A Microtube Laminar Gas Flow Data Analysis Method Based on Edge Computing
By using edge computing methods to adaptively adjust the sliding window length and dynamic baseline, the problems of flow pattern fluctuations and communication effects in the measurement of small gas flow rates are solved, achieving efficient real-time anomaly detection and data continuity.
Patent Information
- Application Number
- CN202611552804.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies for measuring minute gas flow rates, fixed sliding windows cannot adapt to flow fluctuation cycles, independent spectral analysis cannot participate in subsequent signal processing, dynamic baselines are easily contaminated by continuous abnormal samples, and centralized processing is affected by communication status.
An edge computing-based approach is adopted to collect signals from microtube laminar flow sensors, establish an initial dynamic baseline and reference spectrum, adjust the sliding window length using the main frequency component, and adjust the dynamic baseline by combining the proportion of high-frequency disturbance energy. Sensor outlier identification and sliding filtering are performed to generate standard volumetric flow rate, and spectrum analysis and anomaly detection are conducted at the edge nodes.
It implements adaptive sliding window processing, distinguishes between sensor outliers and changes in actual operating conditions, reduces the impact of noise, maintains the stability of the dynamic baseline, and ensures the accuracy of real-time anomaly detection and data continuity during network outages.
Smart Images

Figure CN123065540A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of gas micro-flow measurement, industrial sensor signal processing and edge computing technology, specifically involving a microtube laminar gas micro-flow data analysis method based on edge computing. Background Technology
[0002] The measurement and condition monitoring of minute gas flow rates are widely used in the delivery of specialty gases in semiconductor manufacturing processes, the micro-gas branches of medical gas equipment, carrier gas control in analytical instruments, and gas bypass sampling. Compared with conventional gas flow rate measurements, the measurement signal amplitude corresponding to minute gas flow rates is smaller, and the differential pressure signal is easily affected by sensor zero drift, electrical noise, mechanical vibration, upstream pressure fluctuations, valve operation, changes in ambient temperature, and changes in gas properties.
[0003] For differential pressure measurement structures employing microchannels or laminar flow elements, gas flow information can be obtained based on the deterministic relationship between flow rate and differential pressure within the applicable laminar flow operating range. However, the differential pressure signal output by a microchannel laminar flow sensor does not only contain low-frequency components reflecting the average flow rate, but may also include components varying at different time scales due to periodic pulsations in the gas supply equipment, control valve actions, pipeline vibration, the development of minor leaks, local blockages, and sensor noise.
[0004] Existing signal processing methods typically perform moving average, median processing, or other denoising operations on the differential pressure signal within a fixed-length data window, and then calculate the gas flow rate based on the filtering results. While a fixed moving window can reduce random noise under stable operating conditions, when the flow regime fluctuates periodically, an excessively long moving window may attenuate the true pulsation components that need to be retained, while an excessively short moving window may retain more random noise.
[0005] In the measurement of minute gas flow rates, the amplitude of actual abnormal pulsations and differential pressure sensor noise may be similar. If all out-of-limit data are directly deleted as abnormal data during the preprocessing stage, the signal structure generated by actual leaks, blockages, or flow pulsations may be destroyed before anomaly detection. If isolated sensor spikes are not processed at all, these isolated spikes may form non-real high-frequency energy in subsequent spectrum analysis.
[0006] Spectrum analysis can describe the periodic composition of differential pressure signals and energy variations across different frequency bands. However, if spectrum analysis is only used for display or to generate independent alarms, and the sliding window length and dynamic baseline parameters remain fixed, the spectrum analysis output will not affect subsequent signal processing, and multiple processing algorithms will still lack data dependencies. The dynamic baseline can change slowly under normal operating conditions, but a fixed update rate applies the same processing method to stable operating conditions, slow transitions between operating conditions, and sudden abnormal conditions. When continuous abnormal samples continuously enter the dynamic baseline update process, the dynamic baseline may gradually move towards the abnormal state, and the deviation of the standard volumetric flow rate relative to the dynamic baseline will continuously decrease, leading to the problem of continuous abnormalities being absorbed by the normal reference state.
[0007] Centralized analysis architectures also require the continuous transmission of high-frequency raw signals to remote computing devices. When multiple minute gas flow measurement points operate simultaneously, high-frequency differential pressure signals, absolute pressure signals, and temperature signals form a continuous data stream. Network jitter or communication interruptions can affect the remote alarm time. Simply migrating conventional algorithms to field devices can reduce network transmission distance, but it cannot automatically solve the problems of mismatch between fixed filter scales and flow regime changes, as well as the dynamic baseline of continuous abnormal contamination. Summary of the Invention
[0008] The purpose of this invention is to provide a microtube laminar gas low-flow-rate data analysis method based on edge computing, which solves the technical problems that fixed sliding windows cannot adapt to different flow pattern fluctuation cycles, independent spectrum analysis cannot participate in subsequent signal processing, dynamic baselines may be contaminated by continuous abnormal samples, and centralized processing is affected by communication status.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A microtube laminar gas flow data analysis method based on edge computing includes: The differential pressure signal, absolute pressure signal, and temperature signal output by the microtube laminar flow sensor are collected to form a microtube laminar flow gas measurement sequence with a unified timestamp; During the initial normal monitoring period, establish the initial dynamic baseline, initial dynamic standard deviation, and reference spectrum; The edge computing node determines the sliding window length of the current processing cycle based on the main frequency component of the previous processing cycle, and performs sensor outlier identification and sliding filtering on the differential pressure signal to obtain the filtered differential pressure signal. The actual volumetric flow rate is calculated based on the filtered differential pressure signal, and a standard volumetric flow rate is generated by combining the absolute pressure signal, temperature signal, and gas physical property parameters. The filtered differential pressure signal is subjected to spectral transformation to obtain the main frequency component and spectral energy distribution vector. The sliding window length of the next processing cycle is updated based on the main frequency component. The high-frequency disturbance energy ratio is obtained based on the spectral energy distribution vector. The forgetting coefficient of the dynamic baseline is adjusted using the high-frequency disturbance energy ratio. The dynamic baseline and dynamic standard deviation are updated recursively based on the standard volumetric flow rate. The reference spectrum is also updated recursively based on the spectral energy distribution vector. A joint anomaly is generated based on the standardized flow deviation of the standard volume flow rate relative to the dynamic baseline and the degree of deviation of the spectral distribution relative to the reference spectrum. When the joint anomaly reaches the alarm condition, the updates of the dynamic baseline, dynamic standard deviation and reference spectrum are frozen and a local alarm is triggered. An alarm context data segment is formed based on the microtube laminar gas measurement sequence. The freeze is lifted when the joint anomaly meets the recovery condition. The standard volumetric flow rate, combined anomaly data, local alarm information, and alarm context data segment are written to the local storage unit and output to the host computer system via the edge gateway.
[0010] As a preferred embodiment of the present invention, the formation of the microtube laminar gas measurement sequence includes: The analog differential pressure signal, analog absolute pressure signal, and analog temperature signal output by the microtube laminar flow sensor are conditioned and converted from analog to digital to obtain digital differential pressure signal, digital absolute pressure signal, and digital temperature signal, respectively. Timestamps are assigned to the digital differential pressure signal, digital absolute pressure signal, and digital temperature signal according to a unified sampling clock. The time-aligned data is written into a circular buffer, and differential pressure data blocks, absolute pressure data blocks, and temperature data blocks are generated according to the processing cycle. The differential pressure data blocks are used for sensor outlier identification and sliding filtering, while the absolute pressure data blocks and temperature data blocks are used to obtain the average absolute pressure and average thermodynamic temperature of the processing cycle, respectively. The initialization of the normal monitoring period includes multiple processing cycles. The initial dynamic baseline and initial dynamic standard deviation are established using the standard volumetric flow rate within the initialization of the normal monitoring period, and the reference spectrum is established using the spectral energy distribution vector within the initialization of the normal monitoring period.
[0011] As a preferred embodiment of the present invention, the generation of the spectral energy distribution vector includes: The DC component is removed from the filtered differential pressure signal and then a fast Fourier transform is performed. The maximum spectral peak is searched within the flow analysis frequency band. When the amplitude of the maximum spectral peak reaches the product of the median of the amplitude spectrum of the flow analysis frequency band and the peak prominence coefficient, the frequency corresponding to the maximum spectral peak is determined as the main frequency component. Multiple analysis frequency bands are defined based on the effective response bandwidth of the microtube laminar flow sensor, the pulsation frequency of the gas supply equipment, the valve action frequency, the shortest duration of the target anomaly, and the noise frequency band of the differential pressure sensor. Calculate the proportion of the spectral energy of each analysis frequency band to the total spectral energy of the flow analysis frequency band, and form a spectral energy distribution vector in ascending order of frequency.
[0012] As a preferred embodiment of the present invention, the sliding window length for the next processing cycle is determined according to the following formula:
[0013] in, The length of the sliding window for the next processing cycle; This is the upper limit of the sliding window length; This is the lower limit of the sliding window length; This is the window adjustment coefficient; The sampling frequency for the differential pressure signal; This refers to the main frequency component of the current processing cycle; For processing cycle number; This is for calculating the minimum value; This is for calculating the maximum value. This is for floor function.
[0014] As a preferred embodiment of the present invention, sensor outlier identification and sliding filtering include: Calculate the center value and dispersion of the differential pressure signal within the sliding window; When the deviation of a differential pressure sample value from the center value reaches the outlier determination threshold, the corresponding differential pressure sample value is written to the pending confirmation buffer position. Continue to receive subsequent differential pressure sampling values. When the subsequent differential pressure sampling values return to the normal fluctuation range, the differential pressure sampling value in the buffer position to be confirmed is determined as the sensor outlier point, and linear interpolation is performed based on the nearest normal differential pressure sampling value in front of and behind the sensor outlier point. When the number of consecutive exceedance points reaches the consecutive point threshold, the corresponding differential pressure sample value is retained as the actual operating condition change data and entered into subsequent flow calculation and spectrum transformation; The differential pressure signal after sensor outlier processing is subjected to sliding filtering.
[0015] As a preferred embodiment of the present invention, the proportion of high-frequency disturbance energy is obtained by adding the normalized spectral energy of the mid-to-high frequency disturbance band to the normalized spectral energy of the high-frequency noise band. When the maximum spectral peak amplitude within the flow regime analysis band is lower than the product of the median of the amplitude spectrum and the peak prominence coefficient, the current processing cycle is marked as invalid for the main frequency component, and the next processing cycle maintains the sliding window length of the current processing cycle.
[0016] As a preferred embodiment of the present invention, the forgetting coefficient of the dynamic baseline, the dynamic baseline, and the dynamic standard deviation are derived according to the following formula:
[0017] in, This represents the forgetting coefficient for the current processing cycle; This represents the upper limit of the forgetting coefficient. This is the lower limit of the forgetting coefficient; Basic forgetting coefficient; A positive spectral disturbance adjustment coefficient; This represents the percentage of high-frequency disturbance energy in the current processing cycle. This serves as the dynamic baseline for the current processing cycle; This serves as the dynamic baseline for the previous processing cycle; This is the standard volumetric flow rate for the current processing cycle; This represents the dynamic standard deviation of the current processing cycle. This represents the dynamic standard deviation of the previous processing cycle; This is for calculating the minimum value; This is for calculating the maximum value. This is for square root operations.
[0018] As a preferred embodiment of the present invention, the degree of deviation of the spectral distribution and the joint anomaly are calculated according to the following formula:
[0019] in, The degree of deviation in the spectral distribution during the current processing cycle; To analyze the total number of frequency bands; To analyze the frequency band number; For the current processing cycle number Normalized spectral energy of each analysis frequency band; The first reference spectrum in the previous processing cycle The reference energy ratio of each analysis frequency band; This represents the combined anomaly quantity for the current processing cycle; This represents the percentage of high-frequency disturbance energy in the current processing cycle. This is the standard volumetric flow rate for the current processing cycle; This serves as the dynamic baseline for the previous processing cycle; This represents the dynamic standard deviation of the previous processing cycle; To prevent positive stable terms with a denominator of zero, and It has the same units as the dynamic standard deviation; For summation operations; This is for absolute value operations; When the number of processing cycles in which the combined anomaly continuously reaches the alarm threshold reaches the alarm duration cycle, the updates of the dynamic baseline, dynamic standard deviation, and reference spectrum are frozen and a local alarm is triggered. Continue to calculate standard volumetric flow rate, spectral energy distribution vector, high-frequency disturbance energy ratio, spectral distribution deviation, and joint anomaly under frozen conditions; The freeze is lifted when the number of processing cycles in which the combined abnormal value is continuously below the recovery threshold reaches the recovery duration cycle number, where the recovery threshold is lower than the alarm threshold.
[0020] As a preferred embodiment of the present invention, the generation of standard volumetric flow rate includes: Based on the temperature and absolute pressure signals, the actual dynamic viscosity and actual compressibility factor are obtained from the gas property parameter table corresponding to the type of gas being measured. The reference flow rate coefficient is corrected based on the ratio of the reference dynamic viscosity to the actual dynamic viscosity, and the actual volumetric flow rate is obtained based on the corrected flow rate coefficient and the average pressure difference during the treatment cycle. The standard volumetric flow rate is obtained by standard state conversion based on the actual volumetric flow rate, the average absolute pressure during the treatment cycle, the reference pressure, the average thermodynamic temperature during the treatment cycle, the reference thermodynamic temperature, the actual compressibility factor, and the reference compressibility factor.
[0021] As a preferred embodiment of the present invention, the edge computing node includes a processor, running memory, non-volatile storage medium, data acquisition interface and network interface; Cyclic buffer processing, sliding filtering, spectrum transformation, standard volume flow rate calculation, dynamic baseline update, spectrum distribution deviation calculation, and joint anomaly calculation are all performed locally on the edge computing node; Edge computing nodes run data acquisition tasks, real-time computing tasks, local storage tasks, and data upload tasks according to task priority, with data acquisition tasks having higher priority than data upload tasks; During network connection interruption, the standard volume traffic to be uploaded, joint anomaly data, local alarm information and alarm context data segment are written into the local transmission queue. After the network connection is restored, transmission continues according to the data sequence number.
[0022] As a preferred technical solution of the present invention, the microtube laminar gas small flow rate data analysis method based on edge computing is used for monitoring the special gas micro-transmission branch of semiconductor manufacturing equipment. The gas physical property parameter table establishes corresponding temperature index areas and absolute pressure index areas according to the types of special gases; Based on standardized flow deviation, spectral distribution deviation, and high-frequency disturbance energy ratio, local alarm information is classified into flow deviation anomaly, periodic pulsation anomaly, and mid-to-high frequency disturbance anomaly, and output to the gas delivery monitoring system through the edge gateway.
[0023] As a preferred technical solution of the present invention, the microtube laminar gas small flow rate data analysis method based on edge computing is used for monitoring the micro gas branch of medical gas equipment. Based on the known gas component ratios, average thermodynamic temperature during the treatment cycle, and average absolute pressure during the treatment cycle, the equivalent dynamic viscosity and compressibility factor are obtained from the mixed gas property parameter table. The low-frequency variation band and main pulsation band corresponding to the normal respiratory cycle are retained in the spectral energy distribution vector, and the high-frequency disturbance energy ratio is generated by using the mid-to-high frequency disturbance band and high-frequency noise band located outside the normal respiratory cycle band.
[0024] In addition, the present invention also discloses a microtube laminar gas low-flow data analysis system, including a microtube laminar flow sensor, a signal conditioning and analog-to-digital conversion module, an edge computing node, a local storage unit, an alarm output interface, and an edge gateway; The microtube laminar flow sensor includes a differential pressure sensor, an absolute pressure sensor, and a temperature sensor. The differential pressure sensor, absolute pressure sensor, and temperature sensor output differential pressure signal, absolute pressure signal, and temperature signal to the signal conditioning and analog-to-digital conversion module, respectively. The signal conditioning and analog-to-digital conversion module outputs the digitized differential pressure signal, absolute pressure signal, and temperature signal to the edge computing node; The edge computing node includes a circular buffer processing module, an adaptive window filtering module, a traffic calculation module, a spectrum feature processing module, a dynamic baseline processing module, and a joint anomaly detection module; The circular buffer processing module outputs differential pressure data blocks to the adaptive window filtering module; The spectrum feature processing module outputs the main frequency component of the previous processing cycle to the adaptive window filtering module. The adaptive window filtering module outputs the filtered differential pressure signal to the flow calculation module and the spectrum feature processing module. The flow calculation module receives absolute pressure and temperature signals and outputs standard volume flow to the dynamic baseline processing module, the joint anomaly detection module, the local storage unit, and the edge gateway. The spectrum feature processing module outputs the high-frequency disturbance energy ratio to the dynamic baseline processing module, the spectrum energy distribution vector and the degree of spectrum distribution deviation to the joint anomaly determination module, and the spectrum energy distribution vector to the edge gateway. The dynamic baseline processing module outputs the dynamic baseline and dynamic standard deviation to the joint anomaly detection module; The joint anomaly determination module outputs freeze control signals or unfreeze control signals to the dynamic baseline processing module and the spectrum feature processing module based on the joint anomaly quantity, outputs local alarm information to the alarm output interface, and outputs joint anomaly quantity and local alarm information to the local storage unit and the edge gateway. The edge gateway communicates with the host computer system to receive standard volumetric flow rate, spectral energy distribution vector, joint anomaly quantity and local alarm information output by the edge computing node, and transmits them to the host computer system.
[0025] Compared with the prior art, the present invention has the following beneficial effects: This invention establishes a cross-processing cycle feedback relationship between the differential pressure signal spectrum state and subsequent filtering scales, allowing the dominant frequency component of the current processing cycle to determine the sliding window length for the next processing cycle. The fixed sliding window is replaced by an adaptive sliding window constrained by the dominant frequency component and having upper and lower limits, thereby reducing the attenuation of the true pulsation component by an excessively long sliding window and avoiding the retention of excessive random noise due to an excessively short sliding window. This invention distinguishes between sensor outliers and actual operating condition changes. Isolated spikes undergo linear interpolation after passing through a confirmation buffer, while data that continuously exceeds the limit to a certain threshold is retained for flow calculation, spectrum transformation, and anomaly detection. Therefore, the preprocessing process does not simply delete all out-of-limit signals. This invention does not set up spectral feature extraction and dynamic baseline detection as independent processing branches. The spectral energy distribution vector forms the high-frequency disturbance energy proportion, which further adjusts the dynamic baseline forgetting coefficient and participates in the joint anomaly calculation. The spectral analysis output is transformed from a display result into a control variable for subsequent statistical detection processes.
[0026] This invention freezes the updates of dynamic baseline, dynamic standard deviation, and reference spectrum after alarm conditions are met, reducing baseline migration caused by continuous abnormal samples entering the normal reference state. During anomalies, joint anomaly calculations continue; after the anomaly disappears, the freeze is lifted based on a recovery threshold below the alarm threshold and a recovery duration condition, allowing for adaptive updates to resume in the normal state. The entire real-time calculation process can be executed at edge computing nodes in the measurement field. During network interruptions, data to be uploaded is stored in a local transmission queue, and transmission continues according to the data sequence number after network recovery. Therefore, real-time anomaly detection does not rely on real-time responses from remote servers and maintains data continuity during network interruptions. Attached Figure Description
[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.
[0028] Figure 1 is the main flowchart of the microtube laminar gas low flow rate data analysis of the present invention.
[0029] Figure 2 This is a flowchart of the adaptive sliding window update and sensor outlier identification of the present invention.
[0030] Figure 3 This is a schematic diagram illustrating the standard volumetric flow rate calculation and data flow of this invention.
[0031] Figure 4 This is a schematic diagram of the micro-volume delivery branch monitoring of special gases in semiconductor manufacturing equipment according to the present invention.
[0032] Figure 5 is a flowchart of the dynamic baseline adaptive update sub-process of the present invention.
[0033] Figure 6 is a flowchart of the edge node local storage and disconnected transmission sub-process of the present invention. Detailed Implementation
[0034] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0035] The following is in conjunction with the appendix Figures 1-6 The embodiments of the present invention will be described in detail below.
[0036] This invention processes gas flow measurement signals output by a microtube laminar flow sensor. The microtube laminar flow sensor includes at least a microtube channel, a laminar flow element, a differential pressure sensor, an absolute pressure sensor, and a temperature sensor. The differential pressure sensor obtains the pressure difference between the upstream and downstream sides of the laminar flow element. The absolute pressure sensor obtains the absolute pressure at the flow rate conversion location. The temperature sensor obtains the gas temperature at the flow rate conversion location. The signal conditioning and analog-to-digital conversion module performs range matching, anti-aliasing filtering, and analog-to-digital conversion of the analog signal.
[0037] Edge computing nodes read data according to processing cycles. Within each processing cycle, they complete the construction of differential pressure data blocks, absolute pressure data blocks, and temperature data blocks; sensor outlier identification; sliding filtering; actual volumetric flow rate calculation; standard volumetric flow rate conversion; spectral feature extraction; dynamic baseline processing; joint anomaly calculation; and status determination. Data feedback relationships are established between different processing cycles through the dominant frequency component and the frozen state. In the first formal online monitoring processing cycle after the initial normal monitoring period ends, the edge computing nodes use the initial sliding window length.
[0038] After completing the spectral analysis of the first formal online monitoring processing cycle, the dominant frequency component of the first formal online monitoring processing cycle is obtained. Based on this dominant frequency component, the sliding window length for the second formal online monitoring processing cycle is determined. In subsequent online monitoring processes, the dominant frequency component of the previous processing cycle is always used to determine the sliding window length of the current processing cycle, and the dominant frequency component of the current processing cycle is used to determine the sliding window length of the next processing cycle.
[0039] Under normal update conditions, the dynamic baseline, dynamic standard deviation, and reference spectrum are updated recursively based on normal data. Under suspected anomaly conditions, updates to the dynamic baseline, dynamic standard deviation, and reference spectrum are paused, and the number of suspected anomaly duration cycles is accumulated. Under alarm conditions, updates to the dynamic baseline, dynamic standard deviation, and reference spectrum are frozen. During the frozen state, standard volumetric flow rate, spectral energy distribution vector, high-frequency disturbance energy proportion, spectral distribution deviation, and joint anomaly quantities continue to be calculated. The freeze is lifted once the recovery conditions are met.
[0040] Specifically, the acquisition of the microtube laminar gas measurement sequence is as follows: The microtube channel size is set according to the target flow range, the dynamic viscosity of the measured gas, the allowable pressure loss, and the target Reynolds number range. The laminar flow element can be a capillary bundle, a parallel microchannel structure, or a laminar flow structure capable of establishing a stable laminar pressure drop relationship within the target flow range. The two pressure ports of the differential pressure sensor are connected to the upstream and downstream positions of the laminar flow element, respectively. The absolute pressure sensor is positioned to reflect the pressure converted from the actual flow rate. The temperature sensor is positioned to reflect the actual thermodynamic temperature of the gas.
[0041] In one implementation, the differential pressure signal sampling frequency can be from 100 Hz to 2000 Hz, with the specific sampling frequency determined based on the differential pressure sensor's response bandwidth, the highest effective flow regime change frequency to be detected, and the processing capability of the edge computing nodes. The differential pressure signal sampling frequency should be higher than twice the highest effective frequency to be analyzed, with a buffer band for anti-aliasing filtering. Absolute pressure and temperature signals are allowed to use sampling frequencies lower than those for the differential pressure signal.
[0042] Edge computing nodes employ either the most recently valid sample value hold or time interpolation at the processing cycle boundaries to establish a time correspondence between absolute pressure and temperature signals and differential pressure data blocks. Edge computing nodes maintain raw data circular buffers and processed data circular buffers. The raw data circular buffer stores the raw sample values before outlier replacement and filtering. The processed data circular buffer stores data after sensor outlier identification and sliding filtering. Alarm context data segments are extracted from the raw data circular buffer, thus preserving the original signal variations.
[0043] Furthermore, before formal online monitoring begins, a normal monitoring period needs to be initialized.
[0044] After the microtube laminar flow sensor is installed, a zero-point check and normal flow status confirmation are performed first. The initial normal monitoring period begins when the gas type is configured, the delivery status is stable, and there are no alarms.
[0045] In one implementation, the initialization of the normal monitoring period comprises 60 processing cycles, each 2 seconds long, lasting 120 seconds. Edge computing nodes calculate the standard volumetric flow rate for each of the 60 processing cycles. The arithmetic mean of these 60 standard volumetric flow rates is used as the initial dynamic baseline, and the standard deviation of these 60 standard volumetric flow rates relative to the initial dynamic baseline is used as the initial dynamic standard deviation. Simultaneously, the spectral energy distribution vector is calculated for each of the 60 processing cycles. For each analysis frequency band, the normalized spectral energy of the corresponding analysis frequency band is averaged across the 60 processing cycles. After averaging across all analysis frequency bands, the average results for each analysis frequency band are normalized again to ensure that the sum of all reference energy proportions is 1, thus obtaining the reference spectrum. After the initialization of the normal monitoring period ends, the initial dynamic baseline, initial dynamic standard deviation, and reference spectrum are written to the state parameter area.
[0046] Sensor outlier identification is used to handle isolated electrical spikes, transient jumps in analog-to-digital conversion, or single-point errors in communication data; it is not used to remove continuous real-time flow variations. Edge computing nodes calculate the center value and dispersion of the differential pressure signal within the current sliding window. The center value can be the median, and the dispersion can be the standard deviation or median absolute deviation. The outlier determination threshold is determined based on the dispersion of the differential pressure signal within the current sliding window; the normal fluctuation range is the area centered on the center value and bounded by the outlier determination threshold.
[0047] When the deviation of a differential pressure sample value from the center value exceeds the outlier detection threshold, the corresponding differential pressure sample value is not immediately deleted. Instead, it is written to the pending confirmation buffer. The edge computing node continues to read subsequent differential pressure sample values. If the subsequent differential pressure sample values return to the normal fluctuation range, and there is no continuous unidirectional change between the preceding and following normal differential pressure sample values, then the differential pressure sample value in the pending confirmation buffer is identified as a sensor outlier point. A linear interpolation based on time distance is then performed using the most recent preceding and following normal differential pressure sample values. Let the most recent preceding normal differential pressure sample value be... The corresponding sampling time is The nearest normal differential pressure sample value behind the sensor outlier point is The corresponding sampling time is The sampling time to be interpolated is The interpolated differential pressure sample value is ,but:
[0048] When the number of consecutive exceedances reaches the consecutive point threshold, linear interpolation is not performed. The consecutive point threshold is determined based on the differential pressure signal sampling frequency and the shortest duration of the actual operating condition change that needs to be retained. For example, if the differential pressure signal sampling frequency is 200 Hz and the actual operating condition change that needs to be retained for a duration of not less than 20 ms needs to be retained, the consecutive point threshold can be set to 4 sampling points. When 4 consecutive sampling points exceed the limit, the corresponding data is retained for sliding filtering, flow calculation, and spectrum transformation.
[0049] The length of the adaptive sliding window is determined by the dominant frequency component of the previous processing cycle, and the length of the sliding window in the next processing cycle is calculated according to the following formula:
[0050] Among them, the window adjustment coefficient This is used to determine the proportion of the sliding window relative to the main fluctuation period. In monitoring scenarios where it is necessary to preserve the characteristics of flow pulsation, it can be set between 0.20 and 0.50. Lower limit of sliding window length. To prevent excessive residual random noise caused by excessively short windows, the upper limit... This is used to avoid a decrease in dynamic response caused by an excessively long window. For example, if the differential pressure signal sampling frequency is 200 Hz, the main frequency component in the current processing cycle is 5 Hz, the window adjustment coefficient is 0.25, the lower limit of the sliding window length is 5 sampling points, and the upper limit of the sliding window length is 80 sampling points, then the sliding window length in the next processing cycle will be 10 sampling points. After the sliding filter is completed in the current processing cycle, the filtered differential pressure signal is obtained, which is then simultaneously used in the flow calculation process and the spectral feature extraction process.
[0051] Furthermore, the edge computing nodes remove the DC component from the filtered differential pressure signal and perform a fast Fourier transform; the flow analysis frequency band does not include the zero-frequency DC component.
[0052] The maximum spectral peak is searched within the flow analysis frequency band, and the median of the amplitude spectrum of the flow analysis frequency band is calculated. When the amplitude of the maximum spectral peak reaches the product of the amplitude spectrum median and the peak prominence coefficient, the maximum spectral peak is confirmed as valid, and the frequency containing the maximum spectral peak is determined as the dominant frequency component; otherwise, the current processing cycle is marked as invalid for the dominant frequency component, and the next processing cycle maintains the sliding window length of the current processing cycle to avoid invalid frequencies close to zero from participating in the window length calculation. The peak prominence coefficient is determined based on normal operating condition calibration data and noise floor distribution. In one implementation, it can be set to 3 to 8.
[0053] For calculating actual and standard volumetric flow rates, gas property parameter tables are established separately for each gas type, using at least temperature and absolute pressure as dual indices, and storing dynamic viscosity and compressibility factor. When the actual temperature and actual absolute pressure fall between adjacent parameter nodes, bilinear interpolation is used to obtain the actual dynamic viscosity and actual compressibility factor. The reference operating condition flow coefficient is obtained through reference operating condition calibration of the microtube laminar flow sensor. The edge computing node multiplies the reference operating condition flow coefficient by the ratio of the reference dynamic viscosity to the actual dynamic viscosity to obtain the current operating condition flow coefficient, and then multiplies it by the average pressure difference of the processing cycle to obtain the actual volumetric flow rate of the current processing cycle. The actual volumetric flow rate is further converted into standard volumetric flow rate:
[0054] in, , These are the standard volumetric flow rate and the actual volumetric flow rate for the current processing cycle, respectively. To handle the cycle mean absolute pressure, For reference pressure, For reference thermodynamic temperature, To handle the average thermodynamic temperature over the cycle, , These are the reference compressibility factor and the actual compressibility factor, respectively. For low-pressure conditions approaching ideal gas states, the compressibility factor is close to one; for conditions with a large pressure variation range, the actual compressibility factor should be obtained from the gas property parameter table without omitting corrections.
[0055] Regarding the spectral energy distribution vector, the flow analysis frequency band is set according to the specific monitoring object. The frequency band boundary should at least consider the effective response bandwidth of the microtube laminar flow sensor, the basic pulsation frequency of the gas supply equipment, the operating frequency of the control valve, the shortest duration of the target anomaly, and the main distribution frequency band of the differential pressure sensor noise.
[0056] For example, in an implementation of a semiconductor specialty gas micro-transmission branch with a differential pressure signal sampling frequency of 200 Hz, the flow analysis frequency band can be divided into a low-frequency variation band of 0.5 Hz to 3 Hz, a main pulsation band of 3 Hz to 8 Hz, a mid-to-high frequency disturbance band of 8 Hz to 25 Hz, and a high-frequency noise band of 25 Hz to 100 Hz. Edge computing nodes accumulate the power of the frequency units contained in each analysis band. The spectral energy of each analysis band is divided by the total spectral energy of the flow analysis frequency band to obtain the normalized spectral energy of the corresponding analysis band, and these are arranged in ascending order of frequency to form a spectral energy distribution vector.
[0057] The normalized spectral energy of the mid-to-high frequency disturbance band is added to the normalized spectral energy of the high-frequency noise band to form the high-frequency disturbance energy proportion. The value of the high-frequency disturbance energy ratio is between 0 and 1.
[0058] The dynamic baseline and dynamic standard deviation are updated recursively using a variable forgetting coefficient. The proportion of high-frequency perturbation energy is used to adjust the forgetting coefficient, specifically recursively calculated as follows:
[0059] Among them, the basic forgetting coefficient Based on the rate of change under normal operating conditions, for applications with processing cycles shorter than 2 seconds and slow changes in normal operating conditions, the value can be set between 0.90 and 0.98; Spectrum disturbance adjustment coefficient. The value is positive, determined based on the range of change in the proportion of high-frequency disturbance energy between normal and abnormal disturbance states. As the proportion of high-frequency disturbance energy increases, the forgetting coefficient increases, and the proportion of the current standard volumetric flow rate entering the dynamic baseline decreases, thus slowing down the dynamic baseline's tracking speed in response to disturbance states.
[0060] The reference spectrum is established during the initial normal monitoring period and updated via a band-by-band recursive method under normal online update conditions. For each analysis band, the reference energy ratio of the previous processing cycle is multiplied by the historical reference weight, and the normalized spectral energy of the corresponding analysis band in the current processing cycle is multiplied by one minus the historical reference weight. The two are added together to obtain the current candidate reference energy ratio. After all analysis bands have been recursively updated, the candidate reference energy ratios are renormalized to form the updated reference spectrum. The historical reference weight can be set between 0.90 and 0.99. The reference spectrum is not updated under suspected abnormal conditions and alarm freeze conditions.
[0061] The joint anomaly is synthesized by weighting the standardized flow deviation with the degree of deviation from the spectral distribution. The formulas for calculating the degree of deviation from the spectral distribution and the joint anomaly are as follows:
[0062] in, The degree of deviation in the spectral distribution during the current processing cycle; To analyze the total number of frequency bands; To analyze the frequency band number; For the current processing cycle number Normalized spectral energy of each analysis frequency band; The first reference spectrum in the previous processing cycle The reference energy ratio of each analysis frequency band; This represents the combined anomaly quantity for the current processing cycle; This represents the percentage of high-frequency disturbance energy in the current processing cycle. This is the standard volumetric flow rate for the current processing cycle; This serves as the dynamic baseline for the previous processing cycle; This represents the dynamic standard deviation of the previous processing cycle; To prevent positive stable terms with a denominator of zero, and It has the same units as the dynamic standard deviation.
[0063] Regarding anomaly detection and status management, normal update thresholds, alarm thresholds, and recovery thresholds are set, with the recovery threshold lower than the normal update threshold and the normal update threshold lower than the alarm threshold. When the combined anomaly quantity is lower than the normal update threshold, updates to the dynamic baseline, dynamic standard deviation, and reference spectrum are allowed. When the normal update threshold is reached but the alarm threshold is not, a suspected anomaly state is entered, normal reference status updates are suspended, and the number of suspected anomaly duration cycles is accumulated. When the alarm threshold is reached consecutively and the number of consecutive processing cycles reaches the alarm duration cycle count, an alarm freeze state is entered. The dynamic baseline, dynamic standard deviation, and reference spectrum retain their values before entering the alarm freeze state, but outlier processing, sliding filtering, flow calculation, spectral feature extraction, and combined anomaly quantity calculation continue to be performed. When the combined anomaly quantity is continuously lower than the recovery threshold and the number of consecutive processing cycles reaches the recovery duration cycle count, the alarm freeze state is lifted, and subsequent processing cycles re-participate in the updates of the dynamic baseline, dynamic standard deviation, and reference spectrum. The above thresholds and duration cycle counts are calibrated based on the normal data fluctuation range, the allowable false alarm frequency, and the shortest duration of the target anomaly.
[0064] Preferably, an edge computing node includes at least a processor, RAM, non-volatile storage, a data acquisition interface, and a network interface. Internally, it includes a raw data circular buffer, a processed data circular buffer, a status parameter area, an alarm context buffer, and a local transmission queue. The data acquisition task reads digital differential pressure signals, digital absolute pressure signals, and digital temperature signals; the real-time computing task executes the core algorithm; the local storage task saves standard volumetric flow rate, joint anomalies, local alarm information, and alarm context data segments; and the data upload task writes the data to be uploaded to the edge gateway's transmission channel. The data acquisition task has a higher priority than the data upload task. When the network connection is interrupted, the data to be uploaded is written to the local transmission queue. Each set of data includes a data sequence number and a timestamp. After the network connection is restored, transmission continues from the position of the data sequence number that was not successfully uploaded. When an alarm occurs, the data segments before the alarm, the alarm duration, and the alarm after the alarm are extracted from the raw data circular buffer to form the alarm context data segment. The edge computing node can be configured with a local web data viewing function, but this function does not participate in adaptive sliding window, dynamic baseline updates, or joint anomaly calculation.
[0065] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.
[0066] Example 1: In this example, the differential pressure signal sampling frequency is set to 200 Hz, the processing cycle length is 2 s, and each processing cycle contains 400 differential pressure sampling values. The initial normal monitoring period contains 60 processing cycles, based on which the initial dynamic baseline, initial dynamic standard deviation, and reference spectrum are established.
[0067] The first formal online monitoring and processing cycle used an initial sliding window length of 16 sampling points. After completing the spectrum analysis, a 5 Hz effective main frequency component was obtained within the flow regime analysis frequency band. The window adjustment coefficient was set to 0.25, with a lower limit of 5 sampling points and an upper limit of 80 sampling points for the sliding window length. Therefore, the sliding window length for the second formal online monitoring and processing cycle was determined to be 10 sampling points.
[0068] In this embodiment, the flow analysis frequency band is divided into 4 analysis frequency bands, and the spectral energy distribution vector of a normal processing cycle is: The high-frequency perturbation energy ratio is 0.080. The basic forgetting coefficient is set to 0.92, the spectral perturbation adjustment coefficient is 0.10, the forgetting coefficient is limited from 0.90 to 0.99, and the current forgetting coefficient is 0.928.
[0069] The dynamic baseline of the previous processing cycle was 49.4414 mL / min, the current standard volumetric flow rate was 49.5650 mL / min, and the dynamic standard deviation was 0.0800 mL / min. Since the spectral energy distribution vector was the same as the reference spectrum, the spectral distribution deviation was 0, the standardized flow rate deviation was 1.545, and the joint anomaly was 1.4214. The normal update threshold was set to 1.5. Since the current joint anomaly was below the threshold, an update was allowed. After the update, the dynamic baseline was 49.4503 mL / min, and the dynamic standard deviation was 0.0839 mL / min. This embodiment fully demonstrates the data closed loop of main frequency component feedback adjustment window, high-frequency disturbance adjustment forgetting coefficient, and joint anomaly control reference state update.
[0070] Example 2: This example is a further optimization of Example 1 for monitoring the micro-nitrogen delivery branch in semiconductor manufacturing equipment, as detailed below: The microtube laminar flow sensor has a measurement range of 0 to 500 mL / min, a differential pressure sampling frequency of 500 Hz, and a processing cycle of 2 s. The gas property parameter table is based on nitrogen gas, and real-time properties are obtained using bilinear interpolation.
[0071] In this embodiment, local alarms are divided into flow offset anomalies, periodic pulsation anomalies, and mid-to-high frequency disturbance anomalies. The classification is based on the degree of deviation of the standardized flow deviation, the energy ratio of the main pulsation frequency band from the reference spectrum, and whether the energy ratio of the high frequency disturbance reaches the corresponding judgment threshold. This classification is only used for alarm information identification and does not change the freezing process of the joint anomaly.
[0072] In one calculation cycle, the normal reference spectrum is After the disturbance, the spectrum vector becomes The proportion of high-frequency disturbance energy increased to 0.1382, the forgetting coefficient increased accordingly, and the dynamic baseline following speed decreased. After the alarm is triggered, the edge computing node saves the original data and key feature quantities before and after the alarm, and sends the alarm information to the gas transport monitoring system through the edge gateway.
[0073] Example 3: This example provides an implementation method for a micro-gas branch circuit in a medical gas device, as detailed below: In this embodiment, the microtube laminar flow sensor is installed in the low-flow branch to be monitored, and the component ratio of the mixed gas is provided by the device configuration parameters. A table of mixed gas physical property parameters is established for a fixed component ratio range, and the equivalent dynamic viscosity and compressibility factor are stored using component ratio range, temperature, and absolute pressure as indexes. The current operating condition parameters are obtained through multidimensional interpolation.
[0074] The normal respiratory cycle components are the real signals that need to be preserved. The flow regime analysis band sets the normal respiratory frequency range within the low-frequency variation band and the main pulsation band, while the mid-to-high frequency disturbance band and the high-frequency noise band are used to generate the high-frequency disturbance energy ratio.
[0075] This embodiment follows the same main frequency feedback, adaptive window, forgetting coefficient adjustment, joint anomaly determination, and freeze-recovery closed loop as Embodiment 1.
[0076] Comparative Example 1: Used to compare the effects of a fixed sliding window and the adaptive window of this invention. Comparative Example 1 uses the exact same parameter configuration and input signal as Example 1, the only difference being that it does not use the main frequency component feedback, and always uses a fixed arithmetic moving average window with 100 sampling points. The input differential pressure signal is a deterministic multi-frequency signal:
[0077] During the initial normal monitoring period, only a 5Hz main pulsating flow state exists, while 11Hz and 31Hz are anomalous disturbances. The normal reference spectrum established accordingly is... After processing using a sliding window with 10 sampling points, the spectral energy distribution vector is: The deviation was 0.3125902; while after processing with a fixed window of 100 sampling points, the spectrum vector was... The deviation was 0.1734289, a decrease of approximately 44.5%. The results indicate that a fixed long window attenuates dynamic frequency components, thereby reducing the detection sensitivity for anomalous spectral shifts.
[0078] Comparative Example 2: This comparative example is used to illustrate the necessity of alarm freezing. Comparative Example 2 has the same configuration as Example 1, but alarm freezing is canceled, that is, the dynamic baseline and dynamic standard deviation continue to be updated after the combined abnormal quantity meets the alarm conditions.
[0079] The dynamic baseline before the anomaly was set at 49.4414 mL / min, with a dynamic standard deviation of 0.0800 mL / min. Under continuous anomaly, the standard volumetric flow rate for each cycle was 49.8122 mL / min, and the forgetting coefficient was maintained at 0.95.
[0080] Recursively, the standardized flow deviations at the start of the 1st to 5th anomaly handling cycles were 4.6350, 3.0949, 2.4596, 2.0879, and 1.8353, respectively, showing a clear decreasing trend. Example 1, by freezing the baseline, prevented the normal reference state from migrating to outliers. This demonstrates that continuously entering updates with outlier samples gradually contaminates the reference baseline.
[0081] To facilitate further understanding in this field, the complete data chain of this invention is presented below, and continuous numerical extrapolation is performed.
[0082] Specifically as follows: The differential pressure sampling frequency is 200 Hz, the processing cycle includes 400 sampling points, and the effective dominant frequency component is 5 Hz. Based on this, the next cycle window length is calculated to be 10 sampling points. After flow rate conversion, the standard volumetric flow rate is 49.4414 mL / min, and the spectral energy distribution vector at this time is... The high-frequency perturbation energy percentage was 0.0812, and the forgetting coefficient was 0.92812. The standard volumetric flow rate for the next cycle became 49.5650 mL / min. With a deviation of 0, the standardized flow rate deviation was 1.545, and the combined outlier was 1.4214, below the normal update threshold of 1.5. Therefore, the baseline was updated to 49.4503 mL / min, with a standard deviation of 0.0839 mL / min. Subsequently, the standard volumetric flow rate increased to 49.8122 mL / min, and the spectral vector changed... The high-frequency disturbance energy percentage is 0.3476, the spectral distribution deviation is 0.5344, the standardized flow deviation is 4.6351, and the combined anomaly is 3.2097. The alarm threshold is set to 2.2, and the baseline is frozen and an alarm is triggered after three consecutive cycles of exceeding the limit.
[0083] During the freeze period, all parameters continue to be calculated. Once the anomaly is resolved, the freeze is lifted when the combined anomaly value is below the recovery threshold of 1.0 for five consecutive cycles, and normal updates resume. This process fully demonstrates the closed loop from data acquisition, filtering, feedback, anomaly detection to freeze recovery.
[0084] Example 4: This example discloses a microtube laminar gas low-flow data analysis system, including a microtube laminar flow sensor, a signal conditioning and analog-to-digital conversion module, an edge computing node, a local storage unit, an alarm output interface, and an edge gateway.
[0085] The differential pressure sensor, absolute pressure sensor, and temperature sensor in the microtube laminar flow sensor output signals to the signal conditioning and analog-to-digital conversion module, and then enter the edge computing node after being digitized.
[0086] The edge computing node integrates a circular buffer processing module, an adaptive window filtering module, a flow calculation module, a spectrum feature processing module, a dynamic baseline processing module, and a joint anomaly detection module. The circular buffer processing module organizes differential pressure data blocks; the spectrum feature processing module outputs the main frequency component of the previous processing cycle to the adaptive window filtering module; the adaptive window filtering module outputs the filtered differential pressure signal to the flow calculation module and the spectrum feature processing module; the flow calculation module outputs the standard volumetric flow rate based on physical property parameters; the spectrum feature processing module outputs the main frequency component, the spectrum energy distribution vector, and the proportion of high-frequency disturbance energy; the dynamic baseline processing module outputs the dynamic baseline and dynamic standard deviation based on the forgetting coefficient; and the joint anomaly detection module outputs freeze or unfreeze control signals based on the joint anomaly quantity and outputs local alarm information to the alarm output interface, local storage unit, and edge gateway. The edge gateway communicates with the host computer system, transmitting the standard volumetric flow rate, spectrum energy distribution vector, joint anomaly quantity, and alarm information. Each module in the system has a clearly defined input source, output destination, and data interaction relationship.
[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A microtube laminar gas low-flow-rate data analysis method based on edge computing, characterized by including: The differential pressure signal, absolute pressure signal, and temperature signal output by the microtube laminar flow sensor are collected to form a microtube laminar flow gas measurement sequence with a unified timestamp; During the initial normal monitoring period, establish the initial dynamic baseline, initial dynamic standard deviation, and reference spectrum; The edge computing node determines the sliding window length of the current processing cycle based on the main frequency component of the previous processing cycle, and performs sensor outlier identification and sliding filtering on the differential pressure signal to obtain the filtered differential pressure signal. The actual volumetric flow rate is calculated based on the filtered differential pressure signal, and the standard volumetric flow rate is generated by combining the absolute pressure signal, temperature signal and gas physical property parameters. The filtered differential pressure signal is subjected to spectral transformation to obtain the main frequency component and the spectral energy distribution vector. The sliding window length of the next processing cycle is updated based on the main frequency component. The high-frequency disturbance energy ratio is obtained based on the spectral energy distribution vector. The forgetting coefficient of the dynamic baseline is adjusted using the high-frequency disturbance energy ratio. The dynamic baseline and dynamic standard deviation are updated recursively based on the standard volumetric flow rate. The reference spectrum is also updated recursively based on the spectral energy distribution vector. A joint anomaly is generated based on the standardized flow deviation of the standard volume flow rate relative to the dynamic baseline and the degree of deviation of the spectral distribution relative to the reference spectrum. When the joint anomaly reaches the alarm condition, the updates of the dynamic baseline, dynamic standard deviation and reference spectrum are frozen and a local alarm is triggered. An alarm context data segment is formed based on the microtube laminar gas measurement sequence. The freeze is lifted when the joint anomaly meets the recovery condition. The standard volumetric flow rate, combined anomaly data, local alarm information, and alarm context data segment are written to the local storage unit and output to the host computer system via the edge gateway.
2. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 1, characterized in that, The formation of the microtube laminar gas measurement sequence includes: The analog differential pressure signal, analog absolute pressure signal, and analog temperature signal output by the microtube laminar flow sensor are conditioned and converted from analog to digital to obtain digital differential pressure signal, digital absolute pressure signal, and digital temperature signal, respectively. Timestamps are assigned to the digital differential pressure signal, digital absolute pressure signal, and digital temperature signal according to a unified sampling clock. The time-aligned data is written into a circular buffer, and differential pressure data blocks, absolute pressure data blocks, and temperature data blocks are generated according to the processing cycle. The differential pressure data blocks are used for sensor outlier identification and sliding filtering, while the absolute pressure data blocks and temperature data blocks are used to obtain the average absolute pressure and average thermodynamic temperature of the processing cycle, respectively. The initialization of the normal monitoring period includes multiple processing cycles. The initial dynamic baseline and initial dynamic standard deviation are established using the standard volumetric flow rate within the initialization of the normal monitoring period, and the reference spectrum is established using the spectral energy distribution vector within the initialization of the normal monitoring period.
3. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 1, characterized in that, The generation of the spectral energy distribution vector includes: The DC component is removed from the filtered differential pressure signal and then a fast Fourier transform is performed. The maximum spectral peak is searched within the flow analysis frequency band. When the amplitude of the maximum spectral peak reaches the product of the median of the amplitude spectrum of the flow analysis frequency band and the peak prominence coefficient, the frequency corresponding to the maximum spectral peak is determined as the main frequency component. Multiple analysis frequency bands are defined based on the effective response bandwidth of the microtube laminar flow sensor, the pulsation frequency of the gas supply equipment, the valve action frequency, the shortest duration of the target anomaly, and the noise frequency band of the differential pressure sensor. Calculate the proportion of the spectral energy of each analysis frequency band to the total spectral energy of the flow analysis frequency band, and form a spectral energy distribution vector in ascending order of frequency.
4. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 3, characterized in that the sliding window length of the next processing cycle is determined according to the following formula: in, The length of the sliding window for the next processing cycle; This is the upper limit of the sliding window length; This is the lower limit of the sliding window length; This is the window adjustment coefficient; The sampling frequency for the differential pressure signal; This refers to the main frequency component of the current processing cycle; For processing cycle number; This is for calculating the minimum value; This is for calculating the maximum value. This is for floor function.
5. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 1, characterized in that, Sensor outlier identification and sliding filtering include: Calculate the center value and dispersion of the differential pressure signal within the sliding window; When the deviation of a differential pressure sample value from the center value reaches the outlier determination threshold, the corresponding differential pressure sample value is written to the pending confirmation buffer position. Continue to receive subsequent differential pressure sampling values. When the subsequent differential pressure sampling values return to the normal fluctuation range, the differential pressure sampling value in the buffer position to be confirmed is determined as the sensor outlier point, and linear interpolation is performed based on the nearest normal differential pressure sampling value in front of and behind the sensor outlier point. When the number of consecutive exceedance points reaches the consecutive point threshold, the corresponding differential pressure sample value is retained as the actual operating condition change data and entered into subsequent flow calculation and spectrum transformation; The differential pressure signal after sensor outlier processing is subjected to sliding filtering.
6. The microtube laminar gas small flow rate data analysis method based on edge computing according to claim 3, characterized in that the proportion of high frequency disturbance energy is obtained by adding the normalized spectral energy of the mid-to-high frequency disturbance band and the normalized spectral energy of the high frequency noise band; When the maximum spectral peak amplitude within the flow regime analysis band is lower than the product of the median of the amplitude spectrum and the peak prominence coefficient, the current processing cycle is marked as invalid for the main frequency component, and the next processing cycle maintains the sliding window length of the current processing cycle.
7. The microtube laminar gas flow rate data analysis method based on edge computing according to claim 6, characterized in that the forgetting coefficient, dynamic baseline, and dynamic standard deviation of the dynamic baseline are recursively derived according to the following formula: in, This represents the forgetting coefficient for the current processing cycle; This represents the upper limit of the forgetting coefficient. This is the lower limit of the forgetting coefficient; Basic forgetting coefficient; A positive spectral disturbance adjustment coefficient; This represents the percentage of high-frequency disturbance energy in the current processing cycle. This serves as the dynamic baseline for the current processing cycle; This serves as the dynamic baseline for the previous processing cycle; This is the standard volumetric flow rate for the current processing cycle; This represents the dynamic standard deviation of the current processing cycle. This represents the dynamic standard deviation of the previous processing cycle; This is for calculating the minimum value; This is for calculating the maximum value. This is for square root operations.
8. The microtube laminar gas flow rate data analysis method based on edge computing according to claim 7, characterized in that the degree of spectral distribution deviation and the joint anomaly are calculated according to the following formula: in, The degree of deviation in the spectral distribution during the current processing cycle; To analyze the total number of frequency bands; To analyze the frequency band number; For the current processing cycle number Normalized spectral energy of each analysis frequency band; The first reference spectrum in the previous processing cycle The reference energy ratio of each analysis frequency band; This represents the combined anomaly quantity for the current processing cycle; This represents the percentage of high-frequency disturbance energy in the current processing cycle. This is the standard volumetric flow rate for the current processing cycle; This serves as the dynamic baseline for the previous processing cycle; This represents the dynamic standard deviation of the previous processing cycle; To prevent positive stable terms with a denominator of zero, and It has the same units as the dynamic standard deviation; For summation operations; This is for absolute value operations; When the number of processing cycles in which the combined anomaly continuously reaches the alarm threshold reaches the alarm duration cycle, the updates of the dynamic baseline, dynamic standard deviation, and reference spectrum are frozen and a local alarm is triggered. Continue to calculate standard volumetric flow rate, spectral energy distribution vector, high-frequency disturbance energy ratio, spectral distribution deviation, and joint anomaly under frozen conditions; The freeze is lifted when the number of processing cycles in which the combined abnormal value is continuously below the recovery threshold reaches the recovery duration cycle number, where the recovery threshold is lower than the alarm threshold.
9. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 1, characterized in that the generation of standard volumetric flow rate includes: Based on the temperature and absolute pressure signals, the actual dynamic viscosity and actual compressibility factor are obtained from the gas property parameter table corresponding to the type of gas being measured. The reference flow rate coefficient is corrected based on the ratio of the reference dynamic viscosity to the actual dynamic viscosity, and the actual volumetric flow rate is obtained based on the corrected flow rate coefficient and the average pressure difference during the treatment cycle. The standard volumetric flow rate is obtained by standard state conversion based on the actual volumetric flow rate, the average absolute pressure during the treatment cycle, the reference pressure, the average thermodynamic temperature during the treatment cycle, the reference thermodynamic temperature, the actual compressibility factor, and the reference compressibility factor.
10. The microtube laminar gas low-flow-rate data analysis method based on edge computing according to claim 1, characterized in that the edge computing node includes a processor, running memory, non-volatile storage medium, data acquisition interface and network interface; Cyclic buffer processing, sliding filtering, spectrum transformation, standard volume flow rate calculation, dynamic baseline update, spectrum distribution deviation calculation, and joint anomaly calculation are all performed locally on the edge computing node; Edge computing nodes run data acquisition tasks, real-time computing tasks, local storage tasks, and data upload tasks according to task priority, with data acquisition tasks having higher priority than data upload tasks; During network connection interruption, the standard volume traffic to be uploaded, joint anomaly data, local alarm information and alarm context data segment are written into the local transmission queue. After the network connection is restored, transmission continues according to the data sequence number.
