Industrial data dynamic sensing and acquisition strategy adjusting method

Abstract

The invention provides an industrial data dynamic sensing and acquisition strategy adjusting method, and belongs to the technical field of industrial data analysis. An unsupervised label generation module based on pattern learning generates high-quality phase labels for unlabeled data in batches through multi-dimensional features; the multi-feature fusion hybrid detection network decouples an input signal into a periodic component and a residual component based on a signal decomposition algorithm, extracts depth features of the periodic component and the residual component by adopting a parallel double-branch structure, and comprehensively senses and predicts a system state through a feature fusion network; the reinforcement learning-driven adaptive sampling system is based on a deep reinforcement learning technology, an intelligent decision-making system for dynamically adjusting sampling frequency and acquisition strategies according to the real-time state of the system is constructed, sampling decisions are continuously optimized through interactive learning with the environment, and on the premise of ensuring that key information is not lost, the sampling efficiency is greatly improved. And the collection efficiency and the resource utilization rate are maximized.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, specifically to a method for dynamically sensing and adjusting industrial data acquisition strategies. Background Technology

[0002] Early industrial data acquisition primarily relied on fixed-frequency sampling and simple threshold-based monitoring. These methods are characterized by their simple structure and ease of deployment, but they exhibit poor adaptability to complex and ever-changing industrial environments. Traditional phase detection methods mainly depend on preset threshold judgments and simple statistical analysis. These methods perform reasonably well in simple scenarios where industrial systems operate in relatively stable and well-defined modes, but they demonstrate significant shortcomings in complex, nonlinear, and time-varying industrial environments.

[0003] Deep learning technology has achieved breakthroughs in many fields, and its advantages in nonlinear feature extraction and complex pattern recognition make it a powerful tool for processing time series data. However, its application in intelligent sensing and phase recognition of industrial data still needs further development. Most existing technical solutions only classify time series as a whole, lacking refined modeling of the inherent physical laws of industrial signals, such as decoupling periodic and non-periodic changes.

[0004] While deep learning offers powerful capabilities in complex pattern recognition, its insufficient application in industrial time series analysis highlights a gap between general-purpose deep learning models and the specific needs of industrial data. This means that simply applying off-the-shelf deep learning models is inadequate. To address the inefficiency of fixed sampling strategies, some research has begun exploring adaptive sampling methods based on rules or simple feedback mechanisms. These methods attempt to adjust the sampling frequency according to certain preset conditions, but they are essentially still combinations of static rules and struggle to cope with the various uncertainties and unforeseen circumstances that arise during the operation of industrial systems.

[0005] Traditional manual annotation is not only time-consuming and costly, but more seriously, its quality is highly susceptible to subjective factors when processing time-series data. Furthermore, industrial field data often suffers from missing labels and class imbalance, further increasing the difficulty of effective annotation and model training. Traditional methods primarily rely on preset thresholds and simple statistical analysis, based on the assumption of stable system operation. However, the operating modes of real industrial systems are influenced by multiple factors such as load variations, environmental conditions, and equipment aging, exhibiting high nonlinearity and time-varying characteristics. The static, rule-based nature of traditional methods makes them ill-suited to this complexity, leading to phase misjudgments and response lags.

[0006] Most existing data acquisition systems use fixed-frequency sampling, which cannot be dynamically adjusted according to system status. When the system is running smoothly, an excessively high sampling frequency will result in a huge waste of computing and storage resources; while when the system status undergoes critical changes, a fixed low sampling frequency may miss important transient information, creating a monitoring blind spot. Summary of the Invention

[0007] This invention proposes a method for dynamic sensing and acquisition strategy adjustment of industrial data, which constructs an unsupervised label generation module based on pattern learning, a hybrid detection network based on multi-feature fusion, and an adaptive sampling system driven by reinforcement learning.

[0008] The pattern learning-based unsupervised label generation module automatically learns from historical data and combines multi-dimensional features such as numerical values, gradients, and positions within periods to generate phase labels in batches for unlabeled data.

[0009] The multi-feature fusion hybrid detection network is based on a signal decomposition algorithm, which decouples the input signal into periodic components and residual components. It uses a parallel dual-branch structure to extract the deep features of the periodic components and residual components respectively, and predicts the system state through a feature fusion network.

[0010] The reinforcement learning-driven adaptive sampling system is based on deep reinforcement learning technology and dynamically adjusts the sampling frequency and acquisition strategy according to the real-time state of the system.

[0011] In a preferred embodiment, the pattern learning-based unsupervised label generation module extracts multiple complete periodic samples from historical operating data, performs time series periodicity detection, and determines the system's period length; predefines four operating phases, constructs a template containing three types of feature constraints for each phase; and performs phase labeling on the time series data to be labeled, generating phase labels.

[0012] In a preferred embodiment, the four predefined operating phases include: rising phase, peak stable phase, falling phase, and trough stable phase;

[0013] Three types of feature constraints are constructed for each phase: numerical range constraints, gradient range constraints, and location prior constraints. By analyzing historical average patterns, the typical location points of each phase within the period are statistically determined, and a priority mechanism is introduced to assign high priority to peak stable phases and trough stable phases.

[0014] In a preferred embodiment, for the time series data to be labeled, the system uses a sliding window technique to perform phase labeling point by point, and fuses numerical matching scores S. value (t), gradient matching score S gradient (t) and the prior score of the periodic position S position(t), calculate the matching score between the data point at time t and each phase template. i (t):

[0015] Score i (t)=w1·S value (t)+w2·S gradient (t)+w3·S position (t);

[0016] Where w1 = 0.4, w2 = 0.4, and w3 = 0.2 are the weights of numerical matching, gradient matching, and location prior, respectively; the scores of each component are calculated using the following formula:

[0017] Numerical matching score:

[0018] Gradient matching score:

[0019] Prior score of position:

[0020] In a preferred embodiment, a multi-stage global optimization mechanism is introduced; for regions with values ​​higher than the first threshold or lower than the second threshold and with gradients less than the third threshold, labels of peak stable phase or trough stable phase are assigned; a sliding window is used to smooth the initially generated label sequence; the entire label sequence is traversed to check for unreasonable phase jumps, and if any exist, the phase of the previous point is retained based on the confidence score.

[0021] In a preferred embodiment, the multi-feature fusion hybrid detection network adopts a deep network architecture based on signal decomposition, modeling the input industrial time series signal X(t) as a linear superposition of three parts:

[0022] X(t)=X cyclic (t)+X residual (t)+∈(t),

[0023] Among them, X cyclic (t) represents the periodic component in the signal; X residual (t) represents the residual component remaining after removing the periodic component, and ∈(t) represents random noise;

[0024] Construct a two-branch network structure to decouple the input signal into a periodic component X. cyclic (t) and residual component X residual (t), using the periodic component processing branch and the residual component processing branch, deep feature extraction is performed on the periodic component and the residual component respectively.

[0025] In a preferred embodiment, the learnable periodic extractor uses a learnable periodic template matrix W∈R T×D The core periodic component is extracted from the original input signal, where T is a predetermined signal period length and D is the feature dimension of the data; for any time t in the input sequence, its corresponding periodic component X cyclic (t) is calculated by indexing the periodic template Q matrix:

[0026] X cyclic (t) = Q[t mod T,:]

[0027] The periodic feature analyzer uses a sinusoidal position encoding matrix. The purified periodic component X was extracted. cyclic (t) performs in-depth analysis, d model It is the hidden layer dimension of the model;

[0028] The periodic component X is encoded using an encoder. cyclic (t) is encoded to obtain the periodic feature F. cyc Based on the current data point's index within the cycle (cycle_index), the corresponding position vector F is retrieved from the sinusoidal position encoding matrix P. pos The two feature vectors are concatenated to form a joint feature vector F. joint The data is fed into the feedforward network to initially determine the phase of the cycle.

[0029] In a preferred embodiment, the residual component processing branch removes periodic components from the original signal to obtain a residual signal sequence. By analyzing the current value of the residual signal sequence and its changes before and after, it determines whether the current trend of the system is rising, stable, or falling. Combining the residual signal sequence characteristics and trend characteristics, it quantifies the current volatility of the system and outputs a normalized volatility intensity value.

[0030] In a preferred embodiment, the deep features from the periodic branch, the deep features from the residual branch, and the trend classification probability and normalized fluctuation intensity value output by the residual branch are concatenated to generate a fused feature vector. This fused feature vector is then fed into a deep fusion network. The final feature vector after deep fusion is distributed to three parallel final phase classifiers, an intra-period position regressor, and a confidence estimator, simultaneously completing three independent prediction tasks:

[0031] The final phase classifier predicts the current phase prediction result of the system based on all the fused final feature vectors;

[0032] The periodic position regressor regresses the relative position of the current data point within the period and outputs a continuous value from 0 to 1.

[0033] Confidence estimator: Estimates the confidence level of the current phase prediction result and outputs a value between 0 and 1.

[0034] In a preferred embodiment, the reinforcement learning-driven adaptive sampling system constructs a 9-dimensional state vector from the current phase, detection confidence, system volatility, and time since the last sampling output of the multi-feature fusion hybrid detection network. The sampling decision policy π(s) is learned through a deep Q-network to maximize long-term cumulative reward.

[0035] π(s)=argmax a Q(s,a);

[0036] The trained master Q network Q main As the optimal strategy π * (s)=argmax a Q main (s,a;θ * This enables intelligent adaptive sampling decision-making, where π(s) refers to the sampling decision-making strategy learned by the system. * (s) refers to the optimal strategy.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. Overcome the shortcomings of high cost and difficulty in guaranteeing the quality of data annotation.

[0039] This invention achieves automated, unsupervised label generation through the PatternBasedPhaseAnalyzer module. The automated labeling process is over a thousand times faster than traditional manual methods. This significantly shortens project cycles, reducing manual labeling work that previously took months to just hours, thus substantially lowering time and financial costs.

[0040] 2. Overcoming the limitations of traditional phase detection methods

[0041] The HybridCyclePhaseNet hybrid detection network proposed in this invention exhibits superior performance, achieving a technological breakthrough. Test results on the Electricity validation set show that the model achieves an overall phase detection accuracy of up to 95.2%, with an average prediction confidence of 0.996.

[0042] 3. Solve the inefficiency problem of fixed sampling strategies.

[0043] The dynamic adaptive sampling strategy introduced by this invention through the AdaptiveSamplingSystem module demonstrates an overwhelming advantage in comparative tests. Compared to rule-based strategies, the reinforcement learning (RL) agent of this invention achieves a 100% capture rate of key events, i.e., zero omissions.

[0044] 4. To compensate for the insufficient application of deep learning

[0045] The HybridCyclePhaseNet module of this invention achieves deeper and more effective applications through its domain-aware structured design. HybridCyclePhaseNet outperforms the advanced general-purpose time series model PrecTime (93.88%) in accuracy (95.2%), proving that the dedicated architecture designed for the characteristics of industrial data is more effective than the general-purpose architecture.

[0046] 5. Establish a truly adaptive decision-making mechanism

[0047] The AdaptiveSamplingSystem module of this invention constructs an intelligent decision-making system with learning and self-optimization capabilities by introducing reinforcement learning. In a dynamic testing environment, the learning-based RL agent achieves a 100% key event capture rate.

[0048] 6. Achieve comprehensive multi-scale temporal feature processing

[0049] The HybridCyclePhaseNet architecture of this invention, through its unique design, achieves effective processing and fusion of features at different time scales. The model not only exhibits extremely high recognition accuracy for stable phases with long durations but also performs exceptionally well for transitional phases ("rising" and "falling") with shorter durations and faster changes. This demonstrates that the model can simultaneously and accurately identify system states defined by features at different time scales. The effectiveness of this invention's multi-scale processing mechanism enables it to make more accurate judgments based on a contextual understanding of multi-scale features. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the unsupervised label generation module based on pattern learning.

[0051] Figure 2 This is a diagram of a hybrid detection network architecture that integrates multiple features.

[0052] Figure 3 A flowchart for a reinforcement learning-driven adaptive sampling system;

[0053] Figure 4 Power grid load data from three specific sensors;

[0054] Figure 5 Four types of phase tag load curves and their time-varying logical sequences;

[0055] Figure 6 This is a performance evaluation diagram of the phase recognition model;

[0056] Figure 7 A graph showing the average round reward curve for the agent;

[0057] Figure 8 To explore the rate curve;

[0058] Figure 9 This is a graph showing the average Q value.

[0059] Figure 10 This is a confusion matrix diagram;

[0060] Figure 11 A confidence distribution chart for phase prediction;

[0061] Figure 12 For the accuracy of each phase detection;

[0062] Figure 13 This refers to the real-time decision-making process of an adaptive sampling system. Detailed Implementation

[0063] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0064] This invention constructs a multi-layered intelligent sensing system, which includes three mutually cooperating technical modules:

[0065] Pattern Based Phase Analyzer: This unsupervised label generation module specifically addresses the labeling challenges of large-scale industrial time series data. It automatically learns typical system operating patterns from historical data and combines multi-dimensional features such as numerical values, gradients, and intra-cycle positions to generate high-quality phase labels in batches for unlabeled data. When handling multivariate data, this module also includes a voting mechanism to integrate judgments from multiple data sources, enhancing the robustness of the labels.

[0066] This unsupervised label generation module avoids the high cost and inconsistency of traditional manual labeling, providing sufficient and reliable labeled data for the training of subsequent deep learning models.

[0067] Hybrid Cycle Phase Net, a multi-feature fusion hybrid detection network, is designed to address the multimodal characteristics of industrial system data. Based on signal decomposition, this network decouples the input signal into periodic and residual components, and employs a parallel dual-branch structure to extract their respective deep features. Finally, through an efficient feature fusion network, comprehensive perception and accurate prediction of the system state are achieved.

[0068] The Reinforcement Learning-Driven Adaptive Sampling System (DQN) is an intelligent decision-making system built on deep reinforcement learning (DQN) technology. It dynamically adjusts the sampling frequency and acquisition strategy based on the system's real-time state. The system treats the phase detection model as part of the environment, continuously optimizing sampling decisions through interactive learning. This maximizes acquisition efficiency and resource utilization while ensuring no key information is lost.

[0069] The strength of reinforcement learning-driven adaptive sampling systems lies not only in the technological advancement of each module, but also in the tight coupling and virtuous cycle formed between them.

[0070] First, the unsupervised label generation module based on pattern learning solves the problem of data source for subsequent deep learning model training. It automatically generates high-quality phase labels for massive amounts of unlabeled industrial data, completely breaking the bottleneck of data labeling.

[0071] Secondly, the multi-feature fusion hybrid detection network, fed by these high-quality labels, uses its innovative signal decomposition architecture to achieve high-precision dynamic perception of the system state. The output is no longer a single classification result, but a multi-dimensional, three-dimensional state description including phase, confidence, and volatility. Finally, the reinforcement learning-driven adaptive sampling system uses this rich state information as a decision-making basis. Through interactive learning with the environment, it dynamically optimizes the data acquisition frequency, maximizing resource utilization efficiency while ensuring no key information is lost. This optimized acquisition system, in turn, provides higher-quality input data for the entire system, thus forming an intelligent closed loop capable of continuous self-optimization and iteration.

[0072] I. Unsupervised Label Generation Module Based on Pattern Learning

[0073] This unsupervised label generation module learns system operation patterns from historical data in an automated and unsupervised manner, generating high-quality, highly consistent phase labels in batches. It systematically integrates various technologies such as historical pattern learning, multi-dimensional feature matching, priority mechanisms, and global consistency optimization, simulating and automating the annotation logic of domain experts. A flowchart of this module is shown below. Figure 1 As shown. The specific steps are as follows:

[0074] Step 1: Historical pattern extraction and learning.

[0075] The system first identifies and extracts multiple complete periodic samples from historical operational data. Using the Autocorrelation Function (ACF) analysis method, it detects time series periodicity for multiple data features and automatically determines the dominant period length, `cycle_len`. The autocorrelation function measures the correlation between a time series and its lagged versions. For periodic data, the autocorrelation function peaks when the lag is an integer multiple of the period length. The system calculates the autocorrelation function values ​​for different lags and identifies the lag corresponding to the first significant peak as the dominant period length. For each identified complete period, independent min-max normalization is performed to eliminate the influence of absolute numerical magnitude and focus on the pattern itself. Through statistical analysis, the average pattern and standard deviation pattern of multiple periods are calculated to identify the typical operating patterns of the system.

[0076] Accurately identify the inherent periodicity of the data and extract stable, dimensionless typical periodic patterns, providing a data-driven basis for subsequent phase template construction; automated period detection and pattern normalization processing avoid the errors and tediousness of manual experience judgment, ensuring the objectivity and consistency of pattern extraction.

[0077] Step 2: Construction and optimization of phase templates based on domain knowledge.

[0078] Step 21: Phase definition.

[0079] Based on general understanding of industrial systems, the system predefines four basic operating phases and clarifies their physical meanings:

[0080] Rising phase: The value shows an upward trend, and the gradient is positive.

[0081] Peak stability: The value is at a high level and changes slowly, with the gradient close to zero.

[0082] The decreasing phase: the value shows a decreasing trend and the gradient is negative.

[0083] Low-level stability: The value is at a low level and changes slowly, with the gradient close to zero.

[0084] Step 22: Multidimensional feature constraints.

[0085] A template containing three types of feature constraints was constructed for each phase, thereby enabling multi-dimensional comprehensive judgment:

[0086] Numerical range constraint: Defines the reasonable range of normalized data values ​​under this phase, i.e., V min ≤v(t)≤V max V min and V max These refer to the reasonable minimum and maximum values ​​of the normalized data value v(t) under a specific phase, respectively. v(t): represents the normalized data value at time t.

[0087] Gradient range constraint: Defines the reasonable range of the rate of change (gradient) of the data at this phase, i.e. Among them G min and G max These refer to the reasonable minimum and maximum values ​​of the rate of change of data, i.e., the gradient, at a specific phase, dv(t).

[0088] / dt: represents the rate of change (gradient) of the data at time t.

[0089] Specifically:

[0090] Ascending phase: Numerical range: (0.1, 0.8), Gradient range: (0.01, 1.0)

[0091] Peak stability: numerical range: (0.7, 1.0), gradient range: (-0.025, 0.025)

[0092] Descent phase: Numerical range: (0.2, 0.9), Gradient range: (-1.0, -0.01)

[0093] Valley stability: numerical range: (0.0, 0.3), gradient range: (-0.025, 0.025)

[0094] Prior location constraints: By analyzing the historical average patterns learned previously, the system statistically analyzes the set P of the most frequently occurring typical location points within the period for each phase. typical =p1, p2, ..., p k This provides important contextual information for phase determination; for example, peaks typically occur in the middle of the cycle.

[0095] It should be noted that the typical location points are determined by analyzing the historical average patterns learned previously, and the system statistically analyzes the set P of the most frequently occurring typical location points for each phase within the period. typicalThis means that the system analyzes average patterns learned from historical data to determine which time points or time periods a particular phase (such as the "rising phase" or "peak stabilization") most frequently occurs within the entire cycle (e.g., a 24-hour cycle). This "position" is relative to a complete cycle.

[0096] For example, if a period is 24 hours, the system might find that the "peak stabilization" phase typically occurs between the 12th and 16th hours of the period. These hourly points (or more precise time points) can then be considered the typical set of locations for "peak stabilization".

[0097] Step 23: Priority Mechanism

[0098] To address the potential for ambiguous classification in phase transition regions—for example, a point that resembles both the end of a "rise" and the beginning of a "peak"—a priority mechanism is introduced. Phases with clearly defined states and significant characteristics, such as "peak stability" and "trough stability," are assigned higher priority.

[0099] Specifically, when a point's score is within 60% of the highest score, additional points will be awarded based on priority:

[0100] priority_bonus=(3-priority)*0.15;

[0101] Priority Bonus: This is a calculated bonus value. When the difference between the highest and second-highest scores among the matching scores of a data point and each phase template is within a threshold, that is, when a point's score is within 60% of the highest score, the system will add this bonus value to the score according to the phase priority to solve the fuzzy classification problem that may occur in the phase transition region.

[0102] Priority: This is a preset integer value used to represent the priority of different phases.

[0103] Priority values ​​are pre-set to 1 for peak and trough phases, and 2 for rising and falling phases. This ensures that when a data point has similar matching scores to multiple templates, the higher-priority phase is selected first, effectively improving the accuracy of annotation in key stable regions.

[0104] Integrating domain knowledge into the label generation process ensures that the generated labels conform to their actual physical meaning. A priority mechanism resolves ambiguity issues arising from overlapping phase features. Combining data-driven learning with expert knowledge constraints improves label accuracy and robustness. The priority mechanism is an effective means of addressing the ambiguous boundaries caused by the complexity of industrial data.

[0105] Step 3: Calculate the matching score

[0106] Having prepared the "templates" and "rules" for annotation through the preceding steps, the system then uses a sliding window technique to perform phase annotation point by point for the time series data to be annotated.

[0107] The historical pattern extraction and learning in step 1 and the domain knowledge-based phase template construction and optimization in step 2 are to prepare the "templates" and "rules" for annotation. The "time series data to be labeled" is the "new data" or "unlabeled data" that will be automatically assigned phase labels by the system using these constructed phase templates and rules. The preceding steps provide the foundation and tools for this annotation process.

[0108] For each data point at time t, calculate its matching score with each phase template. i (t), where i refers to the index of different phase templates, i = 0, 1, 2, 3, representing the four phase templates respectively. This matching score formula is the first of its kind proposed in this invention specifically for the characteristics of industrial data, by fusing numerical matching scores S. value (t), gradient matching score S gradient (t) and the prior score of the periodic position S position (t) aims to address the problems of poor robustness of traditional thresholding methods under complex conditions and high cost of manual annotation:

[0109] Score i (t)=w1·S value (t)+w2·S gradient (t)+w3·S position (t)

[0110] Where w1 = 0.4, w2 = 0.4, and w3 = 0.2 are the weights for numerical matching, gradient matching, and location prior, respectively. The scores for each component are calculated based on the proximity of the data point features to the range and typical locations defined in the template.

[0111] Numerical matching score:

[0112] Where v(t) refers to the normalized data value at time t, center i The range refers to the center point of the numerical range defined by the i-th phase template. i The width of the numerical range defined by the i-th phase template (i.e., V) max,i -V min,i V min,i V refers to the reasonable minimum value of the normalized data defined by the i-th phase template. max,iThis refers to the reasonable maximum value of the normalized data value defined by the i-th phase template.

[0113] Gradient matching score:

[0114] Where g(t) refers to the rate of change (gradient) of the data at time t, center g,i The range refers to the center point of the gradient range defined by the i-th phase template. g,i The width of the gradient range defined by the i-th phase template (i.e., G) max,i -G min,i ), G min,i G refers to the reasonable minimum value of the gradient defined by the i-th phase template. max,i This refers to the reasonable maximum value of the gradient defined by the i-th phase template.

[0115] Prior score of position:

[0116]

[0117] Where T represents the period length of the input sequence, cycle_len,min distance ((tmodT),P typical,i The term P refers to the set of typical position points P of the current time point t within the period and the i-th phase template. typical,i The closest distance in the middle. The similarity between the quantified data points and each phase template provides a basis for preliminary phase labeling. The multi-dimensional matching score formula solves the problem of poor robustness of traditional threshold methods under complex working conditions and effectively utilizes periodic position information.

[0118] Step 4: Global consistency optimization.

[0119] To ensure logical consistency in tag generation, the system introduces a multi-stage global optimization mechanism:

[0120] Forced logical constraints: In the time series data, for regions with extremely high values ​​(value>0.85) or extremely low values ​​(value<0.15) and gentle gradients (|grad|<0.03), the labels "peak stable" or "trough stable" are forcibly assigned to ensure the correctness of the physical meaning.

[0121] Sliding window smoothing: A sliding window of size max(3, cycle_len / / 8) is used to smooth the initially generated label sequence. cycle_len / / 8 is the cycle length cycle_len divided by 8 and rounded down.

[0122] This effectively eliminates isolated erroneous labels caused by noise or outliers in a single data point, such as a sudden "falling" label appearing in a "rising" label sequence, making the label sequence more coherent.

[0123] Phase transition logic check: Traverse the initially generated time series label sequence to check for unreasonable phase jumps, such as jumping directly from "rising" to "falling". If such a jump exists, the phase of the previous point is retained based on the confidence score, thus ensuring that the final label sequence follows a predefined logical order, such as rising phase → {rising phase, peak stability}; peak stability → {peak stability, falling phase}; falling phase → {falling phase, trough stability}; trough stability → {trough stability, rising phase}.

[0124] The system improves the physical plausibility and temporal continuity of generated labels, eliminating noise and illogical tags. Multi-stage global optimization is introduced, particularly mandatory logical constraints and phase transition logic checks, significantly improving the accuracy and usability of labels and resolving common local inconsistencies in automated labeling.

[0125] Step 5: Multi-column data fusion and voting mechanism.

[0126] When the input data contains multiple feature columns, such as data from different sensors, the pattern-learning-based unsupervised label generation module independently runs a complete phase detection process for each column, generating an independent set of phase labels and confidence scores for each column. At each time point t, the prediction results for all columns are collected, and a majority voting principle (implemented via np.argmax(np.bincount(votes))) is used to select the phase predicted most frequently at that time point as the final unified label. For example, if two of the three sensors predict the current phase as "rising," then the final label for that time point is "rising."

[0127] By integrating information from multiple data sources, the robustness of label generation is enhanced, reducing the impact of noise or anomalies from a single data source. The voting mechanism is a key innovation for generating robust labels when processing multivariate industrial data, effectively improving the reliability of the labels.

[0128] This unsupervised label generation module, based on pattern learning, incorporates a multi-step heuristic algorithm that combines historical pattern learning, multi-dimensional feature (numerical, gradient, and intra-period position) matching, a priority mechanism, and global consistency optimization. It systematically integrates various data processing techniques to automatically generate high-quality phase labels.

[0129] II. Hybrid Detection Network with Multi-Feature Fusion

[0130] To address the issues mentioned in the background art, such as insufficient application of deep learning, difficulties in multi-scale feature processing, and limitations of traditional methods, a core objective of this invention is to provide a multi-feature fusion hybrid detection network, named Hybrid Cycle Phase Net. This network aims to perform refined analysis of complex industrial time-series data through an innovative deep learning architecture based on signal decomposition, thereby achieving comprehensive, accurate, and robust perception and prediction of system operating status.

[0131] The multi-feature fusion hybrid detection network adopts a deep network architecture based on the signal decomposition idea. The theoretical basis of this architecture is to model the input complex industrial time series signal X(t) as a linear superposition of three parts, and its mathematical model can be expressed as:

[0132] X(t)=X cyclic (t)+X residual (t)+∈(t),

[0133] Among them, X cyclic (t) represents the regular and predictable periodic components in the signal; X residual (t) represents the residual component remaining after removing the periodic component. This component usually contains key information such as irregular fluctuations, external disturbances, long-term trends, and abnormal events of the system; while ∈(t) represents random noise.

[0134] Based on this mathematical model, this invention designs a parallel dual-branch network structure, firstly decoupling the input signal into the two core components X mentioned above. cyclic (t), X residual (t), and then using two dedicated network branches with different structures and functions, deep feature extraction is performed on the periodic component and the residual component respectively, so that each branch can focus more on processing its corresponding signal characteristics, thereby achieving an analysis depth and accuracy far exceeding that of a single model.

[0135] 2.1. Dual-branch signal processing architecture

[0136] The overall architecture of Hybrid Cycle Phase Net is built around the idea of ​​signal decomposition, comprising two parallel core processing streams: a periodic component processing branch and a residual component processing branch. The input time-series data first passes through a learnable periodic extractor, decomposing it into periodic and residual components. These two parts are then fed into their respective dedicated branches for parallel analysis. Finally, the deep features extracted by the two branches are integrated in an efficient fusion module to produce the final, comprehensive state-aware result.

[0137] 2.1.1. Learnable Cycle Extractor

[0138] The learnable periodic extractor is responsible for separating the core periodic component from the original input signal. Its core is a learnable periodic template matrix W∈R. T×D Where T is a predetermined signal period length (e.g., for daily load data, T = 24), and D is the feature dimension of the data.

[0139] For any time t in the input sequence, its corresponding periodic component X cyclic (t) is calculated by indexing the periodic template matrix, and the specific formula is as follows:

[0140] X cyclic (t) = Q[t mod T,:]

[0141] The t mod T operation here ensures that no matter how far away the current time point t is in the sequence, it can be mapped back to the corresponding position in the periodic template.

[0142] For an input sequence of length L, its corresponding periodic component sequence c t-L+1:t The calculation is performed by aligning and repeating the template matrix. Specifically, the model shifts the template matrix Q to the left based on the current time step's relative position within the period (i.e., t mod T), resulting in the aligned periodic template Q(t). Then, Q(t) is repeated according to the length L of the input sequence. Next, and splice it together. To construct a periodic component c of the same length as the input sequence. t-L+1:t .

[0143] This process can be represented by the following formula:

[0144]

[0145] c t-L+1:t It is a periodic component sequence of length L, which is constructed by aligning and repeating the learned periodic template, covering the time range from t-L+1 to t.

[0146] And X cyclic (t) refers to the specific periodic component value at time t in a periodic sequence of length L. The relationship between the two is: X cyclic (t)=c t-L+1:t [L-1].

[0147] Crucially, the periodic template matrix Q is not a manually preset fixed waveform, but rather an automatically learned network parameter acquired through end-to-end backpropagation training. This allows the model to capture the most representative periodic patterns unique to specific industrial processes, regardless of their complexity or non-standard shapes. Periodicity and smoothness constraints are incorporated into the learning process to ensure that the final learned template possesses good physical properties. The design of the learnable periodic extractor originates from CycleNet's core technology, Residual Cycle Forecasting (RCF), which utilizes learnable cycle periods to explicitly model periodic patterns in time series data.

[0148] 2.2. Periodic Component Treatment Branch

[0149] The mission of this branch is to gain a deep understanding of the stable, repetitive periodic patterns in signals and to accurately determine their macroscopic phase.

[0150] After extracting the pure periodic component X cyclic After (t), the periodic feature analyzer is responsible for in-depth analysis. Its unique feature is that it effectively combines the characteristics of the periodic component itself with its precise position information within the period.

[0151] To achieve position awareness, this module employs a standard sinusoidal position coding matrix. Where d model This is the hidden layer dimension of the periodic feature analyzer. The formula for calculating the element in the j-th column of the t-th row (representing the t-th position within the period) of the sinusoidal position encoding matrix is ​​as follows:

[0152]

[0153] The feature fusion process is as follows: First, an encoder is used to process the periodic component X. cyclic (t) is encoded to obtain the periodic feature F. cyc Then, the index of the current data point within the period, cycle_index, is obtained by periodically calculating t mod T. The corresponding position vector F is then retrieved from the sinusoidal position encoding matrix P using this index. pos Finally, these two feature vectors are concatenated to form a joint feature vector F. joint The vector is fed into a feedforward network consisting of a linear layer (nn.Linear), a ReLU activation function, and Dropout, to initially determine the period phase to which the current data point belongs.

[0154] 2.3. Residual Component Processing Branch

[0155] This residual component processing branch is specifically responsible for analyzing the residual signal X obtained after removing periodic components from the original signal. residual (t). Although this part of the signal usually has a small amplitude, it often contains crucial information such as abnormal state of the system, sudden interference from the external environment, and long-term performance degradation trend.

[0156] The core architecture of this residual component processing branch is a multi-scale one-dimensional convolutional neural network (1D-CNN). By using the nn.Conv1d module and convolutional kernels of different sizes, it can effectively capture local patterns and trends of change at different time scales in the residual sequence.

[0157] In addition to extracting deep features, this branch also performs two specific analysis tasks:

[0158] Trend classification probability analysis: By analyzing the current value of the residual sequence and its changes before and after, a specialized classifier is trained to determine whether the current macro trend of the system is "rising", "stable" or "falling".

[0159] Volatility intensity estimation: Combining sequence features and trend features, a regression network is used to quantify the current volatility of the system, outputting a normalized volatility intensity value between 0 and 1.

[0160] 2.4. Feature Fusion and Multi-Task Output Layer

[0161] The final stage of the network involves intelligently fusing information from the two parallel branches to generate a final, multi-dimensional state description.

[0162] Feature fusion mechanism: Deep features from the periodic branch, deep features from the residual branch, the trend classification probability output by the residual branch, and the normalized fluctuation intensity value are all concatenated into a higher-dimensional, more comprehensive fused feature vector. This vector is then fed into a deep fusion network consisting of a fully connected layer (Linear), layer normalization (LayerNorm), activation function (ReLU), and Dropout to learn the complex nonlinear interactions between features from different sources.

[0163] Multi-task output layer: The final feature vector after deep fusion is distributed to three parallel output heads (Prediction Heads) to simultaneously complete three independent prediction tasks:

[0164] Final phase classifier: Based on all fused information, predicts the most likely operating phase of the system at present (e.g., "rising", "peak stable", "falling", "trough stable").

[0165] Intra-cycle position regressor: Performs precise regression on the relative position of the current data point within the cycle, outputting a continuous value from 0 to 1. This provides more refined positioning information than a simple cycle_index.

[0166] It's important to note the range of 0 to 1: the numerical range predicted by this regressor is normalized to between 0 and 1. 0 represents the start of a cycle, while 1 represents the end of a cycle (or the start of the next cycle).

[0167] Continuous value: This value is not a discrete integer (like the cycle index, which is an integer), but a floating-point number that can be any value between 0 and 1. This means it can more precisely indicate the "position" of the current data point within the cycle; for example, 0.25 might indicate that a quarter of the cycle has elapsed, and 0.5 might indicate that the cycle is halfway through. This is much more precise than simply knowing which point in the cycle it is (cycle_index).

[0168] Confidence estimator: Estimates the model's confidence in the current phase prediction result, and also outputs a value between 0 and 1.

[0169] This multi-task output design elevates the model's role from a simple "classifier" to a "state description generator." It no longer outputs an isolated label, but a rich state vector encompassing multiple dimensions such as phase, position, confidence, volatility, and trend. This comprehensive state vector provides downstream intelligent decision-making systems (such as Module 3 in the overall scheme of this invention: reinforcement learning adaptive sampling strategy) with unprecedented breadth and depth of information, forming the cornerstone for achieving true intelligent perception.

[0170] 2.5. Composite Loss Function

[0171] To effectively train the Hybrid Cycle Phase Net with its complex architecture and multiple outputs, this invention designs and employs an original composite loss function. This loss function embodies the core idea of ​​multi-task learning, combining the losses of multiple key prediction tasks in the network through a weighted sum to jointly drive the model's optimization process. This approach not only improves the performance of the main task but also, by introducing supervisory information from auxiliary tasks, acts as a regularization mechanism, forcing the model to learn more fundamental and robust feature representations.

[0172] The composite loss function L total The mathematical expression is:

[0173] L total =λ1L phase +λ2Lposition +λ3L trend

[0174] The components and their weights are defined as follows:

[0175] L phase The main task, phase classification loss, is calculated using the cross-entropy loss function. Its weight λ1 = 0.3, which is the highest among all tasks, to ensure that the main direction of model optimization is to improve the accuracy of phase recognition.

[0176] L position The position regression loss within the auxiliary task period is calculated using the Mean Squared Error (MSE) function. Its weight λ² = 0.1. This task forces the model to learn the fine temporal structure of the data within the period.

[0177] L trend The auxiliary task residual trend classification loss is also calculated using the mean squared error loss function (MSE). Its weight λ3 = 0.1. This task-guided model explicitly learns latent trend information from the non-periodic residual components.

[0178] The design of this loss function itself embodies the invention's core concept. The weight allocation scheme (λ1 = 0.3, λ2 = 0.1, λ3 = 0.1) clearly reveals the invention's intent: phase classification is the core objective, but understanding the position within the period and grasping the residual trend are two key supports for achieving robust classification. The existence of these two auxiliary tasks effectively prevents the model from taking "shortcuts" when optimizing the main task, i.e., relying solely on some superficial or spurious correlations in the data. By simultaneously optimizing these three tasks, the model is guided to build a more fundamental and comprehensive internal model of the data's inherent physical laws, thereby achieving stronger generalization ability.

[0179] 2.6. Specific Architecture of the Hybrid Cycle Phase Net Model

[0180] 1) Signal decomposition module

[0181] Function: Receives the original input sequence and decomposes it into periodic components and residual components according to the periodic index.

[0182] Implementation: This decomposition is achieved through a learnable periodic extractor.

[0183] 2) Periodic Feature Extraction Branch

[0184] Function: Focuses on analyzing the periodic components extracted from signals.

[0185] Implementation: Composed of a periodic feature analyzer. It combines the characteristics of periodic components with positional information within the period (through positional encoding) to preliminarily determine the periodic phase to which the data belongs.

[0186] 3) Residual Trend Analysis Branch

[0187] Function: Specifically designed to process and analyze the residual components obtained after removing periodic components from the original signal. This portion of the signal typically contains key information such as irregular fluctuations, external interference, long-term trends, and anomalous events.

[0188] Implementation: This branch uses a one-dimensional convolutional network to capture local patterns and trends in the residual sequence. It also performs trend analysis and fluctuation intensity estimation.

[0189] 4) Feature blending layer (feature fusion)

[0190] Function: Fuse all deep features from the periodic feature extraction branch (such as periodic features) and the residual trend analysis branch (such as residual features, trend probability, and volatility intensity).

[0191] Implementation: These multi-source features are concatenated to form a more comprehensive and information-rich joint feature vector, which is then processed through linear layers, layer normalization, and activation functions to learn the complex nonlinear interactions between features from different sources.

[0192] 5) Multi-task output layer

[0193] Function: This is the final prediction layer of the model, which performs multiple independent prediction tasks simultaneously based on the fused features.

[0194] The output tasks include:

[0195] Phase classification: Predicts the most likely operating phase of the system and outputs the probability distribution of the phase.

[0196] Location regression: accurately regresses the relative position of the current data point within a period, outputting a continuous value from 0 to 1, providing more refined location information than simple periodic indexing.

[0197] Confidence estimation: Evaluates the model's confidence in the current phase prediction result and outputs a confidence score between 0 and 1.

[0198] The specific structure of the HybridCyclePhaseNet model is as follows: Figure 2 As shown:

[0199] By decoupling the signals, the model can perform specialized and refined modeling of periodic and residual components, thereby capturing key details that are easily overlooked by a single general-purpose model. The algorithm's multi-task output layer can simultaneously predict phase, precise intra-period position, and the confidence level of the prediction results. This provides downstream decision-making and control systems with far richer and more comprehensive state information than a single classification result, enhancing the overall situational awareness of the system.

[0200] III. Reinforcement Learning-Driven Adaptive Sampling System

[0201] The reinforcement learning-driven adaptive sampling system proposed in this invention transforms the rich, multi-dimensional state-aware information output by Module 2 (HybridCyclePhaseNet) into optimal data acquisition actions in real time. Through reinforcement learning, the agent can learn the long-term value of states and actions, thereby acquiring forward-looking decision-making capabilities. It can not only identify the current state but also predict the probability of certain state transition sequences leading to critical events, and adjust the sampling strategy in advance to avoid risks.

[0202] This invention rigorously models the dynamic sampling problem as a Markov Decision Process (MDP). The theoretical basis for this choice is that at any given time, the optimal sampling decision depends only on the current state of the system, without needing to trace the entire historical state sequence, which is entirely consistent with the Markov property.

[0203] To solve this MDP, this reinforcement learning-driven adaptive sampling system employs the Deep Q-Network (DQN) algorithm from Deep Reinforcement Learning (DRL). Given the high dimensionality and continuous variable nature of the state space constructed in this invention, traditional tabular Q-learning methods are no longer applicable. Therefore, it is necessary to leverage the powerful fitting capabilities of deep neural networks to approximate the action value function Q(s,a), thereby finding the optimal policy within the complex decision space.

[0204] 3.1 Environment Modeling: State-Space Design

[0205] The design of the state space is one of the core features of this invention. To enable the agent to make accurate decisions, this invention innovatively designs a state vector St containing nine dimensions of features. This vector organically combines the output of the HybridCyclePhaseNet model with key elements such as periodic information and sampling history, providing an unprecedented comprehensive information foundation for intelligent decision-making.

[0206] This 9-dimensional state vector is not simply a collection of inputs; it extracts and transforms the massive amounts of raw information (such as various probabilities and values) generated by the second-level perception layer into a structured format most favorable for decision-making. It clearly answers four key questions for the agent during decision-making:

[0207] Where is the current data point located in the cycle? (Periodic position information), What is happening in the system? (Core operating phase), How confident is the system in its judgment? (Model self-confidence), How stable is the system at present? (Volatility and sampling interval).

[0208] This transformation from raw perception to a structured "situational awareness" vector enables a relatively simple DQN network architecture to learn efficient and robust control policies. Its superiority over traditional rule-based policies that only use partial information fully validates the advanced nature of this design. The specific composition of the state vector St is as follows:

[0209] (1) Feature dimensions 1-2 are periodic positions

[0210] Calculate based on the current time step t and the known period length T:

[0211]

[0212] Its role and contribution in decision-making: It provides a continuous and unambiguous representation of the system's position within a cycle, avoiding the jump problem caused by linear indexing at the end of the cycle. It provides the agent with prior knowledge about the expected behavior at the current position.

[0213] (2) Feature dimensions 3-6 represent the current phase.

[0214] One-hot encoding of predicted phase If the predicted phase is i, i represents the index corresponding to the specific phase category predicted. If "peak stability" is the first phase, then i is 1.

[0215] Its role and contribution in decision-making: It provides the most crucial state information, directly informing the intelligent agent system of its current operating mode (such as "peak stability", "rising phase", etc.).

[0216] (3) Feature dimension 7 is the detection confidence s conf :

[0217]

[0218] Where N represents the number of most recent predictions used to calculate the average, and i represents the time step index in the summation operation, from t-N+1 to t.

[0219] confidence i This represents the confidence level of the HybridCyclePhaseNet model in predicting the phase at time step i.

[0220] Its role and contribution in decision-making: It serves as a metacognitive signal, informing the agent of the reliability of the current phase judgment. Low confidence itself is a critical state, indicating the need for increased vigilance and sampling density.

[0221] (4) Characteristic Dimension 8: System Volatility s vol

[0222]

[0223] volatility i This represents the system fluctuation intensity estimated by the HybridCyclePhaseNet model at time step i.

[0224] N represents the number of most recent predictions (or the number of samples) used to calculate the average; t represents the current time step; i represents the time step index in the summation operation, from t-N+1 to t.

[0225] volatility i This represents the intensity of system fluctuations estimated by the HybridCyclePhaseNet model at time step i. This value is typically between 0 and 1.

[0226] Its role and contribution in decision-making: Informing the agent about the system's stability. Even in a "stable" phase, high volatility can be a precursor to state changes or anomalies, providing a basis for increasing sampling.

[0227] (5) Feature dimension 9: time since last sampling s time :

[0228]

[0229] t lastsample This indicates the time step of the last sampling action.

[0230] t represents the current time step.

[0231] Role and contribution in decision-making: The higher the value, the longer it has been without acquiring new data, and the greater the risk of missing key events, thus incentivizing the agent to sample periodically even in a stable state.

[0232] 3.2 Definition of Action Space

[0233] To achieve clear and efficient decision-making, this invention defines the action space as a discrete binary set A = {0, 1}, where:

[0234] a=0: Implement a low sampling rate strategy. This is a resource-saving action, for example, data is collected only once every 2 time steps.

[0235] a=1: Execute a high sampling rate strategy. This is an information-gathering action, such as collecting data at each time step.

[0236] The sampling decision policy π(s) is learned through a deep Q-network, with the goal of maximizing long-term cumulative reward.

[0237] π(s)=argmax a Q(s,a).

[0238] For any given state s, the optimal policy π(s) is chosen by finding the action a that produces the highest Q-value. Essentially, DQN learns Q-values, and the optimal policy is derived directly from these learned Q-values, choosing the action that brings the maximum expected future reward.

[0239] 3.3 Reward Function Design

[0240] The action space defines what the agent "can do," while the reward function defines the "good or bad" evaluation criteria for the agent's different actions. By selecting actions in the defined action space and receiving feedback from the reward function, the agent learns how to maximize cumulative rewards and ultimately forms an optimal sampling strategy.

[0241] 's' refers to the current state of the environment, and 'a' refers to the action taken. In the policy definition, 's' is the basis for the decision, and 'a' is the result of the decision; in reward calculation, 's' and 'a' are the environmental conditions and agent behavior that lead to the reward.

[0242] Reward function R(s) t a t The design quantifies the complex requirements and business logic of industrial scenarios into scalar signals that the agent can learn. This function aims to guide the agent to find the optimal balance between the two conflicting goals of "ensuring no loss of critical information" and "maximizing resource utilization".

[0243] First, the system defines what constitutes a "critical moment." A state St is considered critical if it meets one of the following conditions: the system is in the "peak stability" phase. Or the prediction confidence of the perception model is lower than a preset threshold (s) conf <0.7). This logic innovatively combines physical state risk with model uncertainty risk, forming a comprehensive assessment of system risk.

[0244] Four-quadrant reward and punishment structure: Based on the decision at critical moments, the reward function is designed with the following structure:

[0245] R = +1.0: At a critical moment, a high sampling rate (a = 1) was correctly chosen. This is the highest reward for effective alertness.

[0246] R = +0.8: A low sampling rate (a = 0) was correctly chosen during non-critical times. This is a significant reward for high efficiency and cost-effectiveness.

[0247] R = -0.8: At a critical moment, a low sampling rate (a = 0) was mistakenly chosen. This is the most severe penalty for missed information.

[0248] R = -0.5: A high sampling rate (a = 1) was incorrectly chosen during a non-critical period. This is a relatively mild penalty for wasting resources.

[0249] This asymmetric design of reward and penalty values ​​(|-0.8| > |-0.5|) mathematically encodes a "reliability-first" industrial safety criterion. This means that in critical industrial systems, the potential losses from missing a critical event (such as equipment damage or production interruption) far outweigh the cost of consuming additional computing resources. In this way, the reward function transforms high-level, qualitative industrial demands into a quantifiable, learnable signal, directly shaping the agent's behavior and naturally influencing it to construct a safe and reliable decision-making strategy.

[0250] 3.4 Algorithm Implementation and Architecture

[0251] This invention employs a feedforward fully connected neural network to approximate the Q-function, and its specific architecture is designed to balance expressive power and computational efficiency:

[0252] Input layer: Contains 9 neurons, which perfectly match the dimensions of the 9-dimensional state vector St.

[0253] Hidden layers: consist of three fully connected (Linear) layers with 128, 64, and 32 neurons respectively;

[0254] Activation function: A modified linear unit (ReLU) is used as the activation function after each hidden layer to introduce nonlinearity and enhance the expressive power of the model.

[0255] Regularization: A Dropout layer (dropout probability p = 0.1) is added after the first two hidden layers to effectively prevent the model from overfitting during training.

[0256] Output layer: Contains 2 neurons, which output the Q values ​​that can be obtained by performing actions a=0 (low sampling rate) and a=1 (high sampling rate) in the current state, respectively.

[0257] 3.5 Training Protocol and Strategy Optimization

[0258] This invention employs an advanced DQN training algorithm that incorporates key technologies such as experience replay and target network to ensure the stability and efficiency of the learning process.

[0259] Network initialization: At the start of training, the main network Q is initialized. main (s; θ) and the target network Q target (s;θ - And make the network parameters of the two identical.

[0260] Exploration and Exploitation: An ε-greedy strategy is employed to balance exploration and exploitation. In the early stages of training, the agent randomly selects actions with high probability ε (initial value ε = 1.0) to explore the environment. As training progresses, the ε value decays at a fixed rate (ε). decay =0.995) gradually decreases until a minimum value (ε) is reached. min =0.01), which allows the agent to smoothly transition from exploration to utilizing the learned optimal strategy.

[0261] Experience replay: Experience tuples (s) generated by the interaction between the agent and the environment. t ,a t ,r t ,s t+1 ,done t The data is stored in a fixed-capacity (e.g., 10,000) experience replay buffer. During training, small batches of data are randomly drawn from the buffer for learning. This breaks the temporal correlation between experiences and greatly improves the stability of training.

[0262] Furthermore, the present invention can also employ Prioritized Experience Replay technology based on the sampling of the temporal difference error (TD-error) magnitude, so that unexpected and more informative experiences are learned more frequently, thereby accelerating convergence.

[0263] Learning steps: For each sampled experience, first calculate the target Q value y. t This value consists of the immediate reward rt and an estimate of the maximum future return, and is calculated using the following formula: Here, γ is a discount factor used to weigh current and future rewards.

[0264] Loss function and optimization: The target Q-value yt is compared with the Q-value predicted by the main network. main (s t a t The mean square error between θ and θ is used as the loss function: The gradient of the loss function with respect to the network parameters θ is calculated using the backpropagation algorithm, and the parameters are updated using an optimizer (such as Adam) to minimize the loss.

[0265] Target network update: To provide a stable learning target, the target network Q... target The parameter θ - It is not updated synchronously, but only at fixed intervals τ=100 from the main network Q. main The parameters are copied completely once.

[0266] The flow of the reinforcement learning-driven adaptive sampling algorithm in this invention is as follows: Figure 3 As shown, the overall system design and environmental modeling of this adaptive sampling algorithm construct a refined 9-dimensional state vector from the phase, confidence, and volatility information output by the preceding HybridCyclePhaseNet model, providing the agent with comprehensive information needed for decision-making; the reward mechanism is used to guide the agent to find the optimal balance between the two objectives of "ensuring that key information is not lost" and "maximizing resource utilization".

[0267] Example 2

[0268] Intelligent monitoring of power grid load

[0269] This embodiment selects three data columns with similar diurnal periodic patterns from the widely used Electricity dataset in the time series field: sensor_212, sensor_227, and sensor_252, to form a multivariate time series as the experimental object. This dataset has obvious 24-hour periodicity, making it an ideal choice for verifying the present invention.

[0270] This embodiment will provide a complete, step-by-step operating procedure to guide you on how to set up and use this system to monitor power consumption.

[0271] System initial setup and configuration

[0272] Objective: To initialize the system in order to monitor grid load data from three specific sensors.

[0273] Step 1: Connect to the data source

[0274] In the system's "Data Input" interface, select and connect the data streams from the following three sensors: sensor_212, sensor_227, and sensor_252. These sensors provide the raw power usage data that the system will analyze.

[0275] Step 2: Set core operating parameters

[0276] Enter the values ​​listed in the table below in the configuration panel. These settings are determined based on the specific pattern of this power data and have been optimized for best performance.

[0277]

[0278]

[0279] Properly setting initial parameters is fundamental to all subsequent high-precision operations of the system. These parameters are not arbitrarily set, but are interrelated and derived through in-depth analysis of data characteristics. For example, the "cycle length" of 24 hours is determined by the natural daily rhythm of electricity usage. The "sequence length" of 96 data points (i.e., 4 cycles) is to provide the system with a sufficiently long historical background (a full 4 days of data) to enable it to make accurate predictions about future trends.

[0280] II. Understanding Real-Time Monitoring Dashboard

[0281] Objective: To help operators become familiar with the system's main screen and its key metrics.

[0282] The system's main dashboard clearly displays all key information, allowing operators to monitor equipment status in real time.

[0283] Main data curve: The chart in the center of the screen displays the grid load curve from the sensors in real time.

[0284] Stage Indicator: A clear, color-coded text box displaying the current operational stage. Based on the system's data analysis, it will display one of four states: Stable Low (green), Rising Stage (red), Stable Peak (orange), or Declining Stage (blue).

[0285] Confidence gauge: A dashboard ranging from 0% to 100%, displaying how "confident" the system is of its current assessment. During normal operation, this gauge should remain at a very high level. A sudden drop in confidence is an important signal that requires operator attention.

[0286] System stability gauge: A simple indicator that reflects the "volatility" or "stability" of the input data. This indicator is calculated by the system's internal residual analysis module. A lower reading indicates stable system operation; a higher reading indicates unusual volatility.

[0287] Sampling rate indicator: An indicator light that shows whether the system is currently in low-frequency (normal) sampling mode or high-frequency (critical) sampling mode.

[0288] Model and training configuration:

[0289] Hardware environment: GPU: 2x NVIDIA GeForce RTX 3090, each with 24GB of video memory, CPU: Intel(R) Xeon(R) Platinum 8124M CPU@3.00GHz, RAM: 125GB.

[0290] Software environment: Python 3.8, PyTorch 1.7.1, CUDA 11.1, Scikit-learn 1.2.2.

[0291] Automated label generation: The PatternBasedPhaseAnalyzer class is used to independently learn and perform preliminary phase pattern detection on the three columns of data. Subsequently, through a majority voting mechanism, the three candidate labels at each time point are merged to generate a unified and robust final phase label.

[0292] Model building: Instantiate the HybridCyclePhaseNet model with a sequence length seq_len of 96 (4 cycles) to provide sufficient historical information.

[0293] Training Procedure: The dataset with generated labels is divided into training and validation sets in an 80 / 20 ratio. Training is performed using the Adam optimizer and a multi-task loss function L0. total =0.3*L phase +0.1*L position +0.1*L trend Perform backpropagation. Apply an early stopping strategy during training to prevent overfitting.

[0294] Training the Adaptive Sampling System: After training HybridCyclePhaseNet, it is used as the core component to build and train the AdaptiveSamplingSystem. The DQN agent optimizes its sampling decision strategy through approximately 800 episodes of iterative learning.

[0295] Performance evaluation and results analysis:

[0296] To fully verify the effectiveness and advancement of the method proposed in this invention, a series of rigorous experiments and comparative analyses were conducted on three aspects: unsupervised label generation quality, phase recognition model performance, and reinforcement learning adaptive sampling strategy.

[0297] (1) Unsupervised label generation and data visualization

[0298] High-quality training data is fundamental to the success of deep learning models. This invention first uses an unsupervised method based on pattern learning to automatically generate phase labels for the raw data.

[0299] Figure 4 The results of the voting mechanism were demonstrated when processing data from three columns (three sensors) in the Electricity dataset.

[0300] from Figure 4 It can be seen from this:

[0301] High consistency: The consistency between the independent analysis results of each column of data and the unified label generated by the final vote exceeded 92%, which proves the strong correlation between different data sources and the effectiveness of the voting mechanism.

[0302] High consensus: Among all 26,229 time points, as many as 77.4% of the samples (20,307) obtained voting results with completely consistent data across the three columns, indicating that in most cases, the judgment of the system state is clear and unambiguous.

[0303] Robust label generation: Through this voting process, the system ultimately generates a unified and reliable set of phase distribution labels for multivariate time series data, providing a high-quality data foundation for the training of subsequent deep learning models.

[0304] like Figure 5 As shown, the upper subplot overlays automatically generated four phase labels (rising, peak, falling, and trough) as scatter dots of different colors onto the standardized load curve. It can be seen that the label assignment closely matches the actual data trend: in the trough and peak regions of the load, the system stably labels them as "trough stable" and "peak stable"; in regions of rapid load change, it accurately labels them as "rising phase" and "falling phase." The lower subplot clearly shows the logical sequence of phase changes over time. This result demonstrates that the unsupervised label generation module of this invention can effectively, accurately, and logically consistently empower large-scale time series data, providing a high-quality foundation for subsequent model training.

[0305] (2) Performance evaluation of phase recognition model

[0306] Training process analysis

[0307] The training process of the HybridCyclePhaseNet model is smooth and efficient, such as Figure 6 As shown.

[0308] from Figure 6 It can be seen from this:

[0309] Loss convergence: The loss curve on the left shows that the loss on both the training and validation sets drops rapidly and converges to a very low level with a small difference between them, indicating that the model is highly efficient in learning and has not overfitted.

[0310] Accuracy and stability: The phase detection accuracy on the right shows that the model quickly reached a high level of accuracy (over 94%) in the early stages of training and has remained stable, demonstrating the effectiveness of the model architecture and the stability of learning.

[0311] Example 3

[0312] Performance comparison with baseline methods

[0313] To further verify the advancement of the method of this invention, it was compared with the advanced time series segmentation model PrecTime proposed in 2023.

[0314] The HybridCyclePhaseNet model proposed in this invention outperforms the state-of-the-art PrecTime model (93.88%) with an accuracy of 95.2%. This is attributed to the unique signal decomposition architecture of this invention, which is better at understanding and processing industrial data with physical periodicity than general sequence models.

[0315] Based on accurate phase recognition, this invention further constructs an intelligent sampling system through reinforcement learning.

[0316] The training process of the intelligent agent is as follows Figures 7-9 As shown in the figure, this diagram illustrates the learning progress of the DQN agent over 800 training rounds:

[0317] Rewards: such as Figure 7 As shown, the agent's average round reward (red smooth curve) steadily increases, indicating that it has successfully learned a decision-making strategy that yields higher rewards.

[0318] Exploration Rate: such as Figure 8 As shown, the Epsilon value smoothly decreases from 1.0 to 0.01, which conforms to the classic training paradigm of moving from "exploration" to "exploitation".

[0319] Q-value: such as Figure 9 As shown, the steady increase in the average Q value indicates that the agent is becoming more optimistic and accurate in assessing the value of its behavior.

[0320] Performance comparison test

[0321] The trained RL agent is compared with a "rule-based" policy. The "rule-based" policy is a pre-defined, non-learning decision-making method that follows a pre-defined logic: high-frequency sampling is only used when the system is in a "peak-stable" phase or when the model prediction confidence is below 0.7, and a series of if-then logics determine whether to use a high or low sampling rate.

[0322] In-depth analysis of the two strategies reveals that the RL agent trained in this invention demonstrates significant advantages over rule-based hard-coded logic in terms of decision-making intelligence, system reliability, and cost-effectiveness trade-offs.

[0323] The fundamental reason for the performance difference between the two strategies lies in the essential difference in their decision-making mechanisms. Rule-based strategies have a static and reactive decision-making mechanism. This mechanism judges based on preset, fixed thresholds, and once the current state is determined to be low-risk, it enters a fixed, low-frequency sampling period. This inevitably leads to periodic monitoring blind spots. If a critical event occurs within this blind spot, the system is inactive and unable to execute its judgment logic, resulting in the permanent loss of critical information.

[0324] In contrast, the RL agent of this invention constructs a dynamic and forward-looking decision-making model. Through deep learning of a high-dimensional state space, it gains the ability to predictive risk assessment. This model does not passively respond to deteriorating states that have already occurred, but rather predicts the probability of future critical events based on subtle evolutions in temporal characteristics. Based on this prediction, the agent can proactively adjust its observation frequency, increasing monitoring intensity in advance when potential risks rise, thereby effectively avoiding monitoring blind spots and ensuring timely and forward-looking decision-making.

[0325] The overall phase detection accuracy of the model in this invention reached 95.2%, and the average prediction confidence was as high as 0.996, which proves the reliability of its prediction results.

[0326] Figures 10-12 The diagram provides deeper performance insights:

[0327] Figure 10 The confusion matrix shown has values ​​on the diagonal (e.g., 2461 samples were correctly classified during the peak stabilization phase) that are much higher than those in other positions, indicating that the model has a very strong ability to distinguish between the four phase types.

[0328] Figure 11 The phase prediction confidence distribution shown is as follows: the confidence of phase detection for almost all samples is above 95%, which reflects that the model is very confident in phase detection.

[0329] Figure 12 The detection accuracy of each phase is shown: the classification accuracy of all phases is very high, especially the accuracy of the "peak stability" and "valley stability" phases, which exceeds 97%.

[0330] like Figure 13 As shown, a continuous scenario is used to demonstrate the real-time decision-making process of the entire adaptive sampling system.

[0331] The agent uses a low sampling rate (blue dots) most of the time, but in areas where the system determines a "critical event" (peak stabilization phase or low confidence), it decisively switches to a high sampling rate (red dots). This demonstration proves the completeness and effectiveness of the closed-loop system constructed in this invention, from high-precision state perception to intelligent decision-making and adjustment.

[0332] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for adjusting dynamic sensing and acquisition strategies for industrial data, characterized in that, We construct an unsupervised label generation module based on pattern learning, a hybrid detection network that integrates multiple features, and an adaptive sampling system driven by reinforcement learning. The pattern learning-based unsupervised label generation module automatically learns from historical data and combines multi-dimensional features such as numerical values, gradients, and positions within periods to generate phase labels in batches for unlabeled data. The multi-feature fusion hybrid detection network is based on a signal decomposition algorithm, which decouples the input signal into periodic components and residual components. It uses a parallel dual-branch structure to extract the deep features of the periodic components and residual components respectively, and predicts the system state through a feature fusion network. The reinforcement learning-driven adaptive sampling system is based on deep reinforcement learning technology and dynamically adjusts the sampling frequency and acquisition strategy according to the real-time state of the system.

2. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 1, characterized in that, The unsupervised label generation module based on pattern learning extracts multiple complete periodic samples from historical operational data, performs time series periodicity detection, and determines the period length of the system. Four operating phases are predefined, and a template containing three types of feature constraints is constructed for each phase; Phase labeling is performed on the time series data to be labeled, generating phase labels.

3. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 2, characterized in that, The four predefined operating phases include: rising phase, peak stable phase, falling phase, and trough stable phase; Three types of feature constraints are constructed for each phase: numerical range constraints, gradient range constraints, and location prior constraints. By analyzing historical average patterns, the typical location points of each phase within the period are statistically determined, and a priority mechanism is introduced to assign high priority to peak stable phases and trough stable phases.

4. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 3, characterized in that, For the time series data to be labeled, the system uses a sliding window technique to perform phase labeling point by point, and fuses numerical matching scores S. value (t), gradient matching score S gradient (t) and the prior score of the periodic position S position (t), calculate the matching score between the data point at time t and each phase template. i (t): Score i (t)=w1·S value (t)+w2·S gradient (t)+w3·S position (t); Where w1 = 0.4, w2 = 0.4, and w3 = 0.2 are the weights of numerical matching, gradient matching, and location prior, respectively; the scores of each component are calculated using the following formula: Numerical matching score: Gradient matching score: Prior score of position:

5. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 4, characterized in that, A multi-stage global optimization mechanism is introduced; for regions where the value is higher than the first threshold or lower than the second threshold and the gradient is less than the third threshold, a label of peak stable phase or trough stable phase is assigned; a sliding window is used to smooth the initially generated label sequence; the entire label sequence is traversed to check for unreasonable phase jumps, and if they exist, the phase of the previous point is retained based on the confidence score.

6. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 1, characterized in that, The multi-feature fusion hybrid detection network adopts a deep network architecture based on signal decomposition, modeling the input industrial time series signal X(t) as a linear superposition of three parts: X(t)=X cyclic (t)+X residual (t)+∈(t), Among them, X cyclic (t) represents the periodic component in the signal; X residual (t) represents the residual component remaining after removing the periodic component, and ∈(t) represents random noise; Construct a two-branch network structure to decouple the input signal into a periodic component X. cyclic (t) and residual component X residual (t), using the periodic component processing branch and the residual component processing branch, deep feature extraction is performed on the periodic component and the residual component respectively.

7. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 6, characterized in that, The learnable periodic extractor uses a learnable periodic template matrix W∈R T×D The core periodic component is extracted from the original input signal, where T is a predetermined signal period length and D is the feature dimension of the data; for any time t in the input sequence, its corresponding periodic component X cyclic (t) is calculated by indexing the periodic template Q matrix: X cyclic (t)=Q[tmodT,:] The periodic feature analyzer uses a sinusoidal position encoding matrix. The purified periodic component X was extracted. cyclic (t) performs in-depth analysis, d model It is the hidden layer dimension of the model; The periodic component X is encoded using an encoder. cyclic (t) is encoded to obtain the periodic feature F. cyc Based on the current data point's index within the cycle (cycle_index), the corresponding position vector F is retrieved from the sinusoidal position encoding matrix P. pos The two feature vectors are concatenated to form a joint feature vector F. joint The data is fed into the feedforward network to initially determine the phase of the cycle.

8. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 7, characterized in that, The residual component processing branch removes periodic components from the original signal to obtain the residual signal sequence. By analyzing the current value of the residual signal sequence and its changes before and after, it determines whether the current trend of the system is rising, stable, or falling. Combining the residual signal sequence characteristics and trend characteristics, it quantifies the current volatility of the system and outputs a normalized volatility intensity value.

9. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 8, characterized in that, The deep features from the periodic branch, the deep features from the residual branch, and the trend classification probability and normalized fluctuation intensity value output by the residual branch are concatenated to generate a fused feature vector. This fused feature vector is then fed into a deep fusion network. The final feature vector after deep fusion is distributed to three parallel final phase classifiers, intra-period position regressors, and confidence estimators, simultaneously completing three independent prediction tasks: The final phase classifier predicts the current phase prediction result of the system based on all the fused final feature vectors; The periodic position regressor regresses the relative position of the current data point within the period and outputs a continuous value from 0 to 1. Confidence estimator: Estimates the confidence level of the current phase prediction result and outputs a value between 0 and 1.

10. The method for adjusting the dynamic sensing and acquisition strategy of industrial data according to claim 9, characterized in that, The reinforcement learning-driven adaptive sampling system constructs a 9-dimensional state vector from the current phase, detection confidence, system volatility, and time since the last sampling, based on the output of a multi-feature fusion hybrid detection network. The sampling decision policy π(s) is learned through a deep Q-network to maximize long-term cumulative reward. π(s)=argmax a Q(s,a); The trained master Q network Q main As the optimal strategy π * (s)=argmax a Q main (s,a;θ * This enables intelligent adaptive sampling decision-making, where π(s) refers to the sampling decision-making strategy learned by the system. * (s) refers to the optimal strategy.

Images