Marine scientific research project fund execution progress monitoring model construction method

Abstract

The invention provides a marine scientific research project fund execution progress monitoring model construction method, which belongs to the technical field of large model construction, and comprises the following steps: extracting a potential expenditure mode, identifying a project execution mode category by adopting a dynamic time warping algorithm, constructing a multi-task learning neural network, and predicting fund and progress at the same time. Starting a CUDA (Compute Unified Device Architecture) parallel computing framework to run a cumulative sum control chart, a Hotelling T square statistic and a wavelet multi-scale decomposition task in a three-layer thread block to realize anomaly detection, calculating an anomaly confidence coefficient through sequential probability ratio test to trigger an early warning mechanism, and adopting a multi-interpolation method for missing data, so as to complete the detection of the missing data. Robust M estimation is adopted for the abnormal observation value to reduce the influence weight, the optimal estimation result is output through the Kalman filtering fusion model prediction value and the noise observation value, and the technical problem that the marine scientific research project fund execution progress monitoring lacks the abnormal detection and prediction capability is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large model construction, and in particular, relates to a marine scientific research project fund execution progress monitoring model construction method. BACKGROUND

[0002] The marine scientific research project fund execution progress monitoring is an important link of project management, and the traditional monitoring method mainly relies on manual auditing of financial statements and periodic checking of progress nodes, and the execution status of the project is judged by comparing the deviation degree of the budget subject expenditure amount and the planned value. In the current marine scientific research project management practice, due to the long project cycle, complex fund subjects and various execution modes, the traditional manual monitoring method is difficult to timely find the abnormal trend in fund execution, and the problem is often not detected until it is serious, resulting in lagging correction measures. Although the existing automatic monitoring system can realize data collection and statistical analysis, it lacks the ability to deeply mine the historical execution mode, cannot adaptively select the monitoring strategy according to the characteristics of the project, and cannot capture weak abnormal signals in the multi-dimensional space-time correlation level. That is, there is a technical problem of lack of abnormal detection and prediction ability in the marine scientific research project fund execution progress monitoring in the prior art. SUMMARY

[0003] Therefore, the application provides a marine scientific research project fund execution progress monitoring model construction method, which can solve the technical problem of lack of abnormal detection and prediction ability in the marine scientific research project fund execution progress monitoring in the prior art.

[0004] The application is implemented in the following manner: the marine scientific research project fund execution progress monitoring model construction method provided by the application collects historical fund execution data and historical progress record data of a marine scientific research project, organizes the historical fund execution data into a three-dimensional tensor structure according to the subject category dimension, the time dimension and the project dimension, extracts a low-rank factor matrix group from the three-dimensional tensor structure by using a CP decomposition method, performs non-negative matrix decomposition on the low-rank factor matrix group to obtain a basic expenditure mode dictionary, generates a compact feature subset by using a local linear embedding algorithm, constructs a dynamic time warping reference template library based on the historical fund execution data, calculates the cumulative distance matrix of the real-time fund execution data of the current monitoring project and each reference sequence in the dynamic time warping reference template library, identifies the execution mode category by using a dynamic programming algorithm to find the optimal time warping path, constructs a multi-task learning neural network architecture to extract space-time feature representations and output fund prediction and progress prediction, and allocates three layers of thread blocks to respectively run the cumulative and control chart calculation task, the Hotelling T2 test and the cumulative sum control chart calculation task and the cumulative sum control chart calculation task of the GPU by starting all CUDA streams of the GPU. The tasks of calculating the squared statistic and wavelet multi-scale decomposition are carried out. The sequential probability ratio test method is used to calculate the anomaly confidence to trigger the early warning mechanism. For data with missing values, multiple imputation method is used to generate multiple candidate complete datasets. For noisy anomaly observations, robust M estimation is used to reduce the influence weight. The optimal estimation result is output by fusing the model prediction value and noisy observation value through Kalman filtering.

[0005] Furthermore, the three-dimensional tensor structure refers to a third-order array formed by arranging historical expenditure data according to the subject category dimension, time dimension, and project dimension. Each element of the third-order array represents the expenditure amount of the project in the project dimension at a certain point in time in the time dimension and the subject in the subject category dimension.

[0006] Furthermore, the optimal time curvature path refers to the path from the top left corner to the bottom right corner in the cumulative distance matrix that minimizes the cumulative distance. Each point on the optimal time curvature path represents the time alignment relationship between the corresponding positions in the query sequence and the reference sequence.

[0007] Furthermore, the multi-task learning neural network architecture specifically sets up shared convolutional layers and long short-term memory layers at the bottom layer to extract spatiotemporal feature representations, sets up a funding prediction branch and a schedule prediction branch at the top layer, and uses the CUDA parallel computing framework to accelerate the training process.

[0008] Furthermore, the first-level thread block continuously accumulates the deviation of real-time funding execution data from the execution mode category benchmark value to obtain a cumulative sum statistic. When the cumulative sum statistic exceeds the cumulative sum control limit, an alarm is triggered.

[0009] Furthermore, the second-layer thread block comprehensively monitors real-time funding execution data and historical progress record data, calculating multi-dimensional indicators for Hotelling. Square statistic, Hotelling The squared statistic measures the squared Mahalanobis distance between a multidimensional observation vector and the mean vector.

[0010] Furthermore, the third-layer thread block performs wavelet multi-scale decomposition on the real-time funding execution data to extract weak anomaly features, and avoids data competition between thread blocks by allocating independent memory regions for the first-layer, second-layer, and third-layer thread blocks.

[0011] Furthermore, the sequential probability ratio test method calculates the likelihood ratio point by point in the observation sequence and compares it with a preset threshold. When the likelihood ratio exceeds the upper threshold, it is determined that there is an anomaly, and when the likelihood ratio is lower than the lower threshold, it is determined that it is normal.

[0012] Furthermore, when the anomaly confidence level exceeds the judgment threshold, an early warning mechanism is triggered and the anomaly pattern characteristics are recorded. When the anomaly confidence level does not exceed the judgment threshold, real-time funding execution data is added to the training sample set of the normal behavior baseline model.

[0013] Furthermore, the local linear embedding algorithm specifically involves finding the nearest neighbor set for each data point in the basic expenditure pattern dictionary, calculating the weight coefficients of the data points linearly represented by the neighboring points, and finding the embedding coordinates that maintain the same weight relationship in the low-dimensional space as a compact feature subset.

[0014] Furthermore, the element in the i-th row and j-th column of the cumulative distance matrix represents the minimum cumulative distance between the first i points of the query sequence and the first j points of the reference sequence. The entire cumulative distance matrix is ​​filled using dynamic programming. The CP decomposition method, also known as the CANDECOMP-PARAFAC decomposition method, specifically decomposes the three-dimensional tensor structure into the outer product of three low-rank factor matrices. The row vectors of the factor matrices in the decomposed low-rank factor matrix group serve as the latent expenditure pattern representation vectors. The non-negative matrix decomposition specifically decomposes the non-negative matrix formed by the latent expenditure pattern representation vectors into the product of two non-negative matrices. Iterative optimization minimizes the reconstruction error, thereby extracting the basic expenditure pattern dictionary and activation coefficient matrix.

[0015] Furthermore, the multiple imputation method specifically generates multiple candidate complete datasets to fill in missing values ​​based on the distribution of the observed data, and uses the expectation-maximization algorithm to estimate parameters and calculate confidence intervals on the multiple candidate complete datasets.

[0016] Furthermore, the robust M-estimation is specifically achieved by reducing the weight of outlier observations using a loss function with a growth rate lower than that of the square function, and by adjusting the influence weights to achieve robustness to outliers. The estimation process is solved through weighted least squares iteration.

[0017] Furthermore, the Kalman filter includes a prediction step and an update step. The prediction step calculates a priori estimates based on the state transition equation and process noise, and the update step calculates a posterior estimate based on the observation equation and measurement noise. The Kalman gain determines the fusion weight of the model prediction and the noise observation.

[0018] This invention addresses the technical problem of insufficient anomaly detection and prediction capabilities in monitoring the progress of marine research project funding by constructing an intelligent monitoring model based on tensor decomposition and multi-task learning. It utilizes CANDECOMP-PARAFAC decomposition to extract potential expenditure patterns and combines this with a dynamic time warping algorithm to identify the execution mode category of the project. This allows the monitoring strategy to adaptively adjust according to project characteristics, overcoming the limitations of traditional methods in handling diverse execution modes. Furthermore, this invention employs a multi-task learning neural network to simultaneously predict funding and progress, extracting spatiotemporal feature representations through a shared layer, thus achieving correlation modeling between funding expenditure and progress. Finally, this invention designs a three-layer anomaly detection mechanism based on CUDA parallel computing. The first layer detects process drift through cumulative sum and control charts; the second layer monitors multi-dimensional indicators using the Hotelling T-squared statistic; and the third layer captures weak anomaly features through wavelet multi-scale decomposition. The three layers work collaboratively to achieve comprehensive anomaly signal capture. In summary, this invention solves the technical problem mentioned in the background art of insufficient anomaly detection and prediction capabilities in monitoring the progress of marine research project funding. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 This is a time-series graph showing the cumulative and statistical changes of ship charter fees.

[0021] Figure 3 This is a comparison chart of the observed values ​​and the optimal estimation results during the Kalman filter fusion process. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0023] like Figure 1 The diagram shown is a flowchart of a method for constructing a monitoring model for the execution progress of marine scientific research project funding provided by this invention. This method includes the following steps: S01. Collect historical funding execution data and historical progress record data of marine scientific research projects. Organize the historical funding execution data into a three-dimensional tensor structure according to the dimensions of subject category, time, and project. Organize the historical progress record data into a two-dimensional matrix structure according to the dimensions of task node and completion. Perform standardization preprocessing on the three-dimensional tensor structure and the two-dimensional matrix structure and remove obvious erroneous records. S02. The CANDECOMP-PARAFAC decomposition method is used to extract the low-rank factor matrix group from the three-dimensional tensor structure. The row vectors of the factor matrix in the low-rank factor matrix group are used as the latent expenditure pattern representation vectors. The latent expenditure pattern representation vectors are decomposed into a non-negative matrix to obtain the basic expenditure pattern dictionary. The local linear embedding algorithm is used to perform manifold structure preservation processing on the basic expenditure pattern dictionary to generate a compact feature subset. S03. Construct a dynamic time warping reference template library based on historical funding execution data. Use the real-time funding execution data of the current monitored project as the query sequence. Calculate the cumulative distance matrix between the query sequence and each reference sequence in the dynamic time warping reference template library. Use a dynamic programming algorithm to find the optimal time warping path in the cumulative distance matrix to achieve non-linear time alignment. Identify the execution mode category of the current monitored project based on the optimal time warping path. S04. Construct a multi-task learning neural network architecture. Set up a shared convolutional layer and a long short-term memory layer at the bottom layer of the multi-task learning neural network architecture. Input a compact feature subset and historical progress record data into the shared convolutional layer and the long short-term memory layer to extract spatiotemporal feature representations. Set up a funding prediction branch and a progress prediction branch at the top layer of the multi-task learning neural network architecture. Use the CUDA parallel computing framework to accelerate the training process of the multi-task learning neural network architecture. S05. Start all CUDA streams of the graphics processing unit. Within each CUDA stream, allocate a first-layer thread block to run the cumulative sum and control chart calculation task, a second-layer thread block to run the Hotelling T-squared statistic calculation task, and a third-layer thread block to run the wavelet multi-scale decomposition calculation task. The first-layer thread block continuously accumulates the deviation of real-time funding execution data relative to the execution mode category benchmark value to obtain the cumulative sum statistic. The second-layer thread block comprehensively monitors the multi-dimensional indicators of real-time funding execution data and historical progress record data to calculate the Hotelling T-squared statistic. The third-layer thread block performs wavelet multi-scale decomposition on the real-time funding execution data to extract weak anomaly features. Data competition between thread blocks is avoided by allocating independent memory regions for the first-layer thread block, the second-layer thread block, and the third-layer thread block. S06. For real-time funding execution data where the cumulative sum statistic exceeds the cumulative sum control limit or the Hotelling T-squared statistic exceeds the Hotelling control limit, the sequential probability ratio test method is used to calculate the anomaly confidence level. When the anomaly confidence level exceeds the judgment threshold, an early warning mechanism is triggered and the anomaly pattern characteristics are recorded. When the anomaly confidence level does not exceed the judgment threshold, the real-time funding execution data is added to the training sample set of the normal behavior baseline model. S07. For real-time funding execution data and historical progress record data with missing values, multiple interpolation methods are used to generate multiple candidate complete datasets. The expectation-maximization algorithm is used to estimate parameters and calculate confidence intervals on multiple candidate complete datasets. Robust M-estimation is used to reduce the influence weight of abnormal observations on model parameters by noisy outlier observations. The optimal estimation result is output by fusing model predictions and noisy observations through Kalman filtering.

[0024] Among them, the three-dimensional tensor structure refers to a third-order array formed by arranging historical expenditure data according to the subject category dimension, time dimension, and project dimension. Each element of the third-order array represents the expenditure amount of the project in the project dimension at a certain point in time in the time dimension and the subject in the subject category dimension.

[0025] The CANDECOMP-PARAFAC decomposition method is a tensor decomposition technique that decomposes a three-dimensional tensor structure into the sum of the outer products of three low-rank factor matrices. The formula is expressed as follows: each element of the three-dimensional tensor structure is equal to the sum of the products of the corresponding row vector elements of each low-rank factor matrix, where the number of accumulated terms is the rank parameter. The three factors in each accumulated term are divided by the maximum element value of their respective low-rank factor matrices to achieve dimensionlessness. The row vectors of the factor matrices in the decomposed low-rank factor matrix group are the potential expenditure pattern representation vectors.

[0026] Nonnegative matrix factorization is a method that decomposes a nonnegative matrix into the product of two nonnegative matrices. It is used to extract the basic expenditure pattern dictionary and activation coefficient matrix from the latent expenditure pattern representation vector. The decomposition process minimizes the reconstruction error through iterative optimization. In each iteration, the error term is divided by the Frobenius norm of the original matrix to achieve dimensionlessness.

[0027] Locally linear embedding (LANE) is a manifold learning method that maps high-dimensional data to a low-dimensional space by preserving the local linear relationships within the neighborhood of data points. The LANE first finds the nearest neighbor set for each data point in the basic expenditure pattern dictionary, then calculates the weight coefficients of the data points in linear representation of the neighborhood points, and finally finds the embedding coordinates in the low-dimensional space that maintain the same weight relationship as a compact feature subset.

[0028] The Dynamic Time Warping Reference Template Library is a database that stores the time series of funding execution for multiple historical projects. Each reference sequence in the Dynamic Time Warping Reference Template Library records the funding expenditure of a historical project from start to finish at each point in time, and is used to compare the similarity with the query sequence of the currently monitored project.

[0029] The cumulative distance matrix is ​​a two-dimensional matrix used in the dynamic time warping algorithm. The element in the i-th row and j-th column of the cumulative distance matrix represents the minimum cumulative distance between the first i points of the query sequence and the first j points of the reference sequence. During calculation, the distance between the current point pairs is divided by the sum of the lengths of the two sequences to achieve dimensionlessness. The entire cumulative distance matrix is ​​filled by dynamic programming.

[0030] The optimal time curvature path is a path from the top left corner to the bottom right corner in the cumulative distance matrix. Each point on the optimal time curvature path represents the time alignment relationship between the corresponding positions in the query sequence and the reference sequence. The optimal time curvature path minimizes the cumulative distance.

[0031] The execution mode category is a classification based on the historical patterns of project funding execution. Different execution mode categories have different funding expenditure rhythms and progress characteristics. The execution mode category of the currently monitored project is determined by the similarity of the optimal time curvature path.

[0032] Multi-task learning neural network architecture refers to a deep network structure that includes shared layers and task-specific layers. The bottom shared convolutional layer performs convolution operations on a compact subset of input features and historical progress record data to extract local patterns. The long short-term memory layer captures the long-term dependencies of the time series and outputs spatiotemporal feature representations. The top funding prediction branch outputs the remaining funding amount for future periods based on the spatiotemporal feature representations, and the progress prediction branch outputs the project completion percentage based on the spatiotemporal feature representations.

[0033] The CUDA parallel computing framework refers to a programming model that uses graphics processing units for general-purpose computing. It accelerates computation by launching multiple parallel threads to process different data simultaneously, thus avoiding the problems of gradient vanishing and slow convergence that occur during the training of multi-task learning neural network architectures.

[0034] CUDA streams are a mechanism in CUDA programming used to manage asynchronous operations. Operations in different CUDA streams are executed concurrently on the graphics processing unit, and each CUDA stream maintains an independent task queue and synchronization state.

[0035] The first-level thread block refers to the thread organization unit allocated on the graphics processing unit for performing cumulative sum and statistic updates. The cumulative sum control chart detects process drift by accumulating the deviation between the current observation and the target value. The formula for calculating the cumulative sum statistic is: the current cumulative sum statistic equals the previous cumulative sum statistic plus the dimensionless value of the current deviation divided by the standard deviation of the historical deviation. An alarm is triggered when the cumulative sum statistic exceeds the cumulative sum control limit.

[0036] The second-level thread block refers to the thread organization unit allocated on the graphics processing unit for calculating the Hotelling T-squared statistic in multivariate statistical process control. The Hotelling T-squared statistic measures the squared Mahalanobis distance between the multidimensional observation vector and the mean vector. The formula for calculating the Hotelling T-squared statistic is: the statistic equals the transpose of the deviation vector multiplied by the inverse of the covariance matrix multiplied by the deviation vector, where each component of the deviation vector is divided by the standard deviation of the corresponding variable to achieve dimensionlessness.

[0037] The third-layer thread block refers to the thread organization unit allocated on the graphics processing unit for performing wavelet transform operations. Wavelet multi-scale decomposition decomposes real-time data into wavelet coefficients of different frequency scales. By using wavelet transform, signal features can be located in both the time and frequency domains. It is suitable for detecting weak anomalies submerged in noise. During the decomposition process, the wavelet coefficients of each scale are divided by the energy of the corresponding scale to achieve dimensionlessness.

[0038] Data races refer to the unpredictable behavior that may occur when multiple thread blocks access the same memory location simultaneously and at least one thread block performs a write operation. Data races can be avoided by allocating independent memory regions for the first-level thread blocks, the second-level thread blocks, and the third-level thread blocks, and by using atomic operations to ensure data consistency.

[0039] The benchmark value for the execution mode category refers to the average funding execution level of historical projects within the corresponding execution mode category, serving as a reference standard for judging whether the funding execution of the currently monitored project is normal.

[0040] Cumulative sum control limits are thresholds used in cumulative sum control charts to determine whether a process is out of control. They are determined based on the statistical distribution of historical data and are typically set to 3 to 5 times the standard deviation of historical deviations.

[0041] The Hotelling control limit is a threshold used in Hotelling T-squared control charts to determine whether a multidimensional process is abnormal. It is calculated based on the quantiles of the F-distribution and the degrees of freedom of historical data.

[0042] The sequential probability ratio test is a statistical hypothesis testing method that calculates the likelihood ratio point by point in the observation sequence and compares it with a preset threshold. When the likelihood ratio exceeds the upper threshold, the null hypothesis is rejected and an anomaly is determined. When the likelihood ratio is lower than the lower threshold, the null hypothesis is accepted and a normal result is determined. The likelihood ratio is calculated by taking the logarithm of the probability density ratio under the two hypotheses and summing them to achieve dimensionless representation.

[0043] Anomaly confidence is a numerical value output by the sequential probability ratio test method, representing the likelihood of an anomaly. The higher the anomaly confidence, the greater the possibility that the real-time funding execution data deviates from the normal pattern.

[0044] The judgment threshold is a critical value used to determine whether to trigger the early warning mechanism. It is set based on statistical analysis of historical abnormal cases and the risk tolerance of managers.

[0045] Anomaly pattern characteristics are a set of parameters that describe the features of abnormal real-time funding execution data, including the time point of the anomaly, the category of the subject involved, the degree of deviation, and the duration, which are used for subsequent anomaly cause analysis and prevention measure formulation.

[0046] The normal behavior baseline model is a statistical model that describes the normal execution of project funds and progress. It is established by learning from historical normal project data and is used to identify abnormal deviations in the currently monitored project.

[0047] The training sample set is the dataset used by the normal behavior baseline model to learn normal patterns, and it is continuously updated as new normal real-time funding execution data is added.

[0048] Multiple imputation is a statistical technique for processing missing data. It generates multiple candidate complete datasets to fill in missing values ​​based on the distribution of observed data. Each candidate complete dataset reflects different possible values ​​of the missing values. By combining the analysis results of multiple candidate complete datasets, an inference that takes into account uncertainty is obtained.

[0049] The expectation-maximization algorithm is an iterative algorithm for maximum likelihood estimation in the presence of missing data or latent variables. The expectation step calculates the posterior distribution of the latent variables based on the current parameters, and the maximization step updates the parameters based on the expected complete data likelihood function. The process iterates until convergence.

[0050] A confidence interval is the range of uncertainty in a parameter estimate, indicating that the true parameter value falls within the interval with a certain probability. The width of the confidence interval reflects the accuracy of the estimate.

[0051] Abnormal observations refer to values ​​in real-time funding execution data or historical progress records that deviate significantly from the normal range, which may be caused by system failures, manual input errors, or transmission interruptions.

[0052] Robust M-estimation is a parameter estimation method that resists the influence of outliers. It reduces the weight of outlier observations by using a loss function with a growth rate lower than that of the square function. Commonly used loss functions include the Huber loss function and the graph basis double weight function. The estimation process is solved by weighted least squares iteration.

[0053] Influence weights are weight coefficients assigned to each observation in robust M-estimation. The influence weights of outlier observations are smaller, while those of normal observations are larger. Robustness to outliers is achieved by adjusting the influence weights.

[0054] Kalman filtering is a recursive estimation algorithm that estimates the system state by fusing the dynamic model predictions and noisy observations. The Kalman filtering process includes a prediction step and an update step. The prediction step calculates the prior estimate based on the state transition equation and process noise, while the update step calculates the posterior estimate based on the observation equation and measurement noise. The Kalman gain determines the fusion weight of the model predictions and the noisy observations.

[0055] The model predictions are future funding and schedule values ​​extrapolated by a multi-task learning neural network architecture based on historical data and the current state.

[0056] Noise observations are real-time budget execution data and historical progress records that contain measurement errors and random interference.

[0057] The optimal estimation result is the state estimate output by fusing the Kalman filter model prediction and the noisy observation, and it has the characteristic of minimum mean square error.

[0058] The detailed structure of the constructed monitoring model includes: a data acquisition layer responsible for acquiring raw historical expenditure execution data and historical progress record data from the financial system and project management system; a preprocessing layer standardizing the historical expenditure execution data and organizing it into a three-dimensional tensor structure, and standardizing the historical progress record data and organizing it into a two-dimensional matrix structure; a feature extraction layer converting the three-dimensional tensor structure into a low-rank factor matrix group through CANDECOMP-PARAFAC decomposition, extracting the basic expenditure pattern dictionary from the low-rank factor matrix group through non-negative matrix decomposition, and reducing the dimensionality of the basic expenditure pattern dictionary into a compact feature subset through a local linear embedding algorithm; a pattern recognition layer calculating the cumulative distance matrix between the query sequence and the reference sequence in the dynamic time warping reference template library based on the dynamic time warping algorithm, and identifying the execution pattern category through the optimal time warping path; and a prediction layer using a multi-task learning neural network architecture to process the compact feature subset and historical progress record data, with the bottom layer sharing convolutional layers and long short-term memory layers to extract spatiotemporal features. The system is characterized by two layers: a top-level funding prediction branch and a progress prediction branch, which output the remaining funding amount and the percentage of project completion, respectively. The anomaly detection layer utilizes the CUDA parallel computing framework to launch multiple CUDA streams. The first-layer thread block runs the cumulative sum control chart calculation task and outputs the cumulative sum statistic. The second-layer thread block runs the Hotelling T-squared statistic calculation task, and the third-layer thread block runs the wavelet multi-scale decomposition task to extract weak anomaly features. Anomaly confidence is calculated using the sequential probability ratio test method and compared with a judgment threshold to trigger an early warning mechanism. The robustness enhancement layer performs multiple interpolation on real-time funding execution data and historical progress record data to generate multiple alternative complete datasets. Parameter estimation and confidence interval calculation are performed using the expectation-maximization algorithm. Robust M-estimation reduces the influence weight of anomaly observations. Kalman filtering fuses model predictions and noisy observations to output the optimal estimation result. The decision output layer integrates the results from each layer to generate a project health assessment report and risk warning information, providing decision support for managers.

[0059] In addition, the present invention also provides a method for implementing a marine scientific research project funding execution progress monitoring system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned method when the computer is run.

[0060] The specific implementation methods of the above steps are described in detail below.

[0061] The specific implementation of step S01 involves extracting historical funding execution data and historical progress record data for marine scientific research projects from the financial system and project management system. The historical funding execution data includes expenditure records for each project at different points in time for each category, while the historical progress record data includes completion information for each project at different task nodes. The historical funding execution data is arranged into a three-dimensional tensor structure in the form of a three-dimensional array, categorized by category, time, and project. The category dimension corresponds to categories such as equipment costs, material costs, and personnel costs; the time dimension corresponds to monthly or quarterly time scales; and the project dimension corresponds to different historical project numbers. The historical progress record data is arranged into a two-dimensional matrix structure, categorized by task node and completion degree. The task node dimension corresponds to key milestones in the project plan, and the completion degree dimension corresponds to completion rates from 0% to 100%. Standardization preprocessing is performed on the three-dimensional tensor structure and the two-dimensional matrix structure. The mean value of each dimension is subtracted, and then divided by the standard deviation to achieve zero-mean unit variance normalization. Clearly erroneous records with values ​​exceeding three times the reasonable standard deviation are removed to ensure the reliability of the data used in subsequent analysis. The purpose of these steps is to construct a structured multidimensional data representation, laying the foundation for feature extraction from high-dimensional sparse data.

[0062] The specific implementation of step S02 involves using the CANDECOMP-PARAFAC decomposition method to perform low-rank approximate decomposition of the three-dimensional tensor structure. The decomposition algorithm iteratively optimizes three low-rank factor matrices using alternating least squares, ensuring that the outer product of these three low-rank factor matrices optimally approximates the original three-dimensional tensor structure. During the iteration process, two low-rank factor matrices are fixed while solving for the least squares solution of the third low-rank factor matrix. This process is repeated until the reconstruction error converges to a set threshold below 0.01 or the number of iterations reaches 100. The resulting low-rank factor matrix group includes subject dimension factor matrices, time dimension factor matrices, and project dimension factor matrices. The row vectors of the subject dimension factor matrix are extracted as potential expenditure pattern representation vectors, capturing the expenditure correlation patterns between different subjects. The matrix formed by the potential expenditure pattern representation vectors is then subjected to non-negative matrix decomposition, decomposing it into a product of a basic expenditure pattern dictionary and an activation coefficient matrix. The two non-negative matrices are iteratively optimized using a multiplication update rule to minimize the reconstruction error. The iteration terminates when the relative error change is less than 0.001 or after 50 iterations. The column vectors of the basic expenditure pattern dictionary represent typical expenditure pattern types. A local linear embedding algorithm is used to reduce the dimensionality of the basic expenditure pattern dictionary. The algorithm first finds the 10 nearest neighbors of each data point by Euclidean distance. Then, it solves for linear reconstruction weights to achieve an optimal weighted combination of neighbors that approximates the center point. Finally, it solves for the embedding coordinates in the low-dimensional space while maintaining the same linear reconstruction relationship. The dimensionality-reduced embedding coordinates form a compact feature subset, reducing the dimensionality from hundreds of dimensions to 20-50 dimensions. The purpose of these steps is to transform high-dimensional, sparse expenditure data into a low-dimensional, dense feature representation through tensor decomposition and manifold learning techniques, thus solving the difficulties in feature extraction caused by the curse of dimensionality and sparsity.

[0063] The specific implementation of step S03 involves extracting time-series data from the start to the end of each historical project from historical funding execution data to construct a dynamic time warping reference template library. Each reference sequence records the cumulative expenditure amount of a historical project at each time point. Real-time funding execution data of the currently monitored project is obtained as a query sequence, containing the cumulative expenditure time series of the current monitored project from its start to the present. The cumulative distance matrix between the query sequence and each reference sequence in the dynamic time warping reference template library is calculated. The matrix elements are calculated using a dynamic programming method. The element in the i-th row and j-th column equals the minimum value of the Euclidean distance between the i-th point of the query sequence and the j-th point of the reference sequence plus the cumulative distances of the three candidate predecessor positions. The three candidate predecessor positions are the adjacent elements on the left, below, and lower left, respectively. The path that minimizes the cumulative distance from the start to the end in the cumulative distance matrix is ​​searched as the optimal time warping path. The path search uses a backtracking method, tracing back from the end position along the predecessor direction with the minimum cumulative distance to the start position. The similarity between the query sequence and each reference sequence is determined based on the minimum cumulative distance corresponding to the optimal time warping path. The execution mode categories of the top 5 reference sequences with the smallest cumulative distances are then voted on, and the category with the most votes is the execution mode category of the current monitored item. The purpose of this step is to achieve non-linear time alignment between the query sequence and the reference sequences through a dynamic time warping algorithm, overcoming the time scale inconsistency caused by differences in execution rhythms, and accurately identifying the execution mode affiliation of the current monitored item.

[0064] The specific implementation of step S04 involves constructing a multi-task learning neural network architecture comprising an input layer, a shared layer, and an output layer. The input layer receives a concatenated vector of a compact feature subset and historical progress record data. The shared layer comprises two one-dimensional convolutional layers and one long short-term memory layer. The first one-dimensional convolutional layer uses 64 convolutional kernels to extract local temporal patterns, with a kernel size of 3 and a modified linear unit activation function. The second one-dimensional convolutional layer uses 128 convolutional kernels to further extract high-level features. The long short-term memory layer contains 256 hidden units to capture long-term dependencies in the time series. The long short-term memory layer uses a gating mechanism to control the inflow, outflow, and forgetting of information. The output layer comprises two task-specific branches: a funding prediction branch maps the output of the long short-term memory layer to a predicted value of the remaining funding for the next 12 months through a fully connected layer; and a progress prediction branch outputs a predicted percentage of project completion through another fully connected layer. The training of the multi-task learning neural network architecture employs backpropagation and an adaptive moment estimation optimizer. The loss function is the weighted sum of the prediction errors of the two branches. The budget prediction branch uses mean squared error loss, while the schedule prediction branch uses smoothed L1 loss, with weight coefficients set to 0.6 and 0.4, respectively. Using the CUDA parallel computing framework, training data is batched across multiple streaming multiprocessors in the graphics processing unit for parallel forward and backward propagation computation. The batch size is set to 128, and the initial learning rate is 0.001, decaying to 0.5 times the original rate every 30 epochs. The training converges after 200 epochs. The purpose of these steps is to achieve joint optimization of budget and schedule prediction through a multi-task learning mechanism, utilizing shared layers to capture the intrinsic correlation between budget and schedule, thereby improving prediction accuracy.

[0065] The specific implementation of step S05 involves activating all CUDA streams on the graphics processing unit to achieve parallel processing of the anomaly detection task. Each CUDA stream is allocated three layers of thread blocks to execute different statistical monitoring calculations. The first layer of thread blocks is responsible for the calculation of the cumulative sum and control chart. Each thread processes the update of the cumulative sum and statistics of one item in the real-time expenditure execution data. The thread reads the deviation between the real-time expenditure amount of the current item and the baseline value of the execution mode category, divides the deviation by the historical deviation standard deviation to obtain the standardized deviation, and adds the standardized deviation to the cumulative sum and statistics of the previous moment. When the cumulative sum and statistics exceed the cumulative sum and control limit of 5.0, it is marked as a preliminary anomaly. The second layer of thread blocks is responsible for the calculation of the Hotelling T-squared statistic. The thread reads the multidimensional observation vector composed of real-time expenditure execution data and historical progress record data, calculates the deviation vector between the observation vector and the mean vector, calculates the Hotelling T-squared statistic by matrix multiplication, multiplies the transpose of the deviation vector by the inverse of the covariance matrix, and multiplies the deviation vector. When the statistic exceeds the Hotelling control limit of 15.0, it is marked as a preliminary anomaly. The third-layer thread block is responsible for calculating wavelet multi-scale decomposition. The thread performs discrete wavelet transform on the real-time funding data, using Daubechies wavelet basis functions for three-level decomposition to obtain approximation coefficients and detail coefficients at different scales. The abrupt change characteristics of the detail coefficients are used to identify weak anomalous signals. Independent memory is allocated to the first-layer thread block to store the cumulative and statistical arrays, to the second-layer thread block to store the covariance matrix and intermediate calculation results, and to the third-layer thread block to store the wavelet coefficients. Independent memory allocation and atomic operations avoid data contention and write conflicts between thread blocks. The purpose of these steps is to utilize the massively parallel computing capabilities of the graphics processing unit to achieve real-time calculation of various statistical monitoring methods, improving the detection sensitivity of early weak anomalous signals through accumulating small deviations and multi-scale decomposition.

[0066] The specific implementation of step S06 involves further verifying the real-time funding execution data that triggers preliminary anomalies using the cumulative sum statistic or Hotelling's T-squared statistic. An anomaly confidence level is calculated using the sequential probability ratio test method. This method maintains the cumulative likelihood ratio values ​​for two hypotheses: the null hypothesis is that the data follows a normal distribution, and the alternative hypothesis is that the data follows an abnormal distribution. For each newly arrived observation, the probability density function value under the two hypotheses is calculated, and the ratio of these values ​​is taken and its logarithm is accumulated in the cumulative likelihood ratio value. When the cumulative likelihood ratio value exceeds the upper threshold of 3.0, the null hypothesis is rejected, indicating an anomaly. The anomaly confidence level is calculated as the ratio of the cumulative likelihood ratio value to the upper threshold. When the anomaly confidence level exceeds the judgment threshold of 0.85, an early warning mechanism is triggered. The time of anomaly occurrence, the involved subjects, the degree of deviation, and the abrupt change position of wavelet detail coefficients are recorded as anomaly pattern features and stored in the anomaly database. When the cumulative likelihood ratio is below the lower threshold of -3.0 or the anomaly confidence level does not exceed the judgment threshold, the null hypothesis is accepted and the fluctuation is judged as normal. The real-time funding execution data is then added to the training sample set of the normal behavior baseline model to update the model parameters. The purpose of this step is to reduce the false alarm rate through sequential hypothesis testing and achieve rapid and reliable anomaly detection under the constraint of a fixed false alarm probability.

[0067] The specific implementation of step S07 involves identifying the locations of missing values ​​and noise outliers in real-time funding execution data and historical progress record data. Missing values ​​are identified by their empty values ​​or by marking them as missing. Noise outliers are identified using a box plot method as extreme values ​​exceeding 1.5 times the interquartile range. For data containing missing values, a multiple imputation method is used to generate five candidate complete datasets. Each imputation generates imputation values ​​for missing locations by randomly sampling from a multivariate normal distribution model fitted to the observed data. The five candidate complete datasets reflect different possible realizations of missing values. On each candidate complete dataset, the expectation-maximization algorithm is applied to estimate model parameters. The expectation step calculates the conditional expectation of the missing data based on the current parameters, and the maximization step updates the parameters based on the likelihood function of the complete data. After 20 iterations, the parameter estimates converge. The average of the parameter estimates obtained from the five candidate complete datasets is calculated as the final parameter estimate. The standard deviation of the parameter estimates is calculated, and a 95% confidence interval is constructed using a t-distribution to quantify the uncertainty. Robust M-estimation is used to reduce the impact of identified noise outliers. M-estimation uses the Huber loss function instead of the least squares squared loss. Squared loss is used when the absolute value of the residual is less than a threshold of 1.345 times the standard deviation; linear loss is used when the absolute value of the residual exceeds the threshold. Parameters are solved iteratively using weighted least squares to minimize the weighted sum of squared residuals, thus assigning a smaller impact weight to noise outliers. The parameters obtained from robust M-estimation are substituted into the state-space model, and the Kalman filter algorithm is applied to fuse model predictions and noisy observations. The prediction step predicts the state and error covariance at the next time step based on the state transition equation and process noise covariance. The update step calculates the Kalman gain based on the observation equation and updates the state estimate and error covariance. The Kalman gain automatically balances the reliability of model predictions and observations, and the optimal estimation result is output as the denoised funding and progress data. The purpose of these steps is to process missing and noisy data through multiple interpolation and robust estimation techniques, ensuring the accuracy of parameter estimation and the stability of predictions in monitoring models under conditions of suboptimal data quality.

[0068] It should be noted that one of the key technical approaches of this invention is to address the challenge of representing high-dimensional sparse funding data by employing a feature extraction method that combines tensor decomposition and manifold learning. Traditional statistical methods are prone to the curse of dimensionality and struggle to capture latent patterns in high-dimensional sparse data, while machine learning models face the risk of overfitting. This invention decomposes the three-dimensional funding tensor into a set of low-rank factor matrices using CANDECOMP-PARAFAC decomposition. It utilizes the multilinear structure of the tensor to discover the coupling relationships between different dimensions. Nonnegative matrix decomposition further extracts a dictionary of basic expenditure patterns with partial interpretability. The locally linear embedding algorithm preserves the inherent manifold geometry of the data during dimensionality reduction, avoiding the loss of nonlinear features inherent in linear dimensionality reduction methods. The resulting compact feature subset reduces computational complexity while retaining key discriminative information.

[0069] The second key technical approach is to construct an early anomaly detection mechanism based on a CUDA multi-layer thread block parallel architecture to enhance the identification capability of weak signals. Traditional threshold alarm methods trigger only after the anomaly has fully manifested, resulting in delayed intervention. Increasing sensitivity introduces a large number of false alarms. This invention designs three-layer thread blocks to run cumulative sum control charts, Hotelling's T-squared statistic, and wavelet multi-scale decomposition tasks respectively. The cumulative sum method amplifies the statistical evidence of early anomalies by continuously accumulating small deviations. Hotelling's T-squared statistic enables joint monitoring of multi-dimensional indicators to avoid missing comprehensive anomalies by univariate methods. Wavelet multi-scale decomposition extracts signal features at different frequency scales to identify weak anomaly patterns submerged in noise. The sequential probability ratio test minimizes the detection delay under a fixed false alarm rate constraint. The parallel computation and complementary fusion of multiple methods significantly improve the detection sensitivity and reliability of early anomalies.

[0070] The third key technical approach is to develop a data quality assurance mechanism that integrates multiple imputation and robust estimation to address the challenges of missing and noisy data. Directly deleting missing samples leads to bias and information loss, while simple imputation introduces spurious patterns, and noisy data contaminates model parameters. This invention uses multiple imputation to generate multiple alternative complete datasets to quantify the uncertainty of missing values. The expectation-maximization algorithm performs unbiased parameter estimation under incomplete data, robust M-estimation reduces the destructive impact of outliers through an adaptive weighting function, and Kalman filtering recursively fuses model prediction and noisy observations to achieve optimal state estimation. The entire mechanism forms a complete robust processing chain from data completion, parameter estimation, anomaly suppression to state fusion.

[0071] The synergistic effect of the three key technological approaches described above has enabled the optimization of the entire process for monitoring the progress of marine scientific research project funding. Tensor decomposition and manifold learning provide high-quality feature representations, laying the foundation for subsequent analysis; the multi-layer parallel anomaly detection mechanism quickly and accurately identifies early risk signals in the feature space; and the data quality assurance mechanism ensures the stability and reliability of computation at each stage. These three elements work together to form a complete technological chain from data preprocessing, feature extraction, pattern recognition to anomaly detection. Compared to traditional methods that only focus on improving a single stage, the systematic collaborative design of this invention achieves an overall breakthrough in three key performance dimensions: feature representation capability, anomaly detection sensitivity, and model robustness, providing reliable technical support for the refined management of marine scientific research projects.

[0072] It should be noted that this invention also solves the following technical problem: Existing technologies suffer from reduced monitoring accuracy due to missing values ​​and noise interference in marine scientific research project funding data. This invention generates multiple candidate complete datasets from real-time funding execution data containing missing values ​​using a multiple imputation method. It then uses the expectation-maximization algorithm to estimate parameters and calculate confidence intervals on these candidate complete datasets. This method does not simply fill in missing data with a single value, but generates multiple possible complete datasets to reflect the uncertainty of missing values, thus avoiding the bias that may be introduced by traditional single imputation methods. This invention uses robust M-estimation to reduce the impact of noisy outlier observations on model parameters. By using a loss function with a growth rate lower than the square function, outliers do not have an excessive impact on the overall parameter estimation. Furthermore, this invention fuses model predictions and noisy observations using Kalman filtering, adaptively adjusting the fusion weights using Kalman gain. Random interference is gradually eliminated during the recursive processes of the prediction and update steps, outputting the optimal estimation result with the minimum mean square error characteristic, thereby ensuring the robustness and accuracy of the monitoring model even under conditions of suboptimal data quality.

[0073] Specifically, the principle of this invention is as follows: This invention solves the technical problem of lacking anomaly detection and prediction capabilities in the monitoring of the execution progress of marine scientific research project funding. Its principle lies in reducing the dimensionality of three-dimensional historical funding data to a low-rank factor matrix group through tensor decomposition, extracting hidden basic expenditure patterns from these patterns. These patterns reflect the correlation expenditure patterns of different subjects at different time periods, laying the foundation for subsequent pattern recognition. The dynamic time warping algorithm achieves non-linear time alignment between the current project and historical projects by calculating the cumulative distance matrix and finding the optimal time warping path. This allows even projects with different execution rhythms to find similar reference templates, thereby accurately determining the execution mode category of the current project. The multi-task learning neural network extracts spatiotemporal feature representations through the bottom-level shared convolutional layers and long short-term memory layers, predicting funding and progress at the top level. This architecture utilizes the inherent correlation between funding execution and progress advancement, improving prediction accuracy. The three-layer parallel anomaly detection mechanism monitors anomaly signals from different angles: the cumulative and control charts focus on process trends, the Hotelling T-squared statistic focuses on multi-dimensional synthesis, and wavelet decomposition focuses on frequency domain features. These three complement each other to form a comprehensive anomaly capture capability, consistent with the complex characteristics of marine scientific research project funding execution.

[0074] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0075] The specific implementation of step S01 is to collect historical funding execution data and historical progress record data of marine scientific research projects, and organize the historical funding execution data into a three-dimensional tensor structure according to the dimensions of subject category, time, and project. ,in This represents the total number of subject categories. The total number of time points. Tensor elements represent the total number of projects. Indicates the first The project in the The time point at the The expenditure amount for each item, expressed in ten thousand yuan, is organized into a two-dimensional matrix structure based on historical progress records. ,in This represents the total number of task nodes. The total number of completion levels, matrix elements Indicates the first The task node reached the first The number of records at each completion level is used to standardize and preprocess the three-dimensional tensor structure and the two-dimensional matrix structure, and to remove obvious erroneous records.

[0076] The specific implementation of step S02 is to extract the low-rank factor matrix group from the three-dimensional tensor structure using the CANDECOMP-PARAFAC decomposition method. The decomposition formula is expressed as follows: ; In the formula, For the third three-dimensional tensor structure Subject category, No. The time point, the first The expenditure amount for each project is in ten thousand yuan. This is the rank parameter, with a value ranging from 5 to 15; Subject category factor matrix The Line number Column elements, dimensionless; Time factor matrix The Line number Column elements, dimensionless; Project factor matrix The Line number Column elements, dimensionless; For matrix The largest element value in the array, dimensionless; For matrix The largest element value in the array, dimensionless; For matrix The largest element value in the array, dimensionless; represents the maximum expenditure amount in the three-dimensional tensor structure, in ten thousand yuan. The row vectors of the factor matrices in the low-rank factor matrix group are used as the representation vectors of the potential expenditure patterns. The potential expenditure pattern representation vectors are then subjected to non-negative matrix decomposition, as expressed in the following formula: ; In the formula, The matrix consisting of vectors representing potential spending patterns, with dimension . The matrix elements are dimensionless. The basic expenditure pattern dictionary matrix, with dimensions of The matrix elements are dimensionless. Here is the activation coefficient matrix, with dimension 1. The matrix elements are dimensionless. The number of basic patterns is empirically estimated to be 8 to 12. The decomposition process minimizes the reconstruction error through iterative optimization, and the reconstruction error formula is expressed as follows: ; In the formula, The reconstruction error is dimensionless. For matrix The Frobenius norm is dimensionless; For matrix The Frobenius norm is dimensionless; the Frobenius norm The calculation formula is ,in For matrix the number of rows, For matrix The number of columns, For matrix The Line number Column elements. The basic expenditure pattern dictionary is processed using a locally linear embedding algorithm to preserve its manifold structure, generating a compact feature subset. Weight coefficients are calculated by minimizing the reconstruction error, which is expressed as follows: ; In the formula, For the first The reconstruction error of each data point is dimensionless. For the first The feature vector of each data point is dimensionless. For the first The data point pair The weight coefficients of the neighboring points are dimensionless and satisfy the constraints. ; For the first The first data point The feature vector of each of the neighborhood points is dimensionless; This is the number of neighboring points, with a default value of 5 to 10. It is the Euclidean norm.

[0077] The specific implementation of step S03 is to construct a dynamic time-warping reference template library based on historical expenditure execution data, and use the real-time expenditure execution data of the currently monitored project as the query sequence. ,in For query sequence number The expenditure amount at each point in time is expressed in ten thousand yuan. To calculate the length of the query sequence, compare the query sequence with each reference sequence in the dynamic time warping reference template library. The cumulative distance matrix, where For the reference sequence number The expenditure amount at each point in time is expressed in ten thousand yuan. The formula for calculating the elements of the cumulative distance matrix, taking the reference sequence length, is as follows: ; In the formula, The cumulative distance matrix is ​​the first Line number Column elements, dimensionless; For query sequence number The expenditure amount is expressed in ten thousand yuan; For the reference sequence number The expenditure amount is expressed in ten thousand yuan; To query the maximum expenditure amount in the sequence, in ten thousand yuan; This represents the maximum expenditure amount in the reference series, expressed in ten thousand yuan. The first cumulative distance matrix Line number Column elements, dimensionless; The first cumulative distance matrix Line number Column elements, dimensionless; The first cumulative distance matrix Line number Column elements, dimensionless; initial conditions are: A dynamic programming algorithm is used to find the optimal time curvature path in the cumulative distance matrix to achieve nonlinear time alignment. The execution mode category of the current monitored project is then identified based on the optimal time curvature path.

[0078] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0079] The specific implementation of step S05 is as follows: all CUDA streams of the graphics processing unit are started, and a first-level thread block is allocated in each CUDA stream to run the cumulative sum control graph calculation task. The formula for calculating the cumulative sum statistics is expressed as follows: ; In the formula, For the first The accumulation and statistics of time points are dimensionless. For the first The accumulation and statistics of time points are dimensionless. For the first Real-time expenditure execution data at any given moment, in ten thousand yuan; The benchmark value for the execution mode category is in ten thousand yuan. The standard deviation is the historical deviation, expressed in ten thousand yuan; the initial conditions are: The second-level thread block is allocated to run the task of calculating the Hotelling T-squared statistic. The formula for calculating the Hotelling T-squared statistic is as follows: ; In the formula, The T-squared statistic is the Hotelling T-squared statistic, which is dimensionless. For the standardized multidimensional observation vector, its first... Each component is Dimensionless For the original observation vector, the first... Each component, with units determined based on specific indicators. For the first The standard deviation of each variable, in units of same; This is the standardized mean vector, and the vector elements are dimensionless. This is the inverse of the standardized covariance matrix, and its elements are dimensionless. This is a deviation vector, with dimensionless elements; superscript... This represents the matrix transpose. The third-level thread block is allocated to run the wavelet multi-scale decomposition calculation task. The wavelet coefficient calculation formula is as follows: ; In the formula, For the first The first scale The wavelet coefficients at each position are dimensionless. For the first Energy at a scale, measured in tens of thousands of yuan. moon; This is a time series of real-time expenditure execution data, in ten thousand yuan. For wavelet basis functions, the unit is months. ; For time variables, the unit is months; integral The unit is 10,000 yuan. moon Data contention between thread blocks is avoided by allocating independent memory regions for the first-level thread block, the second-level thread block, and the third-level thread block.

[0080] The specific implementation of step S06 is as follows: for real-time funding execution data where the cumulative sum statistic exceeds the cumulative sum control limit or the Hotelling T-squared statistic exceeds the Hotelling control limit, the sequential probability ratio test method is used to calculate the anomaly confidence level. The likelihood ratio calculation formula is expressed as follows: ; In the formula, For the first The cumulative log-likelihood ratio at time step 1 is dimensionless. Under the anomalous assumption, the first The probability density function of a number of observations, in units of ten thousand yuan. ; Under normal assumptions, the first The probability density function of a number of observations, in units of ten thousand yuan. ; For the first Individual observations, in units of ten thousand yuan; It is a function of the natural logarithm. The formula for calculating the anomaly confidence level is as follows: ; In the formula, For the first The anomaly confidence level at time 1 is dimensionless and ranges from 0 to 1. It is an exponential function. When the anomaly confidence level exceeds the judgment threshold... The system triggers an early warning mechanism and records abnormal pattern characteristics, among which... The threshold is dimensionless and has an empirical value of 0.85 to 0.95. When the anomaly confidence does not exceed the threshold, the real-time funding execution data is added to the training sample set of the normal behavior baseline model.

[0081] The specific implementation of step S07 is as follows: Multiple candidate complete datasets are generated from real-time funding execution data and historical progress record data containing missing values ​​using a multiple imputation method. The expectation-maximization algorithm is then used to estimate parameters and calculate confidence intervals on these candidate complete datasets. Robust M-estimation is applied to noisy outlier observations to reduce their impact on model parameters. The formula for calculating the weight function of the robust M-estimation is as follows: ; In the formula, For the first The influence weight of each observation is dimensionless. The derivative of the loss function is dimensionless. For the first The residuals of each observation are expressed in ten thousand yuan. The standard deviation of the residuals is expressed in ten thousand yuan. The residuals are standardized and dimensionless. The optimal estimate is output by fusing the model predictions and noisy observations using Kalman filtering. The Kalman gain calculation formula is as follows: ; In the formula, For the first The Kalman gain matrix at time t, where the matrix elements are dimensionless; This is the prior estimation error covariance matrix, with matrix elements in units of ten thousand yuan. ; The matrix is ​​an observation matrix, and its elements are dimensionless. This is the transpose of the observation matrix; To measure the noise covariance matrix, the matrix elements are in units of ten thousand yuan. ; The matrix element unit is ten thousand yuan. The state update formula is expressed as follows: ; In the formula, For the first The posterior state estimation vector at time t, with vector elements in units of ten thousand yuan; For the first The prior state estimation vector at time t, with vector elements in units of ten thousand yuan; For the first The observation vector at time t, with vector elements in units of ten thousand yuan; The vector is used to predict observations, with vector elements in units of ten thousand yuan. To observe the residual vector, the vector elements are in units of ten thousand yuan.

[0082] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: The funding for a certain deep-sea resource exploration project involves more than two categories, including vessel leasing fees, equipment procurement fees, material consumption fees, personnel labor costs, data processing fees, and experimental analysis fees. The project cycle is 36 months, with a total budget of 85 million yuan. The technical team collected historical funding execution data and historical progress records for 45 similar deep-sea resource exploration projects completed in the past 5 years. For each project, the expenditure amounts for 27 categories and the completion status information for 12 key task nodes were recorded for each month from start to finish. A three-dimensional tensor structure with 45×36×27 dimensions and a two-dimensional matrix structure with 45×12 dimensions were constructed.

[0083] The technical team performed standardized preprocessing on the 3D tensor structure, removing records from three projects with more than 30% missing data due to financial system malfunctions, ultimately obtaining a dataset of 42 valid historical projects. The CANDECOMP-PARAFAC decomposition method was used to decompose the 3D tensor structure, setting the rank parameter to 8. After 87 iterations, the reconstruction error converged to 0.008, resulting in a 27×8 subject dimension factor matrix, a 36×8 time dimension factor matrix, and a 42×8 project dimension factor matrix. The eight column vectors of the subject dimension factor matrix were extracted as potential expenditure pattern representation vectors. Non-negative matrix decomposition was performed on the 27×8 matrix formed by these vectors, iterating for 48 times to obtain a 27×5 basic expenditure pattern dictionary and a 5×8 activation coefficient matrix. The five column vectors of the basic expenditure pattern dictionary represent centralized equipment procurement, stable personnel input, cyclical material consumption, experimental phase bursts, and comprehensive balanced allocation, respectively. The basic expenditure pattern dictionary is dimensionality reduced by using the local linear embedding algorithm. The number of neighborhood points is set to 10, and the 27-dimensional features are reduced to a compact feature subset of 35 dimensions. The compact feature subset retains 96.3% of the variance information of the original data.

[0084] The technical team extracted funding execution time series from 42 historical projects to construct a dynamic time-normalized reference template library. Each reference series contains 36 monthly cumulative expenditure data points. The real-time funding execution data of the currently monitored deep-sea resource exploration project, which has been running for 12 months, was used as the query series. The query series records the cumulative expenditure amounts for each month since the project's inception: 2.35 million yuan, 5.78 million yuan, 9.56 million yuan, 14.23 million yuan, 18.67 million yuan, 23.45 million yuan, 27.89 million yuan, 32.56 million yuan, 36.98 million yuan, 41.23 million yuan, 45.67 million yuan, and 50.12 million yuan. The cumulative distance matrix between the query sequence and 42 reference sequences in the reference template library is calculated. The matrix size is 12×36. The dynamic programming algorithm is used to find the optimal time curvature path. The cumulative distances between the current project and the 7th, 15th, 23rd, 31st and 38th projects in the reference template library are identified as the smallest, which are 127.8, 135.2, 142.6, 148.9 and 156.3 respectively. All five reference projects belong to the centralized equipment procurement mode category. The execution mode category of the current monitoring project is determined by voting to be the centralized equipment procurement mode.

[0085] The technical team constructed a multi-task learning neural network architecture. The input layer receives a 47-dimensional input vector, which is a concatenation of a 35-dimensional compact feature subset and 12-dimensional historical progress record data. The shared layer consists of two one-dimensional convolutional layers and one long short-term memory layer. The first convolutional layer uses 64 convolutional kernels to extract local patterns, and the second convolutional layer uses 128 convolutional kernels to extract high-level features. The long short-term memory layer contains 256 hidden units. The output layer's funding prediction branch outputs a predicted value for the remaining funding amount in the next 12 months, and the progress prediction branch outputs a predicted value for the project completion percentage. Training was conducted on a training set containing 42 historical projects, with a batch size of 128 and an initial learning rate of 0.001. The training converged after 200 epochs. The root mean square error of the funding prediction branch on the validation set was 1.568 million yuan, and the mean absolute error of the progress prediction branch was 3.7%.

[0086] The technical team initiated parallel processing of the anomaly detection task using 16 CUDA streams on the graphics processing unit, with each CUDA stream allocated to three layers of thread blocks. The first layer of thread blocks contained 27 threads, each handling the calculation of cumulative sums and statistics for 27 different items. A cumulative sum control limit of 5.0 was set. In the 13th month, the cumulative sum statistic for the ship charter fee item reached 5.3, exceeding the control limit and triggering a preliminary anomaly. The second layer of thread blocks calculated the Hotelling T-squared statistic, setting a Hotelling control limit of 15.0. In the 13th month, the calculated statistic was 16.8, exceeding the control limit. The third layer of thread blocks performed wavelet multi-scale decomposition on the ship charter fee expenditure sequence, using the Daubechies wavelet basis function for a three-level decomposition. The level 2 detail coefficients exhibited abrupt changes in the 13th month, with the amplitude jumping from 0.8 in the previous month to 2.3. Figure 2 As shown, the ship charter fee item showed a clear upward trend in cumulative total and statistical value in the 13th month. The cumulative total and statistical value increased from 3.1 in the 10th month to 4.2 in the 11th month, then to 4.8 in the 12th month, and finally reached 5.3 in the 13th month, exceeding the control limit.

[0087] The technical team used a sequential probability ratio test to verify the vessel charter fee data that triggered the initial anomaly. An upper threshold of 3.0 and a lower threshold of -3.0 were set. The calculated cumulative likelihood ratio was 3.5, exceeding the upper threshold. The anomaly confidence level of 3.5 divided by 3.0 equaled 1.17, exceeding the judgment threshold of 0.85, thus triggering the early warning mechanism. The recorded anomaly pattern was an overspending of the vessel charter fee item in the 13th month, a deviation from the baseline value of the execution mode category reaching 32.6%, and a sudden change in the wavelet detail coefficients occurring in the 13th month. Further investigation by the technical team revealed that the vessel charter contract in the 13th month resulted in an extension of operating days due to deteriorating sea conditions, incurring additional charter fees of 1.58 million yuan.

[0088] During data processing, the technical team discovered missing material consumption data for months 8 and 9, and noisy outliers in the experimental analysis cost data for month 14. A multiple imputation method was used to generate five candidate complete datasets. Each imputation randomly sampled data based on the statistical characteristics of material consumption costs from other months to fill in the missing values. The imputed values ​​for material consumption costs in month 8 in the five candidate complete datasets were 873,000 yuan, 916,000 yuan, 852,000 yuan, 938,000 yuan, and 895,000 yuan, respectively. The expectation-maximization algorithm was applied to the five candidate complete datasets to estimate model parameters. The algorithm converged after 20 iterations, and the average of the five parameter estimates was used as the final parameter estimate. The 95% confidence interval was calculated to be 856,000 to 942,000 yuan. For the noisy outlier of 2.36 million yuan in the experimental analysis cost for month 14, robust M-estimation was used to reduce its impact weight. The absolute value of the residual of this outlier was 780,000 yuan, exceeding the threshold of 1.345 times the standard deviation of 420,000 yuan. A linear loss function was used to calculate the impact weight, reducing it from 1.0 to 0.35. Figure 3 As shown, during the process of fusing the model prediction and the noisy observations using Kalman filtering, the observation value of 2.36 million yuan in the 14th month deviated significantly from the model prediction value of 1.58 million yuan. The Kalman gain was automatically adjusted to 0.28, and the optimal estimate was 1.79 million yuan, effectively suppressing the interference of the noisy observations.

[0089] Based on the results output by the monitoring model, the technical team generated a project health assessment report. The report showed that the project was generally operating normally under the centralized equipment procurement model, but the vessel leasing fee item showed an abnormal risk of overspending in the 13th month. It recommended optimizing the work plan to avoid vessel operations during periods of deteriorating sea conditions. The remaining budget for the next 12 months was predicted to be 34.88 million yuan, and the project completion rate was expected to reach 68.5%. As shown in Table 1, the anomaly detection results of the monitoring model for 27 items showed that the cumulative sums and statistics of the vessel leasing fee, equipment procurement fee, and material consumption fee were relatively high, at 5.3, 4.6, and 4.2 respectively. The Hotelling T-squared statistic was 16.8, exceeding the control limit. The anomaly confidence level of the sequential probability ratio test was 1.17, and the number of items triggering the early warning mechanism was 1.

[0090] Table 1. Statistical Table of Subject Anomaly Detection Results

[0091] This invention represents a significant advancement over traditional methods for monitoring marine research project funding. Traditional methods rely heavily on manual review by financial personnel and simple threshold comparisons, which struggle to handle high-dimensional, sparse funding data and fail to capture patterns of correlation between different items. This results in severe monitoring lag, often only revealing problems after budget overruns. This invention utilizes tensor decomposition technology to reduce high-dimensional funding data into a compact feature representation, fully leveraging the multi-linear coupling relationships across item, time, and project dimensions. The basic expenditure pattern dictionary extracted through non-negative matrix decomposition is interpretable, facilitating the identification of typical funding usage patterns for different project types. The dynamic time warping algorithm achieves non-linear time alignment, overcoming the limitations of traditional linear matching methods in handling differences in execution pace. It accurately identifies the execution pattern category of the current project, providing a personalized reference benchmark for subsequent anomaly detection. The multi-task learning neural network architecture captures the intrinsic correlation between funding and schedule through shared layers. Compared to traditional methods that independently predict funding and schedule, the joint modeling mechanism improves prediction accuracy and reveals the coupling patterns between the two. The anomaly detection mechanism based on a CUDA multi-layer thread block parallel architecture simultaneously runs three methods: cumulative sum control chart, Hotelling's T-squared statistic, and wavelet multi-scale decomposition. The cumulative sum method continuously accumulates small deviations, amplifying the statistical evidence of early anomalies. Hotelling's T-squared statistic enables joint monitoring of multi-dimensional indicators, avoiding the limitations of univariate methods. Wavelet multi-scale decomposition extracts signal features at different frequency scales, effectively identifying weak anomaly patterns submerged in noise. Compared with traditional single-threshold alarm methods, this significantly improves the sensitivity and reliability of early anomaly detection. The sequential probability ratio test minimizes detection latency under a fixed false alarm rate constraint, avoiding the dilemma that increasing sensitivity inevitably increases false alarms in traditional methods. The data quality assurance mechanism that integrates multiple imputation and robust M-estimation effectively handles missing and noisy data. Traditional methods, such as directly deleting missing samples or simple imputation, introduce bias. This invention quantifies uncertainty by generating multiple alternative complete datasets and combines robust estimation to reduce the impact of outliers, ensuring the stability and reliability of the model under conditions of suboptimal data quality.

[0092] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0093] Table 2. Variable Explanation Table (Part 1)

[0094] Table 3. Variable Explanation Table (Part Two)

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a monitoring model for the execution progress of marine scientific research project funding, characterized in that, Historical funding execution data and historical progress records of marine scientific research projects are collected. The historical funding execution data is organized into a three-dimensional tensor structure according to the dimensions of subject category, time, and project. The CP decomposition method is used to extract low-rank factor matrix groups from the three-dimensional tensor structure. Non-negative matrix decomposition is then performed on the low-rank factor matrix groups to obtain a basic expenditure pattern dictionary. A compact feature subset is generated using a local linear embedding algorithm. A dynamic time warping reference template library is constructed based on the historical funding execution data. The cumulative distance matrix between the real-time funding execution data of the currently monitored project and each reference sequence in the dynamic time warping reference template library is calculated. A dynamic programming algorithm is used to find the optimal time warping path to identify the execution pattern category. A multi-task learning neural network architecture is constructed to extract spatiotemporal feature representations and output funding and progress predictions. All CUDA streams of the graphics processing unit are allocated to three-layer thread blocks to run the cumulative and control chart calculation tasks, as well as the Hotelling task. The tasks of calculating the squared statistic and wavelet multi-scale decomposition are carried out. The sequential probability ratio test method is used to calculate the anomaly confidence to trigger the early warning mechanism. For data with missing values, multiple imputation method is used to generate multiple candidate complete datasets. For noisy anomaly observations, robust M estimation is used to reduce the influence weight. The optimal estimation result is output by fusing the model prediction value and noisy observation value through Kalman filtering.

2. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 1, characterized in that, The three-dimensional tensor structure refers to a third-order array formed by arranging historical expenditure data according to the subject category dimension, time dimension, and project dimension. Each element of the third-order array represents the expenditure amount of the project in the project dimension at a certain point in time in the time dimension, and the subject in the subject category dimension.

3. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 2, characterized in that, The optimal time curvature path refers to the path from the top left corner to the bottom right corner in the cumulative distance matrix that minimizes the cumulative distance. Each point on the optimal time curvature path represents the time alignment relationship between the corresponding positions in the query sequence and the reference sequence.

4. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 3, characterized in that, The multi-task learning neural network architecture specifically sets up shared convolutional layers and long short-term memory layers at the bottom layer to extract spatiotemporal feature representations, sets up funding prediction branches and progress prediction branches at the top layer, and uses the CUDA parallel computing framework to accelerate the training process.

5. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 4, characterized in that, The first-level thread block continuously accumulates the deviation of real-time funding execution data from the execution mode category baseline value to obtain a cumulative sum statistic. When the cumulative sum statistic exceeds the cumulative sum control limit, an alarm is triggered.

6. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 5, characterized in that, The second-layer thread block comprehensively monitors real-time funding execution data and historical progress record data, calculating multi-dimensional indicators. (Hotlin) Square statistic, Hotelling The squared statistic measures the squared Mahalanobis distance between a multidimensional observation vector and the mean vector.

7. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 6, characterized in that, The third-layer thread block performs wavelet multi-scale decomposition on real-time funding execution data to extract weak anomaly features. Data competition between thread blocks is avoided by allocating independent memory regions for the first, second, and third-layer thread blocks.

8. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 7, characterized in that, The sequential probability ratio test method calculates the likelihood ratio point by point in the observation sequence and compares it with a preset threshold. When the likelihood ratio exceeds the upper threshold, it is determined that there is an anomaly, and when the likelihood ratio is lower than the lower threshold, it is determined that it is normal.

9. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 8, characterized in that, When the anomaly confidence level exceeds the judgment threshold, an early warning mechanism is triggered and the anomaly pattern characteristics are recorded. When the anomaly confidence level does not exceed the judgment threshold, the real-time funding execution data is added to the training sample set of the normal behavior baseline model.

10. The method for constructing a monitoring model for the execution progress of marine scientific research project funding according to claim 9, characterized in that, The local linear embedding algorithm specifically involves finding the nearest neighbor set for each data point in the basic expenditure pattern dictionary, calculating the weight coefficients of the data points linearly represented by the neighboring points, and finding embedding coordinates in the low-dimensional space that maintain the same weight relationship as a compact feature subset.

Images