Radiological equipment testing data and report management system

Abstract

The present invention relates to the field of radiation equipment detection technology, and specifically provides a radiation equipment detection data and report management system, which includes a data collection and processing module, a self-attention encoding module, a data completion module and a deviation detection module, wherein the data collection and processing module generates a time embedding vector, the self-attention encoding module uses a multi-head self-attention layer to capture the long-term and short-term dependencies of data, thereby improving the richness of data representation; the data completion module innovatively uses the attention mechanism to identify and complete missing points, thereby significantly improving data integrity; the deviation detection module trains a time series prediction model based on complete time-coded data, and realizes accurate anomaly detection by calculating the difference between the prediction result and the average value, and generates a comprehensive report to provide comprehensive support for equipment management.

Description

Technical Field

[0001] The present invention relates to the field of radiation equipment detection technology, in particular to a radiation equipment detection data and report management system. Background Art

[0002] Traditional radiological equipment inspection data and report management systems have many technical limitations when processing time series data.

[0003] First, these systems often use simple timestamp processing methods, which cannot effectively capture the periodic characteristics and long-term and short-term dependencies of the data, resulting in insufficient use of time series information. Second, in the data encoding process, traditional methods find it difficult to consider both local and global contexts at the same time, resulting in insufficient data representation and affecting the accuracy of subsequent analysis. When dealing with missing data, commonly used interpolation or average filling methods fail to consider the time series characteristics and feature similarity of the data, resulting in inaccurate completion results.

[0004] In addition, traditional systems usually use fixed thresholds or simple statistical methods for anomaly detection and prediction, which are difficult to adapt to complex data patterns and dynamically changing environments. In the report generation stage, traditional systems often only provide basic statistical information and lack in-depth analysis of data trends and potential risks.

[0005] These technical problems combined have resulted in obvious deficiencies in traditional systems in terms of data integrity, analysis accuracy, predictive capabilities and decision support, making it difficult to meet the needs of efficient and precise management of modern radiation equipment. Summary of the invention

[0006] The purpose of the present invention is to provide a radiation equipment detection data and report management system to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a radiation equipment detection data and report management system, including a data collection and processing module, a self-attention encoding module, a data completion module and a deviation detection module, wherein:

[0008] The data collection and processing module collects radiation equipment detection data, and performs time processing on it to generate a time embedding vector containing time steps;

[0009] The self-attention encoding module inputs the time embedding vector into the multi-head self-attention layer, calculates the attention score of each time step and all other time steps, and generates time encoding data based on the attention score;

[0010] The data completion module uses a time step missing point detection method to identify missing points in the time coding data, calculates the attention score between each missing point and the known data point, and calculates the missing value of the missing point based on the attention score using the time coding data of the known data point to generate complete time coding data;

[0011] The deviation detection module trains a time series prediction model based on the complete time coding data; uses the trained time series prediction model to predict the next time step of the complete time coding data, and calculates the difference between the prediction result and the average result of the complete time coding data, and uses the difference result as the deviation result; and outputs a comprehensive report based on the time step corresponding to the deviation result.

[0012] As a further improvement of the technical solution, the data collection and processing module collects the detection data of the radiation equipment and performs time processing on it, specifically including:

[0013] Collect radiation equipment detection data from various sensors of the radiation equipment and set the data collection frequency;

[0014] By converting the timestamp into a time standard format, using sine and cosine functions to periodically encode the timestamp, generating periodic coding features, calculating the relative position of each data point relative to the start time, and assigning a unique position code to each time step according to the data acquisition frequency; where relative position = (current timestamp - start timestamp) / total time span;

[0015] The temporally processed data is organized into a time embedding vector containing time steps. Each row represents a time step, represented by position encoding, and each column represents a time embedding feature. The time embedding features include time standard format, periodic encoding features, relative position and measurement value.

[0016] As a further improvement of the technical solution, the self-attention encoding module inputs the time embedding vector into the multi-head self-attention layer, calculates the attention score of each time step and all other time steps, and generates the time encoding data based on the attention score. The process specifically includes:

[0017] Set the number of attention heads and initialize the weight matrix of query, key and value for each attention head;

[0018] Apply a linear transformation to the time embedding vector at each time step to generate query, key, and value vectors as follows:

[0019] Query = time embedding vector * W_Q;

[0020] Key = time embedding vector * W_K;

[0021] Value = time embedding vector * W_V; where W_Q, W_K, W_V are weight matrices, which are continuously updated through the back propagation algorithm;

[0022] For each attention head, calculate the dot product between the query and the key to get the raw attention score, divide the raw attention score by the square root of the key vector dimension, scale it, and apply the softmax function to normalize the raw attention score, where the softmax function represents the normalized exponential function; use the normalized attention score to weighted sum the value vector to get the context vector for each time step;

[0023] Concatenate the context vectors of each time step of all attention heads together, apply a linear transformation, and project the concatenated vectors into the original dimensional space to generate the output of the multi-head self-attention;

[0024] The output of the multi-head self-attention is residually connected to the time embedding vector and layer normalized. The result of the layer normalization is passed through a feedforward neural network, and residual connection and layer normalization are applied again to generate time encoding data, where the time encoding data maintains the same time order and structure as the time embedding vector.

[0025] As a further improvement of the technical solution, the data completion module includes a missing point identification unit, and the process of the time step missing point detection method in the missing point identification unit specifically includes:

[0026] Determine the expected time interval between adjacent time steps in the time-encoded data based on the sampling frequency of the data;

[0027] According to the start and end of the timestamps in the time-coded data, each pair of adjacent time steps is checked in turn, and for each pair of adjacent time steps, the actual time interval between them is calculated, and the actual interval is compared with the expected interval;

[0028] If the actual interval is greater than the expected interval, it is considered that there is a missing point between the two time steps; the position of each identified missing point is recorded using the position code in the time step until the entire time code data is checked to generate the missing point of the time code data.

[0029] As a further improvement of the technical solution, the data completion module includes a missing point completion unit, and the process of generating complete time coding data by the missing point completion unit specifically includes:

[0030] Initialize the attention weight matrix W_A, use the time embedding features of the missing points as the query vector Query_missing, use the time encoding data of all known data points as the key vector Key_known and the value vector Value_known, where the known data points represent the data of the non-missing points; calculate the dot product of the query vector Query_missing and the key vector Key_known;

[0031] The dot product is scaled and the softmax function is applied to normalize the score to obtain the missing point attention score. The missing point attention score is used to perform weighted summation on the time coding data of the known data points to obtain the missing value of the missing point. The calculated missing value is filled into the corresponding position of the time coding data according to the position of the missing point to obtain the complete time coding data.

[0032] As a further improvement of the technical solution, the deviation detection module trains a time series prediction model based on the complete time coding data; uses the trained time series prediction model to predict the next time step of the complete time coding data, and calculates the difference between the prediction result and the average result of the complete time coding data, and uses the difference result as the deviation result, which specifically includes:

[0033] Use the complete time-coded data as the training set for the time series prediction model; create training samples using the sliding window technique by dividing the complete time-coded data into an input sequence and a target sequence, where the target sequence is the data of the next time step; optimize the loss function in the time series prediction model until the model converges or reaches a preset number of training rounds;

[0034] Use the trained time series prediction model to predict each time window in the complete time-coded data. For each time step t, use the data from t-10 to t-1 to predict the value at time t. Save all prediction results to form a prediction sequence with the same length as the complete time-coded data.

[0035] The average value of each feature in the complete time-coded data is calculated as the average result of the complete time-coded data; for each time step, the difference between the predicted result and the average result of the complete time-coded data is calculated, the difference of each time step is sorted into a sequence, a deviation threshold is set, the difference exceeding the deviation threshold is marked as a significant deviation, and the difference is used as the deviation result.

[0036] As a further improvement of the technical solution, the deviation detection module selects the corresponding time embedding features and related tags according to the time step corresponding to the deviation result, and outputs them as a comprehensive report. The comprehensive report is as follows:

[0037] Time information: including timestamp and location code to identify the specific time of each deviation;

[0038] Deviation result: the difference calculated at each time step;

[0039] Original data: time embedding vector corresponding to the time step;

[0040] Prediction value: the prediction result of the time series prediction model for each time step;

[0041] Mean: average value used for comparison;

[0042] Significant deviation marking: Marks which time steps have deviations exceeding the preset deviation threshold.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The radiation equipment detection data and report management system fully utilizes the temporal characteristics and internal associations of time series data by inputting the time embedding vector into the multi-head self-attention layer; the multi-head self-attention mechanism can simultaneously calculate the attention scores of each time step and all other time steps from multiple perspectives, capturing the long-term and short-term dependencies and complex patterns in the data; this time-coded data generated based on the attention score not only retains the temporal information of the original data, but also incorporates the global context, making the data representation richer and more meaningful; in addition, the system innovatively applies the attention mechanism to the missing point completion process; by calculating the attention scores of the missing points and the known data points, the system can intelligently identify the known data most relevant to the missing points, and perform weighted completion based on these correlations. This method not only considers the temporal proximity, but also the similarity of data features, so that it can more accurately estimate the missing values ​​and generate high-quality complete time-coded data; this process significantly improves the integrity and reliability of the data, laying a solid foundation for subsequent analysis.

[0045] 2. The radiation equipment detection data and report management system uses the generated complete time-coded data to train the time series prediction model, giving full play to the rich time series information and global context contained in the time-coded data. This model trained based on a complete and information-rich data set can more accurately capture the inherent laws and trends of the data, thereby improving the accuracy and stability of the prediction; in the prediction process, the system uses historical data to predict the value of the next time step through the sliding window technology. This method takes into account both short-term fluctuations and long-term trends; by calculating the difference between the predicted results and the average results of the data, the system can effectively identify abnormal fluctuations and deviations; this method based on time series prediction and statistical comparison can not only find obvious anomalies, but also detect subtle trend changes; finally, the system generates a comprehensive report based on the deviation results, including time information, deviation values, original data, predicted values ​​and other multi-dimensional information, providing a comprehensive and intuitive basis for equipment monitoring and decision-making. This time-based full-process analysis method forms a complete closed loop from data collection, processing, completion to prediction and report generation, greatly improving the analysis efficiency and accuracy of radiation equipment detection data, and providing strong support for the safe operation and preventive maintenance of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall module of the present invention;

[0047] Figure 2 Schematic diagram of a data completion module unit of the present invention.

[0048] In the figure: 100, data collection and processing module; 200, self-attention encoding module; 300, data completion module; 301, missing point recognition unit; 302, missing point completion unit; 400, deviation detection module. DETAILED DESCRIPTION

[0049] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0050] Unless otherwise specified, the terms (including technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is understood that the terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

[0051] The following are some definitions of terms:

[0052] Time embedding vector: Radiological equipment detection data after time processing, including time steps and time embedding features. Each row represents a time step, represented by position encoding; each column represents a time embedding feature, including time standard format, periodic encoding features, relative position and measurement value;

[0053] Multi-head self-attention layer: A neural network structure that simultaneously learns the correlation of input sequences from different representation subspaces through multiple parallel attention calculations (heads), thereby capturing richer feature information and long-range dependencies;

[0054] Attention score: In the self-attention mechanism, it is a numerical value used to measure the degree of association between different time steps. It is obtained by calculating the dot product between the query vector and the key vector, and then scaling and softmax normalization.

[0055] Time-encoded data: data processed by the multi-head self-attention layer maintains the same temporal order and structure as the temporal embedding vector, but the feature representation of each time step becomes richer and more abstract;

[0056] Known data points: In time-coded data, non-missing data points are used to calculate missing values ​​for missing points;

[0057] Complete time-coded data: time-coded data processed by the data completion module, in which the missing points have been filled to form a complete time series data set;

[0058] Deviation result: The result obtained by predicting the complete time-coded data through the time series prediction model and calculating the difference between the predicted result and the average result of the complete time-coded data. It is used to identify and mark the time steps with significant deviations.

[0059] Next, see Figure 1-Figure 2 The present invention provides a technical solution: a radiation equipment detection data and report management system, including a data collection and processing module 100, a self-attention encoding module 200, a data completion module 300 and a deviation detection module 400.

[0060] The data collection and processing module 100 collects the radiation equipment detection data, and performs time processing on it to generate a time embedding vector containing time steps, specifically including:

[0061] Collect radiation equipment detection data from various sensors of the radiation equipment and set the data collection frequency (such as every second, every minute or every hour). The sensors include dose rate sensors, temperature sensors and pressure sensors, etc. The radiation equipment detection data includes timestamps, sensor IDs and measurement values, etc.;

[0062] The radiation equipment detection data is processed in time by converting the timestamp into a time standard format (such as Unix timestamp), using sine and cosine functions to periodically encode the timestamp (such as time of day, date of week), generating periodic coding features, calculating the relative position of each data point relative to the start time, and assigning a unique position code to each time step according to the data acquisition frequency; where relative position = (current timestamp - start timestamp) / total time span;

[0063] The temporally processed data is organized into a time embedding vector containing time steps. Each row represents a time step, represented by position encoding, and each column represents a time embedding feature. The time embedding features include time standard format, periodic encoding features, relative position and measurement value.

[0064] The self-attention encoding module 200 inputs the time embedding vector into the multi-head self-attention layer, where the multi-head self-attention layer is a neural network structure that simultaneously learns the correlation of the input sequence from different representation subspaces through multiple parallel attention calculations (heads), thereby capturing richer feature information and long-distance dependencies, calculating the attention score of each time step with all other time steps, and generating time encoding data based on the attention score, specifically including:

[0065] Set the number of attention heads (e.g. 8 heads), and initialize the weight matrices of query, key, and value for each attention head;

[0066] Apply a linear transformation to the time embedding vector at each time step to generate query, key, and value vectors as follows:

[0067] Query = time embedding vector * W_Q;

[0068] Key = time embedding vector * W_K;

[0069] Value = time embedding vector * W_V; where W_Q, W_K, W_V are weight matrices, which are continuously updated through the back propagation algorithm;

[0070] For each attention head, calculate the dot product between the query and the key to get the raw attention score, divide the raw attention score by the square root of the key vector dimension, scale it, and apply the softmax function to normalize the raw attention score; use the normalized attention score to weighted sum the value vector to get the context vector for each time step;

[0071] Concatenate the context vectors of each time step of all attention heads together, apply a linear transformation, and project the concatenated vectors into the original dimensional space to generate the output of the multi-head self-attention;

[0072] The output of the multi-head self-attention is residually connected to the time embedding vector and layer normalized. The result of the layer normalization is passed through a feedforward neural network, and residual connection and layer normalization are applied again to generate time-encoded data, where the time-encoded data maintains the same time order and structure as the time embedding vector, but the feature representation of each time step becomes richer and more abstract.

[0073] The missing point identification unit 301 in the data completion module 300 uses a time step missing point detection method to identify missing points in the time coded data, specifically including:

[0074] Determine the expected time interval between adjacent time steps in the time-encoded data based on the sampling frequency of the data, e.g., if the data is sampled once a minute, the expected interval is 60 seconds;

[0075] According to the start and end of the timestamps in the time-coded data, each pair of adjacent time steps is checked in turn, and for each pair of adjacent time steps, the actual time interval between them is calculated, and the actual interval is compared with the expected interval;

[0076] If the actual interval is greater than the expected interval, it is considered that there is a missing point between the two time steps; the position of each identified missing point is recorded using the position code in the time step until the entire time code data is checked to generate the missing point of the time code data.

[0077] The missing point completion unit 302 in the data completion module 300 calculates the attention score of each missing point in the time code data with the known data point, and based on the attention score, calculates the missing value of the missing point using the time code data of the known data point to generate complete time code data, specifically including:

[0078] Initialize the attention weight matrix W_A (similar to W_Q, W_K, W_V in the self-attention encoding module 200), take the temporal embedding features of the missing points (excluding the measured values) as the query vector Query_missing, take the temporal encoding data of all known data points as the key vector Key_known and the value vector Value_known, where the known data points represent the data of the non-missing points; calculate the dot product of the query vector Query_missing and the key vector Key_known;

[0079] The dot product is scaled and the softmax function is applied to normalize the score to obtain the missing point attention score. The missing point attention score is used to perform weighted summation on the time coding data of the known data points to obtain the missing value of the missing point. The calculated missing value is filled into the corresponding position of the time coding data according to the position of the missing point to obtain the complete time coding data.

[0080] The deviation detection module 400 trains a time series prediction model based on the complete time coding data; uses the trained time series prediction model to predict the next time step of the complete time coding data, and calculates the difference between the prediction result and the average result of the complete time coding data, and uses the difference result as the deviation result, which specifically includes:

[0081] Use the complete time-coded data as the training set for the time series prediction model; create training samples by dividing the complete time-coded data into input sequences and target sequences, where the target sequence is the data of the next time step, using the sliding window technique, for example, using the past 10 time steps to predict the next time step; optimize the loss function (such as mean square error) in the time series prediction model until the model converges or reaches the preset number of training rounds;

[0082] Use the trained time series prediction model to predict each time window in the complete time-coded data. For each time step t, use the data from t-10 to t-1 to predict the value at time t. Save all prediction results to form a prediction sequence with the same length as the complete time-coded data.

[0083] The average value of each feature in the complete time-coded data is calculated as the average result of the complete time-coded data; for each time step, the difference between the predicted result and the average result of the complete time-coded data is calculated, the difference of each time step is sorted into a sequence, a deviation threshold is set, the difference exceeding the deviation threshold is marked as a significant deviation, and the difference is used as the deviation result.

[0084] The deviation detection module 400 selects the corresponding time embedding features and related tags according to the time step corresponding to the deviation result, and outputs them as a comprehensive report. The comprehensive report is as follows:

[0085] Time information: including timestamp and location code to identify the specific time of each deviation;

[0086] Deviation result: the difference calculated at each time step;

[0087] Original data: time embedding vector corresponding to the time step;

[0088] Prediction value: the prediction result of the time series prediction model for each time step;

[0089] Mean: The overall average used for comparison;

[0090] Significant deviation marking: Marks which time steps have deviations exceeding the preset deviation threshold.

[0091] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Radiation equipment detection data and report management system, characterized in that: The method comprises a data collection and processing module (100), a self-attention encoding module (200), a data completion module (300) and a deviation detection module (400), wherein: The data collection and processing module (100) collects radiation equipment detection data, performs time processing on the data, and generates a time embedding vector containing time steps; The self-attention encoding module (200) inputs the time embedding vector into the multi-head self-attention layer, calculates the attention score of each time step and all other time steps, and generates time encoding data based on the attention score; The data completion module (300) uses a time step missing point detection method to identify missing points in the time coding data, calculates the attention score of each missing point with the known data point, and calculates the missing value of the missing point based on the attention score of the missing point using the time coding data of the known data point to generate complete time coding data; The deviation detection module (400) trains a time series prediction model based on the complete time coding data; uses the trained time series prediction model to predict the next time step of the complete time coding data, and calculates the difference between the prediction result and the average result of the complete time coding data, and uses the difference result as the deviation result; and outputs a comprehensive report based on the time step corresponding to the deviation result.

2. The radiation equipment detection data and report management system according to claim 1, characterized in that: The data collection and processing module (100) collects radiation equipment detection data and performs time processing on it, specifically including: Collect radiation equipment detection data from various sensors of the radiation equipment and set the data collection frequency; By converting the timestamp into a time standard format, using sine and cosine functions to periodically encode the timestamp, generating periodic coding features, calculating the relative position of each data point relative to the start time, and assigning a unique position code to each time step according to the data acquisition frequency; where relative position = (current timestamp - start timestamp) / total time span; The temporally processed data is organized into a time embedding vector containing time steps. Each row represents a time step, represented by position encoding, and each column represents a time embedding feature. The time embedding features include time standard format, periodic encoding features, relative position and measurement value.

3. The radiation equipment detection data and report management system according to claim 1, characterized in that: The self-attention encoding module (200) inputs the time embedding vector into the multi-head self-attention layer, calculates the attention score of each time step and all other time steps, and the process of generating time encoding data based on the attention score specifically includes: Set the number of attention heads and initialize the weight matrix of query, key and value for each attention head; Apply a linear transformation to the time embedding vector at each time step to generate query, key, and value vectors as follows: Query = time embedding vector * W_Q; Key = time embedding vector * W_K; Value = time embedding vector * W_V; where W_Q, W_K, W_V are weight matrices, which are continuously updated through the back propagation algorithm; For each attention head, calculate the dot product between the query and the key to get the raw attention score, divide the raw attention score by the square root of the key vector dimension, scale it, and apply the softmax function to normalize the raw attention score, where the softmax function represents the normalized exponential function; use the normalized attention score to weighted sum the value vector to get the context vector for each time step; Concatenate the context vectors of each time step of all attention heads together, apply a linear transformation, and project the concatenated vectors into the original dimensional space to generate the output of the multi-head self-attention; The output of the multi-head self-attention is residually connected to the time embedding vector and layer normalized. The result of the layer normalization is passed through a feedforward neural network, and residual connection and layer normalization are applied again to generate time encoding data, where the time encoding data maintains the same time order and structure as the time embedding vector.

4. The radiation equipment detection data and report management system according to claim 3, characterized in that: The data completion module (300) comprises a missing point identification unit (301), and the process of the time step missing point detection method in the missing point identification unit (301) specifically comprises: Determine the expected time interval between adjacent time steps in the time-encoded data based on the sampling frequency of the data; According to the start and end of the timestamps in the time-coded data, each pair of adjacent time steps is checked in turn, and for each pair of adjacent time steps, the actual time interval between them is calculated, and the actual interval is compared with the expected interval; If the actual interval is greater than the expected interval, it is considered that there is a missing point between the two time steps; the position of each identified missing point is recorded using the position code in the time step until the entire time code data is checked to generate the missing point of the time code data.

5. The radiation equipment detection data and report management system according to claim 4, characterized in that: The data completion module (300) comprises a missing point completion unit (302), and the process of generating complete time coding data by the missing point completion unit (302) specifically comprises: Initialize the attention weight matrix W_A, use the time embedding features of the missing points as the query vector Query_missing, use the time encoding data of all known data points as the key vector Key_known and the value vector Value_known, where the known data points represent the data of the non-missing points; calculate the dot product of the query vector Query_missing and the key vector Key_known; The dot product is scaled and the softmax function is applied to normalize the score to obtain the missing point attention score. The missing point attention score is used to perform weighted summation on the time coding data of the known data points to obtain the missing value of the missing point. The calculated missing value is filled into the corresponding position of the time coding data according to the position of the missing point to obtain the complete time coding data.

6. The radiation equipment detection data and report management system according to claim 5, characterized in that: The deviation detection module (400) trains a time series prediction model based on the complete time coding data; uses the trained time series prediction model to predict the next time step of the complete time coding data, and performs a difference calculation between the prediction result and the average result of the complete time coding data, and uses the difference result as the deviation result, specifically including: Use the complete time-coded data as the training set for the time series prediction model; create training samples using the sliding window technique by dividing the complete time-coded data into an input sequence and a target sequence, where the target sequence is the data of the next time step; optimize the loss function in the time series prediction model until the model converges or reaches a preset number of training rounds; Use the trained time series prediction model to predict each time window in the complete time-coded data. For each time step t, use the data from t-10 to t-1 to predict the value at time t. Save all prediction results to form a prediction sequence with the same length as the complete time-coded data. The average value of each feature in the complete time-coded data is calculated as the average result of the complete time-coded data; for each time step, the difference between the predicted result and the average result of the complete time-coded data is calculated, the difference of each time step is sorted into a sequence, a deviation threshold is set, the difference exceeding the deviation threshold is marked as a significant deviation, and the difference is used as the deviation result.

7. The radiation equipment detection data and report management system according to claim 6, characterized in that: The deviation detection module (400) selects the corresponding time embedding features and related tags according to the time step corresponding to the deviation result, and outputs them as a comprehensive report. The comprehensive report is as follows: Time information: including timestamp and location code to identify the specific time of each deviation; Deviation result: the difference calculated at each time step; Original data: time embedding vector corresponding to the time step; Prediction value: the prediction result of the time series prediction model for each time step; Mean: average value used for comparison; Significant deviation marking: Marks which time steps have deviations exceeding the preset deviation threshold.

Images