AI data scheduling distribution system and method based on cloud computing

Abstract

The invention discloses an AI data scheduling distribution system and method based on cloud computing, and relates to the technical field of intelligent cloud computing, and the system comprises an AI data collection module, a feature extraction module, a decision parameter generation module, a scheduling distribution module and a visual feedback module. The AI data acquisition module acquires equipment video data and temperature data, the feature extraction module extracts video temperature features, the decision parameter generation module outputs equipment vibration intensity and temperature rise rate, and the scheduling distribution module calculates equipment shutdown probability and node load score and triggers a shutdown instruction and a task migration decision. And the visual feedback module maps the equipment health state and node load distribution, and optimizes the feature weight. Through multi-modal data fusion and a closed-loop feedback mechanism, accurate early warning of equipment abnormity and dynamic optimal allocation of resources are realized, visual decision assistance and adaptive weight adjustment are combined, equipment safety and cluster load balancing efficiency in a cloud computing environment are improved, and manual intervention requirements and operation and maintenance cost are reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent cloud computing technology, and in particular to a cloud computing-based AI data scheduling and distribution system and method. Background Art

[0002] AI data scheduling and allocation refers to a technical mechanism that dynamically optimizes the allocation of data and computing resources based on AI task requirements in a distributed computing environment. Its core goal is to achieve efficient resource utilization through intelligent decision-making and ensure the efficiency and quality of task execution. The scheduling system needs to comprehensively consider task priority, data dependency, model computing volume and hardware heterogeneity, break down tasks and allocate them to appropriate nodes, while monitoring resource load for dynamic adjustment. During model training, the scheduler will divide large-scale data sets into blocks and allocate them to different GPU nodes to balance computing pressure. In inference services, computing power resources are allocated in real time based on request traffic. This technology is directly related to the scalability, cost control and response speed of the AI system, and is a key infrastructure supporting the implementation of complex AI applications.

[0003] To address the issues of dynamic resource allocation lag and cross-scenario collaborative optimization in AI data scheduling and allocation, existing technologies employ static threshold allocation based on a single data source. However, this approach fails to integrate dynamic changes in multimodal data in real time, leading to delayed resource allocation responses and imbalanced resource utilization across scenarios. To address these issues, a cloud computing-based AI data scheduling and allocation system and method are proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI data scheduling and distribution system and method based on cloud computing to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, the technical solutions adopted by the present invention are as follows: First, a cloud computing-based AI data scheduling and allocation system includes an AI data acquisition module, a feature extraction module, a decision parameter generation module, a scheduling and allocation module, and a visual feedback module;

[0006] The AI data acquisition module collects and pre-processes video data of devices in the cloud computing target scene and temperature data on the surface of the corresponding server hardware devices;

[0007] The feature extraction module extracts video features and temperature features from the preprocessed video data and temperature data;

[0008] The decision parameter generation module combines video features and temperature features to build a video state perception model and a temperature change analysis model, and outputs the device vibration intensity and the device surface temperature rise rate;

[0009] The scheduling and allocation module obtains the equipment shutdown probability and cloud computing node load score based on the equipment vibration intensity and the equipment surface temperature rise rate, and then triggers the equipment shutdown instruction and cloud computing task migration decision;

[0010] The visual feedback module maps the device health status and node load distribution through heat maps and topology maps, and performs feedback optimization to adjust feature weights.

[0011] A further improvement of the technical solution of the present invention is that in the AI data acquisition module, the acquisition and preprocessing process of the video data of the device in the cloud computing target scene and the temperature data of the surface of the corresponding server hardware device includes:

[0012] Install high-resolution camera arrays around devices in the cloud computing target scenario to cover the entire working area of the devices. Mount micro-thermocouple sensors on the surfaces of heat-generating components. Install wide-angle infrared thermal imagers on the front and back of server racks in the cloud computing target scenario to fully cover the temperature field of the heat dissipation channel. Deploy surface-mount temperature sensors on the surfaces of each server's power module and hard drive tray, installing them close to the metal casing.

[0013] The camera array captures device operation video data at a rate of 30 frames per second. The edge computing node performs key frame extraction, extracting reference frames at a time interval of Δt = 1 / 30 second. When the difference between adjacent frames exceeds Δt, dynamic frame supplementary acquisition is triggered. The micro-thermocouple sensor and the wide-angle infrared thermal imager synchronously collect temperature data. The edge computing node performs sliding window and normalization processing on the raw temperature data.

[0014] The high-resolution camera array is arranged at a vertical downward angle, 0.5-1.5 meters from the device surface, with a resolution of no less than 3840×2160 pixels and an 8-bit color depth. The spacing between the micro-thermocouple sensors is no more than 10 cm. The vertical height of the wide-angle infrared thermal imager corresponds to the position of the GPU heat sink module and the CPU heat sink, with a horizontal spacing of 1 meter. The sampling frequency of the chip temperature sensor is no less than 1000 Hz.

[0015] A non-local mean filtering algorithm is applied to the dynamic frames in the video data. The search window is set to 7×7 pixels, the similarity threshold is set to 0.8, and the discrete cosine transform is used to compress the reference frame in blocks. The compression ratio is not less than 10:1, and the low-frequency components are retained. The cubic spline interpolation method is used to fill the data points in the temperature data detection window that deviate from the mean by ±3 times the standard deviation. The video data and temperature data are synchronized with the timestamp through the NTP protocol. When the deviation exceeds 1 millisecond, the edge computing node time domain interpolation compensation is triggered.

[0016] A further improvement of the technical solution of the present invention is that: in the feature extraction module, the process of extracting video features and temperature features from the preprocessed video data and temperature data includes:

[0017] A dense optical flow algorithm is used to obtain the displacement vectors of adjacent frames of preprocessed video data pixel by pixel. Each frame is divided into a 32×32 grid. The mean of the displacement vectors in each grid is calculated to generate a motion vector field feature matrix. Based on the key frame device area, a gray-level co-occurrence matrix is constructed to obtain the texture entropy values in four directions including 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The mean of the four-directional texture entropy values is taken to generate a texture entropy feature vector.

[0018] The preprocessed temperature data were subjected to 8-layer wavelet packet decomposition using the Daubechies-4 wavelet basis function to extract the scale coefficient and detail coefficient. The scale coefficient and detail coefficient were then concatenated into a frequency domain energy eigenvector. At the same time, the mean and variance of the temperature data within a sliding window with a window length of 60 seconds and a step size of 10 seconds were statistically analyzed.

[0019] A further improvement of the technical solution of the present invention is that: in the decision parameter generation module, the video state perception model is constructed by combining video features, and the process of outputting the vibration intensity of the device includes:

[0020] The preprocessed motion vector field feature matrix and texture entropy feature vector are spatially aligned and channel-fused. The motion vector field feature matrix retains the original spatial dimension and is expanded into a two-channel tensor. The texture entropy feature vector is reshaped to the same spatial size through upsampling and used as the third channel to generate a 32×32×3 three-channel joint input tensor H1.

[0021] A video state perception model is constructed by generating a 3D convolutional network containing 6 residual blocks. The input layer of the 3D convolutional network receives a four-channel joint input tensor of 32×32×3×1, where the fourth dimension is the time step, initialized to a single frame. Each residual block of the 3D convolutional network consists of two layers of convolution kernels and skip connections. The sizes of the first and second layer residual blocks are 5×5×5 and 3×3×3, respectively. Spatiotemporal features are gradually extracted by alternating stacking convolution and pooling operations. The process of extracting spatiotemporal features includes a forward propagation stage, in which the four-channel joint input tensor is subjected to the residual block to extract the local spatiotemporal pattern, the spatial dimension is compressed and the time step is expanded by the pooling layer, and a 128-dimensional spatiotemporal fusion feature vector is generated by global average pooling.

[0022] The 128-dimensional spatiotemporal fusion feature vector is input into a two-layer fully connected network. The first layer of the fully connected network uses the ReLU activation function to reduce the 128-dimensional spatiotemporal fusion feature vector to 64 dimensions. The second layer of the fully connected network uses Sigmoid activation to output the normalized device vibration intensity A. vib , and then map the normalized device vibration intensity to [0,10] mm / s through linear scaling 2 The physical dimension range is calculated as follows:

[0023] A vib =10·σ(W2·σ(W1·H1+b1)+b2);

[0024] Where W1∈R 64×128 , W2∈R 1×64 is the weight matrix, b1∈R 64 ,b2∈R 1 is the bias term, σ is the Sigmoid function, and the coefficient 10 normalizes the output to [0,10] mm / s 2 scope.

[0025] A further improvement of the technical solution of the present invention is that: in the decision parameter generation module, the process of building a temperature change analytical model in combination with temperature characteristics and outputting the device surface temperature rise rate includes:

[0026] The preprocessed frequency domain energy feature vector and the mean and variance of the temperature data are concatenated into an 18-dimensional joint feature, arranged into a time series with a time step of 1 sample per second, forming a two-dimensional tensor H2 with a dimension of time steps × 18, which serves as the input of the dilated causal convolutional network;

[0027] A temperature change parsing model is constructed by generating an 8-layer dilated causal convolutional network. Each layer of the dilated causal convolutional network uses an exponentially increasing dilation coefficient and a 3×1 convolution kernel. Causal padding is used to ensure historical data dependencies. The output of the dilated causal convolutional layer is processed through a 4-head self-attention mechanism to capture long-term dependencies. Global temporal pooling is used to take the average along the time dimension, compressing the two-dimensional tensor into a 64-dimensional vector to extract the spatiotemporal pattern of temperature changes.

[0028] The 64-dimensional vector is input into a two-layer fully connected network, and the first layer compresses the 64-dimensional vector into a 32-dimensional vector H through the ReLU activation function. mid The second layer outputs the normalized device surface temperature rise rate through the Sigmoid activation function The normalized device surface temperature rise rate is linearly scaled and mapped to the range of [-5,5]℃ / min to obtain the device surface temperature rise rate The calculation process is as follows:

[0029] H mid =ReLU(W1·H2+b1);

[0030]

[0031] Where W1∈R 32×64 , W2∈R 1×32 is the weight matrix, b1∈R 32 ,b2∈R 1is the bias term, σ is the Sigmoid function, and the coefficient 10 expands the normalized device surface temperature rise rate to [0,10] and maps it to the actual physical dimension ℃ / min. By subtracting 5, the [0,10] range is shifted to [-5,5], so that negative values represent the temperature drop rate and positive values represent the temperature rise rate.

[0032] A further improvement of the technical solution of the present invention is that, in the scheduling and allocation module, the process of obtaining the equipment downtime probability and the cloud computing node load score based on the equipment vibration intensity and the equipment surface temperature rise rate includes:

[0033] The device vibration intensity output by the video state perception model and the device surface temperature rise rate output by the temperature change analysis model are combined into a two-dimensional feature vector. The adaptive sliding window is used to calculate the variance of the device vibration intensity within the historical 30 minutes. Variance of the temperature rise rate on the equipment surface Define the vibration weight α and temperature rise weight β, and compare the equipment vibration intensity v with the forward temperature rise rate r of the equipment surface. + The linear combination is mapped to the interval [0,1] to obtain the joint health index H. The modified logistic function is used to convert the joint health index into the equipment downtime probability P. The calculation process is as follows:

[0034]

[0035] β=1-α;

[0036] r + =max(r,0);

[0037]

[0038]

[0039] Where ∈ = 10 -5 To prevent the denominator from being zero, Normalize the device vibration intensity to [0,1], The surface temperature rise rate of the equipment is normalized to [0, 1], k = 8 controls the steepness of the equipment shutdown probability curve, and θ = 0.6 is the equipment health status threshold;

[0040] Introducing cloud computing node temperature T j Node computing utilization U j and network delay D j , calculate the cloud computing node load score L j , the calculation process is as follows:

[0041]

[0042] Among them, 0.5, 0.3 and 0.2 are the node temperature weight, node computing utilization weight and network delay weight respectively. If the historical failure rate of the cloud computing node exceeds 10%, the cloud computing node temperature weight is increased to 0.6, the node computing utilization weight and the network delay weight are increased to 0.2.

[0043] A further improvement of the technical solution of the present invention is that: in the scheduling and allocation module, the process of triggering the device shutdown instruction and the cloud computing task migration decision includes:

[0044] When P ≥ 0.7, a shutdown command is sent to the device controller, power is cut off, and an emergency cooling program is initiated. When 0.4 ≤ P ≤ 0.7, the device load is reduced to 80% of the rated value. Vibration and temperature data are collected every 5 minutes, a periodic health report is generated, and manual inspections are conducted. When P < 0.4, the device maintains full load operation, continuously records vibration and temperature data, and updates the joint health index every 30 minutes. If P < 0.2 for 12 consecutive hours, performance optimization mode is activated, 10% redundant computing resources are allocated to high-load nodes, and device operating parameters are adjusted to explore the optimal energy efficiency range.

[0045] Use the best fit algorithm to select the optimal target node with the lowest load as the migration target. When the node temperature is ≥85°C and the node load score is ≥0.8, forced migration is triggered. If the current node load score is ≥0.5 and there is a low-load node with a score below 0.3, preventive migration is triggered, proactively migrating non-real-time tasks to balance the cluster load.

[0046] Through containerization technology, task shards are migrated in seconds, reducing the end-to-end latency of real-time task migration to ≤50ms and that of non-real-time tasks to ≤200ms. During the migration process, redundant transmission and breakpoint-resume mechanisms are implemented. In the event of a failure, the system automatically rolls back to the original node, records the abnormal status, and triggers incremental model training.

[0047] After the task migration is completed, the original node resources are released and reallocated to the high-priority queue. The load changes of the target node are monitored. If the target node load still exceeds 0.75 and 0.85, secondary migration and elastic expansion are initiated respectively to maintain the overall cluster load rate ≤ 75%. At the same time, the temperature safety threshold is adaptively adjusted based on ambient temperature fluctuations.

[0048] A further improvement of the technical solution of the present invention is that: in the visual feedback module, the process of mapping the device health status and node load distribution through the heat map and topology map includes:

[0049] The joint health index is mapped to HSV hue to generate a heat map of the equipment health status. The transparency is adjusted along the gradient of the joint health index to highlight abnormal areas.

[0050] Combining node load scores and node location information, the repulsive and attractive forces between nodes are defined, and the force-directed layout algorithm is used to calculate the spatial distribution of nodes. The cloud computing node load distribution topology is displayed in layers according to the device surface temperature and node load score, with high-temperature nodes placed on the top layer and low-load nodes on the bottom layer. The radius of the circular node is proportional to the node computing utilization rate, and the color changes gradually according to the temperature. The temperature threshold is set to 90 degrees. The width of the connection between nodes is proportional to the network delay, and the load and task migration status are marked in real time.

[0051] A further improvement of the technical solution of the present invention is that: in the visual feedback module, the process of performing feedback optimization to adjust feature weights includes:

[0052] When users select an abnormal area in the device health status heat map and manually drag the weight slider, they enter the vibration weight and temperature rise weight, and update the formula for calculating the joint health index. When the AI data scheduling and allocation system's historical false alarm rate is ≥5%, the weight of high false alarm features is reduced in proportion to the false alarm rate. When the ambient temperature fluctuates ≥10°C, the temperature rise weight is increased by 0.1 for every 10°C fluctuation, and an upper limit constraint is set.

[0053] Based on the adjusted vibration weight and temperature rise weight, the equipment downtime probability is recalculated. At the same time, the vibration weight and temperature rise weight are written to the distributed configuration center. Version control and consistency protocols are used to ensure that the weights of nodes in the cluster are consistent.

[0054] The updated vibration weights and temperature rise weights are divided into an experimental group and a control group. The false alarm rates are compared and a t-test is used to assess whether the decrease in the false alarm rate in the experimental group is significant. If the false alarm rate in the experimental group decreases by 10% or more, the updated vibration weights and temperature rise weights are enabled in the entire cluster. If the false alarm rate in the experimental group decreases by less than 10%, the system is rolled back to the historical stable version.

[0055] In a second aspect, a cloud computing-based AI data scheduling and allocation method includes the following steps:

[0056] S1. Collect and pre-process the video data of the devices in the cloud computing target scene and the temperature data of the surface of the corresponding server hardware devices;

[0057] S2, extracting video features and temperature features from the preprocessed video data and temperature data;

[0058] S3. Combining video features and temperature features, a video state perception model and a temperature change analysis model are constructed to output the device vibration intensity and the device surface temperature rise rate.

[0059] S4. Based on the equipment vibration intensity and the equipment surface temperature rise rate, the equipment shutdown probability and cloud computing node load score are obtained, thereby triggering the equipment shutdown instruction and cloud computing task migration decision;

[0060] S5. Draw a heat map of the equipment health status and a topology map of the node load distribution, and perform feedback optimization to adjust the feature weights.

[0061] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art:

[0062] 1. The present invention provides an AI data scheduling and allocation system and method based on cloud computing. By integrating the collaborative analysis of multimodal data of equipment vibration and equipment surface temperature, and combining spatiotemporal convolutional networks with frequency domain feature analysis models, the system significantly improves the detection accuracy of equipment abnormal conditions, while achieving dynamic and elastic allocation of cloud computing resources, effectively solving the decision-making lag and resource waste problems of traditional single-source data-driven systems.

[0063] 2. The present invention provides an AI data scheduling and distribution system and method based on cloud computing. Based on a dynamic weight fusion algorithm and an adaptive threshold trigger mechanism, it generates equipment downtime probability and node load scores in real time, supports millisecond-level task migration and hierarchical response strategies, and significantly improves the resource utilization efficiency and task execution stability of cloud computing clusters while ensuring the safe operation of industrial equipment.

[0064] 3. The present invention provides an AI data scheduling and distribution system and method based on cloud computing, which intuitively feedbacks the health status of equipment and node load distribution through visual heat maps and dynamic topology maps, combines false alarm rate-driven weight optimization and environmental adaptation mechanisms to form a closed-loop feedback control, greatly enhancing the adaptability of the AI data scheduling and distribution system to complex industrial scenarios and reducing the need for manual operation and maintenance intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0066] Figure 1 A block diagram of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] Example 1, as Figure 1 As shown, the present invention provides an AI data scheduling and allocation system based on cloud computing, including an AI data acquisition module, a feature extraction module, a decision parameter generation module, a scheduling and allocation module, and a visual feedback module;

[0069] The AI data acquisition module collects and pre-processes the video data of the devices in the cloud computing target scene and the temperature data on the surface of the corresponding server hardware devices. A high-resolution camera array is installed around the devices in the cloud computing target scene to cover the entire working area of the equipment. Micro thermocouple sensors are mounted on the surface of the heat-generating components of the equipment. Wide-angle infrared thermal imagers are installed on the front and rear sides of the server racks in the cloud computing target scene to fully cover the temperature field of the heat dissipation channel. Chip temperature sensors are deployed on the surface of each server power module and hard disk bracket and installed close to the metal casing. The camera array captures the device operation video data at a rate of 30 frames per second. The edge computing node performs key frame extraction and extracts the reference frame at a time interval of Δt = 1 / 30 seconds. When the difference between adjacent frames exceeds Δt, dynamic frame supplementary acquisition is triggered. The micro thermocouple sensor and the wide-angle infrared thermal imager synchronously collect temperature data. The edge computing node performs sliding window and standard Standardized processing, in which the high-resolution camera array is arranged at a vertical downward angle, 0.5-1.5 meters from the device surface, with a resolution of not less than 3840×2160 pixels and an 8-bit color depth. The spacing between micro-thermocouple sensors does not exceed 10 cm. The vertical height of the wide-angle infrared thermal imager corresponds to the position of the GPU heat dissipation module and the CPU heat sink, with a horizontal interval of 1 meter. The sampling frequency of the chip temperature sensor is not less than 1000 Hz. The non-local mean filtering algorithm is applied to the dynamic frames in the video data, the search window is set to 7×7 pixels, the similarity threshold is set to 0.8, and the discrete cosine transform is used to compress the reference frame in blocks with a compression ratio of not less than 10:1. The low-frequency component is retained, and the cubic spline interpolation method is used to fill the data points within the temperature data detection window that deviate from the mean by ±3 times the standard deviation. The video data and temperature data are synchronized with the timestamp through the NTP protocol. When the deviation exceeds 1 millisecond, the edge computing node time domain interpolation compensation is triggered;

[0070] The feature extraction module extracts video features and temperature features from the preprocessed video data and temperature data. It uses a dense optical flow algorithm to obtain the displacement vectors of adjacent frames of the preprocessed video data pixel by pixel, divides each frame into a 32×32 grid, and calculates the mean of the displacement vectors in each grid to generate a motion vector field feature matrix. Based on the key frame device area, a grayscale co-occurrence matrix is constructed to obtain the texture entropy values in four directions including 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The mean of the four-directional texture entropy values is taken to generate a texture entropy value feature vector. The preprocessed temperature data is decomposed into an 8-layer wavelet packet using the Daubechies-4 wavelet basis function to extract the scale coefficient and detail coefficient. The scale coefficient and detail coefficient are spliced into a frequency domain energy feature vector. At the same time, the mean and variance of the temperature data in a sliding window with a window length of 60 seconds and a step size of 10 seconds are calculated.

[0071] The decision parameter generation module combines video features and temperature features to build a video state perception model and a temperature change analysis model, outputs the device vibration intensity and the device surface temperature rise rate, and performs spatial alignment and channel fusion on the preprocessed motion vector field feature matrix and the texture entropy feature vector. The motion vector field feature matrix retains the original spatial dimension and is expanded into a two-channel tensor. The texture entropy feature vector is reshaped to the same spatial size through upsampling and used as the third channel to generate a 32×32×3 three-channel joint input tensor H1. By generating a 3D convolutional network containing 6 residual blocks, a video state perception model is constructed. The input layer of the 3D convolutional network receives a 32×32×3×1 four-channel joint input tensor, where the fourth dimension is the time step, initialized to a single Frame, each residual block of the 3D convolutional network consists of two layers of convolution kernels and jump connections. The sizes of the first and second residual blocks are 5×5×5 and 3×3×3, respectively. The spatiotemporal features are gradually extracted by alternating stacked convolution and pooling operations. The process of extracting spatiotemporal features includes a forward propagation stage, in which the four-channel joint input tensor is subjected to the residual block to extract the local spatiotemporal pattern, the spatial dimension is compressed and the time step is expanded through the pooling layer, and a 128-dimensional spatiotemporal fusion feature vector is generated by global average pooling. The 128-dimensional spatiotemporal fusion feature vector is input into a two-layer fully connected network. The first-layer fully connected network uses the ReLU activation function to reduce the 128-dimensional spatiotemporal fusion feature vector to 64 dimensions. The second-layer fully connected network outputs the normalized device vibration intensity A through Sigmoid activation. vib , and then map the normalized device vibration intensity to [0,10] mm / s through linear scaling 2 The physical dimension range is calculated as follows:

[0072] A vib =10·σ(W2·σ(W1·H1+b1)+b2);

[0073] Where W1∈R 64×128, W2∈R 1×64 is the weight matrix, b1∈R 64 ,b2∈R 1 is the bias term, v is the Sigmoid function, and the coefficient 10 normalizes the output to [0,10] mm / s 2 The preprocessed frequency domain energy feature vector and the mean and variance of the temperature data are concatenated into an 18-dimensional joint feature, and arranged into a time series at a time step of 1 sample per second to form a two-dimensional tensor H2 with a dimension of time steps × 18, which is used as the input of the dilated causal convolutional network. By generating an 8-layer dilated causal convolutional network, a temperature change analysis model is constructed. Each layer of the dilated causal convolutional network uses an exponentially growing dilation coefficient and a 3×1 convolution kernel. Causal filling is used to ensure the dependency of historical data. The output of the dilated causal convolutional layer captures long-term dependencies through a 4-head self-attention mechanism. Global time pooling is used to take the average along the time dimension, compressing the two-dimensional tensor into a 64-dimensional vector to extract the spatiotemporal pattern of temperature changes. The 64-dimensional vector is input into a two-layer fully connected network. The first layer compresses the 64-dimensional vector into a 32-dimensional vector H through the ReLU activation function. mid The second layer outputs the normalized device surface temperature rise rate through the Sigmoid activation function The normalized device surface temperature rise rate is linearly scaled and mapped to the range of [-5,5]℃ / min to obtain the device surface temperature rise rate The calculation process is as follows:

[0074] H mid =ReLU(W1·H2+b1);

[0075]

[0076] Where W1∈R 32×64 , W2∈R 1×32 is the weight matrix, b1∈R 32 ,b2∈R 1 is the bias term, σ is the Sigmoid function, and the coefficient 10 expands the normalized device surface temperature rise rate to [0,10] and maps it to the actual physical dimension ℃ / min. By subtracting 5, the [0,10] range is shifted to [-5,5], so that negative values represent the temperature drop rate and positive values represent the temperature rise rate.

[0077] The scheduling and allocation module obtains the equipment shutdown probability and cloud computing node load score based on the equipment vibration intensity and equipment surface temperature rise rate, and then triggers the equipment shutdown instruction and cloud computing task migration decision. The equipment vibration intensity output by the video state perception model and the equipment surface temperature rise rate output by the temperature change analysis model are combined into a two-dimensional feature vector, and an adaptive sliding window is used to calculate the equipment vibration intensity variance within the historical 30 minutes. Variance of the temperature rise rate on the equipment surface Define the vibration weight α and temperature rise weight β, and compare the equipment vibration intensity v with the forward temperature rise rate r of the equipment surface. + The linear combination is mapped to the interval [0,1] to obtain the joint health index H. The modified logistic function is used to convert the joint health index into the equipment downtime probability P. The calculation process is as follows:

[0078]

[0079] β=1-α;

[0080] r + =max(r,0);

[0081]

[0082] Where ∈ = 10 -5 To prevent the denominator from being zero, Normalize the device vibration intensity to [0,1], The device surface temperature rise rate is normalized to [0, 1], k = 8 controls the steepness of the device shutdown probability curve, θ = 0.6 is the device health threshold, and the cloud computing node temperature T is introduced. j Node computing utilization U j and network delay D j , calculate the cloud computing node load score L j , the calculation process is as follows:

[0083]

[0084] Among them, 0.5, 0.3 and 0.2 are the node temperature weight, node computing utilization weight and network delay weight respectively. If the historical failure rate of the cloud computing node exceeds 10%, the cloud computing node temperature weight is increased to 0.6, the node computing utilization weight and the network delay weight are increased to 0.2. When P≥0.7, the equipment shutdown command is sent to the equipment controller, the power is cut off and the emergency cooling program is started. When 0.4≤P≤0.7, the equipment load is reduced to 80% of the rated value. Vibration and temperature data are collected every 5 minutes, a periodic health report is generated, and manual inspections are carried out. When P<0.4, the equipment maintains full load operation, continuously records vibration and temperature data, and updates the joint health index every 30 minutes. If P<0.2 for 12 consecutive hours, the performance optimization mode is started, 10% redundant computing resources are allocated to high-load nodes, the equipment operating parameters are adjusted, the optimal energy efficiency range is explored, and the best adaptation algorithm is used to screen the load. The lowest optimal target node is used as the migration target. When the node temperature is ≥85°C and the node load score is ≥0.8, forced migration is triggered. If the current node load score is ≥0.5 and there is a low-load node with a score below 0.3, preventive migration is triggered, and non-real-time tasks are actively migrated to balance the cluster load. Task sharding is migrated in seconds through containerization technology, making the end-to-end delay of real-time task migration ≤50ms and the delay of non-real-time tasks ≤200ms. During the migration process, redundant transmission and breakpoint resumption mechanisms are added. In case of failure, the system automatically rolls back to the original node and records the abnormal status, triggering incremental model training. After the task migration is completed, the original node resources are released and reallocated to the high-priority queue. The load changes of the target node are monitored. If the target node load still exceeds 0.75 and 0.85, secondary migration and elastic expansion are initiated respectively to maintain the overall cluster load rate ≤75%. At the same time, the temperature safety threshold is adaptively adjusted based on ambient temperature fluctuations.

[0085] The visual feedback module maps the device health status and node load distribution through heat maps and topology maps, and performs feedback optimization to adjust the feature weights. The joint health index is mapped through HSV hue to generate a device health status heat map. The transparency is adjusted with the gradient of the joint health index to highlight the abnormal area. The node load score and node location information are combined to define the repulsion and attraction between nodes. The force-directed layout algorithm is used to calculate the node spatial distribution. The cloud computing node load distribution topology map is displayed in layers according to the device surface temperature and node load score. High-temperature nodes are placed on the top layer and low-load nodes are placed on the bottom layer. The radius of the circular node is proportional to the node computing utilization rate. The color changes by temperature, and the temperature threshold is set to 90 degrees. The width of the connection between nodes is proportional to the network delay, and the load and task migration status are marked in real time. When the user selects the abnormal area in the device health status heat map and manually drags the weight slider, the vibration weight and temperature are input. Increase the weight and update the formula for calculating the joint health index. When the historical false alarm rate of the AI data scheduling and allocation system is ≥5%, the weight of high false alarm features is attenuated according to the false alarm rate ratio. When the ambient temperature fluctuates ≥10℃, the temperature rise weight is increased by 0.1 for every 10 degrees Celsius fluctuation, and an upper limit constraint is set. Based on the adjusted vibration weight and temperature rise weight, the probability of equipment downtime is recalculated. At the same time, the vibration weight and temperature rise weight are written to the distributed configuration center. Version control and consistency protocols are used to ensure consistency of node weights in the cluster. The updated vibration weight and temperature rise weight are divided into the experimental group and the vibration weight and temperature rise weight before the update are divided into the control group. The difference in their false alarm rates is compared, and the t-test is used to evaluate whether the false alarm rate of the experimental group has decreased significantly. If the false alarm rate of the experimental group is reduced by ≥10%, the updated vibration weight and temperature rise weight are enabled in the entire cluster. If the false alarm rate of the experimental group is reduced by <10%, it is rolled back to the historical stable version.

[0086] Example 2, as Figure 1 As shown, based on Example 1, the present invention also provides a technical solution: a cloud computing-based AI data scheduling and allocation method, comprising the following steps:

[0087] S1. Collect and pre-process the video data of the devices in the cloud computing target scene and the temperature data of the surface of the corresponding server hardware devices;

[0088] S2, extracting video features and temperature features from the preprocessed video data and temperature data;

[0089] S3. Combining video features and temperature features, a video state perception model and a temperature change analysis model are constructed to output the device vibration intensity and the device surface temperature rise rate.

[0090] S4. Based on the equipment vibration intensity and the equipment surface temperature rise rate, the equipment shutdown probability and cloud computing node load score are obtained, thereby triggering the equipment shutdown instruction and cloud computing task migration decision;

[0091] S5. Draw a heat map of the equipment health status and a topology map of the node load distribution, and perform feedback optimization to adjust the feature weights.

[0092] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A cloud computing-based AI data scheduling and distribution system, characterized by: It includes AI data acquisition module, feature extraction module, decision parameter generation module, scheduling allocation module and visual feedback module; The AI data acquisition module collects and pre-processes video data of devices in the cloud computing target scene and temperature data on the surface of the corresponding server hardware devices; The feature extraction module extracts video features and temperature features from the pre-processed video data and temperature data; The decision parameter generation module combines video features and temperature features to build a video state perception model and a temperature change analysis model, and outputs the device vibration intensity and the device surface temperature rise rate; The scheduling and allocation module obtains the equipment shutdown probability and cloud computing node load score based on the equipment vibration intensity and the equipment surface temperature rise rate, and then triggers the equipment shutdown instruction and cloud computing task migration decision; The visual feedback module maps the device health status and node load distribution through heat maps and topology maps, and performs feedback optimization to adjust feature weights.

2. The cloud computing-based AI data scheduling and distribution system according to claim 1, characterized in that: In the AI data acquisition module, the acquisition and preprocessing process of the video data of the device in the cloud computing target scene and the temperature data of the surface of the corresponding server hardware device includes: Install high-resolution camera arrays around devices in the cloud computing target scenario to cover the entire working area of the devices. Mount micro-thermocouple sensors on the surfaces of heat-generating components. Install wide-angle infrared thermal imagers on the front and back of server racks in the cloud computing target scenario to fully cover the temperature field of the heat dissipation channel. Deploy surface-mount temperature sensors on the surfaces of each server's power module and hard drive tray, installing them close to the metal casing. The camera array captures device operation video data at a rate of 30 frames per second. The edge computing node performs key frame extraction, extracting reference frames at a time interval of Δt = 1 / 30 second. When the difference between adjacent frames exceeds Δt, dynamic frame supplementary acquisition is triggered. The micro-thermocouple sensor and the wide-angle infrared thermal imager synchronously collect temperature data. The edge computing node performs sliding window and normalization processing on the raw temperature data. A non-local mean filtering algorithm is applied to the dynamic frames in the video data. The search window is set to 7×7 pixels, the similarity threshold is set to 0.8, and the discrete cosine transform is used to compress the reference frame in blocks. The compression ratio is not less than 10:1, and the low-frequency components are retained. The cubic spline interpolation method is used to fill the data points in the temperature data detection window that deviate from the mean by ±3 times the standard deviation. The video data and temperature data are synchronized with the timestamp through the NTP protocol. When the deviation exceeds 1 millisecond, the edge computing node time domain interpolation compensation is triggered.

3. The cloud computing-based AI data scheduling and distribution system according to claim 2, characterized in that: In the feature extraction module, the process of extracting video features and temperature features from the preprocessed video data and temperature data includes: A dense optical flow algorithm is used to obtain the displacement vectors of adjacent frames of preprocessed video data pixel by pixel. Each frame is divided into a 32×32 grid. The mean of the displacement vectors in each grid is calculated to generate a motion vector field feature matrix. Based on the key frame device area, a gray-level co-occurrence matrix is constructed to obtain the texture entropy values in four directions including 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The mean of the four-directional texture entropy values is taken to generate a texture entropy feature vector. The preprocessed temperature data were subjected to 8-layer wavelet packet decomposition using the Daubechies-4 wavelet basis function to extract the scale coefficient and detail coefficient. The scale coefficient and detail coefficient were then concatenated into a frequency domain energy eigenvector. At the same time, the mean and variance of the temperature data within a sliding window with a window length of 60 seconds and a step size of 10 seconds were statistically analyzed.

4. The cloud computing-based AI data scheduling and distribution system according to claim 3, characterized in that: In the decision parameter generation module, the process of building a video state perception model by combining video features and outputting the device vibration intensity includes: The preprocessed motion vector field feature matrix and texture entropy feature vector are spatially aligned and channel-fused. The motion vector field feature matrix retains the original spatial dimension and is expanded into a two-channel tensor. The texture entropy feature vector is reshaped to the same spatial size through upsampling and used as the third channel to generate a 32×32×3 three-channel joint input tensor H1. A video state perception model is constructed by generating a 3D convolutional network containing 6 residual blocks. The input layer of the 3D convolutional network receives a four-channel joint input tensor of 32×32×3×1, where the fourth dimension is the time step, initialized to a single frame. Each residual block of the 3D convolutional network consists of two layers of convolution kernels and skip connections. The sizes of the first and second layer residual blocks are 5×5×5 and 3×3×3, respectively. Spatiotemporal features are gradually extracted by alternating stacking convolution and pooling operations. The process of extracting spatiotemporal features includes a forward propagation stage, in which the four-channel joint input tensor is subjected to the residual block to extract the local spatiotemporal pattern, the spatial dimension is compressed and the time step is expanded by the pooling layer, and a 128-dimensional spatiotemporal fusion feature vector is generated by global average pooling. The 128-dimensional spatiotemporal fusion feature vector is input into a two-layer fully connected network. The first layer of the fully connected network uses the ReLU activation function to reduce the 128-dimensional spatiotemporal fusion feature vector to 64 dimensions. The second layer of the fully connected network uses Sigmoid activation to output the normalized device vibration intensity A. vib , and then map the normalized device vibration intensity to [0,10] mm / s through linear scaling 2 Physical dimension range.

5. The cloud computing-based AI data scheduling and distribution system according to claim 4, characterized in that: In the decision parameter generation module, the process of building a temperature change analysis model based on temperature characteristics and outputting the device surface temperature rise rate includes: The preprocessed frequency domain energy feature vector and the mean and variance of the temperature data are concatenated into an 18-dimensional joint feature, arranged into a time series with a time step of 1 sample per second, forming a two-dimensional tensor H2 with a dimension of time steps × 18, which serves as the input of the dilated causal convolutional network; A temperature change parsing model is constructed by generating an 8-layer dilated causal convolutional network. Each layer of the dilated causal convolutional network uses an exponentially increasing dilation coefficient and a 3×1 convolution kernel. Causal padding is used to ensure historical data dependencies. The output of the dilated causal convolutional layer is processed through a 4-head self-attention mechanism to capture long-term dependencies. Global temporal pooling is used to take the average along the time dimension, compressing the two-dimensional tensor into a 64-dimensional vector to extract the spatiotemporal pattern of temperature changes. The 64-dimensional vector is input into a two-layer fully connected network, and the first layer compresses the 64-dimensional vector into a 32-dimensional vector H through the ReLU activation function. mid The second layer outputs the normalized device surface temperature rise rate through the Sigmoid activation function The normalized device surface temperature rise rate is linearly scaled and mapped to the range of [-5,5]℃ / min to obtain the device surface temperature rise rate 6. The cloud computing-based AI data scheduling and distribution system according to claim 5, characterized in that: In the scheduling and allocation module, the process of obtaining the equipment downtime probability and the cloud computing node load score based on the equipment vibration intensity and the equipment surface temperature rise rate includes: The device vibration intensity output by the video state perception model and the device surface temperature rise rate output by the temperature change analysis model are combined into a two-dimensional feature vector. The adaptive sliding window is used to calculate the variance of the device vibration intensity within the historical 30 minutes. Variance of the temperature rise rate on the equipment surface Define the vibration weight α and temperature rise weight β, and compare the equipment vibration intensity v with the forward temperature rise rate r of the equipment surface. + Linear combination, mapped to the [0,1] interval, to obtain the joint health index H, using the modified logic function to convert the joint health index into the equipment downtime probability P, at the same time, the cloud computing node temperature T is introduced j , node computing utilization U j and network delay D j , calculate the cloud computing node load score L j .

7. The cloud computing-based AI data scheduling and distribution system according to claim 6, characterized in that: In the scheduling and allocation module, the process of triggering the device shutdown instruction and cloud computing task migration decision includes: When P ≥ 0.7, a shutdown command is sent to the device controller, power is cut off, and an emergency cooling program is initiated. When 0.4 ≤ P ≤ 0.7, the device load is reduced to 80% of the rated value. Vibration and temperature data are collected every 5 minutes, a periodic health report is generated, and manual inspections are conducted. When P < 0.4, the device maintains full load operation, continuously records vibration and temperature data, and updates the joint health index every 30 minutes. If P < 0.2 for 12 consecutive hours, performance optimization mode is activated, 10% redundant computing resources are allocated to high-load nodes, and device operating parameters are adjusted to explore the optimal energy efficiency range. Use the best fit algorithm to select the optimal target node with the lowest load as the migration target. When the node temperature is ≥85°C and the node load score is ≥0.8, forced migration is triggered. If the current node load score is ≥0.5 and there is a low-load node with a score below 0.3, preventive migration is triggered, proactively migrating non-real-time tasks to balance the cluster load. Through containerization technology, task shards are migrated in seconds, reducing the end-to-end latency of real-time task migration to ≤50ms and that of non-real-time tasks to ≤200ms. During the migration process, redundant transmission and breakpoint-resume mechanisms are implemented. In the event of a failure, the system automatically rolls back to the original node, records the abnormal status, and triggers incremental model training. After the task migration is completed, the original node resources are released and reallocated to the high-priority queue. The load changes of the target node are monitored. If the target node load still exceeds 0.75 and 0.85, secondary migration and elastic expansion are initiated respectively to maintain the overall cluster load rate ≤ 75%. At the same time, the temperature safety threshold is adaptively adjusted based on ambient temperature fluctuations.

8. The cloud computing-based AI data scheduling and distribution system according to claim 7, characterized in that: In the visual feedback module, the process of mapping the device health status and node load distribution through heat maps and topology maps includes: The joint health index is mapped to HSV hue to generate a heat map of the equipment health status. The transparency is adjusted along the gradient of the joint health index to highlight abnormal areas. Combining node load scores and node location information, the repulsive and attractive forces between nodes are defined, and the force-directed layout algorithm is used to calculate the spatial distribution of nodes. The cloud computing node load distribution topology is displayed in layers according to the device surface temperature and node load score, with high-temperature nodes placed on the top layer and low-load nodes on the bottom layer. The radius of the circular node is proportional to the node computing utilization rate, and the color changes gradually according to the temperature. The temperature threshold is set to 90 degrees. The width of the connection between nodes is proportional to the network delay, and the load and task migration status are marked in real time.

9. The cloud computing-based AI data scheduling and distribution system according to claim 8, characterized in that: In the visualization feedback module, the process of performing feedback optimization to adjust feature weights includes: When users select an abnormal area in the device health status heat map and manually drag the weight slider, they enter the vibration weight and temperature rise weight, and update the formula for calculating the joint health index. When the AI data scheduling and allocation system's historical false alarm rate is ≥5%, the weight of high false alarm features is reduced in proportion to the false alarm rate. When the ambient temperature fluctuates ≥10°C, the temperature rise weight is increased by 0.1 for every 10°C fluctuation, and an upper limit constraint is set. Based on the adjusted vibration weight and temperature rise weight, the equipment downtime probability is recalculated. At the same time, the vibration weight and temperature rise weight are written to the distributed configuration center. Version control and consistency protocols are used to ensure that the weights of nodes in the cluster are consistent. The updated vibration weights and temperature rise weights are divided into an experimental group and a control group. The false alarm rates are compared and a t-test is used to assess whether the decrease in the false alarm rate in the experimental group is significant. If the false alarm rate in the experimental group decreases by 10% or more, the updated vibration weights and temperature rise weights are enabled in the entire cluster. If the false alarm rate in the experimental group decreases by less than 10%, the system is rolled back to the historical stable version.

10. A cloud computing-based AI data scheduling and distribution method, implemented based on the cloud computing-based AI data scheduling and distribution system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Collect and pre-process the video data of the devices in the cloud computing target scene and the temperature data of the surface of the corresponding server hardware devices; S2, extracting video features and temperature features from the preprocessed video data and temperature data; S3. Combining video features and temperature features, a video state perception model and a temperature change analysis model are constructed to output the device vibration intensity and the device surface temperature rise rate. S4. Based on the equipment vibration intensity and the equipment surface temperature rise rate, the equipment shutdown probability and cloud computing node load score are obtained, thereby triggering the equipment shutdown instruction and cloud computing task migration decision; S5. Draw a heat map of the equipment health status and a topology map of the node load distribution, and perform feedback optimization to adjust the feature weights.

Images