Signal processing method based on thermal flow sensor

Abstract

The invention discloses a signal processing method based on a thermal flow sensor, and relates to the technical field of digital signal processing. The signal processing method based on the thermal flow sensor comprises the following steps: S1, collecting multiple types of flow and context data sets, and carrying out unified time base alignment; s2, sliding window segmented modeling and multi-scale statistical feature extraction are carried out; s3, constructing a thermal flow sequence rule driving and data quality scoring model; s4, performing weighted filtering and resampling on the thermal flow sequence; and S5, constructing a hierarchical archiving and data pedigree structure, and performing model evolution management. According to the method, the data quality controllability and the result stability of the traffic signal in the whole process are effectively improved, and the problems that an existing method lacks unified data quality marking and confidence coefficient output, and differential processing cannot be carried out on abnormal samples are solved.

Description

Technical Field

[0001] This invention relates to the field of digital signal processing technology, specifically to a signal processing method based on a thermal flow sensor. Background Technology

[0002] With the development of thermal flow sensors in process condition monitoring, environmental emission monitoring, and metering and accounting scenarios, enabling multi-channel, high-frequency, and long-term online data acquisition, the demand for traceability of acquired data, precise identification of abnormal samples, and quantitative evaluation of the effects of different signal processing versions continues to increase. Existing methods only perform verification at the communication level and use fixed thresholds or simple filtering to uniformly process flow sequences, treating all data samples that pass verification the same. They neither construct sample-level data quality scores by combining differential pressure, temperature, power supply status, and preheating status, nor do they have explicit confidence labels and data spectrum management structures. This leads to low-quality samples being directly used for integral statistics and model training, which easily introduces systematic biases and anomaly misjudgments, making it difficult to support continuous optimization and objective comparison of signal processing schemes.

[0003] For example, invention patent CN114637960B discloses a time-frequency domain digital turbine flow sensor signal processing method, including: analog turbine flow sensors directly shape sensor signals into pulse signals for output and measurement, exhibiting good real-time performance. However, when operating in environments with strong noise interference, the turbine flow sensor output signal is prone to distortion, leading to inaccurate pulse output and reduced precision. Digital turbine flow sensors can acquire sensor signals in real time, perform spectrum analysis to calculate the flow frequency, and have good anti-interference performance, but require high frequency resolution to ensure measurement accuracy. When the flow meter diameter is small and the measured flow rate is large, the signal frequency is high, resulting in a large amount of data involved in the calculation, long computation time, and poor real-time performance. Therefore, this patent proposes a new digital turbine flow sensor signal processing method, employing a time-frequency domain combined signal processing approach to improve anti-interference performance, ensure measurement accuracy, reduce computational load, and guarantee real-time performance.

[0004] In existing technologies, as long as the communication CRC check passes, all traffic samples are considered equally reliable and directly used for accumulation, statistics, and decision-making. There is a lack of a mechanism to fine-grainedly evaluate sample reliability based on the acquisition context. When short-term power supply drops, bus congestion, unstable sensor self-heating, recent range switching, or internal self-test flags indicating non-steady-state conditions occur, although the output values ​​still exist, their statistical reliability is far lower than during normal operation. Directly including these values ​​in accumulation or model training will introduce implicit biases.

[0005] Therefore, in order to address the above problems, there is an urgent need for a signal processing method based on thermal flow sensors. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a signal processing method based on a thermal flow sensor, which solves the problem that existing methods lack unified data quality labels and confidence outputs, and are unable to differentiate abnormal samples.

[0008] Technical solution

[0009] To achieve the above objectives, this invention provides the following technical solution: a signal processing method based on a thermal flow sensor, comprising: S1, the sensor collaboratively acquires multiple types of flow and context datasets, and performs unified time base alignment and context labeling on the multiple types of flow and context datasets to form a thermal flow acquisition data frame; S2, the thermal flow acquisition data frame is subjected to basic filtering and outlier screening, and a quality assessment feature frame is constructed by performing sliding window segmentation modeling and multi-scale statistical feature extraction; S3, a thermal flow sequence rule-driven and data quality scoring model is constructed based on the quality assessment feature frame, the sample-level quality score is derived, and confidence labels and diagnostic information are output; S4, the thermal flow sequence is weighted and resampled based on the confidence labels, and multi-scale integration and statistical calculations are performed to output multi-granularity flow results and anomaly event records; S5, a hierarchical archiving and data hierarchy structure is constructed based on the multi-granularity flow results and quality event records to perform evolution management and quality playback diagnostic support for the thermal flow signal processing link.

[0010] Furthermore, the specific steps for sensor collaborative acquisition of multiple types of flow and context datasets, and the unified time base alignment and context labeling of these datasets to construct a thermal flow acquisition data frame are as follows: Multiple types of flow and context datasets are acquired through sensor collaborative acquisition. These datasets include: thermal flow sensor output dataset, differential pressure sensor dataset, environmental operating condition dataset, power supply monitoring dataset, communication link status dataset, equipment operating status dataset, and timestamp and sequence identifier dataset. Historical acquisition sequences already stored under the same installation scenario and configuration conditions are used as historical operating data. Acquisition strategy management and context classification labeling are performed on the acquired multiple types of flow and context datasets. Timing scheduling and unified triggering of various sensor interfaces are performed using the acquisition control unit. Multi-channel data is time-aligned using the local unified clock of the acquisition node to obtain the processed multiple types of flow and context datasets. The multiple types of flow and context datasets acquired in each sampling period are encapsulated using a unified data structure to construct a thermal flow acquisition data frame. Finally, the thermal flow acquisition data frame is segmented, scaled, normalized, and format standardized.

[0011] Furthermore, the specific steps for basic filtering and outlier screening of the thermal flow acquisition data frame are as follows: Input the thermal flow acquisition data frame, perform time consistency verification and multi-channel alignment on the global acquisition queue, mark and correct out-of-order, dropped frames and delayed sampling to obtain the preprocessed thermal flow acquisition data frame; Basic filtering and outlier screening include: For instantaneous flow values, differential pressure values ​​and continuous quantities of bridge voltage or drive current, use small window moving average, moving median filtering and first-order exponential smoothing lightweight filtering methods within the segment to suppress high-frequency quantization noise and isolated spikes while ensuring that the flow change trend and edge morphology are not significantly weakened, and generate smoothed values ​​and residual values; and form a preliminary estimate of the anomaly density.

[0012] Furthermore, the specific steps for constructing quality assessment feature frames through sliding window segmentation modeling and multi-scale statistical feature extraction are as follows: input the preprocessed thermal flow acquisition data frame, perform sliding window segmentation modeling on the time axis, extract multi-scale statistical features from the flow and context data within each window, and encode the context state quantity into a feature vector to form a feature sequence; add statistically unreliable labels to windows with a missing sample ratio greater than the missing threshold, and encode the labels into the feature vector; generate quality assessment feature frames for each window.

[0013] Furthermore, the specific steps for constructing a rule-driven and data quality scoring model for hot traffic sequences based on quality assessment feature frames are as follows: Input the quality assessment feature frames and historical operating data, construct a data quality scoring model for hot traffic sequences based on rule-driven principles and supplemented by lightweight statistics, and generate basic quality scores for each time window: perform cluster analysis and interval statistics on the historical distribution of traffic fluctuation, anomaly density, missing proportion, saturation proportion, power supply fluctuation amplitude, link error count, preheating flag, and range switching flag features, divide the feature space into normal intervals, suspicious intervals, and abnormal intervals, and assign basic quality scores and penalty weights to each interval; on top of the rule-driven basic quality scoring model, introduce a lightweight statistical model to finely distinguish samples near the rule boundaries, forming a hybrid quality scoring mechanism; use the quality assessment feature frames and historical label categories as training data, and employ lightweight models such as logistic regression, decision trees, or gradient boosting trees to learn the nonlinear relationship between each feature and the quality category.

[0014] Furthermore, the specific steps for deriving the sample-level quality score and outputting confidence labels and diagnostic information are as follows: Input the window-level quality score value and the corresponding quality assessment feature frame; calculate the sample-level quality score and confidence label for each time window and each sampling point within the window; and bind the quality score result and diagnostic information together to the hot flow data output; for each sliding window, based on the inherited window-level quality score value, perform fine-grained correction on the sample-level quality score of each sampling point within the window according to the abnormal distribution and context state; map the sample-level quality score value to numerical confidence labels and discrete quality status labels; generate diagnostic information based on the sample-level quality score result; output the dominant factors causing quality degradation in the form of cause codes along with the flow data; aggregate time periods where the window-level quality score is continuously below the low score threshold and the sample-level confidence is continuously at the low confidence threshold to form quality events, and record them as part of the quality time series; output the hot flow data frame with sample-level confidence labels, quality status labels, and diagnostic information.

[0015] Furthermore, the specific steps for weighted filtering and resampling of hot flow sequences based on confidence labels are as follows: Input hot flow data frames with sample-level confidence labels and quality status labels; perform confidence-weighted filtering and resampling on the hot flow sequences to reconstruct information-enhanced fine-grained flow waveforms that balance waveform authenticity and data quality on the time axis; perform confidence-weighted filtering on the hot flow sequences based on sample-level confidence labels to suppress mutations and noise introduced by low-quality samples, obtaining smooth flow waveforms while maintaining real physical changes; perform confidence-weighted sampling on the hot flow sequences based on confidence-weighted filtering to convert irregular sequences into a uniform time step, generating a well-structured fine-grained flow sequence with resampling confidence indicators; maintain the mapping relationship and weight allocation information between original sampling points and resampling points during the resampling process, record the original sample index set and weight vector corresponding to each resampling point as resampling traceability information, and output the confidence-weighted fine-grained flow waveform.

[0016] Further, the specific steps for performing multi-scale integration and statistical calculations to output multi-granularity flow results and abnormal event records are as follows: Based on the confidence-weighted fine-grained flow waveform, perform confidence-weighted integration and multi-scale statistical calculations on the thermal flow sequence to generate flow statistical results with overall confidence evaluation at each time scale; construct a multi-granularity flow result set carrying an overall confidence index. The multi-scale statistics include: dividing the time axis into continuous non-overlapping or slidingly overlapping statistical windows according to time granularity; calculating the flow mean, variance, maximum, minimum, peak, occurrence time, and first-order difference statistics within each window based on the confidence-weighted fine-grained flow waveform; introducing a quality-aware anomaly detection and labeling mechanism based on multi-scale statistics; using a joint criterion of numerical anomaly degree and confidence information to distinguish between physical flow events and measurement link anomalies, and outputting abnormal event records; outputting the multi-granularity flow result records and abnormal event records in a unified data format, while retaining the index relationship from multi-scale results to fine-grained samples.

[0017] Furthermore, the specific steps for constructing a hierarchical archiving and data hierarchy structure based on multi-granularity flow results and quality event records are as follows: Input fine-grained flow waveforms with confidence levels, multi-granularity flow results, and quality event records; perform hierarchical storage and indexing of flow data of different granularities and corresponding quality labels, diagnostic information, acquisition configuration, and algorithm version information; map the original acquisition sequence, preprocessing results, quality assessment results, statistical results, and event records to a unified data hierarchy structure; divide the data related to the acquisition link into multi-layer data views according to time, channel, and processing stage, and construct inter-layer mapping relationships through unified index fields to form a tree-like or graph-like data hierarchy structure; assign a unique processing hierarchy identifier to each continuous time interval of data when data is entered into the database and archived.

[0018] Furthermore, the specific steps for evolution management and quality playback diagnostic support of the hot traffic signal processing link are as follows: Based on the data spectrum structure, perform coexistence and difference analysis of the results of the old and new versions, without directly overwriting the historical results, but storing the new results as an additional layer of derived results; construct a quality playback and diagnostic support mechanism for operation and maintenance and analysis, and use multi-layer archived data and spectrum information to achieve multi-dimensional playback and comparison by time, version, and event; by statistically analyzing the quality scoring results and abnormal event results of different algorithm versions, compare the differences of each version in window-level quality scoring standard deviation, abnormal event detection recall rate and false alarm rate indicators, and form an evaluation report including indicator curves and summary tables.

[0019] Beneficial effects

[0020] The present invention has the following beneficial effects: (1) This invention performs time consistency verification, basic filtering and outlier screening on the collected data, so that spike pulses that violate physical continuity, segments that are close to the upper and lower limits of the range for a long time and short interruptions are identified and marked separately before entering the statistical analysis. This avoids the unreasonable impact of extreme abnormal samples on the subsequent window features and cumulative calculation, and improves the smoothness and physical rationality of the flow baseline curve.

[0021] (2) This invention maps window-level quality scores to sample-level quality scores and numerical confidence scores, and automatically generates high confidence, medium confidence and low confidence labels according to the confidence interval. At the same time, it outputs the dominant cause code that leads to quality decline, so that each traffic sampling point has clear data quality label and interpretable quality cause. The integral calculation, threshold judgment and model training can be weighted or eliminated according to the confidence, realizing differentiated processing of abnormal samples and low quality samples.

[0022] (3) In this invention, by using sample-level confidence as weight, weighted filtering, weighted interpolation and resampling are performed on the flow sequence, so that high-confidence samples contribute more to the reconstruction results and the influence of low-confidence samples in the reconstruction results is compressed. Under the premise of maintaining the overall trend of flow signal change and quality conservation relationship, the interference of the preheating stage, the range switching transition area and the power supply fluctuation period on the statistical results and trend analysis is effectively suppressed, and the stability and physical reliability of the reconstructed flow sequence are improved.

[0023] (4) This invention assigns algorithm version numbers and processing link identifiers to the quality scoring model, preprocessing strategy and confidence weighting algorithm, stores the original results and new version derived results in the data genealogy structure respectively, and supports comparing window-level quality score distribution, sample-level confidence distribution and abnormal event records by version on the same time segment, thereby realizing the coexistence and difference quantitative analysis of the results of the old and new versions.

[0024] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0025] Figure 1 This is a flowchart of a signal processing method based on a thermal flow sensor according to the present invention. Figure 2 This is a comparison chart of the basic filtering and outlier screening effects of this invention; Figure 3 This is a comparison chart of the confidence-weighted filtering effects of the present invention; Figure 4 This is a schematic diagram of the distribution of the feature intervals of the present invention; Figure 5 This is a schematic diagram of the version coexistence and quality playback process based on the data genealogy structure of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figures 1-5 This invention provides a technical solution: a signal processing method based on a thermal flow sensor, comprising: S1, the sensor collaboratively collects multiple types of flow and context datasets, and performs unified time base alignment and context labeling on the multiple types of flow and context datasets to form a thermal flow acquisition data frame; S2, the thermal flow acquisition data frame is subjected to basic filtering and outlier screening, and a quality assessment feature frame is constructed by performing sliding window segmentation modeling and multi-scale statistical feature extraction; S3, a thermal flow sequence rule-driven and data quality scoring model is constructed based on the quality assessment feature frame, the sample-level quality score is derived, and confidence labels and diagnostic information are output; S4, the thermal flow sequence is weighted and resampled based on the confidence labels, and multi-scale integration and statistical calculation are performed to output multi-granularity flow results and abnormal event records; S5, a hierarchical archiving and data hierarchy structure is constructed based on the multi-granularity flow results and quality event records to perform evolution management and quality playback diagnostic support for the thermal flow signal processing link.

[0028] Specifically, the steps for sensor collaborative acquisition of multiple types of flow and context datasets, and the unified time base alignment and context labeling of these datasets to construct a thermal flow acquisition data frame are as follows: Multiple types of flow and context datasets are acquired collaboratively by sensors. These datasets include: thermal flow sensor output dataset, differential pressure sensor dataset, environmental condition dataset, power supply monitoring dataset, communication link status dataset, equipment operating status dataset, and timestamp and sequence identifier dataset. The thermal flow sensor output dataset includes: instantaneous mass flow rate value, sensor internal temperature, bridge temperature, bridge voltage, and raw readings of drive current. These are basic measurements periodically read from the digital output interface by an integrated thermal flow sensor or micro-flow chip installed on the measured pipeline. The differential pressure sensor dataset includes: instantaneous differential pressure value, sensor internal temperature, and differential pressure zero-point offset self-check mark. This is synchronously acquired by a differential pressure sensor installed in the same pipeline or related channels at a sampling frequency that is the same as or an integer multiple of that of the thermal flow sensor. The environmental condition dataset includes: ambient temperature, ambient humidity, and safety... The temperature of the housing at the installation location is periodically collected by ambient temperature and humidity sensors and housing temperature probes installed inside and outside the equipment cavity; the power supply monitoring dataset includes: instantaneous power supply voltage value, power supply voltage change rate, and power supply voltage over-limit flag, which are electrical operating conditions synchronously recorded each time thermal flow data is collected by acquiring the power supply rail voltage in real time by the MCU built-in ADC or dedicated power monitoring chip; the communication link status dataset includes: CRC check result of this reading, number of retransmissions, timeout flag, and bus error count within the sampling period, which are obtained by statistically analyzing each data interaction process on the UART bus; the equipment operation status dataset includes: sensor power-on time count, preheating stage flag, range indicator, self-test result flag, and internal error code, which are summarized by reading the internal status registers, firmware status words, and external control logic flags of the thermal flow sensor and differential pressure sensor; the timestamp and sequence identifier dataset includes: local timestamp of the acquisition node, sampling sequence number, sensor channel number, and sampling configuration version number, which are timing and identification information uniformly assigned by the acquisition control unit after each multi-channel data reading.

[0029] Historical data collected under the same installation scenario and configuration conditions is considered historical operational data. This historical operational data includes: thermal flow time series, differential pressure time series, environmental condition change trajectories, power supply voltage fluctuation records, and historical communication error statistics for different time periods. The statistical scope of historical operational data is based on summarizing all online acquisition records that have been continuously running for at least several days, assuming the installation location, sensor model, and sampling configuration remain unchanged. These records are then uniformly converted to a time axis in seconds or milliseconds during archiving. The system manages acquisition strategies and classifies and labels contextual data for various types of traffic and contextual datasets. Based on the physical characteristics and update rates of different sensor channels, thermal flow sensor outputs and differential pressure sensor outputs are set as high-priority periodic acquisition channels, environmental conditions and power supply monitoring channels are set as medium-priority periodic acquisition channels, and device self-test status and range switching flags are set as event-triggered acquisition channels. Contextual status labels are added to samples in the acquisition buffer during the preheating, range switching transition, and power supply anomaly phases. Contextual status labels use a discrete enumeration encoding method, assigning a unique integer code to each status and providing a unified meaning in the system configuration table. This ensures that the calculation process only needs to process dimensionless integer codes when judging contextual status without directly parsing physical event descriptions. The system also performs statistical analysis on the communication link status dataset, aggregating CRC failures, retransmissions, and multiple timeouts into link anomaly segments and adding link risk flags to traffic samples within the corresponding time range.

[0030] The acquisition and control unit performs timing scheduling and unified triggering of various sensor interfaces. Using the unified local clock of the acquisition node as a reference, multi-channel data is time-aligned to obtain processed multi-type flow and context datasets. The unified local clock of the acquisition node is jointly provided by a temperature-compensated crystal oscillator and an upstream time synchronization service, with a time resolution of at least milliseconds. Long-term time deviations are controlled through periodic synchronization with the upstream time source, making the multi-channel sampling time error controllable within the target range. This ensures that thermal flow, differential pressure, environmental conditions, power supply monitoring, and equipment status channels obtain a unified acquisition timestamp within the same sampling period. By verifying the consistency between the sampling sequence number and the timestamp, cross-channel missed sampling, duplicate sampling, and delayed sampling are detected and corrected. In the event of a sampling period adjustment... When configuration changes occur, configuration change records are automatically inserted, clearly marking the boundaries between the previous and current configurations on the timeline. Multiple types of flow and context datasets collected within each sampling period are encapsulated using a unified data structure to construct a hot flow acquisition data frame. This hot flow acquisition data frame includes: flow acquisition device ID, sensor type identifier, sampling sequence number, sampling timestamp, instantaneous flow value, differential pressure value, sensor internal temperature, bridge voltage or drive current, ambient temperature, ambient humidity, power supply voltage, sampling configuration version number, range identifier, warm-up stage flag, range switching flag, self-test result flag, communication CRC result, retransmission count, and link error count window value. A quality flag field is reserved for recording sample quality scores and confidence calculation results. On the acquisition nodes, a global acquisition queue ordered by time and a local acquisition queue divided by sensor channel are maintained. For queues approaching their buffer limits, a strategy of prioritizing the retention of the latest data frames and data frames marked as critical contexts is adopted. Data frames are bucketed and archived according to sensor type and sampling period. The cache limit refers to the maximum queue length allocated to the acquisition queue on the acquisition node. Its value is determined by a combination of the available memory capacity of the acquisition node, the average byte length of data frames acquired through various hot traffic acquisition methods, and the target cache time window. The target cache time window is measured in seconds or minutes and is set according to the time span requirements for historical playback and reconstruction in different application scenarios.

[0031] The thermal flow acquisition data frames are then segmented, scaled, normalized, and formatted. Segment scaling normalization refers to, while maintaining the clear dimensions of physical quantities, first converting the flow rate, differential pressure, temperature, and voltage values ​​output by different sensor models to a unified physical unit range according to calibration coefficients, and then linearly or piecewise mapping the corresponding numerical ranges to a dimensionless range according to their respective nominal ranges or the minimum and maximum values ​​obtained from historical operating data. For channels with multiple ranges, mapping is performed within each range using the upper and lower bounds of that range, ensuring that the normalized features are comparable in numerical space and preserving the relative proportions of the original physical quantities. The flow rate, differential pressure, temperature, and voltage values ​​output by different sensor models are linearly or piecewise scaled according to their respective calibrated ranges and unit conversion rules, mapping to a unified range and eliminating scale differences caused by different devices and different range configurations. The timestamp field is uniformly represented using a precision continuous time axis, and the sampling sequence number is overflow protected and marked with rounds. Abstract sensor type identifiers and channel type codes are used to uniformly encapsulate data from different physical channels such as thermal flow, differential pressure, environment, and power supply into a consistent acquisition data frame format.

[0032] In this implementation scheme, by collecting multiple types of data and completing time alignment, context state enumeration and labeling, cache management, and segmented scaling normalization under a unified clock reference, a hot-type flow acquisition data frame with consistent structure and clear dimensions is constructed. This enables subsequent processing steps to simultaneously access the complete combination of physical quantities and context information at any sampling moment. It significantly reduces the impact of cross-channel timing misalignment and sampling loss on flow analysis results, providing a stable and traceable input foundation for data quality assessment, confidence calculation, and anomaly diagnosis. It achieves an overall improvement in the multi-source acquisition link in terms of time consistency, context visibility, and numerical scale comparability.

[0033] Specifically, the steps for basic filtering and outlier screening of thermal flow acquisition data frames are as follows: Input the thermal flow acquisition data frame, perform time consistency verification and multi-channel alignment on the global acquisition queue, and mark and correct out-of-order, dropped, and delayed sampling to obtain the preprocessed thermal flow acquisition data frame. Time consistency verification includes: sorting and checking the thermal flow acquisition data frame using the sampling timestamp and sampling sequence number as keys; detecting records with backward, jump, or duplicate sampling sequence numbers; marking detected abnormal records as time-series anomalies; reordering samples with slight out-of-order issues but whose timestamps are still within acceptable error range using a stable timestamp-based sorting strategy to ensure that the acquisition records maintain a monotonically increasing trend on a unified time axis; marking instances where the time interval between adjacent samples exceeds the upper limit as time gap events and recording the missing duration; the upper limit of the time interval is determined by historical data analysis. The data includes statistical analysis of normal and abnormal time intervals, and cross-validation is used to select the threshold value that minimizes the combined index of false positive and false negative rates from the candidate threshold set. Multi-channel alignment includes: aligning instantaneous flow rate, differential pressure, sensor internal temperature, bridge voltage or drive current, ambient temperature, ambient humidity, power supply voltage, and communication link status fields frame by frame according to a unified sampling timestamp; recording channel missing due to bus retries or temporary timeouts with channel missing markers; when the difference between adjacent timestamps is less than the interpolation threshold, using neighboring sample values ​​or linear interpolation to locally complete a single channel; when the time interval exceeds the interpolation threshold or the number of consecutive missing samples exceeds the missing proportion, keeping the missing markers unchanged to prevent misjudging long periods of inactivity as a steady, gradual change process; segmenting segments with time axis breaks after multi-channel alignment to generate continuous subsequences with segment numbers.

[0034] Basic filtering and outlier screening include: for instantaneous flow rate, differential pressure, and continuous quantities such as bridge voltage or drive current, small-window moving average, moving median filtering, and first-order exponential smoothing lightweight filtering methods are used within the segment to suppress high-frequency quantization noise and isolated spikes while ensuring that the flow change trend and edge morphology are not significantly weakened, generating smoothed values ​​and residual values; samples with a difference exceeding a reasonable threshold before and after filtering are marked with a numerical jump mark, and samples near the upper or lower limit of the sensor range and continuously for multiple sampling periods are marked with a saturation mark; the upper and lower limits of the sensor range are taken from the maximum and minimum flow rates within the nominal range of the thermal flow sensor, and are determined by statistical analysis of historical operating data to identify the starting point of the output... The flow range with clipped peaks, waveform compression, or significant nonlinear distortion is considered, and the impact of different intervals on saturation identification results is compared among a set of candidate intervals. The interval with the optimal comprehensive index of false positive rate and false negative rate is selected as the interval near the upper limit and lower limit of the flow range. Continuous multi-sampling period is defined as continuously satisfying the saturation interval condition in continuous sampling points, and the number of continuous sampling points is not less than the lower limit of sampling points. The lower limit of sampling points is determined by setting a set of candidate sampling point numbers on historical operating data and comparing the detection rate and false positive rate of saturation markers under different candidate numbers. The sampling point number with the optimal comprehensive index is selected as the lower limit of sampling points to ensure that the saturation marker is insensitive to single-point spikes but has a high identification ability for truly continuous saturation periods. Combining the power supply voltage overrun flag, self-test error flag, and CRC failure record in the communication link status, the corresponding sampling points are uniformly marked as original abnormal samples. The proportion of abnormal samples in each time segment and sliding window is statistically analyzed to form a preliminary estimate of the abnormal density. For time areas with short-term power supply drops or sudden increases in bus errors, additional link risk flags are added.

[0035] In this implementation scheme, through the aforementioned time consistency verification, multi-channel alignment, and basic anomaly screening, the original thermal flow acquisition data frames are converted into data records with complete structure, unified time axis, and those marked as self-test failures and serious fault states, providing clean and reliable input for feature construction and quality scoring.

[0036] Specifically, the steps for constructing quality assessment feature frames through sliding window segmentation modeling and multi-scale statistical feature extraction are as follows: Input the preprocessed thermal flow acquisition data frame; perform sliding window segmentation modeling on the time axis; extract multi-scale statistical features from the flow and context data within each window; and encode the context state variables into feature vectors to form a feature sequence. Sliding window segmentation modeling includes: constructing a fixed-length or adaptive-length sliding time window based on a unified sampling timestamp for each continuous time segment. The fixed-length window uses either the time width or the number of sampling points, while the adaptive-length window dynamically adjusts its size based on the flow rate of change and anomaly density. The time window length is determined through spectral analysis based on the flow rate of change characteristics and sampling period in the target frequency band. Within each window, calculate the window's average flow, standard deviation, maximum value, minimum value, peak-to-valley difference, first-order difference mean, first-order difference absolute value mean, and statistical values ​​of the rate of change of adjacent sampling points for the instantaneous flow sequence; and calculate the differential pressure sequence, sensor internal temperature, and other parameters. The ambient temperature and power supply voltage sequences are used to calculate the corresponding window average, variation amplitude, and short-term slope, forming window-level features describing the thermal balance environment and power supply stability. The number of CRC failures, retransmissions, timeouts, and link error count increments within the communication link status sequence window are statistically analyzed to form window-level features describing bus health. The proportions of original abnormal samples, saturated samples, and missing samples within the window are statistically analyzed to form abnormal density and missing density features. For each window, the start and end times, window length, and number of valid samples are recorded. Context state encoding includes: aggregating and statistically analyzing discrete state quantities such as preheating stage flags, range switching flags, self-test result flags, link risk flags, and time gap event flags within the window; compressing information such as whether preheating samples appear within the window, whether range switching boundaries are crossed, whether self-test failures or internal errors exist, and whether the time interval falls within a power supply abnormality or link abnormality into a set of context state features, which are added to the feature vector through Boolean values ​​or multi-hot encoding.

[0037] For windows with a missing sample ratio greater than the missing threshold, a statistically unreliable label is added, and the label is encoded into the feature vector. A quality assessment feature frame is generated for each window. The quality assessment feature frame includes: a window time index, continuous feature fields describing flow patterns and fluctuation characteristics, feature fields describing differential pressure and temperature environment, feature fields describing power supply and communication link stability, feature fields describing abnormal distribution and missing conditions, and discrete encoded fields describing preheating, range switching, and link risk context states. This allows the sample quality assessment and confidence labeling module to comprehensively consider the three dimensions of signal pattern, environmental conditions, and system state on a single structured input, and to perform refined calculations on the confidence of each window or each sampling point.

[0038] As shown in Table 1, the acquisition strategy priority classification table lists the acquisition strategies, priorities, and status label rules for the seven channel types: For thermal flow channels, the acquisition method is periodic acquisition, the priority is set to high, and the status labels are divided into three categories: normal, preheating, and range switching, corresponding to three operating conditions: stable sensor operation, incomplete power-on preheating, and range switching process, respectively. For differential pressure channels, the acquisition method is periodic acquisition synchronized with the thermal flow channel, the priority is set to high, and the status labels are divided into two categories: normal and self-test failure, corresponding to two states: normal differential pressure sensor measurement and internal self-test failure, respectively. For ambient temperature channels, the acquisition method is periodic acquisition, the priority is set to medium, and the status labels are divided into two categories: normal and environmental fluctuation, corresponding to two situations: stable ambient temperature and temperature fluctuations exceeding the reasonable range, respectively. For ambient humidity channels, the acquisition method is periodic acquisition, with a priority of... When the channel type is set to medium, the status label is consistent with the ambient temperature channel, i.e., normal and ambient fluctuation, corresponding to the operating conditions of stable ambient humidity and abnormal humidity fluctuation, respectively. When the channel type is the power supply voltage channel, the acquisition method is periodic acquisition, the priority is set to medium, and the status label is divided into two categories: normal and voltage exceeding the limit, corresponding to the two electrical operating conditions of power supply voltage within the standard range and voltage exceeding the rated threshold, respectively. When the channel type is the equipment self-test channel, the acquisition method is event-triggered (acquiring only during self-test), the priority is set to high, and the status label is divided into two categories: self-test passed and self-test failed, corresponding to the sensor and related modules passing the self-test and having a fault, respectively. When the channel type is the range switching channel, the acquisition method is event-triggered (acquiring only during range switching), the priority is set to high, and the status label is divided into two categories: stable and transition phase, corresponding to the two situations of stable operation after range switching and not reaching a stable state during the switching process, respectively.

[0039] Table 1. Priority Classification of Data Acquisition Strategies

[0040] In this implementation scheme, by transforming the original single-channel flow time series into a multi-channel, multi-scale quality assessment feature sequence carrying contextual information, a unified, complete, and scalable feature foundation is provided for sample-level quality scoring and confidence output in the thermal flow signal processing method.

[0041] Specifically, the steps for constructing a rule-driven and data quality scoring model for hot flow sequences based on quality assessment feature frames are as follows: Input the quality assessment feature frames and historical operating data, construct a data quality scoring model for hot flow sequences based on rule-driven principles and supplemented by lightweight statistics, and generate basic quality score values ​​for each time window; Perform statistical analysis and rule extraction on quality assessment features based on historical operating data, construct a rule-driven basic quality scoring model, perform cluster analysis and interval statistics on the historical distribution of features such as flow fluctuation, anomaly density, missing proportion, saturation proportion, power supply fluctuation amplitude, link error count, preheating flag, and range switching flag, divide the feature space into normal intervals, suspicious intervals, and abnormal intervals, and assign basic quality scores and penalty weights to each interval.

[0042] The feature interval division adopts histogram statistics and cluster analysis of long-term stable operating samples and confirmed abnormal samples in the feature space. First, the candidate interval boundary set is obtained in each feature dimension. Then, different interval combinations are cross-validated on historical operating data. The interval boundary that minimizes the comprehensive index of false positive rate and false negative rate is selected as the final division criterion for normal interval, suspicious interval and abnormal interval. The penalty weight is calculated by performing grid search or random search on different weight vectors on historical operating data with quality labels. The correlation coefficient and identification index between the basic quality score and the target quality label are calculated. The weight combination with the best comprehensive score is selected. The weight vector is normalized to ensure that the penalty intensity of each feature has a clear and reproducible source.

[0043] By statistically analyzing the characteristic distribution of long-term stable operating samples, typical ranges of flow fluctuation, differential pressure change amplitude, ambient temperature slope, and power supply stability indicators under normal operating conditions are determined. These ranges are used as basic high-quality intervals, and windows falling into these intervals are assigned a basic score. Replay analysis is performed on historically confirmed periods of sensor contamination, unstable power supply, or severe link anomalies to extract the corresponding feature change patterns, such as continuously increasing anomaly density, continuous accumulation of link error counts, prolonged failure to clear preheating flags, and frequent triggering of range switching flags. Corresponding feature intervals are marked as low-quality or unreliable intervals, and windows falling into these intervals are directly judged as unreliable. Transitional areas between these two are classified as suspicious intervals, and further refined through joint evaluation with other features. The results of the above interval classification are compiled into a rule table, defining a basic score contribution, penalty strategy, and priority resolution rules for conflicting features for each feature. This allows the rule engine to quickly provide a basic quality score for window-level feature vectors. High-quality samples refer to samples with a quality score greater than the high-quality threshold. Medium-quality samples refer to samples between the high-quality threshold and the low-quality threshold. Low-quality samples refer to samples with a quality score lower than the low-quality threshold but not reaching the unusable condition. Unusable samples refer to samples with a quality score lower than the unusable threshold.

[0044] Building upon the rule-driven basic quality scoring model, a lightweight statistical model is introduced to finely distinguish samples near the rule boundaries, forming a hybrid quality scoring mechanism. Quality assessment feature frames, along with historical label categories, are used as training data. Lightweight models such as logistic regression, decision trees, or gradient boosting trees are employed to learn the non-linear relationship between each feature and the quality category. Training labels are derived from multi-source aggregation and manual judgment of historical operational data: on the one hand, time periods with long-term operation and verified by comparative measurement or external reference instruments are selected as high-quality sample labels; on the other hand, time periods accompanied by sensor maintenance records, fault work orders, power supply anomaly records, or long-term communication error records are selected as low-quality or unusable sample labels. The remaining time periods with minor anomalies but no serious faults are marked as medium-quality samples. When the data scale is large, initial labels can be generated by the rule model first, and then the label quality can be improved through sampling and manual review and correction to ensure that the training labels have clear physical basis and consistency.

[0045] During training, cross-validation is used to select hyperparameters such as regularization strength, class weights, and tree depth to achieve a balance between high-quality sample recognition rate and abnormal sample rejection rate, while controlling complexity to facilitate operation in edge or embedded processing units. During runtime, the rule-based model first quickly classifies abnormal windows. Only windows near rule boundaries or with rule scores falling within the suspicious interval are supplemented with a lightweight model for judgment. The output class probability or score is weighted and synthesized with the base score of the rule-based model to form a mixed quality score for the window, avoiding excessive misjudgments or overly conservative processing caused by hard threshold rules at the boundaries. The weighted synthesis strategy uses a confidence-based weighted summation method. The mixed quality score is represented as a weighted sum of the regular model quality score and the statistical model quality score, with the weighting coefficients monotonically adjusted based on the position of the regular model quality score relative to the boundaries of the normal and abnormal intervals: when the regular model quality score is far from the interval boundary and within the normal or abnormal interval, the weighting coefficient is close to one, and the mixed quality score is mainly determined by the regular model quality score; when the regular model quality score is close to the boundary between the normal and suspicious intervals, or between the suspicious and abnormal intervals, the weighting coefficient decreases, increasing the proportion of the statistical model quality score in the mixed quality score, thus relying more on the lightweight statistical model's ability to distinguish complex patterns near the boundaries.

[0046] The specific curve form of the weight coefficients changing with the quality score of the rule model was determined by setting multiple candidate function forms and parameter combinations on historical running data with quality labels. The performance of different candidate schemes in terms of quality identification accuracy, anomaly removal effect, and score stability was compared. Cross-validation was used to select the function form and parameter values ​​with the optimal comprehensive index, ensuring the fusion strategy has good robustness and interpretability under different operating conditions. To prevent extreme scores under training data bias, a truncation and smoothing strategy was introduced for the mixed quality score values, limiting the rate of change and maximum adjustment range of the quality score in a single window, ensuring the smooth evolution of the quality score over time.

[0047] In this implementation scheme, by combining the above-mentioned rule-driven basic scoring with the fine supplementation of a lightweight learning model, a window-level quality score value reflecting the overall credibility of the current measurement link is generated for each sliding window, and the key feature contribution information and rule hit information constituting the score are retained, providing basic support for the generation of sample-level scoring and diagnostic information.

[0048] Specifically, the steps for deriving the sample-level quality score and outputting confidence labels and diagnostic information are as follows: Input the window-level quality score value and the corresponding quality assessment feature frame; calculate the sample-level quality score and confidence label for each time window and each sampling point within the window; and bind the quality score result and diagnostic information together to the thermal flow data output; the window-level quality score value comprehensively considers the overall level of continuous features within the window, the density of abnormal and missing samples, the distribution of context status labels, and the health of power supply and communication links, reflecting the overall reliability of the measurement link within the time period; the quality assessment feature frame provides anomaly markers, preheating flags, range switching flags, link risk flags, and other context statuses for each sampling point within the window, which are used to further distinguish the quality differences of different sampling points within the window based on the window-level score.

[0049] For each sliding window, based on the inherited window-level quality score, the sample-level quality score of each sampling point within the window is finely modified according to the anomaly distribution and context state. For sampling points within the window that are not marked as original anomaly samples, numerical jump samples, or saturated samples and are not at the moment of power supply anomaly or link error burst, the sample-level quality score is set to be close to the window-level quality score or adjusted within a narrow range near the window-level score. For sampling points marked as original anomaly samples, numerical jump samples, or at the moment of short-term power supply drop or bus error burst, the sample-level quality score is significantly reduced based on the window-level score by an attenuation coefficient, and if necessary, directly marked as unusable. For sampling points located in the range switching transition zone, the sample-level quality score is reduced proportionally so that they are automatically in a low-weight state in statistical calculation and training rather than being completely eliminated. A time smoothing strategy is adopted for the quality scores of adjacent sampling points within the window to prevent isolated spikes in the quality score of a single sampling point due to instantaneous noise.

[0050] The sample-level quality scores are mapped to numerical confidence labels and discrete quality status labels to achieve explicit confidence labeling of thermal flow sequences. The confidence labels can use a continuous interval from 0 to 1, where 0 represents completely unreliable and 1 represents high confidence. They are obtained by transforming the sample-level quality scores through a nonlinear mapping function. At the same time, quality status labels are divided according to the interval in which the sample-level quality scores fall. For example, samples with quality scores above the high-quality threshold are marked as normal, samples with quality scores between the high-quality threshold and the low-quality threshold are marked as suspicious, samples with scores below the low-quality threshold but not completely masked are marked as low-quality, and samples with extremely low scores or those directly judged as unusable by rules are marked as unusable. The quality status labels are used by downstream modules to quickly filter data of different quality levels.

[0051] Diagnostic information is generated based on sample-level quality scoring results. The dominant factors causing quality degradation are output as cause codes along with traffic data. The diagnostic information is automatically inferred by analyzing sample-level feature vectors and rule hits. For quality degradation caused by the warm-up phase, the cause code is set to "warm-up incomplete". For quality instability caused by range switching, the cause code is set to "range switching transition". For quality degradation caused by power outages or power fluctuations, the cause code is set to "power supply anomaly". For quality degradation caused by frequent link errors, the cause code is set to "communication anomaly". For quality degradation caused by excessive anomaly density or excessive missing data ratio, the cause code is set to "data anomaly or excessive missing data". Cause codes can be output individually or in combination. The cause code dictionary uses a discrete enumeration encoding method, assigning a fixed integer code to each dominant cause. The system configuration table uniformly provides the correspondence between the code and its meaning. This ensures that causes such as "warm-up incomplete", "range switching transition", "power supply anomaly", "communication anomaly", "data anomaly", and "excessive missing data" can be matched one-to-one through cause codes during internal processing, operation and maintenance analysis, and external system integration, avoiding ambiguity caused by inconsistent text descriptions. For samples that are unavailable, a data masking marker is added to the output to explicitly warn downstream modules that they are prohibited from participating in cumulative calculations, threshold judgments, and training. For samples that are in a suspicious or low-quality state, they are allowed to be used in scenarios that require event playback or extreme working condition analysis, but alarm prompts are added during statistical calculations and modeling.

[0052] Time periods where the window-level quality score consistently falls below a low threshold and the sample-level confidence level remains consistently below a low confidence threshold are aggregated to form quality events, which are recorded as part of the quality time series. The aggregation window refers to the time interval used to determine whether the quality score and confidence level remain consistently low. This time interval can be measured in seconds or by several consecutive sliding windows. The specific length is selected by statistically analyzing the duration of quality issues in historical operational data and combining it with cross-validation. This ensures that the aggregation window is sufficient to cover a complete quality degradation process without merging multiple independent short-term fluctuations into a single quality event, thus guaranteeing that the boundaries of the quality events have clear physical meaning. Hot traffic data frames with sample-level confidence labels, quality status labels, and diagnostic information are output. The quality event log includes the event start and end time, affected channel, main anomaly characteristic type, maximum anomaly density, minimum confidence value, and corresponding window number, used for quick backtracking of the problem occurrence interval during maintenance and fault diagnosis; quality events can be used as screening conditions during retraining and parameter optimization to avoid mixing severely contaminated data into the training samples; the confidence-based hot flow data frame includes: sampling timestamp, instantaneous flow value, sample-level quality score value, numerical confidence label, quality status label, and diagnostic cause code fields for preheating, range switching, power supply anomaly, and communication anomaly. These are calculated and added for each sampling point based on the quality assessment feature frame and the hybrid quality scoring model, giving each flow sample quality measurement information in addition to its numerical value.

[0053] like Figure 2 The comparison charts shown here illustrate the effects of basic filtering and outlier suppression. The top chart displays the original flow waveform, while the bottom chart shows the baseline flow waveform after basic filtering and outlier suppression. In the top chart, the horizontal axis represents time, and the vertical axis represents instantaneous flow. The light yellow curve represents the collected original flow sequence, showing multiple sharp pulses and drops superimposed on a generally smooth, undulating background trend. The two horizontal lines marked with red dashed lines represent the upper and lower limits of the flow range, determined based on the nominal range of the thermal flow sensor and statistical results of the intervals where clipping and distortion began to appear in historical operating data. Samples exceeding these limits are considered out-of-range samples. Near the upper and lower limits, highlighted rectangles and scatter dots mark the sampling points initially identified as anomalous pulses and saturated samples, indicating that although these points have relatively large or small values, they do not conform to the continuous change trend of the surrounding time. In the figure below, the horizontal and vertical axes have the same meaning as in the figure above. The blue curve represents the result of applying basic filtering and outlier suppression to the original flow sequence on the same time axis. The basic filtering uses smoothing processing centered on the window midpoint or mean to suppress high-frequency noise. Sampling points marked as out-of-bounds samples and saturated samples are replaced by interpolation of normal samples in the neighborhood or truncated at the baseline value, providing a smoother and physically reasonable baseline flow input for window-based statistical feature extraction and quality score calculation.

[0054] In this implementation scheme, hot flow data frames with sample-level confidence labels, quality status labels and diagnostic information are output and called by confidence-driven multi-scale flow reconstruction, statistical analysis and anomaly detection modules. This enables the processing to automatically downweight or remove low-quality data based on the confidence level, thereby achieving end-to-end quality perception and robust processing of hot flow signals.

[0055] Specifically, the steps for weighted filtering and resampling of hot flow sequences based on confidence labels are as follows: Input hot flow data frames with sample-level confidence labels and quality status labels; perform confidence-weighted filtering and resampling on the hot flow sequences to reconstruct information-enhanced fine-grained flow waveforms that balance waveform authenticity and data quality on the time axis; perform confidence-weighted filtering on the hot flow sequences based on sample-level confidence labels to suppress abrupt changes and noise introduced by low-quality samples, obtaining a smooth flow waveform while maintaining real physical changes; the confidence-weighted filtering includes: constructing a sliding time window on a unified time axis, and performing weighted averaging, weighted median, or weighted exponential smoothing on the instantaneous flow values ​​within the window to smooth the samples... The confidence level label is used as a weighting factor in the calculation; samples with an unusable quality status have their weights reset to zero and are removed from the filtering calculation; high-confidence samples refer to samples with a confidence level greater than or equal to the high-confidence threshold, medium-confidence samples refer to samples with a confidence level greater than or equal to the low-confidence threshold but less than the high-confidence threshold, and low-confidence samples refer to samples with a confidence level less than the low-confidence threshold; the weighting function changes monotonically with confidence and can be configured as a linear function, a power-law function, or a piecewise function with threshold truncation. By comparing the performance of different function forms on root mean square error, smoothness index, and time delay index on historical running data, cross-validation is used to automatically select the weighting function form and parameters with the best overall performance. Samples with questionable or low quality status are weighted according to confidence level, so that high-confidence samples dominate the filtering results. For each window, the effective weight sum and weight normalization factor are calculated simultaneously. When the effective weight sum is lower than the effective threshold, the filtering result output by the window is marked as a low-confidence filtering value, and a filtering confidence index is attached to prompt the statistical module to reduce the weight or ignore the result. During the filtering process, the difference between the original instantaneous flow value and the filtered flow value is retained as a residual sequence, which is used to identify the difference between noise caused by data quality problems and the sudden change in the actual working condition in anomaly detection.

[0056] Based on confidence-weighted filtering, the thermal flow sequence is sampled with confidence-weighted sampling to convert irregular sequences into a uniform time step, generating a fine-grained flow sequence with a regular structure and a resampling confidence index. The target sampling period is determined based on the variation characteristics of the thermal flow signal in the target frequency band and the original sampling period. It is automatically obtained by performing spectrum analysis and energy concentration analysis on historical operating data, so that the target sampling period can cover the typical rising and falling edges of the flow, but is not too large to cause excessive smoothing of short-term fluctuations, thus achieving a balance between time resolution and noise suppression. Confidence-weighted sampling includes: constructing a target time point sequence based on the target time granularity or target sampling period; selecting original sampling points near each target time point to form a resampling window; interpolating or reconstructing the instantaneous flow rate values ​​within the resampling window using sample-level confidence labels as weights; and obtaining the resampled flow rate values ​​at the target time points using weighted linear interpolation, weighted spline interpolation, or weighted local fitting methods. The specific interpolation method used is determined by calculating the root mean square error, the smoothness index of the first-order difference variance of the curve, and the relative time offset delay index of the characteristic peak on historical operating data with high-precision reference curves or high-confidence dense sampling segments. Cross-validation is then used to select the interpolation method and parameters that offer the best overall performance across the three categories of indicators: error, smoothness, and time delay. For target time points where the sum of effective confidence weights within the window is lower than the effective threshold, the resampled values ​​are marked as low-confidence interpolation values ​​or directly set as missing. Depending on the configuration, the output value is selected either with low confidence or only a missing label, to avoid interpolation results supported by low-quality samples from participating in the decision-making process. The effective threshold is determined by statistically analyzing the impact of the number of samples in different neighborhoods in historical operating data on the stability of resampling error and confidence estimation. Cross-validation is used to select the minimum number of samples that optimizes the combined index of error and coverage.

[0057] During the resampling process, the mapping relationship and weight allocation information between the original sampling points and the resampling points are maintained. The original sample index set and weight vector corresponding to each resampling point are recorded together as resampling traceability information, and a confidence-weighted fine-grained flow waveform is output. The confidence-weighted fine-grained flow waveform includes: the resampling flow value at the target time step, the corresponding sample-level comprehensive confidence, the filtered confidence and the resampling confidence, as well as the resampling traceability information. It is reconstructed on a unified time axis through confidence-weighted filtering and resampling. This allows for the rapid location of the original sample set and confidence distribution that constitute the value from the resampling flow value when performing event playback or tracing the statistical results at a certain time point, supporting the interpretation and analysis of the credibility of the reconstructed waveform.

[0058] like Figure 3The comparison chart of confidence-weighted filtering effects is shown, with time on the horizontal axis and flow rate on the vertical axis. The light yellow curve represents the original flow rate sequence without confidence-weighting, showing a large number of high-frequency spikes superimposed on the overall slow, fluctuating trend. The dark blue curve represents the filtered flow rate result after sample-level confidence-weighting on the same time series. For most of the time period, the dark blue curve maintains the same overall trend as the light yellow curve, but sharp short-term fluctuations are significantly weakened or suppressed. The two time intervals marked with light red shading correspond to periods where the sample-level confidence level is consistently below the low confidence threshold. These periods are often accompanied by power supply fluctuations, link errors, or abnormal sensor preheating. It can be seen that the original flow rate curve exhibits significant random jitter within the shaded areas, while the flow rate curve after confidence-weighted filtering is smoother and more continuous, avoiding the excessive influence of low-confidence samples on the overall flow rate trend. By comparing the original flow rate with the flow rate after confidence-weighted filtering in the same coordinate system, it can be intuitively shown that confidence-weighted filtering does not change the macroscopic change pattern of the flow rate signal, and can automatically reduce the weight of abnormal samples in the low confidence period, thereby improving the stability and physical reliability of the reconstruction results.

[0059] In this implementation scheme, the original thermal flow sequence is transformed from a simple sequence of single values ​​and time into a well-structured, fine-grained quality-aware flow waveform with filtered confidence and resampling confidence through the aforementioned confidence-weighted filtering and resampling processing. This provides stable and interpretable basic data for upper-level multi-scale integration, statistical analysis, and anomaly detection.

[0060] Specifically, the steps for performing multi-scale integration and statistical calculations to output multi-granularity flow results and abnormal event records are as follows: Based on the confidence-weighted fine-grained flow waveform, perform confidence-weighted integration and multi-scale statistical calculations on the thermal flow sequence to generate flow statistical results with overall confidence evaluation at each time scale; construct a multi-granularity flow result set carrying an overall confidence index. The confidence-weighted integration includes: accumulating the fine-grained resampled flow values ​​covered by the specified time interval with adjacent time steps; considering the product of the flow value and the time interval within each time step as the volume contribution of the step; weighting the accumulation using the comprehensive confidence corresponding to the resampled flow value as the weighting coefficient; and setting the volume contribution to zero for time steps with a comprehensive confidence close to zero or an unusable quality state. For time steps with low overall confidence but not completely unusable, the volume contribution is reduced proportionally to the confidence level to avoid excessive impact of abnormal peaks or erroneous readings on the cumulative flow results. After integration, the sum of effective weights, effective time coverage ratio, and number of effective steps are calculated to construct an interval confidence index for the cumulative flow results over a time interval, which characterizes the statistical reliability of the cumulative results. The interval confidence index is calculated by weighting the overall confidence of each time step within the interval by time interval to obtain a time-weighted average confidence. Then, the effective time coverage ratio is multiplied by the time-weighted average confidence as a decay factor to obtain an interval confidence value ranging from zero to one, thus simultaneously reflecting the sample confidence level and the completeness of time coverage. The multi-granularity flow result set includes: quality-sensed flow waveforms at the millisecond or sampling period level, confidence-weighted statistics at the second and minute or adaptive window level, and abnormal event records based on quality perception, used to support trend analysis, training, and equipment operation status assessment of flow signals. The time granularity is set at several levels, including sampling period, second, minute, and window level determined by event adaptation. The time length of each level is determined by performing spectrum analysis and energy distribution statistics on historical operating data. Cross-validation is used to select the optimal value between waveform detail retention and noise suppression effect from a set of candidate time lengths, ensuring that fine granularity can characterize transient changes and coarse granularity can reflect long-term trends.

[0061] Multi-scale statistics include: dividing the time axis into continuous non-overlapping or slidingly overlapping statistical windows according to time granularity; calculating the flow mean, variance, maximum, minimum, peak, occurrence time, and first-order difference statistics based on confidence-weighted fine-grained flow waveforms within each window; when calculating the above statistics, the flow values ​​of each fine-grained sample are weighted by the overall confidence level for weighted average and weighted variance calculations, ensuring that high-confidence samples dominate the statistical results while the influence of low-confidence samples is automatically compressed; recording the effective weight coverage ratio and the number of statistical samples for each statistical window; when the number of samples is insufficient, the statistical results of the window are marked as having insufficient statistical confidence, indicating that caution should be exercised when using these statistics; and uniformly encapsulating statistical results at different time granularities (such as second-level, minute-level, or adaptive event-level) into multi-granularity flow result records, recording the time range, time scale type, statistical type, numerical results, and overall confidence index, providing selectable multi-granularity inputs. The overall confidence index is obtained by time-weighted average of the comprehensive confidence of fine-grained samples within the corresponding statistical window, and multiplied by the effective weight coverage ratio of the window. When the coverage ratio is low, the overall confidence is automatically reduced so that the statistical results are not mistakenly considered to be highly reliable when the sample is sparse.

[0062] Based on multi-scale statistics, a quality-aware anomaly detection and labeling mechanism is introduced. This mechanism uses a joint criterion of numerical anomaly degree and confidence level information to distinguish between physical flow events and measurement link anomalies, outputting anomaly event records. Quality-aware anomaly detection includes: applying mutation detection, rate of change thresholding, or residual thresholding algorithms to confidence-weighted fine-grained flow waveforms at high-frequency scales to identify rising edges, falling edges, and peak candidate anomalies in the flow. This is then combined with the comprehensive confidence level, quality status label, and diagnostic cause code at the corresponding time point for screening. The joint criterion specifically includes three parts: numerical deviation degree condition, auxiliary channel consistency condition, and confidence level lower limit condition. When the flow instantaneous... If the absolute value of the hourly increment or residual exceeds the numerical deviation threshold determined by statistical analysis of historical operating data, and the direction and magnitude of the changes in the differential pressure channel and temperature channel in the same time neighborhood match the flow rate changes, and the overall confidence level and window-level quality score at the corresponding time point are both higher than the lower confidence level threshold and the lower quality level threshold, the candidate anomaly is determined to be a physical flow event. If the candidate anomaly meets the numerical deviation conditions, but the differential pressure or temperature channel does not show corresponding changes, or the overall confidence level is lower than the lower confidence level threshold, and is accompanied by power supply anomalies, frequent communication errors, or self-test failure diagnostic reason codes, the candidate anomaly is determined to be a measurement link anomaly.

[0063] When candidate anomalies are concentrated on high-confidence samples and match the variation patterns of differential pressure and temperature auxiliary channels, they are identified as physical flow events. When candidate anomalies are mainly triggered by low-confidence samples and are accompanied by power supply anomalies, frequent communication errors, or self-test failure diagnostic markers, they are identified as measurement link anomalies to avoid misreporting measurement noise or link faults as actual operating condition changes. At mid-to-low frequency scales, multi-scale statistical results and window-level quality scores are combined to classify and label long-term drift, periodic fluctuations, and short-term abnormal fluctuations. A sliding window average is applied to the difference between the confidence-weighted mean sequence and the reference mean sequence at mid-to-low frequency scales. Compared with the offset threshold, when the moving average offset continuously exceeds the offset threshold and the corresponding window-level quality score is higher than the lower quality limit threshold, the phenomenon is marked as a slow change in operating conditions, which is used for calibration and compensation strategy design. When the variance or peak value increases in a short period of time, and the window-level quality score decreases or is accompanied by link risk marking, the phenomenon is marked as measurement link instability, and a quality alarm record is generated. The above anomaly detection results are output in the form of anomaly event records. Each record includes event type, time range, significance index, average and minimum confidence during the event period, dominant anomaly features, and diagnostic labels, providing high-value structured input for maintenance and training.

[0064] Multi-granularity flow results and anomaly records are output in a unified data format, preserving the index relationship from multi-scale results to fine-granular samples. Multi-granularity flow results records include time interval identifiers, time scale type, flow statistics, overall confidence index, effective weight coverage ratio, and number of effective samples, used for filtering and weighting by confidence level. Anomaly records include event type, start and end times, corresponding multi-scale statistical segment index and fine-granular sample index, key indicator values, and diagnostic labels. This allows analysts to quickly trace back to the original flow samples and confidence distribution that constitute the results when viewing a statistical result or anomaly, achieving bidirectional linkage from the statistical view to the original data view.

[0065] Table 2 shows the correspondence between quality status labels and confidence levels, outlining seven quality statuses, their corresponding parameters, and processing rules. When the quality status is "high confidence normal," the sample quality score range is [0.8, 1.0], corresponding to a confidence level range of [0.9, 1.0], indicating extremely high sample confidence. Downstream processing involves full participation in score calculation, training, and business decision-making, without any weight discounts or usage restrictions. When the quality status is "low confidence normal," the sample quality score range is [0.6, 0.8), corresponding to a confidence level range of [0.7, ...]. 0.9) indicates that the sample is basically reliable but has slight uncertainty. The downstream processing method is to participate in various calculations normally, only appropriately reducing the weight to weaken the potential impact; when the quality status is doubtful - power supply impact, the sample quality score range is [0.4, 0.6), and the corresponding confidence range is [0.5, 0.7), indicating that the sample reliability is moderate and affected by power supply factors. The downstream processing method is to reduce the weight during statistical calculations, and whether to include it in training can be selected according to needs; when the quality status is doubtful - communication impact, the sample quality score range is [ A score of 0.4, 0.6), corresponding to a confidence range of [0.5, 0.7), indicates that the sample confidence is moderate and affected by communication link factors. The downstream processing method is consistent with the suspected power supply impact, i.e., statistical weighting and modeling are optional. When the quality status is low quality - non-critical anomaly, the sample quality score range is [0.2, 0.4), corresponding to a confidence range of [0.3, 0.5), indicating that the sample confidence is low but the anomaly does not affect the core business. The downstream processing method is only used for event replay analysis and does not participate in training. When the quality status is low quality When the sample quality score is [0.0, 0.2), the corresponding confidence level is [0.0, 0.3), indicating that the sample has extremely low credibility and contains a key anomaly. The downstream processing method is to reduce the weight by 90% during statistical calculation and remove the sample directly during training. When the quality status is unusable, the sample quality score is [0.0, 0.0], the corresponding confidence level is [0.0, 0.0], indicating that the sample is completely unreliable. The downstream processing method is to completely block the sample from participating in any integral calculation, threshold judgment, or training.

[0066] Table 2 Correspondence between Quality Status Labels and Confidence Levels

[0067] like Figure 4The diagram showing the distribution of characteristic intervals indicates that the horizontal axis represents the characteristic value of flow fluctuation, which is a dimensionless index obtained from the standard deviation of the flow series within a given time window, used to characterize the severity of flow changes within that window; the vertical axis represents the number of samples when the corresponding characteristic value falls into each interval. The bars in the graph are displayed in different colors according to the interval category: the green bars on the left correspond to samples classified as "normal intervals". When the flow fluctuation is less than or equal to the upper limit of the normal interval, the samples are mainly concentrated around 0.3 to 0.7, showing a high peak and narrow distribution, indicating that the flow fluctuation is small and the distribution is stable under normal operating conditions; the yellow bars in the middle correspond to samples in "suspicious intervals". When the flow fluctuation is between the upper limit of the normal interval and the upper limit of the suspicious interval, the height of the bars decreases significantly and the distribution widens, indicating that the flow fluctuation in this interval is significantly higher than the normal operating level, but has not yet reached the level of extreme abnormality, and further judgment is needed in combination with other characteristics; the red bars on the right correspond to samples in "abnormal intervals". When the flow fluctuation exceeds the upper limit of the suspicious interval, although the number of samples is small, the characteristic values ​​are distributed above 1.5 or even close to 2.5, indicating that the samples generally have violent fluctuations or strong interference.

[0068] In this implementation scheme, by using the aforementioned confidence-weighted integral and statistical calculation, as well as quality-perceived anomaly detection and identification, the numerical results and quality information are deeply coupled, enabling the thermal flow signal to simultaneously consider both measurement values ​​and data quality during multi-scale analysis and anomaly detection, thereby significantly improving the robustness and interpretability of the flow results.

[0069] Specifically, the steps for constructing a hierarchical archiving and data hierarchy structure based on multi-granularity flow results and quality event records are as follows: Input fine-grained flow waveforms with confidence levels, multi-granularity flow results, and quality event records; perform hierarchical storage and indexing of flow data of different granularities and corresponding quality labels, diagnostic information, acquisition configuration, and algorithm version information; map the original acquisition sequence, preprocessing results, quality assessment results, statistical results, and event records into a unified data hierarchy structure; divide the data related to the acquisition link into multi-level data views according to time, channel, and processing stage, and use a unified index... The system constructs inter-layer mapping relationships, forming a tree-like or graph-like data hierarchy structure. In the raw data layer, it stores the thermal flow acquisition data frames output by the acquisition and control unit, retaining the original fields of sampling timestamps, instantaneous flow values, differential pressure values, sensor internal temperature, ambient temperature and humidity, power supply voltage, communication link status, and self-test results, as well as CRC error flags and buffer overflow flags, used to reconstruct arbitrary processing flows from the lowest-level signal. In the preprocessing layer, it stores the preprocessed dataset after time consistency verification, multi-channel alignment, segment division, and basic filtering, along with time segment encoding... Structured information such as time-series anomaly markers, time-gap event markers, and long-term missing segment markers is used to support sliding window feature calculation and segment-level analysis. In the quality assessment layer, a quality view is stored, consisting of quality assessment feature frames, rule hit information, window-level quality scores, and sample-level confidence labels. This encapsulates the feature vector, basic rule score, and mixed score output by the lightweight model for each time window, supporting offline analysis of quality model behavior and threshold rationality. In the statistical results layer, a statistical view is stored, consisting of multi-granularity traffic result records, showing the traffic at various time scales. The mean, variance, cumulative amount, peak value, and overall confidence index are archived for long-term trend analysis and training sample selection. At the event layer, all quality event records and physical traffic event records are stored to form an event index layer on the time series, which is used to quickly locate abnormal segments and key operating conditions. By uniformly introducing global time index, channel identifier, time segment number, and processing link ID index fields into the data of the above layers, any statistical result or event record can be associated with its source quality assessment results, preprocessing records, and original collection messages, forming a two-way tracking path from top to bottom and from bottom to top.

[0070] The global time index uses a monotonically increasing timestamp at the millisecond level under a unified time base. The channel identifier uses a combination of device ID and sensor channel number. The time segment number increments from zero in chronological order within each channel. The processing link ID uniquely identifies a processing path with a fixed algorithm and configuration combination in the system. The above four fields maintain the same naming and encoding rules in the data of each layer, and are the smallest set of fields to achieve inter-layer mapping and consistent association. When an abnormal offset is found in a second-level or minute-level statistical result, the time index and channel identifier can be used to quickly trace back to the fine-grained traffic samples and confidence distribution covered by the statistical interval, and then trace back to the corresponding original acquisition data frame and link status record from the fine-grained sample. When a power supply jitter or bus error burst is found in a certain time period in the original acquisition layer, the segment number and time interval can be used to reverse query the corresponding changes and alarms in the statistical result layer and event layer, thereby realizing cross-level and multi-perspective comprehensive analysis in the complete spectrum structure.

[0071] Each layer of data is appended with metadata for acquisition configuration, calibration parameters, and algorithm model version. This metadata describes both the origin and processing of the data within the data hierarchy, ensuring traceability of the signal processing process at the version level. Acquisition configuration metadata includes: sensor model, measurement range, installation location identifier, sampling period, warm-up time setting, differential pressure interface type, bridge or drive circuit configuration, and communication interface type and rate, describing the hardware and sampling strategy corresponding to the current data. Calibration parameter metadata includes: zero-point offset of the thermal flow sensor, sensitivity coefficient, temperature compensation coefficient, differential pressure-flow conversion parameters, most recent calibration time, and calibration conditions. The system includes calibration residual indices to reflect the physical calibration status of the current data. Algorithm model version metadata includes: preprocessing algorithm version number, quality assessment model version number, confidence-weighted filtering and resampling algorithm version number, confidence-weighted integration and statistics algorithm version number, and anomaly detection algorithm version number. Each version number corresponds to a unique algorithm configuration summary and parameter hash value, enabling accurate reproduction of the processing flow in an offline environment. During data entry and archiving, the acquisition configuration metadata, calibration parameter metadata, and algorithm model version metadata are combined using the processing link ID to assign a unique processing lineage identifier to each continuous time interval. The processing lineage identifier is generated by hashing four parts: data source identifier, channel identifier, time interval identifier, and processing link ID, ensuring no duplication across the entire database. The time interval identifier consists of the start and end timestamps of the interval, clearly defining the time range covered by the lineage identifier. This allows the same channel to form multiple independent but traceable lineage units before and after different acquisition configuration switches, facilitating filtering and comparison by lineage identifier during analysis.

[0072] In this implementation plan, any traffic sample, quality score, statistical result, or event record can be traced back to the sensor configuration, calibration parameter version, and signal processing algorithm version combination at that time through genealogy identification, thereby achieving precise correlation and visual management in both time and version dimensions.

[0073] Specifically, the steps for evolution management and quality playback diagnostic support of the hot flow signal processing link are as follows: Figure 5 The diagram illustrates the version coexistence and quality replay process based on data hierarchy structure. Based on this structure, it analyzes the coexistence and differences between old and new version results, without directly overwriting historical results. Instead, new results are stored as an additional layer of derived results. When the quality scoring model is retrained, the preprocessing strategy is optimized, or the confidence weighting rules are adjusted, a new version number and processing link ID are assigned to the new version of the model and algorithm. After the new version is launched, newly collected traffic data automatically carries the new version identifier and the new hierarchy identifier into the archiving system. For historical time intervals, if it is necessary to evaluate the performance of the new version model relative to the old version model, representative time segments can be selected from the original data layer or the preprocessing layer for offline loop analysis. The system applies a new version of the quality assessment model, confidence-weighted algorithm, and anomaly detection algorithm to generate new versions of quality scores, multi-scale statistical results, and anomaly event records for the same original data. These results are then appended and stored as derived results in the quality assessment layer, statistical result layer, and event layer, and are labeled with the new version's algorithm version number and processing link ID. Queries support filtering or comparison by algorithm version. For example, it compares the window-level quality score distribution, sample-level confidence distribution, and the number and type of anomalies between the old and new versions within the same time interval. This quantifies the improvements of the new version in terms of false positive rate, false negative rate, and score stability, providing quantifiable evaluation criteria for the continuous iteration of signal processing algorithms.

[0074] A quality playback and diagnostic support mechanism for operation and maintenance (O&M) and analysis is constructed. Utilizing multi-layered archived data and spectral information, it enables multi-dimensional playback and comparison by time, version, and event, supporting rapid location and interpretation of sensor and data acquisition link issues. Time-based playback includes: when O&M personnel are troubleshooting sensor anomalies or communication link problems within a specific time period, they select a time interval and channel number in the archive system, extracting corresponding segments from fine-grained flow waveforms, sample-level confidence sequences, quality status labels, multi-scale statistical curves, and quality event markers. These segments are then overlaid on a unified timeline to display changes in flow values, quality scores, and link status, helping to distinguish between sudden changes in actual operating conditions and measurement link interference. Version replay includes: retrieving quality assessment results and anomaly event records generated by different algorithm versions within the same time interval, comparing window-level quality score curves, sample-level confidence distribution heatmaps, and anomaly event lists in a unified interface, analyzing the differences between old and new versions in anomaly sensitivity and robustness, and providing a basis for whether to switch or revert to an algorithm version; event-based replay includes: selecting a quality event or physical traffic event from the event layer, automatically locating the start and end times and the corresponding fine-grained sample range, retrieving the detection results and scores of the event under different algorithm versions, and displaying the traffic waveform, confidence evolution, and link state evolution before and after the event, providing multi-dimensional evidence for root cause analysis and pattern recognition.

[0075] The archiving system provides multi-dimensional query interfaces by time, channel, quality level, and algorithm version, enabling analysts to quickly filter high-confidence, stable datasets for flow and differential pressure characteristic modeling, temperature compensation strategy optimization, and quality scoring model training. It also allows for filtering low-quality datasets with high anomaly density for link hardening and anomaly pattern mining. Through long-term statistical analysis of the quality scoring results, confidence distribution, and anomaly detection differences of multiple algorithm versions on the same historical data, and by statistically analyzing the quality scoring results and anomaly event results of different algorithm versions, the system compares the differences in window-level quality scoring standard deviation, anomaly event detection recall rate, and false alarm rate among the versions, generating an evaluation report that includes indicator curves and summary tables.

[0076] This implementation scheme manages the coexistence of different versions of output results within the same data hierarchy structure, and combines multi-dimensional playback and comparison capabilities to achieve objective evaluation and visual comparison between versions, thereby ensuring the continuity and interpretability of evaluation indicators over time. It can quickly identify key links in the sensor and acquisition chain that lead to quality degradation, providing accurate samples and quantitative basis for calibration parameter correction, acquisition strategy adjustment, and quality scoring model retraining, thus improving the overall stability, maintainability, and problem diagnosis efficiency of the thermal flow signal processing chain.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0078] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A signal processing method based on a thermal flow sensor, characterized in that: Includes the following steps: S1, sensors collaboratively collect multiple types of traffic and context datasets, and perform unified time base alignment and context labeling on the multiple types of traffic and context datasets to form a hot traffic acquisition data frame; S2, perform basic filtering and outlier screening on the thermal flow acquisition data frame, and construct quality assessment feature frames by sliding window segmentation modeling and multi-scale statistical feature extraction. S3, based on the quality assessment feature frame, constructs a hot flow sequence rule-driven and data quality scoring model, derives sample-level quality scores and outputs confidence labels and diagnostic information; S4 performs weighted filtering and resampling of the hot flow sequence based on confidence labels, and performs multi-scale integration and statistical calculations to output multi-granularity flow results and abnormal event records. S5 constructs a hierarchical archiving and data hierarchy structure based on multi-granularity flow results and quality event records, and provides evolution management and quality playback diagnosis support for the hot flow signal processing link.

2. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for the sensors to collaboratively collect multiple types of traffic and context datasets, and to perform unified time-base alignment and context labeling on the multiple types of traffic and context datasets to form a hot traffic acquisition data frame are as follows: Multiple types of flow and context datasets are collected through sensor collaboration. These datasets include: thermal flow sensor output datasets, differential pressure sensor datasets, environmental condition datasets, power supply monitoring datasets, communication link status datasets, equipment operation status datasets, and timestamp and sequence identifier datasets. Historical acquisition sequences stored under the same installation scenario and configuration conditions are used as historical operating data. The collected multiple types of flow and context datasets are managed by acquisition strategies and classified and labeled with context. The acquisition control unit performs timing scheduling and unified triggering of various sensor interfaces. The multi-channel data is time-aligned based on the local unified clock of the acquisition node to obtain the processed multiple types of flow and context datasets. The multiple types of flow and context datasets collected in each sampling period are encapsulated with a unified data structure to construct thermal flow acquisition data frames. The thermal flow acquisition data frames are then segmented, scaled, normalized, and format standardized.

3. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for performing basic filtering and outlier screening on the thermal flow acquisition data frame are as follows: Input a hot flow acquisition data frame, perform time consistency verification and multi-channel alignment on the global acquisition queue, mark and correct out-of-order, frame-dropped and delayed sampling, and obtain a preprocessed hot flow acquisition data frame; Basic filtering and outlier screening include: for instantaneous flow rate, differential pressure, and continuous quantities of bridge voltage or drive current, small window moving average, moving median filtering, and first-order exponential smoothing lightweight filtering methods are used within the segment to suppress high-frequency quantization noise and isolated spikes while ensuring that the flow rate change trend and edge morphology are not significantly weakened, generating smoothed values ​​and residual values; and forming a preliminary estimate of anomaly density.

4. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for constructing quality assessment feature frames through sliding window segmentation modeling and multi-scale statistical feature extraction are as follows: Input the preprocessed thermal flow acquisition data frame, perform sliding window segmentation modeling on the time axis, extract multi-scale statistical features from the flow and context data in each window, and encode the context state variables into feature vectors to form a feature sequence; For windows where the proportion of missing samples is greater than the missing threshold, add statistically unreliable labels and encode the labels into the feature vector; Generate a quality assessment feature frame for each window.

5. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for constructing a hot flow sequence rule-driven and data quality scoring model based on quality assessment feature frames are as follows: Inputting quality assessment feature frames and historical operational data, a data quality scoring model based on rules and supplemented by lightweight statistics is constructed for hot flow sequences. A basic quality score value is generated for each time window. Cluster analysis and interval statistics are performed on the historical distribution of features such as flow fluctuation, anomaly density, missing proportion, saturation proportion, power supply fluctuation amplitude, link error count, preheating indicator, and range switching indicator. The feature space is divided into normal intervals, suspicious intervals, and abnormal intervals, and a basic quality score and penalty weight are assigned to each interval. On top of the rule-driven basic quality scoring model, a lightweight statistical model is introduced to finely distinguish samples near the rule boundaries, forming a hybrid quality scoring mechanism. The quality assessment feature frames and historical label categories are used as training data, and lightweight models such as logistic regression, decision trees, or gradient boosting trees are used to learn the nonlinear relationship between each feature and the quality category.

6. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for deriving the sample-level quality score and outputting confidence labels and diagnostic information are as follows: The system inputs a window-level quality score and the corresponding quality assessment feature frame. For each time window and each sampling point within the window, it calculates a sample-level quality score and confidence label, and binds the quality score results and diagnostic information to the hot flow data output. For each sliding window, based on the inherited window-level quality score, it performs fine-grained correction on the sample-level quality score of each sampling point within the window according to the abnormal distribution and context state. The sample-level quality score is mapped to a numerical confidence label and a discrete quality state label. Diagnostic information is generated based on the sample-level quality score results, and the dominant factors causing quality degradation are output as cause codes along with the flow data. Aggregate time periods where the window-level quality score is consistently below the low score threshold and the sample-level confidence level is consistently at the low confidence threshold to form quality events, and record them as part of the quality time series; output hot traffic data frames with sample-level confidence labels, quality status labels and diagnostic information.

7. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for weighted filtering and resampling of the hot flow sequence based on confidence labels are as follows: Inputting a hot flow data frame with sample-level confidence labels and quality status labels, the system performs confidence-weighted filtering and resampling on the hot flow sequence to reconstruct an information-enhanced fine-grained flow waveform on the time axis that balances waveform authenticity and data quality: Confidence-weighted filtering is applied to the hot flow sequence based on sample-level confidence labels to suppress abrupt changes and noise introduced by low-quality samples, obtaining a smooth flow waveform while preserving the true physical changes; Based on the confidence-weighted filtering, confidence-weighted sampling is performed on the hot flow sequence to convert irregular sequences into a uniform time step, generating a well-structured fine-grained flow sequence with resampling confidence indicators. During the resampling process, the mapping relationship and weight allocation information between the original sampling points and the resampling points are maintained. The original sample index set and weight vector corresponding to each resampling point are recorded together as resampling traceability information, and the confidence-weighted fine-grained flow waveform is output.

8. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for performing multi-scale integration and statistical calculations, and outputting multi-granularity flow results and abnormal event records are as follows: Based on the confidence-weighted fine-grained flow waveform, confidence-weighted integral and multi-scale statistical calculations are performed on the thermal flow sequence to generate flow statistics results with overall confidence evaluation at each time scale; Construct a multi-granularity flow result set carrying an overall confidence index. The multi-scale statistics include: dividing the time axis into continuous non-overlapping or sliding overlapping statistical windows according to the time granularity; and calculating the flow mean, variance, maximum value, minimum value, peak value, occurrence time and first difference statistics based on the confidence-weighted fine-granular flow waveform within each window. Based on multi-scale statistics, a quality-aware anomaly detection and identification mechanism is introduced. The joint criterion of numerical anomaly degree and confidence information is used to distinguish physical flow events from measurement link anomalies and output anomaly event records. Multi-granularity flow result records and anomaly event records are output in a unified data format, and the index relationship from multi-scale results to fine-granular samples is retained.

9. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for constructing a hierarchical archiving and data hierarchy structure based on multi-granularity flow results and quality event records are as follows: Input fine-grained flow waveforms with confidence levels, multi-granularity flow results, and quality event records. Flow data of different granularities, along with corresponding quality labels, diagnostic information, acquisition configurations, and algorithm version information, are stored and indexed hierarchically. The original acquisition sequence, preprocessing results, quality assessment results, statistical results, and event records are mapped to a unified data hierarchy structure. Data related to the acquisition chain is divided into multi-layered data views according to time, channel, and processing stage, and inter-layer mapping relationships are constructed through unified index fields, forming a tree-like or graph-like data hierarchy structure. During data entry and archiving, a unique processing hierarchy identifier is assigned to each continuous time interval of data.

10. The signal processing method based on a thermal flow sensor according to claim 1, characterized in that: The specific steps for supporting the evolution management and quality playback diagnostics of the hot-type flow signal processing link are as follows: Based on the data hierarchy structure, we conduct coexistence and difference analysis of results from old and new versions. Instead of directly overwriting historical results, we store new results as an additional layer of derived results. We construct a quality playback and diagnostic support mechanism for operation and maintenance and analysis. By utilizing multi-layered archived data and hierarchy information, we achieve multi-dimensional playback and comparison by time, version, and event. By statistically analyzing the quality scoring results and abnormal event results of different algorithm versions, we compare the differences of each version in window-level quality scoring standard deviation, abnormal event detection recall rate, and false alarm rate indicators, and generate an evaluation report that includes indicator curves and summary tables.

Images