the noise level later.

4. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 3, wherein, Step S401 specifically includes the following sub-steps: S401a: Generate a series of noise samples from a Gaussian distribution and represent the standard deviation of the Gaussian distribution, reflecting the noise intensity Dynamically adjust the intensity of the noise according to the following formula, ； where k is a dynamic adjustment factor, denotes the error of the time-series characteristic data of the navigation system obtained by real-time monitoring; The calculation formula of the dynamic adjustment factor is: ； Among them, is the time-series characteristic value of the i-th navigation system, which is used to dynamically adjust the noise level, is the weight coefficient of the time-series characteristic data of the i-th navigation system, which is used to balance the influence of the volatility and trend characteristics of the time-series data of the navigation system on the noise level; S401b: Gradually superimpose the generated noise samples on the preprocessed time series data of the navigation system to generate noisy sample data with different noise levels, ； Among them represents the sample data before adding noise, represents the sample data with added noise, represents noise; S401c: Generate a noisy data sample set of the navigation system for training the improved Transformer model.

5. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 4, wherein, The said step S402 specifically includes the following sub-steps: S402a: Use the inverted self-attention mechanism of the improved Transformer model to predict the noise component of each step according to the long-range dependence relationship of the input data; encode the time series data into independent Tokens, and adopt inverted Tokens to invert the time dimension and space dimension of the time series data, and use the attention mechanism to model the correlation between different variables in the multivariate time series data; S402b: Starting from the highest noise level, gradually remove the noise through the improved Transformer model to generate time series data close to the original data distribution. ； S402c: Calculate the error between the denoised navigation system data and the original noise-free data, and use the mean squared error (MSE) as the loss function for optimization. Iteratively adjust the parameters of the improved Transformer model. ； In the formula represents the loss function; represents the optimization parameter; N represents the number of samples; represents the true value of the sample, s represents the sample serial number, and t represents the time step; represents the predicted value; represents the noise term; S402d: Evaluate the performance of the improved Transformer model on an independent validation set to ensure its generalization ability and make necessary parameter adjustments.

6. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 5, characterized in that, The specific process of step S402a is as follows: First, obtain the sequence encoding of a single variable in the multivariate time series data of the navigation system. Through the linear MLP layer, obtain the time feature information of the noisy sample data. ； ； In the formula, represents the time series encoding output by the linear MLP layer, and MLP( ) is the mapping function of the linear MLP layer, and are both parameters of the linear MLP layer; Normalize the time series data using the LayerNorm layer. The calculation formula of the LayerNorm layer is: ； where LayerNorm( ) represents the mapping function of the LayerNorm layer; Mean( ) is the mean function; Var( ) is the variance function; represents the time series encoding of the output of the LayerNorm layer; After that, is input into the attention layer: ； In the formula represents the output of the attention layer; Transformer( ) represents the mapping function of the attention layer; Q, K, and V represent the query, key, and value of the attention layer respectively. Finally, obtain the correlation between different time steps within the time series data through a linear layer: ; where denotes the output of the linear layer; FFN( ) denotes the mapping function of the linear layer.

7. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 6, wherein Step S5 specifically includes the following sub-steps: S501: Use the trained improved Transformer model to denoise the real-time acquired time series data of the navigation system and predict the future time series trend. This step gradually restores the data through reverse denoising. S502: Compare the predicted time series data obtained by the improved Transformer model with the actually observed healthy time series data, calculate the error, and use a detection method based on the error threshold to identify the time points with large errors and determine whether there is an abnormal state. The anomaly detection formula is: ； Among them represents the error threshold represents the abnormal state S503: If an abnormal state is detected, send a warning signal to remind the operation and maintenance personnel to take corresponding measures in time to avoid equipment failures or other risks.

8. The abnormal monitoring method for navigation equipment based on the inverted attention mechanism and diffusion model according to claim 7, wherein, In step S502, the error threshold Threshold is dynamically adjusted according to the average prediction error calculated by the navigation system. ； where represents the threshold reference; represents the error reference; represents the number of time steps, and MSE( ) is the mean square error function.

9. The method for abnormal monitoring of navigation equipment based on the inverted attention mechanism and diffusion model according to claim 8, wherein, The navigation equipment anomaly monitoring method further includes step S6: adaptively adjust and optimize the model used for navigation equipment anomaly monitoring. S601: Continuously update the parameters of the improved Transformer model according to the newly collected time series data of the navigation system to enhance the adaptability of the improved Transformer model to the new data distribution. S602: Based on the feedback information in the actual navigation system application scenario, automatically adjust the noise intensity of the diffusion model and the parameters of the attention mechanism in the improved Transformer model to cope with the data characteristics in different environments. The parameter adjustment is achieved through the following optimization objective function: ； Among them, represents the regularization term, is the regularization strength coefficient; represents the total loss function; represents the optimization parameter; S603: Through periodic verification and fine-tuning, ensure the robustness and prediction accuracy of the improved Transformer model in various different scenarios of the navigation system.

10. The system for the abnormal monitoring method of the navigation equipment according to claims 1-9, characterized in that, It includes the following modules: Data processing module: Collect multi-source data from the navigation informatization business system, preprocess the time series data of the navigation system therein to obtain multivariate time series data. Improved Transformer Module: Construct an improved Transformer model that employs an inverted attention mechanism; Model Training Module: Use a diffusion model to train the improved Transformer model; Anomaly Detection Module: Invoke the data preprocessing module and the improved Transformer module to obtain predicted time series data based on the real-time acquired time series of the navigation system; Use a detection method based on an error threshold to compare the predicted time series data with the healthy time series data and detect and identify anomalies. If an anomaly is detected and identified, a warning signal is issued; Model Optimization Module: Used to optimize and adjust the trained improved Transformer model and diffusion model; Continuously learn from the time series data of different navigation subsystems and continuously adjust the model parameters according to different situations in the actual application of the navigation system to ensure the robustness and accuracy of the improved Transformer model in different scenarios and environments of the navigation system.

Images