Artificial intelligence-based department data operation management method and device, and medium

Abstract

The invention discloses a department data operation management method and device based on artificial intelligence and a medium, and belongs to the technical field of data operation management, and the method specifically comprises the steps: firstly collecting department historical concurrent requests and resource consumption data, and precisely splitting the data into two groups of data pairs of smooth operation and burst pressure according to an instantaneous arrival rate; then, respectively extracting key features of the two to construct a long-tail memory matrix and a conventional baseline vector; in decision model training, parameters are innovatively updated based on burst data, and then gradients generated by stationary data are projected to an orthogonal null space of a long-tail memory matrix for iteration, so that disastrous forgetting of the model to long-tail burst features is effectively prevented; and finally, inputting a real-time request into the model, matching in combination with the long-tail memory, outputting a resource load prediction tensor, and generating a scheduling instruction to drive a bottom layer node to execute elastic capacity expansion and contraction according to the tensor.

Description

Technical Field

[0001] This invention relates to the field of data operation and management technology, specifically to a method, equipment, and medium for departmental data operation and management based on artificial intelligence. Background Technology

[0002] The rapid popularization and in-depth development of medical informatization have made the daily operations of various hospital departments highly dependent on underlying data systems and network computing resources. To improve the response efficiency and resource utilization of medical services, artificial intelligence-based departmental data operation management methods are gradually becoming a core technological means for the construction of modern smart hospitals. Existing management methods typically rely on big data analysis and deep learning algorithms. This involves deploying data collection probes at departmental gateways and computing nodes to continuously aggregate large-scale business interaction logs, electronic medical record transfer records, and underlying server operating consumption indicators. After acquiring this multi-dimensional data, existing technologies primarily utilize historical stable data accumulated under normal departmental operation to continuously train and fit neural network models. This allows for periodic predictions of the departmental system's routine resource load trends, and based on this, basic computing power allocation and network scheduling are performed to maintain the orderly operation of the department's daily business systems.

[0003] However, existing AI-powered departmental operation management methods heavily rely on historical, stable data distributions, lacking specific feature retention mechanisms for long-tail, low-frequency, abnormal operational events. During the training or fine-tuning phase of the operational decision-making model, massive amounts of daily stable operational data severely suppress the gradient updates of features during sudden resource surges, causing the neural network to catastrophically forget extreme concurrent load scenarios. This imbalance in feature weights makes the model exhibit severe predictive sluggishness when facing sudden increases in complex business concurrent requests, failing to plan ahead and output flexible emergency resource scheduling instructions, ultimately leading to instantaneous operational paralysis of the department under pressure. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device, and medium for departmental data operation and management based on artificial intelligence, thereby solving the problems in the background art: The objective of this invention can be achieved through the following technical solutions: An artificial intelligence-based method for departmental data operation and management. It includes the following steps: S1. Collect concurrent request data and corresponding resource consumption data from the department system for historical periods, and bind time tags to the concurrent request data and resource consumption data; S2. Calculate the instantaneous arrival rate of concurrent request data based on the time tag. Based on the threshold of the instantaneous arrival rate, synchronously split the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs. S3. Extract the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extract the periodic characteristics of stable operation data pairs to construct a regular baseline vector. S4. During the training phase of the decision model, the parameters are first updated based on the sudden pressure data to calculate the first update gradient; then, when backpropagating using the stable operation data, the generated second update gradient is projected onto the orthogonal null space of the long-tail memory matrix for parameter iteration, thus obtaining the trained decision model. S5. Collect the current concurrent request data of the department system in real time, input it into the trained decision model, combine it with the long-tail memory matrix for feature matching, and output the resource load prediction tensor of the current concurrent request data. S6. Generate emergency resource scheduling instructions based on the resource load prediction tensor, and send the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling up and down.

[0005] As a further aspect of the present invention: In step S1, the process of collecting concurrent request data and corresponding resource consumption data from historical periods of the department system, and binding time tags to the concurrent request data and resource consumption data, is as follows: Intercept business messages received by the department system gateway during historical time periods and parse the business messages into concurrent request data; Based on the task scheduling records of business messages, extract the underlying running logs of task execution nodes to generate corresponding resource consumption data; Read the standard clock stamp when the business message enters the gateway, and bind the standard clock stamp as a timestamp to the concurrent request data and resource consumption data.

[0006] As a further aspect of the present invention: In step S2, the process of calculating the instantaneous arrival rate of concurrent request data based on the time stamp, and synchronously splitting the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs based on the threshold of the instantaneous arrival rate, is as follows: Divide the time window according to the time label, count the total number of concurrent data requests within the time window, and divide the total number of requests by the time window span to obtain the instantaneous arrival rate. The instantaneous arrival rate is compared with the instantaneous arrival rate threshold. When the instantaneous arrival rate is greater than the instantaneous arrival rate threshold, a burst flag is generated for the corresponding time window; otherwise, a stable flag is generated. Based on the sudden and stable indicators, extract the concurrent request data and the associated resource consumption data within the corresponding time window, and combine and split them into sudden pressure data pairs and stable operation data pairs.

[0007] As a further aspect of the present invention: in step S3, the process of extracting the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extracting the periodic characteristics of stable operation data pairs to construct a conventional baseline vector is as follows: For sudden pressure data, perform sliding window difference operation to extract time series fluctuation features, and obtain concurrency features through fully connected layer mapping operation; The temporal fluctuation features and concurrency features are concatenated into tensors, and the principal component feature array of the concatenated tensors is extracted to construct a long-tailed memory matrix; For stable operation data, perform discrete Fourier transform to generate frequency domain data, and extract the core frequency components in the frequency domain data as periodic features; A smoothed value is obtained by performing a moving average calculation on the periodic features along the time axis, and the smoothed values ​​are arranged in sequence to construct a regular baseline vector.

[0008] As a further aspect of the present invention: in step S4, the process of updating parameters based on the sudden pressure data to calculate the first update gradient during the training phase of the decision model is as follows: Extract concurrent request data from sudden pressure data pairs and input them into the decision model to perform forward computation, outputting an initial resource prediction tensor; Substitute the resource consumption data from the initial resource prediction tensor and the sudden pressure data pair into the loss function to calculate the sudden pressure error value. The first update gradient is obtained by performing back-derivative on the parameters of the decision model based on the sudden pressure error value, and the parameters of the decision model are updated according to the first update gradient.

[0009] As a further aspect of the present invention: In step S4, the process of projecting the generated second update gradient onto the orthogonal null space of the long-tailed memory matrix for parameter iteration during backpropagation using stationary operating data to obtain the trained decision model is as follows: The decision model with updated input parameters based on stable operating data is subjected to forward inference and backward differentiation to calculate and generate the second update gradient. Calculate the eigenvectors of the long-tailed memory matrix to construct an orthogonal projection matrix, and multiply the second update gradient with the orthogonal projection matrix to generate an orthogonal gradient tensor; The parameters of the decision model are summed and iterated based on the orthogonal gradient tensor to generate the trained decision model.

[0010] As a further aspect of the present invention: in step S5, the process of outputting the resource load prediction tensor of the current concurrent request data is as follows: The system analyzes departmental system messages in real time to generate current concurrent request data, and inputs it into the trained decision model to extract current request features. Calculate the inner product of the current request features and the long-tail memory matrix to generate matching weights, and extract the data of the long-tail memory matrix based on the matching weights to construct a feature matching tensor; The feature matching tensor is concatenated with the current request features and substituted into the prediction layer of the decision model to calculate and output the resource load prediction tensor of the current concurrent request data.

[0011] As a further aspect of the present invention: in step S6, the process of generating emergency resource scheduling instructions according to the resource load prediction tensor and issuing the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling is as follows: The predicted resource demand is extracted by parsing the resource load prediction tensor. The predicted resource demand is then subtracted from the current running resource quantity to generate a resource difference vector. Based on the pre-set scheduling action template of the resource difference vector retrieval system, the resource difference vector and the scheduling action template are compiled to generate emergency resource scheduling instructions; Emergency resource dispatch instructions are sent to the resource control nodes of the department system, and the resource control nodes perform flexible scaling up and down according to the emergency resource dispatch instructions.

[0012] On the other hand, an artificial intelligence-based departmental data operation and management device is also provided, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method as described in any one of claims 1-8.

[0013] Furthermore, a non-volatile computer storage medium is also provided, storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: Collect concurrent request data and corresponding resource consumption data from historical periods in the department system, and bind time tags to the concurrent request data and resource consumption data; The instantaneous arrival rate of concurrent request data is calculated based on the time stamp. Based on the threshold of the instantaneous arrival rate, the concurrent request data and resource consumption data are synchronously split into stable operation data pairs and sudden pressure data pairs. Extract the temporal fluctuation and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extract the periodic characteristics of stable operation data pairs to construct a regular baseline vector. During the training phase of the decision model, the parameters are first updated based on the sudden stress data to calculate the first update gradient; then, when backpropagating using stable operation data, the generated second update gradient is projected onto the orthogonal null space of the long-tail memory matrix for parameter iteration, thus obtaining the trained decision model. Real-time collection of current concurrent request data from the department system is input into the trained decision model. Feature matching is performed using the long-tailed memory matrix, and the resource load prediction tensor of the current concurrent request data is output. Emergency resource scheduling instructions are generated based on the resource load prediction tensor, and then sent to the resource control nodes of the departmental system for elastic scaling up and down.

[0014] The beneficial effects of this invention are: This invention fundamentally solves the problem of feature weight imbalance during model training by precisely separating sudden pressure data from stable operational data and constructing a dedicated long-tail memory matrix accordingly. Particularly in the decision model training phase, this invention innovatively introduces an orthogonal gradient projection mechanism. When backpropagating using massive amounts of stable data, the resulting second update gradient is forcibly projected onto the orthogonal null space of the long-tail memory matrix for parameter iteration. This mechanism mathematically isolates the impact of stable data on the gradients of sudden features, allowing the model to grasp daily cyclical operational patterns while perfectly avoiding catastrophic forgetting of extreme concurrent load scenarios. This achieves permanent and sensitive exclusive preservation of the characteristics of long-tail, low-frequency abnormal operational events.

[0015] Thanks to the effective solidification of long-tail burst features, this invention significantly eliminates the predictive sluggishness of the model when facing sudden surges in complex requests, completely breaking the passive situation where departmental systems are prone to instantaneous paralysis under high-pressure scenarios. During real-time operation, the model matches the current request features with the long-tail memory matrix, enabling it to rapidly recall feature memories for handling extreme loads, accurately capture precursors to resource surges, and output high-precision resource load prediction tensors in advance. This allows the system to pre-generate and issue emergency resource scheduling instructions before business surges substantially impact the underlying computing power, driving resource control nodes to complete pre-emptive elastic scaling of physical resources, thereby endowing the departmental data system with high availability and strong resilience under extreme pressure scenarios. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart illustrating a departmental data operation and management method based on artificial intelligence according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a departmental data operation and management method based on artificial intelligence, comprising the following steps: S1. Collect concurrent request data and corresponding resource consumption data from the department system for historical periods, and bind time tags to the concurrent request data and resource consumption data; S2. Calculate the instantaneous arrival rate of concurrent request data based on the time tag. Based on the threshold of the instantaneous arrival rate, synchronously split the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs. S3. Extract the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extract the periodic characteristics of stable operation data pairs to construct a regular baseline vector. S4. During the training phase of the decision model, the parameters are first updated based on the sudden pressure data to calculate the first update gradient; then, when backpropagating using the stable operation data, the generated second update gradient is projected onto the orthogonal null space of the long-tail memory matrix for parameter iteration, thus obtaining the trained decision model. S5. Collect the current concurrent request data of the department system in real time, input it into the trained decision model, combine it with the long-tail memory matrix for feature matching, and output the resource load prediction tensor of the current concurrent request data. S6. Generate emergency resource scheduling instructions based on the resource load prediction tensor, and send the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling up and down.

[0020] In one embodiment of the present invention, step S1, which involves collecting concurrent request data and corresponding resource consumption data from historical time periods of the department system, and binding time tags to the concurrent request data and resource consumption data, is as follows: In a specific embodiment, the gateway serves as the unified entry point for all traffic within the department, carrying all business interaction data. To intercept business packets received by the gateway during historical periods, a traffic mirroring mechanism is deployed at the gateway node. This mechanism copies the incoming data packets completely through bypass replication, allowing the acquisition of the most original interaction records without affecting backbone business transmission and communication. These intercepted business packets contain actual operational instructions from doctors accessing electronic medical records or nurses uploading vital sign data, while also including a large amount of underlying communication protocol headers and redundant control information. Because the original business packets are unstructured and highly mixed, they cannot be directly used for concurrent analysis; therefore, they need to be parsed into concurrent request data. The specific parsing process employs deep packet inspection technology, peeling away the network and transport layer envelopes of the business packets layer by layer to extract the pure application layer payload content. By identifying the request initiator address, target service interface, and specific operational actions within the application layer payload, the originally complex packets are transformed into structured request entries. Each parsed concurrent request triggers a series of computational tasks, which are then assigned by the scheduling center to specific computing nodes for execution. Based on the task scheduling records of the business messages, the specific physical or virtual machine handling the request can be precisely traced. The task scheduling record is similar to a delivery slip, detailing the complete routing path of the task from the gateway to the specific execution terminal. After locating the task execution node along this path, the underlying runtime logs of that node are extracted to generate corresponding resource consumption data. Specifically, a lightweight acquisition probe is activated on the located task execution node, and the probe reads the underlying runtime logs generated by the operating system kernel within the task processing time window. These underlying runtime logs faithfully record the dynamic heartbeat changes of the hardware, covering the CPU computation cycles used during task execution, the allocated and released random access memory capacity, and the hard drive's I / O I / O throughput. By aggregating and statistically analyzing these hardware operating indicators recorded in the logs, the physical cost incurred in completing the task can be quantified.

[0021] When a network packet arrives at the gateway's first contact point, the device immediately timestamps it. Reading the standard clock stamp of the service packet entering the gateway requires capturing the precise time generated by a high-precision hardware oscillator and globally synchronized. This standard clock stamp ensures absolute time alignment throughout the environment, avoiding time drift caused by discrepancies in the internal clocks of different computing nodes. After obtaining the absolutely accurate time, the standard clock stamp is bound as a timestamp to concurrent request data and resource consumption data. The binding process establishes a unique time anchor for cause and effect, where concurrent request data represents the cause triggering the computation, and resource consumption data represents the objective result of the computation. By forcibly attaching the same standard clock stamp as a universal marker to these two sets of data, a strict time sequence correspondence is established.

[0022] In one embodiment of the present invention, in step S2, the process of calculating the instantaneous arrival rate of concurrent request data based on the time stamp, and then, based on a threshold of the instantaneous arrival rate, synchronously splitting the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs, is as follows: After acquiring concurrent request data bound with absolute time-series features, time windows need to be divided based on time tags to construct discrete observation intervals. The principle of time window division is to set a fixed-length observation step on a continuously flowing timeline, thereby truncating the uncountable data stream into multiple finite independent data blocks to facilitate fine-grained measurement of local features. For example, if the observation step is set to 5 minutes, starting from 8:00 AM, multiple consecutive and non-overlapping time windows will be successively divided, such as 8:00 AM to 8:05 AM and 8:05 AM to 8:10 AM. After the time windows are divided, the total number of concurrent request data within each time window is counted. The specific statistical principle is to check the time tag attached to each data point one by one and assign it to the corresponding time window, then accumulate them to obtain the absolute number of requests occurring within that interval. For example, if a total of 500 registration query requests carrying this time segment tag are included in the specific time window from 8:00 AM to 8:05 AM, then the total number of requests in this time window is 500. To eliminate the bias caused by different observation intervals in the pressure assessment, the total number of requests is divided by the time window span to obtain the instantaneous arrival rate.

[0023] After accurately quantifying the intensity of concurrent service inflows for each independent observation interval, the instantaneous arrival rate needs to be compared with the instantaneous arrival rate threshold to define the current service operation status. The instantaneous arrival rate threshold represents the critical warning level of service capacity under normal conditions. The principle of comparing the two is to identify abnormal peaks exceeding normal capacity expectations through direct numerical comparison, because only data exceeding this critical level has the destructive power to impact the underlying computing architecture. After completing the comparison calculation, when the instantaneous arrival rate is greater than the instantaneous arrival rate threshold, a burst identifier is generated for the corresponding time window. The principle of generating burst identifiers is to use the idea of ​​binary classification to separate potentially congestion-risk dangerous intervals from the normal time axis and mark them with a high-risk mark in order to accurately locate the anchor point of abnormal pressure.

[0024] As each observation interval on the timeline is marked with a clear state identifier, it is necessary to further extract concurrent request data and associated resource consumption data within the corresponding time window based on burst and stable indicators. The extraction principle is to use the aforementioned state identifier as a data filter to collect all original requests with time tags belonging to a specific identifier window and their resulting underlying hardware overhead. Because concurrent request data and resource consumption data are tightly bound together by time tags in the initial stage, extraction based on time window identifiers can capture the causes of stress and the results of stress absorption. After completing the refined extraction operation, these data are combined and split into burst stress data pairs and stable operation data pairs. The principle behind combining and splitting these two types of data pairs is based on the physical isolation of the identifiers, which spatially separates the originally mixed daily mild traffic and abnormal peak traffic and solidifies them in the form of cause-and-effect pairings. Causes and results with burst identifiers are combined and assembled into burst stress data pairs, while causes and results with stable identifiers are assembled into stable operation data pairs. Through this identifier-based physical isolation and pairing assembly operation, the originally disordered massive interaction logs are deconstructed into two sets of independent materials with different business characteristics and attributes, thus providing a basis for differentiated feature processing for different traffic patterns.

[0025] In one embodiment of the present invention, step S3 involves extracting the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extracting the periodic characteristics of stable operation data pairs to construct a conventional baseline vector. After acquiring the burst pressure data pairs, it is necessary to perform sliding window differential operations to extract time-series fluctuation characteristics. The principle of sliding window differential operations is to set two adjacent observation intervals of fixed size on the time axis, and subtract the corresponding values ​​of the previous interval from the resource and request values ​​of the latter interval to calculate the gradient of data change within a unit time span. The reason why time-series fluctuation characteristics can be obtained through differential operations is that the absolute number of concurrent requests often cannot truly reflect the burst speed of pressure, while the difference between adjacent time windows precisely represents the acceleration of traffic inflow and the intensity of resource consumption. For example, if the concurrency in the first 5 minutes is 100 and the concurrency in the current 5 minutes jumps to 500, the 400 obtained by differential operations represents an extremely strong upward slope. This slope indicator that reflects the intensity of change is the time-series fluctuation characteristic. While extracting the temporal fluctuation features, it is also necessary to derive concurrency features through fully connected layer mapping operations. The specific approach of fully connected layer mapping operations is to construct a multi-layer artificial neural network structure, input various basic hardware indicators contained in the sudden pressure data pairs into this network structure, and connect each input node to all hidden nodes in the next layer to perform weight multiplication and bias addition calculations. The reason why concurrency features can be derived through fully connected layer mapping operations is that there is a complex non-linear coupling relationship between the original CPU utilization and memory consumption. The dense multiplication and addition operations of the fully connected layer can map and transform these low-dimensional isolated data into a high-dimensional abstract space. In this high-dimensional space, the originally hidden correlation patterns are amplified and extracted. For example, the concurrency pattern of multiple service interfaces encountering high-frequency calls at the same time can be abstracted into a set of high-dimensional vector features, thus forming concurrency features.

[0026] After obtaining the features reflecting the slope of time-series changes and the features reflecting high-dimensional spatial coupling, it is necessary to perform tensor concatenation of the temporal fluctuation features and the concurrency features. Tensor concatenation involves combining feature vectors representing different dimensions along specified coordinate axes in a mathematical matrix. For example, a temporal fluctuation feature vector of length 128 and a concurrency feature vector of length 256 are concatenated end-to-end to merge into a new comprehensive vector of length 384. Tensor concatenation is necessary because a single temporal change rate or a single spatial concurrency state cannot fully describe the true destructive power of sudden flows. Only by concatenating the features of both time and space dimensions into a single comprehensive tensor can a complete, multi-faceted characterization of sudden stress events be formed. After feature fusion, it is necessary to extract the principal component feature array of the spliced ​​tensor to construct a long-tailed memory matrix. The method for extracting the principal component feature array is to perform an orthogonal transformation on the massive spliced ​​tensor, calculate the covariance matrix of the tensor, and solve for its eigenvalues ​​and eigenvectors. Then, the top few core eigenvectors are selected in descending order of eigenvalues ​​and combined into an array. The reason why the long-tailed memory matrix can be obtained through this principal component extraction method is that sudden stress events are often accompanied by a large amount of random noise and useless interference. Orthogonal transformation can project the data onto several mutually orthogonal directions with the largest variance, thereby eliminating redundant low-value information and retaining the core skeleton that best represents the extreme sudden mode. The long-tailed memory matrix is ​​a mathematical structure composed of these purified core skeletons. It is specifically used to persistently store the extreme traffic characteristics and patterns that occur with extremely low frequency but have extremely high destructive power, ensuring that the corresponding resource carrying mode can be quickly recalled when facing similar rare concurrent shocks.

[0027] After processing the abnormal data representing extreme situations, the focus shifts to processing traffic records under normal, stable conditions. This involves performing a Discrete Fourier Transform (DFT) on the stable operational data to generate frequency domain data. The DFT operation uses a series of sine and cosine waves of different frequencies to fit the daily concurrent request volume and resource consumption records that fluctuate over time, transforming the original chronological line graph into a spectrum graph with frequency as the horizontal axis. After generating the frequency domain data, the core frequency components need to be extracted as periodic features. This is achieved by setting an amplitude threshold on the generated spectrum graph, retaining only a few specific frequencies with peak amplitudes far exceeding the average level, and filtering out noisy high-frequency signals representing random, minor fluctuations. After accurately capturing the cyclical patterns of the department's daily operations, it is necessary to perform a moving average calculation on the periodic characteristics along the time axis to obtain smoothed values. The specific operation of the moving average calculation is to set an observation window with a fixed span and let this window slide gradually from front to back on the data sequence of the periodic characteristics. Each slide calculates the arithmetic mean of all periodic values ​​within the current window and uses this mean to replace the original value at the center of the window. Through this calculation method, the original fluctuating periodic characteristics are transformed into a smooth and stable continuous stream of values. These value streams filter out all random disturbances and retain only the most solid macroscopic trends. After obtaining the smoothed values, the smoothed values ​​are arranged in sequence to construct a regular baseline vector. The method of constructing the baseline vector is to fill a one-dimensional mathematical array with each average result calculated over time in chronological order. The reason why a regular baseline vector can be obtained by arranging them in sequence is that these smoothed values, which are strictly ordered by time, together depict an ideal reference trajectory representing the absolute stable state of the department's business. When these values ​​are solidified into a vector, it is equivalent to drawing a standard scale line in the complex daily traffic changes. This baseline vector will serve as an absolute reference criterion for measuring whether the current traffic has deviated from the normal track in future resource assessments.

[0028] In one embodiment of the present invention, in step S4, during the training phase of the decision model, the process of updating the parameters based on the sudden pressure data to calculate the first update gradient is as follows: When training the decision model, it's crucial to first process rare samples representing extreme load conditions to allow the model to prioritize crisis management. Concurrent request data from burst stress data pairs is extracted and input into the decision model for forward computation. Specifically, this extraction involves separately identifying the causal components that trigger a surge in computing power from the burst stress data pairs that have already been stripped of regular traffic. These causal components include scenarios such as hundreds of doctors simultaneously initiating 3D image reconstruction requests or thousands of monitoring devices concurrently uploading high-frequency vital sign data. These numerical matrices representing extreme concurrency states are then used as initial excitation signals input into the input layer nodes of the decision model. The principle behind the forward computation of the input decision model is to utilize a multi-layered artificial neural network structure built within the model to perform layer-by-layer information transmission and feature transformation on the input signal. When concurrent request data enters the first layer of neurons, each node multiplies it with its own initial weights and adds a bias constant. Then, nonlinear activation processing is used to pass the result to the next layer of neurons. This dense matrix multiplication and addition operation, which progresses layer by layer from the input to the output, is the forward computation. It can simulate the complex nonlinear mapping between request features and hardware load in a high-dimensional space. As the forward computation progresses layer by layer in the network, it eventually reaches the output layer and outputs the initial resource prediction tensor. The principle behind outputting the initial resource prediction tensor is that the output layer neurons integrate all the deeply refined hidden layer features and transform them into prediction results that match the dimensions of the physical machine's hardware indicators. This tensor is essentially a comprehensive matrix containing multiple values. For example, it may predict that in order to cope with the wave of 3D image reconstruction requests that were just input, the physical machine group needs to schedule and allocate 80 virtual computing cores and 256 gigabytes of running memory. Since the various parameters inside the decision model have not yet been refined and corrected by the real results, the matrix containing the expected resource scheduling amounts obtained from the forward calculation is called the initial resource prediction tensor. It represents the model's original intuitive estimate of the computing power cost required for extreme burst traffic in an immature state.

[0029] After obtaining the predictions from the decision-making model based on raw intuition, a rigorous quantitative evaluation of their accuracy is needed to identify the gap between the predictions and reality. The initial resource prediction tensor and the resource consumption data from the burst stress data pair are substituted into the loss function to calculate the burst stress error value. The resource consumption data in the burst stress data pair is an objective historical record of actual events occurring on the physical units. It represents the 85 virtual computing cores and 280 gigabytes of running memory actually consumed when responding to that wave of 3D image reconstruction requests—an irrefutable and true metric. The approach of substituting the estimated tensor and the actual consumption data into the loss function calculation is to align the various resource scheduling quantities predicted by the model with the corresponding resource consumption quantities recorded in the historical data, and then use specific mathematical metrics to compare the degree of deviation between them one by one. The reason why the sudden pressure error value can be generated by substituting into the loss function is that the loss function is specifically designed to measure the distance between two high-dimensional vectors in a spatial geometric sense. It calculates the difference between the virtual computing core prediction value and the actual value, as well as the difference between the running memory prediction value and the actual value, and then integrates these scattered dimensional differences into a single scalar value by summing the absolute values ​​or the squares.

[0030] After obtaining a clearly quantified penalty signal for the model's prediction bias, the crucial step of correcting internal cognitive errors within the model begins. The first update gradient is derived by performing back-derivative on the decision model's parameters based on the sudden stress error value. This back-derivative process starts from the output layer of the decision model and proceeds backward along the forward computation path, using the chain rule to calculate the partial derivative of the error value with respect to the connection weights and biases of each neuron layer by layer. The collection of all these calculated partial derivatives forms the first update gradient. This gradient matrix acts like a highly precise correction navigation map, telling each node in the model which direction and by how much to fine-tune to minimize the large sudden stress error value as quickly as possible. Once this correction navigation map is understood, the decision model's parameters are updated based on the first update gradient. The principle of gradient-based parameter updates is to subtract the product of the first update gradient and a small learning step from the existing weight matrix within the current decision model. This subtraction operation forces the weights of neurons to shift in the direction of error reduction. Since the entire process of error backpropagation and parameter shifting is based entirely on burst pressure data pairs representing extreme peaks, the decision model after updating parameters based on the first update gradient has deeply imprinted the memory characteristics of how to deal with concurrent storms on its internal connection weights. This ensures that when faced with a massive influx of abnormal requests, it can make the most accurate and sufficient computing power scheduling predictions and avoid making the fatal mistake of underestimating resource consumption.

[0031] In one embodiment of the present invention, in step S4, when backpropagating using stationary operating data, the process of projecting the generated second update gradient onto the orthogonal null space of the long-tailed memory matrix for parameter iteration to obtain the trained decision model is as follows: After the decision-making model has initially grasped the characteristics of handling extreme traffic through sudden pressure data, it needs to continue learning the characteristics of normal daily traffic to acquire comprehensive coordination capabilities. The decision-making model, with its input parameters updated using stable operation data, undergoes forward inference and backward differentiation. Specifically, the updated input model extracts concurrent request data representing routine operations from the extracted stable operation data set and inputs it into the network input layer, whose weights have been adjusted by the first update gradient. The forward inference principle involves performing matrix multiplication and bias stacking layer by layer on the characteristics of these normal requests across each neuron node according to the existing weight distribution, ultimately deriving a predicted resource demand for daily business load at the output layer. After obtaining this prediction, the backward differentiation process compares it with the actual hardware resource consumption objectively recorded in the stable operation data set to calculate the prediction deviation value under stable conditions. The principle for calculating the second update gradient is to use this newly generated prediction deviation value to deduce, step by step, from the hidden layers to the input layer, the magnitude and direction of adjustment needed for the connection weights of each neuron to eliminate this daily prediction deviation.

[0032] After obtaining the second update gradient representing the direction of daily learning patterns, directly applying it to the model parameters would inevitably wash away the hard-won experience of handling emergencies. Therefore, mathematical isolation is necessary. The eigenvectors of the long-tail memory matrix are calculated to construct an orthogonal projection matrix. The method for calculating the eigenvectors of the long-tail memory matrix involves performing matrix eigenvalue decomposition on the long-tail memory matrix extracted from the burst data, extracting several eigenvectors that represent the core multidimensional spatial skeleton of this matrix. The principle of constructing the orthogonal projection matrix is ​​to use these eigenvectors representing the burst memory spatial skeleton to construct a mathematical transformation matrix with special spatial filtering properties; this orthogonal projection matrix is ​​like an extremely precise directional filter, capable of accurately identifying any interference signals attempting to intrude into or change the direction of the burst memory space. After preparing this filter, the second update gradient is multiplied by the orthogonal projection matrix to generate an orthogonal gradient tensor. The principle of multiplying the two is to use matrix projection calculation to forcibly project the second update gradient, which originally intended to directly change the model weights, into a null space completely perpendicular to the long-tail memory matrix space.

[0033] After successfully obtaining the pure adjustment direction stripped of destructive components, the network can be safely upgraded for final cognitive purposes. Parameter iteration is completed by summing the parameters of the decision model based on the orthogonal gradient tensor. Specifically, this summing calculation involves extracting the connection weights and biases of each currently used neuron within the decision model and adding them one by one to the corresponding fine-tuning values ​​in the orthogonal gradient tensor. The principle behind parameter iteration is that this safe incremental summing forces a translation of the spatial coordinates of the network weights. Since the summing is done on an orthogonal tensor, this translation occurs entirely in an orthogonal dimension that does not affect existing burst memory features. For example, in a 3-dimensional space, burst memories occupy the horizontal planes of length and width, while the parameter iteration based on the orthogonal gradient tensor only moves the position in the vertical dimension of height. Therefore, regardless of changes in height, the network's projected coordinates on the horizontal planes of length and width remain unchanged. The reason why the trained decision model can be generated through parameter iteration is that after this superposition, the internal structure of the network reaches an extremely ideal dual-track balance state. The model not only firmly retains the memory of dealing with extreme concurrency pressure, but also perfectly integrates the new understanding of predicting daily stable resource consumption in a completely independent dimension. Thus, the model has the complete ability to give accurate computing power scheduling prediction under any traffic pattern, thereby generating the final trained decision model.

[0034] In one embodiment of the present invention, step S5, which involves real-time collection of current concurrent request data from the department system, inputting it into the trained decision model, performing feature matching in conjunction with the long-tail memory matrix, and outputting the resource load prediction tensor of the current concurrent request data, is as follows: In a real-world network environment where business operations are continuously running, it is necessary to parse departmental messages in real time to capture every ongoing interaction. The specific approach to real-time parsing of departmental messages involves deploying a persistent listening component on the data gateway port. Whenever a doctor's or nurse's terminal device issues any business instruction, the listening component instantly intercepts the rapidly passing network packets and immediately strips away the outer routing protocol shell. After acquiring these real-time request records, they need to be input into a trained decision model to extract the current request features. This is achieved by injecting these structured records, containing the number and type of requests, as initial signals into the input nodes of a multi-layered neural network. Within this model, which already possesses mature cognitive capabilities, the initial signals undergo intensive matrix multiplication and addition operations along the connection weights between neurons. After undergoing nonlinear spatial transformations within the multi-layered network structure, the originally simple quantitative indicators are projected into a high-dimensional mathematical space. After accurately extracting the high-dimensional vector representing the current pressure pattern, it is necessary to calculate the inner product of the current request feature and the long-tail memory matrix. The specific method for calculating the inner product of the current request feature and the long-tail memory matrix is ​​to multiply the feature vector representing the current real-time pressure element-wise with the matrix array specifically responsible for storing extreme historical experiences, and then sum all the multiplications to obtain specific numerical values. The reason why matching weights can be generated by calculating the inner product is that, in multidimensional spatial geometry, the size of the inner product of two vectors directly reflects the degree of overlap and physical similarity between their directions. When the current request feature shows signs of extreme concurrency, its direction in mathematical space will be highly parallel to a feature axis in the long-tail memory matrix that records historical crises; in this case, the calculated inner product result will be a very large positive number. Conversely, if the current situation is merely a routine smooth operation, its inner product with the crisis memory will approach zero. Through this rigorous spatial similarity measurement calculation, the model can accurately determine the extent to which the current real-time event has awakened those dormant extreme concurrency memories. These sets of inner product values ​​reflecting the degree of similarity constitute the matching weights. After generating this set of precise weight metrics, it is necessary to extract data from the long-tail memory matrix based on the matching weights to construct a feature matching tensor. The method for extracting data from the long-tail memory matrix based on the matching weights involves multiplying each weight value calculated earlier by the corresponding historical experience vector in the long-tail memory matrix, and then summing and fusing these scaled vectors. The reason why it is possible to construct a feature matching tensor by extracting and fusing this data is that the matching weights essentially act as attention allocation; the higher the weight of the historical experience vector, the more dominant it becomes in the final fused tensor. For example, if the current concurrency situation is extremely similar to a sudden peak in server performance where the server nearly crashed in the past, the weight assigned to that memory will be extremely high. Therefore, the extracted feature matching tensor will replicate the knowledge from that crisis event to the greatest extent possible, thus perfectly reshaping those static, underlying memories into a dynamic reference that highly matches the current real-time pressure.

[0035] After extracting dynamic reference data that highly fits the current pressure, the feature matching tensor needs to be concatenated with the current request features and substituted into the prediction layer of the decision model. Specifically, this concatenation involves mathematically linking a vector array representing historical crisis experience with a high-dimensional vector representing the current real-time business situation, merging and expanding it into a comprehensive tensor matrix containing longer dimensions and richer information. Substituting this concatenated comprehensive tensor into the prediction layer of the decision model allows this ultimate data stream, integrating reality and history, to enter the network's final processing unit. In this prediction layer, responsible for making the final decision, dense neural synapses perform a final dimensionality reduction mapping operation on the concatenated tensor according to the optimal computational rules embedded within the network. The principle behind calculating and outputting the resource load prediction tensor for the current concurrent request data is that the prediction layer smoothly compresses and projects the high-dimensional composite carrying massive features onto a low-dimensional coordinate system corresponding to real physical computing resources. Because the calculation process at this time not only considers the instantaneous impact of the current real-time concurrency, but also fully integrates the extreme case protection strategy provided by the long-tail memory matrix, the final output resource load prediction tensor accurately indicates the physical defense indicators actually needed to resist the current wave of traffic.

[0036] In one embodiment of the present invention, step S6, which involves generating emergency resource scheduling instructions according to the resource load prediction tensor and sending the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling, is as follows: After obtaining the mathematical results output by the decision model, it is necessary to parse the resource load prediction tensor to extract the predicted resource requirements. The specific approach to parsing is to perform a reverse dimensional mapping on the high-dimensional mathematical structure, transforming the abstract matrix coordinates into hardware indicator values ​​that the physical machine cluster can directly recognize. The reason why the predicted resource requirements can be extracted by parsing the resource load prediction tensor is that the model, in its output, has already condensed the required configuration into specific positions in the matrix according to a predetermined correspondence rule. For example, the values ​​in the first row of the matrix are set to correspond to the number of central processing units, and the values ​​in the second row are set to correspond to the capacity of random access memory. Through this rigorous coordinate extraction, the tensor, which originally represented the abstract computational scale, is transformed into a detailed list of requirements specifying 120 virtual computing cores and 500 gigabytes of running memory. After obtaining the target demand list, the predicted resource demand needs to be subtracted from the current operating resource quantity. The specific method for obtaining the current operating resource quantity is to send real-time status probe commands to the hardware management layer to collect the hardware quotas currently being used and operated by all computing nodes. For example, the probe might indicate that only 80 computing cores and 300 gigabytes of memory are currently active in the department's architecture. After clarifying the current situation, the predicted total demand list is subtracted item by item from the currently available physical quotas using numerical subtraction. The reason why a resource difference vector can be generated through subtraction is that the subtraction operation directly strips away the already satisfied resource base, exposing only the gaps that urgently need to be filled or the redundancies that need to be eliminated. The set of positive and negative differences calculated in this process constitutes a mathematical vector with a clear physical orientation.

[0037] After clarifying the specific physical resource discrepancies, they need to be translated into concrete action procedures that machines can understand and execute. The specific method for retrieving preset scheduling action templates based on resource discrepancy vectors involves extracting the numerical signs and magnitudes contained in the discrepancy vectors. These features are then used as search criteria to traverse a pre-configured hardware action strategy repository. This repository contains basic response skeletons for hardware scheduling at different scales. For example, in an emergency situation involving a sudden surge of dozens of computing cores, the repository contains a set of action skeletons that skip some initialization checks in exchange for extremely fast boot response. The reason why preset scheduling action templates can be retrieved through feature comparison is that these templates cover various standard operation paths, ranging from gentle memory quota fine-tuning to aggressive large-scale computing cluster startups. By comparing the currently calculated physical discrepancy scale with the applicable boundaries of each skeleton in the repository, the underlying operation plan that best suits the current crisis level can be selected. After selecting the most suitable underlying operation scheme, the resource difference vector and scheduling action template need to be compiled to generate emergency resource scheduling instructions. The specific method of compilation is to directly inject the precise hardware increase and decrease values ​​in the difference vector into the blank parameter positions reserved in the action template, turning the original general operation framework into a customized execution script for the current specific peak. Once the customized execution code set is compiled, it must be pushed to the front line that can directly access the hardware as quickly as possible. The specific method for issuing emergency resource scheduling instructions to the department's resource control node is to use a high-priority out-of-band network channel that is independent of the daily medical record business data flow to transmit this instruction directly to the master control server responsible for coordinating the entire department's physical computer room; this master control server, which is also the resource control node, has the highest management authority across all underlying computing devices and can directly intervene in the working allocation status of each physical motherboard and memory chip. After the command safely arrives at its destination, the control resource control node executes elastic scaling based on the emergency resource scheduling command. The specific method for the control node to execute elastic scaling is to parse the received customized script and issue specific micro-manipulation actions to the underlying virtualization management platform, forcibly changing the physical topology of the existing computing cluster. For example, when the command requests to add 40 computing cores, the control node will quickly scan and find the spare physical devices in a dormant state on the rack, allocate fixed running memory for them, power them up and wake them up, and seamlessly integrate them into the existing departmental computing resource pool to jointly bear the massive data interaction pressure. Conversely, if the command indicates that there is a low traffic and redundancy needs to be eliminated, the control node will first migrate the scattered processes running on the hardware to be reclaimed, and then safely disconnect the power supply to the excess physical cores.

[0038] This specification also provides an artificial intelligence-based departmental data operation and management device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0039] This specification also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: Collect concurrent request data and corresponding resource consumption data from historical periods in the department system, and bind time tags to the concurrent request data and resource consumption data; The instantaneous arrival rate of concurrent request data is calculated based on the time stamp. Based on the threshold of the instantaneous arrival rate, the concurrent request data and resource consumption data are synchronously split into stable operation data pairs and sudden pressure data pairs. Extract the temporal fluctuation and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix; and extract the periodic characteristics of stable operation data pairs to construct a regular baseline vector. During the training phase of the decision model, the parameters are first updated based on the sudden stress data to calculate the first update gradient; then, when backpropagating using stable operation data, the generated second update gradient is projected onto the orthogonal null space of the long-tail memory matrix for parameter iteration, thus obtaining the trained decision model. Real-time collection of current concurrent request data from the department system is input into the trained decision model. Feature matching is performed using the long-tailed memory matrix, and the resource load prediction tensor of the current concurrent request data is output. Emergency resource scheduling instructions are generated based on the resource load prediction tensor, and then sent to the resource control nodes of the departmental system for elastic scaling up and down.

[0040] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A departmental data operation and management method based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect concurrent request data and corresponding resource consumption data from the department system for historical periods, and bind time tags to the concurrent request data and resource consumption data; S2. Calculate the instantaneous arrival rate of concurrent request data based on the time tag. Based on the threshold of the instantaneous arrival rate, synchronously split the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs. S3. Extract the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs, and construct a long-tailed memory matrix; And extract the periodic features of stable operation data pairs to construct a regular baseline vector; S4. During the training phase of the decision model, the parameters are updated based on the sudden pressure data to calculate the first update gradient. Subsequently, during backpropagation using stable operational data, the generated second update gradient is projected onto the orthogonal null space of the long-tailed memory matrix for parameter iteration, resulting in a trained decision model. S5. Collect the current concurrent request data of the department system in real time, input it into the trained decision model, combine it with the long-tail memory matrix for feature matching, and output the resource load prediction tensor of the current concurrent request data. S6. Generate emergency resource scheduling instructions based on the resource load prediction tensor, and send the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling up and down.

2. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S1, the process of collecting concurrent request data and corresponding resource consumption data from historical periods of the departmental system, and binding time tags to the concurrent request data and resource consumption data, is as follows: Intercept business messages received by the department system gateway during historical time periods and parse the business messages into concurrent request data; Based on the task scheduling records of business messages, extract the underlying running logs of task execution nodes to generate corresponding resource consumption data; Read the standard clock stamp when the business message enters the gateway, and bind the standard clock stamp as a timestamp to the concurrent request data and resource consumption data.

3. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S2, the process of calculating the instantaneous arrival rate of concurrent request data based on the time stamp, and then, based on a threshold of the instantaneous arrival rate, synchronously splitting the concurrent request data and resource consumption data into stable operation data pairs and sudden pressure data pairs, is as follows: Divide the time window according to the time label, count the total number of concurrent data requests within the time window, and divide the total number of requests by the time window span to obtain the instantaneous arrival rate. The instantaneous arrival rate is compared with the instantaneous arrival rate threshold. When the instantaneous arrival rate is greater than the instantaneous arrival rate threshold, a burst flag is generated for the corresponding time window; otherwise, a stable flag is generated. Based on the sudden and stable indicators, extract the concurrent request data and the associated resource consumption data within the corresponding time window, and combine and split them into sudden pressure data pairs and stable operation data pairs.

4. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S3, the process of extracting the temporal fluctuation characteristics and concurrency characteristics of sudden pressure data pairs to construct a long-tailed memory matrix, and extracting the periodic characteristics of stable operation data pairs to construct a conventional baseline vector, is as follows: For sudden pressure data, perform sliding window difference operation to extract time series fluctuation features, and obtain concurrency features through fully connected layer mapping operation; The temporal fluctuation features and concurrency features are concatenated into tensors, and the principal component feature array of the concatenated tensors is extracted to construct a long-tailed memory matrix; For stable operation data, perform discrete Fourier transform to generate frequency domain data, and extract the core frequency components in the frequency domain data as periodic features; A smoothed value is obtained by performing a moving average calculation on the periodic features along the time axis, and the smoothed values ​​are arranged in sequence to construct a regular baseline vector.

5. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S4, the process of updating parameters based on the sudden pressure data to calculate the first update gradient during the training phase of the decision model is as follows: Extract concurrent request data from sudden pressure data pairs and input them into the decision model to perform forward computation, outputting an initial resource prediction tensor; Substitute the resource consumption data from the initial resource prediction tensor and the sudden pressure data pair into the loss function to calculate the sudden pressure error value. The first update gradient is obtained by performing back-derivative on the parameters of the decision model based on the sudden pressure error value, and the parameters of the decision model are updated according to the first update gradient.

6. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S4, the process of projecting the generated second update gradient onto the orthogonal null space of the long-tailed memory matrix for parameter iteration during backpropagation using stationary operating data to obtain the trained decision model is as follows: The decision model with updated input parameters based on stable operating data is subjected to forward inference and backward differentiation to calculate and generate the second update gradient. Calculate the eigenvectors of the long-tailed memory matrix to construct an orthogonal projection matrix, and multiply the second update gradient with the orthogonal projection matrix to generate an orthogonal gradient tensor; The parameters of the decision model are summed and iterated based on the orthogonal gradient tensor to generate the trained decision model.

7. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S5, the process of outputting the resource load prediction tensor of the current concurrent request data is as follows: The system analyzes departmental system messages in real time to generate current concurrent request data, and inputs it into the trained decision model to extract current request features. Calculate the inner product of the current request features and the long-tail memory matrix to generate matching weights, and extract the data of the long-tail memory matrix based on the matching weights to construct a feature matching tensor; The feature matching tensor is concatenated with the current request features and substituted into the prediction layer of the decision model to calculate and output the resource load prediction tensor of the current concurrent request data.

8. The departmental data operation and management method based on artificial intelligence according to claim 1, characterized in that, In step S6, the process of generating emergency resource scheduling instructions according to the resource load prediction tensor and sending the emergency resource scheduling instructions to the resource control nodes of the department system for elastic scaling is as follows: The predicted resource demand is extracted by parsing the resource load prediction tensor. The predicted resource demand is then subtracted from the current running resource quantity to generate a resource difference vector. Based on the pre-set scheduling action template of the resource difference vector retrieval system, the resource difference vector and the scheduling action template are compiled to generate emergency resource scheduling instructions; Emergency resource dispatch instructions are sent to the resource control nodes of the department system, and the resource control nodes perform flexible scaling up and down according to the emergency resource dispatch instructions.

9. A departmental data operation and management device based on artificial intelligence, characterized in that, The device includes: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: Collect concurrent request data and corresponding resource consumption data from historical periods in the department system, and bind time tags to the concurrent request data and resource consumption data; The instantaneous arrival rate of concurrent request data is calculated based on the time stamp. Based on the threshold of the instantaneous arrival rate, the concurrent request data and resource consumption data are synchronously split into stable operation data pairs and sudden pressure data pairs. Extract the temporal fluctuation and concurrency characteristics of sudden pressure data pairs and construct a long-tailed memory matrix; And extract the periodic features of stable operation data pairs to construct a regular baseline vector; During the training phase of the decision model, the parameters are first updated based on the sudden stress data to calculate the first update gradient; Subsequently, during backpropagation using stable operational data, the generated second update gradient is projected onto the orthogonal null space of the long-tailed memory matrix for parameter iteration, resulting in a trained decision model. Real-time collection of current concurrent request data from the department system is input into the trained decision model. Feature matching is performed using the long-tailed memory matrix, and the resource load prediction tensor of the current concurrent request data is output. Emergency resource scheduling instructions are generated based on the resource load prediction tensor, and then sent to the resource control nodes of the departmental system for elastic scaling up and down.

Images