A filter fault detection method, device, and medium based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based filter fault detection method, device, and medium, relating to the field of filter material detection and analysis technology. The method includes: acquiring a raw physical signal set and constructing a filter material distribution correlation set; preprocessing the raw physical signal set to generate a standardized analysis dataset; extracting multi-dimensional feature vectors from the standardized analysis dataset and performing topological aggregation and weighted fusion to generate a fused feature tensor; constructing a material performance analysis model and inputting the fused feature tensor into the model to perform material performance parameter mapping, outputting a filter material performance parameter vector, and simultaneously identifying the filter material aging degree to generate an effective material performance state; based on the effective material performance state, obtaining the evolution trend of filter material pore blockage rate, matching it with a maintenance strategy library, and generating a filter material performance evaluation report. This invention achieves the effect of improving the accuracy of fault detection and the interpretability of results.

Description

Technical Field

[0001] This invention relates to the field of filter material testing and analysis technology, and in particular to a filter fault detection method, equipment and medium based on artificial intelligence. Background Technology

[0002] Filter performance testing and fault detection fall under the category of fluid filtration material condition identification and industrial testing and analysis technology. Current filter testing typically revolves around operational information such as differential pressure, particulate matter concentration, airflow velocity, and air volume. This involves deploying detection points to acquire multi-source physical signals, which are then combined with statistical analysis, frequency domain analysis, or empirical criteria to identify the filter media's flow state, permeability characteristics, and contaminant accumulation status. Some solutions further incorporate data-driven methods, performing feature extraction, parameter mapping, and state discrimination on continuously sampled signals to improve the automation level of filter operation status measurement.

[0003] However, existing conventional methods still have two limitations. Firstly, conventional methods often rely on single signal characteristics or shallow statistical results for state assessment, making it difficult to comprehensively characterize filter media permeability, flow resistance coefficient, porosity, and pressure drop response coefficient. This results in insufficient adaptability of material performance measurements to complex operating conditions. Secondly, conventional methods typically focus on static discrimination for aging identification and degradation assessment, rarely integrating material performance parameter mapping results with temporal evolution trends and maintenance strategy matching processes. Therefore, there is still room for improvement in characterizing clogging rate evolution and supporting maintenance decision-making. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based filter fault detection method to address the problems of insufficient collaborative characterization of material performance parameters and insufficient integrated assessment of aging and degradation processes.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-based filter fault detection method, comprising: acquiring a raw physical signal set and constructing a filter media distribution correlation set; simultaneously preprocessing the raw physical signal set to generate a standardized analysis dataset; extracting multi-dimensional feature vectors from the standardized analysis dataset and performing topological aggregation and weighted fusion to generate a fused feature tensor; constructing a material performance analysis model and inputting the fused feature tensor into the material performance analysis model to perform material performance parameter mapping, outputting a filter media performance parameter vector, and simultaneously identifying the degree of filter media aging to generate an effective material performance state; and based on the effective material performance state, obtaining the evolution trend of filter media pore blockage rate, matching it with a maintenance strategy library, and generating a filter media performance evaluation report.

[0007] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the steps of acquiring the original physical signal set and constructing the filter media distribution correlation set are as follows: Collect a set of raw physical signals including differential pressure signals and particulate matter concentration signals, and construct a set of spatial location coordinates; Based on the original set of physical signals and spatial coordinates, and combined with the physical adjacency relationships of the filter media, a set of correlation relationships for filter media distribution is constructed.

[0008] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the preprocessing of the original physical signal set includes the following steps: Based on the original physical signal set, outlier points are identified by an outlier discrimination threshold, and missing signals are filled in using linear interpolation to generate a complete physical signal set. Based on the complete physical signal set, combined with the law of conservation of mass and Darcy's law, the filter material permeability error is calculated, and high-frequency noise is suppressed by wavelet thresholding to generate a denoised physical signal set. The denoised physical signal set is normalized to generate a standardized analysis dataset.

[0009] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the steps of extracting multi-dimensional feature vectors from a standardized analysis dataset and performing topological aggregation and weighted fusion to generate a fused feature tensor are as follows: Statistical moments and trend terms of time-series data are extracted from the standardized analysis dataset to construct time-domain feature vectors, and spatial feature vectors are obtained by combining the filter material distribution correlation set. Frequency domain indices and time-frequency features are extracted from the standardized analysis dataset using Fast Fourier Transform, forming frequency domain feature vectors and time-frequency feature vectors. The filter material distribution correlation set and spatial feature vector are aggregated through a spatial topology aggregation mechanism, and the time-domain feature vector, frequency-domain feature vector, time-frequency feature vector and aggregated spatial feature vector are fused to generate a fused feature tensor.

[0010] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the steps for constructing the material performance analysis model are as follows: The feature input layer is constructed by linear mapping and positional encoding, and high-order coupling relationships and long-range dependencies are extracted based on the Transformer encoder structure and multi-head self-attention mechanism to construct the signal representation channel; Based on the mass conservation equation, Darcy's law equation, and continuity constraints, the physical behavior of the filter material is characterized in layers, and physical constraint channels are constructed. The output feature vector of the signal characterization channel is concatenated with the output feature vector of the physical constraint channel, and a feature coupling layer is constructed through linear transformation and nonlinear activation processing. A parametric regression output layer is built using a fully connected structure, and the mass conservation equation, Darcy's law equation, and continuity constraints are embedded as joint constraint terms into the loss function to construct a multi-layer physical constraint loss function. The feature input layer, signal representation channel, physical constraint channel, feature coupling layer, and parameter regression output layer are directed by sequential forward connection through a deep neural network, and a material performance analysis model is constructed by combining a multi-layer physical constraint loss function.

[0011] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the step of inputting the fused feature tensor into the material performance analysis model to perform material performance parameter mapping is as follows: Input the historical labeled material sample library into the material performance analysis model, and adjust the weights of the material performance analysis model through iterative optimization algorithm until the loss converges, thereby generating the trained material performance analysis model. The fused feature tensor is input into the trained material performance analysis model to perform material performance parameter mapping and output filter material performance parameter vector.

[0012] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the steps of identifying the aging degree of the filter material and generating an effective material performance status are as follows: Key physical property indicators are extracted from the filter media performance parameter vector, and the degree of filter media aging is identified by combining dynamic confidence thresholds to generate material performance measurement labels. The material performance testing labels are voted on through a time-series verification window to obtain the stability variance of the material performance testing labels, eliminate transient interference, and generate an effective material performance status.

[0013] As a preferred embodiment of the AI-based filter fault detection method of the present invention, the steps of obtaining the evolution trend of filter media pore blockage rate based on the effective material performance status, matching it with a maintenance strategy library, and generating a filter media performance evaluation report are as follows: Based on standardized analysis datasets and effective material performance status, the evolution trend of filter media pore blockage rate is obtained by utilizing time-series evolution, and a predicted value of filter media performance degradation is generated. Based on the predicted values ​​of filter media performance degradation and the material performance measurement labels, a maintenance strategy library is matched to generate a filter media performance evaluation report.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the artificial intelligence-based filter fault detection method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based filter fault detection method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by spatial topology aggregation, a unified expression of spatiotemporal correlation information of multiple collection points of the filter is realized, thereby enhancing the integrity of fault characterization and spatial identification capability; by constructing a material performance analysis model, the filter material performance parameter vector is determined and the filter material aging degree is identified, thereby deeply characterizing the filter's operating status and improving the accuracy of fault detection and the interpretability of results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-based filter fault detection method.

[0019] Figure 2 This is a flowchart of data acquisition and preprocessing.

[0020] Figure 3 This is a flowchart of feature extraction and fusion.

[0021] Figure 4 This is a flowchart for material performance analysis and evaluation. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based filter fault detection method, including the following steps: S1. Collect the original physical signal set and construct the filter material distribution correlation set. At the same time, preprocess the original physical signal set to generate a standardized analysis dataset.

[0026] The raw physical signal set containing differential pressure signal and particulate matter concentration signal is collected, and a spatial location coordinate set is constructed.

[0027] Furthermore, differential pressure sampling points, particulate matter concentration sampling points, airflow velocity sampling points, and airflow volume sampling points are respectively set up on the windward side, the middle of the filter material, and the air outlet side of the filter material. The sampling is synchronized according to a unified clock to obtain the original physical signal set containing differential pressure signal, particulate matter concentration signal, airflow velocity signal, and airflow volume signal. The geometric center of the filter material is set as the origin of the coordinate system. An X-axis is established along the airflow direction, a Y-axis is established along the width direction of the filter material, and a Z-axis is established along the thickness direction of the filter material. Three-dimensional coordinate values ​​are written according to the installation position of each sampling point to construct a spatial position coordinate set corresponding to each point of the original physical signal set. For example, the sampling frequency is 100Hz, the sampling duration is 300s, and the coordinate spacing is 20mm.

[0028] Based on the original set of physical signals and spatial coordinates, and combined with the physical adjacency relationships of the filter media, a set of correlation relationships for filter media distribution is constructed.

[0029] Furthermore, the time-series records of each acquisition point in the original physical signal set are used as node signal inputs, and the three-dimensional coordinate values ​​in the spatial location coordinate set are used as node location inputs. Node connection edges are established according to the coordinate spacing threshold and airflow propagation direction. Adjacent acquisition points from the windward side to the air outlet side, adjacent acquisition points on the same layer, and corresponding acquisition points across layers are written into the connection table to form the physical adjacency relationship of the filter material. Based on the physical adjacency relationship of the filter material, the node number, connection edge number, edge direction mark, and edge distance value of each group are combined to construct the filter material distribution association relationship set. Among them, the edge direction mark corresponding to the adjacent acquisition point from the windward side to the air outlet side is defined as the main airflow direction mark, the edge direction mark corresponding to the adjacent acquisition point on the same layer is defined as the intra-layer adjacent direction mark, and the edge direction mark corresponding to the cross-layer corresponding acquisition point is defined as the cross-layer corresponding direction mark. The main airflow direction mark is set along the direction from the windward side to the air outlet side, the intra-layer adjacent direction mark is set along the direction of increasing coordinate number on the same layer, and the cross-layer corresponding direction mark is set along the direction from the upper layer acquisition point to the lower layer corresponding acquisition point (for example, the coordinate spacing threshold is 20mm).

[0030] It should be noted that the physical adjacency relationship of the filter media is a connection relationship established based on the spatial adjacency of each sampling point in the thickness direction, intralayer direction, and cross-layer corresponding position of the filter media. It is used to characterize the airflow propagation path and the local state transmission path. The coordinate spacing threshold is set based on the filter media measuring point layout interval and spatial resolution requirements. It is obtained by statistically analyzing the actual installation spacing of adjacent sampling points and taking the upper limit of adjacent connections. An exemplary value range is 10mm to 30mm.

[0031] Based on the original physical signal set, outlier points are identified by an outlier discrimination threshold, and missing signals are filled in using linear interpolation to generate a complete physical signal set.

[0032] Furthermore, within a fixed-length (e.g., 20 sampling points) sliding window of the original physical signal set, the median value of the differential pressure signal, the median value of the particulate matter concentration signal, and the window deviation value are calculated. The absolute value of the deviation between the current sampled value and the corresponding median value is compared point by point, and sampling points that exceed the anomaly discrimination threshold are marked as signal anomalies and removed. Signal interruption points and null points are recorded as missing signals, and interpolation is calculated point by point according to the two valid sampled values ​​before and after the missing segment of the same sampling point and the sampling time interval. The interpolation is written into the corresponding missing position to generate a complete physical signal set.

[0033] It should be noted that the anomaly detection threshold is set based on the sampling fluctuation amplitude and the sensor measurement dispersion. It is obtained by statistically analyzing the sliding window deviation value and taking the upper limit of the anomaly boundary. For example, the value range is 2 to 4 times the window deviation value.

[0034] Based on the complete physical signal set, combined with the law of conservation of mass and Darcy's law, the filter material permeability error is calculated, and high-frequency noise is suppressed by wavelet thresholding to generate a denoised physical signal set.

[0035] Furthermore, the pressure difference signal, airflow velocity signal, and air volume signal at the same acquisition time are extracted and supplemented from the physical signal set. The flow conservation difference is calculated based on the difference between the gas volume throughput on the windward side and the gas volume throughput on the outlet side. The permeability of each acquisition point is calculated by combining the fluid dynamic viscosity, filter material thickness, effective flow area of ​​the filter material, and the corresponding pressure difference on both sides of the filter material. The permeability of each acquisition point is compared with the reference permeability to obtain the filter material permeability error. Then, the signal sequences containing the filter material permeability error are subjected to, for example, three-level wavelet decomposition to obtain the low-frequency and high-frequency components of each level. The parts of the high-frequency components whose absolute values ​​are lower than the wavelet threshold are set to zero. The retained high-frequency components and low-frequency components are reconstructed to generate a denoised physical signal set.

[0036] It should be noted that the baseline permeability is set as the baseline permeability value under the factory calibration state of the filter material; the wavelet threshold is set based on the distribution of the absolute value of the high-frequency component coefficients and the sampling fluctuation level, and is obtained by statistically analyzing the distribution of the absolute value of the high-frequency coefficients of the wavelet and taking the upper limit of the noise cutoff. An exemplary value range is 1.5 times to 3 times the standard deviation of the high-frequency coefficients.

[0037] Darcy's Law is expressed mathematically as follows: ; in, To indicate the first The collection point is at the [number]th [location]. The filter media permeability corresponding to each sampling time, where The sampling time number is used. Number the collection points; To indicate the first The collection point is at the [number]th [location]. The gas volume throughput at each sampling time is obtained based on the local flow velocity and the local effective flow area. The fluid dynamic viscosity characterizes the magnitude of the viscous resistance inside the gas medium in the sampling environment, and is determined based on the correspondence between the type of gas medium in the sampling environment and the temperature value at the sampling time. For filter media thickness; The effective flow area of ​​the filter media is determined by the local geometric area based on the outer contour dimensions of the filter media, the local coverage area of ​​the sampling point, and the obstruction area of ​​the filter media installation frame. The local effective flow area is then obtained by deducting the obstruction area of ​​the frame, the sealing edge area, and the non-flow coverage area. To indicate the first The collection point is at the [number]th [location]. The pressure difference across the filter material at each sampling time.

[0038] The denoised physical signal set is normalized to generate a standardized analysis dataset.

[0039] Furthermore, using the maximum and minimum values ​​in the differential pressure signal sequence, particulate matter concentration signal sequence, airflow velocity signal sequence, and air volume signal sequence as numerical boundaries, normalization processing is performed to convert each sampled value into a dimensionless value. The samples are then rearranged according to the sampling point number, spatial location coordinates, and sampling time to generate a standardized analysis dataset.

[0040] S2. Extract multidimensional feature vectors from the standardized analysis dataset and perform topological aggregation and weighted fusion to generate a fused feature tensor.

[0041] Statistical moments and trend terms of time-series data are extracted from the standardized analysis dataset to construct time-domain feature vectors. Spatial feature vectors are obtained by combining the filter material distribution correlation set.

[0042] Furthermore, the standardized analysis dataset is divided into fixed-length time segments (e.g., 50 sampling points) according to the sampling point number and sampling time. The mean, variance, skewness, kurtosis, and linear trend slope are extracted from each time segment and arranged in order according to signal type and statistical order to construct a time-domain feature vector. Based on the node connection edges, edge direction markers, and edge distance values ​​in the filter material distribution correlation set, the signal difference, distance weighting value, and direction transfer value between the current sampling point and the adjacent sampling points (e.g., 4 adjacent sampling points) at the same sampling time are sequentially written into the feature sequence to obtain the spatial feature vector.

[0043] Frequency domain indices and time-frequency features are extracted from the standardized analysis dataset using Fast Fourier Transform, forming frequency domain feature vectors and time-frequency feature vectors.

[0044] Furthermore, fast Fourier transforms are performed on the differential pressure signal sequence, particulate matter concentration signal sequence, airflow velocity signal sequence, and wind volume signal sequence in the standardized analysis dataset (e.g., the spectrum calculation sequence length is 64 sampling points) to obtain the frequency sequence and amplitude sequence corresponding to each sequence; the dominant frequency, dominant frequency amplitude, spectral energy, and spectral centroid are obtained sequentially according to the frequency position, and arranged according to the signal type and index order to form a frequency domain feature vector; the amplitude corresponding to each frequency position is continuously arranged along the sampling time to form a time-frequency energy distribution sequence (e.g., short-time spectral energy is calculated using a sliding window of 16 sampling points and a sliding step size of 8 sampling points), forming a time-frequency feature vector.

[0045] The spatial topology aggregation mechanism aggregates the filter media distribution correlation set and spatial feature vectors, and fuses the time domain feature vectors, frequency domain feature vectors, time-frequency feature vectors and the aggregated spatial feature vectors to generate a fused feature tensor. The fused feature tensor can synchronously retain time changes, frequency distribution, time-frequency energy and spatial transmission information, thereby enhancing the integrity of filter media fault characterization and the stability of subsequent parameter mapping.

[0046] Furthermore, the spatial feature vectors are mapped to node feature matrices according to the collection point numbers, and a spatial topology propagation sequence is constructed based on the node connection edges, edge direction markers, and edge distance values ​​in the filter material distribution association set (e.g., the number of spatial topology propagation layers is 2). Layered propagation and neighborhood difference accumulation are performed on the spatial feature vectors of adjacent collection points along the windward side to the windward side. The contribution value of each adjacent collection point is determined by the inverse of the edge distance and the direction consistency coefficient (e.g., the number of adjacent collection points is 4), generating an aggregated spatial feature vector. The time-domain feature vector, frequency-domain feature vector, time-frequency feature vector, and aggregated spatial feature vector are dimensionally aligned with the same collection point and the same sampling time, and stacked sequentially along the feature dimensions to generate a fused feature tensor.

[0047] It should be noted that the directional consistency coefficient is set based on the angle between the direction of the line connecting adjacent collection points and the direction of the main airflow of the filter material. It is obtained by normalizing the deviation of the angle direction. The exemplary value range is 0 to 1, where 1 indicates that the directions are completely consistent and 0 indicates that the directions are completely opposite. The spatial topology aggregation mechanism is a construction method that uses neighborhood transfer and weighted aggregation of spatial feature vectors based on the spatial connection relationship, directional relationship and distance relationship between each collection point in the filter material distribution association relationship set.

[0048] S3. Construct a material performance analysis model, input the fused feature tensor into the material performance analysis model, perform material performance parameter mapping, output filter material performance parameter vector, identify the aging degree of filter material, and generate effective material performance status.

[0049] The feature input layer is constructed by linear mapping and positional encoding, and high-order coupling relationships and long-range dependencies are extracted based on the Transformer encoder structure and multi-head self-attention mechanism to construct the signal representation channel.

[0050] Furthermore, the fused feature tensor is written into the input sequence according to the sampling point number and sampling time order, and a fixed-dimensional linear projection is performed on each set of feature components to generate an input mapping vector; the position encoding vector is written according to the sampling time order and spatial position order, and is superimposed with the input mapping vector bit by bit to form a feature input layer; the feature input layer is sequentially fed into a Transformer encoder (e.g., 3 layers), and in each layer, the input vector is mapped into a query vector, key vector, and value vector respectively, and the association weight between each sampling point is calculated according to the multi-head self-attention mechanism; the outputs of each attention head are concatenated and iteratively updated by a feedforward network to extract high-order coupling relationships and long-range dependencies, and to construct a signal representation channel.

[0051] It should be noted that higher-order coupling relationships are cross-signal, cross-space, and cross-temporal joint correlations formed after multiple feature components in the fused feature tensor undergo multi-layer attention mapping; long-range dependencies are long-distance correlations retained between non-adjacent acquisition points or non-adjacent sampling times.

[0052] Based on the mass conservation equation, Darcy's law equation, and continuity constraints, the physical behavior of filter media is characterized in layers, and physical constraint channels are constructed.

[0053] Furthermore, the sequences corresponding to the pressure difference signal, airflow velocity signal, and airflow volume signal in the fused feature tensor are expanded according to the sampling point number and sampling time sequence. Based on the mass conservation equation, the flow conservation difference value of each sampling layer is obtained to form a flow constraint sequence. Based on Darcy's law equation, the permeability, pressure difference response value, and flow response value of each sampling point are calculated to form a seepage constraint sequence. The pressure difference response value is defined as the ratio of the change in pressure difference across the filter material to the corresponding change in flow rate at two adjacent sampling times, and the flow response value is defined as the ratio of the change in gas volume throughput to the corresponding change in pressure difference at two adjacent sampling times. Based on the pressure difference signal difference, flow velocity signal difference, and airflow volume signal difference of adjacent sampling points at the same sampling time, the pressure difference continuity residual value, flow velocity continuity residual value, and airflow continuity residual value are calculated to form a continuity constraint sequence. The flow constraint sequence, seepage constraint sequence, and continuity constraint sequence are arranged in physical hierarchy order to generate a physical constraint channel, which is used to constrain the subsequent feature characterization results to conform to the actual physical behavior of the filter material.

[0054] It should be noted that the continuity constraint is a constraint relationship that limits the degree of continuity of changes in pressure difference, airflow velocity and air volume at adjacent sampling points at the same sampling time. It is used to characterize that the transition of physical state between adjacent positions inside the filter material cannot be abrupt.

[0055] The output feature vector of the signal representation channel is concatenated with the output feature vector of the physical constraint channel, and a feature coupling layer is constructed through linear transformation and nonlinear activation processing.

[0056] Furthermore, the output feature vectors of the signal characterization channel and the physical constraint channel are aligned sequentially at the same acquisition point and the same sampling time, and then concatenated along the beginning and end of the feature dimension to form a joint feature sequence (e.g., the joint feature dimension is 256). A fixed-dimensional linear mapping is performed on the joint feature sequence to convert it into a unified coupled feature dimension, generating a coupled feature vector. The coupled feature vector is then fed bit by bit into a nonlinear activation function to enhance the joint response strength across signals, spaces, and physical constraints, thus constructing a feature coupling layer.

[0057] A parametric regression output layer is built using a fully connected structure, and the mass conservation equation, Darcy's law equation, and continuity constraints are embedded as joint constraint terms into the loss function to construct a multi-layer physical constraint loss function.

[0058] Furthermore, the output feature vectors of the feature coupling layer are written into the regression sequence according to the same acquisition point and the same sampling time, and then sequentially fed into the fully connected mapping relationship (e.g., two layers). The coupling feature dimension is compressed layer by layer into the target parameter dimension (e.g., compressed from 128 dimensions to 64 dimensions, and then from 64 dimensions to 4 dimensions) to form the parameter regression output layer. The filter material physical performance parameter vector output by the parameter regression output layer, the flow conservation difference in the flow constraint sequence, the permeability constraint value in the seepage constraint sequence, and the continuity residual value (including the pressure difference continuity residual value, the flow velocity continuity residual value, and the air volume continuity residual value) together constitute the loss function. The mass conservation equation corresponds to the flow conservation term, the Darcy's law equation corresponds to the permeability constraint term, and the continuity constraint condition corresponds to the continuity constraint term. These are superimposed according to the physical level to form a multi-layer physical constraint loss function.

[0059] The feature input layer, signal representation channel, physical constraint channel, feature coupling layer, and parameter regression output layer are directed by sequential forward connection through a deep neural network, and a material performance analysis model is constructed by combining a multi-layer physical constraint loss function.

[0060] Furthermore, the output of the feature input layer is connected to the signal characterization channel and the physical constraint channel, respectively. The output of the signal characterization channel and the output of the physical constraint channel are connected together to the feature coupling layer. The output of the feature coupling layer is connected to the parameter regression output layer. The output of the parameter regression output layer is connected to the mass conservation loss term, the Darcy's law loss term, and the continuity constraint loss term, respectively. After weighted summation, a multi-layer physical constraint loss function is formed, thus completing the sequential forward connection construction of the material performance analysis model.

[0061] Input the historical labeled material sample library into the material performance analysis model, and adjust the weights of the material performance analysis model through iterative optimization algorithm until the loss converges, thereby generating the trained material performance analysis model.

[0062] Furthermore, the historical labeled material sample library is input into the material performance analysis model, and the sample fusion feature tensor, filter media permeability, filter media flow resistance coefficient, filter media porosity (the ratio between the pore volume and the total volume of the filter media), and filter media pressure drop response coefficient corresponding to each sample in the historical labeled material sample library are simultaneously written into the training sequence. The material performance analysis model sequentially completes the forward propagation of the feature input layer, signal representation channel mapping, physical constraint channel mapping, feature coupling layer mapping, and parameter regression output layer mapping to obtain the filter media physical performance parameter vector. Using the loss value output by the multi-layer physical constraint loss function as the optimization objective, the Adam iterative optimization algorithm is used to update the weights of each layer round by round until the loss difference between two adjacent rounds is lower than the loss convergence threshold, thus generating the trained material performance analysis model.

[0063] It should be noted that the historical labeled material sample library includes sample fusion feature tensors and corresponding tags for filter media permeability, filter media flow resistance coefficient, filter media porosity, and filter media pressure drop response coefficient. The filter media flow resistance coefficient is obtained based on the relationship between the pressure difference across the filter media and the unit flow rate, characterizing the resistance encountered by the airflow when passing through the filter media; an exemplary value range is 50 Pa·s / m³ to 500 Pa·s / m³. The filter media pressure drop response coefficient is obtained based on the response relationship between the change in pressure difference and the change in flow rate during the flow process, characterizing the pressure drop response coefficient of the filter media. The sensitivity to changes in flow is reduced, with an exemplary value range of 0.1 to 10; the loss convergence threshold is set based on training stability and the magnitude of loss reduction (the difference between the loss value in the current iteration and the loss value in the previous iteration), and is obtained by statistically analyzing the range of changes in the loss difference between adjacent iterations and taking the stopping update boundary value, with an exemplary value range of 0.0001 to 0.001; the Adam iterative optimization algorithm is a gradient optimization method that adaptively adjusts the learning step size of each parameter and combines first-order moment estimation and second-order moment estimation for weight updates.

[0064] The fused feature tensor is input into the trained material performance analysis model to perform material performance parameter mapping and output filter material performance parameter vector.

[0065] Furthermore, the fused feature tensor is input into the feature input layer of the trained material performance analysis model according to the acquisition point number and sampling time sequence. Through linear mapping and position encoding, an input feature sequence is generated. The input feature sequence is then input into the signal representation channel, where associated features are extracted using a Transformer encoder structure and a multi-head self-attention mechanism to generate a signal representation feature vector. Simultaneously, the fused feature tensor is input into the physical constraint channel, where a physical constraint feature vector is generated based on the mass conservation equation, Darcy's law equation, and continuity constraints. The signal representation feature vector and the physical constraint feature vector are then input into the feature coupling layer and concatenated to generate a coupled feature vector. Finally, the coupled feature vector is input into the parameter regression output layer to map material performance parameters, outputting a filter material performance parameter vector.

[0066] Key physical property indicators are extracted from the filter media performance parameter vector, and the degree of filter media aging is identified by combining dynamic confidence thresholds to generate material performance measurement labels.

[0067] Furthermore, filter media permeability, flow resistance coefficient, porosity, and pressure drop response coefficient are extracted from the filter media performance parameter vector and weighted by a fixed weight to synthesize aging judgment values. The mean and standard deviation of the aging judgment values ​​for the current batch (a continuous sample set formed within the same continuous sampling task for the same filter media object) are calculated. The aging judgment values ​​are compared one by one with the dynamic confidence threshold. Filter media samples that are higher than the dynamic confidence threshold are marked as aged filter media, and filter media samples that are not higher than the dynamic confidence threshold are marked as non-aged filter media, thus generating material performance measurement labels.

[0068] It should be noted that the key physical properties include filter media permeability, filter media flow resistance coefficient, filter media porosity, and filter media pressure drop response coefficient. The fixed weights are set based on the contribution of each filter media performance parameter in the historical labeled material sample library to aging identification, for example, filter media permeability 0.35, filter media flow resistance coefficient 0.30, filter media porosity 0.20, and filter media pressure drop response coefficient 0.15. The dynamic confidence threshold is set by using the historical percentile (e.g., 90th percentile) of the aging judgment value in the historical labeled material sample library as the basic threshold, and simultaneously adjusting the basic threshold in combination with the mean offset and standard deviation fluctuation of the current batch aging judgment value. An exemplary value range is 0.60 to 0.90.

[0069] The material performance testing labels are voted on through a time-series verification window to obtain the stability variance of the material performance testing labels, eliminate transient interference, and generate an effective material performance status.

[0070] Furthermore, the material performance measurement labels are written into the timing verification window in the order of sampling time (e.g., the timing verification window length is 5 sampling times), and the occurrence frequency of aging labels and non-aging labels in the timing verification window is counted; the labels with more occurrence frequency are used as voting decision labels, and the stability variance of the material performance measurement labels is calculated based on the label sequence corresponding to each sampling time in the timing verification window; material performance measurement labels with stability variance higher than the stability variance threshold and label duration lower than the duration threshold are judged as transient interference and are removed, and the retained voting decision labels are used as valid material performance states.

[0071] It should be noted that the stability variance threshold is set based on the label fluctuation amplitude in the historical labeled material sample library, and is obtained by statistically analyzing the upper limit of the label value dispersion within the time-series verification window under normal conditions. The exemplary value range is 0.10 to 0.30. The duration threshold is set based on the distribution of instantaneous disturbance duration, and is obtained by statistically analyzing the upper limit of the continuous sampling duration corresponding to historical instantaneous disturbances. The exemplary value range is 2 sampling times to 5 sampling times.

[0072] S4. Based on the effective material performance status, obtain the evolution trend of filter media pore blockage rate, match the maintenance strategy library, and generate a filter media performance evaluation report.

[0073] Based on standardized analysis datasets and effective material performance status, the evolution trend of filter media pore blockage rate is obtained by utilizing time-series evolution, and a predicted value of filter media performance degradation is generated.

[0074] Furthermore, the differential pressure signal growth rates corresponding to each sampling time in the standardized analysis dataset are arranged into a growth rate sequence in chronological order, and the effective material performance states are arranged into a state sequence in chronological order of the same sampling time. The growth rate sequence and state sequence of multiple consecutive sampling times (e.g., a continuous sampling length of 10 sampling times) are subjected to time-series evolution. Specifically, the basic clogging rate increment is calculated based on the difference in growth rates between adjacent sampling times in the growth rate sequence, and the basic clogging rate increment is corrected according to the effective material performance state at the corresponding sampling time. The filter media pore clogging rate increment is obtained moment by moment, and then the filter media pore clogging rate increment at each moment is continuously accumulated along the time direction to obtain the filter media pore clogging rate evolution trend. Based on the continuous change direction and continuous change amplitude at the end of the filter media pore clogging rate evolution trend, the clogging rate value for the next prediction period is calculated to generate the filter media performance degradation prediction value.

[0075] It should be noted that the differential pressure signal growth rate is calculated from the differential pressure signal difference between adjacent sampling times in the standardized analysis dataset and the corresponding time interval, and is used to characterize how fast the differential pressure signal changes over time.

[0076] Based on the predicted values ​​of filter media performance degradation and the material performance measurement labels, a maintenance strategy library is matched to generate a filter media performance evaluation report.

[0077] Furthermore, the predicted values ​​of filter media performance degradation are written into the evaluation sequence according to the corresponding numbers of the material performance measurement labels. Based on the predicted value range and label category corresponding to each aging level in the maintenance strategy library, each value is compared. When the predicted value of filter media performance degradation falls into the target predicted value range and the material performance measurement label is consistent with the target label category, the corresponding maintenance strategy is written into the current maintenance record. The filter media permeability, filter media flow resistance coefficient, filter media porosity, filter media pressure drop response coefficient, predicted value of filter media performance degradation, material performance measurement label and current maintenance record are combined in the order of fields to generate a filter media performance evaluation report.

[0078] It should be noted that the maintenance strategy library is set based on the filter media aging degree, filter media performance degradation prediction value range, and corresponding maintenance records in the historically labeled material sample library. It is obtained by statistically analyzing the maintenance methods corresponding to different material performance measurement labels and different filter media performance degradation prediction value ranges and then writing them in a hierarchical manner. For example, it can be set as continuous monitoring, time-limited maintenance, and immediate replacement. The prediction value range is set based on the distribution of filter media performance degradation prediction values ​​and the boundary division of filter media performance degradation prediction values ​​in the historically labeled material sample library. For example, it can be divided into 0.00 to 0.30 (corresponding to continuous monitoring), 0.30 to 0.60 (corresponding to time-limited maintenance), and 0.60 to 1.00 (corresponding to immediate replacement). The target label category is the material performance measurement label category preset in the maintenance strategy library. It is obtained by statistically classifying the material performance measurement labels corresponding to each maintenance record in the historically labeled material sample library and then writing them into the maintenance strategy library. For example, aged filter media and non-aged filter media.

[0079] This embodiment also provides a computer device applicable to the case of an artificial intelligence-based filter fault detection method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based filter fault detection method proposed in the above embodiment.

[0080] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0081] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based filter fault detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0082] In summary, this invention achieves a unified expression of spatiotemporal correlation information of multiple collection points of the filter through spatial topology aggregation, thereby enhancing the completeness of fault characterization and spatial identification capabilities. Furthermore, by constructing a material performance analysis model, it enables the determination of filter material performance parameter vectors and the identification of filter material aging levels, providing a deeper characterization of the filter's operating status and improving the accuracy of fault detection and the interpretability of results.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A filter fault detection method based on artificial intelligence, characterized in that, include: The raw physical signal set is collected, and a set of correlation relationships of filter material distribution is constructed. At the same time, the raw physical signal set is preprocessed to generate a standardized analysis dataset. The following steps are taken to extract multidimensional feature vectors from the standardized analysis dataset and perform topological aggregation and weighted fusion to generate a fused feature tensor: Statistical moments and trend terms of time-series data are extracted from the standardized analysis dataset to construct time-domain feature vectors, and spatial feature vectors are obtained by combining the filter material distribution correlation set. Frequency domain indices and time-frequency features are extracted from the standardized analysis dataset using Fast Fourier Transform, forming frequency domain feature vectors and time-frequency feature vectors. The filter media distribution correlation set and spatial feature vector are aggregated through a spatial topology aggregation mechanism, and the time-domain feature vector, frequency-domain feature vector, time-frequency feature vector and aggregated spatial feature vector are fused to generate a fused feature tensor. The multidimensional feature vectors include time-domain feature vectors, frequency-domain feature vectors, time-frequency feature vectors, and spatial feature vectors. The spatial feature vectors are topologically aggregated based on the node connections, edge directions, and edge distances in the filter material distribution relationship. The four types of feature vectors are aligned and fused according to the collection point and sampling time. A material performance analysis model is constructed, and the fused feature tensor is input into the material performance analysis model to perform material performance parameter mapping, output filter material performance parameter vector, identify the aging degree of filter material, and generate effective material performance status. The filter media performance parameter vector includes filter media permeability, filter media flow resistance coefficient, filter media porosity, and filter media pressure drop response coefficient. The degree of filter media aging is identified based on the weighted result of the four parameters, dynamic confidence threshold, and time-series verification result. The aging degree of the filter material is determined based on the aging judgment value obtained by weighting four parameters and the dynamic confidence threshold, and instantaneous interference is eliminated by the time-series verification of the material performance test label. Based on the effective material performance status, the evolution trend of filter media pore blockage rate is obtained, and a maintenance strategy library is matched to generate a filter media performance evaluation report.

2. The filter fault detection method based on artificial intelligence as described in claim 1, characterized in that, The steps for acquiring the original physical signal set and constructing the filter material distribution correlation set are as follows: Collect a set of raw physical signals including differential pressure signals and particulate matter concentration signals, and construct a set of spatial location coordinates; Based on the original set of physical signals and spatial coordinates, and combined with the physical adjacency relationships of the filter media, a set of correlation relationships for filter media distribution is constructed.

3. The filter fault detection method based on artificial intelligence as described in claim 2, characterized in that, The preprocessing of the original physical signal set includes the following steps: Based on the original physical signal set, outlier points are identified by an outlier discrimination threshold, and missing signals are filled in using linear interpolation to generate a complete physical signal set. Based on the complete physical signal set, combined with the law of conservation of mass and Darcy's law, the filter material permeability error is calculated, and high-frequency noise is suppressed by wavelet thresholding to generate a denoised physical signal set. The denoised physical signal set is normalized to generate a standardized analysis dataset.

4. The filter fault detection method based on artificial intelligence as described in claim 1, characterized in that, The steps for constructing the material performance analysis model are as follows: The feature input layer is constructed by linear mapping and positional encoding, and high-order coupling relationships and long-range dependencies are extracted based on the Transformer encoder structure and multi-head self-attention mechanism to construct the signal representation channel; Based on the mass conservation equation, Darcy's law equation, and continuity constraints, the physical behavior of the filter material is characterized in layers, and physical constraint channels are constructed. The output feature vector of the signal characterization channel is concatenated with the output feature vector of the physical constraint channel, and a feature coupling layer is constructed through linear transformation and nonlinear activation processing. A parametric regression output layer is built using a fully connected structure, and the mass conservation equation, Darcy's law equation, and continuity constraints are embedded as joint constraint terms into the loss function to construct a multi-layer physical constraint loss function. The feature input layer, signal representation channel, physical constraint channel, feature coupling layer, and parameter regression output layer are directed by sequential forward connection through a deep neural network, and a material performance analysis model is constructed by combining a multi-layer physical constraint loss function.

5. The filter fault detection method based on artificial intelligence as described in claim 1 or 4, characterized in that, The steps for inputting the fused feature tensor into the material property analysis model and mapping the material property parameters are as follows: Input the historical labeled material sample library into the material performance analysis model, and adjust the weights of the material performance analysis model through iterative optimization algorithm until the loss converges, thereby generating the trained material performance analysis model. The fused feature tensor is input into the trained material performance analysis model to perform material performance parameter mapping and output filter material performance parameter vector.

6. The filter fault detection method based on artificial intelligence as described in claim 5, characterized in that, The steps for identifying the aging degree of the filter material and generating an effective material performance status are as follows: Key physical property indicators are extracted from the filter media performance parameter vector, and the degree of filter media aging is identified by combining dynamic confidence thresholds to generate material performance measurement labels. The material performance testing labels are voted on through a time-series verification window to obtain the stability variance of the material performance testing labels, eliminate transient interference, and generate an effective material performance status.

7. The filter fault detection method based on artificial intelligence as described in claim 1, characterized in that, The steps for obtaining the evolution trend of filter media pore blockage rate based on the effective material performance status, matching it with the maintenance strategy library, and generating a filter media performance evaluation report are as follows: Based on standardized analysis datasets and effective material performance status, the evolution trend of filter media pore blockage rate is obtained by utilizing time-series evolution, and a predicted value of filter media performance degradation is generated. Based on the predicted values ​​of filter media performance degradation and the material performance measurement labels, a maintenance strategy library is matched to generate a filter media performance evaluation report.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based filter fault detection method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based filter fault detection method according to any one of claims 1 to 7.

Images