A suspended particle detection system based on multi-feature data analysis

Abstract

This invention discloses a suspended particle detection system based on multi-feature data analysis, belonging to the field of environmental monitoring technology. It includes a sampling module, a main flow channel detection module, a fluid diversion module, an array-type detection module, and a data reconstruction module working collaboratively to achieve high-precision suspended particle detection. The system first performs wide-baseline detection on the full-flux aerosol, then proportionally splits it into multiple parallel airflows, which are then detected by an array-type sensor using dimensionality reduction and high-resolution. The data reconstruction module aggregates the array detection results into baseline data, compares it in real-time with the main flow channel data to determine the overlap distortion state, and extracts the true small-size baseline, applies a probabilistic statistical model to calculate the error correction amount, and performs algebraic subtraction to reconstruct the large-size data. Finally, it fuses and outputs a distribution result that combines the confidence level of large-size particles with the resolution of small-size particles. The system does not require an external dilution gas source, effectively solving the particle overlap error problem in high-concentration environments and improving detection accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a suspended particle detection system based on multi-feature data analysis. Background Technology

[0002] In high-end clean environment monitoring fields such as semiconductor manufacturing, biopharmaceuticals, and precision electronic assembly, high-precision real-time detection of the concentration and particle size distribution of aerosol suspended particles is a key technological link to ensure process yield and product quality. Current mainstream suspended particle detection systems are typically based on the principle of single-channel optical particle counting (OPC), which uses a laser to illuminate an airflow passing through a fixed optically sensitive area and determines the particle size by measuring the amplitude of the scattered light pulse generated by a single particle. However, with increasingly stringent environmental cleanliness requirements in industrial processes and the continuous emergence of new pollution sources, existing technologies have revealed the following significant technical deficiencies and industry pain points in dealing with ultra-high concentration detection, particle physical property characterization, and long-term system reliability: Firstly, existing systems generally employ a design where the full-throughput airflow passes directly through a single optically sensitive area. This design operates well in a cleanroom under steady-state conditions, but when the concentration of the sampled environment surges to tens or even hundreds of millions (e.g., during equipment maintenance or abnormal venting), a large number of particles pass through the sensitive area simultaneously within a very short time interval, causing severe coincidence error. This error directly leads to a significant underestimation of small-diameter particle counts, while the distribution of large-diameter particles is abnormally overestimated due to the spurious merging of multiple small-particle pulses, ultimately resulting in completely distorted particle size distribution data. To mitigate this problem, some high-end systems introduce pneumatic dilution modules, which introduce external pure high-pressure gas to proportionally dilute the sample. However, such solutions not only result in bulky and expensive systems, but the introduction of external gas sources also easily leads to secondary contamination. More importantly, large-diameter particles with high inertia are prone to settling and loss in the bends and splitting structures of the diluter, creating a structural deadlock between "reducing the overlap rate" and "preserving the probability of capturing large particles."

[0003] Secondly, to address the aforementioned overlap error, the industry has proposed software post-processing correction algorithms based on the Poisson statistical model, attempting to mathematically deduce and reconstruct the true particle size distribution. However, the effectiveness of such algorithms heavily relies on a crucial premise—that the spatial distribution of particles within the sensitive region completely follows a uniform random distribution. In complex working conditions with extremely high concentrations and highly dispersed particle size spectra, this premise is often not met. When the overlap distortion rate reaches a high level, the Poisson correction model loses the physical constraint of iterative convergence due to the low signal-to-noise ratio of the input data, leading to divergent correction results and outputting erroneous pseudo-peak data. In other words, traditional software correction schemes are merely a "mathematical game" based on statistical assumptions, lacking an independent and absolutely true physical measurement benchmark unaffected by overlap as a calibration anchor, and thus cannot fundamentally solve the data accuracy problem in extreme environments.

[0004] Third, traditional optical particle counters calibrate equivalent particle size solely based on the intensity of scattered light. This principle has inherent physical limitations: a tiny metal particle and a large organic fiber may produce the exact same light scattering signal at a specific angle. However, in many high-value manufacturing processes (such as lithium-ion battery coating and wafer lithography), the material of the contaminant (inorganic metal shavings or organic shavings) has drastically different impacts on process yield. Existing technologies rely entirely on the single dimension of optical equivalent particle size, lacking the ability to perceive and invert the aerodynamic equivalent density or material characteristics of particles. This means that environmental monitoring personnel can only know "how many particles there are," but not "what these particles are," greatly limiting the efficiency and accuracy of pollution source tracing and investigation.

[0005] Fourth, during long-term continuous operation, mainstream optical sensors inevitably face two major hidden threats: first, the accumulation of trace dust on the surface of optical lenses due to aerosol adhesion causes background baseline drift; second, the semiconductor laser, as the core light source, experiences power attenuation due to aging of the driving current. Existing single-channel detection architectures lack an effective internal redundant reference system, making it impossible to fundamentally distinguish between the aforementioned hardware performance degradation and real-world concentration fluctuations. This often leads to the system outputting continuously drifting erroneous data without any warning, until batch yield anomalies occur or are discovered by third-party calibration, resulting in irreversible economic losses. Traditional solutions rely on high-frequency manual calibration and maintenance, which not only incurs high lifecycle costs (LCC) but also fails to fundamentally eliminate monitoring blind spots caused by latent instrument failures. Summary of the Invention

[0006] The purpose of this invention is to provide a suspended particle detection system based on multi-feature data analysis to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a suspended particle detection system based on multi-feature data analysis, comprising: The sampling module is used to acquire the aerosol sample to be tested and establish a continuous directional airflow path; The main channel detection module is located downstream of the sampling module and in fluid communication with it along the airflow dynamic path. It has a built-in first particle sensing component for performing first-level wide baseline detection on the aerosol sample flowing through it under full flux conditions and outputting first sensing data containing full particle characteristics. The fluid splitting module receives the exhaust gas flow after full detection and physically divides the exhaust gas flow into N parallel split gas flows in equal proportions. The array-type detection module includes N parallel-connected second particle sensing components. Each of the N second particle sensing components is connected to a transmission channel of one of the N split airflows. It is used to perform second-level high-resolution detection in a low-density state of spatial dimensionality reduction and to simultaneously output the second sensing data set. The data reconstruction module receives the first sensing data and the second sensing data set, extracts the second sensing dataset, merges and performs mathematical aggregation to generate the second aggregated benchmark data, and compares the first sensing data and the second aggregated benchmark data in real time to determine whether the main channel detection module has triggered a signal overlap distortion state. In response to the distortion state, the reconstruction feedback logic is triggered: the target feature fragment in the second aggregated reference data is extracted as the real physical reference, the error correction amount of the main channel detection module at the current concentration due to the overlap of multiple particle signals is calculated, and the error correction amount is used to perform algebraic subtraction on the first sensing data, and the fused aerosol distribution data is output by splicing the corrected large particle size data before dimensionality reduction and the measured small particle size data after dimensionality reduction.

[0008] Preferably, the method for calculating the error correction amount of the main channel detection module at the current concentration due to multi-particle signal overlap includes the following steps: Extract the small particle size range that is not affected by overlap from the second aggregated baseline data as the true particle concentration parameter; The actual particle concentration parameters are correlated with the volume characteristics of the fixed optical sensitive area of ​​the first particle sensing component in a spatiotemporal scale, and then input together into the spatial distribution probability statistical algorithm model. Based on the model output, the probability of overlapping distortion caused by multiple particles coexisting simultaneously in the fixed optical sensitive area and triggering photoelectric signal merging is output, and then quantified to generate the amount of overlapping pseudo-peak signal used to correct the first sensing data as an error correction amount.

[0009] As a preferred method, the method for outputting fused aerosol distribution data by splicing the corrected large-particle-size data before dimensionality reduction and the measured small-particle-size data after dimensionality reduction includes the following steps: A dynamic particle size feature threshold is constructed, and the system output chain is truncated based on the dynamic particle size feature threshold. The small particle size signal range below the dynamic particle size feature threshold is forcibly called with the second aggregated reference data output by the array detection module, and the large particle size signal range above the dynamic particle size feature threshold is forcibly called with the real large particle size feature data output by the main channel detection module and reconstructed by algebraic subtraction.

[0010] Preferably, the data reconstruction module further includes a multi-dimensional inversion unit, which interacts with the fluid diversion module to perform physical loss and data inversion functions, including the following steps: The pipe wall of the fluid diversion module serves as a physical capture surface for high-inertia particles, causing large particles to lose inertia as they flow through it. The multidimensional inversion unit extracts the dynamic aerodynamic penetration rate of the aerosol sample as it flows through the pipe wall by comparing the total amount of real large-diameter particles after subtraction and reconstruction with the total amount of large particles remaining in the second aggregated reference data, and then inverts the aerodynamic equivalent density attribute of the sample.

[0011] As a preferred method, the method for inverting the aerodynamic equivalent density properties of a sample includes the following steps: A multidimensional mapping matrix including equivalent particle size, aerodynamic penetration rate and particle density is pre-made, and the dynamic aerodynamic penetration rate calculated in real time is used as an index to input the matrix. When the penetration rate index drops sharply, a high-density inorganic label is output, and when the penetration rate index is maintained at the theoretical fluid threshold, a low-density organic label is output.

[0012] Preferably, the data reconstruction module further includes a synchronous cross-validation mechanism constructed with the array detection module and the main channel detection module on the time axis, including continuously differentially comparing the background features of the first sensing data and the second aggregated reference data in a steady state without triggering signal overlap distortion; in response to detecting a unidirectional baseline drift in the first sensing data and the second aggregated reference data remaining stable, triggering and outputting a pollution alarm for the physical interface of the main channel detection module.

[0013] Preferably, in response to the comparison finding that the signal peak of the first sensing data shows a proportional attenuation across all channels, it is determined that the power of the detection source is aging, and a control command is generated and sent to the main channel detection module to forcibly increase the driving current injected into the detection source to compensate for the attenuation. At the same time, a remaining lifetime prediction work order is output according to the time-series evolution model of the compensation current.

[0014] Preferably, the data reconstruction module integrates a data-driven signal deconstruction prediction model. During training, the second aggregated benchmark data serves as the convergence of the distortion-free supervised label constraint model, and the distorted multi-peak signal in the first sensing data serves as the feature input. During inference, the original distorted waveform of the main channel detection module is directly captured, and the multi-granular independent features after deep deconstruction are output.

[0015] Preferably, the current particulate matter concentration gradient is evaluated based on the first sensor data and converted into an actuation electrical signal fed back to the fluid diversion module; in the low concentration state, a throttling command is executed to actively lock some diversion channels to reduce the overall power consumption of the array-type detection module; in the high concentration state, a full-pass command is executed to automatically turn on all N channels and start the extreme diversion dimensionality reduction mode.

[0016] Preferably, the fluid diversion module includes a flow equalization region with isodynamic expansion buffering characteristics and N impedance-matched diversion channels connected downstream of the flow equalization region; the fluid dynamic structure of the flow equalization region and the diversion channels is configured such that the diverted airflow is kept in balance in terms of fluid dynamic pressure, so as to reduce the frequency of aerosol particles entering any of the second particle sensing components by utilizing a physical space isolation multiplexing mechanism.

[0017] In summary, the beneficial effects of this invention are: This invention creatively introduces the concept of spatial dimension reduction and reuse, which can uniformly split extremely high concentration samples into array microchannels without the need to introduce external high-pressure pure gas. This avoids the defects of traditional pneumatic dilution devices, such as large size and susceptibility to secondary pollution from gas sources, and completely eliminates the congestion and overlap errors of small-diameter particles at extreme concentrations from the underlying physical structure. Traditional detection technologies face the dilemma that preserving the sample size will result in overlap, while dilution will result in the loss of large particles. This invention performs full-throughput sampling in the first detection zone, preserving the capture probability of rare large particles. At the same time, it uses a rear array-type splitting to obtain a distortion-free real small particle concentration benchmark, which is then substituted into a probabilistic statistical model to reverse the distortion signal from the front side. The fused data has both absolute confidence in the large particle size range and ultra-high resolution in the small particle size range. This invention turns waste into treasure, transforming the unavoidable "airflow pipe wall inertial loss" in the gas flow splitting structure into an analytical probe with physical significance. By establishing a dynamic penetration rate model between two-stage large particle data, it enables aerosols to not only be counted and their diameter measured, but also their density calculated and their materials identified, providing a new dimension with great industrial application value for tracing pollution sources in ultra-clean environments. Based on the physical principle of independent detection of the same airflow source by the front and rear stages, the system constructs a digital twin-level self-interlocking verification network. When the instrument experiences latent attenuation or a small amount of dust accumulation on the lens, it can perform sensitive self-sensing and adaptive compensation of the drive current, which greatly extends the system calibration cycle and completely eliminates the blind spots of error monitoring and process contamination caused by instrument latent drift. The system robustness and automated maintenance capabilities are greatly improved. Attached Figure Description

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

[0019] Figure 1 This is a schematic diagram of the overall process framework of a suspended particle detection system based on multi-feature data analysis according to the present invention. Figure 2 For the present invention Figure 1 A schematic diagram of the data processing steps and workflow framework for the intermediate error reconstruction and fusion output stage; Figure 3 For the present invention Figure 1 A schematic diagram illustrating the working logic of the multidimensional inversion and self-diagnosis extension process. Figure 4 This is a diagram showing the detection data terminal interface of the main channel detection module of the suspended particle detection system based on multi-feature data analysis according to the present invention. Figure 5 This is a diagram showing the detection data terminal interface of the array-type detection module of the suspended particle detection system based on multi-feature data analysis according to the present invention. Figure 6 This is a terminal display interface diagram of a suspended particle detection system based on multi-feature data analysis according to the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. These drawings are simplified schematic diagrams, which are only used to illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0021] To facilitate understanding of the present invention, a more complete description of the invention will be given below with reference to the accompanying drawings, which illustrate several embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the invention will be more thorough and complete.

[0022] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0023] Any feature disclosed in this specification (including any appended claims, abstract, and drawings) may be replaced by other equivalent or similar features for a similar purpose, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0024] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of at least two elements or the interaction relationship of at least two elements, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Please see Figures 1-6 The present invention provides an embodiment of a suspended particle detection system based on multi-feature data analysis. In terms of hardware physical structure, it innovatively integrates full-throughput airflow capture and array-based physical diversion dimensionality reduction mechanisms. At the software and algorithm level, it constructs a composite processing architecture based on a posterior probability benchmark for inverse error stripping, cross-domain feature fusion, multi-dimensional aerodynamic material inversion, and closed-loop self-monitoring and adaptive compensation. Specifically, it includes: The data acquisition module includes a high-precision brushless constant-flow air pump, a mass flow controller (MFC), and an isodynamic sampling probe with a funnel-shaped inlet. The air pump operates at a standard constant flow rate. ,Right now Continuously extract aerosol samples to be tested from external cleanrooms or closed process environments.

[0026] The intake pipe is made of anti-static stainless steel with a polished inner wall. The pipe diameter is matched with the flow rate to ensure that the airflow is in a fully developed and stable state before entering the first detection zone.

[0027] Mainstream detection module: After leaving the sampling inlet, the airflow passes vertically downwards through the first detection area. This area houses the mainstream optical particle sensing chamber, which contains a semiconductor laser (wavelength...). After being shaped by an asymmetric cylindrical mirror group, it forms a sheet-like laser beam with extremely thin thickness and a flat cross-section.

[0028] The full-flux optical sensitive area of ​​the first particle sensing component completely covers the entire 28.3 L / min airflow cross-section. When the aerosol sample to be tested passes through this sensitive area without damage at 100% raw flux, the scattered light from the suspended particles is captured by the photomultiplier tube (PMT) and converted into continuous high-speed analog electrical pulses. The system's high-speed analog-to-digital converter (ADC) samples these pulses and outputs first sensing data containing full particle characteristics. At this point, if the environmental concentration is at a normal or low to medium level, the data directly reflects the total particle size distribution; however, when the environmental concentration soars to tens of millions of ultra-high concentrations, the main channel will experience severe overlap errors due to multiple particles passing through the sensitive area simultaneously, resulting in distortions in the original data such as "lower counts of small particles and falsely higher counts of large particles".

[0029] Fluid Diversion Module: After the airflow has been detected at full throughput in the first detection zone, it flows into the fluid diversion module. This component is the core mechanical structure for achieving physical dimensionality reduction of the system. It consists of a gradually expanding buffer flow equalization cavity and N evenly distributed subdivided flow channels at the tail end, for example, N=4.

[0030] The cross-sectional area of ​​the buffer flow equalization cavity expands smoothly at a specific cone angle, using fluid dynamics principles to eliminate turbulence and pressure abrupt changes brought about by the main flow channel. At the end of the flow equalization cavity, four micro-diversion pipes with absolutely symmetrical geometric dimensions, lengths, and inner wall roughness are arranged in a ring array.

[0031] This structure divides the total flow rate Q into four completely equivalent split flow streams through fluid dynamic flow restriction interaction, with each stream having a flow rate of... This physical diversion action directly reduces the spatial distribution density of particles entering the rear channel proportionally to one-quarter of the original density, thereby greatly diluting the congestion of microscopic particles passing through the sensitive area at the physical source.

[0032] Array-type detection module: The four streams of airflow output from the manifold are precisely matched and injected into the four parallel micro-sensor chambers within the array-type detection module.

[0033] Each sensing chamber houses an independent second particle sensing component, comprising miniaturized laser diodes and photodiode arrays. Because each channel handles only a quarter of the total airflow and its optically sensitive area is miniaturized, the effective particle passage frequency of a single sensor drops drastically in extremely high-concentration environments, fundamentally eliminating overlap errors. These four array units operate in parallel, each outputting a high-precision subset of second sensing data, which the system synchronously acquires in real time to form a second sensing dataset.

[0034] Data reconstruction module: As the core of the entire system, this unit executes a rigorous algorithm with closed-loop feedback characteristics to address signal distortion in high-concentration environments. The specific implementation steps are as follows: Step 1: Two-level synchronous data acquisition and streaming data alignment The heterogeneous processing unit synchronously captures the first set of sensing data from the main channel detection module via a high-bandwidth parallel bus at a fixed clock cycle (e.g., every 100 milliseconds as a sampling window T). And the second sensor data subsets independently output by the four parallel channels in the array detection module. .

[0035] Align the data packets of each channel on the timeline to ensure that all data correspond to the same batch of transient aerosol samples.

[0036] Step Two: Mathematical Aggregation and Cross-Domain Cross-Monitoring Interaction The processing unit performs summation and aggregation operations on the four subsets of data from the array-type detection module to generate the second aggregated benchmark data. : ; Establish cross-monitoring interaction: The processing unit continuously compares the total count of the first sensor data of the main channel in real time. Total aggregated second aggregated baseline data Theoretically, at low concentrations without overlap, the ratio between the two should maintain a strictly linear conversion relationship. .

[0037] Trigger determination: If the ratio deviates from a preset threshold, for example... If the value is significantly lower than the theoretical expectation, and the count in the large particle size range of the first sensing data shows an abnormally expanded distortion tail, the processing unit determines that the main channel detection module has entered a signal overlap distortion state, and then automatically triggers the cross-domain reconstruction feedback logic.

[0038] Step 3: Physical benchmark extraction and application of spatial distribution probability statistical algorithm model After confirming the entry into the overlap distortion state, the processing unit performs baseline extraction: due to the physical splitting and dimensionality reduction of the array-type detection module, its aggregated data... Small particle size segments, such as Completely unaffected by overlap errors, the data is absolutely true and reliable.

[0039] Extract the actual particle concentration parameters of this small particle size range and record them as the baseline concentration. .

[0040] Spatiotemporal scale correlation and probabilistic inference: The effective volume of the fixed optical sensitive area of ​​the first particle sensing component By correlating the flow velocity with the spatiotemporal scale, the expected arrival rate parameter of particles within a unit sensitive area is calculated. ,in .

[0041] parameters Substituting this into the spatial distribution probability statistical algorithm model (Poisson distribution mathematical model), according to the Poisson statistical principle, the probability expression for the simultaneous random occurrence of k small particles within a small sensitive area volume is: ; Further quantification calculations of multi-particle (i.e.) The probability of overlap distortion, where both exist simultaneously within the fixed sensitive area, causing photoelectric pulse signals to become indistinguishable on the time axis and merge into a single "false large particle pulse," is high. : ; Step 4: Error Correction Quantization and Signal Reverse Stripping The processing unit is based on the calculated overlap distortion probability. Combined with the total number of pulse trigger events recorded in the first detection area within this sampling window The amount of overlapping pseudo-peak signal generated by signal merging is quantized and calculated. : ; Perform algebraic reduction: from the total number of large-diameter particles on the surface recorded in the first sensor data. In the process of subtracting the overlapping spurious peak signal, the actual total amount of large-diameter particles captured by the main channel detection module can be reconstructed and restored. : ; Step 5: Cross-domain feature splicing and fusion particle distribution output After the aforementioned reverse stripping and algebraic reduction, the processing unit constructs a dynamic particle size feature threshold in memory, for example, setting the dividing point as... .

[0042] Data splicing and interaction strategy: For values ​​below the dynamic particle size characteristic threshold In the small-particle-size signal range, the second aggregated reference data is forcibly invoked and aggregated by the array detection module, taking advantage of its non-overlapping congestion and extremely high resolution.

[0043] For values ​​above the dynamic particle size characteristic threshold For large-particle-size signal regions, the actual large-particle-size feature data after full-throughput detection and algebraic reduction reconstruction of the main channel is forcibly invoked. It leverages its advantages of 100% sampling volume, avoiding large particle loss in the pipeline, and extremely high statistical confidence.

[0044] Finally, the processing unit smoothly merges the two data segments at the threshold and outputs a final aerosol distribution fusion data report that combines large-particle-size absolute confidence with small-particle-size ultra-high resolution.

[0045] In another embodiment, to endow traditional optical particle counters with insights into the material and density properties of aerosol particles, this invention incorporates a multi-dimensional inversion unit in the data reconstruction module. This multi-dimensional inversion unit, together with the fluid diversion module, constitutes a functional interaction between physical loss and data inversion, specifically including the following: 1. Inertial Separation and Loss Capture Mechanism When an aerosol sample passes through the flow equalization and diversion microstructure of the fluid diversion module, the airflow direction is slightly deflected and flows through the microchannel wall.

[0046] According to aerodynamic principles, the inertia of a particle is directly proportional to the square of its aerodynamic equivalent diameter and its particle density. With the same optically equivalent particle size, high-density particles have significantly greater mass and inertia than low-density particles, such as plant fibers, human skin flakes, and organic dust.

[0047] Large particles with high inertia are prone to being trapped and lost when passing through bends and junctions of the manifold, due to inertial collisions and pipe wall settling, because they cannot follow the gas flow line in time.

[0048] 2. Dynamic aerodynamic penetration rate calculation and material inversion The aerodynamic loss and material multidimensional inversion unit runs continuously in the background. This module obtains the total amount of true large-diameter particles in the main channel after inversion correction in step four. This represents the initial number of large particles before entering the diversion manifold, and simultaneously obtains the total number of large-diameter particles remaining in the second aggregation baseline data. .

[0049] Calculate the dynamic aerodynamic penetration rate of the aerosol sample as it flows through the splitter assembly. : ; An internally pre-defined cross-modal feature mapping matrix of "equivalent particle size - aerodynamic penetration rate - particle density" is used, based on the hydrodynamic Stokes number. The lookup table or machine learning regression model is calibrated.

[0050] Output category label: Dynamic aerodynamic penetration rate calculated in real time When a precipitous drop occurs, significantly below the theoretical fluid limit, it indicates severe particle loss due to impact at the wall surface, resulting in extremely large mass. The system automatically outputs a classification label indicating that the sample is high-density inorganic matter, such as heavy metal particles or mechanical wear debris. Conversely, if the dynamic aerodynamic penetration rate... Maintaining a level close to the theoretical fluid transport threshold, the system outputs a classification label indicating that the sample is low-density organic matter, such as cleanroom garment fibers or light environmental dust. This function provides indispensable physical evidence for the accurate tracing of cleanroom contamination sources.

[0051] In another embodiment, the present invention utilizes the physical symmetry of the system's built-in "two-stage independent detection of the same airflow source" to achieve digital twin-level self-monitoring, adaptive compensation, and lifetime prediction, specifically including: Under steady-state operation or low-concentration conditions in the cleanroom, the redundant interlocking self-test diagnostic module and the two-level detection module form a synchronous cross-verification interaction on the time axis.

[0052] The module continuously performs differential comparison of the background noise parameters or low-concentration pulse broadening parameters of the first sensing data and the second aggregated reference data.

[0053] If the system detects a continuous unidirectional upward drift in the background ambient noise parameter of the first particle sensing component or an increase in weak stray light, while the aggregated reference data of the array detection module remains stable, it proves that the second channel is unaffected. Therefore, the system automatically triggers and determines that a small amount of aerosol dust contamination has occurred at the physical interface of the optical lens of the main channel detection module, and outputs a location-based contamination alarm for this main channel interface in the host computer's FMS plant monitoring system.

[0054] If, during long-term operation, the system finds that the peak values ​​of the optical sensing signals in each characteristic channel of the first sensing data exhibit a proportional attenuation across all channels, and the array-type detection module also exhibits the corresponding characteristics after calibration, the processing unit determines that the core power of the primary detection source in the main channel has aged.

[0055] The processing unit immediately generates hardware control commands and sends them to the hardware drive circuit of the detector source, forcibly and dynamically increasing the constant current source drive current parameter injected into the detector source proportionally, and forcibly pulling the emitted light intensity back to the initial calibration baseline, thus achieving "hardware self-healing and online light intensity calibration" without manual intervention.

[0056] The system records the time-series data of the historical evolution of the compensation current, and combines it with embedded time series prediction models such as exponential smoothing or long short-term memory networks (LSTM) to accurately calculate the remaining available time before the detector source completely fails. It automatically generates and pushes early warning work orders for the remaining lifespan of the detector source to the equipment maintenance personnel, thus completely eliminating blind spots in process monitoring caused by the instrument's hidden light decay.

[0057] In another embodiment, to further optimize system energy consumption and improve dynamic adaptability over ultra-wide flow ranges, the fluid splitting module of the present invention is internally coupled with a dynamic flow resistance distribution array: The array has a miniature electronically controlled throttle valve driven by a microelectromechanical system embedded at the entrance of each shunt channel.

[0058] The dynamic flow resistance distribution array and the data reconstruction module form an electromechanical negative feedback regulation closed loop: the processing unit calculates the current environmental load level based on the particulate matter concentration gradient parameters fed back in real time by the first particle sensing component, and converts it into an actuation electrical signal that is fed back to the dynamic flow resistance distribution array.

[0059] Adaptive switching of operating conditions: Under low-concentration steady-state conditions, the processing unit issues a throttling command to actively block some diversion channels, for example, keeping only one channel open and closing the other three channels, thereby significantly reducing the load on the air pump and the overall power consumption of the array detection module.

[0060] Once the processing unit detects a sharp increase in the concentration gradient parameter, exceeding the preset high-throughput overlap congestion threshold, the system instantly triggers a full-pass command, automatically opening all N=4 fluid physical diversion channels, enabling the system to seamlessly and quickly switch into the extreme high-concentration physical diversion array detection mode.

[0061] The above description is merely a specific embodiment of the invention, but the scope of protection of the invention is not limited thereto. Any variations or substitutions conceived without inventive effort should be included within the scope of protection of the invention. Therefore, the scope of protection of the invention should be determined by the scope defined in the claims.

Claims

1. A suspended particle detection system based on multi-feature data analysis, characterized in that: include: The sampling module is used to acquire the aerosol sample to be tested and establish a continuous directional airflow path; The main channel detection module is located downstream of the sampling module and in fluid communication with it along the airflow dynamic path. It has a built-in first particle sensing component for performing first-level wide baseline detection on the aerosol sample flowing through it under full flux conditions and outputting first sensing data containing full particle characteristics. The fluid splitting module receives the exhaust gas flow after full detection and physically divides the exhaust gas flow into N parallel split gas flows in equal proportions. The array-type detection module includes N parallel-connected second particle sensing components. Each of the N second particle sensing components is connected to a transmission channel of one of the N split airflows. It is used to perform second-level high-resolution detection in a low-density state of spatial dimensionality reduction and to simultaneously output the second sensing data set. The data reconstruction module receives the first sensing data and the second sensing data set, extracts the second sensing dataset, merges and performs mathematical aggregation to generate the second aggregated benchmark data, and compares the first sensing data and the second aggregated benchmark data in real time to determine whether the main channel detection module has triggered a signal overlap distortion state. In response to the distortion state, the reconstruction feedback logic is triggered: the target feature fragment in the second aggregated reference data is extracted as the real physical reference, the error correction amount of the main channel detection module at the current concentration due to the overlap of multiple particle signals is calculated, and the error correction amount is used to perform algebraic subtraction on the first sensing data, and the fused aerosol distribution data is output by splicing the corrected large particle size data before dimensionality reduction and the measured small particle size data after dimensionality reduction.

2. The suspended particle detection system based on multi-feature data analysis according to claim 1, characterized in that: The method for calculating the error correction amount of the main channel detection module at the current concentration due to multi-particle signal overlap includes the following steps: Extract the small particle size range that is not affected by overlap from the second aggregated baseline data as the true particle concentration parameter; The actual particle concentration parameters are correlated with the volume characteristics of the fixed optical sensitive area of ​​the first particle sensing component in a spatiotemporal scale, and then input together into the spatial distribution probability statistical algorithm model. Based on the model output, the probability of overlapping distortion caused by multiple particles coexisting simultaneously in the fixed optical sensitive area and triggering photoelectric signal merging is output, and then quantified to generate the amount of overlapping pseudo-peak signal used to correct the first sensing data as an error correction amount.

3. The suspended particle detection system based on multi-feature data analysis according to claim 2, characterized in that: The method for outputting fused aerosol distribution data by splicing together the corrected large-particle-size data before dimensionality reduction and the measured small-particle-size data after dimensionality reduction includes the following steps: A dynamic particle size feature threshold is constructed, and the system output chain is truncated based on the dynamic particle size feature threshold. The small particle size signal range below the dynamic particle size feature threshold is forcibly called with the second aggregated reference data output by the array detection module, and the large particle size signal range above the dynamic particle size feature threshold is forcibly called with the real large particle size feature data output by the main channel detection module and reconstructed by algebraic subtraction.

4. The suspended particle detection system based on multi-feature data analysis according to claim 2, characterized in that: The data reconstruction module further includes a multi-dimensional inversion unit, which, together with the fluid diversion module, constitutes a functional interaction between physical loss and data inversion, including the following steps: The pipe wall of the fluid diversion module serves as a physical capture surface for high-inertia particles, causing large particles to lose inertia as they flow through it. The multidimensional inversion unit extracts the dynamic aerodynamic penetration rate of the aerosol sample as it flows through the pipe wall by comparing the total amount of real large-diameter particles after subtraction and reconstruction with the total amount of large particles remaining in the second aggregated reference data, and then inverts the aerodynamic equivalent density attribute of the sample.

5. A suspended particle detection system based on multi-feature data analysis according to claim 4, characterized in that: The method for inverting the aerodynamic equivalent density properties of a sample includes the following steps: A multidimensional mapping matrix including equivalent particle size, aerodynamic penetration rate and particle density is pre-made, and the dynamic aerodynamic penetration rate calculated in real time is used as an index to input the matrix. When the penetration rate index drops sharply, a high-density inorganic label is output, and when the penetration rate index is maintained at the theoretical fluid threshold, a low-density organic label is output.

6. The suspended particle detection system based on multi-feature data analysis according to claim 1, characterized in that: The data reconstruction module also includes a synchronous cross-validation mechanism built on the time axis with the array detection module and the main channel detection module, including continuously differentially comparing the background features of the first sensing data and the second aggregated reference data under steady state without triggering signal overlap distortion; In response to the detection of a unidirectional baseline drift in the first sensor data and the stability of the second aggregated reference data, a pollution alarm is triggered and output for the physical interface of the main channel detection module.

7. A suspended particle detection system based on multi-feature data analysis according to claim 6, characterized in that: In response to the comparison finding that the signal peak of the first sensor data shows a proportional attenuation across all channels, it is determined that the power of the detection source is aging, and a control command is generated and sent to the main channel detection module to forcibly increase the driving current injected into the detection source to compensate for the attenuation. At the same time, a remaining lifetime prediction work order is output according to the time-series evolution model of the compensation current.

8. The suspended particle detection system based on multi-feature data analysis according to claim 1, characterized in that: The data reconstruction module integrates a data-driven signal deconstruction prediction model. During training, the second aggregated benchmark data serves as the convergence of the distortion-free supervised label constraint model, and the distorted multi-peak signal in the first sensing data serves as the feature input. During inference, the original distorted waveform of the main channel detection module is directly captured, and the multi-particle independent features after deep deconstruction are output.

9. A suspended particle detection system based on multi-feature data analysis according to claim 1, characterized in that: The current particulate matter concentration gradient is assessed based on the first sensor data and converted into an actuation electrical signal fed back to the fluid diversion module. In the low concentration state, a throttling command is executed to actively lock some diversion channels to reduce the overall power consumption of the array-type detection module. In the high concentration state, a full-pass command is executed to automatically turn on all N channels and start the extreme diversion dimensionality reduction mode.

10. A suspended particle detection system based on multi-feature data analysis according to claim 9, characterized in that: The fluid diversion module includes a flow equalization region with isodynamic expansion buffering characteristics and N impedance-matched diversion channels connected downstream of the flow equalization region; the fluid equalization region and the diversion channels are configured to maintain the diverted airflow in equilibrium in terms of fluid dynamic pressure, so as to reduce the frequency of aerosol particles entering any of the second particle sensing components by utilizing a physical space isolation multiplexing mechanism.

Images