Calculation power resource scheduling mode identification method

Abstract

The invention discloses a computing power resource scheduling mode identification method, and relates to the technical field of data processing. Comprising 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; preprocessing the computing power resource scheduling mode recognition training data; and training a computing power resource scheduling mode recognition model, and performing computing power resource scheduling mode recognition. By designing a space-time joint encoder, fusing event type embedding and resource state differential, and accurately modeling a synchronous response relationship between event triggering and resource state change, the limitation of simply splicing events and state vectors in the traditional technology is broken through, the instantaneous influence of the events on the resource state is more accurately captured, and the resource state change is more accurate. Therefore, the pattern recognition precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a computing resource scheduling mode recognition method. BACKGROUND

[0002] With the rapid development of cloud computing, big data and artificial intelligence technologies, computing resource scheduling has become a key core capability in data centers, edge computing platforms and supercomputing clusters and other infrastructures. The scheduling system needs to make real-time decisions in a high-concurrency, high-heterogeneity and high-dynamic environment according to changing task requirements, resource states and service level agreements, to ensure efficient use of system resources and continuous availability of services. However, scheduling behavior usually reflects a complex event sequence and dynamic response process of resource state, with significant heterogeneity, nonlinearity and multiscale characteristics.

[0003] In the prior art, the Chinese invention patent with publication number CN119903293A proposes a flight mode intelligent recognition method based on IFPRM-SBLFS deep learning, which first solves the flight state recognition problem from the perspective of time detection, proposes an intelligent flight mode recognition method combining sensitive boundary recognition and long flight sequence processing, solves the limitations of traditional methods in single mode recognition and composite maneuver recognition, and can accurately locate the temporal boundary of the flight mode, realize online automatic recognition, and greatly improve the practicability and operation convenience of the model, but still has the following shortcomings: 1. Traditional methods such as directly splicing event and resource state vectors do not fully consider the timing relationship between event triggering and resource state changes, which may lead to inaccurate capture of the immediate impact of events when processing scheduling modes, affecting the accuracy of mode recognition; 2. Conventional masking strategies usually generate masks based on fixed probabilities or positions, and have weak modeling ability for low-frequency events and abnormal patterns, especially when dealing with unbalanced event distribution, which may lead to model performance degradation or neglect of key events; 3. Existing gated recurrent structures are difficult to effectively handle long-term evolution of resource states and instantaneous impact of scheduling events, resulting in insufficient capture of multi-time scale features and inability to accurately model the spatio-temporal dynamics of scheduling behavior; 4. Traditional static curriculum learning strategies fail to dynamically adjust the complexity of samples according to training progress, which may lead to insufficient learning of some difficult samples, thereby affecting the final generalization ability and adaptability. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the above-mentioned shortcomings of the prior art and provide a computing resource scheduling mode recognition method.

[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: , 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 .

[0008] 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: , , 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.

[0009] 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.

[0010] 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.

[0011] 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.

[0012] 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: , 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 weight sparsity constraint are... For the first The weight matrix of the layer, This is the index for the network layer.

[0013] 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.

[0014] 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: , , 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: , This is the cumulative entropy distribution function.

[0015] 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.

[0016] (2) The application combines an adaptive masking mechanism based on event importance, and generates an importance mask vector after layer normalization processing, enhances the recognition ability of rare events and abnormal patterns, effectively avoids the performance bottleneck of traditional masking strategies when processing low-frequency events, and improves the performance of the model in complex scheduling scenarios.

[0017] (3) The application adopts a double-scale gating unit, adjusts the information flow through the reset gate and update gate mechanism, dynamically balances long-term resource evolution and short-term event impact, and further improves the capture ability of different time scale scheduling behaviors, can more accurately identify the immediate response and long-term trend of resource state changes, and avoids the time sequence modeling deficiency of traditional methods.

[0018] (4) The application adopts a dynamic curriculum learning scheduling mechanism, dynamically adjusts the difficulty of training samples according to the verification accuracy, gradually improves the complexity of learning tasks, ensures the learning ability of the model when facing samples of different complexity, improves the generalization ability of the model and the adaptability in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flow chart of the computing power resource scheduling mode recognition method of the application.

[0020] Figure 2 is a performance performance line chart of different gating units in complex scheduling scenarios.

[0021] Figure 3 is a bar chart of the influence of dynamic feature enhancement on rare scheduling mode recognition.

[0022] Figure 4 is a mode recognition accuracy rate distribution comparison curve.

[0023] Figure 5 is a recognition performance scatter plot in different complexity scenarios. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the application clearer and more understandable, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0025] As shown in Figure 1 , a computing power resource scheduling mode recognition method of the embodiment includes the following steps: S1, collecting scheduling log data and resource state data.

[0026] Through the deployment of lightweight collection agents on each node of the computing power cluster, real-time capture of task allocation, resource application and release event logs in the scheduling system, log content includes event type (such as task start, resource expansion), trigger timestamp, associated resource identifier and operation parameter, etc. Core fields are written in structured format to distributed message queue.

[0027] Synchronize the collection of resource state time series data of physical servers and virtualization layers, covering CPU utilization, memory occupancy, network bandwidth and storage IO, etc. Key indicators, generate multi-dimensional state sequence through preset sampling frequency (such as 1 second / time), and mark accurate timestamp.

[0028] Establish a time series alignment mechanism for event logs and resource states, use high-precision clock service to ensure the consistency of the time base of the two types of data, provide synchronized data sources for subsequent space-time joint analysis, and store the original data into the time series database after verification, retain the original precision and complete context.

[0029] In this embodiment, a piece of collected data is: “Event type: task start; Timestamp: 2025-06-0210:00:05.123; Associated resource identifier: Node_001; Operation parameters: CPU=4 cores, Memory=16GB; Resource state time series data: CPU utilization=45.2% (sampling time: 10:00:05.000), memory occupancy=8.1GB (sampling time: 10:00:05.000), network bandwidth=102.4Mbps (sampling time: 10:00:05.000), storage IO=512IOPS (sampling time: 10:00:05.000)”.

[0030] S2, build computing resource scheduling mode recognition training data.

[0031] Based on the collected original logs and state data, extract the scheduling event sequence and its corresponding resource state slice within the continuous time window, each training sample contains all scheduling events occurred within a fixed time, such as all scheduling events occurred within 10 minutes, and the uniform sampling state sequence of all resource types within that period.

[0032] Associate events and resource entities, mark the resource state mutation point triggered by the same event as a key state anchor point, such as marking the CPU utilization jump at the moment of task start as a key state anchor point, and build “event-state response pair” as the smallest unit of pattern analysis, forming the basic component of the sample library.

[0033] In this embodiment, one piece of training data is: "Time window: 2025-06-02 10:00:00 to 10:10:00; Scheduled event sequence: [Event 1: Type = Task Start, 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]; Resource state slices: [Time point = 10:00:00, CPU utilization = 40.1%, memory usage = 6.0 GB, network bandwidth = 100.0 Mbps, storage IO = 500 IOPS; Time point = 10:00:01, CPU utilization = 40.3%, memory usage = 6.0 GB,... (uniformly sampled to 10:10:00)]; Key state anchor point: Event 1 triggers CPU utilization to jump to 60.5% (mutation point time: 10:00:05.000), forming an event-state response pair".

[0034] S3, annotated computing resource scheduling mode recognition training data.

[0035] According to the scheduling strategy manual and historical operation records, the event sequence in each training sample is manually annotated with mode classification, and typical modes such as "elastic expansion and contraction", "load balancing migration", "fault transfer reorganization", etc. are defined as the label of each piece of training data.

[0036] In this embodiment, one piece of annotated training data is: "Time window: 2025-06-02 10:00:00 to 10:10:00; Scheduled event sequence: [Event 1: Type = Task Start, Timestamp = 10:00:05.123, Resource Identifier = Node_001; Event 2: Type = Resource Expansion, Timestamp = 10:05:30.456, Resource Identifier = Node_002]; Resource state slices: [Time point = 10:00:00, CPU utilization = 40.1%, memory usage = 6.0 GB, network bandwidth = 100.0 Mbps, storage IO = 500 IOPS; Time point = 10:00:01, CPU utilization = 40.3%, memory usage = 6.0 GB,... (uniformly sampled to 10:10:00)]; Key state anchor point: Event 1 triggers CPU utilization to jump; Mode classification label: elastic expansion and contraction".

[0037] S4, pre-process the training data of the computing resource scheduling mode recognition: S401, perform data cleaning operation, eliminate repeated event records due to abnormal collection, fill in missing values of resource state caused by sensor failure, and the specific method can adopt linear interpolation of previous and subsequent time points, and filter isolated events without resource change association.

[0038] S402, perform feature normalization, respectively perform Z-score standardization on resource state sequence according to resource type, eliminate dimension difference, and use frequency statistics weighting on event type to reduce the dominant influence of high-frequency events on the model.

[0039] S403, implement sample balancing processing, increase the sample amount of rare mode by time window sliding resampling, generate synthetic samples by using SMOTE sampling technology, and alleviate the class imbalance problem in mode recognition task.

[0040] S5, perform spatio-temporal joint feature coding and dynamic feature enhancement of training data.

[0041] The scheduling log data exists in the form of discrete event sequence, and the resource state data exists in the form of continuous time sequence. The two types of data are heterogeneous in format, granularity and expression mechanism. The conventional processing method such as directly splicing event and state vector or separately coding and then fusing cannot effectively model the synchronous response relationship between event triggering behavior and resource state change, and thus cannot accurately depict the influence mechanism of event on resource dynamics.

[0042] The method for spatio-temporal joint feature coding of training data in the embodiment is: constructing a spatio-temporal joint encoder, obtaining a low-dimensional vector expression of discrete event type through event embedding, simultaneously capturing resource instantaneous response trend based on time series differentiation of resource state tensor at event occurrence time, and then constructing a joint feature representation across scheduling events. The process takes event type identifier as input, generates event embedding vector through category embedding layer, and splices the event embedding vector with resource state differentiation at the corresponding time point to form a joint feature vector, which is used for subsequent task modeling, and is represented as: , In the above formula, is the joint feature vector of the i th scheduling event, which is a comprehensive representation of event type and resource state change, is a category embedding layer, represents mapping discrete event type to a d-dimensional vector, capturing the semantic features of the event, such as task start, resource expansion, etc. is the dimension of the event embedding vector, is the joint feature vector of the i th scheduling event, which is a comprehensive representation of event type and resource state change, is a category embedding layer, represents mapping discrete event type to a d-dimensional vector, capturing the semantic features of the event, such as task start, resource expansion, etc. a type identifier of a scheduling event, such as "task start" or "resource scaling", for vector concatenation, for the differential of the resource state tensor at the time point of the event occurrence, reflecting the instantaneous resource change trend triggered by the event, that is, characterizing the change trend of the resource state at the time point of the event occurrence, for the resource state tensor at the time , and , describing the state change of the resource within the time window, describing the state change of the resource within the time window, for the number of resource types, for the time window, denotes the parameter dimension, denotes assigning to , for the specific time point of the occurrence of the th scheduling event.

[0043] In this embodiment, the annotated training data is used for spatio-temporal joint feature encoding calculation, such as: time window: 2025-06-0210:00:00 to 10:10:00, event sequence contains event 1 and event 2, and the label is elastic scaling.

[0044] In this embodiment, the specific process of spatio-temporal joint feature encoding of the training data is as follows: 1) Extract the event type identifier for each scheduling event, such as the type of event 1 is "task start" and the type of event 2 is "resource scaling"; 2) Map the discrete event type to a low-dimensional vector representation through a category embedding layer: The "task start" type is mapped to a fixed-dimensional vector with dimensions, which captures the event semantic features, and at the same time, the differential of the resource state tensor is calculated at the time point of the event occurrence, such as the timestamp of event 1 is 10:00:05.123, which captures the instantaneous change trend of the CPU utilization, memory occupation and other indicators at that moment, such as the differential value of the CPU utilization from 40.1% to 60.5%, reflecting the impact of the event on the resource state.

[0045] 3) Concatenate the event embedding vector with the resource state differential vector at the corresponding time point to form a joint feature vector, such as for event 1, its event embedding vector is concatenated with the resource state differential at 10:00:05.123 to generate a fusion vector, and event 2 is processed in the same way.

[0046] 4) Output the joint feature vector of each event as the base element of the subsequent model input.

[0047] Scheduling behavior has characteristics such as uneven event density and strong state signal noise. Traditional masking strategies based on fixed probability or random position are prone to破坏关键事件链条结构,尤其在低频事件和异常模式识别中存在显著性能瓶颈。

[0048] The method for dynamically enhancing the features of training data in this embodiment is as follows: An adaptive masking mechanism guided by event importance is adopted. After performing layer normalization on each joint feature vector, an importance masking vector is generated through learnable parameters. According to the masking values, the original feature vector and the batch mean vector are fused to achieve enhanced expression of rare but important event features, thereby improving the model's ability to model unbalanced event distributions, expressed as: , , In the above formula, [[ID=​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​The core components of the deep neural network include a plurality of cascaded double-scale gating units and a hybrid attention mechanism. The plurality of cascaded double-scale gating units are used to process time series dependencies, and the hybrid attention mechanism is used to fuse heterogeneous event and state features, so as to ensure that the model can capture the spatiotemporal dynamics of scheduling behaviors.

[0053] Specifically, the double-scale gating unit is used as a basic module, and the information flow is adjusted through reset gate and update gate mechanisms. The unit uses the differential amplitude of the resource state as a transient change intensity coefficient, dynamically balances long-term resource evolution and short-term event impact, and outputs a hidden state sequence.

[0054] According to the hybrid attention mechanism, the event sequence and the resource state time series are processed respectively. The event attention path extracts key events such as high-frequency task start, and the state attention path focuses on resource mutation points such as CPU utilization jump. After the two paths are spliced, the global context features are formed through a multi-layer perception fusion.

[0055] The model output layer adopts a full connection network to map the global features output by the hybrid attention to a predefined mode category, such as elastic scaling and load balancing migration.

[0056] The deep neural network combines double regularization constraints to suppress overfitting during training.

[0057] S602, define a gated recurrent unit; Traditional gated recurrent structures such as GRU have difficulty in simultaneously processing long-term evolution of resource states and transient impact of scheduling events when modeling scheduling behaviors, resulting in insufficient capture of multi-time scale features.

[0058] In this embodiment, a double-scale gating mechanism is adopted, the differential amplitude of the resource state is used as a transient change intensity coefficient, the response weights of the reset gate and the update gate to the features are dynamically adjusted, at each time step, the gating signal is generated by using the previous hidden state and the current enhanced feature vector, and the transient change intensity is used to adjust the gating activation amplitude, the historical dependence and the short-term impact are balanced, and the updated hidden state is generated, which is represented as: , , , , , In the above formula, is the reset gate vector of the time step , which is used to control the information retention degree of the previous hidden state, is a Sigmoid activation function, is the weight matrix of the reset gate, is the hidden state of time step , is the enhanced feature vector of time step , is the instantaneous change intensity coefficient, used to dynamically adjust the activation amplitude of the gating signal, is the bias vector of the reset gate.

[0059] is the update gate vector of time step , used to control the degree of introduction of new information, is the weight matrix of the update gate, is the bias vector of the update gate.

[0060] is the candidate hidden state vector of time step , is the hyperbolic tangent function, is the weight matrix of the candidate hidden state, is the element-wise multiplication.

[0061] is the hidden state of time step .

[0062] is the L2 norm, is the maximum value of the L2 norm of the resource state differential tensor at all time points within the entire time window, is the differential of the resource state tensor at the event occurrence time point, is the resource state tensor at time .

[0063] S603, power resource scheduling mode recognition model weight initialization.

[0064] Because the scheduling behavior data has high complexity in both the resource dimension and the time dimension, if the initialization is unreasonable, the deep network training process is prone to fall into local optimum, which is manifested as slow training convergence and decreased generalization ability.

[0065] The embodiment adopts the spectral characteristics of the feature covariance matrix to design the initialization strategy, extracts the joint features of the training samples, constructs the covariance matrix thereof, calculates the maximum and minimum eigenvalues, and sets the weight initialization range accordingly, so that the initialized weights are closer to the sample distribution structure, thereby improving the network convergence efficiency and stability, which is expressed as: , In the above formula, is the weight matrix of the fully connected layer, obeys the distribution, is a uniform distribution, The largest eigenvalue of the characteristic covariance matrix. is the smallest eigenvalue of the characteristic covariance matrix.

[0066] S604, constructing a hybrid attention mechanism.

[0067] 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.

[0068] 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: , , , , , 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.

[0069] The output vector of event attention. For the first A vector of values ​​for each event.

[0070] For time step Attention weights in state attention mechanisms This is the transpose of the state query vector. For time step state key vector, For time steps 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.

[0071] The output vector of state attention. For time steps The state value vector.

[0072] This is the final output vector of the hybrid attention mechanism. This represents a multilayer perceptron.

[0073] S605, double regularization constraint; 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.

[0074] 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: , 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.

[0075] 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.

[0076] S606 performs forward data propagation; 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.

[0077] 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.

[0078] 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.

[0079] S607, Dynamic Course Learning Scheduling.

[0080] 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.

[0081] 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: , , 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 data set of the wheel, for the first the training data set of the wheel, the information entropy of the sample measures the complexity of the sample , is the training sample, is the dynamic threshold, and the calculation method is represented as ; is the cumulative entropy distribution function.

[0082] S608, error back propagation is performed.

[0083] Calculate the error between the global context feature expression output by the model and the labeled scheduling mode, which reflects the accuracy of the model in identifying the scheduling mode.

[0084] The error is back propagated through the hybrid attention mechanism module to adjust the weights of event attention and state attention, optimize the attention output, and enable it to more accurately extract key events and highlight key time point resource changes.

[0085] The error continues to be back propagated to the double-scale gated recurrent unit and the spatio-temporal joint encoder to update 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.

[0086] S609, stop iteration condition judgment: when the verification set accuracy improvement amplitude is less than 0.5% in the last 5 training cycles, or the total training period reaches the preset upper limit, terminate the training, save the model parameters with the best verification set performance, and prevent overfitting; S7, perform computing resource scheduling mode recognition.

[0087] Perform data preprocessing operations on newly collected raw scheduling logs and resource state data, including data cleaning and feature normalization.

[0088] Perform feature encoding and enhancement, i.e., use the spatio-temporal joint feature encoder to process event sequences and resource states to generate joint feature vectors, and use the dynamic feature enhancement mechanism to adjust the vectors based on importance masks to highlight key event features.

[0089] The preprocessed enhanced feature vector sequence is input into the trained pattern recognition model, and the model performs forward propagation, i.e., processes the sequence through the double-scale gated unit to capture event-driven resource dynamics, the hybrid attention mechanism to fuse event and state features, and the global context to output the probability distribution of the mode category.

[0090] According to the output probability, the scheduling mode category such as "elastic scaling" or "failover reorganization" is determined, and the identification result is fed back to the scheduling system for real-time optimization of resource allocation.

[0091] The experiment verification of the embodiment is as follows: As Figure 2 shown, in order to verify the robustness advantage of the double-scale gating unit in the dynamic scheduling environment, five typical scenes with progressive complexity are simulated, which are low-load stable period, medium-load fluctuation period, high-load peak period, sudden failure recovery period and mixed mode complex period. The pattern recognition accuracy of the traditional gating recurrent unit, the long short-term memory network, the transformer and the double-scale gating unit of the embodiment is compared, Figure 2 The background gray scale gradient intuitively reflects the increase of scene complexity, and the curve trend reflects the degree of performance decay of the model with the deterioration of the environment. The experimental results show that the traditional method has a cliff performance drop in the sudden failure and mixed mode scene, while the curve of the embodiment always maintains a gentle downward trend, and still maintains a significant advantage in the high-load peak period, indicating that the double-scale gating mechanism takes the resource state differential amplitude as the instantaneous intensity coefficient, dynamically adjusts the activation weight of the reset gate and the update gate, so that the model can adaptively balance the modeling needs of long-term resource evolution and short-term event impact.

[0092] As Figure 3 shown, in order to analyze the effect of the dynamic feature enhancement mechanism on the identification of rare scheduling modes, four low-frequency modes are selected, which are failover reorganization, cross-cluster migration, resource isolation and emergency expansion. The identification performance of three strategies of no feature enhancement, random mask and the adaptive mask of the embodiment is compared, Figure 3 The cylinder height represents the harmonic mean, and the performance of the method of the embodiment on the four types of rare modes is significantly better than that of the control group. In particular, the advantage is most prominent in the emergency expansion mode with the lowest frequency. The random mask strategy even performs worse than the no enhancement method in the resource isolation mode, proving that blind masking can destroy the key event chain, and the performance improvement amplitude is negatively correlated with the mode frequency. The experimental results show that by generating importance weights through learnable parameters, the expression of rare event features is enhanced while the integrity of the key event sequence is protected, fundamentally solving the modeling defects of traditional methods for unbalanced event distribution.

[0093] As Figure 4As shown, by analyzing the accuracy rate distribution characteristics of different algorithms in multiple tests through kernel density estimation curve, the stability and reliability of the algorithm recognition are verified, the horizontal axis represents the accuracy rate percentage, the vertical axis represents the probability density, the curve peak reflects the most common accuracy level, and the curve width reflects the fluctuation range, the corresponding blue curve of the present application presents a high and narrow distribution form, indicating that the accuracy rate is highly concentrated in the high level interval, and the fluctuation range is extremely small, compared with other comparative algorithms, the distribution curve is more flat and left, showing that the mean accuracy rate is low and the fluctuation is large, especially the curve of the feature splicing method presents a significant right bias form, there are a large number of low accuracy rate outliers, which reflects the deep fusion ability of the spatio-temporal joint encoder to heterogeneous data, and the strengthening effect of the dynamic feature enhancement mechanism on the key event features, ensuring stable and superior performance under various operating conditions.

[0094] As shown in the formula (I), in order to evaluate the robustness of the algorithm in four typical scheduling scenarios, the experiment adopts the way of box plot combined with scatter plot, the box position reflects the median of the accuracy rate, the box height represents the data dispersion degree, and the scatter distribution shows the original test sample. Figure 5 In the conventional task scenario, all algorithms perform well, but in high complexity scenarios such as burst traffic and fault transfer, the present application can still maintain a compact high box, the median is more than 90%, and the scatter is concentrated in the upper quartile area, while the time sequence long short-term memory network and event gate recurrent unit appear a large box in the fault transfer scenario, the median is less than 70%, and the scatter is vertically diffused, verifying the role of the double-scale gate unit, dynamically adjusting the information flow through the resource state differential amplitude, accurately balancing long-term resource evolution and short-term event impact, so that the model can still maintain stable recognition ability in extreme scenarios, overcoming the performance decay problem of traditional methods in complex scheduling modes.

[0095] The above is only a preferred embodiment of the present application, and is not used to limit the protection scope of the present application.

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; 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 data forward propagation; S607, Dynamic Course Learning Scheduling; 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.

10. The computing resource scheduling pattern recognition method according to claim 1, characterized in that, 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 changes in verification accuracy. The probability of difficulty increase is calculated using the Sigmoid activation function. 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. This is expressed 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: , This is the cumulative entropy distribution function.

Images