Time series data processing method and system

Abstract

The invention relates to the technical field of industrial data analysis, provides a time series data processing method and system, and solves the problem of insufficient integrated coherent analysis and co-processing capability for engineering equipment abnormity. The method comprises the steps of obtaining a multi-source time sequence collected by an Internet of Things platform, performing protocol conversion and time alignment through an aggregation gateway to obtain a time sequence data stream, and then identifying an abnormal data segment through a stream calculation engine; performing state deduction on the abnormal data segment by using a state space model of a physical information neural network to generate an ideal state trajectory; according to the deviation between the ideal trajectory and the actual data, a sensitive feature set is obtained by combining an attention mechanism and causal inference traceability; and finally, according to a tenant mode, packaging the abnormal data segment, the ideal state track and the sensitive feature set into a tenant data packet for storage. According to the method, the detection accuracy and the source tracing capability of the abnormal state of the industrial equipment and the safety management efficiency of the analysis result can be improved.

Description

Technical Field

[0001] This application relates to the technical field of industrial data analysis, and in particular to a method and system for processing time-series data. Background Technology

[0002] In industrial sectors such as large-scale engineering infrastructure and equipment manufacturing, real-time status monitoring and intelligent early warning of key equipment such as tunnel boring machines and tower cranes are core requirements for ensuring construction safety and improving operation and maintenance efficiency. The foundation for this lies in the efficient and accurate analysis of the massive amounts of multi-source time-series data generated by the equipment.

[0003] Currently, in common industrial time-series data processing workflows, each step is usually relatively independent. For example, after data collection, it is mostly stored and processed in batches. Anomaly detection relies on comparison with historical statistical thresholds or triggering of fixed rules. The analysis of the causes of anomalies often requires cross-system log queries or manual correlation and inference based on expert experience.

[0004] However, when faced with complex working conditions of engineering machinery groups, existing technical solutions struggle to achieve end-to-end coherent analysis, from data aggregation and precise anomaly location to root cause tracing and flexible application of analysis results, thus limiting overall diagnostic efficiency and cross-departmental collaboration capabilities. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for processing time-series data, so as to solve the problem that the existing technology lacks the ability to perform integrated and coherent analysis and collaborative handling of engineering equipment anomalies.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for processing time-series data, comprising: Acquire multi-source time series data of the target monitoring object collected through the IoT data development platform; The multi-source time series is converted and time-aligned by the aggregation gateway to obtain a time series data stream. The time series data stream is then subjected to pattern recognition within an online sliding window by a stream computing engine to obtain abnormal data segments. The state-space model built on a physical information neural network is invoked to perform state deduction on the abnormal data segment and its corresponding context time series, generating an ideal state trajectory. Based on the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, the abnormal data segment is subjected to feature tracing based on attention mechanism and causal inference to obtain a sensitive feature set. Based on the tenant mode of the IoT data development platform, the abnormal data segment, the ideal state trajectory and the sensitive feature set are encapsulated to obtain a tenant data packet, and the tenant data packet is stored in the data storage area of ​​the corresponding tenant.

[0007] Optionally, the step of invoking a state-space model constructed based on a physical information neural network to perform state deduction on the abnormal data segment and its corresponding context time series, and generating an ideal state trajectory, includes: The abnormal data segments and their corresponding context time series are combined into a model input sequence; An initial state sequence is obtained by performing forward propagation calculation on the input sequence of the model using a pre-built state space model based on a physical information neural network. The state space model integrates the equipment operation mechanism and data-driven capability of the target monitoring object. During the forward propagation calculation, the residuals of the physical constraint terms in the fusion process are calculated, and the state adjustment amount corresponding to the residuals is generated. The state adjustment amount is superimposed on the initial state sequence to obtain the intermediate state trajectory. Extract the normal time sequence segment before the occurrence of the abnormal data segment from the context time sequence, and combine the state difference signal between the intermediate state trajectory and the normal time sequence segment to calibrate the parameters characterizing the individual characteristics of the device in the state space model to obtain the calibrated state space model. Using the calibrated state-space model, the final state deduction is performed on the input sequence of the model, and the ideal state trajectory that conforms to the individual characteristics of the device is output.

[0008] Optionally, the step of calibrating the parameters characterizing individual device characteristics in the state-space model by combining the state difference signal between the intermediate state trajectory and the normal time segment includes: Perform multi-resolution wavelet transform on the state difference signal to obtain a first difference component representing the long-term trend state and a second difference component representing the short-term fluctuation state. The first difference component and the second difference component are encoded and mapped respectively to obtain the first encoding vector and the second encoding vector; The first encoding vector and the second encoding vector are fused to generate a comprehensive difference feature that describes the degree of deviation of the current state of an individual device. Based on the comprehensive difference characteristics, a set of update instructions for the predefined calibrable parameters of the state space model are generated through a lightweight parameter adjustment network. The parameters to be updated are then calibrated in a targeted manner according to the parameter identifier, update direction, and update magnitude of the parameters to be updated in the update instructions.

[0009] Optionally, based on the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, the abnormal data segment is subjected to feature tracing based on attention mechanism and causal inference to obtain a sensitive feature set, including: Using the features in the multi-source time series corresponding to the abnormal data segment as nodes, an initial causal graph is constructed, wherein the edges between the nodes represent the potential causal influence directions; The multidimensional bias is injected into the initial causal graph to obtain the initial energy state of each node; The initial causal graph is used as the current generation causal graph, and the initial energy state is used as the current generation state; The attention weights are calculated based on the current generation state, and the nodes in the current generation causal graph are dynamically masked according to the attention weights to obtain the key inference subgraph. On the key inference subgraph, the direction and strength of the edges are verified and corrected by a preset causal inference rule to obtain correction instructions. The initial energy state of the relevant nodes and the structure of the initial causal graph are updated according to the correction instructions to obtain the updated energy state and the updated causal graph. The updated energy state and the updated causal graph are used as the current generation causal graph and the current generation state, respectively. The step of calculating attention weights to update the energy state is repeated until the updated energy state and the updated causal graph both reach the corresponding preset stability conditions, and a stable energy state and a stable causal graph are obtained. In the stable causal graph, the final source node of all energy inflows under stable energy states is identified and traced, and all the final source nodes are integrated to obtain a sensitive feature set.

[0010] Optionally, the step of using a stream computing engine to perform pattern recognition within an online sliding window on the time-series data stream to obtain abnormal data segments includes: In the stream computing engine, a sliding time window with a configurable length is defined for the time-series data stream, and the time-series data stream is segmented online through the sliding time window to generate multiple data units; An online knowledge distillation framework is built within a stream computing engine, the online knowledge distillation framework comprising a teacher branch and a student branch; The teacher branch and the student branch synchronously perform forward computation on the data unit to obtain the first representation vector and the second representation vector. Calculate the distillation loss value between the first representation vector and the second representation vector, and based on the distillation loss value, calculate the anomaly score corresponding to each data unit to obtain the anomaly score sequence; Concept drift detection is performed on the abnormal score sequence to obtain candidate abnormal time periods. The original data corresponding to all the candidate abnormal time periods are extracted from the time-series data stream to obtain abnormal data segments.

[0011] Optionally, based on the tenant mode of the IoT data development platform, the abnormal data segment, the ideal state trajectory, and the sensitive feature set are encapsulated to obtain a tenant data packet, including: Based on the tenant's unique identifier defined by the tenant pattern, extract the data encapsulation components that the corresponding tenant has subscribed to from the IoT data development platform; The data encapsulation component performs automated encapsulation processing on the abnormal data segment, the ideal state trajectory, and the sensitive feature set to obtain a structured data packet. The service access interface corresponding to the structured data packet is determined according to the preset interface specification, and a data list is generated according to the preset data management requirements; The structured data packet, the service access interface, and the data list are integrated to obtain the tenant data packet.

[0012] Optionally, the step of performing protocol conversion and time alignment on the multi-source time series data through the aggregation gateway to obtain a time series data stream includes: Create a corresponding protocol parsing instance for each communication protocol within the aggregation gateway; The multi-source time series is parsed in real time through the protocol parsing instance to obtain the original data points. Based on the time synchronization service accessed by the aggregation gateway, a time correction value is calculated for each of the original data points. The time correction value is applied to the corresponding original data point to obtain a calibration data point with a standard timestamp; All the calibration data points are written into a time-aligned buffer in chronological order according to the standard timestamps and then sorted to obtain a time-series data stream.

[0013] Secondly, this application provides a time-series data processing system, including: The acquisition module is used to acquire multi-source time series data of the target monitoring object collected through the IoT data development platform. The identification module is used to perform protocol conversion and time alignment on the multi-source time series through the aggregation gateway to obtain a time series data stream, and to use a stream computing engine to perform pattern recognition within an online sliding window on the time series data stream to obtain abnormal data segments. The deduction module is used to call the state space model built based on the physical information neural network to perform state deduction on the abnormal data segment and the corresponding context time sequence, and generate an ideal state trajectory. The inference module is used to perform feature tracing based on attention mechanism and causal inference on the abnormal data segment according to the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, and obtain a sensitive feature set. The encapsulation module is used to encapsulate the abnormal data segment, the ideal state trajectory, and the sensitive feature set according to the tenant mode of the IoT data development platform to obtain a tenant data packet, and store the tenant data packet in the data storage area of ​​the corresponding tenant.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, configured to implement the steps of a time-series data processing method as described in the first aspect above when executing the computer program.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of a time-series data processing method as described in the first aspect above.

[0016] The time-series data processing method provided in this application has the following beneficial effects: First, heterogeneous data is unified through an aggregation gateway to lay the foundation for real-time analysis. Then, a stream computing engine is used to realize online anomaly detection and capture device malfunctions in a timely manner. Next, a model that integrates physical mechanisms is called to perform state inference to generate a reliable and interpretable theoretical trajectory. Subsequently, feature tracing is performed based on trajectory deviation and combined with attention and causal inference to accurately locate the key feature set that caused the anomaly. Finally, the analysis results are standardized, encapsulated, and stored in isolation according to the tenant model so that different business units can call them securely and flexibly, thereby improving the efficiency of data collaborative management.

[0017] Furthermore, this application quantifies the residuals of physical constraint terms into state adjustment quantities and superimposes them with the initially calculated state trajectory, enabling the extrapolation process to actively conform to physical laws. In turn, by using the difference between the normal data segments before the anomaly occurred and the current calculated trajectory to calibrate the model, the general model can be quickly adapted to the individual operating characteristics of specific equipment, ultimately outputting a more accurate and personalized ideal state trajectory, thereby providing a reliable comparison benchmark for subsequent accurate tracing. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a time-series data processing method provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating a specific implementation of a time-series data processing method provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a time-series data processing system provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In the intelligent operation and maintenance scenario of large-scale engineering machinery, it is crucial to achieve end-to-end analysis from massive data perception to precise decision support. However, the current common technical path usually separates the processes of data aggregation, anomaly alarm, cause investigation and result application, and relies on manual connection between different tools. This results in a disjointed analysis chain, slow response, and difficulty in forming standardized analysis results that can be directly called by different business units, thus restricting the efficiency of in-depth equipment status management and collaborative handling.

[0021] To address this, this application proposes a time-series data processing method. Its core concept lies in constructing a coherent technical process from unified data access, real-time anomaly detection, and deep state deduction based on physical mechanisms, to key feature tracing, and finally, packaged delivery according to business needs. This solution can transform equipment monitoring data into a structured and interpretable analysis result package containing anomaly fragments, theoretical trajectories, and root cause characteristics in real time, and supports secure on-demand distribution. This not only achieves integrated and accurate diagnosis and tracing of complex anomalies, but also effectively solves the pain point of difficult cross-departmental collaborative application of analysis results through standardized packaging, thereby overcoming the limitations of existing technical processes being disconnected and having low processing efficiency.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide a method for processing time-series data, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Obtain the multi-source time series of the target monitoring object collected through the Internet of Things data development platform.

[0024] Among them, the Internet of Things (IoT) data development platform refers to a software system that integrates device connection, data acquisition, protocol parsing and preliminary processing. As a unified data entry point, it is responsible for converting device signals in the physical world into digital information that can be recognized by the computing system. The target monitoring object refers to the engineering machinery equipment whose operating status needs to be continuously monitored. Multi-source time series refers to the sequence of measurement data collected synchronously from multiple different types of sensors deployed on the same monitoring device and arranged in chronological order.

[0025] In step S101, the various sensors pre-installed on the target monitoring object are first connected to the Internet of Things data development platform via a network. Then, the platform establishes a stable data link with these sensors through built-in multiple communication protocol adapters.

[0026] Subsequently, the platform actively or passively receives raw data messages reported in real time from each sensor according to the preset acquisition frequency. Each data message usually contains the sensor number, the measurement value, and the timestamp of its own recording.

[0027] Next, the platform performs preliminary parsing of the received raw messages to extract valid measurement point identifiers, data values, and time information. It also logically associates and bundles data from different sensors of the same device to form a set of parallel data sequences that are roughly aligned in time. This completes the acquisition and preliminary organization of "multi-source time series" data, preparing for subsequent centralized processing.

[0028] S102. The multi-source time series is converted and time-aligned through the aggregation gateway to obtain a time series data stream. The time series data stream is then subjected to pattern recognition within an online sliding window using a stream computing engine to obtain abnormal data segments.

[0029] In one specific implementation, step S102 includes: Step 1021: Create a corresponding protocol parsing instance for each communication protocol within the aggregation gateway.

[0030] Among them, the protocol parsing instance refers to a parsing program generated in the aggregation gateway software to deal with the unique data format rules, i.e., communication protocols, such as Modbus and MQTT, used by different brands or models of sensors.

[0031] In step 1021, when the aggregation gateway receives multi-source time-series data streams from the IoT platform, it automatically identifies the communication protocol types followed by each of these multi-source time-series data streams. Subsequently, the core processing unit of the aggregation gateway dynamically allocates computing resources and loads the corresponding protocol parsing template for each identified protocol type, thereby initializing an independent protocol parsing instance focused on interpreting that specific "data language".

[0032] Step 1022: Perform real-time parsing of the multi-source time series using the protocol parsing instance to obtain the original data points, and calculate the time correction value for each of the original data points based on the time synchronization service accessed by the aggregation gateway.

[0033] The time synchronization service refers to a high-precision and unified standard clock signal source provided by a time server in the network. By accessing this time synchronization service, the aggregation gateway obtains an authoritative time reference to correct time inconsistencies caused by individual sensors' own clock errors or network transmission delays.

[0034] In step 1022, each protocol parsing instance can work in parallel, that is, according to the built-in rules, the messy raw binary or text data packets are decoded in real time into a structured record containing three key elements: "sensor ID", "physical quantity measurement value" and "time of local sensor recording", that is, the raw data point. Next, the aggregation gateway calls its synchronized time service and obtains the current accurate standard time. Then, for each raw data point, it compares its built-in sensor local timestamp with the standard time and calculates the difference between the two using a time synchronization algorithm, such as estimating network latency and clock offset. This difference is the time correction value required to calibrate the data point to a unified time axis.

[0035] Step 1023: Apply the time correction value to the corresponding original data point to obtain a calibration data point with a standard timestamp.

[0036] In step 1023, the local timestamp of each raw data point is added to or subtracted from its corresponding time correction value to generate calibration data points.

[0037] Step 1024: Sort all the calibration data points according to the order of the standard timestamps into a time-aligned buffer to obtain a time-series data stream.

[0038] The time alignment buffer is a temporary storage area in memory specifically used to receive and temporarily store calibration data points. Its core function is to act as a sorting queue to ensure that all points are output in strict chronological order.

[0039] In step 1024, all generated calibration data points are successively sent into the time alignment buffer. The buffer then acts like a highly efficient "time sorting machine" to automatically insert them into the correct time position in the queue according to the standard timestamp carried by each data point. Subsequently, the buffer reads out and sends out these data points in the order from morning to night to output a time-series data stream.

[0040] Step 1025: In the stream computing engine, define a configurable sliding time window for the time-series data stream, and perform online segmentation of the time-series data stream through the sliding time window to generate multiple data units.

[0041] Sliding time window is a core processing model in stream computing. It defines a fixed-length time interval, such as the past 5 seconds. This window is not static, but slides forward on the time axis as new data arrives, covering the latest continuous data each time.

[0042] In step 1025, after receiving the time-series data stream through the stream computing engine, a sliding time window is logically set according to the preset analysis granularity, such as analyzing the data of the most recent 10 seconds each time. This window is closely attached to the beginning of the data stream and continuously "frames" the latest incoming data within the window duration. Subsequently, whenever a new data point flows in and causes the window to slide forward, the stream computing engine extracts all the data in the current window as an independent analysis batch, which is called a data unit.

[0043] It should be noted that the specific process of the stream computing engine can be referred to relevant technologies, and the specific implementation process of the stream computing engine will not be described in detail in the embodiments of this application.

[0044] Step 1026: Build an online knowledge distillation framework within the stream computing engine. The online knowledge distillation framework includes a teacher branch and a student branch.

[0045] The online knowledge distillation framework refers to a lightweight machine learning model structure suitable for data streams. Its teacher branch is a relatively complex, parameter-fixed, and highly representative pre-trained model that is responsible for providing relatively stable "experience judgments." The student branch is a relatively simple model with parameters that can be updated online. Its task is to learn and adapt to the latest characteristics of the data stream in real time.

[0046] It should be noted that this embodiment does not limit the structure of the online knowledge distillation framework, and can be set accordingly based on the actual situation.

[0047] In step 1026, to achieve adaptive anomaly detection, the stream computing engine loads two related neural network models during initialization to form an online knowledge distillation framework. Specifically, the teacher branch is usually a model pre-trained on a large amount of historical normal data, and its parameters are locked. The student branch is initialized as a model with a similar structure to the teacher but a simpler one, and its parameters are allowed to be updated in subsequent processes. Both are deployed in parallel to process each input data unit.

[0048] Step 1027: Perform forward computation on the data unit synchronously through the teacher branch and the student branch to obtain the first representation vector and the second representation vector.

[0049] In step 1027, each newly generated data unit is simultaneously input into the teacher branch and the student branch. Then, the two branches independently perform forward propagation calculations and convert the data unit into a high-dimensional mathematical vector that condenses the intrinsic characteristics of the data. The output of the teacher branch is called the first representation vector, which contains its understanding based on historical experience; the output of the student branch is called the second representation vector, which reflects its current interpretation of the data.

[0050] Step 1028: Calculate the distillation loss value between the first representation vector and the second representation vector. Based on the distillation loss value, calculate the anomaly score corresponding to each data unit to obtain the anomaly score sequence.

[0051] The distillation loss value refers to the difference between the output of the student branch and the output of the teacher branch. In this application, this difference is innovatively used as an indicator of the "abnormality" of the data: the greater the difference, the more the current data deviates from the normal pattern familiar to the teacher model, and the more difficult it is for the student model to learn quickly.

[0052] In step 1028, the distance between the first representation vector and the second representation vector, such as the cosine distance, is calculated. This scalar value is the distillation loss value. Then, this distillation loss value is transformed into an anomaly score within a specific range through a mapping function, such as the Sigmoid function. As the data unit is continuously processed, a time-varying anomaly score sequence is output in real time.

[0053] Step 1029: Perform concept drift detection on the abnormal score sequence to obtain candidate abnormal time periods, and extract the original data corresponding to all the candidate abnormal time periods from the time series data stream to obtain abnormal data segments.

[0054] Concept drift detection refers to a monitoring algorithm for changes in the statistical characteristics of streaming data. It does not focus on the instantaneous highs and lows of individual score points, but analyzes the overall distribution of abnormal score sequences, such as whether the mean and variance have undergone continuous and significant shifts, corresponding to substantial changes in the operating status of the equipment.

[0055] In step 1029, the real-time generated abnormal score sequence is continuously monitored, and a concept drift detection algorithm, such as a sliding window-based statistical test, is applied. This algorithm determines whether the abnormal score sequence has continuously risen from a low-value stable range over a period of time and entered and remained in a high-value range. The start and end times of such a continuous and significant change in the score distribution identified by the algorithm are marked as candidate abnormal periods. Then, based on the start and end timestamps of these candidate abnormal periods, we trace back to the time-aligned time-series data stream and accurately locate and extract all the original sensor data within the corresponding time period. These extracted data segments are the final output, clearly defined abnormal data segments.

[0056] This application achieves standardization and high-precision time synchronization of multi-source heterogeneous data through an aggregation gateway; and further utilizes online knowledge distillation and concept drift detection under the stream computing framework to achieve real-time and adaptive perception of abnormal equipment operating status, which can sensitively capture gradual or new abnormal patterns that are easily missed by traditional threshold methods, thereby further improving the intelligence level and early warning timeliness of status monitoring.

[0057] S103. Call the state space model built on the physical information neural network to perform state deduction on the abnormal data segment and the corresponding context time sequence, and generate an ideal state trajectory.

[0058] In one specific implementation, such as Figure 2 As shown, step S103 includes: Step 1031: Combine the abnormal data segment and the corresponding context time sequence into a model input sequence.

[0059] In step 1031, firstly, based on the precise start and end timestamps recorded in the identified abnormal data segments, the continuous data before and after the occurrence of the abnormal period is located and extracted from the time-aligned complete time-series data warehouse. These two parts of data together constitute the context time series. The abnormal data segment located at the center of the time axis is then spliced ​​together with the extracted context time sequence according to the natural chronological order of the events, in order to integrate it into a longer, continuous single data sequence that contains the complete context before and after the abnormal event. This newly constructed data sequence is the model input sequence.

[0060] For example, an abnormal fluctuation in the current of the main drive motor of a tunnel boring machine is identified between 2:00 PM and 2:10 PM to form an abnormal data segment. Then, data from 30 minutes before and after this period, namely data from 1:30 PM to 2:00 PM and data from 2:10 PM to 2:40 PM, are extracted as context time series. The abnormal segment is then concatenated with the context to obtain a model input sequence that lasts 70 minutes, from 1:30 PM to 2:40 PM.

[0061] Step 1032: Perform forward propagation calculation on the input sequence of the model using a pre-constructed state space model based on physical information neural network to obtain an initial state sequence. The state space model integrates the equipment operation mechanism and data-driven capability of the target monitoring object.

[0062] Among them, physical information neural networks refer to a type of machine learning model that embeds domain knowledge, especially physical laws expressed in the form of differential equations or systems of algebraic equations, as structured constraints into the neural network architecture or training objective function. A state-space model is a mathematical model framework used to characterize how the internal latent variables of a dynamic system evolve with time and external inputs. The process of pre-constructing a state-space model based on a physical information neural network can be found in relevant construction techniques, which will not be elaborated here.

[0063] In step 1032, a pre-trained state-space model specifically designed for engineering machinery is invoked. Its "fusion" capability stems from a unique training method: during the model building phase, physical conservation laws describing the core operating mechanism of the equipment, such as the torque balance equation of the cutterhead drive system and the force balance equation of the hydraulic propulsion system, are encoded as regularization constraints that must be satisfied during training. Simultaneously, the main body of the model adopts a deep neural network structure suitable for processing time-series data, such as a long short-term memory network, to learn complex dynamic patterns from observation data. Then, the model uses historical normal operation data and is trained through a joint loss function, which simultaneously optimizes the network's fitting accuracy to historical data sequences and the degree to which its internal state deduction results comply with the aforementioned physical constraint equations.

[0064] Then, after training is completed, the input sequence of the model is fed into the trained state space model. The state space model then performs forward propagation calculation and, based on its learned fusion knowledge, calculates a state estimation sequence of the same length as the input sequence, reflecting the internal health and performance of the device, time by time. This sequence is the initial state sequence. For example, at 2:05 pm, the model initially calculates the overall efficiency of the device to be 85%.

[0065] Step 1033: During the forward propagation calculation, calculate the residual of the physical constraint term in the fusion process, generate the state adjustment amount corresponding to the residual, and superimpose the state adjustment amount with the initial state sequence to obtain the intermediate state trajectory.

[0066] The residual refers to the unbalance quantity of the equation calculated when the model performs forward calculations, after substituting the currently derived state estimate into the preset physical law equations. The residual is a scalar or vector used to quantify the degree to which the current state estimate deviates from the physical laws. The state adjustment variable is a vector determined based on the calculated residual and through a predefined mapping relationship. The function of this vector is to correct the current state estimate in terms of direction and magnitude, so as to reduce its deviation from the physical laws.

[0067] In step 1033, during the same process of the model calculating and generating the initial state sequence based on the model input sequence, an online physical consistency verification and correction mechanism is simultaneously initiated. Specifically, when deducing the state at each moment, the estimated initial state value obtained at that moment is substituted into the preset physical constraint equations describing the operating mechanism of the equipment for verification. For example, for energy conservation constraints, the verification process is to calculate the difference between the predicted input energy, dissipated energy and output energy, and this difference is the physical constraint residual at that moment. Subsequently, the physical constraint residual is converted into a corresponding state adjustment vector through a preset, differentiable adjustment amount generation function, such as a simple linear scaling function adjustment=α×residual. Here, adjustment represents the state adjustment amount, α is a preset sensitivity coefficient that is greater than zero, which is set to 0.1 here, and residual represents the physical constraint residual. Finally, this state adjustment amount is added to the initial state estimate at that moment to make corrections, thereby obtaining the intermediate state trajectory.

[0068] For example, at 2:05 PM, based on the initial state sequence, the model estimates the input power to be 1000 kW and the fixed loss to be 150 kW. Then, according to the law of conservation of energy, the output power should be 850 kW, but the model currently calculates the output power to be 840 kW, resulting in a residual of 10 kW. Subsequently, based on the preset linear adjustment function, a power adjustment of 10 × 0.1 = 1 kW is obtained, which leads to the correction of the efficiency state value, that is, the efficiency estimate at this moment is corrected from 84% to 84.1%, thus obtaining the intermediate state value at this moment. Similarly, after performing this operation on all time points, an intermediate state trajectory is formed.

[0069] Step 1034: Extract the normal time sequence segment before the occurrence of the abnormal data segment from the context time sequence, and combine the state difference signal between the intermediate state trajectory and the normal time sequence segment to calibrate the parameters representing the individual characteristics of the device in the state space model, so as to obtain the calibrated state space model.

[0070] Step 1034 may specifically include the following steps: Step a1: Perform multi-resolution wavelet transform on the state difference signal to obtain the first difference component representing the long-term trend state and the second difference component representing the short-term fluctuation state.

[0071] Among them, multi-resolution wavelet transform refers to a mathematical transformation method that uses a set of scaled and translated wavelet basis functions to decompose the original signal into different time scales; Long-term trend states refer to the state deviation trends that gradually accumulate and manifest on a timescale of hours or even days, caused by the slow degradation of equipment, such as wear of main bearings and aging of seals. Short-term fluctuations refer to state deviations that fluctuate rapidly on a timescale of seconds to minutes, caused by instantaneous changes in external load, hydraulic shocks, or measurement noise.

[0072] In step a1, a multi-resolution wavelet transform is performed on the obtained state difference signal using a preset type of wavelet basis function, such as the Daubechies wavelet. This transform process decomposes the original one-dimensional time series signal into a series of sub-band spaces of different scales. Then, by reconstructing the signal components mainly contributed by low-frequency, wide-scale sub-bands, the first difference component is obtained. This component signal is smooth and changes slowly, mainly capturing the long-term and gradual change trend of the individual characteristic deviation of the equipment. At the same time, by reconstructing the signal components mainly contributed by high-frequency, narrow-scale sub-bands, the second difference component is obtained. This component signal oscillates frequently and changes rapidly, mainly capturing the short-term and transient fluctuations and noise in the deviation.

[0073] Step a2: Encode and map the first difference component and the second difference component respectively to obtain the first encoding vector and the second encoding vector.

[0074] Encoding mapping refers to the process of compressing a variable-length time-series signal into a fixed-length, high-dimensional, dense numerical vector through a neural network. This vector aims to preserve the most representative feature patterns in the original signal in a compact form.

[0075] In step a2, two structurally independent and pre-trained encoder neural networks are used. The first encoder receives the first difference component as input, extracts the key pattern in the trend signal through its internal convolution and pooling layers, and finally outputs a first encoding vector of fixed dimension. Then the second encoder processes the second difference component in the same but independent process and outputs a second encoding vector of fixed dimension.

[0076] For example, an encoder network is used to encode the first difference component into a 128-dimensional vector with values ​​such as [0.12, -0.05, ..., 0.08] as the first encoding vector; and the second difference component is encoded into another 128-dimensional vector such as [0.80, 0.01, ..., -0.10] as the second encoding vector.

[0077] Step a3: Perform feature fusion between the first encoding vector and the second encoding vector to generate a comprehensive difference feature that describes the degree of deviation of the current state of an individual device.

[0078] In step a3, the first encoding vector and the second encoding vector are concatenated in the dimensional direction to form a longer joint feature vector. Then, this joint feature vector is input into a preset feature fusion network. This network learns how to adaptively weigh and integrate the bias information from both long-term trends and short-term fluctuations through its fully connected layers and non-linear activation functions. The final output of the network is a new, unified, dense vector as a comprehensive difference feature.

[0079] Step a4: Based on the comprehensive difference features, a set of update instructions for the predefined calibrable parameters of the state space model are generated through a lightweight parameter adjustment network, and the parameters to be updated are calibrated in a targeted manner according to the parameter identifier, update direction and update magnitude of the parameters to be updated in the update instructions.

[0080] Among them, the parameter tuning network refers to a lightweight fully connected neural network whose function is to map high-dimensional feature descriptions into tuning strategies for specific parameters of downstream models. The update instruction is structured data output by the network. For each parameter to be adjusted, the instruction contains at least three elements: a parameter identifier used to locate the target parameter in the model, an update direction indicating whether the parameter value should be increased or decreased, and an update magnitude specifying the specific numerical value to be adjusted.

[0081] In step a4, a pre-trained lightweight parameter tuning network is invoked. The input dimension of this network matches the dimension of the comprehensive difference feature, and the output dimension is equal to the number of all predefined calibrable parameters multiplied by the dimension of the instruction information required for each parameter. Then, the comprehensive difference feature is input into this network, which outputs a complete set of update instructions through forward computation. These instructions explicitly specify which specific parameters in the state-space model need to be calibrated, such as a parameter labeled "transmission efficiency coefficient". The instructions also include the specific calibration direction for each target parameter, such as increasing the coefficient value by 0.05. Finally, based on these instructions, arithmetic operations are performed on the corresponding parameter values ​​located by parameter identifiers in the state space model to increase or decrease by a specified amount, thereby obtaining the calibrated state space model.

[0082] For example, the comprehensive difference characteristics are input into the parameter adjustment, and two update instructions are output: the first is for the parameter identified as "main bearing friction coefficient", the direction is to increase, and the magnitude is 0.04; the second is for the parameter identified as "hydraulic response gain", the direction is to decrease, and the magnitude is 0.15. Then, according to the instructions, the current value of "main bearing friction coefficient" in the model is increased by 0.04, and "hydraulic response gain" is decreased by 0.15 to complete the calibration.

[0083] Step 1035: Using the calibrated state-space model, perform the final state deduction on the input sequence of the model and output the ideal state trajectory that conforms to the individual characteristics of the device.

[0084] In step 1035, the same model input sequence constructed in step 1031 is used as input data and fed back into the state space model after fine calibration in step 1034. At this time, the model uses the updated parameter set that incorporates the individual characteristics of the current device and performs a complete forward propagation state deduction calculation on the model input sequence again. Since the key parameters of the model have been personalized and adapted according to the actual normal performance of the device before the anomaly, the state sequence deduced this time can more accurately depict the theoretical state evolution path that the device should present under specific conditions if it is always in a healthy and stable operating state. The final output state sequence is the ideal state trajectory.

[0085] This application achieves in-depth and personalized simulation of abnormal equipment states by introducing physical information neural networks and online parameter calibration. It not only provides a state reference trajectory that conforms to physical laws, but also makes the simulation results closely match the actual operating characteristics of specific equipment.

[0086] S104. Based on the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, perform feature tracing based on attention mechanism and causal inference on the abnormal data segment to obtain a sensitive feature set.

[0087] In one specific implementation, step S104 includes: Step 1041: Using the features in the multi-source time series corresponding to the abnormal data segment as nodes, construct an initial causal graph, wherein the edges between the nodes represent the potential causal influence direction; The initial causal graph refers to a graphical model used to represent and analyze causal hypotheses between variables. In this graph, nodes refer to the individual sensor measurement features extracted from multi-source time series, and edges refer to the directed line segments connecting two nodes, with the direction indicating the causal influence that one feature may have on another feature.

[0088] In step 1041, all sensor feature time series that are perfectly aligned with the abnormal data segment in time are first extracted, and each independent feature is defined as a node. Then, based on the preset causal assumption, directed edges are added between node pairs that are considered to have a causal relationship, thereby forming an initial causal graph.

[0089] For example, for an abnormal tunneling data of a certain tunnel boring machine, three key sensor features are selected as nodes: the vibration value V at the front end of the cutterhead, the temperature value T of the main bearing, and the current value C of the drive motor. Then, based on the domain knowledge that "overheating of the bearing may cause abnormal vibration" and "changes in the drive load may affect the vibration", the initial cause-effect graph is preset to contain two directed edges: from node T to node V, and from node C to node V.

[0090] Step 1042: Inject the multidimensional bias into the initial causal graph to obtain the initial energy state of each node.

[0091] Among them, energy state refers to the degree of deviation of the corresponding feature of each node in the abnormal segment. The higher the score, the more obvious the deviation.

[0092] In step 1042, the deviation between the ideal trajectory and the actual data is calculated for each feature dimension; then, for each node in the initial causal graph, the deviation of its corresponding feature is processed by a preset standardization function, which maps the original deviation to a standardized value and uses it as the initial energy state of the node.

[0093] For example, to calculate the deviation between the actual observed values ​​of each feature and the ideal trajectory during an abnormal period, assume the original deviations are: the actual vibration value exceeds the ideal value by an average of 20 units, the temperature by 10 units, and the current by 6 units. Then, divide these three deviation values ​​by the maximum value of 20 to obtain preliminary scaling values: vibration is 1.0, temperature is 0.5, and current is 0.3. However, to facilitate subsequent calculations and reflect the differences, the preliminary scaling values ​​are multiplied by 0.8 to obtain the final normalized energy values. Therefore, the initial energy of the vibration node is 0.8, the temperature node is 0.4, and the current node is 0.24.

[0094] Step 1043: Use the initial causal graph as the current generation causal graph and the initial energy state as the current generation state.

[0095] In step 1043, the constructed initial causal graph is designated as the object to be operated on in the first round of iterative analysis, i.e., the current generation causal graph; at the same time, the initial energy state of each node calculated in step 1042 is designated as the energy value carried by each node at the beginning of the first round of iteration, i.e., the current generation state.

[0096] For example, the initial causal graph containing three nodes V, T, and C and two preset edges is set as the current generation causal graph; then the energy state set (V: 0.8, T: 0.4, C: 0.24) calculated in step 1042 is set as the current generation state.

[0097] Step 1044: Calculate attention weights based on the current generation state, and dynamically mask the nodes in the current generation causal graph according to the attention weights to obtain the key inference subgraph.

[0098] The preset attention weight threshold is a critical value used to determine whether a node is important, and it can usually be set as the average weight of all nodes.

[0099] In step 1044, an attention allocation function, such as the Softmax function, is used to calculate the attention weight of each node based on the current generation energy state value. The higher the energy state value of a node, the greater the attention weight it receives. Then, all the calculated weights are compared with a preset attention weight threshold, and nodes with attention weights lower than the threshold, along with all edges connected to them, are temporarily removed from the current generation causal graph, thereby obtaining the key inference subgraph. It should be noted that the specific calculation process of the Softmax function can be found in relevant technologies, and will not be elaborated here.

[0100] For example, suppose we use the Softmax function to calculate the energy values ​​of three nodes: 0.8, 0.4, and 0.24, i.e.: e 0.8 ≈2.23, e 0.4 ≈1.49, e 0.24 The sum of these values ​​is 4.99, and the attention weights are then calculated as approximately 2.23÷4.99≈0.45, 1.49÷4.99≈0.30, and 1.27÷4.99≈0.25. The preset attention weight threshold is then calculated as the arithmetic mean of these three weights, i.e., (0.45+0.30+0.25) / 3=0.33. Since the weight of the current node C is 0.25, which is lower than this threshold, it is dynamically masked. The resulting key inference subgraph contains only the vibration node V and the temperature node T, which have weights higher than the threshold, as well as the preset edge from T to V between them.

[0101] Step 1045: On the key inference subgraph, apply the preset causal inference rules to verify and correct the direction and strength of the edges to obtain correction instructions. Update the initial energy state of the relevant nodes and the structure of the initial causal graph according to the correction instructions to obtain the updated energy state and the updated causal graph.

[0102] Among them, causal inference rules refer to a set of algorithmic logic used to test and quantify causal relationships between variables, such as the PC algorithm based on conditional independence tests or the Granger causality test in time series data.

[0103] In step 1045, a preset causal inference rule is applied to each directed edge on the obtained key inference subgraph. This rule is based on the actual observation data sequence of the source node and the target node on the edge, and performs statistical tests and calculations to determine whether the causal relationship is valid, whether the direction is correct, and how strong it is. The judgment results are then converted into specific correction instructions. Subsequently, these instructions are executed to modify the initial energy state and the structure of the current generation causal graph, thereby obtaining the updated energy state.

[0104] For example, on a key inference subgraph containing only nodes V and T, a causal inference rule based on a temporal conditional independence test is applied. This rule then analyzes the historical data sequences of V and T and finds that changes in V significantly lead and can be used to predict changes in T, while the reverse is not true. Therefore, a correction instruction is generated: the direction of the edge from T to V is reversed to V to T, and the strength of this causal relationship is quantified to 0.85. The causal graph structure is then updated according to this instruction. At the same time, energy is redistributed according to the new causal direction and strength: part of the energy of the source node V, 0.15, is transferred along the edge from V to T to the target node T. Therefore, the updated energy state is: the energy of node V decreases to 0.65, the energy of node T increases to 0.55, and the energy of node C remains at 0.24.

[0105] Step 1046: Take the updated energy state and the updated causal graph as the current generation causal graph and the current generation state, respectively, and repeat the step of calculating attention weights to update the energy state until the updated energy state and the updated causal graph both reach the corresponding preset stability conditions, thus obtaining a stable energy state and a stable causal graph.

[0106] Among them, the preset stability condition refers to the convergence criterion for determining that the iterative process can be terminated. It usually includes that the change in the energy state of the nodes is less than a minimum threshold and that the structure of the causal graph no longer changes in continuous iterations.

[0107] In step 1046, the updated energy state and the updated causal graph are used as inputs for the next iteration, i.e., the new current generation state and the current generation causal graph. Then, steps 1044 and 1045 are repeated. This process is repeated until a preset stability condition is met. For example, when the change in the energy state value of all nodes in consecutive iterations is less than 0.001, and the set of edges in the causal graph no longer increases, decreases, or changes direction, the system is considered to have converged. At this time, the final causal graph is a stable causal graph, and the energy state to which its nodes are attached is a stable energy state.

[0108] Step 1047: Identify and trace the final source node of all energy inflows under stable energy states in the stable causal graph, and integrate all the final source nodes to obtain a sensitive feature set.

[0109] In step 1047, the transmission path of stable energy states is analyzed based on the stable causal graph. That is, along the direction of the directed edge, the nodes with high energy states but almost no incoming edges or extremely low incoming edge energy contribution are traced. These nodes are regarded as the starting point of energy flow, that is, the final source node that leads to the propagation of abnormal deviation. Then, the original sensor feature identifiers corresponding to these source nodes are summarized into a set, which is the final output sensitive feature set.

[0110] For example, although the temperature node T has a high energy of 0.55, its energy mainly comes from the transmission from the vibration node V to the edge T via V. The vibration node V has an energy of 0.65, and there is no directed edge pointing to it in the entire stable causal graph. That is, no node is inferred to be the energy source of V. Therefore, V is identified as the final source node of the energy flow, and the sensor feature corresponding to this node, "vibration at the front end of the cutter head", is determined to be a sensitive feature that is fundamentally responsible for this anomaly and is included in the set of sensitive features output.

[0111] This application achieves accurate tracing of the root cause characteristics of anomalies by constructing a causal graph and iteratively fusing attention focusing and causal inference, which effectively improves the accuracy and interpretability of fault diagnosis.

[0112] S105. Based on the tenant mode of the IoT data development platform, the abnormal data segment, the ideal state trajectory and the sensitive feature set are encapsulated to obtain a tenant data packet, and the tenant data packet is stored in the data storage area of ​​the corresponding tenant.

[0113] In one specific implementation, step S105 includes: Step 1051: Based on the tenant's unique identifier defined in the tenant mode, extract the data encapsulation components that the corresponding tenant has subscribed to from the IoT data development platform.

[0114] In step 1051, the IoT data development platform first maintains a configuration library that records each tenant and the specific data processing services they subscribe to. Then, when a delivery package needs to be generated for a certain anomaly analysis result, it is first necessary to determine which tenant the result belongs to. Subsequently, based on the unique identifier of the tenant, a query is performed in the platform configuration library to locate the data encapsulation component program that the tenant has pre-ordered and configured, and load it from the component repository into the runtime environment.

[0115] Step 1052: Through the data encapsulation component, perform automated encapsulation processing on the abnormal data segment, the ideal state trajectory, and the sensitive feature set to obtain a structured data packet.

[0116] In step 1052, the data encapsulation component has pre-set data transformation and integration rules, and it automatically receives three core outputs from upstream analysis. Then, the data encapsulation component cleans, converts the format of abnormal data segments, ideal state trajectories and sensitive feature sets according to the established rules. For example, it aligns and binds the timestamps of abnormal data segments with the time axis of ideal state trajectories, and attaches the sensitive feature sets as metadata tags to the entire data packet, thereby generating a structured data packet.

[0117] Step 1053: Determine the service access interface corresponding to the structured data packet according to the preset interface specification, and generate a data list according to the preset data management requirements.

[0118] Among them, the service access interface refers to a predefined programming interface used to access and manipulate data, which specifies how external systems or users can obtain or query the contents of this structured data packet through specific network addresses, protocols and commands; A data manifest is a directory or metadata file that describes the detailed contents of a data packet. It typically includes key descriptive information such as the packet's identifier, generation time, data type, data size, and version information.

[0119] In step 1053, a unique network access address and a set of standard application programming interfaces are automatically assigned to the structured data packet according to the platform's unified interface design specifications; at the same time, the internal structure of the structured data packet is scanned, and key information is automatically extracted according to preset data management requirements to generate a structured data list document.

[0120] Step 1054: Integrate the structured data packet, the service access interface, and the data list to obtain the tenant data packet.

[0121] In step 1054, the structured data packet, the service access interface information defining the access method, and the data list describing the content are physically or logically integrated and bundled. Specifically, a dedicated storage directory or archive file is created, and the above three elements are placed in it. This complete directory or archive file is defined as a finally deliverable tenant data packet.

[0122] This application effectively connects the link from in-depth back-end analysis to front-end business applications, thereby improving the efficiency of delivering analysis results, the ease of cross-system sharing, and the collaborative data service capabilities for different business units.

[0123] Figure 3 This is a schematic diagram illustrating a specific implementation of a time-series data processing system provided in this application. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire the multi-source time series of the target monitoring object collected through the Internet of Things data development platform.

[0124] The identification module 32 is used to perform protocol conversion and time alignment on the multi-source time series through the aggregation gateway to obtain a time series data stream, and to use a stream computing engine to perform pattern recognition within an online sliding window on the time series data stream to obtain abnormal data segments.

[0125] The deduction module 33 is used to call the state space model built based on the physical information neural network to perform state deduction on the abnormal data segment and the corresponding context time sequence, and generate an ideal state trajectory.

[0126] The inference module 34 is used to perform feature tracing based on attention mechanism and causal inference on the abnormal data segment according to the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, and obtain a sensitive feature set.

[0127] The encapsulation module 35 is used to encapsulate the abnormal data segment, the ideal state trajectory and the sensitive feature set according to the tenant mode of the IoT data development platform to obtain a tenant data packet, and store the tenant data packet in the data storage area of ​​the corresponding tenant.

[0128] This application provides a time-series data processing system for implementing the aforementioned time-series data processing method. Therefore, the specific implementation of the time-series data processing system can be found in the embodiment section of the time-series data processing method above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0129] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described time-series data processing methods.

[0130] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described time-series data processing methods.

[0131] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0132] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described time-series data processing method embodiments.

[0133] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0134] The methods and systems provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for processing time-series data, characterized in that, include: Acquire multi-source time series data of the target monitoring object collected through the IoT data development platform; The multi-source time series is converted and time-aligned by the aggregation gateway to obtain a time series data stream. The time series data stream is then subjected to pattern recognition within an online sliding window by a stream computing engine to obtain abnormal data segments. The state-space model built on a physical information neural network is invoked to perform state deduction on the abnormal data segment and its corresponding context time series, generating an ideal state trajectory. Based on the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, the abnormal data segment is subjected to feature tracing based on attention mechanism and causal inference to obtain a sensitive feature set. Based on the tenant mode of the IoT data development platform, the abnormal data segment, the ideal state trajectory and the sensitive feature set are encapsulated to obtain a tenant data packet, and the tenant data packet is stored in the data storage area of ​​the corresponding tenant.

2. The method for processing time-series data according to claim 1, characterized in that, The process of calling a state-space model built on a physical information neural network to perform state deduction on the abnormal data segment and its corresponding context time series, and generating an ideal state trajectory, includes: The abnormal data segments and their corresponding context time series are combined into a model input sequence; An initial state sequence is obtained by performing forward propagation calculation on the input sequence of the model using a pre-built state space model based on a physical information neural network. The state space model integrates the equipment operation mechanism and data-driven capability of the target monitoring object. During the forward propagation calculation, the residuals of the physical constraint terms in the fusion process are calculated, and the state adjustment amount corresponding to the residuals is generated. The state adjustment amount is superimposed on the initial state sequence to obtain the intermediate state trajectory. Extract the normal time sequence segment before the occurrence of the abnormal data segment from the context time sequence, and combine the state difference signal between the intermediate state trajectory and the normal time sequence segment to calibrate the parameters characterizing the individual characteristics of the device in the state space model to obtain the calibrated state space model. Using the calibrated state-space model, the final state deduction is performed on the input sequence of the model, and the ideal state trajectory that conforms to the individual characteristics of the device is output.

3. The method for processing time-series data according to claim 2, characterized in that, The step of calibrating the parameters characterizing individual device characteristics in the state space model by combining the state difference signal between the intermediate state trajectory and the normal time segment includes: Perform multi-resolution wavelet transform on the state difference signal to obtain a first difference component representing the long-term trend state and a second difference component representing the short-term fluctuation state. The first difference component and the second difference component are encoded and mapped respectively to obtain the first encoding vector and the second encoding vector; The first encoding vector and the second encoding vector are fused to generate a comprehensive difference feature that describes the degree of deviation of the current state of an individual device. Based on the comprehensive difference characteristics, a set of update instructions for the predefined calibrable parameters of the state space model are generated through a lightweight parameter adjustment network. The parameters to be updated are then calibrated in a targeted manner according to the parameter identifier, update direction, and update magnitude of the parameters to be updated in the update instructions.

4. The method for processing time-series data according to claim 1, characterized in that, The process involves performing feature tracing based on attention mechanisms and causal inference on the abnormal data segments according to the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, to obtain a sensitive feature set, including: Using the features in the multi-source time series corresponding to the abnormal data segment as nodes, an initial causal graph is constructed, wherein the edges between the nodes represent the potential causal influence directions; The multidimensional bias is injected into the initial causal graph to obtain the initial energy state of each node; The initial causal graph is used as the current generation causal graph, and the initial energy state is used as the current generation state; The attention weights are calculated based on the current generation state, and the nodes in the current generation causal graph are dynamically masked according to the attention weights to obtain the key inference subgraph. On the key inference subgraph, the direction and strength of the edges are verified and corrected by a preset causal inference rule to obtain correction instructions. The initial energy state of the relevant nodes and the structure of the initial causal graph are updated according to the correction instructions to obtain the updated energy state and the updated causal graph. The updated energy state and the updated causal graph are used as the current generation causal graph and the current generation state, respectively. The step of calculating attention weights to update the energy state is repeated until the updated energy state and the updated causal graph both reach the corresponding preset stability conditions, and a stable energy state and a stable causal graph are obtained. In the stable causal graph, the final source node of all energy inflows under stable energy states is identified and traced, and all the final source nodes are integrated to obtain a sensitive feature set.

5. The method for processing time-series data according to claim 1, characterized in that, The process employs a stream computing engine to perform online sliding window pattern recognition on the time-series data stream to obtain anomalous data segments, including: In the stream computing engine, a sliding time window with a configurable length is defined for the time-series data stream, and the time-series data stream is segmented online through the sliding time window to generate multiple data units; An online knowledge distillation framework is built within a stream computing engine, the online knowledge distillation framework comprising a teacher branch and a student branch; The teacher branch and the student branch synchronously perform forward computation on the data unit to obtain the first representation vector and the second representation vector. Calculate the distillation loss value between the first representation vector and the second representation vector, and based on the distillation loss value, calculate the anomaly score corresponding to each data unit to obtain the anomaly score sequence; Concept drift detection is performed on the abnormal score sequence to obtain candidate abnormal time periods. The original data corresponding to all the candidate abnormal time periods are extracted from the time-series data stream to obtain abnormal data segments.

6. The method for processing time-series data according to claim 1, characterized in that, Based on the tenant mode of the IoT data development platform, the abnormal data segment, the ideal state trajectory, and the sensitive feature set are encapsulated to obtain a tenant data packet, including: Based on the tenant's unique identifier defined by the tenant pattern, extract the data encapsulation components that the corresponding tenant has subscribed to from the IoT data development platform; The data encapsulation component performs automated encapsulation processing on the abnormal data segment, the ideal state trajectory, and the sensitive feature set to obtain a structured data packet. The service access interface corresponding to the structured data packet is determined according to the preset interface specification, and a data list is generated according to the preset data management requirements; The structured data packet, the service access interface, and the data list are integrated to obtain the tenant data packet.

7. The method for processing time-series data according to claim 1, characterized in that, The step of performing protocol conversion and time alignment on the multi-source time series data through the aggregation gateway to obtain a time series data stream includes: Create a corresponding protocol parsing instance for each communication protocol within the aggregation gateway; The multi-source time series is parsed in real time through the protocol parsing instance to obtain the original data points. Based on the time synchronization service accessed by the aggregation gateway, a time correction value is calculated for each of the original data points. The time correction value is applied to the corresponding original data point to obtain a calibration data point with a standard timestamp; All the calibration data points are written into a time-aligned buffer in chronological order according to the standard timestamps and then sorted to obtain a time-series data stream.

8. A time-series data processing system, characterized in that, include: The acquisition module is used to acquire multi-source time series data of the target monitoring object collected through the IoT data development platform. The identification module is used to perform protocol conversion and time alignment on the multi-source time series through the aggregation gateway to obtain a time series data stream, and to use a stream computing engine to perform pattern recognition within an online sliding window on the time series data stream to obtain abnormal data segments. The deduction module is used to call the state space model built based on the physical information neural network to perform state deduction on the abnormal data segment and the corresponding context time sequence, and generate an ideal state trajectory. The inference module is used to perform feature tracing based on attention mechanism and causal inference on the abnormal data segment according to the multidimensional deviation between the ideal state trajectory and the actual data in the context time series, and obtain a sensitive feature set. The encapsulation module is used to encapsulate the abnormal data segment, the ideal state trajectory, and the sensitive feature set according to the tenant mode of the IoT data development platform to obtain a tenant data packet, and store the tenant data packet in the data storage area of ​​the corresponding tenant.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a time-series data processing method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables a method for processing time-series data as described in any one of claims 1 to 7.

Images