A computing power resource scheduling mode identification method

Abstract

The application discloses a computing power resource scheduling mode recognition method, relates to the technical field of data processing, and specifically discloses a computing power resource scheduling mode recognition method, which comprises the following steps: collecting scheduling log data and resource state data, constructing computing power resource scheduling mode recognition training data, marking the computing power resource scheduling mode recognition training data, pre-processing the computing power resource scheduling mode recognition training data, performing spatio-temporal joint feature coding and dynamic feature enhancement on the training data, training a computing power resource scheduling mode recognition model, and performing computing power resource scheduling mode recognition. Through the design of the spatio-temporal joint encoder, the event type embedding and the resource state differential are fused, the synchronous response relationship between event triggering and resource state change is accurately modeled, the limitation of simply splicing events and state vectors in the traditional technology is broken through, the instantaneous influence of events on the resource state is more accurately captured, and therefore the mode recognition precision is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for identifying computing resource scheduling patterns. Background Technology

[0002] With the rapid development of technologies such as cloud computing, big data, and artificial intelligence, computing resource scheduling has become a critical core capability in infrastructures such as data centers, edge computing platforms, and supercomputing clusters. Scheduling systems must make real-time decisions in high-concurrency, high-heterogeneity, and high-dynamic environments based on constantly changing task requirements, resource status, and service level agreements (SLAs) to ensure efficient utilization of system resources and continuous service availability. However, scheduling behavior typically manifests as a complex sequence of events and a dynamic response process of resource status, exhibiting significant heterogeneity, nonlinearity, and multi-scale characteristics.

[0003] In the prior art, Chinese invention patent CN119903293A proposes an intelligent flight pattern recognition method based on IFPRM-SBLFS deep learning. This method, for the first time, addresses the flight state recognition problem from a time-detection perspective. It proposes an intelligent flight pattern recognition method that combines sensitive boundary recognition and long flight sequence processing, overcoming the limitations of traditional methods in single-pattern recognition and complex maneuver recognition. It can accurately locate the temporal boundaries of flight patterns and achieve online automatic recognition, significantly improving the model's practicality and ease of operation. However, it still has the following drawbacks: 1. Traditional methods, such as directly concatenating event and resource state vectors, fail to fully consider the temporal correlation between event triggering and resource state changes, easily leading to model... 1. The model cannot accurately capture the immediate impact of events when processing scheduling patterns, affecting the accuracy of pattern recognition; 2. Conventional masking strategies are usually based on fixed probabilities or random location-generated masks, which are weak in modeling low-frequency events and abnormal patterns, especially when dealing with imbalanced event distributions, often leading to a decline in model performance or the neglect of key events; 3. Existing gated loop structures are difficult to effectively handle the long-term evolution of resource states and the instantaneous impact of scheduling events, resulting in insufficient ability to capture features at multiple time scales and an inability to accurately model the spatiotemporal dynamics of scheduling behavior; 4. Traditional static course learning strategies fail to dynamically adjust the complexity of samples according to training progress, which can easily lead to insufficient learning of some more difficult samples, thus affecting the final generalization ability and adaptability. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for identifying computing resource scheduling patterns.

[0005] The technical solution adopted to solve the above-mentioned technical problems is: a method for recognizing computing resource scheduling patterns, comprising the following steps: S1, collecting scheduling log data and resource status data; S2, constructing training data for computing resource scheduling pattern recognition; S3, labeling the training data for computing resource scheduling pattern recognition; S4, preprocessing the training data for computing resource scheduling pattern recognition; S5, performing spatiotemporal joint feature encoding and dynamic feature enhancement on the training data; S6, training the computing resource scheduling pattern recognition model; the method for training the computing resource scheduling pattern recognition model comprises the following steps: S601, constructing a computing resource scheduling pattern recognition model; S60 2. Define gated loop units; S603. Initialize the weights of the computing resource scheduling pattern recognition model; S604. Construct a hybrid attention mechanism; S605. Apply double regularization constraints; S606. Perform data forward propagation; S607. Perform dynamic course learning scheduling; S608. Perform error backpropagation: During error backpropagation, adjust the model parameters based on the difference between the model output and the true label; S609. Determine the stopping iteration condition: When the accuracy improvement of the validation set is less than 0.5% within 5 consecutive training cycles, save the model parameters with the best performance on the validation set to prevent overfitting; S7. Perform computing resource scheduling pattern recognition.

[0006] Furthermore, the method for preprocessing the training data for computing resource scheduling pattern recognition in S4 includes the following steps: S401, performing data cleaning operations to remove duplicate event records caused by abnormal acquisition and fill in missing resource status values ​​caused by sensor failures; S402, performing feature normalization by standardizing the resource status sequence according to resource type using Z-score to eliminate dimensional differences; S403, implementing sample balancing processing by increasing the sample size of rare patterns through time window sliding resampling and generating synthetic samples using SMOTE sampling technology.

[0007] Furthermore, the method for spatiotemporal joint feature encoding of training data in S5 is as follows: A spatiotemporal joint encoder is constructed to obtain low-dimensional vector representations of discrete event types through event embedding. Simultaneously, the instantaneous response trend of resources is captured based on the temporal differential of the resource state tensor at the time of event occurrence, thereby constructing a joint feature representation across scheduling events. This process takes the event type identifier as input, generates an event embedding vector through a category embedding layer, and concatenates it with the resource state differential at the corresponding time point to form a joint feature vector for subsequent task modeling, expressed as:

[0008] ,

[0009] In the above formula, For the first The joint feature vector of the scheduling events, For category embedding layer, The representation maps discrete event types to 3D vectors capture the semantic features of events. The dimension of the event embedding vector. For the first The type identifier of a scheduling event. For vector concatenation, The derivative of the resource state tensor at the time of the event. In time The resource state tensor, and , For the number of resource types, For time window, Indicates the parameter dimension. For the first The specific time point when a scheduling event occurs. Indicates will Assign to .

[0010] Furthermore, the method for dynamic feature enhancement of training data in S5 is as follows: An adaptive masking mechanism guided by event importance is employed. After layer normalization of each joint feature vector, an importance mask vector is generated using learnable parameters. The original feature vector and the batch mean vector are then fused based on the mask value, as shown below:

[0011] ,

[0012] ,

[0013] In the above formula, Let be the importance mask vector, representing the th Importance weights of the joint feature vector of each event It is the Sigmoid activation function. For learnable weight vectors, for transpose, Representation layer normalization, For the first The joint feature vector of the scheduling events;

[0014] For the first An enhanced feature vector after masking. For element-wise multiplication, This represents the batch characteristic mean.

[0015] Furthermore, the method for defining the gated loop unit in S602 is as follows: a dual-scale gating mechanism is adopted, the differential amplitude of the resource state is used as the instantaneous change intensity coefficient, the response weights of the reset gate and the update gate to the feature are dynamically adjusted, and at each time step, the previous hidden state and the current enhanced feature vector are used to generate a gating signal, and the gating activation amplitude is adjusted in combination with the instantaneous change intensity to balance historical dependence and short-term impact, thereby generating the updated hidden state.

[0016] Furthermore, the method for initializing the weights of the computing power resource scheduling pattern recognition model in S603 is as follows: the spectral characteristics of the feature covariance matrix are used to design the initialization strategy, the joint features of the training samples are extracted, their covariance matrix is ​​constructed, the maximum and minimum eigenvalues ​​are calculated, and the weight initialization range is set accordingly, so that the initial weights are closer to the sample distribution structure.

[0017] Furthermore, the method for constructing the hybrid attention mechanism in S604 is as follows: a dual-path attention mechanism is adopted, and attention calculations are performed on the event sequence and the resource state time sequence respectively. In event attention, an event query vector is constructed and matched with each event key vector to extract key events. In state attention, a state query vector is constructed and matched with the state key vector to highlight resource changes at key time points. The outputs of the two attention paths are concatenated and fused through a multilayer perceptron to obtain a global contextual feature representation.

[0018] Furthermore, the method of dual regularization constraint in S605 is as follows: a feature-level and weight-level regularization strategy is adopted, and the regularization loss is composed of two weighted sub-terms, which respectively characterize the feature orthogonality and the weight sparsity, expressed as:

[0019] ,

[0020] In the formula, This is the regularization loss term. The coefficients of the characteristic orthogonality constraint are... For the attention output matrix, This is the transpose of the attention output matrix. It is the identity matrix. It is the Frobenius norm. It is the L21 norm. The coefficients of the weighted sparsity constraint are... For the first The weight matrix of the layer, This is the index for the network layer.

[0021] Furthermore, the method for forward propagation of data in S606 is as follows: During the training process, for the input batch data, each sample is an enhanced feature vector sequence within a time window. Each vector in the sequence corresponds to a scheduling event and its resource state response. The input sequence is first fed into a dual-scale gating unit layer. Based on the previous hidden state and the current enhanced feature vector, this unit dynamically generates reset gate and update gate signals in combination with the instantaneous change intensity coefficient, adjusts the candidate hidden state, outputs the updated hidden state sequence, and captures the temporal dependency of the event chain.

[0022] Furthermore, the dynamic course learning scheduling method in S607 is as follows: A dynamic course learning scheduling mechanism is adopted, the learning progress of the current round is estimated based on the change in verification accuracy, the probability of difficulty increase is calculated using the Sigmoid activation function, and an information entropy threshold is dynamically set based on this probability. Samples with high complexity are selected from the current candidate samples and included in the next round of training, gradually increasing the learning difficulty, as expressed in:

[0023] ,

[0024] ,

[0025] In the above formula, For the first The probability of increasing difficulty in each round It is an exponential function with base to natural numbers. For the first Round verification accuracy, For the first The accuracy of wheel verification This is the sensitivity coefficient;

[0026] For the first The training dataset for the rounds, For the first The training dataset for the rounds, For the sample Information entropy, a measure of sample The complexity, As training samples, The dynamic threshold is calculated as follows: , This is the cumulative entropy distribution function.

[0027] The beneficial effects of the present invention are as follows: (1) The present invention designs a spatiotemporal joint encoder, integrates event type embedding and resource state differentiation, accurately models the synchronous response relationship between event triggering and resource state changes, breaks through the limitations of the traditional technology of simply splicing events and state vectors, and more accurately captures the instantaneous impact of events on resource state, thereby improving the accuracy of pattern recognition.

[0028] (2) This invention combines an adaptive masking mechanism based on event importance with an importance masking vector generated after layer normalization, which enhances the ability to identify rare events and abnormal patterns, effectively avoids the performance bottleneck of traditional masking strategies when dealing with low-frequency events, and improves the model’s performance in complex scheduling scenarios.

[0029] (3) The present invention adopts a dual-scale gating unit, which adjusts the information flow through the reset gate and update gate mechanism, dynamically balances the long-term resource evolution and the impact of short-term events, thereby improving the ability to capture scheduling behavior at different time scales, and can more accurately identify the immediate response and long-term trend of resource status changes, avoiding the time series modeling deficiencies of traditional methods.

[0030] (4) The present invention adopts a dynamic course learning scheduling mechanism, which dynamically adjusts the difficulty of training samples according to the changes in verification accuracy, gradually increases the complexity of learning tasks, ensures the learning ability of the model when facing samples of different complexities, and improves the generalization ability and adaptability of the model in practical applications. Attached Figure Description

[0031] Figure 1 This is a flowchart of the computing resource scheduling pattern recognition method of the present invention.

[0032] Figure 2 It is a line graph showing the performance of different gating units in complex scheduling scenarios.

[0033] Figure 3 This is a bar chart showing the impact of dynamic feature enhancement on rare scheduling pattern recognition.

[0034] Figure 4 This is a curve comparing the distribution of pattern recognition accuracy.

[0035] Figure 5 It is a scatter plot of recognition performance under different complexity scenarios. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] like Figure 1 As shown, a computing resource scheduling pattern recognition method in this embodiment includes the following steps:

[0038] S1 collects scheduling log data and resource status data.

[0039] By deploying lightweight acquisition agents on each node of the computing power cluster, event logs such as task allocation, resource application and release in the scheduling system are captured in real time. The log content includes core fields such as event type (e.g., task start, resource expansion), trigger timestamp, associated resource identifier and operation parameters, and is written to a distributed message queue in a structured format.

[0040] Synchronously collect time-series data on the resource status of physical servers and virtualization layers, covering key indicators such as CPU utilization, memory usage, network bandwidth and storage I / O. Generate multi-dimensional status sequences through a preset sampling frequency (e.g., 1 second / time) and add precise timestamps.

[0041] Establish a time-series alignment mechanism between event logs and resource status, and use a high-precision clock service to ensure the consistency of the time base of the two types of data, providing a synchronous data source for subsequent spatiotemporal joint analysis. The original data is stored in the time-series database after verification, preserving the original accuracy and complete context.

[0042] In this embodiment, one piece of data collected is:

[0043] Event type: Task start;

[0044] Timestamp: 2025-06-02 10:00:05.123;

[0045] Associated resource identifier: Node_001;

[0046] Operating parameters: CPU = 4 cores, Memory = 16GB;

[0047] Resource status time-series data: CPU utilization = 45.2% (sampling time: 10:00:05.000), memory usage = 8.1GB (sampling time: 10:00:05.000), network bandwidth = 102.4Mbps (sampling time: 10:00:05.000), storage IO = 512 IOPS (sampling time: 10:00:05.000).

[0048] S2, construct training data for computing resource scheduling pattern recognition.

[0049] Based on the collected raw logs and status data, the scheduling event sequence and its corresponding resource status slice within a continuous time window are extracted. Each training sample contains all scheduling events that occur within a fixed duration, such as all scheduling events that occur within 10 minutes, as well as the uniformly sampled status sequence of all resource types within that time period.

[0050] Associate events with resource entities, mark resource state mutation points triggered by the same event as key state anchor points. For example, mark the mutation point of CPU utilization jump at the moment of task start as a key state anchor point, and construct "event-state response pairs" as the smallest unit of pattern analysis to form the basic components of the sample library.

[0051] In this embodiment, one training data point is:

[0052] "Time window: 2025-06-02 10:00:00 to 10:10:00;

[0053] Scheduling event sequence: [Event 1: Type = Task Startup, Timestamp = 10:00:05.123, Resource Identifier = Node_001, Operation Parameter = CPU = 4 cores; Event 2: Type = Resource Expansion, Timestamp = 10:05:30.456, Resource Identifier = Node_002, Operation Parameter = CPU = 2 cores];

[0054] Resource status slice: [Time point = 10:00:00, CPU utilization = 40.1%, memory usage = 6.0GB, network bandwidth = 100.0Mbps, storage IO = 500IOPS; Time point = 10:00:01, CPU utilization = 40.3%, memory usage = 6.0GB, ... (sampled uniformly to 10:10:00)];

[0055] Critical state anchor point: Event 1 triggers a jump in CPU utilization to 60.5% (mutation point time: 10:00:05.000), forming an event-state response pair.

[0056] S3, labeled training data for computing power resource scheduling pattern recognition.

[0057] Based on the scheduling strategy manual and historical operation and maintenance records, the event sequences in each training sample are classified and labeled by manuals, and typical patterns such as "elastic scaling", "load balancing migration", and "failover and reorganization" are defined as labels for each piece of training data.

[0058] In this embodiment, a labeled training data point is:

[0059] "Time window: 2025-06-02 10:00:00 to 10:10:00;

[0060] Scheduling event sequence: [Event 1: Type = Task Startup, Timestamp = 10:00:05.123, Resource Identifier = Node_001; Event 2: Type = Resource Expansion, Timestamp = 10:05:30.456, Resource Identifier = Node_002];

[0061] Resource status slice: [Time point = 10:00:00, CPU utilization = 40.1%, memory usage = 6.0GB, network bandwidth = 100.0Mbps, storage IO = 500IOPS; Time point = 10:00:01, CPU utilization = 40.3%, memory usage = 6.0GB, ... (sampled uniformly to 10:10:00)];

[0062] Critical state anchor: Event 1 triggers a jump in CPU utilization;

[0063] Pattern category tag: Flexible expansion and contraction.

[0064] S4, preprocessing the training data for computing resource scheduling pattern recognition:

[0065] S401, perform data cleaning operation to remove duplicate event records caused by acquisition anomalies, fill in missing resource status values ​​caused by sensor failures, and use linear interpolation between previous and subsequent time points to filter out isolated events that are not associated with resource changes.

[0066] S402 performs feature normalization by standardizing the resource state sequence according to resource type using Z-score to eliminate dimensional differences and uses frequency statistical weighting for event types to reduce the dominant influence of high-frequency events on the model.

[0067] S403 implements sample balancing processing, increases the number of rare pattern samples through time window sliding resampling, and uses SMOTE sampling technology to generate synthetic samples, thereby alleviating the class imbalance problem in pattern recognition tasks.

[0068] S5 performs spatiotemporal joint feature encoding and dynamic feature enhancement of the training data.

[0069] Scheduling log data exists in the form of discrete event sequences, while resource status data is presented as a continuous time series. The two types of data are heterogeneous in terms of format, granularity, and expression mechanism. Conventional processing methods, such as directly concatenating events and state vectors or encoding them separately and then fusing them, fail to effectively model the synchronous response relationship between event-triggered behavior and resource status changes, and therefore cannot accurately characterize the impact mechanism of events on resource dynamics.

[0070] The method for spatiotemporal joint feature encoding of training data in this embodiment is as follows: A spatiotemporal joint encoder is constructed to obtain low-dimensional vector representations of discrete event types through event embedding. Simultaneously, the instantaneous response trend of resources is captured based on the temporal differential of the resource state tensor at the time of event occurrence, thereby constructing a joint feature representation across scheduling events. This process takes the event type identifier as input, generates an event embedding vector through a category embedding layer, and concatenates it with the resource state differential at the corresponding time point to form a joint feature vector for subsequent task modeling, represented as:

[0071] ,

[0072] In the above formula, For the first The joint feature vector of each scheduling event is a comprehensive representation of the event type and resource state changes. For category embedding layer, The representation maps discrete event types to A dimensional vector captures the semantic features of an event, such as discrete event types like task initiation or resource expansion. The dimension of the event embedding vector. For the first A type identifier for each scheduling event, such as "task start" or "resource expansion". For vector concatenation, The derivative of the resource state tensor at the time of the event reflects the instantaneous trend of resource changes triggered by the event; that is, it characterizes the trend of resource state changes at the time of the event. In time The resource state tensor, and This describes the state changes of a resource within a time window. For the number of resource types, For time window, Indicates the parameter dimension. Indicates will Assign to , For the first The specific time point when a scheduling event occurs.

[0073] In this embodiment, spatiotemporal joint feature encoding calculation is performed using labeled training data, such as: time window: 2025-06-02 10:00:00 to 10:10:00, event sequence includes event 1 and event 2, and label is elastic scaling.

[0074] In this embodiment, the specific process of performing spatiotemporal joint feature encoding of the training data is as follows:

[0075] 1) Extract the event type identifier for each scheduling event, such as event 1 being of type "task start" and event 2 being of type "resource expansion";

[0076] 2) Map discrete event types to low-dimensional vector representations through a category embedding layer:

[0077] The "Task Startup" type is mapped to a fixed-dimensional vector, with dimension 1. This vector captures the semantic features of the event. At the same time, at the time point of the event, such as the timestamp of event 1, 10:00:05.123, the differential of the resource state tensor is calculated, that is, the instantaneous change trend of indicators such as CPU utilization and memory usage at that moment is captured. For example, the differential value of CPU utilization jumping from 40.1% to 60.5% reflects the impact of the event on the resource state.

[0078] 3) Concatenate the event embedding vector with the resource state differential vector at the corresponding time point to form a joint feature vector. For example, for event 1, its event embedding vector is concatenated with the resource state differential at time 10:00:05.123 to generate a fusion vector. The same applies to event 2.

[0079] 4) Output the joint feature vector of each event as the primitive input for the subsequent model.

[0080] Scheduling behavior is characterized by uneven event density and strong noise in state signals. Traditional masking strategies based on fixed probability or random location are prone to disrupting the chain structure of key events, especially in the recognition of low-frequency events and abnormal patterns, where there are significant performance bottlenecks.

[0081] The method for dynamic feature enhancement of training data in this embodiment is as follows: An adaptive masking mechanism guided by event importance is employed. After layer normalization of each joint feature vector, an importance mask vector is generated using learnable parameters. The original feature vector and the batch mean vector are then fused based on the mask value, thereby enhancing the representation of rare but important event features and improving the model's ability to model imbalanced event distributions. This can be expressed as:

[0082] ,

[0083] ,

[0084] In the above formula, This is an importance mask vector, whose element values ​​are in Between, indicating the first Importance weights of the joint feature vector of each event It is the Sigmoid activation function. These are learnable weight vectors used to generate importance mask vectors; their dimensions are the same as the layer-normalized feature vectors. for transpose, Representation layer normalization is used to normalize feature vectors. For the first The joint feature vector of the scheduling events.

[0085] For the first An enhanced feature vector, after being masked, is formed by fusing the current vector with the global mean. For element-wise multiplication, This represents the batch characteristic mean.

[0086] S6, a training method for a computing resource scheduling pattern recognition model, includes the following steps:

[0087] S601, construct a computing resource scheduling pattern recognition model.

[0088] The overall architecture of the model is based on a sequence processing framework and is designed as a deep neural network. The input layer receives a preprocessed sequence of enhanced feature vectors, which consists of dynamically enhanced features. Each vector represents a joint expression of a scheduling event and its associated resource state.

[0089] The core components of deep neural networks include multi-layered cascaded dual-scale gating units and a hybrid attention mechanism. The multi-layered cascaded dual-scale gating units are used to handle time-series dependencies, while the hybrid attention mechanism is used to fuse heterogeneous events and state features to ensure that the model can capture the spatiotemporal dynamics of scheduling behavior.

[0090] Specifically, the dual-scale gating unit serves as the basic module, regulating the information flow through reset and update gate mechanisms. This unit utilizes the differential amplitude of resource states as the instantaneous change intensity coefficient to dynamically balance long-term resource evolution with short-term event impacts, and outputs a hidden state sequence.

[0091] Based on the hybrid attention mechanism, the event sequence and resource state time sequence are processed separately. The event attention path extracts key events, such as the start of high-frequency tasks, while the state attention path focuses on resource mutation points, such as a jump in CPU utilization. The two outputs are concatenated and then fused by a multilayer perceptron to form global context features.

[0092] The model output layer uses a fully connected network to map the global features of the hybrid attention output to predefined pattern categories, such as elastic scaling and load balancing migration.

[0093] Deep neural networks, combined with dual regularization constraints, suppress overfitting during training.

[0094] S602 defines a gated loop unit;

[0095] Traditional gated loop structures such as GRUs struggle to simultaneously handle the long-term evolution of resource states and the instantaneous impact of scheduling events when modeling scheduling behavior, resulting in insufficient capture of features across multiple time scales.

[0096] This embodiment employs a dual-scale gating mechanism, using the differential amplitude of the resource state as the instantaneous change intensity coefficient to dynamically adjust the response weights of the reset and update gates to features. At each time step, a gating signal is generated using the previous hidden state and the current enhanced feature vector, and the gating activation amplitude is adjusted in conjunction with the instantaneous change intensity to balance historical dependence and short-term shocks, generating the updated hidden state, represented as:

[0097] ,

[0098] ,

[0099] ,

[0100] ,

[0101] ,

[0102] In the above formula, For time step The reset gate vector is used to control the degree to which information from the previous hidden state is retained. It is the Sigmoid activation function. To reset the weight matrix of the gate, For time step The hidden state, For time step Enhanced feature vectors, This is a transient intensity coefficient used to dynamically adjust the activation amplitude of the gating signal. This is the offset vector for resetting the gate.

[0103] For time step The update gate vector is used to control the degree to which new information is introduced. To update the gate weight matrix, To update the bias vector of the gate.

[0104] For time step The candidate hidden state vector, It is the hyperbolic tangent function. Let be the weight matrix of the candidate hidden states. This is an element-wise multiplication.

[0105] For time step The hidden state.

[0106] It is the L2 norm. This represents the maximum L2 norm of the resource state differential tensor across all time points within the entire time window. The derivative of the resource state tensor at the time of the event. In time The resource state tensor.

[0107] S603, Initialization of weights for the computing power resource scheduling pattern recognition model.

[0108] Because scheduling behavior data is highly complex in both resource and time dimensions, if the initialization is not reasonable, deep networks are prone to getting stuck in local optima during training, resulting in slow training convergence and decreased generalization ability.

[0109] This embodiment uses the spectral properties of the feature covariance matrix for initialization strategy design. It extracts the joint features of the training samples, constructs their covariance matrix, calculates the maximum and minimum eigenvalues, and sets the weight initialization range accordingly. This makes the initial weights more closely resemble the sample distribution structure, thereby improving the network convergence efficiency and stability. This is expressed as:

[0110] ,

[0111] In the above formula, This is the weight matrix of the fully connected layer. To conform to the distribution, To ensure uniform distribution, The largest eigenvalue of the characteristic covariance matrix. is the smallest eigenvalue of the characteristic covariance matrix.

[0112] S604, constructing a hybrid attention mechanism.

[0113] Scheduling behavior is composed of heterogeneous event sequences and resource states. Conventional attention mechanisms cannot take into account the structural differences between the two types of data, making it difficult to allocate attention resources reasonably to highlight key features.

[0114] This embodiment employs a dual-path attention mechanism, performing attention calculations on the event sequence and the resource state time sequence separately. In event attention, an event query vector is constructed and matched with each event key vector to extract key events. In state attention, a state query vector is constructed and matched with the state key vector to highlight resource changes at key time points. The outputs of the two attention paths are concatenated and fused using a multilayer perceptron to obtain a global contextual feature representation, expressed as:

[0115] ,

[0116] ,

[0117] ,

[0118] ,

[0119] ,

[0120] In the above formula, For the first The attention weight of each event in the event attention mechanism It is an exponential function with base to natural numbers. This is the transpose of the event query vector. For the first The key vector of each event, For feature dimension, For the first The key vector of each event.

[0121] The output vector of event attention. For the first A vector of values ​​for each event.

[0122] For time step Attention weights in state attention mechanisms This is the transpose of the state query vector. For time step The state key vector, For time step The state key vector, For time step indexing, specifically to distinguish it from The time step index is used to traverse the time points in the state attention mechanism.

[0123] The output vector of state attention. For time step The state value vector.

[0124] This is the final output vector of the hybrid attention mechanism. This represents a multilayer perceptron.

[0125] S605, double regularization constraint;

[0126] In scheduling behavior modeling, the high dimensionality of feature vectors and the limited number of samples make it difficult for conventional weight regularization methods to effectively alleviate the overfitting problem, resulting in a decline in the model's generalization ability.

[0127] This embodiment employs a feature-level and weight-level regularization strategy. On the one hand, orthogonal constraints maintain low redundancy among attention outputs, avoiding linear dependence of feature representations. On the other hand, a sparse regularization term is added to each layer's weight matrix to encourage the model to automatically suppress redundant parameters, improving feature selection capability and model compression potential. The regularization loss is composed of the weighted sum of the above two sub-terms, representing feature orthogonality and weight sparsity, respectively, and is expressed as follows:

[0128] ,

[0129] In the formula, This is the regularization loss term. The coefficients of the characteristic orthogonality constraint are... Set to 0.3, For the attention output matrix, This is the transpose of the attention output matrix. It is the identity matrix. It is the Frobenius norm. It is the L21 norm. The coefficients of the weighted sparsity constraint are... For the first The weight matrix of the layer, This is the index for the network layer.

[0130] It should be noted that, The orthogonal constraint on the representation features serves to reduce redundancy in the feature representation, and... It represents the sparsity constraint of the weights, and its function is to promote the sparsity of the weight matrix.

[0131] S606 performs forward data propagation;

[0132] During training, for each batch of input data, each sample is a sequence of enhanced feature vectors within a time window. Each vector in the sequence corresponds to a scheduling event and its resource state response. The input sequence is first fed into a dual-scale gating unit layer. Based on the previous hidden state and the current enhanced feature vector, this unit dynamically generates reset gate and update gate signals, adjusts the candidate hidden state, and finally outputs the updated hidden state sequence to capture the temporal dependency of the event chain.

[0133] The hidden state sequence input hybrid attention mechanism layer performs dual-path attention computation. The event attention path targets the event sequence, constructing an event query vector that matches the event key vector to highlight key events, such as resource expansion events. The state attention path targets the resource state time sequence, constructing a state query vector that matches the state key vector to focus on resource mutations at key time points, such as a jump in CPU utilization. The output vectors of the two attention paths are concatenated and fused through a multilayer perceptron to generate a global context feature vector, enhancing the model's ability to express heterogeneous data.

[0134] The global context feature vector is passed to the fully connected output layer, where it undergoes linear transformation and nonlinear activation to generate the probability distribution of pattern categories, such as the probability of elastic scaling.

[0135] S607, Dynamic Course Learning Scheduling.

[0136] The complexity varies significantly across different scheduling scenarios. A static course learning strategy cannot adapt to the changes in the model's learning ability for samples of varying difficulty during training, thus affecting the final generalization effect.

[0137] This embodiment employs a dynamic course learning scheduling mechanism. It estimates the learning progress of the current round based on changes in verification accuracy, calculates the probability of difficulty increase using the Sigmoid activation function, dynamically sets the information entropy threshold based on this probability, and selects high-complexity samples from the current candidate samples to include in the next round of training. This gradually increases the learning difficulty, ensuring a balance between progressiveness and adaptability in model training. This can be represented as:

[0138] ,

[0139] ,

[0140] In the above formula, For the first The probability of increasing difficulty in each round It is an exponential function with base to natural numbers. For the first Round verification accuracy, For the first The accuracy of wheel verification This is the sensitivity coefficient;

[0141] For the first The training dataset for the rounds, For the first The training dataset for the rounds, For the sample Information entropy, a measure of sample The complexity, As training samples, The dynamic threshold is calculated as follows: ;

[0142] This is the cumulative entropy distribution function.

[0143] S608 performs backpropagation of errors.

[0144] The error between the global context feature representation output by the computational model and the labeled scheduling pattern is used to measure the accuracy of the model's current recognition of the scheduling pattern.

[0145] Errors are backpropagated through the hybrid attention mechanism module, which adjusts the weights of event attention and state attention, optimizes the attention output, and enables it to extract key events and highlight resource changes at key time points more accurately.

[0146] The error continues to propagate back to the dual-scale gated recurrent unit and the spatiotemporal joint encoder, updating the reset gate and update gate weights of the gated recurrent unit, as well as the event embedding and resource state differential parameters of the encoder, thereby optimizing the model's ability to model scheduling behavior, gradually reducing the error, and improving the model's generalization ability.

[0147] S609, Stop Iteration Condition Judgment: When the accuracy improvement of the validation set is less than 0.5% within 5 consecutive training cycles, or when the total training cycles reach the preset upper limit, training is terminated, and the model parameters with the best performance on the validation set are saved to prevent overfitting.

[0148] S7 performs computing resource scheduling pattern recognition.

[0149] Perform data preprocessing operations on the newly collected raw scheduling logs and resource status data, including data cleaning and feature normalization.

[0150] Feature encoding and enhancement are performed, namely, using a spatiotemporal joint feature encoder to process event sequences and resource states, generating joint feature vectors, and using a dynamic feature enhancement mechanism to adjust the vectors based on importance masks to highlight key event features.

[0151] The preprocessed enhanced feature vector sequence is input into the trained pattern recognition model. The model performs forward propagation, which involves processing the sequence through dual-scale gating units to capture event-driven resource dynamics, fusing event and state features through a hybrid attention mechanism, extracting global context, and generating a probability distribution of pattern categories in the output layer.

[0152] The scheduling mode category is determined based on the output probability, such as "elastic scaling up or down" or "failover and reconfiguration", and the identification results are fed back to the scheduling system for real-time optimization of resource allocation.

[0153] The following experimental verification was performed in this embodiment:

[0154] like Figure 2 As shown, to verify the robustness advantage of the dual-scale gating unit in a dynamic scheduling environment, experiments simulated five typical scenarios with progressive complexity: low load stabilization period, medium load fluctuation period, high load peak period, sudden fault recovery period, and mixed-mode complexity period. The pattern recognition accuracy of traditional gated recurrent units, long short-term memory networks, transformers, and the dual-scale gating unit of this invention was compared. Figure 2 The gradual grayscale change in the background visually represents the increase in scene complexity, and the curve trend reflects the degree of performance degradation of the model as the environment deteriorates. Experimental results show that traditional methods experience a precipitous performance drop in sudden failures and mixed-mode scenarios, while the curve of this invention maintains a gradual downward trend and still maintains a significant advantage during high-load peak periods. This indicates that the dual-scale gating mechanism uses the differential amplitude of resource state as an instantaneous intensity coefficient to dynamically adjust the activation weights of the reset gate and update gate, enabling the model to adaptively balance the modeling needs of long-term resource evolution and short-term event impacts.

[0155] like Figure 3 As shown, to analyze the effect of the dynamic feature enhancement mechanism on the identification of rare scheduling patterns, four low-frequency patterns were selected in the experiment: failover and reorganization, cross-cluster migration, resource isolation, and emergency expansion. The identification performance of three strategies—no feature enhancement, random masking, and the adaptive masking method of this invention—was compared. Figure 3 The height of the central column represents the harmonic mean. The method of this invention significantly outperforms the control group in all four rare modes, especially in the emergency expansion mode with the lowest frequency. The random masking strategy is even weaker than the no-enhancement method in the resource isolation mode, proving that blind masking will destroy the critical event chain. The performance improvement is negatively correlated with the mode frequency. Experimental results show that by generating importance weights through learnable parameters, the integrity of the critical event sequence is protected while enhancing the expression of rare event features, fundamentally solving the modeling defects of traditional methods for imbalanced event distribution.

[0156] like Figure 4As shown, the accuracy distribution characteristics of different algorithms in multiple tests are analyzed through kernel density estimation curves to verify the stability and reliability of the algorithm recognition. The horizontal axis represents the percentage of accuracy, and the vertical axis represents the probability density. The peak of the curve reflects the most frequently occurring accuracy level, and the curve width reflects the fluctuation range. The blue curve corresponding to this invention shows a tall and narrow distribution shape, indicating that its accuracy is highly concentrated in the high-level range and the fluctuation range is extremely small. In contrast, the distribution curves of other comparison algorithms are flatter and located to the left, showing that their average accuracy is lower and the fluctuation is larger. In particular, the curve of the feature splicing method shows a significant right-skewed shape, with a large number of low-accuracy outliers. This reflects the deep fusion capability of the spatiotemporal joint encoder of this invention for heterogeneous data, as well as the strengthening effect of the dynamic feature enhancement mechanism on key event features, ensuring stable and superior performance under various operating conditions.

[0157] like Figure 5 As shown, to evaluate the robustness of the algorithms in four typical scheduling scenarios, the experiment uses a combination of box plots and scatter plots. The position of the boxes reflects the median accuracy, the height of the boxes represents the degree of data dispersion, and the scatter plot distribution shows the original test samples. In normal task scenarios, all algorithms perform well. However, in high-complexity scenarios such as burst traffic and failover, this invention can still maintain a compact, high-level box with a median of over 90%, and the scatter plots are concentrated in the upper quartile region. In contrast, the temporal long short-term memory network and the event-gated recurrent unit show a significant downward shift of the box in the failover scenario, with a median of less than 70%, and the scatter plots are vertically diffused. This verifies the role of the dual-scale gating unit, which dynamically adjusts the information flow through the differential amplitude of resource state, accurately balancing long-term resource evolution and short-term event impacts, enabling the model to maintain stable recognition capabilities in extreme scenarios and overcoming the performance degradation problem of traditional methods in complex scheduling modes.

[0158] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A method for recognizing computing resource scheduling patterns, characterized in that, Includes the following steps: S1 collects scheduling log data and resource status data; The scheduling log data includes "event type, timestamp, associated resource identifier, operation parameters, resource status time-series data, and resource status time-series data including CPU utilization, memory usage, and network bandwidth storage; S2, construct training data for identifying computing resource scheduling patterns; S3, labeled training data for identifying computing resource scheduling patterns; S4, preprocesses the training data for computing resource scheduling pattern recognition; S5 performs spatiotemporal joint feature encoding and dynamic feature enhancement on the training data; S6, training of computing resource scheduling pattern recognition model; The training method for the computing resource scheduling pattern recognition model includes the following steps: S601, constructing a computing resource scheduling pattern recognition model; S602 defines a gated loop unit; S603, Initialization of weights for computing resource scheduling pattern recognition model; S604, constructing a hybrid attention mechanism; S605, double regularization constraint; S606 performs forward data propagation; S607, Dynamic Course Learning Scheduling; The dynamic course learning scheduling method is as follows: A dynamic course learning scheduling mechanism is adopted. The learning progress of the current round is estimated based on changes in verification accuracy. The probability of difficulty increase is calculated using the Sigmoid activation function. Based on this probability, an information entropy threshold is dynamically set. Samples with high complexity are selected from the current candidate samples and included in the next round of training, gradually increasing the learning difficulty. This can be represented as: ， ， In the above formula, For the first The probability of increasing difficulty in each round It is an exponential function with base to natural numbers. For the first Round verification accuracy, For the first The accuracy of wheel verification This is the sensitivity coefficient; For the first The training dataset for the rounds, For the first The training dataset for the rounds, For the sample Information entropy, a measure of sample The complexity, As training samples, The dynamic threshold is calculated as follows: , Let be the cumulative entropy distribution function; S608, perform error backpropagation: during the error backpropagation process, adjust the model parameters according to the difference between the model output and the true label; S609, Stop Iteration Condition Judgment: When the accuracy improvement of the validation set is less than 0.5% within 5 consecutive training cycles, save the model parameters with the best performance on the validation set to prevent overfitting; S7 performs computing resource scheduling pattern recognition.

2. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for preprocessing training data for computing resource scheduling pattern recognition in S4 includes the following steps: S401, Perform data cleaning operation to remove duplicate event records caused by abnormal acquisition and fill in missing resource status values ​​caused by sensor failure; S402, perform feature normalization, and perform Z-score standardization on the resource state sequence according to resource type to eliminate dimensional differences; S403, implement sample balancing processing, increase the sample size of rare modes by sliding resampling through time windows, and generate synthetic samples using SMOTE sampling technology.

3. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for spatiotemporal joint feature encoding of training data in S5 is as follows: A spatiotemporal joint encoder is constructed to obtain low-dimensional vector representations of discrete event types through event embedding. Simultaneously, the instantaneous response trend of resources is captured based on the temporal differential of the resource state tensor at the event occurrence time, thereby constructing a joint feature representation across scheduling events. This process takes the event type identifier as input, generates an event embedding vector through a category embedding layer, and concatenates it with the resource state differential at the corresponding time point to form a joint feature vector for subsequent task modeling, represented as: ， In the above formula, For the first The joint feature vector of the scheduling events, For category embedding layer, The representation maps discrete event types to 3D vectors capture the semantic features of events. The dimension of the event embedding vector. For the first The type identifier of a scheduling event. For vector concatenation, The derivative of the resource state tensor at the time of the event. In time The resource state tensor, and , For the number of resource types, For time window, Indicates the parameter dimension. For the first The specific time point when a scheduling event occurs. Indicates will Assign to .

4. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for dynamic feature enhancement of training data in S5 is as follows: An adaptive masking mechanism guided by event importance is used. After layer normalization of each joint feature vector, an importance mask vector is generated through learnable parameters. The original feature vector and the batch mean vector are then fused based on the mask value, as shown below: ， ， In the above formula, Let be the importance mask vector, representing the th Importance weights of the joint feature vector of each event It is the Sigmoid activation function. For learnable weight vectors, for transpose, Representation layer normalization, For the first The joint feature vector of the scheduling events; For the first An enhanced feature vector after masking. For element-wise multiplication, This represents the batch characteristic mean.

5. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for defining the gated loop unit in S602 is as follows: a dual-scale gating mechanism is adopted, the differential amplitude of the resource state is used as the instantaneous change intensity coefficient, the response weights of the reset gate and the update gate to the feature are dynamically adjusted, and at each time step, the previous hidden state and the current enhanced feature vector are used to generate a gating signal, and the gating activation amplitude is adjusted in combination with the instantaneous change intensity to balance historical dependence and short-term impact, thereby generating the updated hidden state.

6. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for initializing the weights of the computing power resource scheduling pattern recognition model in S603 is as follows: the initialization strategy is designed using the spectral characteristics of the feature covariance matrix, the joint features of the training samples are extracted, their covariance matrix is ​​constructed, the maximum and minimum eigenvalues ​​are calculated, and the weight initialization range is set accordingly, so that the initial weights are closer to the sample distribution structure.

7. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for constructing the hybrid attention mechanism in S604 is as follows: a dual-path attention mechanism is adopted, and attention calculations are performed on the event sequence and the resource state time sequence respectively. In event attention, an event query vector is constructed and matched with each event key vector to extract key events. In state attention, a state query vector is constructed and matched with the state key vector to highlight resource changes at key time points. The outputs of the two attention paths are concatenated and fused through a multilayer perceptron to obtain a global contextual feature representation.

8. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The dual regularization constraint method in S605 is as follows: a feature-level and weight-level regularization strategy is adopted, and the regularization loss is composed of two weighted sub-terms, which respectively represent the feature orthogonality and the weight sparsity, expressed as: ， In the formula, This is the regularization loss term. The coefficients of the characteristic orthogonality constraint are... For the attention output matrix, This is the transpose of the attention output matrix. It is the identity matrix. It is the Frobenius norm. It is the L21 norm. The coefficients of the weighted sparsity constraint are... For the first The weight matrix of the layer, This is the index for the network layer.

9. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, The method for forward propagation of data in S606 is as follows: During the training process, for the input batch data, each sample is an enhanced feature vector sequence within a time window. Each vector in the sequence corresponds to a scheduling event and its resource status response. The input sequence is first fed into a dual-scale gating unit layer. Based on the previous hidden state and the current enhanced feature vector, this unit dynamically generates reset gate and update gate signals in combination with the instantaneous change intensity coefficient, adjusts the candidate hidden state, outputs the updated hidden state sequence, and captures the temporal dependency of the event chain.

Images