A differential pressure time sequence feature-based dry filter clogging mode recognition method

Abstract

This invention discloses a method for identifying clogging modes in a desiccant filter, belonging to the field of intelligent operation and maintenance technology for refrigeration equipment. The invention acquires real-time differential pressure time-series signals at the inlet and outlet of the desiccant filter, extracts the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signals, and constructs a multi-dimensional feature vector reflecting the degradation of the desiccant layer's pore structure. Based on this multi-dimensional feature vector, a classification and identification model is established to distinguish three failure modes: desiccant powdering and accumulation, water absorption saturation and agglomeration, and inlet filter screen clogging. This invention utilizes the periodic pressure pulsations generated by the inherent compressor stroke of the refrigeration system as the detection excitation source. By analyzing the spectral energy redistribution pattern before and after the pressure pulsations cross the desiccant layer, the identification and accurate classification of desiccant filter clogging modes are achieved. This invention effectively solves the technical problem that existing differential pressure threshold methods cannot distinguish multiple failure modes, providing differentiated maintenance decision support for refrigeration systems and reducing unplanned downtime losses.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of refrigeration equipment, specifically involving a blockage pattern identification method based on the timing characteristics of the pressure difference between the inlet and outlet of a dryer filter. It is applicable to online fault diagnosis and early warning of dryer filters in refrigeration systems such as cold storage, cold chain transportation equipment, and dehumidifiers. Background Technology

[0002] Currently, refrigeration systems widely use dryer filters to adsorb moisture and filter impurities, preventing ice blockage or dirt blockage in the throttling device. Existing commercial refrigeration systems generally install mechanical differential pressure switches at both ends of the dryer filter, triggering an alarm when the inlet and outlet pressure difference exceeds a preset threshold.

[0003] However, existing solutions have a fundamental flaw in practical applications: they can only detect "excessive differential pressure" and cannot identify the specific failure modes that cause the differential pressure to rise.

[0004] Clogging of a dryer filter is not a single physical process. In actual operation, there are at least three degradation modes with distinct mechanisms: desiccant powdering and accumulation (desiccant particles break down and accumulate, leading to increased flow resistance), water saturation and agglomeration (desiccant absorbs water and expands, causing particle adhesion and distortion of the flow channel), and inlet filter screen clogging (large particles of impurities accumulate on the inlet side). All three modes result in an increased pressure difference between the inlet and outlet, but their impact on the system and the required maintenance strategies are drastically different—powdering and accumulation requires replacing the desiccant and cleaning the downstream, water saturation requires an urgent investigation of the source of moisture in the system, while inlet filter screen clogging only requires cleaning the filter screen.

[0005] It is worth noting that these three modes differ fundamentally in their temporal response characteristics of pressure pulsation. During normal operation of the refrigeration system, the compressor's reciprocating stroke generates periodic pressure pulsations. When the desiccant layer is in a healthy state, the pressure pulsations exhibit stable attenuation characteristics after passing through the desiccant layer. In the pulsation mode, the particle gaps become smaller, making it easier for the high-frequency components of the pulsation to pass through. In the agglomeration mode, the adhered desiccant layer forms a heterogeneous channel, resulting in nonlinear distortion of the pressure response. However, existing pressure-time-based blockage detection technologies (such as the pipeline blockage early warning method disclosed in Chinese patent CN117419280A) are all designed for "pipeline wall adhesion-type blockage," focusing only on the absolute or average value of the pressure difference. They have never utilized the aforementioned differences in the pressure pulsation spectral response as an identification feature, and cannot distinguish the various degradation modes of porous media desiccants.

[0006] Therefore, how to extract the pressure pulsation spectrum response differences that reflect different degradation modes based on the temporal dynamic characteristics of the pressure difference between the inlet and outlet of the dryer filter, and establish an identification method to distinguish between three failure modes: powdering and accumulation, water absorption and agglomeration, and inlet blockage, has become a technical problem that urgently needs to be solved in the field of intelligent operation and maintenance of refrigeration systems. Summary of the Invention

[0007] In a first aspect, embodiments of the present invention provide a method for identifying clogging patterns in a dryer filter based on differential pressure time-series characteristics, including:

[0008] A pre-trained classification and recognition model is provided, which is used to distinguish three failure modes: desiccant powdering and accumulation, water absorption saturation and clumping, and inlet filter clogging.

[0009] Acquire the pressure difference timing signals of continuous sampling at the inlet and outlet of the dryer filter;

[0010] The differential pressure time-series signal is preprocessed to extract the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal, and a multi-dimensional feature vector reflecting the degradation of the desiccant layer pore structure is constructed.

[0011] The multidimensional feature vector is input into the pre-trained classification and recognition model, which outputs the current clogging pattern recognition result of the dryer filter and the corresponding maintenance instructions.

[0012] Preferably, the time-domain statistical features include:

[0013] The average pressure difference and its rate of increase characterize the flow resistance of the desiccant layer;

[0014] Kurtosis and skewness coefficients characterize the uniformity of pulverized particle packing.

[0015] Preferably, the method for extracting the frequency domain energy distribution features includes:

[0016] Perform a Fourier transform on the differential pressure time sequence signal to extract the frequency band energy corresponding to the compressor's fundamental frequency and its harmonics;

[0017] Calculate the energy ratio of the high-frequency component to the fundamental frequency band energy, and denot it as the high-frequency energy ratio;

[0018] The degree of desiccant pulverization is positively correlated with the high-frequency energy ratio.

[0019] Preferably, the method for constructing the multidimensional feature vector includes:

[0020] Based on the aforementioned time-domain statistical characteristics, static feature components characterizing the degree of clogging of the dryer filter are extracted;

[0021] Based on the frequency domain energy distribution characteristics, dynamic feature components characterizing the degradation of the desiccant layer pore structure are extracted.

[0022] The static feature components and the dynamic feature components are fused to construct the multidimensional feature vector.

[0023] Preferably, the classification and recognition model includes:

[0024] A machine learning model is used to output a first recognition result based on the multidimensional feature vector;

[0025] The rule discrimination module has built-in threshold-based discrimination logic, which is used to output a second recognition result based on the multidimensional feature vector;

[0026] The result fusion unit is used to cross-validate the first identification result and the second identification result, and output the final blocking pattern identification result.

[0027] The rule discrimination module uses the following discrimination logic to distinguish three failure modes:

[0028] When the high-frequency energy ratio exceeds the first threshold and the time domain variance increases, it is identified as a pulverization and accumulation mode.

[0029] When the time domain mean continues to rise and the pressure difference response lag time exceeds the second threshold, it is judged as the water absorption saturation agglomeration mode;

[0030] When the time domain mean rises sharply but the high-frequency energy ratio does not change significantly, it is identified as the imported filter screen being clogged.

[0031] Preferably, the machine learning model is a random forest model or a support vector machine model.

[0032] Preferably, the output maintenance instructions include:

[0033] When the system is identified as being in a powdering and accumulation mode, an early warning signal is output to the refrigeration system controller.

[0034] When the system is determined to be in a water-saturated agglomeration mode, an emergency stop command is output to the compressor controller.

[0035] When the imported filter is identified as clogged, a maintenance prompt signal is output, and the refrigeration system is kept running normally.

[0036] Secondly, embodiments of the present invention also provide a dryer filter clogging pattern recognition system based on differential pressure time-series characteristics, the system comprising:

[0037] The data acquisition module is used to acquire the differential pressure timing signals at the inlet and outlet of the dryer filter;

[0038] The signal preprocessing module is used to filter, denoise, and perform time window slicing on the differential pressure time series signal.

[0039] The feature extraction module is used to extract the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal and construct a multi-dimensional feature vector.

[0040] The pattern recognition module has a built-in classification and recognition model, which includes: a machine learning model for outputting a first recognition result based on the multidimensional feature vector; a rule discrimination module for outputting a second recognition result based on the multidimensional feature vector; and a result fusion unit for cross-validating the first recognition result and the second recognition result to output a final clogging pattern recognition result. The pattern recognition module is used to output the clogging pattern recognition result of the current dryer filter based on the multidimensional feature vector.

[0041] The maintenance decision module is used to output corresponding differentiated maintenance suggestions based on the congestion pattern identification results.

[0042] Preferably, the signal preprocessing module includes:

[0043] A low-pass filter is used to filter out high-frequency noise interference in differential pressure signals;

[0044] The sliding window unit is used to slice the differential pressure timing signal into segments with a preset time window, and there is an overlapping area between adjacent windows.

[0045] The beneficial effects of this invention are as follows: Addressing the technical deficiency of existing dryer filter clogging detection methods, which can only detect "excessive differential pressure" but struggle to identify specific failure modes, this invention provides a clogging mode identification method based on differential pressure time-series characteristics. By extracting the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal, a multi-dimensional feature vector is constructed. A classification and identification model based on random forest and discriminant logic dual verification is established, enabling the differentiation of three failure modes: desiccant powdering and accumulation, water absorption saturation and agglomeration, and inlet filter clogging. This provides differentiated maintenance decision support for refrigeration systems, effectively avoiding unplanned downtime losses and wasted maintenance costs caused by a single replacement strategy. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to the present invention and, together with the specification, serve to explain the technical solutions of the present invention. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating the steps of a method for identifying clogging patterns in a dryer filter based on differential pressure time-series characteristics, as provided in this embodiment of the invention.

[0048] Figure 2This is a schematic diagram of a dryer filter clogging pattern recognition system based on differential pressure timing characteristics, provided in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings to make the technical solution of the present invention clearer and more complete. It should be noted that the described embodiments are for illustrative purposes only and are not intended to limit the present invention. Other implementation methods that can be made by those skilled in the art based on the content of the present invention without creative effort should all fall within the protection scope of the present invention.

[0050] Furthermore, the accompanying drawings in this application are merely illustrative and not necessarily drawn to scale. The same reference numerals denote components with the same or similar functions. To clearly illustrate the present invention, specific details are provided in the following embodiments. Those skilled in the art should understand that these details are not essential for implementing the present invention, and other methods may be used to implement it without affecting the basic idea of ​​the invention.

[0051] I. Method and Steps for Identifying Clogging Patterns in Dryer Filters

[0052] The overall process flowchart is as follows: Figure 1 As shown.

[0053] This embodiment uses a cold storage refrigeration system equipped with a reciprocating compressor as an example. The system includes a compressor, a condenser, a dryer filter, an expansion valve, and an evaporator. The dryer filter is installed between the condenser outlet and the expansion valve inlet, and its interior is filled with molecular sieve desiccant to adsorb residual moisture in the refrigerant and filter impurities. A first pressure sensor and a second pressure sensor are respectively installed at the inlet and outlet of the dryer filter. Each sensor is connected to the data acquisition module of this invention via an RS-485 bus, and data communication is performed using the Modbus-RTU protocol.

[0054] S1: Real-time acquisition and preprocessing of multi-source data

[0055] (1) Differential Pressure Timing Signal Acquisition

[0056] The differential pressure timing signal is acquired through differential pressure transmitters deployed at the inlet and outlet of the dryer filter.

[0057] ① Sensor selection and installation:

[0058] A first pressure sensor (model example: Rosemount 3051S, accuracy class 0.04) is installed on the inlet side of the dryer filter, and a second pressure sensor (model example: Rosemount 3051S, accuracy class 0.04) is installed on the outlet side. Both sensors are set to a range of 0–2.5 MPa and output a 4–20 mA analog signal. To eliminate static pressure errors caused by differences in installation height, both sensors are installed at the same horizontal level, and the pressure tapping tube lengths are kept consistent (no more than 1.5 meters).

[0059] Differential pressure signals can also be measured directly using a single differential pressure transmitter (model example: EJA118E). Its high-pressure end is connected to the inlet of the dryer filter, and its low-pressure end is connected to the outlet of the dryer filter. The range is set to 0~500kPa.

[0060] ② Data collection and uploading:

[0061] Deploy an industrial data acquisition unit (such as the ICP-841 series from ICP) on the equipment side. This acquisition unit is equipped with an analog input module and simultaneously acquires analog signals from the first pressure sensor, the second pressure sensor, and the differential pressure transmitter (if configured) at a sampling frequency of 100Hz (i.e., acquiring 100 data points per second).

[0062] The data collector has a built-in circular buffer with a length of 10 seconds. Every 10 seconds of accumulated data (i.e., 1000 data points), the differential pressure time-series data (including timestamp, device ID, and differential pressure value) for that time period is encapsulated into a JSON format message and periodically sent to the data receiving interface of this invention via the workshop industrial Ethernet using an HTTP POST request. To ensure data real-time performance, the data reporting cycle is consistent with the acquisition cycle, i.e., reported once every 10 seconds.

[0063] (2) Data preprocessing

[0064] The raw data received by the data receiving interface first enters a data cleaning pipeline (a stream processing platform built on Apache Kafka and Spark Streaming) to perform the following preprocessing operations:

[0065] ① Filtering and noise reduction:

[0066] The original differential pressure signal contains high-frequency noise from sources such as compressor vibration and electromagnetic interference. The system uses a low-pass filter (Butterworth filter, cutoff frequency 10Hz, order 4) to filter the differential pressure timing signal, retaining the effective signal components that reflect the pressure pulsation characteristics.

[0067] The filter implementation code is as follows:

[0068]

[0069] ②Outlier removal:

[0070] Adopting based on The principle of statistical methods is to identify and remove outliers. For the differential pressure time series data within each time window, its mean is calculated. and standard deviation It will fall Data points outside the interval are marked as outliers and filled using linear interpolation.

[0071] ③ Time window slicing:

[0072] The continuous differential pressure time series data is divided into several time segments using a sliding window for subsequent feature extraction. In this embodiment, the window duration is set to 60 seconds (i.e., 6000 data points), and the overlap between adjacent windows is 50% (i.e., 30 seconds). The sliding window setting ensures the continuity and smoothness of feature extraction, avoiding feature jumps caused by window switching.

[0073] S2: Multidimensional Feature Vector Extraction

[0074] This step is based on the pressure difference time series data after S1 preprocessing, extracting the time-domain statistical features and frequency-domain energy distribution features that can reflect the degradation of the desiccant layer pore structure, and constructing a multi-dimensional feature vector.

[0075] (1) Extraction of Time-Domain Statistical Features

[0076] Time-domain statistical features are directly calculated from the waveform morphology of the differential pressure time-series signal without the need for frequency-domain transformation. This embodiment extracts the following three types of time-domain features:

[0077] ① The average pressure difference and its rate of rise characterize the flow resistance of the desiccant layer:

[0078] For the differential pressure time series within each time window , where n=6000, and calculate their arithmetic mean. , defined as in the formula As shown.

[0079] (1)

[0080] This indicator reflects the average pressure difference between the inlet and outlet of the desiccant filter. When the flow resistance of the desiccant layer increases... It is on the rise.

[0081] To capture the long-term trend of differential pressure, the rate of change of the average differential pressure relative to a historical window is calculated. The rate of increase is defined. The current window mean Compared with the mean of the previous window The difference divided by the time interval Specifically, as shown in formula (2).

[0082] (2)

[0083] in The window sliding step size is 30 seconds in this embodiment. A positive value indicates that the pressure difference is increasing, and its magnitude reflects the rate of increase.

[0084] ② Kurtosis and skewness coefficients of the pressure difference signal characterizing the uniformity of particle packing:

[0085] Kurtosis describes the tail weight of the differential pressure signal distribution. This embodiment uses excess kurtosis, which is the standard kurtosis minus 3, making the excess kurtosis of the normal distribution zero. For the pulverized packing mode, the desiccant particles are more uniformly packed after breakage, the fluctuation amplitude of the differential pressure signal tends to be consistent, the tail of the distribution becomes lighter, and the excess kurtosis is usually negative. The calculation formula is shown in formula (3): (3)

[0086] Subtracting 3 gives the super-kurtosis, which makes the theoretical value of the normal distribution 0.

[0087] Skewness describes the asymmetry of the differential pressure signal distribution. Under normal conditions, the differential pressure signal is approximately symmetrically distributed, with skewness close to 0; however, when blockage occurs, the differential pressure signal may exhibit a skewed distribution. The calculation formula is as follows: As shown: (4)

[0088] ③ Characterizing the pressure difference response hysteresis time of heterogeneous flow channels caused by water absorption and agglomeration:

[0089] In the water-saturated agglomeration mode, the desiccant particles adhere to each other, forming heterogeneous flow channels, resulting in a hysteresis in the pressure response. The response hysteresis time is obtained by calculating the cross-correlation function between the differential pressure signal and the compressor drive voltage signal. The compressor drive voltage signal can be directly obtained from the compressor controller, without the need for additional sensors.

[0090] Let the discrete sampling sequence of the differential pressure signal be... The compressor drive voltage signal is , ,in The number of sampling points. The normalized cross-correlation function of the two. The definition is as follows: (5)

[0091] in, This is the time delay index (unit: number of sampling points). and These are the average values ​​of the differential pressure signal and the driving voltage signal, respectively.

[0092] make To obtain the maximum value This is the response lag time of the pressure difference relative to the driving voltage (divided by the sampling frequency if the unit is seconds). When the desiccant absorbs moisture and clumps together, the refrigerant flow is obstructed, and the pressure wave propagation speed slows down. Significantly increased.

[0093] In this embodiment, the multidimensional feature vector does not include the differential pressure response hysteresis time feature. This feature is an optional extraction feature and can be used to achieve blockage pattern recognition in other embodiments.

[0094] (2) Extraction of frequency domain energy distribution features

[0095] The frequency domain energy distribution characteristics are obtained by performing a Fourier transform on the pressure difference time series signal, which can reveal the filtering characteristics of the desiccant layer on the frequency components of pressure pulsation.

[0096] ①Fourier Transform and Spectrum Calculation:

[0097] For the differential pressure time series within each time window (sampling frequency) (Window duration T = 60 seconds) is used to transform the time domain to the frequency domain using Fast Fourier Transform (FFT). In this embodiment, the FFT is calculated as follows: As shown.

[0098] (6)

[0099] Where N is the number of data points within the window (N=6000), and k is the frequency index. The corresponding frequency resolution is... .

[0100] Further, the power spectral density (PSD) is calculated: This is used to characterize the energy magnitude of each frequency component.

[0101] ②Energy extraction from fundamental and harmonic frequency bands:

[0102] The reciprocating compressor of the refrigeration system has a fixed operating frequency (50Hz in this embodiment). Each stroke of the compressor generates a periodic pressure pulsation in the refrigerant line, with a fundamental frequency of 50Hz and also including harmonic components such as 100Hz and 150Hz.

[0103] Due to frequency resolution limitations, in actual extraction, the target frequency is used as the center. The sum of energy within a narrow band.

[0104] With fundamental frequency energy For example, its calculation is as follows: As shown.

[0105] (7) Wherein, PSD(k) is the power spectral density, which is calculated from the fast Fourier transform result P(k). For the specific definition, please refer to formula (6).

[0106] ③ Calculation of high-frequency energy ratio:

[0107] The ratio of high-frequency components to the fundamental frequency band energy is defined as the high-frequency energy ratio. Among them, the high-frequency components are taken as 5 times or more the fundamental frequency (i.e., The total frequency band energy of ) The fundamental frequency band energy is taken from the vicinity of the fundamental frequency (i.e., ) frequency band energy ,but The definition is as shown in the formula. As shown.

[0108] (8)

[0109] Mechanism Explanation: When the desiccant is in a healthy state, the desiccant particles are large and loosely arranged. The high-frequency components in the pressure pulsations are significantly attenuated as they pass through the desiccant layer, thus reducing the high-frequency energy ratio. Lower. When the desiccant pulverizes, the particles break into finer particles, increasing the bulk density and reducing the interparticle spacing. This alters the desiccant layer's filtering characteristics for pressure pulsations—high-frequency components can actually penetrate the desiccant layer more easily, leading to... The concentration of the desiccant increased significantly. Therefore, the degree of desiccant pulverization is positively correlated with the high-frequency energy ratio.

[0110] This embodiment verifies this mechanism through offline experiments: under healthy desiccant conditions The value is approximately between 0.05 and 0.10; when the desiccant powdering reaches 30%, It rises to 0.25–0.35; when the degree of pulverization reaches 60%, It rose further to above 0.50.

[0111] (3) Construction of multimodal feature vectors

[0112] The extracted time-domain statistical features and frequency-domain energy distribution features are fused to construct a multi-dimensional feature vector.

[0113] ① Static characteristic components:

[0114] Based on time-domain statistical characteristics, static feature components characterizing the degree of clogging of the dryer filter are extracted. Including: average pressure difference Rate of increase of differential pressure Kurtosis coefficient Skewness coefficient .

[0115] ② Dynamic feature components:

[0116] Based on the frequency domain energy distribution characteristics, dynamic feature components characterizing the degradation of the desiccant layer pore structure are extracted. Including: fundamental frequency energy High frequency energy ratio .

[0117] The final constructed multidimensional feature vector V is represented as shown in the formula. As shown: (9)

[0118] This feature vector serves as the input to the subsequent classification and recognition model. Each dimension is Z-score standardized to eliminate the impact of dimensional differences on model training.

[0119] S3: Classification and Recognition Model Establishment and Judgment Logic

[0120] This step, based on the multidimensional feature vector constructed using S2, establishes a classification and identification model capable of distinguishing three failure modes: desiccant powdering and accumulation, water absorption saturation and clumping, and inlet filter clogging. This embodiment uses a random forest algorithm based on decision tree ensembles as the basic architecture of the classification model. Its core innovation lies in the model's built-in discrimination logic—namely, the differentiated discrimination rules for the three failure modes. The machine learning model, implemented using the random forest algorithm, serves as the core component of the classification and identification model.

[0121] (1) Construction of labeled dataset

[0122] First, a labeled dataset needs to be constructed for model training. This dataset comes from three sources:

[0123] ① Normal state data collection:

[0124] For 30 consecutive days after the new desiccant filter was installed, differential pressure time-series data were collected under normal operating conditions. During this period, the desiccant was in a healthy state, with intact and loosely arranged desiccant particles, and the differential pressure signal exhibited stable low-amplitude fluctuations. A total of 2000 normal state samples were collected, with each sample corresponding to a multidimensional feature vector within a time window (60 seconds).

[0125] ② Data collection on pulverization and accumulation status:

[0126] The desiccant pulverization process was simulated using an offline accelerated aging experiment. The desiccant was placed on a vibration table and continuously run for 500 hours under the impact of refrigerant flow, with differential pressure time-series data collected every 24 hours. Simultaneously, the desiccant pulverization rate was measured by gravimetric analysis (the percentage of pulverized particles by sieving). Samples with pulverization rates between 20% and 80% were selected as data for pulverized accumulation state, totaling 1500 samples.

[0127] ③ Data collection on water absorption saturation and clumping status:

[0128] The desiccant saturation process was simulated by quantitatively injecting water into the refrigeration system. Water was injected into the system 10 times at a rate of 5 mL per injection, and the system was run for 24 hours after each injection, collecting differential pressure data over time. Simultaneously, the degree of saturation was determined by observing whether liquid water precipitated at the desiccant outlet. Samples with a water saturation level between 30% and 100% were selected as data for the water saturation and agglomeration state, with a total of 1500 samples collected.

[0129] ④ Data collection on the status of the imported filter screen being dirty or clogged:

[0130] The clogging of the inlet filter screen was simulated by artificially adding impurities (such as welding slag and metal shavings) to the inlet side of the dryer filter. Impurities were added in 10 separate additions at a rate of 0.5g each time, with the filter running for 12 hours after each addition, and differential pressure time-series data were collected. Samples showing an increase in the mean differential pressure but no significant change in the frequency domain characteristics were selected as inlet filter screen clogging status data, for a total of 1500 samples.

[0131] ⑤ Dataset partitioning:

[0132] The 6,500 samples collected were divided into a training set (4,550 samples), a validation set (1,300 samples), and a test set (650 samples) in a ratio of 7:2:1. Each sample contains a 6-dimensional feature vector calculated in S2 and manually labeled status tags (0-normal, 1-powdering and accumulation, 2-water-saturated and clumping, 3-inlet filter screen clogged).

[0133] (2) Model training and hyperparameter optimization

[0134] This embodiment uses the Random Forest algorithm as the classification model. Random Forest consists of multiple decision trees, which improves classification accuracy and generalization ability through ensemble learning, and can output importance scores for each feature, facilitating subsequent model interpretation.

[0135] ① Model initialization:

[0136] We use the RandomForestClassifier class from the scikit-learn machine learning library. The initial parameters are set as follows: number of decision trees n_estimators = 100, maximum depth max_depth = 10, minimum number of leaf node samples min_samples_leaf = 5, and feature selection strategy is "sqrt" (i.e., randomly selecting features for each decision tree). (Splitting based on individual features).

[0137] ② Model training:

[0138] The feature matrix X_train (4550×6) and label vector y_train (4550) of the training set are input into the model, and the fit() method is called for training. During training, each decision tree draws 4550 samples from the training set based on Bootstrap sampling (random sampling with replacement), and at each node split, two features are randomly selected from six features as candidate split features. The core code for model training is as follows:

[0139] ③ Hyperparameter optimization:

[0140] Grid search combined with 5-fold cross-validation was used to optimize the hyperparameters. The search space was set as follows:

[0141] n_estimators∈[50,100,200],

[0142] max_depth∈[5,10,15,20],

[0143] min_samples_leaf∈[2,5,10].

[0144] The F1 score (harmonic mean of precision and recall) on the validation set was used as the optimization objective. The optimal hyperparameters determined by grid search were: n_estimators=150, max_depth=12, min_samples_leaf=4.

[0145] ④ Model validation and testing:

[0146] The model performance was evaluated on the validation set, with precision, recall, and F1-score calculated for each category. Validation results showed an F1-score of 0.98 for normal conditions, 0.95 for powdery accumulation, 0.96 for water-saturated agglomeration, and 0.94 for clogged inlet filter. The overall accuracy on the test set was 95.2%, indicating good generalization ability.

[0147] (3) Core discrimination logic

[0148] The core innovation of this invention lies not in the random forest algorithm itself, but in the mapping relationship between the features learned by the model and the failure modes—that is, the dual verification mechanism constituted by the machine learning model and the threshold-based discrimination logic described below. By analyzing the splitting rules of each decision tree in the trained random forest model, the following interpretable discrimination logic can be extracted:

[0149] ① Logic for identifying the powdering and accumulation mode:

[0150] When the high frequency energy ratio The time-domain variance of the differential pressure signal exceeds the first threshold (calibrated to 0.25 in this embodiment). When the variance increases (i.e., exceeds 1.5 times the normal variance), it is identified as a powdering and packing pattern.

[0151] Physical mechanism: An increase in the time-domain variance indicates a decrease in the desiccant layer's ability to attenuate high-frequency components of pressure pulsations, which is a direct manifestation of smaller particle gaps. An increase in the time-domain variance reflects intensified pressure fluctuations caused by the random motion of pulverized particles in the flow field. The combined occurrence of both uniquely identifies the pulverization and packing pattern.

[0152] Implementation code:

[0153] ② Logic for determining the water-saturated agglomeration mode:

[0154] When the average pressure difference A sustained rise (i.e., more than 5 consecutive windows) And the pressure difference response lag time When the second threshold is exceeded (in this embodiment, the value is calibrated to be 0.5 seconds), it is determined to be a water-saturated clumping mode.

[0155] Physical mechanism: Water saturation causes desiccant particles to expand and stick together, distorting the laminar flow channel of the desiccant and continuously increasing the flow resistance. (Upward); simultaneously, the propagation speed of the pressure wave in the distorted flow channel slows down, manifested as a response lag time. The increase in [value]. When both of these characteristics occur together, they can uniquely identify the water-saturated clumping pattern.

[0156] Implementation code:

[0157] ③ Logic for determining the clogging mode of the imported filter:

[0158] When the average pressure difference A sharp rise (i.e., the current window mean increases by more than 30% compared to the previous window) but with high-frequency energy ratio No significant changes (i.e.) When the filter screen is clogged, it is identified as the imported filter screen being dirty and clogged.

[0159] Physical Mechanism: Inlet filter clogging is the physical accumulation of impurities on the filter surface, which only increases local flow resistance, leading to a transient increase (sharp rise) in pressure differential. However, filter clogging does not change the pore structure of the desiccant layer, so the spectral response characteristics (high-frequency energy ratio) of the pressure pulsation remain essentially unchanged. This combination of "sharp pressure differential increase but unchanged spectrum" uniquely identifies the inlet filter clogging mode.

[0160] Implementation code:

[0161]

[0162] (4) Model inference process

[0163] In practical online applications, the model inference process is as follows:

[0164] ①Feature Extraction: After preprocessing (S1) and feature extraction (S2) of the real-time collected differential pressure time series data, a multi-dimensional feature vector for the current time window is generated. .

[0165] ②Classification prediction: Input the trained random forest model, and the model outputs the predicted probabilities for four categories: , , ,

[0166] The category corresponding to the highest probability is taken as the current state recognition result.

[0167] ③ Discriminant Logic Verification: To enhance the robustness of the recognition, the output of the random forest is cross-validated with the aforementioned threshold-based discriminant logic. If the two outputs are inconsistent and the confidence level of the discriminant logic is higher (i.e., the feature value is far from the threshold), the result of the discriminant logic shall prevail. This dual verification mechanism of "model + rule" is one of the important innovations of this embodiment.

[0168] ④ Output Results: The final identification results (failure mode type and corresponding probability) are output to the maintenance decision module to trigger subsequent maintenance instructions.

[0169]

[0170]

[0171] S4: Maintenance command output

[0172] This step generates corresponding differentiated maintenance instructions based on the blockage mode identification results output by S3. Unlike existing technologies that only output "differential pressure exceeds the limit" alarms, this invention outputs differentiated control instructions for three different failure modes and transforms maintenance suggestions into automatic control actions (rather than simple information prompts) to enhance its technical attributes.

[0173] (1) Maintenance instructions for powdering and stacking mode

[0174] When the desiccant is identified as pulverized and accumulating, it indicates that significant pulverization has occurred. The pulverized particles may enter the downstream throttling device (capillary tube or expansion valve) along with the refrigerant, posing a risk of blockage. At this stage, it is still in the warning phase, and the refrigeration system can still operate for a short period, but maintenance should be scheduled.

[0175] Output instructions:

[0176] ① Output a warning signal to the refrigeration system controller (PLC or microcontroller), triggering the yellow warning indicator light on the equipment control cabinet to alert maintenance personnel to pay attention.

[0177] ② Record the predicted remaining lifespan of the desiccant: based on the current high-frequency energy ratio The remaining usable time of the desiccant is estimated by relating it to the calibrated pulverization threshold. The estimation formula is as follows:

[0178] ③ At the same time, the early warning information is uploaded to the remote operation and maintenance platform via industrial Ethernet. The platform interface displays "Dryer Filter Powdering Early Warning - Maintenance Recommended" and records the equipment ID, early warning time, and powdering degree parameters.

[0179] (2) Maintenance instructions for water-saturated agglomeration mode

[0180] When the system is identified as being in a water-saturated and clumping mode, it indicates that the desiccant has become saturated with water or even clumped together. The desiccant has lost its ability to absorb moisture, and clumping can lead to severe blockage of the flow channels. In this case, moisture may have entered the refrigeration pipes, posing a risk of ice blockage, requiring immediate attention.

[0181] Output instructions:

[0182] ① Output an emergency stop command to the compressor controller: This disconnects the compressor's AC contactor via a relay, cutting off the compressor's power supply. Simultaneously, the command closes the solenoid valve before the expansion valve to prevent liquid refrigerant from continuing to flow into the evaporator. The code for implementing the emergency stop command is as follows:

[0183] ② Output an alarm signal to the remote operation and maintenance platform, trigger the platform's alarm rules, and send an SMS or APP notification to the operation and maintenance personnel with the following content: "[Emergency] The refrigeration system's dryer filter is saturated with water and has automatically shut down. Please immediately check the source of moisture in the system and replace the desiccant."

[0184] ③ Lock the refrigeration system start-up circuit: After an emergency shutdown, the compressor is prevented from restarting via a software lock until maintenance personnel complete the desiccant replacement and unlock it via the system reset button.

[0185] (3) Maintenance instructions for the inlet filter screen clogging mode

[0186] When the system is identified as having a clogged inlet filter, it indicates that the filter screen on the inlet side of the dryer filter is blocked by large particles of impurities. At this time, the desiccant itself is still in good condition and does not need to be replaced; simply cleaning the filter screen will restore its functionality.

[0187] Output instructions:

[0188] ① Output a maintenance prompt signal to the refrigeration system controller, triggering the green maintenance indicator light on the equipment control cabinet (different from the yellow light for powdering warning), indicating "the inlet filter needs to be cleaned".

[0189] ② Maintain normal operation of the refrigeration system: Since a dirty or clogged inlet filter does not affect the desiccant's adsorption function and will not cause immediate damage to the system, no shutdown operation is required; only a maintenance reminder will be given. The system will continue to operate normally to avoid unnecessary downtime losses.

[0190] ③ Upload the maintenance prompt information to the remote operation and maintenance platform. The platform interface will display "Dryer filter inlet screen is dirty and clogged - it is recommended to clean the screen, no need to replace the desiccant", and record the equipment ID, prompt time, and differential pressure surge value.

[0191] The final maintenance instructions are summarized in Table 1:

[0192] Table 1 Summary of Maintenance Commands

[0193] II. Dryer Filter Clogging Pattern Recognition System

[0194] Reference Figure 2The present invention also provides a clogging pattern recognition system for a dryer filter based on differential pressure time-series characteristics, the system comprising:

[0195] The data acquisition module is used to acquire the differential pressure timing signals at the inlet and outlet of the dryer filter. This module includes: a first pressure sensor and a second pressure sensor (or a single differential pressure transmitter) installed at the inlet and outlet of the dryer filter, and an industrial data acquisition unit (such as the ICP-841 series). The data acquisition module acquires the differential pressure signals at a sampling frequency of 100Hz, encapsulates the acquired data into a JSON format message every 10 seconds, and uploads it to the signal preprocessing module via industrial Ethernet.

[0196] The signal preprocessing module is used to filter, denoise, and slice the differential pressure time series signal into time windows. This module includes: a low-pass filter (Butterworth filter, cutoff frequency 10Hz, order 4) to filter out high-frequency noise interference in the differential pressure signal; and a sliding window unit to slice the differential pressure time series signal into segments with a window length of 60 seconds and a step size of 30 seconds, with a 50% overlap between adjacent windows. The preprocessed data is then passed to the feature extraction module in units of sliding windows.

[0197] The feature extraction module extracts the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal to construct a multi-dimensional feature vector. This module executes the calculation logic described in S2, including the calculation of time-domain features such as the mean differential pressure, rate of rise, kurtosis coefficient, skewness coefficient, and response lag time, as well as the calculation of frequency-domain features such as Fourier transform, fundamental frequency energy extraction, and high-frequency energy ratio calculation. The final output is a 6-dimensional feature vector.

[0198] The pattern recognition module has a built-in classification and recognition model used to output the current clogging pattern recognition result of the dryer filter based on the multi-dimensional feature vector. This classification and recognition model includes a machine learning model, a rule discrimination module, and a result fusion unit. In this embodiment, the machine learning model is implemented using a random forest classifier (n_estimators=150, max_depth=12); the rule discrimination module has built-in threshold-based discrimination logic; the result fusion unit is used to cross-validate the output of the random forest classifier with the output of the rule discrimination module and output the final recognition result. The pattern recognition module outputs the probability values ​​of four categories and the final recognition result (0-normal, 1-powdered accumulation, 2-water-saturated clumping, 3-inlet filter screen clogging).

[0199] The maintenance decision module is used to output corresponding differentiated maintenance instructions based on the clogging mode identification results. This module executes the instruction generation logic described in S4: when the pulverization and accumulation mode is identified, a warning signal is output and the remaining lifespan is estimated; when the water absorption saturation and clumping mode is identified, an emergency shutdown command is output and an alarm is pushed; when the inlet filter is identified as clogged, a maintenance prompt signal is output and the system operation is maintained.

[0200] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0201] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0202] In this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this invention are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0203] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0204] It should also be noted that in the system and method of the present invention, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present invention.

[0205] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0206] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for identifying clogging patterns in a dryer filter based on differential pressure time-series characteristics, characterized in that, include: A pre-trained classification and recognition model is provided, which is used to distinguish three failure modes: desiccant powdering and accumulation, water absorption saturation and clumping, and inlet filter clogging. Acquire the pressure difference timing signals of continuous sampling at the inlet and outlet of the dryer filter; The differential pressure time-series signal is preprocessed to extract the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal, and a multi-dimensional feature vector reflecting the degradation of the desiccant layer pore structure is constructed. The multidimensional feature vector is input into the pre-trained classification and recognition model, which outputs the current clogging pattern recognition result of the dryer filter and the corresponding maintenance instructions.

2. The method according to claim 1, characterized in that, The time-domain statistical features include: The average pressure difference and its rate of increase characterize the flow resistance of the desiccant layer; Kurtosis and skewness coefficients characterize the uniformity of pulverized particle packing.

3. The method according to claim 1, characterized in that, The method for extracting the frequency domain energy distribution features includes: Perform a Fourier transform on the differential pressure time sequence signal to extract the frequency band energy corresponding to the compressor's fundamental frequency and its harmonics; Calculate the energy ratio of the high-frequency component to the fundamental frequency band energy, and denot it as the high-frequency energy ratio; The degree of desiccant pulverization is positively correlated with the high-frequency energy ratio.

4. The method according to claim 1, characterized in that, The method for constructing the multidimensional feature vector includes: Based on the aforementioned time-domain statistical characteristics, static feature components characterizing the degree of clogging of the dryer filter are extracted; Based on the frequency domain energy distribution characteristics, dynamic feature components characterizing the degradation of the desiccant layer pore structure are extracted. The static feature components and the dynamic feature components are fused to construct the multidimensional feature vector.

5. The method according to claim 1, characterized in that, The classification and recognition model includes: A machine learning model is used to output a first recognition result based on the multidimensional feature vector; The rule discrimination module has built-in threshold-based discrimination logic, which is used to output a second recognition result based on the multidimensional feature vector; The result fusion unit is used to cross-validate the first identification result and the second identification result, and output the final blocking pattern identification result. The rule discrimination module uses the following discrimination logic to distinguish three failure modes: When the high-frequency energy ratio exceeds the first threshold and the time domain variance increases, it is identified as a pulverization and accumulation mode. When the time domain mean continues to rise and the pressure difference response lag time exceeds the second threshold, it is judged as the water absorption saturation agglomeration mode; When the time domain mean rises sharply but the high-frequency energy ratio does not change significantly, it is identified as the imported filter screen being clogged.

6. The method according to claim 5, characterized in that, The machine learning model is either a random forest model or a support vector machine model.

7. The method according to claim 1, characterized in that, Output maintenance instructions include: When the system is identified as being in a powdering and accumulation mode, an early warning signal is output to the refrigeration system controller. When the system is determined to be in a water-saturated agglomeration mode, an emergency stop command is output to the compressor controller. When the imported filter is identified as clogged, a maintenance prompt signal is output, and the refrigeration system is kept running normally.

8. A clogging pattern recognition system for a dryer filter based on differential pressure time-series characteristics, characterized in that, The system for implementing the method of any one of claims 1-7 comprises: The data acquisition module is used to acquire the differential pressure timing signals at the inlet and outlet of the dryer filter; The signal preprocessing module is used to filter, denoise, and perform time window slicing on the differential pressure time series signal. The feature extraction module is used to extract the time-domain statistical features and frequency-domain energy distribution features of the differential pressure signal and construct a multi-dimensional feature vector. The pattern recognition module, which incorporates the classification and recognition model described in claim 5, is used to output the clogging pattern recognition result of the current dryer filter based on the multidimensional feature vector. The maintenance decision module is used to output corresponding differentiated maintenance suggestions based on the congestion pattern identification results.

9. The system according to claim 8, characterized in that, The signal preprocessing module includes: A low-pass filter is used to filter out high-frequency noise interference in differential pressure signals; The sliding window unit is used to slice the differential pressure timing signal into segments with a preset time window, and there is an overlapping area between adjacent windows.

Images