An intelligent weighted fusion and dynamic threshold early warning method for project implementation

Abstract

The application discloses a kind of intelligent weighted fusion and dynamic threshold early warning method for project implementation.For the problems of insufficient project feature representation capability and static early warning threshold easily causing early warning shock in prior art, the application proposes an algorithm architecture that integrates multi-scale feature analysis and mathematical statistics dynamic threshold control.In particular, the application first constructs a feature sequence, and extracts multi-scale features in parallel input deep separable convolution, selective state space and causal self-attention branch, generates an abnormal score through gated network weighting and combining evolutionary difference items;Then, the dynamic starting threshold is solved by fitting the residual error with Pareto distribution using extreme value theory, and the state comparison is completed by combining the double delay decision boundary to output the early warning signal.The application significantly improves the capture ability of implicit anomalies, effectively overcomes early warning shock, and provides a high-precision closed-loop early warning and supervision scheme for project implementation.

Description

Technical Field

[0001] This relates to the fields of artificial intelligence and data processing technology, specifically to an intelligent weighted fusion and dynamic threshold early warning method for project implementation. Background Technology

[0002] With the profound evolution of information technology and artificial intelligence, intelligent supervision of project implementation has become a key means to improve supervision service capabilities and implementation efficiency. Currently, in scenarios involving multi-dimensional indicator anomaly monitoring and early warning around the project implementation process, static rule matching and shallow machine learning models are typically used for discretized hard threshold determination.

[0003] However, existing anomaly early warning technologies still have significant technical shortcomings: First, the data generated during project implementation is characterized by multi-source heterogeneity, multi-scale temporal evolution, and high nonlinearity. Traditional methods struggle to effectively decouple and fuse local micro-mutations, long-cycle low-frequency evolution trends, and cross-dimensional task correlations, resulting in insufficient representation capacity for extracting latent anomaly features. Second, since project implementation data is often accompanied by unpredictable fluctuations and high-dimensional noise interference, traditional static early warning mechanisms based on fixed absolute values ​​cannot adapt to the dynamic drift of data manifold distribution. When facing boundary states, they are prone to frequent oscillations in the early warning state due to weak residual fluctuations. Furthermore, it is difficult to accurately control the expected false alarm frequency based on strict mathematical statistical distribution, severely weakening the reliability and accuracy of the closed-loop control of intelligent project supervision.

[0004] Therefore, there is an urgent need to propose a novel algorithm architecture that integrates deep learning and mathematical statistics to solve the technical challenges of deep fusion of multidimensional features and adaptive control of early warning thresholds. Summary of the Invention

[0005] To address the technical problems of single feature representation and static thresholds easily causing state oscillations in existing project monitoring and early warning systems, this invention provides an intelligent weighted fusion and dynamic threshold early warning method for project implementation. This method processes multi-dimensional operational feature streams based on dynamic normalization and local temporal reconstruction to generate feature input sequences. These sequences are then fed in parallel into depthwise separable convolution, selective state space, and causal self-attention branches to extract local mutations, long-term evolution, and cross-channel coupling features, respectively. A gated residual network is used to perform weighted aggregation on the concatenated multi-dimensional features to generate a basic fusion representation. The Jensen-Shannon divergence of the current spatially correlated distribution is calculated to obtain a dynamic evolutionary difference term, which is then superimposed with the basic fusion representation to map the anomaly score at the current time step. Extreme value theory is used to perform distribution fitting on the super-threshold residuals of historical stationary phases, and a dynamic initiation threshold is solved inversely by combining false alarm rate constraints. Based on the dynamic initiation threshold, a dual-hysteresis decision boundary including a recovery lower boundary is constructed, and a minimum hold window is used to perform cross-step state judgment on the anomaly score, outputting a closed-loop early warning control signal.

[0006] The present invention provides an intelligent weighted fusion and dynamic threshold early warning method for project implementation, comprising the following steps: S1: Obtain multi-dimensional operational feature streams based on a global discrete-time benchmark, perform normalization processing and time series reconstruction, and construct feature input sequences; wherein, the multi-dimensional operational feature streams include at least progress flow features, quality risk features, and resource consumption features; S2: Input the feature input sequence in parallel into depthwise separable convolution, selective state space and causal self-attention branch, and extract local mutation features, long-term evolution features and cross-channel coupling feature tensors in the multidimensional running feature flow respectively; S3: Channel concatenation is performed on the feature tensors output from the multiple feature extraction branches, and dynamic weighted aggregation is performed using an attention mechanism to output a basic fusion representation; S4: Calculate the evolutionary difference term between the current feature distribution and the historical smooth distribution, and superimpose and map it with the basic fusion representation to output the anomaly score at the current time step; S5: Extract the tail distribution features of the abnormal scores in the historical stable period, perform extreme value fitting, and combine the preset false alarm frequency constraints to solve the dynamic start threshold for the current period. S6: Construct a decision boundary including the recovery lower boundary based on the dynamic start threshold, compare the anomaly score of the current time step with the decision boundary across the step state, and output an early warning signal for project implementation closed-loop control; According to a specific implementation of an embodiment of the present invention, the specific steps of S1 are as follows: Based on global discrete time benchmark Synchronous resampling and concurrent reception of multidimensional running feature streams to obtain the original multidimensional state vector. ,in This represents the total number of feature dimensions. The original multidimensional state vector is mapped and cleaned using a dynamic normalization mechanism based on adaptive tracking.

[0007] in, The standardized, mapped multidimensional feature vector. and These are the mean tracking vector and variance tracking vector, respectively, updated in real time using the exponential moving average algorithm. It is the minimum positive penalty constant; Set the span of the observation time window as The sliding step span is And satisfy Multidimensional features within a local time window are extracted and assembled into a local time series data matrix. :

[0008] in, The discrete order index representing the sliding window, a matrix Each column vector corresponds to a normalized feature cross section sampled continuously within the observation window; Using the projection weight matrix With position encoding matrix Generate feature input sequence :

[0009] in, The learnable linear projective weight matrix is ​​iteratively updated by the network's backpropagation algorithm. The generated feature input sequence is based on the absolute position matrix generated by mapping the relative time step within the observation window to the latent space channel dimension. Synchronously transmitted to the subsequent feature parsing and extraction branch; According to a specific implementation of an embodiment of the present invention, the specific steps of S2 are as follows: For the feature input sequence Perform parallel analysis of multi-scale spatiotemporal features; S21. Depthwise separable convolution branch: Apply depthwise separable convolution to the feature input sequence along the time dimension to generate a local high-frequency mutation feature tensor. :

[0010] in, For discrete convolution operators, The depthwise convolution kernel matrix, and These are the point-by-point mapping weight matrix and the linear bias term, respectively. The activation function for the Gaussian error linear unit; S22. Selective state-space branch: Extracting low-frequency evolution feature tensors under long-period observations using a selective state-space model with discretized mapping. :

[0011] in, This is a cache variable for the hidden state of the previous truncation step. and These are the evolution matrix and injection matrix, obtained by discretizing the continuous-time state master matrix and input mapping matrix, respectively. Output matrix for state observation; S23, Causal Self-Attention Branch: Utilizes a local self-attention mechanism with causal masking to calculate cross-channel correlation dependency weights and outputs a coupled feature tensor. :

[0012] in, , and They are respectively from The query tensor, key tensor, and value tensor generated by linear projection. The transpose operator for the key tensor. As the feature scaling factor, This is a causal mask matrix with lower triangular constraint characteristics. It is a normalized exponential function; According to a specific implementation of an embodiment of the present invention, the specific steps of S3 are as follows: S31. Obtain the tensors output from the three branches, perform hard concatenation and recombination along the feature channel dimension, and construct a global joint feature tensor. :

[0013] in, For tensor splicing operators along the feature channel dimension; S32. For the global joint feature tensor Perform linear affine transformations and nonlinear mappings to compute channel-level attention gating matrices. :

[0014] in, The first linear mapping weight matrix, For bias vectors, It is a non-linear activation function; S33, Attention-based gating matrix With residual network structure, for global joint feature tensor Implement dynamic weighting and output the basic fusion representation of this truncation step. :

[0015] in, For the Hadama product operator, The second linear mapping weight matrix, To and A uniform matrix with consistent dimensions; According to a specific implementation of an embodiment of the present invention, the specific steps of S4 are as follows: Extract the local attention weights of the causal self-attention branch output at the current truncation step, perform pooling aggregation along the feature dimension, and obtain the current spatial correlation distribution tensor. The historical smoothed correlation distribution tensor is read and updated synchronously using the exponential moving average mechanism. Based on a pre-defined uniformly distributed prior pair and Perform regularization mappings to generate anti-singularity smoothing tensors. and ; Calculate the anti-singularity smoothing tensor and The Jensen-Shannon divergence is used to output the dynamic evolutionary difference term at the current truncation step. :

[0016] in, For the Kolbek-Leibler divergence operator, for and The arithmetic mean mixed tensor; S42, Extract the basic fusion representation from step 1 of S33. Based on the linear mapping parameters and the dynamic evolution difference term Perform overlay fusion and nonlinear activation mapping to calculate the anomaly score for the current time step. :

[0017] in, The weight matrix for performing feature dimensionality reduction projection, To control the scaling penalty constant for the joint contribution magnitude of evolutionary difference terms, For biased scalars, To map the results to Activation function for standard intervals; According to a specific implementation of an embodiment of the present invention, the specific steps of S5 are as follows: S51. Extract abnormal scoring sequences that are in a historically stable operating phase to construct a benchmark set, determine the basic baseline based on preset quantiles, and filter out the scoring calculation conditions exceeding the threshold residuals that meet the constraints:

[0018] in, This is the basic baseline. greater than Historical anomaly scores For conditional threshold residuals; S52. Based on the peak threshold model, The tail distribution is fitted with a generalized Pareto distribution to construct the cumulative distribution function. :

[0019] in, The shape parameter of the generalized Pareto distribution. For the corresponding scale parameters; S53. Obtain the preset expected false alarm index. Exceeding the threshold compared to historical benchmarks Perform inverse analytical mapping calculation on the survival function of the generalized Pareto distribution, and output the dynamic start threshold for the current observation period. :

[0020] in, This is the dynamic start threshold for the current observation period. The historical over-threshold frequency corresponding to the aforementioned baseline. The preset expected false alarm frequency; According to a specific implementation of an embodiment of the present invention, the specific steps of S6 are as follows: S61. Obtain the preset hysteresis offset scalar. Regarding the dynamic start threshold Applying a linear bias to calculate and recover the lower boundary To construct a dual-hysteresis decision boundary:

[0021] in, Used to trigger the state reset mechanism Used to constrain state switching sensitivity; S62, based on and The system constructs a dual-hysteresis decision boundary and introduces a minimum hold window parameter to limit the shortest maintenance period of the early warning state. Anomaly scoring at the current time step Perform a step-by-step state comparison and calculate the discretized state warning signal for this truncation step. :

[0022] in, The alert is now active. To reset the closed-loop state, This is a cache warning signal from the previous truncation step. To meet The index of the most recent historical discrete time step of the decision condition. For logical AND.

[0023] The beneficial effects of this invention are as follows: First, it significantly enhances the ability to deeply capture and represent latent anomalies during project implementation. Addressing the multi-source, heterogeneous, and highly nonlinear characteristics of project operational data, this invention abandons traditional shallow or single-feature extraction methods. By deploying depthwise separable convolutions, selective state spaces, and causal self-attention branches in parallel, it performs multi-scale spatiotemporal analysis of multi-dimensional operational flows such as progress, quality, and resources. Combined with the dynamic weighting of attention-gated networks and the calculation of spatial evolutionary differences in Jensen-Shannon divergence, the algorithm can accurately decouple local high-frequency mutations from low-frequency long-term evolutionary trends, effectively filling the capacity gap in the representation of complex coupled features in existing technologies.

[0024] Secondly, it achieves adaptive threshold control under data manifold drift. Addressing the vulnerability of traditional static hard thresholds to noise interference, this invention utilizes a generalized Pareto distribution to fit the tail distribution of the threshold-exceeding residuals during historical stationary periods to dynamically solve for the activation threshold. More importantly, this invention constructs a dual-hysteresis decision boundary including a recovery lower boundary and performs step-by-step state comparisons in conjunction with a minimal hold window parameter. This mechanism endows the system with strong resistance to weak fluctuations, ensuring that the warning state does not frequently flip due to data jitter when crossing the boundary, thus guaranteeing the stable output of the warning signal.

[0025] Third, it achieves quantitative control of false alarm rate based on rigorous mathematical and statistical constraints, significantly improving the reliability of closed-loop project monitoring. Unlike traditional early warning systems that rely on repeated parameter tuning based on human experience, this invention directly incorporates the expected false alarm indicators into the algorithm architecture, deriving dynamic thresholds by performing inverse analytical mapping on the survival function. This inverse solution mechanism based on statistical priors provides a precise mathematical explanation for adjusting the early warning sensitivity, thus providing a high-confidence decision-making basis for subsequent automatically triggered closed-loop control actions. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method; Figure 2 This is the architecture diagram of this method. Detailed Implementation

[0027] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to examples and accompanying drawings.

[0028] As attached Figure 1 and attached Figure 2 As shown, a smart weighted fusion and dynamic threshold early warning method for project implementation includes the following steps: Step 1: Obtain multidimensional running feature stream based on global discrete time benchmark, perform normalization processing and temporal reconstruction, and construct feature input sequence; Through the API interface and log collection probe, based on the set discrete time benchmark. The specific sampling period is set to 5 minutes, and multi-dimensional running feature streams are received synchronously; the state vector formed by the feature streams The business data covers three dimensions: the progress flow characteristics specifically include the dwell time of R&D tasks at specific nodes and the code submission delay rate; the quality risk characteristics specifically include the abnormal error rate of microservice interfaces and the defect trigger frequency of test cases; and the resource consumption characteristics specifically include the CPU load rate of core servers and the working hours saturation of developers. An adaptive dynamic normalization mechanism based on exponential moving average is used to map and clean the original multidimensional state vector:

[0029] in, The standardized, mapped multidimensional feature vector. and These are the mean tracking vector and variance tracking vector, respectively, updated in real time using the exponential moving average algorithm. Minimum positive penalty constant; set observation time window (Specifically, extract the sliding data from the past 24 hours) and the step size. And satisfy The continuously cleaned features are assembled into a local temporal data matrix. :

[0030] in, The discrete order index representing the sliding window, a matrix Each column vector corresponds to a normalized feature cross section sampled continuously within the observation window; By introducing a learnable projection matrix updated by backpropagation and an absolute position encoding, a feature input sequence is generated. :

[0031] in, The learnable linear projective weight matrix is ​​iteratively updated by the network's backpropagation algorithm. The absolute position matrix is ​​generated based on the mapping between the relative time step within the observation window and the dimension of the latent space channel; Step 2: Input the feature input sequence into depthwise separable convolution, selective state space and causal self-attention branch in parallel, and extract local mutation features, long-term evolution features and cross-channel coupling feature tensors in the multidimensional running feature flow respectively; The feature input sequence Multi-scale spatiotemporal feature parallel analysis specifically includes: Depthwise separable convolution branch: Apply depthwise separable convolution along the time dimension to the feature input sequence to generate a local mutation feature tensor representing high-frequency events of abnormal retry in the business module. :

[0032] in, For discrete convolution operators, The depthwise convolution kernel matrix, and These are the point-by-point mapping weight matrix and the linear bias term, respectively. For activation functions; Selective State-Space Branch: Utilizing a selective state-space model based on discretized mapping, a feature tensor representing the slow-delayed, low-frequency evolution of the task is established. :

[0033] in, This is a cache variable for the hidden state of the previous truncation step. and These are the evolution matrix and injection matrix, obtained by discretizing the continuous-time state master matrix and input mapping matrix, respectively. Output matrix for state observation; Causal self-attention branch: Utilizing an attention mechanism with a lower triangular causal mask, it calculates the interdependence weights between different feature channels of resource load and quality defects, and outputs a coupled feature tensor. :

[0034] in, , and They are respectively from The query tensor, key tensor, and value tensor generated by linear projection. For the transpose operator, As the feature scaling factor, It is a causal mask matrix with lower triangular constraint characteristics; Step 3: Perform channel concatenation on the feature tensors output from the multiple feature extraction branches, and use an attention mechanism to perform dynamic weighted aggregation to output the basic fusion representation; The heterogeneous features output from the three branches are hard-joined along the channel dimension to construct a global joint feature tensor. :

[0035] in, Tensor splicing operator; The channel-level attention gating matrix is ​​calculated using the first linear mapping and the nonlinear activation function. :

[0036] in, The first linear mapping weight matrix, It is the bias vector; Based on the attention gating matrix Dynamic weighting is implemented with the residual network structure:

[0037] in, For the Hadama product operator, The second linear mapping weight matrix, It is a one-dimensional matrix; the dynamically weighted output contains the basic fusion representation of high-risk feature components. ; Step 4: Calculate the evolutionary difference term between the current feature distribution and the historical smooth distribution, and overlay and map it with the basic fusion representation to output the anomaly score at the current time step; Extract the local attention weights of the causal self-attention branch output at the current truncation step, perform pooling aggregation along the feature dimension, and obtain the current spatial correlation distribution tensor. Distribution tensor with historical smooth association Based on a pre-defined uniformly distributed prior pair and Perform regularization mappings to generate anti-singularity smoothing tensors. and ; Calculate the anti-singularity smoothing tensor and The Jensen-Shannon divergence as a dynamic evolutionary difference term :

[0038] in, For the Kolbek-Leibler divergence operator, for and The arithmetic mean mixed tensor; The evolutionary difference term is superimposed as a penalty metric into the fused representation. After fully connected dimensionality reduction and Sigmoid activation, the risk mapping interval is calculated as follows: Abnormal scoring :

[0039] in, The weight matrix for performing feature dimensionality reduction projection, Let be the penalty constant. It is a biased scalar; Step 5: Extract the tail distribution features of the abnormal scores in the historical stable period, perform extreme value fitting, and combine them with the preset false alarm frequency constraint to solve the dynamic start threshold for the current period. Extracting outlier scoring sequences from historically stable operating phases to construct a benchmark set, and determining the baseline based on preset quantiles. Based on the peak threshold exceedance model in extreme value theory, the conditional threshold exceedance residual is calculated. ,in To score historical scalar anomalies above the baseline, a generalized Pareto distribution is fitted to the residual tails that meet the conditions, and a cumulative distribution function is constructed. :

[0040] in, For shape parameters, For scale parameters; The preset expected false alarm frequency Set to once a month, based on historical frequency of exceeding the threshold. Using the Pareto parameter and reverse analysis through the survival function, the dynamic start-up threshold is calculated. :

[0041] in, This is the dynamic start threshold for the current observation period; Step 6: Construct a decision boundary including the recovery lower boundary based on the dynamic start threshold, compare the anomaly score of the current time step with the decision boundary across the step state, and output an early warning signal for project implementation closed-loop control; Get Hysteresis Offset Scalar Calculate the recovery of the lower boundary :

[0042] in, Used to trigger the state reset mechanism Used to constrain state switching sensitivity; Introduce a minimum hold window parameter set to a minimum of 3 time steps. Execute the step state decision:

[0043] in, The alert is now active. To reset the closed-loop state, This is a cache warning signal from the previous truncation step. To meet The index of the most recent historical discrete time step of the decision condition. For logical AND; Discrete warning signal output Closed-loop control actions are triggered via the communication link. These actions include intercepting code merging requests, circuit breaking microservice interfaces, and sending alert emails.

Claims

1. A method for intelligent weighted fusion and dynamic threshold early warning for project implementation, characterized in that... Includes the following steps: S1: Obtain multi-dimensional operational feature streams based on a global discrete-time benchmark, perform normalization processing and time series reconstruction, and construct feature input sequences; wherein, the multi-dimensional operational feature streams include at least progress flow features, quality risk features, and resource consumption features; S2: Input the feature input sequence in parallel into depthwise separable convolution, selective state space and causal self-attention branch, and extract local mutation features, long-term evolution features and cross-channel coupling feature tensors in the multi-dimensional running feature flow respectively; S3: Channel concatenation is performed on the feature tensors output from multiple feature extraction branches, and dynamic weighted aggregation is performed using an attention mechanism to output a basic fusion representation; S4: Calculate the evolutionary difference term between the current feature distribution and the historical smooth distribution, and map it to the basic fusion representation to output the anomaly score at the current time step; S5: Extract the tail distribution features of the abnormal scores in the historical stable period, perform extreme value fitting, and combine the preset false alarm frequency constraints to solve the dynamic start threshold for the current period. S6: Construct a decision boundary that includes the recovery lower boundary based on the dynamic start threshold, compare the anomaly score of the current time step with the decision boundary across the step state, and output an early warning signal for project implementation closed-loop control.

2. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method for step S1 is as follows: Based on global discrete time benchmark Synchronous resampling and concurrent reception of multidimensional running feature streams to obtain the original multidimensional state vector. ,in This represents the total number of feature dimensions. The original multidimensional state vector is mapped and cleaned using a dynamic normalization mechanism based on adaptive tracking. in, The standardized, mapped multidimensional feature vector. and These are the mean tracking vector and variance tracking vector, respectively, updated in real time using the exponential moving average algorithm. It is the minimum positive penalty constant; Set the span of the observation time window as The sliding step span is And satisfy Multidimensional features within a local time window are extracted and assembled into a local time series data matrix. : in, The discrete order index representing the sliding window, a matrix Each column vector corresponds to a normalized feature cross section sampled continuously within the observation window; Using the projection weight matrix With position encoding matrix Generate feature input sequence : in, The learnable linear projective weight matrix is ​​iteratively updated by the network's backpropagation algorithm. The generated feature input sequence is based on the absolute position matrix generated by mapping the relative time step within the observation window to the latent space channel dimension. The data is then transmitted synchronously to the subsequent feature parsing and extraction branch.

3. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method for step S2 is as follows: For the feature input sequence Perform parallel analysis of multi-scale spatiotemporal features; S21. Depthwise separable convolution branch: Apply depthwise separable convolution to the feature input sequence along the time dimension to generate a local high-frequency mutation feature tensor. : in, For discrete convolution operators, The depthwise convolution kernel matrix, and These are the point-by-point mapping weight matrix and the linear bias term, respectively. The activation function for the Gaussian error linear unit; S22. Selective state-space branch: Extracting low-frequency evolution feature tensors under long-period observations using a selective state-space model with discretized mapping. : in, This is a cache variable for the hidden state of the previous truncation step. and These are the evolution matrix and injection matrix, obtained by discretizing the continuous-time state master matrix and input mapping matrix, respectively. Output matrix for state observation; S23, Causal Self-Attention Branch: Utilizes a local self-attention mechanism with causal masking to calculate cross-channel correlation dependency weights and outputs a coupled feature tensor. : in, , and They are respectively from The query tensor, key tensor, and value tensor generated by linear projection. The transpose operator for the key tensor. As the feature scaling factor, This is a causal mask matrix with lower triangular constraint characteristics. It is a normalized exponential function.

4. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method in step S3 is as follows: S31. Obtain the tensors output from the three branches, perform hard concatenation and recombination along the feature channel dimension, and construct a global joint feature tensor. : in, For tensor splicing operators along the feature channel dimension; S32. For the global joint feature tensor Perform linear affine transformations and nonlinear mappings to compute channel-level attention gating matrices. : in, The first linear mapping weight matrix, For bias vectors, It is a non-linear activation function; S33, Attention-based gating matrix With residual network structure, for global joint feature tensor Implement dynamic weighting and output the basic fusion representation of this truncation step. : in, For the Hadama product operator, The second linear mapping weight matrix, To and A uniform matrix with consistent dimensions.

5. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method in step S4 is as follows: S41. Extract the local attention weights of the causal self-attention branch output at the current truncation step, perform pooling aggregation along the feature dimension, and obtain the current spatial correlation distribution tensor. The historical smoothed correlation distribution tensor is read and updated synchronously using the exponential moving average mechanism. Based on a pre-defined uniformly distributed prior pair and Perform regularization mappings to generate anti-singularity smoothing tensors. and ; Calculate the anti-singularity smoothing tensor and The Jensen-Shannon divergence is used to output the dynamic evolutionary difference term at the current truncation step. : in, For the Kolbek-Leibler divergence operator, for and The arithmetic mean mixed tensor; S42. Extract the basic fusion representation described in S33. Based on the linear mapping parameters and the dynamic evolution difference term Perform overlay fusion and nonlinear activation mapping to calculate the anomaly score for the current time step. : in, The weight matrix for performing feature dimensionality reduction projection, To control the scaling penalty constant for the joint contribution magnitude of evolutionary difference terms, For bias scalars, To map the results to Activation function for the standard interval.

6. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method in step S5 is as follows: S51. Extract abnormal scoring sequences that are in a historically stable operating phase to construct a benchmark set, determine the basic baseline based on preset quantiles, and filter out the scoring calculation conditions exceeding the threshold residuals that meet the constraints: in, This is the basic baseline. greater than Historical anomaly scores For conditional threshold residuals; S52. Based on the peak threshold model, for The tail distribution is fitted with a generalized Pareto distribution to construct the cumulative distribution function. : in, The shape parameter of the generalized Pareto distribution. For the corresponding scale parameters; S53. Obtain the preset expected false alarm index. Exceeding the historical benchmark threshold Perform inverse analytical mapping calculation on the survival function of the generalized Pareto distribution, and output the dynamic start threshold for the current observation period. : in, This is the dynamic start threshold for the current observation period. The historical over-threshold frequency corresponding to the aforementioned baseline. This represents the preset expected false alarm frequency.

7. The intelligent weighted fusion and dynamic threshold early warning method for project implementation as described in claim 1, characterized in that... The specific method in step S6 is as follows: S61. Obtain the preset hysteresis offset scalar. Regarding the dynamic start threshold Applying a linear bias to calculate and recover the lower boundary To construct a dual-hysteresis decision boundary: in, Used to trigger the state reset mechanism Used to constrain state switching sensitivity; S62, based on and The system constructs a dual-hysteresis decision boundary and introduces a minimum hold window parameter to limit the shortest maintenance period of the early warning state. Anomaly scoring at the current time step Perform a step-by-step state comparison and calculate the discretized state warning signal for this truncation step. : in, The alert is now active. To reset the closed-loop state, This is a cache warning signal from the previous truncation step. To meet The index of the most recent historical discrete time step of the decision condition. For logical AND.

Images