A game operation data multi-dimensional management method

Abstract

The application discloses a game operation data multidimensional management method, mainly relates to the technical field of data multidimensional management, and aims to solve the problem that the traditional standardized method cannot model the predictable periodicity of game operation activities and implements indiscriminate compression on all extreme values. The method comprises the following steps: obtaining predicted game operation index data by using a preset space-time prediction model according to an operation decoupling feature tensor and a dynamic business directed adjacency sequence in a historical input window; obtaining a comprehensive loss by weighting the predicted game operation index data of each preset type and actual game operation index data, and iteratively training to obtain a trained preset space-time prediction model; obtaining an operation decoupling feature tensor and a dynamic business directed adjacency sequence corresponding to the to-be-predicted multidimensional game operation index data, and obtaining predicted game operation index data by using the trained preset space-time prediction model, so as to determine whether a crash warning is triggered.

Description

Technical Field

[0001] This application relates to the field of multi-dimensional data management technology, and in particular to a method for multi-dimensional management of game operation data. Background Technology

[0002] In the field of game server operation and maintenance, to ensure service stability, the industry generally adopts a solution of collecting, modeling, and predicting server operation indicators. Existing technologies typically first standardize and preprocess multi-dimensional operation and maintenance indicators such as CPU utilization, memory usage, and number of online players to eliminate differences in units; then, time series prediction models are used to train the cleaned data, capture the trend of indicator changes, and trigger alarms based on the deviation between predicted and actual values. Some solutions also introduce simple periodic features, such as daily or weekly averages, to help the model identify regular fluctuations.

[0003] The existing solutions have several flaws: First, traditional standardization methods indiscriminately compress all extreme values, causing spike signals caused by sudden failures such as server crashes and network jitter to be excessively smoothed during the preprocessing stage, which severely weakens the model's ability to capture early signs of failure. Second, existing models fail to explicitly model the predictable periodicity of game operation activities (such as version updates and limited-time events), conflating planned indicator spikes with real anomalies, resulting in a surge in false alarm rates during the event. At the same time, the periodic component absorbs the characteristics of failure evolution in reverse, further obscuring the temporal evolution path of real anomalies. Summary of the Invention

[0004] This application provides a multi-dimensional management method for game operation data to solve the aforementioned problems of existing solutions.

[0005] This application provides a method for multi-dimensional management of game operation data, the method including:

[0006] Collect multi-dimensional game operation and maintenance indicator data during the operation of the game server cluster, and construct the original game operation and maintenance indicator matrix;

[0007] The original game operation and maintenance indicator matrix is ​​subjected to impact preservation standardization to obtain a standardized game operation and maintenance indicator matrix; the standardized game operation and maintenance indicator matrix is ​​decoupled from its trend, operation cycle and residual using the operation calendar to obtain the operation decoupling feature tensor.

[0008] Each game operation and maintenance indicator data is treated as a game operation and maintenance indicator node. Based on the prior connection state between game operation and maintenance indicator nodes, a business prior directed adjacency matrix is ​​constructed, and dynamic edge weights corresponding to the prior causal edges in the business prior directed adjacency matrix are generated. Based on the dynamic edge weights and preset self-connected edges, all time steps are traversed to obtain the dynamic business directed adjacency sequence.

[0009] Based on the operational decoupling feature tensor and the dynamic business directed adjacency sequence in the historical input window, the predicted game operation and maintenance index data is obtained by using the preset spatiotemporal prediction model; the predicted game operation and maintenance index data of each preset type and the actual game operation and maintenance index data are weighted to obtain the comprehensive loss, and the preset spatiotemporal prediction model is trained iteratively.

[0010] The system obtains the operational decoupling feature tensor and dynamic business directed adjacency sequence corresponding to the multi-dimensional game operation and maintenance index data to be predicted. It then uses a pre-trained spatiotemporal prediction model to obtain the predicted game operation and maintenance index data, thereby determining whether a crash warning is triggered.

[0011] In one implementation of this application, multi-dimensional game operation and maintenance indicator data are collected during the operation of the game server cluster to construct an original game operation and maintenance indicator matrix, specifically including:

[0012] Based on a preset sampling interval, multi-dimensional game operation and maintenance indicator data are collected during the operation of the game server cluster. The game server cluster includes at least: a game server monitoring system, a log collection system, a crash reporting system, and a network quality monitoring system. The collection objects corresponding to the game operation and maintenance indicator data include at least: number of online players, number of active rooms, CPU utilization, memory utilization, bandwidth utilization, average latency, packet loss rate, number of reconnections, number of interface errors, number of crash events, and operation scenario tags.

[0013] Align the game operation and maintenance metrics data according to the sampling timestamp.

[0014] In one implementation of this application, impact-preserving standardization is performed on the original game operation and maintenance indicator matrix to obtain a standardized game operation and maintenance indicator matrix, specifically including:

[0015] With the current time step as the end position, it is the nth Each game operation and maintenance metric is used to construct a sliding statistical window for the most recent preset time period;

[0016] Calculate the first [number] within the sliding statistical window of the most recent preset time period. Local median of game operation and maintenance metrics and local interquartile range ;

[0017] Calculate the first The first time step Relative deviation of game operation and maintenance metrics ;

[0018] Obtain the core fault-sensitive indicator set; wherein, the core fault-sensitive indicator set consists of game operation and maintenance indicators that have preset abnormal changes corresponding to crash events;

[0019] For each time step, count the number of game operation and maintenance indicators in the core fault sensitive indicator set whose relative deviation exceeds the candidate threshold;

[0020] Verify that each time step meets any of the following conditions:

[0021] The number of crash events is greater than 0;

[0022] Alternatively, the number of game operation and maintenance indicators whose core fault-sensitive indicator set exceeds the candidate threshold is greater than the preset deviation threshold;

[0023] Alternatively, if the average latency and packet loss rate both exceed the 90th percentile of the sliding statistical window of their respective most recent preset time period, then it is marked as a candidate time step for a crash impact.

[0024] When the The first time step is a candidate time step for the collapse impact, and the second time step is a candidate time step for the collapse impact. Relative deviation of game operation and maintenance metrics Greater than the preset candidate threshold At half the speed, the standard impact retains the mark. The value is 1; otherwise, the impact level is retained. The value is 0;

[0025] Through the formula:

[0026] Calculate the first The first time step Robust standardized median values ​​of game operation and maintenance metrics ;

[0027] in, Represents the original collected values. This represents the local median. Indicates the local interquartile range. Preset non-zero values;

[0028] Through the formula:

[0029] Obtain standardized game operation and maintenance indicator values ;

[0030] in, This indicates the parameter for adjusting the impact retention strength. This indicates the upper limit of the relative deviation cutoff. This indicates the upper limit of the truncation value for standardized game operation and maintenance metrics. This represents the truncation function. The marking indicates that the impact level is retained. Indicates the relative deviation;

[0031] Will As a standardized game operation and maintenance indicator matrix The elements in row t and column d are used to construct a standardized game operation and maintenance indicator matrix. .

[0032] In one implementation of this application, the standardized game operation and maintenance indicator matrix is ​​decoupled from its execution trend, operation cycle, and residual using an operation calendar to obtain an operation decoupling feature tensor, specifically including:

[0033] For each time step Calculate the time slot index Among them, time slot index This indicates the current time step's position within a day;

[0034] Standardized game operation and maintenance indicator matrix Each game operation and maintenance metric in the matrix is ​​subjected to an exponential moving average to obtain a trend component matrix. Among them, the trend component matrix The size is ,and Indicates the first The first time step Trend components of game operation and maintenance metrics;

[0035] For each time step and each game's operation and maintenance metrics At the current time step Searching for the current time slot index within the previously preset historical time range. Same or adjacent, and with the same operational scenario tags Same historical time step;

[0036] Iterate through each time step and each game's operation and maintenance metrics The operating cycle component matrix is ​​obtained. Among them, the operating cycle component matrix The size is ,and Indicates the first The first time step The operational cycle component of each game's operation and maintenance metrics;

[0037] Through the formula:

[0038] Calculate the residual component matrix The Middle The first time step Residual components of game operation and maintenance metrics Thus, the residual component matrix is ​​obtained. ;in, Represents a standardized game operation and maintenance indicator matrix The element in row t and column d;

[0039] When the indicator-level shock retains the mark If the value is 1, then the residual components Multiply by the preset residual enhancement factor To obtain enhanced residual components When the indicator-level shock retains the mark. If the value is 0, the residual components are preserved. Without changing, we obtain the strong residual component matrix. ;

[0040] Trend component matrix Operating cycle component matrix and enhanced residual component matrix By concatenating along the channel dimension, we obtain the operational decoupling feature tensor. .

[0041] In one implementation of this application, each game operation and maintenance indicator data is treated as a game operation and maintenance indicator node. Based on the prior connection state between the game operation and maintenance indicator nodes, a business prior directed adjacency matrix is ​​constructed, specifically including:

[0042] Each game operation and maintenance metric is treated as a game operation and maintenance metric node, resulting in a set of game operation and maintenance metric nodes. ,in, Indicates the first Each game operation and maintenance metric corresponds to a game operation and maintenance metric node;

[0043] Construct a business prior directed adjacency matrix Business prior directed adjacency matrix The size is ,in, Indicates the first The game operation and maintenance indicator node points to the first The prior connection state of each game operation and maintenance indicator node is a business prior directed adjacency matrix. The element in the i-th row and j-th column; when the... The game operation and maintenance metrics can directly affect the first When considering game operation and maintenance metrics... The value is 1; otherwise, The value is 0;

[0044] A priori directed adjacency matrix for business purposes For each prior causal edge in the matrix, a business lag step number is set, resulting in a business lag step matrix. Among them, the business lag step matrix The size is ,and Indicates the first The game operation and maintenance indicator node for the first The discrete lag steps when a game operation and maintenance indicator node has an impact.

[0045] In one implementation of this application, generating dynamic edge weights corresponding to prior causal edges in the business prior directed adjacency matrix specifically includes:

[0046] For each time step Construct a graph statistics window that includes the current time step. and the current time step A previously preset historical time step;

[0047] For each game operation and maintenance metric node Calculate source-side activity within the graph statistics window. and node volatility Among them, source-side activity The mean of the first-order absolute differences of standardized game operation and maintenance indicator values ​​within the statistical window; node volatility. The standard deviation of standardized game operation and maintenance indicator values ​​within the graph statistics window is calculated.

[0048] For the business prior directed adjacency matrix Prior causal edges with a value of 1 Based on the number of business lag steps Computational lag consistency ;

[0049] Through the formula:

[0050] Calculate the original dynamic edge score ;

[0051] in, Describing a priori causal edges Basic edge offset Indicates the source-side activity weight. Indicates the target node volatility weight. Indicates the lagged consistency weight;

[0052] For each game operation and maintenance metric node Only when the conditions are met The dynamic edge weights are obtained by performing Softmax normalization on the original dynamic edge scores within the edge range corresponding to the game operation and maintenance indicator nodes. ;like Then directly order Dynamic edge weights Indicates the first The first time step The game operation and maintenance indicator node for the first Dynamic edge weights of game operation and maintenance indicator nodes.

[0053] In one implementation of this application, a dynamic business directed adjacency sequence is obtained by traversing all time steps based on dynamic edge weights and preset self-connecting edges, specifically including:

[0054] Add a self-connecting edge to each game operation and maintenance metric node; where the self-connecting edge represents the historical state that the game operation and maintenance metric node retains itself;

[0055] Iterate through all time steps, stack the dynamic adjacency matrices of all time steps, and obtain the dynamic business directed adjacency sequence. Among them, dynamic services include directed adjacency sequences. The size is And the dynamic service directed adjacency sequence includes ,and Indicates the first The first time step The game operation and maintenance indicator node points to the first Dynamic edge weights of game operation and maintenance indicator nodes.

[0056] In one implementation of this application, predictive game operation and maintenance index data is obtained using a preset spatiotemporal prediction model based on the operation decoupling feature tensor and the dynamic business directed adjacency sequence within the historical input window. Specifically, this includes:

[0057] The first The operational decoupling feature tensor within the historical input window of each training sample is mapped to the initial node features. At the same time, the number of collapse events corresponding to the historical input window is encoded as multi-scale exponentially decaying pulse features and injected into all nodes to obtain a collapse-enhanced node feature tensor carrying the memory of collapse impact.

[0058] From dynamic business directed adjacency sequence Extracting the first The training samples correspond to the directed adjacency sequence of the sample dynamic business. ,in, Indicates the first The training sample of the th training sample The first historical step The game operation and maintenance indicator node points to the first The dynamic edge weights of the game operation and maintenance indicator nodes; in the first... In the coding layer, for the first... Game operation and maintenance metrics nodes at each historical time step According to the sample dynamic business directed adjacency sequence The node features of the game operation and maintenance indicator nodes are obtained, the node features are linearly transformed, multiplied by the dynamic edge weight of the current time step, and finally summed to obtain the graph aggregation features of the game operation and maintenance indicator nodes.

[0059] The preset spatiotemporal prediction model concatenates the graph aggregation features with the current features of the game operation and maintenance index nodes and inputs them into the feedforward transformation layer. After passing through the linear rectified activation function and layer normalization of the preset spatiotemporal prediction model, the graph update node features are obtained.

[0060] The pre-defined spatiotemporal prediction model performs causal temporal convolution along the time dimension on each game operation and maintenance indicator node. The causal temporal convolution only uses the current historical time step and the historical time steps before it, and outputs the temporal convolution output feature tensor. The causal temporal convolution output is added to the graph update node features through residual connections to obtain the node features of the next encoding layer.

[0061] Repeatedly execute directed graph message passing and causal temporal convolution until all encoding layers are completed, obtaining the encoder output feature tensor. ;

[0062] For encoder output feature tensor Attention aggregation is performed along the time dimension to obtain the global node representation of each game operation and maintenance metric node. ;

[0063] For each future time step and each game's operation and maintenance metrics Constructing future query vectors ;

[0064] Using future query vectors Perform cross-attention on all global node representations to obtain the predicted intermediate features. ;

[0065] Within the same future time step, a single-layer lightweight predictive indicator graph convolution is performed on the intermediate features of all game operation and maintenance indicators to map them to real scalars. For non-negative count-type game operation and maintenance indicators, the real scalars are converted to positive numbers using the Softplus function to obtain the corresponding predicted values. For non-negative continuous game operation and maintenance indicators, the corresponding predicted values ​​are output using the Softplus function. For bounded proportional game operation and maintenance indicators, the prediction results are mapped to the range of 0 to 1 using the Sigmoid function, and then multiplied by a preset upper limit of the proportion to obtain the corresponding predicted values. For the number of crash events, the real scalar is fed into the Sigmoid function to obtain the probability of a crash; then fed into the Softplus function to obtain the crash count intensity; the product of the probability of a crash and the crash count intensity is used as the predicted value corresponding to the number of crash events.

[0066] By iterating through all future time steps and all game operation and maintenance metrics, a prediction result matrix containing the predicted values ​​corresponding to the game operation and maintenance metrics is obtained.

[0067] In one implementation of this application, the predicted game operation and maintenance indicator data and the actual game operation and maintenance indicator data of each preset type are weighted to obtain a comprehensive loss, and the trained preset spatiotemporal prediction model is obtained through iterative training, specifically including:

[0068] For the The training sample is taken from the prediction result matrix. The first future time step The predicted values ​​of each game operation and maintenance indicator are used as the predicted game operation and maintenance indicator data, and the corresponding real values ​​are taken from the real result matrix as the actual game operation and maintenance indicator data.

[0069] For game operation and maintenance metrics such as the number of online players, the Poisson loss function term is used to calculate the prediction error between the predicted game operation and maintenance metric data and the actual game operation and maintenance metric data.

[0070] For game operation and maintenance metrics of average latency, the mean squared error loss function term is used to calculate the prediction error.

[0071] For game operation and maintenance metrics such as CPU utilization, memory utilization, and packet loss rate, the mean squared error loss function term is used to calculate the prediction error.

[0072] For game operation and maintenance metrics related to the number of crash events, the prediction error is calculated using a zero-inflated Poisson loss function term.

[0073] The sample comprehensive loss is obtained by multiplying the preset loss weight corresponding to each preset type by the prediction error and summing the results.

[0074] The batch loss is obtained by averaging the sample loss of all training samples in a training batch. Based on the batch loss, the adaptive moment estimation optimizer is used to update all trainable parameters in the event-driven pre-set spatiotemporal prediction model. During the training process, training samples containing collapse events are weighted and random masks are applied to the enhanced residual components in the operation decoupling feature tensor.

[0075] After each training cycle, the validation error of each game operation and maintenance indicator within the future prediction window is calculated on the validation set, and the model parameters with the lowest overall loss on the validation set are saved as the final output of the trained preset spatiotemporal prediction model.

[0076] As can be seen from the above technical solutions, this application has the following advantages:

[0077] To address the impact of crashes and failures in game operation and maintenance data, an "impact retention standardization" method was designed. By constructing crash impact candidate markers and indicator-level impact retention markers, the method retains extreme abrupt change signals with early warning significance before and after the failure while suppressing interference from ordinary activity peaks, thus avoiding the over-compression of key precursors.

[0078] By using the operation calendar to decouple the three components of operation and maintenance indicators—trend, operation cycle, and residual—and combining operation scenario tags and historical time slot statistics, explainable periodic activity fluctuations are attributed to the operation cycle component, so that the residual component mainly carries abnormal information such as fault impacts, reducing the interference of normal operation activities on anomaly detection.

[0079] A dynamic business directed graph was constructed that integrates business prior causality and statistical context-driven approaches. The causal framework and lag steps were fixed using operational experience, and the edge weights were dynamically adjusted based on recent data fluctuations and lag consistency. This enabled adaptive changes in the transmission strength between metrics under different operational scenarios and avoided pseudo-correlation connections generated by purely data-driven graph construction.

[0080] In the time series prediction model, business lag steps are introduced for directed graph message passing, so that indicator information is propagated along the causal edge according to the delay time that conforms to the business logic. At the same time, multiple types of output heads are combined to distinguish the physical constraints of count, continuous and proportional indicators, and a zero-inflation event output head is specially designed to handle the sparse burst characteristics of crash events, thereby improving the rationality of prediction. Attached Figure Description

[0081] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0082] Figure 1 This is a flowchart of a multi-dimensional management method for game operation data provided in an embodiment of this application.

[0083] Figure 2 This is a raw data collection curve of a core game operation and maintenance indicator over time, provided in an embodiment of this application.

[0084] Figure 3 This is a graph provided in this application embodiment that compares the CPU utilization rate after missing value processing with the original true value.

[0085] Figure 4 This application provides a comparison chart of crash impact candidate markers and crash events.

[0086] Figure 5This is a raw sequence-trend component-operating cycle component-residual component diagram provided in the embodiments of this application. Detailed Implementation

[0087] 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.

[0088] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0089] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0090] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0091] This embodiment provides a method for multi-dimensional management of game operation data, such as... Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0092] Step 110: Collect multi-dimensional game operation and maintenance indicator data during the operation of the game server cluster, and construct the original game operation and maintenance indicator matrix.

[0093] In some embodiments, multi-dimensional game operation and maintenance indicator data are collected during the operation of the game server cluster. For example, game operation and maintenance indicator data are read from the game server monitoring system, log collection system, crash reporting system and network quality monitoring system, and aligned according to a unified sampling timestamp. If a game operation and maintenance indicator is missing at a certain time step, the collected value of the most recent valid time step is used for forward filling. If the missing time exceeds two sampling intervals, the data is deleted.

[0094] The data collected includes the number of online players (in people), the number of active rooms (in rooms), CPU utilization (in %), memory utilization (in %), bandwidth utilization (in Mbps), average latency (in ms), packet loss rate (in %), number of reconnections (in times / sampling interval), number of interface errors (in times / sampling interval), and number of crash events (in times / sampling interval). These game operation and maintenance metrics include not only business load metrics, but also system resource metrics, network quality metrics, and fault event metrics, which together reflect the game's operating status.

[0095] The sampling interval is denoted as , This represents the time interval between two consecutive samples, in seconds. For example, 300 seconds would be used, meaning a sample is collected every 5 minutes. After collection, the game operation and maintenance metrics under the same server group, the same game region, or the same logical service instance are organized in chronological order to obtain the original game operation and maintenance metric matrix. Original game operation and maintenance indicator matrix The size is ,in, Indicates the total number of time steps. This represents the total number of game operation and maintenance metrics. The Middle Line number The elements of a column are denoted as , Indicates the first The first time step The raw collected values ​​of each game operation and maintenance metric. The value range is 1 to , The value range is 1 to .

[0096] In practice, the number of crash events recorded in the crash reporting system is organized into a crash event count (sequence). ,in, Indicates the first The number of crash events occurring within a time step is a non-negative integer. If the crash log is stored as a single event, the number of crash events within the same sampling interval is accumulated according to the sampling timestamp to obtain the crash event count (sequence). .

[0097] In the actual implementation, the game's operation calendar is read, and operation scenario tags are generated for each time step. ,in, Indicates the first The operational scenario type corresponding to each time step. Operational scenario tags. This can include regular days, weekends, version update days, maintenance days, scheduled event days, major tournament days, and compensation distribution days. If the same time step belongs to multiple operational scenarios, the primary operational scenario will be selected according to the preset priority, for example, version update days take precedence over weekends, and maintenance days take precedence over regular days.

[0098] It should be noted that this application also iterates through the original game operation and maintenance indicator matrix. The entire available time period is used to obtain the training sample set. The number of training samples is denoted as . , This represents the total number of sliding window samples available for model training. Each training sample stores the historical input window, the future prediction window, the operational scenario label sequence corresponding to the historical input window, the operational scenario label sequence corresponding to the future prediction window, and the sequence of crash event counts.

[0099] In one embodiment, such as Figure 2 As shown, the analysis of raw game operation and maintenance (O&M) metrics data (first 500 time steps) displays the original data collection curves of six core O&M metrics over time, including online player count, CPU utilization, memory utilization, average latency, packet loss rate, and number of crash events. The different metrics exhibit typical O&M data characteristics such as large differences in dimensions, significant periodic fluctuations, and substantial sudden shocks. The horizontal axis represents time steps (dimensionless, each step corresponding to a 5-minute sampling interval), and the vertical axis represents, in order, people (online player count), % (CPU / memory utilization, packet loss rate), ms (average latency), and times (number of crash events).

[0100] In one embodiment, such as Figure 3 As shown, the analysis illustrates the forward filling of missing CPU utilization values ​​(time steps 200-350). By comparing the curves of CPU utilization after missing value processing (forward filling) with the original true values, the strategy of "prioritizing the use of the most recent valid time step for forward filling" can reasonably fill in discontinuous missing values ​​without introducing abnormal jumps. The horizontal axis represents time steps (dimensionless), and the vertical axis represents CPU utilization (%).

[0101] Step 120: Perform impact retention standardization on the original game operation and maintenance indicator matrix to obtain a standardized game operation and maintenance indicator matrix; use the operation calendar to decouple the trend, operation cycle and residual of the standardized game operation and maintenance indicator matrix to obtain the operation decoupling feature tensor.

[0102] It should be noted that game operation and maintenance metrics data have the characteristics of large differences in dimensions, strong periodic fluctuations, obvious activity peaks, and sparse crash impacts. If ordinary mean-variance standardization is directly applied, it is easily affected by activity peaks and crash spikes; if strong robust compression is directly applied, it may weaken the key abrupt change signals before and after the crash event.

[0103] This step involves analyzing the original game operation and maintenance indicator matrix. The execution of impact management retains standardization and decouples the operational cycle, resulting in a standardized game operation and maintenance indicator matrix. Decoupling feature tensors from operations The specific steps are as follows:

[0104] Adaptive robust normalization based on collapse shock candidate tags:

[0105] The number of crash events in game operation and maintenance metrics is usually 0, and only suddenly increases when a failure occurs. Before a crash event occurs, the average latency, packet loss rate, memory usage rate and number of interface errors may have already increased abnormally. If the standardization process performs the same compression on all extreme values, it will reduce the visibility of the precursors to failure.

[0106] This step constructs candidate markers for collapse shocks based on sliding robust normalization, retaining high amplitude only for mutations that are significant for failure. The specific steps are as follows:

[0107] 1> Step at current time The end position is the first position. Each game's operation and maintenance metrics are used to construct a sliding statistical window for the most recent 24 hours; the sliding statistical window for the most recent 24 hours only includes the current time step. and the current time step Previous historical data collection values ​​do not include future data collection values; if the sampling interval is... If we take 300 seconds, then the sliding statistics window for the most recent 24 hours can contain a maximum of 288 time steps.

[0108] Furthermore, the calculation is performed within the sliding statistical window of the most recent 24 hours. Local median of game operation and maintenance metrics and local interquartile range ,in, Indicates the first The first time step The robust central level of each game's operation and maintenance metrics within a local time frame. Indicates the first The first time step The robust fluctuation range of individual game operation and maintenance metrics within a local time frame; local interquartile range It is obtained by subtracting the 25th percentile from the 75th percentile; when Less than the minimum constant At that time, Replace with ;in, Used to prevent division by zero, with the following value example: .

[0109] In one embodiment, as an example, let the sampling interval be... seconds, current time step Game operation and maintenance metrics This represents the "average delay". A sliding statistical window of the most recent 24 hours is used, containing 288 collected values ​​from time steps 213 to 500. Among these 288 values, after sorting in ascending order, the 144th value is 45, the 72nd value is 40, and the 216th value is 50. Therefore, we have the local median. ms, local interquartile range ms. If the calculated Then, with a minimal constant Substitution, i.e. .

[0110] 2> Calculate the first The first time step Relative deviation of game operation and maintenance metrics relative deviation The degree of deviation of the current raw data collection value from the local normal level is represented as:

[0111] ;

[0112] in, Indicates the first The first time step The relative deviation of each game operation and maintenance metric Characterizes the absolute deviation between the current raw collected value and the local median. Indicates the protection threshold, used to avoid When the relative deviation is close to 0, it amplifies abnormally. An example of such a value is... .

[0113] 3> Select the set of core fault-sensitive indicators Core Fault Sensitive Indicators Set This set of game operation and maintenance metrics is used to represent the game operations and maintenance metrics that are most likely to change abnormally before and after a crash event. These metrics may include average latency, packet loss rate, memory usage, number of interface errors, number of reconnections, and number of crash events.

[0114] Specifically, for each time step Statistical set of core fault-sensitive indicators The relative deviation exceeds the candidate threshold Number of game operation and maintenance metrics, candidate threshold A value of 0.2 to 0.5 is acceptable.

[0115] In one embodiment, as an example, a candidate threshold is set. Core Fault Sensitive Indicators Set Includes "average latency", "packet loss rate", "memory usage", "number of interface errors", "number of reconnections", and "number of crashes"; at time step Calculate the relative deviation of each indicator. as follows:

[0116] Average delay (> );

[0117] Packet loss rate (> );

[0118] Memory usage (> );

[0119] Number of interface errors (≤) );

[0120] Reconnection count (≤) );

[0121] Number of crashes (≤) );

[0122] Then the core fault sensitivity index set The relative deviation exceeds The number of metrics is 3 (average latency, packet loss rate, and memory usage).

[0123] 4> Generate time-level crash impact candidate markers.

[0124] Specifically, if the first If the nth time step satisfies any of the following conditions, then the nth time step will be... Each time step is marked as a candidate time step for the collapse shock:

[0125] a> Crash event count sequence Greater than 0;

[0126] b> Set of core fault-sensitive indicators The relative deviation of at least two game operation and maintenance indicators exceeds the candidate threshold. ;

[0127] c>The average latency and packet loss rate both exceed the 90th percentile of their respective 24-hour sliding statistical windows.

[0128] Based on this, even if a crash event has not yet occurred, candidate crash impact markers can be generated based on synchronous anomalies of multiple indicators during online prediction.

[0129] 5> Generate indicator-level shock retention markers ,in, Indicates the first The first time step Whether a game's operation and maintenance metric needs to be enhanced by impact retention, with a value of 0 or 1.

[0130] Specifically, when the first The first time step is a candidate time step for the collapse shock, and the second time step is a candidate time step for the collapse shock. Relative deviation of game operation and maintenance metrics Greater than the candidate threshold When half of it, The value is 1; otherwise, The value is 0.

[0131] In one embodiment, following the example above, the time step Condition a (number of crash events) is met ), which is the candidate time step for the crash impact, at which point:

[0132] For average delay, greater than Therefore ;

[0133] Regarding the number of interface errors No greater than Therefore .

[0134] 6> Calculate the first The first time step Robust standardized median values ​​of game operation and maintenance metrics , is represented as:

[0135] ;

[0136] in, Indicates the first The first time step Robust standardized median values ​​for game operation and maintenance metrics.

[0137] 7> Retain markings based on index-level impact. Adjusting the robust normalized intermediate value A standardized game operation and maintenance indicator matrix was obtained. Standardized game operation and maintenance metrics in , is represented as:

[0138] ;

[0139] in, Indicates the first The first time step The standardized game operation and maintenance indicator values ​​are the standardized game operation and maintenance indicator matrix. The element in row t and column d; This indicates the parameter for adjusting the impact retention strength, which can be set to 0.1; This indicates the upper limit of the relative deviation cutoff, which can be taken as 5; This represents the upper limit of the truncation value for standardized game operation and maintenance metrics, which can be set to 10. This represents a truncation function used to restrict a value to a specified range (i.e., ...). Within the range.

[0140] By iterating through all time steps and all game operation and maintenance metrics, a standardized game operation and maintenance metric matrix is ​​obtained. Standardized game operation and maintenance indicator matrix The size is Standardized game operation and maintenance indicator matrix The game operation and maintenance index data are characterized after dimensional alignment, normal anomaly suppression, and fault impact retention.

[0141] In one embodiment, for example, a game server crashes at the 500th time step, and the average latency increases from 40ms to 160ms; if the average latency is within the local median of the sliding statistical window of the most recent 24 hours... The value is 45, and the local interquartile range is... If it is 10, then the robust normalized median is... The value is 11.5; due to the number of crash events at the 500th time step. Greater than 0, and the relative deviation of the average delay exceeds the candidate threshold. Indicator-level shocks retain markers A value of 1 represents a standardized game operation and maintenance metric. Maintaining a large positive deviation within the cutoff range allows subsequent models to identify the link between abnormally high latency and crash events.

[0142] It should be noted that this step uses the number of crash events, multiple indicator synchronization anomalies, and the core fault-sensitive indicator set to jointly screen candidate time steps for crash impact. This can reduce the interference of ordinary activity peaks on the standardization results, while retaining abnormal mutations that have early warning significance before and after the crash event.

[0143] This step utilizes the operations calendar to standardize the game operations and maintenance indicator matrix. Decoupling execution trends, operational cycles, and residuals involves the following steps:

[0144] 1> For each time step Calculate the time slot index Time slot index This indicates the current time step's position within a day; for example, with a sampling interval of 5 minutes, a day contains 288 time slots. If the current time is 20:00, then... This corresponds to the 240th time slot.

[0145] 2> Standardized game operation and maintenance indicator matrix Each game operation and maintenance metric in the matrix is ​​subjected to an exponential moving average to obtain a trend component matrix. Trend component matrix The size is ,in, Indicates the first The first time step The trend components of each game operation and maintenance metric are the trend component matrix. The element in row t and column d describes the slow-changing portion of the game's operational metrics.

[0146] In the exponential moving average operation, the smoothing coefficient of the exponential moving average is denoted as . , A value of 0.05 to 0.2 is acceptable. The smaller the value, the smoother the trend component.

[0147] In one embodiment, for example, let the smoothing coefficient be... Game operation and maintenance metrics For "number of online players"; in time step Standardized value The trend component of the previous time step The current trend component . The smaller the value, the stronger the trend component. The less sensitive it is to short-term fluctuations, the smoother the overall performance.

[0148] 3> For each time step and each game's operation and maintenance metrics At the current time step Searching for the current time slot index within a previous historical time range. Same or adjacent, and with the same operational scenario tags For the same historical time step, the historical time range can be the most recent 28 days.

[0149] In practical implementation, if the number of available historical time steps is not less than If the number of available historical time steps is less than [a certain number], then the median of the detrended values ​​corresponding to these historical time steps is taken as the operating cycle component; If so, then revert to using the same time slot index but without requiring the same operational scenario label for the historical time step. This represents the minimum number of historical samples, which can be 5 to 10.

[0150] In one embodiment, as an example, let the current time step be... Corresponding to Saturday 20:05, time slot index Operational scenario tags For "major event days," the historical time range is the most recent 28 days. Within these 28 days, look for historical time steps with time slot indices of 240, 241, or 242 (i.e., 20:00 to 20:10) and the same operational scenario tag: "major event day." If four historical time steps that meet the criteria are found (e.g., last Saturday at 20:05, the Saturday before last at 20:10, etc.), then the median of the trend values ​​of the number of online players in these four time steps is taken as the operational cycle component of the current time step. If the number of historical time steps that meet all the criteria is insufficient... If there are fewer than 5, then we will return to the historical time steps that do not require the same operational scenario label, but only require the same or adjacent time slot indexes for statistics.

[0151] 4> Iterate through each time step and each game's operation and maintenance metrics The operating cycle component matrix is ​​obtained. Operating cycle component matrix The size is ,in, Indicates the first The first time step The operational cycle components of each game operation and maintenance metric are the operational cycle component matrix. The element in row t and column d, the operating cycle component, is used to describe the explainable fluctuations caused by daily cycles, weekend patterns, activity day patterns, and version update patterns.

[0152] In the actual implementation, for each time step and each game's operation and maintenance metrics In time step Within a previous historical time range (e.g., the most recent 28 days), look up the index corresponding to the current time slot. Same or adjacent (i.e.) , , ), and operational scenario tags The same historical time step; if the number of valid historical time steps that meet the conditions is not less than Then calculate the "detrended value" (i.e., the standardized value minus the trend component) for these historical time steps. The median of the operating cycle is used as the operating cycle component. If the number of valid historical time steps that meet the criteria is less than If the historical time steps with the same or adjacent time slot indices are used, but whose operational scenario labels are not required to be the same, the median of the historical time steps is taken as the operational cycle component. ; Traverse all and The operating cycle component matrix is ​​obtained. .

[0153] In one embodiment, as an example, let the current time step be... Corresponding to Saturday 20:05, , For "major event days"; within the last 28 days, find four historical time steps that are also "major event days" with time slot indices of 240, 241, or 242; take the median of the trend values ​​of the number of online players at these four time steps to obtain... If only 3 are found (less than 3) If the labels are not identical, then the estimation will revert to the historical time step where only the time slot index is in the range of 240-242.

[0154] 5> Calculate the residual component matrix residual component matrix The size is ,in, Indicates the first The first time step The residual components of each game operation and maintenance metric are the residual component matrix. The element in row t and column d is calculated as follows:

[0155] ;

[0156] in, Indicates the first The first time step The residual components of a game operation and maintenance indicator represent the portion of the standardized game operation and maintenance indicator value that cannot be explained by the trend component and the operation cycle component.

[0157] 6> Perform event-aware enhancement on the residual components to obtain the enhanced residual component matrix. Enhanced residual component matrix The size is ,in, Indicates the first The first time step The enhanced residual components of each game operation and maintenance metric are the enhanced residual component matrix. The element in row t and column d.

[0158] In practical implementation, if the indicator-level shock retains the flag... If the value is 1, then the residual components Multiply by residual enhancement factor To obtain enhanced residual components If the indicator-level shock retains the mark. If the value is 0, the residual components are preserved. Unchanged (equivalent to enhancing residual components) Among them, residual enhancement factor. If it is greater than 1, then 1.5 can be taken.

[0159] In one embodiment, for example, let the residual enhancement factor be... ; at time step For the "average latency" metric residual components And the indicator-level shock retains the mark. This enhances the residual components. For the same time step, the residual components... ,mark This enhances the residual components. This operation amplifies the abnormal residuals of crash impact-related indicators, which helps subsequent models capture early signs of failure, while not changing the residual magnitudes of irrelevant indicators.

[0160] 7> The trend component matrix Operating cycle component matrix and enhanced residual component matrix By concatenating along the channel dimension, we obtain the operational decoupling feature tensor. Operational decoupling feature tensor The size is Among them, the first channel is the trend component, the second channel is the operating cycle component, and the third channel is the enhanced residual component.

[0161] In one embodiment, for example, a game holds a large-scale event every Saturday from 20:00 to 22:00. For the time step of 20:05 on Saturday, the system prioritizes extracting the detrended median of the number of online players from historical time steps in the past 28 days that were also large-scale event days and had time slots close to 20:05, as the operating cycle component of the number of online players. If the increase in the number of online players during this event is consistent with the historical event pattern, then this increase mainly enters the operating cycle component. If there is an abnormal increase in average latency, packet loss rate, and number of crash events at the same time, then the part that cannot be explained by the operating cycle component enters the enhanced residual component, thereby enabling the model to distinguish between normal event peaks and fault impacts.

[0162] It should be noted that this step uses the time slot index and the operational scenario label together for operational cycle estimation, which enables version update date, maintenance date and activity date to form different cycle benchmarks, reducing the interference of operational activities on anomaly prediction.

[0163] In one embodiment, such as Figure 4 As shown, a comparison is made between crash impact candidate markers and crash events. The upper part plots the sequence of crash event occurrences, while the lower part displays crash impact candidate markers (binary signals) generated based on crash events, multi-indicator synchronization anomalies, and extreme value conditions. Even at time steps before a crash event occurs, if the average latency and packet loss rate simultaneously and abnormally increase, candidate markers will be activated. This proves that multi-indicator joint judgment can identify potential risk time steps in advance. The horizontal axis represents time steps (dimensionless), the vertical axis of the upper part represents the number of crash events (times), and the vertical axis of the lower part represents candidate markers (unitless, 0 or 1).

[0164] Step 130: Treat each game operation and maintenance indicator data as a game operation and maintenance indicator node. Based on the prior connection state between game operation and maintenance indicator nodes, construct a business prior directed adjacency matrix and generate dynamic edge weights corresponding to the prior causal edges in the business prior directed adjacency matrix. Based on the dynamic edge weights and preset self-connected edges, traverse all time steps to obtain the dynamic business directed adjacency sequence.

[0165] There are clear business transmission relationships between game operation and maintenance metrics. For example, an increase in the number of online players usually leads to an increase in the number of active rooms. An increase in the number of active rooms further leads to an increase in CPU utilization, memory utilization, and bandwidth utilization. Increased resource load may lead to an increase in average latency and packet loss rate. Worsening latency and packet loss may lead to an increase in player churn or crashes. Relying solely on correlation coefficients to build graphs can easily lead to incorrect connections between game operation and maintenance metrics that change simultaneously but have no direct business causal relationship.

[0166] This step transforms game operation and maintenance experience into a business-prior directed adjacency matrix, and generates a dynamic business-directed adjacency sequence based on the current statistical status. This is used for subsequent event-driven spatiotemporal neural network execution graph message passing. The specific steps are as follows:

[0167] Construction of the business prior directed adjacency matrix and the business lag step matrix:

[0168] 1> Treat each game operation and maintenance metric as a game operation and maintenance metric node to obtain a set of game operation and maintenance metric nodes. ,in, Indicates the first Each game operation and maintenance metric corresponds to a game operation and maintenance metric node. This represents the index of the game operation and maintenance metric node, with a value range of 1 to... .

[0169] 2> Construct a business prior directed adjacency matrix based on game operation and maintenance business logic. Business prior directed adjacency matrix The size is ,in, Indicates the first The game operation and maintenance indicator node points to the first The prior connection state of each game operation and maintenance indicator node is a business prior directed adjacency matrix. The element in the i-th row and j-th column. When the... The game operation and maintenance metrics can directly affect the first When considering game operation and maintenance metrics... The value is 1; otherwise, The value is 0.

[0170] In practical implementation, the business prior directed adjacency matrix The prior connection relationships can include: the number of online players pointing to the number of active rooms; the number of active rooms pointing to CPU utilization, memory utilization, and bandwidth utilization; CPU utilization pointing to average latency; memory utilization pointing to the number of crash events; bandwidth utilization pointing to packet loss rate; average latency pointing to the number of reconnections; packet loss rate pointing to the number of reconnections; the number of reconnections pointing to indicators related to the decline in the number of online players; and the number of interface errors pointing to average latency and the number of crash events.

[0171] In one embodiment, as an example, the business prior directed adjacency matrix Here is a sample build example:

[0172] There are 10 game operation and maintenance metrics, with the following node indices:

[0173] Online player count Number of active rooms Central processing unit utilization;

[0174] Memory usage Bandwidth usage Average delay;

[0175] Packet loss rate Number of reconnections Number of interface errors Number of crash events;

[0176] According to business logic, "the number of online players directly affects the number of active rooms," therefore "CPU utilization does not affect the number of online players," therefore Business Prior Directed Adjacency Matrix for In the matrix, all indicator pairs that do not have a direct causal relationship based on operational experience have elements of 0.

[0177] In another embodiment, as an example, the specific assignment of prior connection relationships is as follows:

[0178] The number of online players indicates the number of active rooms: ;

[0179] The number of active rooms indicates the CPU utilization rate: ;

[0180] The number of active rooms indicates memory usage: ;

[0181] The number of active rooms indicates bandwidth usage: ;

[0182] CPU utilization indicates average latency: ;

[0183] Memory usage indicates the number of crash events: ;

[0184] Bandwidth usage indicates packet loss rate: ;

[0185] Average latency points to reconnection count: ;

[0186] Packet loss rate indicates the number of reconnections: ;

[0187] The number of interface errors indicates the average latency: ;

[0188] The number of interface errors points to the number of crash events: ;

[0189] The above relationships constitute a directed graph. The main causal framework in [the context].

[0190] 3> A priori directed adjacency matrix for business applications For each prior causal edge in the matrix, a business lag step number is set, resulting in a business lag step matrix. Business lag step matrix The size is ,in, Indicates the first The game operation and maintenance indicator node for the first The discrete lag steps when a game operation and maintenance indicator node has an impact.

[0191] In one implementation, the number of business lag steps First, an operational experience-based candidate range is defined, then calibration is performed using lag correlation in historical data; specifically, for prior causal edges... Calculate the first lag within the candidate lag range respectively The first-order difference sequence of the game operation and maintenance indicators and the... The correlation of the first-order difference sequences of various game operation and maintenance metrics is analyzed, and the lag step with the highest correlation is selected as the business lag step. If there are insufficient historical samples, the default lag step number given by the operation and maintenance experience will be used directly.

[0192] It should be noted that the number of business lag steps The candidate range is determined through operational experience, and then calibrated based on historical data. Specifically, firstly, operations personnel, based on their experience with the game's business, analyze each prior causal edge. Provide a candidate lag range, for example, the candidate lag for "memory usage points to the number of crash events" is... The first time step; then, during the stable running period of historical data, the second time step is calculated respectively. The first-order difference sequence of the first indicator and the first... The first-order difference sequence of each indicator in lag The Pearson correlation coefficient at each time step is selected based on the highest correlation. As the business lag step on this side If historical samples are insufficient (e.g., for new games), the default lag step given by operational experience is used directly. For example, for the edge "Number of active rooms → CPU utilization", the candidate lag step is... Historically, the first-difference correlation coefficient was highest (0.68) when lagged by one step. .

[0193] It should also be noted that the business prior directed adjacency matrix Limiting "which game operation and maintenance metrics are allowed to pass information", business lag step matrix By defining "how much time delay is more reasonable for information to be transmitted", and by introducing both direction and lag, we can avoid synchronous related edges that have no business significance and improve the interpretability of the graph structure.

[0194] Statistical context-driven dynamic business directed adjacency sequence generation:

[0195] The strength of the same causal edge for a business operation varies in different operational scenarios. For example, an increase in memory usage on a normal day may not necessarily lead to an increase in crashes, but if memory usage increases and is accompanied by greater fluctuations on a version update day, then the influence of memory usage on the number of crashes should be stronger.

[0196] Based on the recent averages, volatility, and lag consistency of game operation and maintenance metrics, a priori directed adjacency matrix is ​​generated for the business. The specific steps for generating dynamic edge weights from prior causal edges in the algorithm are as follows:

[0197] 1> For each time step Construct a graph statistics window that includes the current time step. and the current time step For a previous historical time step, the window length can be from 30 minutes to 2 hours; for example, when the sampling interval is 5 minutes, the graph statistics window can take 12 time steps, corresponding to 1 hour.

[0198] Furthermore, for each game operation and maintenance metric node Calculate source-side activity within the graph statistics window. and node volatility Source-side activity The mean of the first-order absolute differences of the standardized game operation and maintenance index values ​​within the graph statistics window can be used to represent the... The intensity of recent changes in individual game operation and maintenance metrics; node volatility. The standard deviation of the standardized game operation and maintenance indicator values ​​within the graph statistics window can be used to represent the first... The recent instability of certain game operation and maintenance metrics.

[0199] In one embodiment, as an example, suppose the graph statistics window length is 12 time steps (corresponding to 1 hour), and the current time step... ,index This refers to "memory usage".

[0200] Within the 12 time steps from time step 489 to 500, the standardized game operation and maintenance index values They are: 45, 46, 47, 50, 52, 51, 49, 48, 53, 58, 62, 61;

[0201] Its first-order difference absolute value sequence is: 1,1,3,2,1,2,1,5,5,4,1;

[0202] Then, the source-side activity within this window The mean of the absolute values ​​of these first differences is, i.e. Node volatility This represents the standard deviation of the standardized values ​​within the same window; the mean of the above 12 values ​​is approximately 51.83, and the calculated standard deviation is approximately 5.91. .

[0203] 2> For the business prior directed adjacency matrix Prior causal edges with a value of 1 Based on the number of business lag steps Computational lag consistency .

[0204] In practical implementation, to calculate lag consistency First, obtain the first The and the first The game operation and maintenance metrics are presented as a first-order difference sequence within the graph statistics window; then, based on the number of business lag steps... Align the sequence: align the first... The difference series of each indicator are shifted towards the historical direction. Each time step (even if its corresponding time index is subtracted as a whole) ), and then with the first The Pearson correlation coefficient was calculated from the original difference sequences of the indicators; this correlation coefficient was then linearly scaled to... The interval, that is, the result .

[0205] In one embodiment, as an example, let the prior causal edge be "memory occupancy ( ) Number of crashes ( "Business lag steps" The graph statistics window includes time steps 489-500; first, extract the first-order difference sequence of memory usage within these 12 time steps. Then, take the first-order difference sequence of the crash event count sequence shifted forward by two time steps (i.e., aligned time steps 491-500), and calculate the Pearson correlation coefficient between the two sequences; if the correlation coefficient is 0.72, after linear scaling (this value can be used directly or a simple truncation can be performed), map it to... Within the range, lag consistency is obtained. .

[0206] It should be noted that in calculating the first... Lag consistency at each time step At that time, use time steps Calculate the sequence within the graph statistics window at the end position, thereby... It becomes a quantity that changes over time.

[0207] 3> Based on source-side activity Target node volatility and lag consistency Calculate the original dynamic edge score , is represented as:

[0208] ;

[0209] in, Indicates the first a priori causal edge at each time step The original dynamic edge score, Describing a priori causal edges The basic edge bias is related to the specific a priori causal edge. The bound learnable scalar parameters (which are trainable parameters); Indicates the source-side activity weight. Indicates the target node volatility weight. Represents the lagged consistency weight; where, , and They can be initialized to 0.3, 0.3 and 0.4 respectively, and updated as learnable parameters during training.

[0210] It should be noted that, It refers to the first Individual source indicator nodes The degree of fluctuation within the graph statistics window is used to characterize the stable / unstable state of the information sender; It refers to the first Target indicator nodes The degree of fluctuation within the graph statistics window is used to characterize the stable / unstable state of the information receiver. When causal edges... middle When the indicator itself fluctuates wildly, it means that the target node may be undergoing significant changes due to certain factors. In this case, appropriate causal edge information transmission is more worthy of attention. Therefore, For the target node volatility, and They are independent statistics.

[0211] 4> For each target game operation and maintenance metric node Only when the conditions are met Softmax normalization is performed within the scope of the source game operation and maintenance indicator nodes to obtain dynamic edge weights. ;like Then directly order Dynamic edge weights Indicates the first The first time step The game operation and maintenance indicator node for the first Information transmission strength of each game operation and maintenance indicator node.

[0212] In one embodiment, as an example, let the current time step be... For the target node (Number of crash events), its prior source nodes include (Memory usage) and (Number of interface errors), i.e. , ; Calculate the original dynamic edge scores for each of the two edges. and For example, scores of 2.1 and 1.3; Softmax normalization of these two scores is expressed as:

[0213] ;

[0214] ;

[0215] Furthermore, other edges that do not exist prior to this ( Set it directly to 0.

[0216] 5> Add a self-connected edge to each game operation and maintenance indicator node. The self-connected edge means that the game operation and maintenance indicator node retains its own historical state. The dynamic edge weight of the self-connected edge participates in the normalization of the same target game operation and maintenance indicator node, so as to avoid the target game operation and maintenance indicator node from being completely dependent on other game operation and maintenance indicator nodes.

[0217] 6> Traverse all time steps to obtain the dynamic business directed adjacency sequence. Dynamic service directed adjacency sequence The size is ,in, Indicates the first The first time step The game operation and maintenance indicator node points to the first Dynamic edge weights of game operation and maintenance indicator nodes.

[0218] In practical implementation, for arrive At each time step, repeat operations 1 to 5 of S302 above: at each time step Within the graph statistics window, source-side activity, node volatility, and hysteresis consistency are calculated for all nodes that meet the criteria. The original dynamic edge score is calculated and then normalized using Softmax (including self-connected edges) to obtain the result. Stacking the dynamic adjacency matrices of all time steps yields a matrix of size [size missing]. Dynamic service directed adjacency sequence .

[0219] In one embodiment, as an example, the business prior directed adjacency matrix There exists a prior causal edge where "memory usage points to the number of crash events." If, within a certain hour, memory usage continuously increases and fluctuations intensify, and the number of crash events also increases after a lag of 1 to 2 time steps, then the source-side activity, target-node volatility, and lag consistency of this prior causal edge all increase, and the dynamic edge weight... As this increases, during subsequent graph message passing, the crash event count node will receive more information from the memory usage node.

[0220] It should be noted that this step combines a fixed business causal skeleton with a real-time statistical context. The fixed business causal skeleton ensures that the graph structure does not deviate from game operation and maintenance experience, while the real-time statistical context allows the same causal edge to have different strengths on ordinary days, event days, version update days, and fault impact periods, thereby reducing the risk of pseudo-correlation connections generated by pure correlation graph construction.

[0221] Step 140: Based on the operational decoupling feature tensor and the directed adjacency sequence of dynamic business in the historical input window, obtain the predicted game operation and maintenance index data using the preset spatiotemporal prediction model; weight the predicted game operation and maintenance index data of each preset type and the actual game operation and maintenance index data to obtain the comprehensive loss, and iteratively train to obtain the trained preset spatiotemporal prediction model.

[0222] In some embodiments, the specific steps are as follows:

[0223] Node feature mapping and crash event pulse coding:

[0224] The first The operational decoupling feature tensors within the historical input window of each training sample are mapped to initial node features. Simultaneously, the historical collapse event sequence is encoded into multi-scale exponentially decaying pulse features and injected into all nodes, resulting in initial node representations carrying collapse impact memories, which are then used for graph message passing in subsequent encoding layers. The specific steps are as follows:

[0225] 1> For the first A training sample, from the operational decoupling feature tensor By extracting historical input windows, the sample operation decoupling feature tensor is obtained. Sample operation decoupling feature tensor The size is ,in, Indicates the length of the history input window. This represents the total number of game operation and maintenance metrics, and 3 represents the three channels: trend component, operation cycle component, and enhanced residual component.

[0226] In specific implementation, for the first Each training sample, whose historical input window covers a time step range. Decoupling feature tensors from operational perspectives (size Extract all data within this window along the timeline. The data from each indicator and all three channels yields a size of [size missing]. Sample operation decoupling feature tensor ;

[0227] Right now ,in, This is the index of the relative time step within the historical input window.

[0228] 2> Decouple the feature tensor from the sample operation Input the linear mapping layer to obtain the initial node feature tensor. Initial node feature tensor The size is ,in, This represents the initial node feature dimension, which can be 32 or 64. For each historical time step and each game operation and maintenance metric node, the linear mapping layer maps the values ​​of the three channels to... Dimensional node feature vectors.

[0229] In its implementation, the linear mapping layer is a single-layer fully connected network (including bias), with a weight matrix of size [size missing]. The bias vector dimension is For each time step and each node The result is obtained by multiplying the three input channel values ​​(trend component, operating cycle component, and enhanced residual component) with the weight matrix and then adding a bias. Dimensional initial node feature vector.

[0230] 3> From the first The crash event count sequence is read from the historical input window of each training sample to obtain the sample crash event count sequence. ,in, This represents the relative time step index within the history input window, with a value ranging from 1 to... , Indicates the first The training sample at the th ... The number of crash events at each historical time step.

[0231] In one embodiment, as an example, suppose the first The historical input window for each training sample covers time steps. to ( ); from the sequence of crash events Extracting values ​​from this range yields... ,in Corresponding time step 100, Corresponding time step 387; for example, This indicates that two crash events occurred at the 150th relative time step within the sample's history window.

[0232] Furthermore, the sequence of sample crash events... Perform multi-scale exponential decay recursion to obtain the collapse event impulse state vector. .

[0233] In one implementation, three decay scales are set, corresponding to short-term impact, medium-term impact, and long-term aftershock, respectively; the short-term decay time can be 30 minutes, the medium-term decay time can be 3 hours, and the long-term decay time can be 12 hours. The pulse state of each scale is recursively calculated by adding the pulse state decay value of the previous time step to the current number of crash events.

[0234] In one embodiment, for example, three attenuation scales and attenuation factors are assumed. ,in Time step (if) ), Time step, Time step; for each scale The recursive method is expressed as: ,in, The pulse state vector is obtained by concatenating the pulse states at three scales. For example, suppose There had been no crashes before. ), ,but Step 151 If ,but .

[0235] 4> Convert the crash event pulse state vector Input the pulse projection layer to obtain the pulse feature vector of the collapse event. crash event pulse feature vector Dimensions and node feature dimensions Similarly, it is used to characterize the impact of historical collapse events on the current time step, which still exists.

[0236] In its implementation, the pulse projection layer is a single-layer fully connected network (including bias), and the weight matrix size is [size missing]. The bias vector dimension is , 3D pulse state vector Mapped to dimensional collapse event impulse feature vector .

[0237] 5> Determine the crash event pulse feature vector Add to the Based on the characteristics of all game operation and maintenance metrics nodes at each historical time step, a crash-enhanced node feature tensor is obtained. This operation indicates that crash events not only affect the crash event count node, but also affect multiple game operation and maintenance metrics nodes such as the number of online players, average latency, packet loss rate, reconnection count, CPU utilization, and memory utilization over a period of time.

[0238] In the specific implementation, for the first input in the history input window... Each time step will calculate the results. dimensional collapse event impulse feature vector Directly with all The initial node feature vectors of each game operation and maintenance indicator node are summed element by element. Specifically, let the first node be... The first time step The initial node feature vectors of the nodes are The node feature vector after collapse enhancement for: This is used to obtain the collapse-enhanced node feature tensor.

[0239] In one embodiment, as an example, suppose that at a certain historical time step, the 64-dimensional pulse feature vector output by the crash event pulse coding module... In the model, some dimensions have positive values, representing the impact of crashes at different scales: short, medium, and long. This 64-dimensional vector is added to the initial 64-dimensional feature vectors of all 10 indicator nodes, such as "number of online players" and "average latency". The impact information of the crash event is then explicitly injected into the feature representation of all operational indicators, rather than just affecting the single node of "number of crash events". This allows the model to learn the broad impact of crashes on other indicators.

[0240] Integrating directed graph message passing with causal temporal convolution to address business delays:

[0241] By stacking multiple encoding layers, directed graph message passing with business lag steps and causal temporal convolution are executed sequentially in each layer. The aim is to enable the features of each game operation and maintenance indicator node to propagate and aggregate along the direction specified by the business causal edge with a delay time that conforms to the business logic. At the same time, multi-scale causal temporal patterns are extracted in the time dimension, thereby generating an encoder feature tensor that integrates dynamic spatial dependency and temporal evolution information. The specific steps are as follows:

[0242] 1> From the directed adjacency sequence of dynamic services Extracting the first The training samples correspond to the directed adjacency sequence of the sample dynamic business. Sample dynamic service directed adjacency sequence The size is ,in, Indicates the first The training sample of the th training sample The first historical step The game operation and maintenance indicator node points to the first Dynamic edge weights of game operation and maintenance indicator nodes.

[0243] In practical implementation, the sample dynamic service directed adjacency sequence From the global dynamic business directed adjacency sequence (size is) The data is obtained by truncating along the time axis, specifically for the first... The training samples have a historical input window covering a time step range of [number]. ,but Depend on All slices within this range are composed of, and their dimensions are Among them, elements equal ;

[0244] Indicates the first The first time step The game operation and maintenance indicator node points to the first Dynamic edge weights of game operation and maintenance indicator nodes.

[0245] In one embodiment, as an example, suppose the first The historical input window for each sample starts at the 100th global time step, and the window length is... Its sample dynamic service directed adjacency sequence The first time step slice corresponds to the global dynamic service directed adjacency sequence. The 100th time step slice; the 2nd time step corresponds to the 101st global slice, and so on, until all 288 time steps have been extracted.

[0246] 2> In the In the coding layer, for the first... Target game operation and maintenance metrics nodes at each historical time step According to the sample dynamic business directed adjacency sequence Receive messages from the source game operation and maintenance metric node; the encoding layer index is denoted as... The number of coding layers can be 2 to 4.

[0247] In practical implementation, for a priori causal edges Based on the number of business lag steps From the first Source game operation and maintenance indicator nodes at each historical time step Read node characteristics to avoid mistakenly treating effects that should be delayed as synchronous effects.

[0248] In practical implementation, graph message passing occurs in the first... Implemented in the first encoding layer, for the first Target node at each historical time step Aggregates data from all source nodes. The message. Specifically, first determine the node. arrive Does a priori causal edge exist (i.e.)? If it exists, then it is based on the number of business lag steps. From the first The source node of each historical time step Obtain its node features from the current time step, rather than from the current time step. First, obtain the source node features; then, perform a linear transformation on the features of the source node, multiply them by the dynamic edge weights of the current time step, and finally sum them up.

[0249] In one embodiment, for example, let the number of business lag steps be... The current target node is the [number]th node in the historical window. Each time step; during message passing, the node From the source node In the The node features receive information at each time step; thus ensuring that the change in "memory usage" in step 8, after a delay of 2 time steps, only affects the "number of crash events" in step 10, simulating real business latency.

[0250] 3> Perform a linear transformation on the features of the source game operation and maintenance indicator nodes, multiply by dynamic edge weights, and then sum them to obtain the graph aggregation features of the target game operation and maintenance indicator nodes. The graph aggregation process can be represented as:

[0251] ;

[0252] in, Indicates the first In the coding layer, the first Target game operation and maintenance metrics nodes at each historical time step Graph aggregation features, Indicates the first Each coding layer source game operation and maintenance indicator node Node features at the time step after hysteresis alignment, where This indicates a lag alignment operation, i.e., retrieving the index. The maximum value in 1 ensures that the time index does not go out of bounds; Indicates the first The graph message linear transformation matrix of each coding layer is a trainable weight parameter; Indicates the first The first sample A historical step, from the node Pointing to node The dynamic edge weights are scalars.

[0253] 4> Aggregate graph features The features of the target game operation and maintenance indicator node are concatenated with its current features and then input into the feedforward transform layer. After passing through the linear rectified activation function and layer normalization, the graph update node features are obtained. The linear rectified activation function is used to retain positive and effective responses, and layer normalization is used to stabilize the feature scale between different game operation and maintenance indicator nodes.

[0254] In its implementation, the feedforward transform layer consists of two fully connected layers. Its structure is as follows: first, the input concatenated feature vector is mapped to a higher-dimensional hidden space through a linear transform layer (e.g., doubling the dimension); then, a linear rectified activation function is applied; next, it is mapped back to the original feature dimension through another linear transform layer; finally, layer normalization is performed.

[0255] 5> Perform causal temporal convolution along the time dimension for each game operation and maintenance indicator node. Causal temporal convolution only uses the current historical time step and the historical time steps before it, and does not use future time steps.

[0256] The kernel length can be 3, and the dilation rate can be 1, 2, and 4 respectively, so that the model can simultaneously perceive short-term mutations, medium-term trends, and long-term aftershocks. If the input length is insufficient for the historical length required for convolution, the node features of the earliest effective time step are used for boundary padding. That is, when performing causal temporal convolution, if the length of the historical input window is insufficient to support the receptive field required by the convolution kernel, the node feature vector of the earliest effective time step in the window is used to fill the missing time steps before the starting position of the historical window.

[0257] In one embodiment, for example, suppose the convolution kernel length is 3, the dilation rate is 4, and the current processing is the second time step within the historical window; at this time, the required inputs are the second step, the "2-4=-2" step, and the "2-8=-6" step, but the -2 and -6 steps do not exist, so the input features of these two steps are filled with the node features of the first time step of the window.

[0258] In its implementation, the causal temporal convolution is executed independently along the time dimension for each game operation and maintenance metric node. Its input is the graph update node features, and its output is the node features fused with local temporal context information, as detailed below:

[0259] Input: Graph update node feature tensor, size [value missing] ,in The length of the history input window. This represents the total number of game operation and maintenance metrics. The feature dimension of the current encoding layer;

[0260] b) Operation: For each game operation and maintenance metric node Its feature sequence along the time dimension is regarded as a sequence of length. The number of channels is A one-dimensional signal; a one-dimensional causal convolution is applied to this signal, the convolution using a length of... The convolution kernel has an inflation rate of 0.5%. ; in calculating the first At each time step, the convolution kernel will only cover indices less than or equal to... The input features will not use any Subsequent future information; to generate receptive fields of different scales, multiple convolutional kernels with different dilation rates (e.g., dilation rate of...) can be used. They are processed in parallel, and their outputs are concatenated or summed along the feature dimensions.

[0261] c> Output: The causal temporal convolution output feature tensor has the same size as the input. .

[0262] 6> The causal temporal convolution output and the graph update node features are added together through residual connections to obtain the node features of the next coding layer.

[0263] Furthermore, directed graph message passing and causal temporal convolution are repeated until all encoding layers are completed, yielding the encoder output feature tensor. The encoder outputs a feature tensor. The size is ,in, This indicates the dimension of the encoder output features, which can be 64 or 128.

[0264] In practice, the residual connection operation involves element-wise addition of the output feature tensor of the causal temporal convolutional layer to the initial input feature tensor of that coding layer. The result of this addition becomes the input of the next coding layer. After all the preset number of coding layers (e.g., 2 to 4 layers) have been executed, the node feature tensor output by the last layer becomes the encoder's output feature tensor. .

[0265] It should be noted that this step will involve the business lag step matrix. It can be directly used for graph message passing, making the information propagation order between the number of online players, the number of active rooms, system resources, network quality and crash events more in line with the game operation and maintenance logic; compared with ordinary synchronous graph convolution, it can more accurately model the fault evolution process of "business load changes first, resource pressure changes later, network quality deteriorates again, and crash events occur last".

[0266] Multi-game operation and maintenance metric joint prediction decoder:

[0267] Game operation and maintenance metrics have different numerical attributes. The number of online players, the number of active rooms, and the number of crash events are non-negative counting game operation and maintenance metrics. Average latency and bandwidth usage are non-negative continuous game operation and maintenance metrics. CPU utilization, memory utilization, and packet loss rate are bounded proportional game operation and maintenance metrics. If all game operation and maintenance metrics use the same output layer, it is easy to cause the predicted values ​​to not conform to the physical meaning, such as negative player counts or resource utilization exceeding 100%.

[0268] This step employs a joint predictive decoder for multiple game operation and maintenance metrics. By summarizing the encoder output with attention, constructing a future-oriented cross-attention query, and performing lightweight predictive metric graph convolution, it generates intermediate predictive features that integrate historical data and business dependencies. Finally, it generates multi-step future prediction values ​​that conform to their respective physical meanings through multiple categorized output heads. The specific steps are as follows:

[0269] 1> For the encoder output feature tensor Attention aggregation is performed along the time dimension to obtain the global node representation of each game operation and maintenance metric node. Global node representation Used to summarize the The trends, operating cycles, enhancement residuals, crash event pulses, and business graph propagation information of each game operation and maintenance indicator node within the historical input window.

[0270] In practical implementation, the first step is to define each game operation and maintenance metric node. Introducing a learnable global query vector Its dimension is the same as the encoder output feature dimension. Same; then, for nodes ,Will As a query, it retrieves all instances of that node within the history input window. The feature vectors at each time step are used as keys and values ​​to compute scaled dot product attention. The resulting attention weights are then weighted and summed over the value vectors at all time steps, thus aggregating the information of the node along the time dimension into a single vector, which is the global node representation. , The first All evolution information of each game operation and maintenance indicator node within the historical window.

[0271] 2> For each future time step and each game's operation and maintenance metrics Constructing future query vectors ,in, This represents the index of a future time step, with values ​​ranging from 1 to... , This represents a game operation and maintenance metric index, with values ​​ranging from 1 to... .

[0272] Specifically, future query vectors It is obtained by adding the future time location embedding, the game operation and maintenance indicator type embedding, and the future operation scenario tag embedding. The future operation scenario tag is obtained by looking up the game operation calendar. For example, if the future time step is in the version update window, the embedding vector corresponding to the version update day is used.

[0273] In the specific implementation, three embedding tables are first established (future time location embedding table, game operation and maintenance indicator type embedding table, and future operation scenario tag embedding table), and all embedding vector dimensions are [missing information]. For the first The next time step and the first A game operation and maintenance metric, its future query vector It is obtained by adding three parts, including:

[0274] Time and location embedding: based on Relative position in the future forecast window ( arrive (This can be obtained by looking up the table.)

[0275] Metric type embedding: Based on game operation and maintenance metrics The type (e.g., "number of online players") is obtained by looking up a table;

[0276] Operational scenario embedding: Based on the game's operational calendar, determine the operational scenario label for the future time step, and then look up the corresponding embedding vector in the table; for example, if the predicted future time step falls within the version update maintenance window, then use the embedding vector corresponding to "version update day";

[0277] Then, the three embedding vectors are summed element-wise to obtain the future query vector. .

[0278] 3> Use future query vectors Perform cross-attention on all global node representations to obtain the predicted intermediate features. Predicting intermediate features The model indicates that the prediction is for the first... The first future time step The comprehensive information read from all game operation and maintenance indicator nodes is a single game operation and maintenance indicator.

[0279] In the specific implementation, future query vectors are used. As a query, with all Global node representation of game operation and maintenance metrics nodes As the key and value, perform a scaled dot product attention calculation; this process enables Adaptively filter out historical summary information from all game operation and maintenance indicator nodes to predict the next... Step 1 The most relevant information for each metric is aggregated, and the resulting output is the predicted intermediate feature. .

[0280] For example, when predicting future average latency, cross-attention can read information from the global node representations corresponding to the number of active rooms, CPU utilization, bandwidth utilization, and packet loss rate.

[0281] 4> Within the same future time step, perform a single-layer lightweight predictive metric graph convolution on the intermediate features of all game operation and maintenance metrics predictions, enabling information exchange between future game operation and maintenance metrics; the connection relationships of the predictive metric graph convolutions originate from the business prior directed adjacency matrix. The subgraphs related to game operation and maintenance metrics are extracted from the data.

[0282] For example, the increase in future central processing unit utilization is related to the increase in future average latency, and predictive metric graph convolution can improve the consistency of prediction results for multiple game operation and maintenance metrics.

[0283] In the specific implementation, firstly, the business prior directed adjacency matrix is ​​used. Extracting and The subgraphs related to game operation and maintenance metrics are used to obtain the adjacency matrix of the predicted metrics. Its size is Adjacency matrix of predictive indicators The element in row r and column s Indicators of game operation and maintenance metrics arrive Does a prior business connection exist? Then, for the first... A future time step, to all Intermediate features of the prediction of each indicator As input node features; during graph convolution operations, for the target metric Aggregate all its source metrics (satisfy The features of the aggregated features are linearly transformed and activated to obtain updated intermediate features for prediction, so that the prediction of future central processing unit occupancy will refer to the prediction information of upstream business indicators such as the number of future active rooms.

[0284] 5> For non-negative counting game operation and maintenance metrics, a non-negative counting output header is set up. This header maps the predicted intermediate features to real numbers, and then uses the Softplus function to map them back to non-negative numbers, obtaining the non-negative count prediction value. This output header can be used for online player counts and active room counts.

[0285] In practical implementation, intermediate features will be predicted. Input a linear layer, map it to a real scalar Then, through the Softplus function, i.e. ,Will Convert to positive numbers to obtain non-negative count prediction values. Softplus guarantees that the output is strictly greater than 0, applicable to the number of online players and the number of active rooms.

[0286] 6> Set up a non-negative continuous output header for non-negative continuous game operation and maintenance indicators. The non-negative continuous output header outputs non-negative predicted values ​​through the Softplus function to obtain the predicted results of game operation and maintenance indicators such as average latency and bandwidth usage.

[0287] In practical implementation, intermediate features will be predicted. Input a linear layer, map it to a real scalar Then, through the Softplus function, i.e. Convert it to a non-negative number to obtain a non-negative continuous predicted value. This output header is used for metrics such as average latency and bandwidth usage.

[0288] 7> Set up a proportional output header for bounded proportional game operation and maintenance indicators. The proportional output header maps the prediction result to the range of 0 to 1 through the Sigmoid function, and then multiplies it by the upper limit of the proportion.

[0289] For example, if the CPU utilization rate during data acquisition is in percentage (%), the upper limit is 100; if the packet loss rate is in percentage (%), the upper limit is 100; if the unit is a ratio (0 to 1), the upper limit is 1. The upper limit value of 1 or 100 depends on the data normalization strategy. In this invention, the unit is percentage (%), so the upper limit is 100.

[0290] In practical implementation, intermediate features will be predicted. Input a linear layer, map it to a real scalar Then, the output is mapped to... using the Sigmoid function. The range; finally, multiply the value by the upper limit of the proportion. , to obtain the predicted value In this invention, all proportional indicators are expressed as percentages (%). Therefore, for CPU utilization, memory utilization, and packet loss rate, the upper limit is... All values ​​are 100. This represents the Sigmoid function.

[0291] 8> Set a zero-inflation event output header for the number of crash events. The zero-inflation event output header also outputs the probability of crash occurrence. and crash count intensity ,in, Mapped to the range of 0 to 1 by the Sigmoid function, representing the first... The probability of a crash occurring in a future time step; The Softplus function maps the values ​​to non-negative numbers, representing the crash count intensity when a crash event occurs.

[0292] In practical implementation, it involves predicting intermediate features. The process starts with two branches, including:

[0293] a>Collapse probability branch: After mapping to a scalar through a linear layer, the result is fed into the Sigmoid function to obtain the probability of collapse. ;

[0294] b) Crash Intensity Branch: After being mapped to a scalar through another independent linear layer, it is fed into the Softplus function to obtain the crash count intensity. ,in, Represented by natural constant As the base, the real scalar output of the linear layer An exponential function of an exponent is used to guarantee Nonnegativity;

[0295] Then, the predicted number of crash events at that time step is the product of the two.

[0296] Right now This output header enables most normal time steps... Approaching 0, thus naturally generating a large number of zero-value predictions.

[0297] 9> Traverse all future time steps and all game operation and maintenance indicators to obtain the prediction result matrix. Prediction result matrix The size is ,in, Indicates the first The training sample at the th ... The first future time step The predicted values ​​of each game operation and maintenance indicator are represented by the prediction result matrix. The element in the k-th row and r-th column.

[0298] It should be noted that this step sets the output headers according to the numerical attributes of the game operation and maintenance indicators, thereby ensuring that the predicted results such as the number of online players, the number of crash events, the average latency, and the resource utilization rate meet the business numerical constraints and improve the availability of the predicted results in the actual operation and maintenance system.

[0299] In one embodiment, such as Figure 5 As shown, this paper analyzes the decoupling of online player counts, including trends, operating cycles, and residual components. The original sequence of online player counts is decomposed from top to bottom into a trend component (exponential moving average), an operating cycle component (detrended value based on the median of historical data from the same time slot and operating scenario), and a residual component (original value minus trend and cycle). The operating calendar-driven decoupling effectively assigns explainable activity peaks to the cycle component, while sudden abnormal fluctuations remain in the residual. The horizontal axis represents time steps (dimensionless), and the vertical axis represents the number of online players (people). However, the trend, cycle, and residual components are simulated decompositions before standardization, and their dimensions are consistent with the original values.

[0300] This step designs loss function terms based on the different numerical attributes of game operation and maintenance indicators, and then weights and sums the various losses to form a comprehensive loss function. Based on this, the event-driven spatiotemporal prediction model is trained end-to-end to obtain the game operation and maintenance indicator prediction model. The specific steps are as follows:

[0301] 1> For the first training samples, from the prediction result matrix Take out the first one The first future time step Predicted values ​​of game operation and maintenance metrics and from the actual result matrix Extract the corresponding real value from ,in, Indicates the predicted value. Represents the actual value.

[0302] 2> For the number of online players, the Poisson loss function is used to calculate the prediction error. The Poisson loss is suitable for non-negative count data and can avoid the ordinary mean square error being overly sensitive to peak counts. During the calculation, the predicted value of the number of online players is restricted to non-negative, and a very small constant is added to the logarithmic operation to avoid illegal logarithms when the predicted value is close to 0.

[0303] 3> For average latency, the mean squared error loss function term is used to calculate the prediction error; average latency is a non-negative continuous game operation and maintenance indicator, and mean squared error can directly measure the deviation between the predicted latency and the actual latency.

[0304] 4> For CPU utilization, memory utilization, and packet loss rate, the mean squared error loss function term is used to calculate the prediction error; since the proportional game operation and maintenance indicators have been limited to the range of 0 to 100 through the proportional output head, the mean squared error can stably measure the percentage prediction deviation.

[0305] 5> For the number of crash events, the prediction error is calculated using a zero-inflated Poisson loss function term; the zero-inflated Poisson loss function term also uses the crash occurrence probability. and crash count intensity When the number of actual crash events is 0, the loss encourages a decrease in the probability of crashes occurring; when the number of actual crash events is greater than 0, the loss simultaneously encourages an increase in the probability of crashes occurring and makes the crash count intensity close to the number of actual crash events.

[0306] In practical implementation, if the actual number of crash events is 0, the loss is determined by the probability of no crash and the probability of a crash but a Poisson count of 0. If the actual number of crash events is greater than 0, the loss is determined by the crash probability term and the Poisson count term. Based on this, it can adapt to the distribution characteristics of a large number of time steps with 0 crash events and a small number of time steps with non-zero impacts.

[0307] 6> Set loss weights for different game operation and maintenance metrics ,in, Indicates the first The loss weights of each game operation and maintenance indicator are determined as follows: the number of online players and the number of crash events can be given higher weights, such as 1.5 or 2.0; the average latency, packet loss rate, CPU utilization, and memory utilization can be set to 1.0; the average of the weighted sums of the losses of all future time steps and all game operation and maintenance indicators is used to obtain the sample comprehensive loss.

[0308] 7. The batch loss is obtained by averaging the sample loss of all training samples within a training batch. An adaptive moment estimation optimizer is then used to update all trainable parameters in the event-driven spatiotemporal prediction model.

[0309] In one implementation, the initial learning rate can be set to... The batch size can be 16, 32, or 64, and the maximum number of training epochs can be 100 to 200. If the validation set loss does not decrease for several consecutive training epochs, the learning rate should be reduced or training should be stopped early.

[0310] 8> During training, training samples containing crash events are weighted and sampled so that they have a higher probability of being included in the training batch. Weighted sampling only changes the probability of a training sample being selected, without changing the sample label. Based on this, the problem of insufficient training caused by the scarcity of crash event samples can be alleviated.

[0311] 9> During training, a random mask is applied to the enhanced residual components in the operational decoupling feature tensor. Specifically, the enhanced residual components of some time steps and some game operation and maintenance indicators are randomly set to 0 with a small probability, which can be between 0.05 and 0.1. This operation simulates the acquisition noise, log latency, and missing some abnormal signals, making the game operation and maintenance indicator prediction model more robust to online missing and weak signals.

[0312] 10> After each training cycle, calculate the validation error of each game operation and maintenance indicator within the future prediction window on the validation set, and save the model parameters when the overall loss of the validation set is the lowest, as the final output game operation and maintenance indicator prediction model.

[0313] In one embodiment, as an example, suppose there is a training sample (denoted as the nth sample) with a future prediction window length of... The game operation and maintenance metrics that need to be predicted include the number of online players (denoted as ). ), average delay (denoted as ), CPU utilization (denoted as ), number of crash events (referred to as )common One indicator; now, using this sample in the future... Taking the prediction at the next time step (i.e., the 15th minute in the future) as an example, the loss calculation process is explained in the following steps:

[0314] At this time step, the predicted value output by the model (Number of online players) (Average latency, in milliseconds) (CPU utilization, in %), and two intermediate quantities: the probability of a crash and the number of crash events. Collapse count intensity ,

[0315] Therefore ;

[0316] The corresponding true value , , , The losses for each item are calculated as follows:

[0317] a> The number of online players is calculated using Poisson loss, with the predicted value being used in the calculation. It has been guaranteed to be non-negative; the loss is approximately ;

[0318] b) The average delay uses the mean squared error loss, i.e. ;

[0319] c> CPU utilization is calculated using mean squared error loss, i.e. ;

[0320] The number of crash events is calculated using zero-inflated Poisson loss, for the true value. (Greater than 0), the loss is determined by the probability of the collapse. Together with the Poisson count term (the negative log-likelihood with a true value of 2 under a count intensity of 1.8), it constitutes the total;

[0321] Furthermore, loss weights are set based on the number of online players. Number of crashes Average latency and CPU utilization After weighting and summing the four losses at that time step, then perform a full calculation. The average loss of a sample is obtained by taking the average of the losses of all training samples in a batch. Then, the average loss of all training samples in a batch is obtained by taking the average of the losses of all training samples in a batch. This batch loss is then fed into the optimizer for parameter updates.

[0322] During sampling, if a sample contains If the time step is specified, the sample will be selected with a higher probability; during training, some values ​​in the enhanced residual components of the sample will be randomly set to 0 to enhance the robustness of the model; training continues until the combined loss of the validation set no longer decreases, and the model parameters at this point are saved.

[0323] Step 150: Obtain the operation decoupling feature tensor and dynamic business directed adjacency sequence corresponding to the multi-dimensional game operation and maintenance index data to be predicted. Use the pre-trained spatiotemporal prediction model to obtain the predicted game operation and maintenance index data, and then determine whether a crash warning is triggered.

[0324] After training, the game operation and maintenance indicator prediction model will be deployed to the game operation and maintenance online monitoring system to perform joint prediction and anomaly warning of future game operation and maintenance indicators online. The specific steps are as follows:

[0325] 1> The online monitoring system is based on the sampling interval Continuously read game operation and maintenance metrics data and maintain a length of [length missing]. The historical sliding window; whenever new sampled data arrives, the historical sliding window discards the data from the earliest time step and adds the data from the latest time step to form a new historical input window to be predicted.

[0326] 2> Perform the same shock preservation standardization on the historical input window for prediction as during the training phase.

[0327] Specifically, the local median, local interquartile range, relative deviation, and indicator-level impact retention flag are calculated using historical data up to the current time step, without using future data. Based on the above results, a standardized game operation and maintenance indicator matrix corresponding to the historical input window to be predicted is generated.

[0328] 3. Based on the current game operation calendar, the execution trend, operation cycle, and residual of the historical input window to be predicted are decoupled to obtain the operation decoupled feature tensor corresponding to the historical input window to be predicted. The operation cycle estimation only uses historical operation data before the current time step to ensure that there is no leakage of future information in the online prediction process.

[0329] 4> Based on the business prior directed adjacency matrix Business lag step matrix Based on the statistical context within the current graph statistics window, a dynamic business directed adjacency sequence corresponding to the historical input window to be predicted is generated. If the current date is a version update day or a major event day, the dynamic business directed adjacency sequence will automatically adjust the causal edge weights of the business based on recent fluctuations in game operation and maintenance metrics.

[0330] 5> Input the operational decoupled feature tensor corresponding to the historical input window to be predicted and the directed adjacency sequence of dynamic business into the trained game operation and maintenance index prediction model, and output the future... Prediction result matrix within each time step Prediction Result Matrix Future predictions of game operation and maintenance metrics, including the number of online players, average latency, packet loss rate, CPU utilization, memory utilization, and number of crashes.

[0331] In one embodiment, as an example, let the future prediction window length be... (Next 1 hour) Number of game operation and maintenance metrics During online prediction, the output prediction result matrix Size is The element in the 5th row (corresponding to the 25th minute in the future) and the 1st column. This indicates that the predicted number of online players at the 5th time step is 1050; the element in the 5th row and 6th column is... This indicates that the probability of a crash occurring at this time step is extremely low, and the model did not anticipate a crash event.

[0332] 6> Based on the prediction result matrix An anomaly warning has been triggered.

[0333] Specifically, the probability of a collapse occurring at any future time step. If the crash probability threshold is exceeded, a crash risk warning will be triggered; the crash probability threshold can be set between 0.6 and 0.8. If the average latency in the future exceeds the latency threshold and persists for at least two future time steps, a latency risk warning will be triggered; the latency threshold can be set according to the game type, for example, 80ms or 100ms for real-time competitive games. If the packet loss rate in the future exceeds the packet loss threshold, a network quality warning will be triggered; the packet loss threshold can be set between 1% and 3%.

[0334] 7> Generate explanation information for abnormal warnings.

[0335] Specifically, the anomaly warning explanation information includes the future time step that triggered the warning, the game operation and maintenance indicator that triggered the warning, the corresponding predicted value, several game operation and maintenance indicators with the largest current enhanced residual components, and several business causal edges with the highest dynamic edge weights. For example, when a crash risk warning is triggered, the system can output explanation information such as "the dynamic edge weight pointing to the number of crash events in memory usage increases, the enhanced residual of the number of interface errors increases, and the enhanced residual of average latency increases," to help operation and maintenance personnel locate the source of the risk.

[0336] 8. The online monitoring system automatically repeats the above process at each sampling interval to continuously predict and warn of anomalies in the game server's operating status. If multiple consecutive sliding windows trigger the same type of warning, the warning level is increased; if the warning indicator returns to below the threshold and remains below it for several sampling intervals, the corresponding warning status is lifted.

[0337] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for multi-dimensional management of game operation data, characterized in that, The method includes: Collect multi-dimensional game operation and maintenance indicator data during the operation of the game server cluster, and construct the original game operation and maintenance indicator matrix; The original game operation and maintenance indicator matrix is ​​subjected to impact preservation standardization to obtain a standardized game operation and maintenance indicator matrix; the standardized game operation and maintenance indicator matrix is ​​decoupled from its trend, operation cycle and residual using the operation calendar to obtain the operation decoupling feature tensor. Each game operation and maintenance indicator data is treated as a game operation and maintenance indicator node. Based on the prior connection state between game operation and maintenance indicator nodes, a business prior directed adjacency matrix is ​​constructed, and dynamic edge weights corresponding to the prior causal edges in the business prior directed adjacency matrix are generated. Based on the dynamic edge weights and preset self-connected edges, all time steps are traversed to obtain the dynamic business directed adjacency sequence. Based on the operational decoupling feature tensor and the dynamic business directed adjacency sequence in the historical input window, the predicted game operation and maintenance index data is obtained by using the preset spatiotemporal prediction model; the predicted game operation and maintenance index data of each preset type and the actual game operation and maintenance index data are weighted to obtain the comprehensive loss, and the preset spatiotemporal prediction model is trained iteratively. The system obtains the operational decoupling feature tensor and dynamic business directed adjacency sequence corresponding to the multi-dimensional game operation and maintenance index data to be predicted. It then uses a pre-trained spatiotemporal prediction model to obtain the predicted game operation and maintenance index data, thereby determining whether a crash warning is triggered. The process of collecting multi-dimensional game operation and maintenance indicator data from the game server cluster during operation and constructing an original game operation and maintenance indicator matrix specifically includes: Based on a preset sampling interval, multi-dimensional game operation and maintenance indicator data are collected during the operation of the game server cluster. The game server cluster includes at least: a game server monitoring system, a log collection system, a crash reporting system, and a network quality monitoring system. The collection objects corresponding to the game operation and maintenance indicator data include at least: number of online players, number of active rooms, CPU utilization, memory utilization, bandwidth utilization, average latency, packet loss rate, number of reconnections, number of interface errors, number of crash events, and operation scenario tags. Align the game operation and maintenance metrics data according to the sampling timestamp; The process of performing impact-preserving standardization on the original game operation and maintenance indicator matrix to obtain a standardized game operation and maintenance indicator matrix specifically includes: With the current time step as the end position, it is the nth A sliding statistical window for the most recent preset time period is constructed from each game operation and maintenance metric; Calculate the first [number] within the sliding statistical window of the most recent preset time period. Local median of game operation and maintenance metrics and local interquartile range ; Calculate the first The first time step Relative deviation of game operation and maintenance metrics ; Obtain the core fault-sensitive indicator set; wherein, the core fault-sensitive indicator set consists of game operation and maintenance indicators that have preset abnormal changes corresponding to crash events; For each time step, count the number of game operation and maintenance indicators in the core fault sensitive indicator set whose relative deviation exceeds the candidate threshold; Verify that each time step meets any of the following conditions: The number of crash events is greater than 0; Alternatively, the number of game operation and maintenance indicators whose core fault-sensitive indicator set exceeds the candidate threshold is greater than the preset deviation threshold; Alternatively, if the average latency and packet loss rate both exceed the 90th percentile of the sliding statistical window of their respective most recent preset time period, then it is marked as a candidate time step for a crash impact. When the The first time step is a candidate time step for the collapse impact, and the second time step is a candidate time step for the collapse impact. Relative deviation of game operation and maintenance metrics Greater than the preset candidate threshold At half the speed, the standard impact retains the mark. The value is 1; otherwise, the impact level is retained. The value is 0; Through the formula: Calculate the first The first time step Robust standardized median values ​​of game operation and maintenance metrics ; in, Represents the original collected values. This represents the local median. Indicates the local interquartile range. Preset non-zero values; Through the formula: Obtain standardized game operation and maintenance indicator values ; in, This indicates the parameter for adjusting the impact retention strength. This indicates the upper limit of the relative deviation cutoff. This indicates the upper limit of the truncation value for standardized game operation and maintenance metrics. This represents the truncation function. The marking indicates that the impact level is retained. Indicates the relative deviation; Will As a standardized game operation and maintenance indicator matrix The elements in row t and column d are used to construct a standardized game operation and maintenance indicator matrix. .

2. The multi-dimensional management method for game operation data according to claim 1, characterized in that, By using the operations calendar to decouple the execution trend, operations cycle, and residuals of the standardized game operations and maintenance indicator matrix, an operations decoupling feature tensor is obtained, which specifically includes: For each time step Calculate the time slot index Among them, time slot index This indicates the current time step's position within a day; Standardized game operation and maintenance indicator matrix Each game operation and maintenance metric in the matrix is ​​subjected to an exponential moving average to obtain a trend component matrix. Among them, the trend component matrix The size is ,and Indicates the first The first time step Trend components of game operation and maintenance metrics; For each time step and each game's operational metrics At the current time step Searching for the index of the current time slot within the previously preset historical time range. Same or adjacent, and with operational scenario tags Same historical time step; Iterate through each time step and each game's operational metrics The operating cycle component matrix is ​​obtained. Among them, the operating cycle component matrix The size is ,and Indicates the first The first time step The operational cycle component of each game's operation and maintenance metrics; Through the formula: Calculate the residual component matrix The Middle The first time step Residual components of game operation and maintenance metrics Thus, the residual component matrix is ​​obtained. ;in, Represents a standardized game operation and maintenance indicator matrix The element in row t and column d; When the indicator-level shock retains the mark If the value is 1, then the residual components will be... Multiply by the preset residual enhancement factor To obtain enhanced residual components When the indicator-level shock retains the mark. If the value is 0, the residual components are preserved. Without changing, we obtain the strong residual component matrix. ; Trend component matrix Operating cycle component matrix and enhanced residual component matrix By concatenating along the channel dimension, we obtain the operational decoupling feature tensor. .

3. The multi-dimensional management method for game operation data according to claim 1, characterized in that, Each game operation and maintenance metric data point is treated as a game operation and maintenance metric node. Based on the prior connection status between game operation and maintenance metric nodes, a business prior directed adjacency matrix is ​​constructed, specifically including: Each game operation and maintenance metric is treated as a game operation and maintenance metric node, resulting in a set of game operation and maintenance metric nodes. ,in, Indicates the first Each game operation and maintenance metric corresponds to a game operation and maintenance metric node; Construct a business prior directed adjacency matrix Business prior directed adjacency matrix The size is ,in, Indicates the first The game operation and maintenance metric node points to the first The prior connection state of each game operation and maintenance indicator node is a business prior directed adjacency matrix. The element in the i-th row and j-th column; when the... The game operation and maintenance metrics can directly affect the first When considering game operation and maintenance metrics... The value is 1; otherwise, The value is 0; A priori directed adjacency matrix for business purposes For each prior causal edge in the matrix, a business lag step number is set, resulting in a business lag step matrix. Among them, the business lag step matrix The size is ,and Indicates the first The game operation and maintenance indicator node for the first The discrete lag steps when a game operation and maintenance indicator node has an impact.

4. The multi-dimensional management method for game operation data according to claim 1, characterized in that, The dynamic edge weights corresponding to the prior causal edges in the business prior directed adjacency matrix are generated, specifically including: For each time step Construct a graph statistics window that includes the current time step. and the current time step A previously preset historical time step; For each game operation and maintenance metric node Calculate source-side activity within the graph statistics window. and node volatility Among them, source-side activity The mean of the first-order absolute differences of standardized game operation and maintenance indicator values ​​within the statistical window; node volatility. The standard deviation of standardized game operation and maintenance indicator values ​​within the graph statistics window is calculated. For the business prior directed adjacency matrix Prior causal edges with a value of 1 Based on the number of business lag steps Computational lag consistency ; Through the formula: Calculate the original dynamic edge score ; in, Describing a priori causal edges Basic edge offset Indicates the source-side activity weight. Indicates the target node volatility weight. Indicates the lagged consistency weight; For each game operation and maintenance metric node Only when the conditions are met The dynamic edge weights are obtained by performing Softmax normalization on the original dynamic edge scores within the edge range corresponding to the game operation and maintenance indicator nodes. ;like Then directly order Dynamic edge weights Indicates the first The first time step The game operation and maintenance indicator node for the first Dynamic edge weights of game operation and maintenance indicator nodes.

5. The multi-dimensional management method for game operation data according to claim 1, characterized in that, Based on dynamic edge weights and pre-defined self-connecting edges, the dynamic business directed adjacency sequence is obtained by traversing all time steps, specifically including: Add a self-connecting edge to each game operation and maintenance metric node; where the self-connecting edge represents the historical state that the game operation and maintenance metric node retains itself; Iterate through all time steps, stack the dynamic adjacency matrices of all time steps, and obtain the dynamic business directed adjacency sequence. Among them, dynamic services include directed adjacency sequences. The size is And the dynamic service directed adjacency sequence includes ,and Indicates the first The first time step The game operation and maintenance metric node points to the first Dynamic edge weights of game operation and maintenance indicator nodes.

6. The multi-dimensional management method for game operation data according to claim 1, characterized in that, Based on the operational decoupling feature tensor and dynamic business directed adjacency sequence within the historical input window, predictive game operation and maintenance index data is obtained using a pre-defined spatiotemporal prediction model, specifically including: The first The operational decoupling feature tensor within the historical input window of each training sample is mapped to the initial node features. At the same time, the number of collapse events corresponding to the historical input window is encoded as multi-scale exponentially decaying pulse features and injected into all nodes to obtain a collapse-enhanced node feature tensor carrying the memory of collapse impact. From dynamic business directed adjacency sequence Extracting the first The training samples correspond to the directed adjacency sequence of the sample dynamic business. ,in, Indicates the first The training sample of the th training sample The first historical step The game operation and maintenance metric node points to the first The dynamic edge weights of the game operation and maintenance indicator nodes; in the first... In the coding layer, for the first Game operation and maintenance metrics nodes at each historical time step According to the sample dynamic business directed adjacency sequence The node features of the game operation and maintenance indicator nodes are obtained, the node features are linearly transformed, multiplied by the dynamic edge weight of the current time step, and finally summed to obtain the graph aggregation features of the game operation and maintenance indicator nodes. The preset spatiotemporal prediction model concatenates the graph aggregation features with the current features of the game operation and maintenance index nodes and inputs them into the feedforward transformation layer. After passing through the linear rectified activation function and layer normalization of the preset spatiotemporal prediction model, the graph update node features are obtained. The pre-defined spatiotemporal prediction model performs causal temporal convolution along the time dimension on each game operation and maintenance indicator node. The causal temporal convolution only uses the current historical time step and the historical time steps before it, and outputs the temporal convolution output feature tensor. The causal temporal convolution output is added to the graph update node features through residual connections to obtain the node features of the next encoding layer. Repeatedly execute directed graph message passing and causal temporal convolution until all encoding layers are completed, obtaining the encoder output feature tensor. ; For encoder output feature tensor Attention aggregation is performed along the time dimension to obtain the global node representation of each game operation and maintenance metric node. ; For each future time step and each game's operational metrics Constructing future query vectors ; Using future query vectors Perform cross-attention on all global node representations to obtain intermediate predicted features. ; Within the same future time step, a single-layer lightweight predictive indicator graph convolution is performed on the intermediate features of all game operation and maintenance indicators to map them to real scalars. For non-negative count-type game operation and maintenance indicators, the real scalars are converted to positive numbers using the Softplus function to obtain the corresponding predicted values. For non-negative continuous game operation and maintenance indicators, the corresponding predicted values ​​are output using the Softplus function. For bounded proportional game operation and maintenance indicators, the prediction results are mapped to the range of 0 to 1 using the Sigmoid function, and then multiplied by a preset upper limit of the proportion to obtain the corresponding predicted values. For the number of crash events, the real scalar is fed into the Sigmoid function to obtain the probability of a crash; then fed into the Softplus function to obtain the crash count intensity; the product of the probability of a crash and the crash count intensity is used as the predicted value corresponding to the number of crash events. By iterating through all future time steps and all game operation and maintenance metrics, a prediction result matrix containing the predicted values ​​corresponding to the game operation and maintenance metrics is obtained.

7. The multi-dimensional management method for game operation data according to claim 5, characterized in that, The predicted game operation and maintenance indicator data and the actual game operation and maintenance indicator data of each preset type are weighted to obtain a comprehensive loss. The model is then iteratively trained to obtain a pre-trained preset spatiotemporal prediction model, which specifically includes: For the The training sample is taken from the prediction result matrix. The first step in the future The predicted values ​​of each game operation and maintenance indicator are used as the predicted game operation and maintenance indicator data, and the corresponding real values ​​are taken from the real result matrix as the actual game operation and maintenance indicator data. For game operation and maintenance metrics such as the number of online players, the Poisson loss function term is used to calculate the prediction error between the predicted game operation and maintenance metric data and the actual game operation and maintenance metric data. For game operation and maintenance metrics of average latency, the mean squared error loss function term is used to calculate the prediction error. For game operation and maintenance metrics such as CPU utilization, memory utilization, and packet loss rate, the mean squared error loss function term is used to calculate the prediction error. For game operation and maintenance metrics related to the number of crash events, the prediction error is calculated using a zero-inflated Poisson loss function term. The sample comprehensive loss is obtained by multiplying the preset loss weight corresponding to each preset type by the prediction error and summing the results. The batch loss is obtained by averaging the sample loss of all training samples in a training batch. Based on the batch loss, the adaptive moment estimation optimizer is used to update all trainable parameters in the event-driven pre-set spatiotemporal prediction model. During the training process, training samples containing collapse events are weighted and random masks are applied to the enhanced residual components in the operation decoupling feature tensor. After each training cycle, the validation error of each game operation and maintenance indicator within the future prediction window is calculated on the validation set, and the model parameters with the lowest overall loss on the validation set are saved as the final output of the trained preset spatiotemporal prediction model.

Images