Notebook computer fault trend prediction system based on time sequence modeling

Abstract

The invention discloses a notebook computer fault trend prediction system based on time sequence modeling, and the system comprises a period marking module which is used for constructing a life cycle feature data set, carrying out the clustering operation through an improved HDBSCAN algorithm, and constructing a life cycle state tag; the trend prediction module is used for outputting a standard trend prediction result through the improved DRN model; the state control module is used for executing state change frequency analysis and generating a state smoothing factor; the burst identification module is used for constructing a sudden change signal channel and outputting a burst trend prediction result; and the trend correction module is used for constructing an abnormal memory cache region, generating an abnormal early warning bias vector and outputting a corrected trend prediction result. According to the method, the fine-grained identification capability of trend prediction and the response sensitivity of abnormal change are effectively improved, and continuous, stable and accurate modeling and prediction of the potential fault trend in the full life cycle of the notebook computer are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of failure prediction, and particularly relates to a notebook failure trend prediction system based on time series modeling. BACKGROUND

[0002] With the wide deployment of notebook devices in office, education and industrial scenarios, failure prediction technology based on device operation logs, hardware sensor data and usage behavior information has gradually become a key means to improve device stability and reduce maintenance costs. Existing failure prediction systems mostly rely on static rule bases or single time series models to analyze the trends of device performance indicators, but in actual applications, the following problems are generally present:

[0003] Current systems usually perform single-point anomaly detection based on fixed time windows or discrete events, making it difficult to capture trend deviation signals in the early latent stage of failure, resulting in a lag in identifying early signs of failure; traditional trend prediction methods are mostly based on linear regression, moving average or standard LSTM networks, which have weak generalization ability and accumulated prediction errors when facing complex scenarios with periodic fluctuations, high-frequency mutations and state jumps in the running process of notebook devices; in addition, existing models lack state-aware adjustment capabilities in response mechanisms to sudden signals, and cannot dynamically adjust the prediction path according to the life cycle stage, usage load or historical failure patterns, resulting in a significant decrease in prediction accuracy when the device is close to failure; although some methods introduce memory mechanisms or cache structures, they still lack effective trend correction modules after abnormal state feedback, making it difficult for the system to achieve closed-loop repair of prediction bias.

[0004] Therefore, how to provide a notebook failure trend prediction system based on time series modeling is a problem that those skilled in the art need to solve. SUMMARY

[0005] One objective of this invention is to propose a notebook computer fault trend prediction system based on time-series modeling. This invention constructs a period labeling module, a trend prediction module, a state control module, a burst identification module, and a trend correction module. The system employs an improved HDBSCAN algorithm to perform clustering operations on a lifecycle feature dataset constructed from multi-source operational data, building lifecycle state labels. An improved DRN model is used to perform standard trend prediction on the multi-source operational data and lifecycle state labels, and a state sliding window is constructed to generate a state smoothing factor to control the switching mechanism of the prediction path. The system constructs a mutation signal channel through the burst identification module, activates the abnormal time-series prediction sub-path to perform burst trend prediction tasks, and further constructs an abnormal memory cache through the trend correction module, performs periodic training to generate an abnormal early warning bias vector, performs weighted correction on the standard trend prediction results, and finally outputs the corrected trend prediction results. This invention has advantages such as distinguishable lifecycle stages, dynamic control of prediction paths, rapid response to burst trends, and adaptive correction of prediction results.

[0006] A notebook computer failure trend prediction system based on time series modeling according to an embodiment of the present invention includes:

[0007] The lifecycle labeling module is used to collect multi-source runtime data of the target laptop and construct a lifecycle feature dataset. The improved HDBSCAN algorithm is used to perform clustering operations on the lifecycle feature dataset, and lifecycle state labels are constructed based on the clustering results. The lifecycle state labels include initial state, stable state and aging state.

[0008] The trend prediction module is used to input the multi-source operational data and lifecycle status labels into the improved DRN model and output standard trend prediction results; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths;

[0009] The state control module is used to construct a state sliding window, record the life cycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; based on the state smoothing factor, it controls the switching delay triggering of the time-series prediction sub-path; if the state change frequency is greater than a first preset threshold, the time-series prediction sub-path corresponding to the life cycle state label is kept in an active state.

[0010] The sudden change identification module is used to monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel. If the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, the sudden trend prediction task is executed through the abnormal time series prediction sub-path without the control and judgment of the state smoothing factor, and the sudden trend prediction result is output.

[0011] The trend correction module is used to construct an anomaly memory cache, record multi-source operating data and sudden trend prediction results before and after the triggering of the mutation signal channel, generate anomaly memory samples based on similarity matching, input the anomaly memory samples into the improved DRN model to perform periodic training, generate anomaly warning bias vector, and correct the standard trend prediction results according to the anomaly warning bias vector, and output the corrected trend prediction results.

[0012] Preferably, the steps between modules include the following:

[0013] S1. Collect multi-source runtime data from the target laptop and construct a lifecycle feature dataset;

[0014] S2. Perform clustering operations on the lifecycle feature dataset using the improved HDBSCAN algorithm, and construct lifecycle state labels based on the clustering results; the lifecycle state labels include initial state, stable state, and aging state;

[0015] S3. Input the multi-source operation data and lifecycle status labels into the improved DRN model and output the standard trend prediction result; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths;

[0016] S4. Construct a state sliding window, record the lifecycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; control the switching delay triggering of the time-series prediction sub-path based on the state smoothing factor; if the state change frequency is greater than a first preset threshold, keep the time-series prediction sub-path corresponding to the lifecycle state label in an active state.

[0017] S5. Monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel; if the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, then without the control and judgment of the state smoothing factor, execute the sudden trend prediction task through the abnormal time series prediction sub-path and output the sudden trend prediction result.

[0018] S6. Construct an abnormal memory cache area to record multi-source operation data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generate abnormal memory samples based on similarity matching; input the abnormal memory samples into the improved DRN model to perform periodic training, generate an abnormal warning bias vector, and correct the standard trend prediction results according to the abnormal warning bias vector, and output the corrected trend prediction results.

[0019] Preferably, the multi-source operating data includes cumulative runtime, CPU temperature change, battery voltage fluctuation, fan speed stability index, and device load response index, specifically:

[0020] The cumulative runtime records the continuous running time interval of the laptop from power-on to shutdown, and records it in the time series table according to the set sampling frequency; the CPU temperature change value is obtained by constructing a CPU temperature sampling sequence, calculating the temperature change amplitude between adjacent sampling points, and recording the temperature rise and fall trend information; the battery voltage fluctuation amplitude data is obtained by obtaining the real-time battery voltage value sequence, calculating the difference between the maximum and minimum voltage values ​​per unit time, and recording the voltage fluctuation interval; the fan speed stability index is obtained by constructing a fan speed time series, calculating the ratio of the speed standard deviation to the mean based on a sliding time window, and recording the speed change stability information; the device load response index is obtained by obtaining the system load state sequence, recording CPU utilization, memory utilization, and IO response latency indicators, and constructing corresponding time series; the multi-source running data is aligned according to a unified sampling timestamp to construct a life cycle feature dataset.

[0021] Preferably, the improved HDBSCAN algorithm is as follows:

[0022] It receives partial labeled state information from the lifecycle feature dataset, constructs a sample adjacency constraint matrix based on the lifecycle stage labels, adjusts the distance graph structure between samples based on the constraint matrix, and guides the construction of the minimum spanning tree path.

[0023] Based on the local sample density changes in the life cycle feature dataset, the minimum number of points for sample clustering is dynamically set, the clustering structure parameters are adjusted according to local density differences, and the clustering threshold is adjusted in the density fluctuation region.

[0024] Based on the timestamp information in the lifecycle feature dataset, the time distance between sample pairs is calculated. A time weighting factor is set according to a preset exponential function, and a time weighting operation is performed on the original sample distance to generate a time-weighted distance matrix.

[0025] Based on the density jump amplitude and life cycle state continuity between adjacent clusters in the clustering results, the density boundary fitting residual value is calculated, and the adjacent clusters are merged based on the residual threshold to generate a corrected life cycle state cluster set; the life cycle state label results, sample state confidence scores and clustering residual metric values ​​are output.

[0026] Preferably, the construction of lifecycle state labels based on clustering results specifically includes:

[0027] Based on the revised lifecycle state cluster set, the mean of the running time distribution, the average level of the equipment load response index, and the average value of the battery voltage fluctuation amplitude of the samples in each cluster are statistically analyzed to generate a preliminary state cluster attribute matrix.

[0028] The preliminary state cluster attribute matrix is ​​input into the state determination rule set. Based on the preset operating cycle stage division standard, battery performance stability threshold and load response fluctuation boundary, multi-condition judgment is performed to generate a state label candidate set.

[0029] For each state label in the state label candidate set, a state label confidence factor is constructed by combining its corresponding cluster sample density center and label distribution tilt, and the life cycle state label is weighted and corrected based on the confidence factor.

[0030] The weighted and corrected lifecycle state labels are mapped to the standard state set and marked as initial state, stable state and aging state, generating a lifecycle state label result table; the label result table is output and the label results are written into the state field of the lifecycle feature dataset.

[0031] Preferably, the improved DRN model includes a state-aware controller, a path selector, a set of time-series predicted sub-paths, and anomaly time-series predicted sub-paths, specifically:

[0032] The state-aware controller receives lifecycle state labels, sample state confidence scores, and multi-source operational data features, and generates a state control signal, which includes a path selection identifier, a confidence threshold value, and a state weight vector.

[0033] The path selector performs path activation judgment based on the state control signal and the life cycle state label sequence recorded in the state sliding window. If the state change frequency is less than the first preset threshold, the time-series prediction sub-path corresponding to the state is kept active. If the state change frequency is greater than the first preset threshold, the time-series prediction sub-path is delayed in switching.

[0034] The time series prediction sub-path set includes the initial prediction path, the stable prediction path, and the aging prediction path. The initial prediction path receives short-cycle running samples from the life cycle feature dataset and performs a fast trend prediction task; the stable prediction path receives medium-cycle sample sequences from the life cycle feature dataset and performs a steady trend prediction task; the aging prediction path receives long-cycle running samples from the life cycle feature dataset and performs a decline trend prediction task.

[0035] The abnormal time series prediction sub-path is constructed independently of the time series prediction sub-path set. It receives multi-source running data after the sudden change signal channel is triggered and performs the sudden trend prediction task. The abnormal time series prediction sub-path includes multi-scale residual convolutional layers and causal convolutional network layers, and outputs the sudden trend prediction result.

[0036] The improved DRN model receives the output results of the time series prediction sub-path set and the abnormal time series prediction sub-path, performs result fusion calculation, corrects the standard trend prediction results based on the abnormal warning bias vector, and generates the corrected trend prediction results.

[0037] Preferably, S4 specifically includes:

[0038] Set the time length and sliding step size parameters of the state sliding window, receive the life cycle state label sequence, and record the number of times each state label appears in a continuous time period;

[0039] The frequency of state changes is calculated based on the time sequence of lifecycle state labels, a sequence of state change frequencies is generated, and a histogram of state change frequencies is constructed.

[0040] The state smoothing factor is calculated based on the histogram of state change frequencies, and an inverse mapping relationship is established between the state smoothing factor and the state change frequencies.

[0041] The state smoothing factor is input into the path selector to control the switching delay triggering process of the timing prediction sub-path;

[0042] If the frequency of state changes is greater than the first preset threshold, the time-series prediction sub-path corresponding to the lifecycle state label remains active; if the frequency of state changes is less than the first preset threshold, the path selector is triggered to perform a delayed switching operation to update the active time-series prediction sub-path.

[0043] Preferably, S5 specifically includes:

[0044] Define a set of key operating indicators; the set of key operating indicators includes temperature parameters, voltage parameters, current parameters, and the ratio of fan speed to processor load.

[0045] Collect continuous time series of the key operating indicators, perform difference calculation between adjacent time points, and generate a rate of change series of the key operating indicators;

[0046] Perform sliding window mean and standard deviation analysis on the rate of change sequence to construct a dynamic threshold interval for the rate of change; determine whether the rate of change is greater than a second preset threshold; if the rate of change of any key operating indicator is greater than the second preset threshold, mark it as a mutation state; input the mutation state to the mutation signal channel, and activate the abnormal time series prediction sub-path without the control and judgment of the state smoothing factor;

[0047] Multi-source operational data under abrupt changes is input into the abnormal time series prediction sub-path, and then passed through multi-scale residual convolutional layers and causal convolutional network layers in sequence to output the sudden trend prediction results.

[0048] Preferably, the construction of the abnormal memory cache, which records multi-source operational data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generates abnormal memory samples based on similarity matching, specifically involves:

[0049] Set the time window range and sampling frequency parameters of the abnormal memory buffer, and record the multi-source running data sequence before the triggering of the sudden signal channel and the multi-source running data sequence after the triggering of the sudden signal channel; the time window range includes the first preset time period before the triggering of the sudden signal channel and the second preset time period after the triggering of the sudden signal channel.

[0050] Align the multi-source operational data in the abnormal memory buffer according to the timestamp order to construct time series data pairs; calculate the difference values ​​of each key operational indicator in the multi-source operational data before and after the triggering of the sudden signal channel to generate a multi-source difference feature matrix;

[0051] Based on the difference between the sudden trend prediction result and the standard trend prediction result, a trend offset vector is constructed; the trend offset vector is concatenated with the multi-source difference feature matrix to generate an abnormal feature vector set.

[0052] Based on preset similarity calculation rules, similarity matching is performed on samples in the abnormal feature vector set; the similarity calculation rules include Euclidean distance, cosine similarity and dynamic time warping distance; the similarity score between each sample is calculated to generate a similarity score matrix; sample pairs with similarity scores greater than a third preset threshold are selected from the similarity score matrix to construct an abnormal memory sample set.

[0053] Preferably, the step of inputting the abnormal memory samples into the improved DRN model for periodic training to generate an anomaly warning bias vector, and correcting the standard trend prediction result based on the anomaly warning bias vector to output the corrected trend prediction result, specifically involves:

[0054] Set the periodic training period and sample extraction batch parameters, select an abnormal memory sample set from the abnormal memory cache, and construct the training input sequence; the training input sequence includes multi-source running data, life cycle state labels, sudden trend prediction results, and trend offset vectors;

[0055] The training input sequence is fed into the improved DRN model, and the periodic training process is performed sequentially through the state-aware controller, path selector, set of temporal prediction sub-paths and abnormal temporal prediction sub-paths. The state-aware controller generates a state control signal based on the lifecycle state label and sample state confidence score. The path selector determines the activated temporal prediction sub-path or abnormal temporal prediction sub-path based on the state control signal.

[0056] A training loss set is constructed based on the difference between the sudden trend prediction result and the standard trend prediction result; the training loss set includes a trend offset error term, a confidence weight error term, and a state residual term; the parameter weights of the improved DRN model are updated based on the training loss set;

[0057] After the training period ends, the set of bias parameters used for trend correction in the improved DRN model is extracted to construct the anomaly warning bias vector;

[0058] The anomaly warning bias vector is input into the result fusion process. Based on the state control signal output by the state-aware controller and the activation result of the path selector, the standard trend prediction result is weighted and corrected, and the corrected trend prediction result is output.

[0059] The beneficial effects of this invention are:

[0060] This invention addresses the issues of insufficient state recognition accuracy, unstable trend prediction path switching, and delayed response to sudden faults in multi-source operating data of laptops by jointly integrating the improved HDBSCAN algorithm and the improved DRN model. It employs sample adjacency constraint matrix construction and time-weighted distance correction to guide density clustering path optimization. Combined with state labels from the lifecycle feature dataset, it generates state confidence scores and density residual metrics, outputting lifecycle state labels. These labels, along with multi-source operating data, are then input into the improved DRN model. A state-aware controller and path selector are constructed to dynamically regulate the activation state of the prediction path. A state smoothing factor controls the path switching delay strategy, and a multi-path parallel prediction structure for initial, stable, and aging stages is built to achieve multi-stage trend modeling of the lifecycle. To address the trend shift problem caused by sudden changes in key operating indicators, a mutation signal channel and anomaly time-series prediction sub-path are constructed. Multi-scale residual convolution and causal convolution networks are invoked to respond to sudden trend changes. Simultaneously, an anomaly memory cache records the difference features and trend shift vectors of the operating data before and after the mutation. These are input into the improved DRN model for periodic training, generating anomaly warning bias vectors and weighting and correcting the standard trend prediction results. Ultimately, it achieves stable identification of lifecycle trend status during laptop operation, rapid response to abnormal trends, and dynamic correction of trend output. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Fig. 1 This is a schematic diagram of a notebook computer fault trend prediction system based on time series modeling proposed in this invention;

[0063] Fig. 2 The flowchart shows a notebook computer fault trend prediction system based on time series modeling proposed in this invention.

[0064] Fig. 3 This is a schematic diagram of the improved DRN model proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figs. 1-3 A notebook computer failure trend prediction system based on time series modeling includes:

[0067] The lifecycle labeling module is used to collect multi-source runtime data of the target laptop and construct a lifecycle feature dataset. The improved HDBSCAN algorithm is used to perform clustering operations on the lifecycle feature dataset, and lifecycle state labels are constructed based on the clustering results. The lifecycle state labels include initial state, stable state and aging state.

[0068] The trend prediction module is used to input the multi-source operational data and lifecycle status labels into the improved DRN model and output standard trend prediction results; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths;

[0069] The state control module is used to construct a state sliding window, record the life cycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; based on the state smoothing factor, it controls the switching delay triggering of the time-series prediction sub-path; if the state change frequency is greater than a first preset threshold, the time-series prediction sub-path corresponding to the life cycle state label is kept in an active state.

[0070] The sudden change identification module is used to monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel. If the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, the sudden trend prediction task is executed through the abnormal time series prediction sub-path without the control and judgment of the state smoothing factor, and the sudden trend prediction result is output.

[0071] The trend correction module is used to construct an anomaly memory cache, record multi-source operating data and sudden trend prediction results before and after the triggering of the mutation signal channel, generate anomaly memory samples based on similarity matching, input the anomaly memory samples into the improved DRN model to perform periodic training, generate anomaly warning bias vector, and correct the standard trend prediction results according to the anomaly warning bias vector, and output the corrected trend prediction results.

[0072] This implementation method uses a cycle labeling module to collect multi-source operational data from the target laptop and construct a lifecycle feature dataset. An improved HDBSCAN algorithm is then used to perform clustering on this dataset, generating lifecycle state labels and achieving a structured representation of the laptop's operational cycle. A trend prediction module inputs the multi-source operational data and lifecycle state labels into an improved DRN model to generate standard trend prediction results, enabling state-guided control of the trend modeling process. A state control module constructs a state sliding window based on the lifecycle state labels and performs state change frequency analysis to generate a state smoothing factor, controlling the time-series prediction path switching mechanism and improving the stability and consistency of the trend prediction process. Furthermore, a sudden change identification module monitors the rate of change of the multi-source operational data in real time, constructs a sudden change signal channel, and activates the abnormal time-series prediction sub-path to output sudden trend prediction results, enabling rapid identification and response to abnormal trends. Furthermore, by setting up a trend correction module, an anomaly memory cache is constructed to record relevant data of mutation signal channels. Anomaly memory samples are generated based on similarity matching and input into the improved DRN model for periodic training. This generates an anomaly warning bias vector and performs correction operations on the standard trend prediction results, achieving continuous correction and dynamic optimization of trend prediction results. This improves the system's adaptability and prediction accuracy during its lifecycle evolution. Through the synergistic effect of these modules, it is possible to model trend characteristics at each stage of the laptop's lifecycle, identify sudden states, and correct prediction biases, possessing advantages in lifecycle awareness, mutation perception, and predictive plasticity.

[0073] In this embodiment, the steps between modules include the following:

[0074] S1. Collect multi-source runtime data from the target laptop and construct a lifecycle feature dataset;

[0075] S2. Perform clustering operations on the lifecycle feature dataset using the improved HDBSCAN algorithm, and construct lifecycle state labels based on the clustering results; the lifecycle state labels include initial state, stable state, and aging state;

[0076] S3. Input the multi-source operation data and lifecycle status labels into the improved DRN model and output the standard trend prediction result; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths;

[0077] S4. Construct a state sliding window, record the lifecycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; control the switching delay triggering of the time-series prediction sub-path based on the state smoothing factor; if the state change frequency is greater than a first preset threshold, keep the time-series prediction sub-path corresponding to the lifecycle state label in an active state.

[0078] S5. Monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel; if the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, then without the control and judgment of the state smoothing factor, execute the sudden trend prediction task through the abnormal time series prediction sub-path and output the sudden trend prediction result.

[0079] S6. Construct an abnormal memory cache area to record multi-source operation data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generate abnormal memory samples based on similarity matching; input the abnormal memory samples into the improved DRN model to perform periodic training, generate an abnormal warning bias vector, and correct the standard trend prediction results according to the abnormal warning bias vector, and output the corrected trend prediction results.

[0080] In this embodiment, the multi-source operating data includes cumulative runtime, CPU temperature change, battery voltage fluctuation, fan speed stability index, and device load response index, specifically:

[0081] The cumulative runtime records the continuous running time interval of the laptop from power-on to shutdown, and records it in the time series table according to the set sampling frequency; the CPU temperature change value is obtained by constructing a CPU temperature sampling sequence, calculating the temperature change amplitude between adjacent sampling points, and recording the temperature rise and fall trend information; the battery voltage fluctuation amplitude data is obtained by obtaining the real-time battery voltage value sequence, calculating the difference between the maximum and minimum voltage values ​​per unit time, and recording the voltage fluctuation interval; the fan speed stability index is obtained by constructing a fan speed time series, calculating the ratio of the speed standard deviation to the mean based on a sliding time window, and recording the speed change stability information; the device load response index is obtained by obtaining the system load state sequence, recording CPU utilization, memory utilization, and IO response latency indicators, and constructing corresponding time series; the multi-source running data is aligned according to a unified sampling timestamp to construct a life cycle feature dataset.

[0082] In this embodiment, the improved HDBSCAN algorithm is specifically as follows:

[0083] The system receives partially labeled state information from the lifecycle feature dataset; constructs a sample adjacency constraint matrix based on the lifecycle stage labels, which represents the nearest neighbor constraint relationship between sample pairs with consistent lifecycle state labels; adjusts the distance graph structure between samples in the lifecycle feature dataset based on the sample adjacency constraint matrix, including scaling and expanding the distance weights between adjacent samples to form a sample connection structure that conforms to the lifecycle stage features; and guides the construction of a minimum spanning tree path based on the adjusted distance graph structure, prioritizing the retention of sample connection relationships with similar lifecycles in the tree path to achieve aggregated expression of stage features.

[0084] Based on the local sample density changes in the life cycle feature dataset, the density distribution difference of samples within the neighborhood is calculated; the minimum number of points for sample clustering is dynamically set according to the density distribution difference, and the clustering structure parameters are adjusted; in areas with large density fluctuations, the clustering threshold is corrected according to the density change rate to form an adaptive density partitioning mechanism; the clustering boundary between dense and sparse regions is controlled through this dynamic threshold mechanism to achieve density balance in life cycle state partitioning.

[0085] Based on the timestamp information in the lifecycle feature dataset, the time distance between sample pairs is calculated; a time weighting factor is set and a time weighting operation is performed on the original distance between samples to generate a time-weighted distance matrix; the time weighting factor is distributed in a decreasing relationship according to the size of the time interval between samples, which is used to strengthen the aggregation weight of time-adjacent samples and form a sample clustering structure with time continuity constraints.

[0086] Based on the density jump amplitude and life cycle state continuity between adjacent clusters in the preliminary clustering results, the boundary density information of adjacent clusters is extracted; a density boundary fitting model is constructed, and the density boundary fitting residual value is obtained by jointly fitting the density difference function and the continuity smoothing function; the structural relationship between adjacent clusters is judged based on the residual threshold. If the density boundary fitting residual value is less than the residual threshold, a cluster merging operation is performed to generate a corrected life cycle state cluster set.

[0087] Output the corrected lifecycle state cluster set, and calculate the sample state confidence score based on the weighted distance between the cluster center and the sample point; record the cluster residual metric based on the boundary change of the cluster structure, and generate lifecycle state label results, sample state confidence scores and cluster residual metric.

[0088] This implementation introduces a sample adjacency constraint matrix during the clustering process to guide the construction of the minimum spanning tree path, thereby achieving stage aggregation of the lifecycle structure. It enhances the flexibility of clustering parameters through a dynamic density threshold mechanism, thereby improving the ability to identify multi-stage lifecycle data. It strengthens the temporal consistency of the clustering process through a time weighting mechanism. It improves the accuracy and boundary rationality of lifecycle state identification by correcting the clustering results through density boundary residual judgment.

[0089] In this embodiment, the construction of lifecycle state labels based on clustering results specifically includes:

[0090] Based on the revised set of lifecycle state clusters, the mean of the sample running time distribution within each state cluster is calculated. The mean of the running time distribution represents the concentration of the running time of each sample within the cluster. The average level of the equipment load response indicators within each state cluster is calculated. The equipment load response indicators include temperature response time, current fluctuation amplitude, and load switching response rate. The average value of the battery voltage fluctuation amplitude within each state cluster is calculated. The battery voltage fluctuation amplitude represents the degree of change of the voltage curve within a set time window. The mean of the running time distribution, the average value of the load response, and the average value of the battery voltage fluctuation are combined to form a state cluster attribute matrix, which is used to represent the lifecycle physical characteristics of each state cluster.

[0091] The state cluster attribute matrix is ​​input into the state determination rule set, which includes the operating cycle stage division standard, battery performance stability threshold, and load response fluctuation boundary. The operating cycle stage division standard is based on a set time node to divide the initial stage, stable stage, and aging stage. The battery performance stability threshold is used to determine whether the voltage fluctuation amplitude is within the stable range. The load response fluctuation boundary is set with upper and lower limits based on the statistical fluctuation range. According to the above determination rules, multi-condition judgment is performed on each state cluster. If the stage time interval, battery stability, and load fluctuation conditions are met simultaneously, the corresponding state label candidate is output, generating a state label candidate set.

[0092] For each state label in the candidate state label set, the sample density center position in its corresponding state cluster is extracted. The density center position is fitted based on the principle of minimizing the distance between local density maximums and samples. The distribution skewness of the state label within the state cluster is calculated, and the distribution skewness represents the consistency between the label concentration and the density center. The state label confidence factor is calculated based on the density center position and the distribution skewness. The confidence factor is used to measure the credibility of the state label. The lifecycle state labels are weighted and corrected according to the state label confidence factor to improve the consistency between the state labels and the cluster structure.

[0093] The weighted and corrected lifecycle state labels are mapped to a standard state set, which includes initial state, stable state, and aging state. Each standard state has a one-to-one correspondence with a lifecycle stage. The corrected labels are normalized according to the mapping rules of the standard state set to generate a lifecycle state label result table. The lifecycle state label result table is output, and the label results are written into the state field of the lifecycle feature dataset to update the lifecycle state identifier of the sample.

[0094] This implementation method constructs a multi-dimensional attribute matrix by statistically analyzing the running time, load response, and battery voltage characteristics of state clusters, thereby improving the comprehensiveness of state label determination. It introduces a set of state determination rules to perform multi-condition judgments, realizing rule-driven partitioning of life cycle states. It introduces density centers and label distribution tilt to calculate the confidence factor of state labels, improving the fit between label distribution and clustering structure. It achieves standardized identification of life cycle state labels through standard state set mapping, enhancing label interpretability and consistency with subsequent analysis.

[0095] In this embodiment, the improved DRN model includes a state-aware controller, a path selector, a set of time-series predicted sub-paths, and abnormal time-series predicted sub-paths, specifically:

[0096] The state-aware controller receives lifecycle state labels, sample state confidence scores, and multi-source operational data features. The lifecycle state labels are derived from the state fields in the lifecycle feature dataset. The sample state confidence scores are generated based on the confidence factors of the state labels. The multi-source operational data features include temperature, voltage, current, and load response indicators. The state-aware controller internally constructs a multi-input sensing structure, integrates lifecycle state labels, state confidence scores, and multi-source features, and outputs a state control signal. The state control signal includes a path selection identifier, a confidence threshold value, and a state weight vector. The path selection identifier indicates the currently applicable predicted path type. The confidence threshold value is obtained by fitting the confidence score with the state change trend. The state weight vector is used to express the importance distribution of different state features.

[0097] The path selector receives a state control signal and a lifecycle state label sequence recorded in a state sliding window. The state sliding window records the state change trajectory of the sample within a set time period. The path selector performs path activation judgment based on the state change frequency. If the state change frequency is less than a first preset threshold, the temporal prediction sub-path corresponding to the state is kept active. If the state change frequency is greater than the first preset threshold, path activation delay processing is performed. The delay duration is obtained by fitting a state transition stability model. The path selector ensures that path switching oscillations are avoided when there are frequent state transitions.

[0098] The time-series prediction sub-path set includes an initial prediction path, a stable prediction path, and an aging prediction path. The initial prediction path constructs a short-time perceptual convolutional structure and receives short-period samples from the lifecycle feature dataset. Short-period samples are defined as having a runtime less than the first period threshold, performing a fast trend prediction task and focusing on modeling sudden increases and decreases. The stable prediction path constructs a smooth causal convolutional structure and receives a sequence of medium-period sample sequences from the lifecycle feature dataset. Medium-period samples satisfy the condition that the runtime is between the first and second period thresholds, performing a steady trend prediction task and emphasizing stable modeling of fluctuation patterns. The aging prediction path constructs a long-term residual memory structure and receives long-period running samples from the lifecycle feature dataset. Long-period samples have a runtime greater than the second period threshold, performing a decline trend prediction task and focusing on modeling performance decline trends and frequency changes. Each prediction path outputs an independent trend prediction result and records the prediction bias distribution.

[0099] The anomaly time series prediction sub-path is constructed independently of the time series prediction sub-path set. It receives multi-source operational data after the mutation signal channel is triggered. The mutation signal channel is based on the presence of a drastic deviation signal output by the mutation detector. The detector's judgment criterion is based on the sample index change exceeding a set threshold range. The anomaly time series prediction sub-path includes multi-scale residual convolutional layers and causal convolutional network layers. The multi-scale residual convolutional layers are used to extract mutation patterns at different time scales, and the causal convolutional network layers maintain temporal consistency and construct the mutation trend time series structure. The anomaly time series prediction sub-path outputs the sudden trend prediction results and the anomaly response factor.

[0100] The improved DRN model receives the output results of the time-series prediction sub-path set and the anomaly time-series prediction sub-path, constructs a fusion input tensor and performs a weighted fusion operation. The weights are obtained by jointly fitting the state weight vector in the state control signal and the anomaly response factor. The improved DRN model further performs a correction process on the fused standard trend prediction result based on the anomaly warning bias vector. The anomaly warning bias vector is constructed by estimating the anomaly occurrence probability and the offset direction. The correction process includes fine-tuning the segment corresponding to the trend offset identifier. The corrected trend prediction result is output as the lifecycle prediction output vector.

[0101] This implementation introduces a state-aware controller to achieve lifecycle state-driven predictive path selection, introduces a path selector to ensure the stability of path selection during state transitions, introduces multiple types of predictive paths to share the trend modeling tasks of different lifecycle stages, introduces abnormal predictive paths to improve the ability to identify and respond to sudden states, and finally improves prediction accuracy and robustness through weighted fusion and anomaly correction mechanisms.

[0102] In this embodiment, S4 specifically refers to:

[0103] Set the time length and sliding step size parameters of the state sliding window, construct the sliding window structure, and slide the sliding window on the life cycle state label sequence according to the set time length. Record the frequency of each state label in each sliding window and generate a state sliding count list.

[0104] The state sliding count list is normalized to construct a standardized distribution matrix of state labels in different time periods. The magnitude difference of state changes in continuous time periods is calculated to form a state change frequency sequence.

[0105] A state change frequency histogram is constructed based on the state change frequency sequence. Each column of the histogram corresponds to a state label, each row corresponds to a sliding window time period, and the matrix elements are the change frequency values ​​of the state label within that time period.

[0106] Based on the histogram of state change frequencies, the smoothness of the change trend of each state label is calculated. The smoothness of the change trend is fitted to generate a state smoothing factor through a preset inverse proportional mapping relationship. The state smoothing factor is used to adjust the state response sensitivity of the path selector.

[0107] The state smoothing factor is input into the path selector module as part of the state control signal. Combined with the path selection identifier and the confidence gating value, it controls the switching triggering behavior of the timing prediction sub-path and enhances the stability of path switching.

[0108] The path selector receives the state change frequency sequence and the state smoothing factor, and judges the state change frequency at each prediction time. If the state change frequency is greater than the first preset threshold, the temporal prediction sub-path corresponding to the current life cycle state label is kept in the active state.

[0109] If the frequency of state changes is less than the first preset threshold, the active path is switched with a delay. The path selector executes the delayed switching control logic to update the currently active time-series prediction sub-path, thus avoiding the prediction instability problem caused by frequent path switching.

[0110] In this embodiment, S5 specifically refers to:

[0111] A set of key operating indicators is defined, including temperature parameters, voltage parameters, current parameters, fan speed and processor load ratio; sampling period and time window parameters are defined, a set of continuous sampling sequences of key operating indicators is constructed, and a multi-source operating data time series matrix is ​​generated;

[0112] For each type of key operating indicator in the time series matrix, the difference between adjacent time points is calculated to construct a corresponding set of rate of change sequences. The rate of change sequences reflect the intensity of the trend change of key operating indicators in a continuous time period.

[0113] Based on the set of rate of change sequences, a sliding window structure is constructed. The mean and standard deviation are calculated for the data in each window to generate a dynamic statistical feature list. The dynamic statistical features are combined to construct a dynamic threshold interval for the rate of change. The dynamic threshold interval for the rate of change is obtained by fitting the coupling relationship between multiple sets of sliding statistics.

[0114] The rate of change of each type of key operating indicator is dynamically judged. If the rate of change of any key operating indicator exceeds the upper limit of the corresponding dynamic threshold and the magnitude of the rate of change is greater than the second preset threshold, the current time point is marked as a sudden change state, and the sudden change state label and trigger indicator identifier are recorded.

[0115] A mutation signal channel is constructed, and the mutation state signal is directly input into the mutation signal channel. The mutation signal channel serves as a bypass input path for the state-aware controller, directly controlling the activation behavior of the abnormal time series prediction sub-path without going through the regulation mechanism of the state smoothing factor.

[0116] After activating the abnormal time series prediction sub-path, the multi-source running data corresponding to the time point of the mutation state is used as the input dataset and input into the internal structure of the abnormal time series prediction sub-path. The abnormal time series prediction sub-path sequentially performs multi-scale residual convolution calculation and causal convolution network modeling operations to extract short-term abnormal features and causal propagation features in the mutation signal and output the sudden trend prediction results.

[0117] In this embodiment, the construction of the abnormal memory cache, which records multi-source operational data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generates abnormal memory samples based on similarity matching, specifically involves:

[0118] Set the time window range and sampling frequency parameters for the abnormal memory buffer. The time window range includes the first preset time period before the triggering of the mutation signal channel and the second preset time period after the triggering of the mutation signal channel. The sampling frequency parameter corresponds to the sampling period of key operating indicators and controls the time granularity of data buffering.

[0119] Record the multi-source running data sequence within the first preset time period as the pre-state sequence, and record the multi-source running data sequence within the second preset time period as the post-state sequence, and bind them to a unified timestamp set respectively;

[0120] Perform time alignment on the previous state sequence and the next state sequence to construct time series data pairs. Each time series data pair contains the previous and next state values ​​at each timestamp and the corresponding running indicator type identifier.

[0121] For each pair of timestamps, calculate the difference value of the state values ​​before and after, and construct a multi-source difference feature matrix. The multi-source difference feature matrix represents the magnitude and direction of change of key operating indicators before and after the mutation point.

[0122] The sudden trend prediction results and standard trend prediction results are called up, the prediction deviation is calculated by aligning the timestamps, and a trend offset vector is constructed. The trend offset vector is obtained by fitting the prediction residuals and characterizes the degree of change in the trend response under the sudden change state.

[0123] The trend offset vector is concatenated with the multi-source difference feature matrix to construct an abnormal feature vector set. Each vector in the abnormal feature vector set represents the context state and response features of a mutation event.

[0124] According to the preset similarity calculation rules, pairwise matching operations are performed on all vectors in the abnormal feature vector set. The similarity calculation rules include Euclidean distance, cosine similarity and dynamic time warping distance. The three similarity measurement methods are combined according to weighted rules, and the combined weights are obtained through empirical fitting or Bayesian optimization methods.

[0125] Perform a similarity matching operation to construct a similarity score matrix, which records the degree of similarity between any two abnormal feature vectors; select sample pairs with scores greater than a third preset threshold from the similarity score matrix and mark them as abnormal memory sample pairs;

[0126] Abnormal memory sample pairs that meet the conditions are classified into an abnormal memory sample set. The abnormal memory sample set serves as the memory retrieval basis for abnormal time-series prediction sub-paths and is used to assist in constructing historical state reference vectors and similar mutation pattern comparison templates under abnormal patterns.

[0127] In this embodiment, the step of inputting abnormal memory samples into the improved DRN model for periodic training to generate an anomaly warning bias vector, and then correcting the standard trend prediction result based on the anomaly warning bias vector to output the corrected trend prediction result, specifically involves:

[0128] Set the periodic training cycle parameter and the sample extraction batch parameter. The periodic training cycle parameter is used to control the model weight update interval, and the sample extraction batch parameter is used to set the number of abnormal memory samples extracted in each round of training.

[0129] An abnormal memory sample set is extracted from the abnormal memory cache according to a set batch. The abnormal memory sample set includes multi-source operational data of marked sudden behaviors and corresponding trend offset information.

[0130] Construct a training input sequence, which includes multi-source operational data, lifecycle state labels, sudden trend prediction results and trend offset vectors. The lifecycle state labels include operational phase identifiers and state numbers, and the trend offset vector is the difference vector between the sudden trend prediction results and the standard trend prediction results.

[0131] The training input sequence is fed into the improved DRN model, which includes a state-aware controller, a path selector, a set of time-predicted sub-paths, and anomaly time-predicted sub-paths.

[0132] The state-aware controller generates state control signals based on lifecycle state labels and sample state confidence scores. The state confidence scores are obtained by fitting the confidence interval of abnormal feature vectors in the training samples.

[0133] The path selector invokes the path activation strategy based on the state control signal, selects the corresponding sub-path from the set of time-series predicted sub-paths and the abnormal time-series predicted sub-paths to participate in the training process, and the path activation strategy is fitted by the matching rules between the sample state encoding and the historical error pattern.

[0134] The offset residual between the sudden trend prediction result and the standard trend prediction result in the training input sequence is used as a supervision signal to construct a training loss set, which includes a trend offset error term, a confidence weight error term, and a state residual term.

[0135] The trend offset error term measures the degree of continuity shift of the prediction results before and after the sudden event; the confidence weight error term measures the accuracy of the model's weight adjustment for abnormal samples; and the state residual term is used to compare the difference between the predicted state and the actual lifecycle label.

[0136] Based on the training loss set, gradient backpropagation is performed through the joint loss function to update the parameter weights of the state-aware controller, path selector, and each predicted sub-path in the improved DRN model. The weighting coefficients of the joint loss function are obtained by least squares fitting based on the sample fluctuation amplitude.

[0137] After the training cycle ends, the bias parameter set of the trend correction module in the improved DRN model is extracted and combined to construct an anomaly warning bias vector. The anomaly warning bias vector is input into the result fusion process, and the latest state control signal of the state-aware controller and the activation result of the path selector are called to determine the fusion weight vector used in trend correction.

[0138] The standard trend prediction result and the anomaly warning bias vector are weighted and corrected to output the corrected trend prediction result. The correction process includes weighted summation and anomaly response amplitude limitation. The anomaly response amplitude limitation operation sets the maximum offset range through an exponential decay rule.

[0139] Example 1:

[0140] To verify the feasibility of this invention in practice, it was applied to the whole-machine testing process of a large notebook computer manufacturer. The main objective was to optimize the trend-based fault warning and response mechanism during the notebook computer aging test phase.

[0141] The company's original fault detection method was mainly based on threshold rules and static feature screening, that is, judging whether a fault might occur based on whether values ​​such as CPU temperature, voltage, fan speed, and charging current exceed the limits. However, this method has a weak ability to detect trend changes and often fails to identify devices that have not yet exceeded the threshold but show abnormal trends, causing some hidden problems to only be exposed after the products leave the factory.

[0142] In this embodiment, the system connects to the company's aging test platform to collect multi-source status data during equipment operation, including: CPU and motherboard temperature, battery charging current, voltage curve, fan speed, power consumption change frequency, number of abnormal restarts, and hibernation response latency. Through dynamic generation of status labels, an improved HDBSCAN clustering module, and a DRN trend path prediction mechanism, it can identify sudden changes, accelerations, and abnormal nodes in the operating trend in real time, avoiding the "unidentifiable" blind spots in traditional systems.

[0143] To ensure the objectivity of the verification data, this embodiment selected three notebook product series: X150, P220, and M520, with 120 sample devices selected from each series. The comparative experiment lasted for three months.

[0144] Table 1 Comparison of Trend Fault Identification ; Table 1 shows that, under the same testing conditions, the system of this invention significantly increases the number of trend-related faults identified compared to traditional systems. Taking the X150 series as an example, traditional rules can only identify trend anomalies in 35 devices, while the system of this invention identifies 91, representing a growth rate of 160%. Similarly, the numbers for P220 and M520 increased to 102 and 97 respectively, both more than double the number identified by traditional systems. This indicates that the system of this invention has a stronger trend perception capability, especially when the fault has not yet fully manifested but has already shown a potential evolutionary tendency, it can accurately identify risky devices and avoid missed detections.

[0145] Table 2 Comparison of Trend Forecast Accuracy ; As can be seen from the comparison in Table 2, traditional systems mainly use sliding window averaging and static models for trend estimation, resulting in relatively large prediction errors, especially with slow response when faced with sudden voltage fluctuations or surges in fan load. In contrast, the system of this invention effectively reduces prediction bias by constructing a state-aware path structure and fusing prediction branches. Taking the CPU temperature of the X150 as an example, the error decreased from 6.9% to 3.1%, a reduction of over 55%. The prediction accuracy of the M520 fan speed also showed a significant improvement, indicating that the system of this invention has superior stability and tracking capabilities under dynamic changes.

[0146] In this embodiment, the average lead time for equipment failure warnings has increased by more than 18 hours, enabling engineers to promptly identify potentially faulty equipment during the aging phase and avoid increased maintenance costs after delivery. Simultaneously, the system's trend graphs can serve as a basis for maintenance recommendations, providing preliminary reference for after-sales service.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A notebook computer fault trend prediction system based on time series modeling, characterized in that, include: The lifecycle labeling module is used to collect multi-source runtime data of the target laptop and construct a lifecycle feature dataset. The improved HDBSCAN algorithm is used to perform clustering operations on the lifecycle feature dataset, and lifecycle state labels are constructed based on the clustering results. The lifecycle state labels include initial state, stable state and aging state. The trend prediction module is used to input the multi-source operational data and lifecycle status labels into the improved DRN model and output standard trend prediction results; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths; The state control module is used to construct a state sliding window, record the life cycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; based on the state smoothing factor, it controls the switching delay triggering of the time-series prediction sub-path; if the state change frequency is greater than a first preset threshold, the time-series prediction sub-path corresponding to the life cycle state label is kept in an active state. The sudden change identification module is used to monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel. If the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, the sudden trend prediction task is executed through the abnormal time series prediction sub-path without the control and judgment of the state smoothing factor, and the sudden trend prediction result is output. The trend correction module is used to construct an abnormal memory cache, record multi-source operating data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generate abnormal memory samples based on similarity matching. The abnormal memory samples are input into the improved DRN model for periodic training to generate an anomaly warning bias vector. The standard trend prediction result is then corrected based on the anomaly warning bias vector, and the corrected trend prediction result is output.

2. The notebook computer fault trend prediction system based on time series modeling according to claim 1, characterized in that, The steps between modules are as follows: S1. Collect multi-source runtime data from the target laptop and construct a lifecycle feature dataset; S2. Perform clustering operations on the lifecycle feature dataset using the improved HDBSCAN algorithm, and construct lifecycle state labels based on the clustering results; the lifecycle state labels include initial state, stable state, and aging state; S3. Input the multi-source operation data and lifecycle status labels into the improved DRN model and output the standard trend prediction result; the improved DRN model includes a state-aware controller, a path selector, a set of time-series prediction sub-paths and abnormal time-series prediction sub-paths; S4. Construct a state sliding window, record the lifecycle state label sequence, perform state change frequency analysis, and generate a state smoothing factor; control the switching delay triggering of the time-series prediction sub-path based on the state smoothing factor; if the state change frequency is greater than a first preset threshold, keep the time-series prediction sub-path corresponding to the lifecycle state label in an active state. S5. Monitor the rate of change of the multi-source operating data in real time and construct a sudden change signal channel; if the rate of change of any key operating indicator in the multi-source operating data is greater than the second preset threshold, then without the control and judgment of the state smoothing factor, execute the sudden trend prediction task through the abnormal time series prediction sub-path and output the sudden trend prediction result. S6. Construct an abnormal memory cache area to record multi-source operation data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generate abnormal memory samples based on similarity matching. The abnormal memory samples are input into the improved DRN model for periodic training to generate an anomaly warning bias vector. The standard trend prediction result is then corrected based on the anomaly warning bias vector, and the corrected trend prediction result is output.

3. The notebook computer fault trend prediction system based on time series modeling according to claim 2, characterized in that, The multi-source operating data includes cumulative runtime, CPU temperature change, battery voltage fluctuation, fan speed stability, and device load response. Specifically: The cumulative runtime records the continuous running time interval from power-on startup to shutdown of the laptop, and records it in the time series table according to the set sampling frequency; the CPU temperature change value is obtained by constructing a CPU temperature sampling sequence, calculating the temperature change amplitude between adjacent sampling points, and recording the temperature rise and fall trend information; the battery voltage fluctuation amplitude data is obtained by acquiring the real-time battery voltage value sequence, calculating the difference between the maximum and minimum voltage values ​​per unit time, and recording the voltage fluctuation interval; the fan speed stability index is obtained by constructing a fan speed time series, calculating the ratio of the speed standard deviation to the mean based on a sliding time window, and recording the speed change stability information; The device load response metrics are obtained by acquiring the system load status sequence, recording CPU utilization, memory usage and IO response latency metrics, and constructing corresponding time series; the multi-source operating data are aligned according to a unified sampling timestamp to construct a lifecycle feature dataset.

4. The notebook computer fault trend prediction system based on time series modeling according to claim 3, characterized in that, The improved HDBSCAN algorithm is as follows: It receives partial labeled state information from the lifecycle feature dataset, constructs a sample adjacency constraint matrix based on the lifecycle stage labels, adjusts the distance graph structure between samples based on the constraint matrix, and guides the construction of the minimum spanning tree path. Based on the local sample density changes in the life cycle feature dataset, the minimum number of points for sample clustering is dynamically set, the clustering structure parameters are adjusted according to local density differences, and the clustering threshold is adjusted in the density fluctuation region. Based on the timestamp information in the lifecycle feature dataset, the time distance between sample pairs is calculated. A time weighting factor is set according to a preset exponential function, and a time weighting operation is performed on the original sample distance to generate a time-weighted distance matrix. Based on the density jump amplitude and life cycle state continuity between adjacent clusters in the clustering results, the density boundary fitting residual value is calculated, and the adjacent clusters are merged based on the residual threshold to generate a corrected life cycle state cluster set; the life cycle state label results, sample state confidence scores and clustering residual metric values ​​are output.

5. A notebook computer fault trend prediction system based on time series modeling according to claim 4, characterized in that, The construction of lifecycle status labels based on clustering results specifically involves: Based on the revised lifecycle state cluster set, the mean of the running time distribution, the average level of the equipment load response index, and the average value of the battery voltage fluctuation amplitude of the samples in each cluster are statistically analyzed to generate a preliminary state cluster attribute matrix. The preliminary state cluster attribute matrix is ​​input into the state determination rule set. Based on the preset operating cycle stage division standard, battery performance stability threshold and load response fluctuation boundary, multi-condition judgment is performed to generate a state label candidate set. For each state label in the state label candidate set, a state label confidence factor is constructed by combining its corresponding cluster sample density center and label distribution tilt, and the life cycle state label is weighted and corrected based on the confidence factor. The weighted and corrected lifecycle state labels are mapped to the standard state set and marked as initial state, stable state and aging state, generating a lifecycle state label result table; the label result table is output and the label results are written into the state field of the lifecycle feature dataset.

6. A notebook computer fault trend prediction system based on time series modeling according to claim 5, characterized in that, The improved DRN model includes a state-aware controller, a path selector, a set of time-series predicted sub-paths, and abnormal time-series predicted sub-paths, specifically: The state-aware controller receives lifecycle state labels, sample state confidence scores, and multi-source operational data features, and generates a state control signal, which includes a path selection identifier, a confidence threshold value, and a state weight vector. The path selector performs path activation judgment based on the state control signal and the life cycle state label sequence recorded in the state sliding window. If the state change frequency is less than the first preset threshold, the time-series prediction sub-path corresponding to the state is kept active. If the state change frequency is greater than the first preset threshold, the time-series prediction sub-path is delayed in switching. The time series prediction sub-path set includes the initial prediction path, the stable prediction path, and the aging prediction path. The initial prediction path receives short-cycle running samples from the life cycle feature dataset and performs a fast trend prediction task; the stable prediction path receives medium-cycle sample sequences from the life cycle feature dataset and performs a steady trend prediction task. The aging prediction path receives long-cycle running samples from the life cycle feature dataset and performs a decline trend prediction task. The abnormal time series prediction sub-path is constructed independently of the time series prediction sub-path set. It receives multi-source running data after the sudden change signal channel is triggered and performs the sudden trend prediction task. The abnormal time series prediction sub-path includes multi-scale residual convolutional layers and causal convolutional network layers, and outputs the sudden trend prediction result. The improved DRN model receives the output results of the time series prediction sub-path set and the abnormal time series prediction sub-path, performs result fusion calculation, corrects the standard trend prediction results based on the abnormal warning bias vector, and generates the corrected trend prediction results.

7. A notebook computer fault trend prediction system based on time series modeling according to claim 6, characterized in that, Specifically, S4 is: Set the time length and sliding step size parameters of the state sliding window, receive the life cycle state label sequence, and record the number of times each state label appears in a continuous time period; The frequency of state changes is calculated based on the time sequence of lifecycle state labels, a sequence of state change frequencies is generated, and a histogram of state change frequencies is constructed. The state smoothing factor is calculated based on the histogram of state change frequencies, and an inverse mapping relationship is established between the state smoothing factor and the state change frequencies. The state smoothing factor is input into the path selector to control the switching delay triggering process of the timing prediction sub-path; If the frequency of state changes is greater than the first preset threshold, the time-series prediction sub-path corresponding to the lifecycle state label remains active; if the frequency of state changes is less than the first preset threshold, the path selector is triggered to perform a delayed switching operation to update the active time-series prediction sub-path.

8. A notebook computer fault trend prediction system based on time series modeling according to claim 7, characterized in that, Specifically, S5 is: Define a set of key operating indicators; the set of key operating indicators includes temperature parameters, voltage parameters, current parameters, and the ratio of fan speed to processor load. Collect continuous time series of the key operating indicators, perform difference calculation between adjacent time points, and generate a rate of change series of the key operating indicators; Perform sliding window mean and standard deviation analysis on the rate of change sequence to construct a dynamic threshold interval for the rate of change; determine whether the rate of change is greater than a second preset threshold; if the rate of change of any key operating indicator is greater than the second preset threshold, mark it as a mutation state; input the mutation state to the mutation signal channel, and activate the abnormal time series prediction sub-path without the control and judgment of the state smoothing factor; Multi-source operational data under abrupt changes is input into the abnormal time series prediction sub-path, and then passed through multi-scale residual convolutional layers and causal convolutional network layers in sequence to output the sudden trend prediction results.

9. A notebook computer fault trend prediction system based on time series modeling according to claim 8, characterized in that, The construction of the abnormal memory cache area records multi-source operational data and sudden trend prediction results before and after the triggering of the mutation signal channel, and generates abnormal memory samples based on similarity matching, specifically as follows: Set the time window range and sampling frequency parameters of the abnormal memory buffer, and record the multi-source running data sequence before the triggering of the sudden signal channel and the multi-source running data sequence after the triggering of the sudden signal channel; the time window range includes the first preset time period before the triggering of the sudden signal channel and the second preset time period after the triggering of the sudden signal channel. Align the multi-source operational data in the abnormal memory buffer according to the timestamp order to construct time series data pairs; calculate the difference values ​​of each key operational indicator in the multi-source operational data before and after the triggering of the sudden signal channel to generate a multi-source difference feature matrix; Based on the difference between the sudden trend prediction result and the standard trend prediction result, a trend offset vector is constructed; the trend offset vector is concatenated with the multi-source difference feature matrix to generate an abnormal feature vector set. Based on preset similarity calculation rules, similarity matching is performed on samples in the abnormal feature vector set; the similarity calculation rules include Euclidean distance, cosine similarity, and dynamic time warping distance; Calculate the similarity score between each sample and generate a similarity score matrix; Select sample pairs with similarity scores greater than a third preset threshold from the similarity score matrix to construct an abnormal memory sample set.

10. A notebook computer fault trend prediction system based on time series modeling according to claim 9, characterized in that, The process involves inputting abnormal memory samples into an improved DRN model for periodic training, generating an anomaly warning bias vector, and then correcting the standard trend prediction result based on this bias vector, outputting the corrected trend prediction result. Specifically: Set the periodic training period and sample extraction batch parameters, select an abnormal memory sample set from the abnormal memory cache, and construct the training input sequence; the training input sequence includes multi-source running data, life cycle state labels, sudden trend prediction results, and trend offset vectors; The training input sequence is fed into the improved DRN model, and the periodic training process is performed sequentially through the state-aware controller, path selector, set of temporal prediction sub-paths and abnormal temporal prediction sub-paths. The state-aware controller generates a state control signal based on the lifecycle state label and sample state confidence score. The path selector determines the activated temporal prediction sub-path or abnormal temporal prediction sub-path based on the state control signal. A training loss set is constructed based on the difference between the sudden trend prediction result and the standard trend prediction result; the training loss set includes a trend offset error term, a confidence weight error term, and a state residual term; the parameter weights of the improved DRN model are updated based on the training loss set; After the training period ends, the set of bias parameters used for trend correction in the improved DRN model is extracted to construct the anomaly warning bias vector; The anomaly warning bias vector is input into the result fusion process. Based on the state control signal output by the state-aware controller and the activation result of the path selector, the standard trend prediction result is weighted and corrected, and the corrected trend prediction result is output.

Images