Internet of Things video monitoring resource scheduling method for optimizing AI model by using big data time sequence analysis

Abstract

The invention discloses an Internet of Things video monitoring resource scheduling method for optimizing an AI model through big data time sequence analysis, and relates to the technical field of crossing of Internet of Things, artificial intelligence and big data. The method comprises the following steps: firstly, acquiring multi-dimensional time sequence data of video stream transmission, equipment operation and the like, and constructing a feature set through adaptive dimensionless processing and dynamic feature screening; predicting a resource demand through an LSTM-attention mechanism model, and constructing a scheduling decision model in combination with multi-objective optimization to obtain an optimal allocation scheme; then realizing change point detection based on an improved BOLL algorithm, and dynamically adjusting a scheduling scheme; and finally, dividing a security domain by means of a TrustZone technology, and guaranteeing the scheduling security in combination with a Freivald algorithm. The method can accurately predict the resource demand, dynamically respond to the system change, balance the resource utilization rate, service stability and safety, and improve the overall performance of the monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things, artificial intelligence and big data, and in particular to resource scheduling optimization of an Internet of Things video monitoring system, and more particularly to a method for optimizing AI model performance based on big data time series analysis technology, realizing efficient allocation and dynamic adjustment of multi-dimensional resources such as video stream transmission, device operation, network transmission and video quality in a video monitoring system. BACKGROUND

[0002] The Internet of Things video monitoring system is increasingly widely used in modern security and intelligent management fields, and its scale is continuously expanding with the increase of monitoring points and the improvement of video clarity. System operation involves the coordinated scheduling of multi-dimensional resources such as video stream transmission, device operation and network bandwidth. At present, traditional resource scheduling methods are mostly based on fixed rules or simple threshold judgments, lacking deep mining and analysis of multi-dimensional time series data in the system operation process, and being difficult to capture the dynamic rules and potential correlations of resource demand changes over time. For example, video stream code rate will show periodic or burst fluctuations due to factors such as personnel density in the monitoring area and changes in light, and device CPU utilization and memory occupancy will also be dynamically adjusted with changes in video analysis task load, while traditional methods cannot accurately predict such dynamic resource demand, easily leading to unbalanced resource allocation - either video stuttering, frame loss in core monitoring areas due to insufficient resources, or waste of non-core area resources. At the same time, existing scheduling schemes mostly use static decision-making mode, which is difficult to quickly respond to sudden changes in system operation state, and has strong adjustment lag, further reducing resource utilization efficiency. In addition, the core parameters and decision logic of resource scheduling of the Internet of Things video monitoring system often face security risks such as data tampering and malicious attacks, and traditional scheduling methods lack a trusted protection mechanism for core data and operation process, making it difficult to ensure the security and reliability of the scheduling scheme. These problems together make it difficult for the existing resource scheduling of the Internet of Things video monitoring system to adapt to complex and changing operating scenarios, and it is difficult to achieve an effective balance between resource utilization, service stability and operation safety, restricting the improvement of the overall performance of the system.

[0003] Therefore, the present application proposes an Internet of Things video monitoring resource scheduling method and system using big data time series analysis to optimize AI models to solve the above problems. SUMMARY

[0004] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] In view of the above and / or problems existing in the prior art, the present application is proposed.

[0006] Therefore, the purpose of the present application is to overcome the deficiencies in the prior art, and to provide an Internet of Things video monitoring resource scheduling method for optimizing AI models using big data time series analysis, characterized by comprising the following steps:

[0007] Step S1, Internet of Things video monitoring time series data preprocessing and feature engineering construction, collecting multi-dimensional time series data of the Internet of Things video monitoring system, the multi-dimensional time series data including real-time code rate of video stream transmission dimension, CPU utilization rate and memory occupation rate of device running dimension, bandwidth occupation rate of network transmission dimension, and frame definition score and interframe difference of video quality dimension, setting data collection time interval to form an original time series data set; adopting adaptive dimensionless processing on the original time series data, using change rate-revenue normalization combined with step Sigmoid dynamic adjustment factor for indexes with trend, using logarithmic yield rate transformation combined with exponential weighted smoothing for video quality indexes; constructing feature engineering based on the preprocessed data, extracting autoregressive features, multi-scale sliding window statistical features, and Fourier transform features, combining mutual information calculation and change point detection for dynamic feature selection to form a final feature set;

[0008] Step S2, video monitoring resource demand prediction model training based on LSTM-attention mechanism, taking the final feature set formed in step S1 as input, constructing a double-layer model architecture of "LSTM feature extraction + attention weight distribution", the LSTM network adopts a three-layer stacking structure, capturing time series dependence through forget gate, input gate, cell state, and output gate, the calculation formulas of the forget gate, input gate, cell state, and output gate are respectively:

[0009]

[0010] wherein, is a weight matrix, is a bias term, is a step Sigmoid activation function, is an element-wise multiplication, is a hidden state; introducing an improved attention mechanism in the LSTM output layer, calculating the correlation score and attention weight combining the dynamic feature selection weight of step S1, training the model using a hybrid loss function, and realizing the prediction of future time step video stream code rate demand, CPU resource demand, and bandwidth demand;

[0011] Step S3, fusion of prediction and classification results of video monitoring resource scheduling decision model construction, based on the resource demand prediction results of step S2, combined with the business domain division of the Internet of Things video monitoring system to construct the resource scheduling decision matrix; design a multi-objective optimization objective function, which combines the mean square error and the timing smoothing loss, considers the total resource constraint and the individual constraint of the device group, uses the Lagrange multiplier method to convert the constraint into an unconstrained optimization problem, and solves the optimal resource allocation scheme by the alternating direction multiplier method;

[0012] Step S4, design of resource scheduling dynamic adjustment mechanism based on change point detection, based on the optimal resource allocation scheme of step S3 and real-time running data to construct a multi-dimensional monitoring index set, and use the improved BOLL band algorithm combined with variable variance detection to realize change point recognition. The moving average of the BOLL band and the standard deviation calculation formula modified by IQR are respectively:

[0013]

[0014]

[0015] Wherein, is a monitoring index, is a sliding window size, is the interquartile range; based on the change point detection result, the prediction model parameters are corrected by using the sequential updating strategy, and the resource scheduling scheme is updated by lightweight iterative optimization;

[0016] Step S5, resource scheduling security protection and trusted verification based on Tru step StZone technology, divide the system hardware resources and software modules into a secure world (TEE) and a normal world (REE), deploy the resource scheduling core parameters and the trusted verification module in the secure world, and deploy the non-sensitive data processing module in the normal world; design a bidirectional authentication data interaction architecture based on shared memory, dynamically calculate the shared memory size, and the shared memory size calculation formula is:

[0017]

[0018] Wherein, is the adjusted scheduling scheme, is the updated demand prediction result, is a monitoring index set, is a redundant space; use Freivald algorithm to perform trusted verification on resource scheduling core matrix operation, and combine memory reuse and quantization processing to adapt to low-power computing environment.

[0019] Preferably, in the step S1, when dynamically screening the features, the mutual information of each candidate feature and the resource scheduling target is calculated, and the mutual information calculation formula is:

[0020]

[0021] Meanwhile, a variable point selection algorithm based on variable variance is used to detect data structure changes, dynamically adjust feature weights, and filter features with weights higher than a threshold to form a final feature set.

[0022] Preferably, the improved attention mechanism in step S2 introduces a feature importance factor when calculating the correlation score The correlation score calculation formula is:

[0023]

[0024] wherein, is a weight parameter, is a bias term, obtained by normalizing the feature filtering weight in step S1; the attention weight calculation introduces a temperature coefficient The attention weight calculation formula is:

[0025]

[0026] By dynamically adjusting to adapt to different data patterns.

[0027] Preferably, the multi-objective optimization objective function in step S3 is:

[0028]

[0029] wherein, is the prediction time step, is the number of device groups, is the weight, is the demand prediction value, is the fairness penalty coefficient, is the coefficient of variation.

[0030] Preferably, in step S4, when sequentially updating, the mean and variance of the prediction model are updated in real time, and the mean and variance update formulas are respectively:

[0031]

[0032] When lightweight iterative optimization is performed, only the sub-problems affected by the variable points are re-solved, and the iteration parameters are adjusted using the optimal decision matrix in step S3 as the initial value.

[0033] Preferably, in step S5, when the Freivald algorithm is verified, a binary vector is randomly generated, and are calculated, respectively, if then it is determined to be credible, is a check error threshold; a linear quantization is used in a low-power consumption environment adaptation, and the linear quantization formula is: wherein, is a floating-point parameter, is a quantization bit number, is a rounding function.

[0034] The application discloses an Internet of Things video monitoring resource scheduling method using big data time sequence analysis to optimize an AI model, which has the following beneficial effects:

[0035] 1. By multi-dimensional time sequence data preprocessing and feature engineering construction, combined with a prediction model of an LSTM-attention mechanism, the dynamic change law of resource demand of a video monitoring system can be accurately captured, and accurate prediction of resource demand such as video stream code rate, CPU and bandwidth can be realized; based on the prediction result and a multi-objective optimization scheduling decision model, optimal allocation of resources can be realized under total resource constraints and individual device constraints, which not only avoids insufficient resources in core areas, but also reduces idle resources in non-core areas, and significantly improves resource utilization efficiency.

[0036] 2. Based on a variable point identification mechanism of an improved BOLL algorithm and variable variance detection, the mutation of system running state can be quickly captured, and combined with a sequential update strategy and lightweight iterative optimization, the prediction model parameters can be corrected and the scheduling scheme can be updated in real time, so that the resource scheduling can quickly respond to abnormal conditions such as device failure and network fluctuation, effectively solve the problem of adjustment lag of traditional static scheduling scheme, and guarantee the stability and continuity of system service.

[0037] 3. By dividing the security world and the ordinary world through the TrustZone technology, sensitive components such as scheduling core parameters and trusted verification modules are deployed in the security world, combined with a bidirectional authentication data interaction architecture based on shared memory, data tampering and malicious attacks can be effectively prevented; the Freivald algorithm is used for trusted verification of core matrix operation, and memory reuse and quantization processing are used to adapt to low-power consumption environment, which guarantees the core logic and data security of resource scheduling, considers the operation demand of edge devices, and improves the overall operation credibility of the system. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the drawings shown.

[0039] Figure 1 It is a kind of big data time sequence analysis optimization AI model of Internet of Things video monitoring resource scheduling method diagram;

[0040] Figure 2 The flow chart for pre-processing the Internet of Things video monitoring time series data and constructing feature engineering in step S1 of the application;

[0041] Figure 3 The flow chart for training the video monitoring resource demand prediction model based on the LSTM-attention mechanism in step S2 of the application;

[0042] Figure 4 The flow chart for constructing the video monitoring resource scheduling decision model fused with the prediction and classification results in step S3 of the application;

[0043] Figure 5 The flow chart for designing the resource scheduling dynamic adjustment mechanism based on the change point detection in step S4 of the application;

[0044] Figure 6 The flow chart for the resource scheduling security protection and trusted verification based on the Tru step StZone technology in step S5 of the application;

[0045] The purposes, functional features and advantages of the application will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0046] The technical solutions in the application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the application. It should be understood that the specific embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0047] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the application, therefore the application is not limited by the specific embodiments disclosed below.

[0048] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is the embodiment independent or selectively excluded from other embodiments.

[0049] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the embodiments of the specification.

[0050] The application proposes an Internet of Things video monitoring resource scheduling method using big data time series analysis to optimize AI model, characterized by comprising the following steps:

[0051] Step S1: Preprocess IoT video surveillance time-series data and construct feature engineering;

[0052] Step S2: Train a video surveillance resource demand prediction model based on LSTM-attention mechanism;

[0053] Step S3: Construct a video surveillance resource scheduling decision model that integrates prediction and classification results;

[0054] Step S4: Design a dynamic adjustment mechanism for resource scheduling based on change point detection;

[0055] Step S5: Resource scheduling security protection and trusted verification based on Tru step StZone technology;

[0056] As a preferred embodiment of the present invention, step S1, preprocessing IoT video surveillance time-series data and constructing feature engineering, includes the following steps: first, performing step S11, multi-dimensional time-series data acquisition and adaptive dimensionless processing, and then performing step S12, constructing time-series feature engineering and dynamic feature filtering.

[0057] In step S11, multi-dimensional time-series data acquisition and adaptive dimensionless processing, multi-dimensional time-series data is acquired through the Internet of Things video surveillance system, including core indicators such as video stream transmission, equipment operation, network transmission, and video quality. Combined with the business attributes of the monitoring scene, information such as camera location and area priority is supplemented. Differentiated acquisition priorities are set according to different areas to ensure the real-time integrity of data in core areas, and finally, an original time-series dataset containing timestamps, multi-dimensional indicators, and business attributes is formed.

[0058] Considering that the difference in the magnitude of data across different dimensions can interfere with the feature learning of the subsequent A1 model, an adaptive dimensionless strategy is introduced. For indicators with obvious trends, the traditional rate of change payoff normalization is improved by calculating the mean through a sliding window. with standard deviation A dynamic adjustment factor based on the step-wise Sigmoid function is constructed, with the following formula:

[0059]

[0060] The factor is combined with the rate of change to achieve adaptive adjustment of normalization intensity: when the data deviates from the normal range, the factor reduces the rate of change to avoid interference from extreme values; when the data fluctuates normally, the factor retains the original trend of the rate of change, thus eliminating the influence of magnitude and preserving the time series characteristics.

[0061] For the frame sharpness score and inter-frame difference in video quality, a logarithmic yield transformation combined with exponential weighted smoothing is used: first, the data is converted into relative change using the logarithmic yield, and then smoothed by the sliding window exponential weighted average. This approach both absorbs the impact of temporal fluctuations and preserves the true trend of change, providing stable basic data for subsequent feature engineering.

[0062] In step S12, the construction of time-series feature engineering and dynamic feature selection, based on the preprocessed dimensionless data, a multi-dimensional feature engineering is constructed from the perspectives of time-series correlation, periodicity, and business adaptability. First, autoregressive features are extracted, and at the same time, multi-scale sliding window statistical features are introduced, setting small, medium, and large windows, and calculating the moving average for each. Sliding standard deviation With sliding coefficient of variation The moving mean reflects the overall trend of the data, the sliding standard deviation reflects the degree of fluctuation, the sliding coefficient of variation eliminates the influence of the mean to accurately measure relative fluctuation, and the multi-scale window design can adapt to the needs of time series patterns of different durations.

[0063] To uncover the periodic characteristics of data, Fourier transform is introduced to convert the time-domain features of time-series data into frequency-domain features: Fourier transform is performed on the fixed window data of each indicator to extract the amplitude and phase of the first few major frequency components, quantify the periodic pattern, make up for the problem of insufficient periodicity of traditional statistical features, and adapt to the scenario requirements of long-term stable operation of IoT monitoring.

[0064] Based on this, a dynamic feature selection mechanism that integrates temporal characteristics and business requirements is designed. On the one hand, the correlation strength between each candidate feature and the resource scheduling target is calculated through mutual information, using the following formula:

[0065]

[0066] Strongly correlated features with mutual information values ​​exceeding a threshold are selected. On the other hand, a variable variance-based change-point selection algorithm is used to detect changes in the temporal data structure, adaptively adjusting feature weights when encountering change points to ensure the feature set dynamically adapts to changes in data patterns. Finally, the mutual information and dynamic weights are combined to select core features, forming the final feature set used for AI model training.

[0067] As a preferred embodiment of the present invention, step S2, the video surveillance resource demand prediction model based on the trained LSTM-attention mechanism, includes the following steps: first, step S21, model structure design and input-output mapping construction; then, step S22, attention mechanism fusion and weight calculation optimization; and finally, step S23, model training and hyperparameter optimization.

[0068] In step S21, model structure design and input-output mapping construction, the final feature set constructed based on step S1... This clarifies the input-output mapping relationship of the video surveillance resource demand prediction model. The input layer dimension corresponds to the feature set. Number of features The output layer needs to predict the core resource requirements for future time steps to determine the video stream bitrate requirements for the output. CPU resource requirements With bandwidth requirements This forms the output vector. .

[0069] The model adopts a two-layer architecture of "LSTM feature extraction + attention weight allocation". The bottom layer is an LSTM network with a three-layer stacked structure to enhance feature capture capability. The number of neurons from the input layer to the first hidden layer, from the first hidden layer to the second hidden layer, and from the second hidden layer to the third hidden layer are set to [values ​​to be filled in]. Furthermore, it adapts to changes in feature set dimensionality by dynamically adjusting the number of neurons in the hidden layer. The forget gate in the LSTM unit... Controlling the degree to which historical information is retained, input gate With cell state Responsible for updating the current timing information, output gate The output ratio of the current information is determined by the following formula:

[0070]

[0071] in, Here is the weight matrix for each gate. For bias terms, The step is the Sigmoid activation function. For element-wise multiplication, The hidden state of the LSTM unit can effectively capture the long-term temporal dependencies in the feature set of step S1.

[0072] In step S22, attention mechanism fusion and weight calculation optimization, the hidden state at each time step is first calculated. Association score with the predicted target Introducing feature importance factors The weights are based on the dynamic feature selection in step S1. Normalization yields the corrected score calculation:

[0073]

[0074] in, These are the weight parameters for the attention mechanism. This is a bias term.

[0075] Based on association score Calculate attention weights The step-wise Softmax function is used to achieve weight normalization, while a temperature coefficient is introduced. The formula for adjusting the concentration of the weight distribution is:

[0076]

[0077] when When the weights are reduced, the weights are more concentrated on time steps with high correlation scores; when As the weights increase, the weight distribution becomes more uniform, and the final prediction result is obtained by attention-weighted summation, as shown in the formula: This process achieves deep integration of the feature set and attention mechanism in step S1.

[0078] In step S23, model training and hyperparameter optimization, a hybrid loss function is used to train the model. This combines the "accuracy" and "stability" requirements of video surveillance resource demand prediction, fusing mean squared error and temporal smoothing loss to avoid resource scheduling imbalances caused by sudden changes in predicted values. The total loss function formula is:

[0079]

[0080] in, The number of samples in the training set. This represents the actual resource demand vector. The weights for the time-series smoothing loss are determined through cross-validation to find the optimal values. The optimizer uses the Adam algorithm with a stepped learning rate adjustment strategy, and the initial learning rate is set to... Each training session indivual The learning rate then decays back to its original value. The formula is: This strategy ensures high convergence speed in the early stages of training, while fine-tuning parameters in the later stages to reduce loss.

[0081] The model hyperparameter optimization employs nested k-fold cross-validation. The outer cross-validation layer divides the training and validation sets, while the inner cross-validation layer optimizes the hyperparameters. The mean absolute percentage error (MAPE) on the validation set is used as the evaluation metric, as shown in the formula: ,in To minimize the value, the denominator is kept to zero. Through multiple rounds of cross-validation, a combination of hyperparameters that performs stably across different data partitions is selected. This ensures that the model maintains high prediction accuracy across different time segments of the feature set in step S1, ultimately resulting in a resource demand prediction model with strong generalization ability and adaptability to IoT video surveillance scenarios.

[0082] As a preferred embodiment of the present invention, step S3, constructing a video surveillance resource scheduling decision model that integrates prediction and classification results, includes the following steps: first, step S31, designing the resource scheduling decision matrix and objective function; then, step S32, constructing and transforming multiple constraints; and finally, step S33, integrating optimization solution and dynamic adjustment mechanism.

[0083] In step S31, the design of the resource scheduling decision matrix and objective function, the resource scheduling decision matrix is ​​first constructed by classifying the devices in the IoT video surveillance system according to the priority of the monitoring area and the device type, forming... A group of monitoring devices, a decision matrix Dimension set to ,in Representing the The equipment group in the first The time step for the first The amount of resources allocated.

[0084] To achieve a balance between maximizing resource utilization and ensuring scheduling fairness, a multi-objective optimization function is designed. The first part of the objective function is a weighted sum of resource utilization rates, with weights... The second part is the fairness constraint, which uses the coefficient of variation to measure the balance of resource allocation among the equipment groups. The objective function formula is as follows:

[0085]

[0086] in, For the first The coefficient of variation of resource allocation for all equipment groups at each time step is calculated as follows: ,in, For standard deviation, For the mean, To minimize the value, avoid a denominator of zero. The optimal value of the fairness penalty coefficient is determined through resource scheduling simulation experiments on the model validation set in step S2, ensuring that the objective function can both maximize resource utilization and suppress allocation imbalance.

[0087] In step S32, the construction and transformation of multiple constraints, the first step is the total resource constraint, and the system in the... Total bitrate available per time step Total CPU resources Total bandwidth The constraints are determined by the hardware configuration and network bandwidth planning of the IoT monitoring system, and the formula is as follows:

[0088]

[0089] Secondly, there are individual constraints for each equipment group. The resource allocation for each equipment group must meet its own hardware capacity limit and minimum operating requirements. The constraint formula is as follows:

[0090]

[0091] To facilitate subsequent solution, the Lagrange multiplier method is used to transform the inequality constraints into an unconstrained optimization problem. The Lagrange multiplier is introduced: total resource constraint. Individual constraints of equipment groups Construct the Lagrange function The constraints are incorporated into the objective function.

[0092] In step S33, the integration of optimization and dynamic adjustment mechanisms, the Alternating Direction Multiplier Method (ADMM) is used to solve the transformed unconstrained optimization problem, and the high-dimensional decision matrix is... The problem is decomposed into multiple low-dimensional subproblems, and distributed solution is achieved through iterative updates to reduce computational complexity. First, the objective function is broken down by time step and device group. Each subproblem corresponds to the resource allocation of a single device group in a single time step. During iteration, dual variables and penalty parameters are introduced. To achieve information exchange between subproblems, the iterative formula is as follows:

[0093]

[0094] in, For the number of generations, Using dual variables, the optimal resource allocation scheme for each time period is obtained through multiple rounds of selection until convergence.

[0095] To adapt to the dynamic changes in time series data, a dynamic resource scheduling adjustment strategy is designed by integrating the update mechanism of the prediction model in step S2. When the model in step S2 updates its prediction results using newly acquired time series data... At that time, calculate the prediction deviation. If the deviation exceeds the threshold This triggers a re-solution of the resource scheduling scheme, and the decision matrix is ​​updated rapidly through ADMM iterative updates. This ensures that resource allocation plans are always aligned with the latest resource demand forecasts.

[0096] As a preferred embodiment of the present invention, step S4, the design of a dynamic adjustment mechanism for resource scheduling based on change point detection, includes the following steps: first, step S41, multi-dimensional change point detection and anomaly determination, and then step S42, sequential update and iterative optimization of the scheduling scheme.

[0097] In step S41, multi-dimensional change point detection and anomaly judgment, the core monitoring indicators closely related to resource scheduling are first selected, including the resource allocation deviation indicator: bitrate allocation deviation. CPU allocation deviation Bandwidth allocation deviation and system operating status indicators: video frame loss rate Device response delay This forms a set of monitoring indicators: ,in The first The equipment group in the first The actual resource consumption value at each time step.

[0098] A two-layer change detection logic is designed for each monitoring indicator. The first layer uses an improved BOLL band algorithm to calculate the dynamic threshold. In this implementation, the bitrate is allocated based on the deviation. For example, first calculate the sliding window size as... moving average Standard deviation adjusted by IQR The formula is:

[0099]

[0100]

[0101] in Interquartile range ( quantiles and The difference in quantiles is used, and IQR correction is applied to avoid interference from extreme values ​​in the standard deviation calculation. The upper limit of the BOLL band is set based on the above parameters. and lower boundary :

[0102]

[0103]

[0104] when Exceeding When the range is reached, the second layer of variable variance detection is triggered; the second layer uses a variable point selection algorithm based on variable variance to calculate the variance change of the monitoring index within the sliding window. The formula is: ,like ,in, The variance change threshold is determined by the coefficient of variation of the resource allocation scheme in step S3. If this time step is determined to be a change point, it indicates that there is a significant deviation between the resource scheduling scheme and the actual operating requirements of the system, and the dynamic adjustment process needs to be triggered.

[0105] In step S42, sequential update and iterative optimization of the scheduling scheme, based on the change point detection results of step S401, a resource scheduling adjustment mechanism integrating sequential update and lightweight solution is designed to achieve seamless connection with the resource scheduling decision model of step S300. Firstly, for abnormal indicators at change points, a sequential update strategy is used to correct the parameters of the resource demand prediction model in step S2 and the input of the decision model in step S3, adjusting the mean of the prediction model. With variance Real-time updates are performed, taking the mean of bitrate demand prediction as an example, when new change point times are added to the observed data. At that time, the update formula is:

[0106]

[0107] By sequentially updating the model, the high computational overhead of retraining is avoided, ensuring that the prediction results can be quickly adapted to the system operation mode after the change point, and providing accurate demand input for scheduling and adjustment.

[0108] Based on this, a lightweight scheduling scheme iterative optimization process is constructed. For the device groups and resource types affected by the change point, the ADMM solution process in step S3 is pruned and optimized—only the subproblems affected by the change point are resolved, rather than a full iteration. Simultaneously, a historical optimal solution guidance mechanism is introduced to guide the optimal decision matrix in step S3. As an initial value, combined with the updated resource demand forecast results Adjusting the Lagrange multipliers and dual variables, the alternative formula is corrected to:

[0109]

[0110] in The adjusted coefficient of variation (calculated only for the variance and mean of the equipment group affected by the change point) is used to output the final adjusted resource scheduling scheme. The decision matrix in step S3 is updated simultaneously to achieve closed-loop linkage of change point detection, parameter update, and scheme optimization.

[0111] As a preferred embodiment of the present invention, step S5, resource scheduling security protection and trusted verification based on Tru step StZone technology, includes the following steps: first, step S51, Tru step StZone security domain division and data interaction architecture design, and then step S52, trusted verification of resource scheduling results and quantitative environment adaptation.

[0112] In step S51 and the Tru step StZone security domain partitioning and data interaction architecture design, the resource scheduling scheme dynamically adjusted in step S4 is used as a basis. A comprehensive security protection architecture is constructed: following the Tru step StZone specification, the system hardware and software are divided into a secure world and a normal world. The secure world deploys core sensitive modules, including the core parameters of the scheduling decision model in step S3, the threshold parameters for change point detection in step S4, and the module for encrypted storage and trusted verification of the scheduling scheme. The normal world deploys non-sensitive modules, including the data acquisition and preprocessing in step S1, the feature input layer of the prediction model in step S2, and the preliminary calculation of the raw data for change point detection in step S4, forming a security domain partitioning mode of "sensitive logic isolation and non-sensitive tasks externalization".

[0113] To achieve efficient data interaction between the secure world and the normal world, a two-way authenticated shared memory communication architecture is designed. First, the resource scheduling client in the normal world initiates a shared memory request to the trusted application in the secure world. This is triggered by the SMC instruction in step S4, which initiates a world switch. The secure world's TA then adjusts the scheduling scheme according to step S4. The amount of data, dynamically calculate the shared memory size. The formula is:

[0114]

[0115] in To adjust the data volume of the scheduling scheme, The data volume of the updated resource demand forecast results in step S4 The data volume of the monitoring index set for change point detection in step S4. Redundant space reserved (for storing data verification information). Shared memory address. Dynamically allocated by the TEE kernel and access permissions set through the MMU—only the TA in the secure world and the designated CA in the normal world are allowed to read and write, while other processes are prohibited from accessing the data, ensuring the security of data interaction.

[0116] Simultaneously, semantic consistency checks are added to the interactive data: before the CA in the ordinary world writes data to shared memory, it generates a semantic description tag containing data type, timestamp, and data length; when the TA in the secure world reads data, it uses a semantic parsing function. Verify the consistency between the data and the label. If the parsing result does not match the expected semantics, refuse to read and trigger an abnormal alarm to avoid malicious data injection attacks and achieve full protection for the transmission of data to the security domain after dynamic adjustment in step S4.

[0117] In step S52, the reliable verification of resource scheduling results and the adaptation to the quantification environment, the adjusted resource scheduling scheme output in step S4 is... A reliable verification mechanism integrating matrix verification and quantization calculation is designed to ensure that the scheduling results are not tampered with and the calculation process is reliable. First, the core matrix operations in the resource scheduling decision model of step S3 are verified. Randomly generated dimensions and... Binary vectors matching column count Calculate separately: , ,like ,in, To verify the error threshold, the matrix operation result of the scheduling scheme is determined to be reliable; otherwise, the exception handling process is triggered, and the dynamic adjustment in step S4 and the optimization solution in step S3 are re-executed to ensure the integrity of the scheduling result calculation.

[0118] Considering the application scenarios of low-power computing units in IoT edge devices, a lightweight adaptation of the trusted verification mechanism is implemented. For the verification module in low-power environments, a memory reuse strategy is adopted to reduce resource consumption: two fixed-size memory blocks are allocated. and Intermediate data during the alternating storage and verification process is used, and the memory block size is calculated using the following formula:

[0119]

[0120] in This represents the total available memory for the low-power unit, avoiding fragmentation issues caused by frequent memory allocation and deallocation through memory reuse. Simultaneously, the matrix and vector parameters during the verification process are quantized using a linear quantization formula: ,in For floating-point parameters, For the number of quantization bits, The rounding function converts high-precision floating-point numbers into low-precision fixed-point numbers, achieving a balance between security verification and low power consumption requirements, and providing reliable protection for the resource scheduling scheme dynamically adjusted in step S4 across all scenarios.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for scheduling IoT video surveillance resources by optimizing AI models using big data time-series analysis, characterized in that, Includes the following steps: Step S1: Preprocessing and Feature Engineering of IoT Video Surveillance Time-Series Data. Collect multi-dimensional time-series data from the IoT video surveillance system. This multi-dimensional time-series data includes real-time bitrate (video stream transmission dimension), CPU utilization and memory usage (device operation dimension), bandwidth usage (network transmission dimension), and frame clarity score and inter-frame difference (video quality dimension). Set a data collection time interval to form the original time-series dataset. Adaptive dimensionless processing is applied to the original time-series data. For trend-oriented indicators, a rate-of-change normalization combining the Sigmoid dynamic adjustment factor is used. For video quality indicators, a logarithmic return transformation combined with exponential weighted smoothing is used. Based on the preprocessed data, feature engineering is constructed, extracting autoregressive features, multi-scale sliding window statistical features, and Fourier transform features. Dynamic feature selection is performed using mutual information calculation and change point detection to form the final feature set. Step S2: Training the video surveillance resource demand prediction model based on the L-step STM-attention mechanism. Using the final feature set formed in Step S1 as input, a two-layer model architecture of "L-step STM feature extraction + attention weight allocation" is constructed. The L-step STM network adopts a three-layer stacked structure, capturing temporal dependencies through forget gates, input gates, cell state gates, and output gates. The calculation formulas for the forget gate, input gate, cell state gate, and output gate are as follows: in, This is the weight matrix. For bias terms, The step is the Sigmoid activation function. For element-wise multiplication, The hidden state is used; an improved attention mechanism is introduced into the output layer of STM in step L. The association score and attention weight are calculated by combining the dynamic feature selection weight in step S1. The model is trained using a hybrid loss function to predict the video stream bitrate requirement, CPU resource requirement and bandwidth requirement in future time steps. Step S3: Constructing a video surveillance resource scheduling decision model by fusing prediction and classification results. Based on the resource demand prediction results of Step S2, a resource scheduling decision matrix is ​​constructed by combining the business domain division of the IoT video surveillance system. A multi-objective optimization objective function is designed, which combines mean square error and time-series smoothing loss, while considering both total resource constraints and individual constraints of equipment groups. The constraints are transformed into an unconstrained optimization problem using the Lagrange multiplier method, and the optimal resource allocation scheme is obtained by solving the problem using the alternating direction multiplier method. Step S4: Design of a dynamic resource scheduling adjustment mechanism based on change point detection. Based on the optimal resource allocation scheme from step S3 and real-time operational data, a multi-dimensional monitoring index set is constructed. An improved BOLL band algorithm combined with variable variance detection is used to identify change points. The moving mean and IQR-corrected standard deviation of the BOLL band are calculated as follows: in, For monitoring indicators, To adjust the sliding window size, The interquartile range is used; based on the change point detection results, a sequential update strategy is adopted to correct the prediction model parameters, and the resource scheduling scheme is updated through lightweight iterative optimization. Step S5: Based on Tru step StZone technology, resource scheduling security protection and trusted verification are implemented by dividing the system hardware resources and software modules into a secure world (TEE) and a normal world (REE). The secure world deploys the core resource scheduling parameters and trusted verification module, while the normal world deploys the non-sensitive data processing module. A bidirectional authentication data interaction architecture based on shared memory is designed, and the shared memory size is dynamically calculated. The formula for calculating the shared memory size is: in, To adjust the scheduling scheme, For the updated demand forecast results, For the monitoring indicator set, To provide redundant space, the Freivald algorithm is used to perform reliable verification of the core matrix operations of resource scheduling, and memory reuse and quantization processing are combined to adapt to low-power computing environments.

2. The method according to claim 1, characterized in that, In step S1, during dynamic feature filtering, the mutual information between each candidate feature and the resource scheduling target is calculated. The formula for calculating the mutual information is: Meanwhile, a variable point selection algorithm based on variable variance is used to detect changes in the data structure, dynamically adjust feature weights, and select features with weights higher than the threshold to form the final feature set.

3. The method according to claim 1, characterized in that, In step S2, the improved attention mechanism introduces a feature importance factor when calculating the association score. The formula for calculating the correlation score is: in, For weight parameters, For bias terms, The weights are obtained by normalizing the feature selection weights in step S1; the attention weight calculation incorporates a temperature coefficient. The attention weight calculation formula is: Through dynamic adjustment Adaptable to different data modes.

4. The method according to claim 1, characterized in that, The multi-objective optimization objective function in step S3 is: in, To predict the time step, For the number of equipment groups, As weight, This is the demand forecast. For fairness penalty coefficient, is the coefficient of variation.

5. The method according to claim 1, characterized in that, In step S4, during the sequential update, the mean and variance of the prediction model are updated in real time. The update formulas for the mean and variance are as follows: In lightweight iterative optimization, only the subproblems affected by the change point are resolved, and the iteration parameters are adjusted with the optimal decision matrix in step S3 as the initial value.

6. The method according to claim 1, characterized in that, In step S5, during the Freivald algorithm verification, a binary vector is randomly generated. Calculate separately and ,like Then it is deemed credible. To verify the error threshold; linear quantization is used for low-power environment adaptation, and the linear quantization formula is: ,in, For floating-point parameters, For the number of quantization bits, This is a rounding function.

Images