Automobile air conditioner filter element health prediction system based on multi-source data and deep learning

Abstract

The invention relates to the field of vehicle intelligent health monitoring, and discloses an automobile air conditioner filter element health prediction system based on multi-source data and deep learning, and the system comprises a multi-source sensor data collection module which is used for collecting environment data and vehicle state data through a multi-source sensor, and collecting filter pressure difference data through a pressure difference sensor; the edge calculation module is connected with the multi-source sensor data acquisition module; the deep learning prediction module is connected with the edge calculation module and is provided with a double-flow neural network model; and the closed-loop control module is connected with the deep learning prediction module. A double-flow neural network model comprising a real-time data flow branch and a historical data flow branch is constructed, and an attention gating fusion unit is used for dynamically adjusting the weights of the two branches according to environmental risk indexes, so that short-term dynamic working condition characteristics and long-term physical aging trends of the filter element can be captured at the same time; and the robustness and the accuracy of the residual life prediction of the filter element are improved.

Description

Technical Field

[0001] This invention relates to the field of vehicle intelligent health monitoring technology, specifically to a health prediction system for automotive air conditioning filters based on multi-source data and deep learning. Background Technology

[0002] Automotive cabin air filters (filter elements) are key components for ensuring in-vehicle air quality and intercepting particulate matter and harmful gases. With the increasing intelligence of vehicles and growing public awareness of healthy travel, monitoring the performance of cabin air filters has become a crucial aspect of in-vehicle environmental management. Currently, cabin air filter maintenance is primarily based on vehicle mileage or fixed time intervals for replacement. Some high-end models are beginning to adopt electronic monitoring solutions based on differential pressure principles. This involves placing pressure sensors at both ends of the filter to monitor changes in pressure difference and determine the filter's dust-holding capacity. This type of technology achieves a degree of digital monitoring of the filter's condition, aiming to ensure the ventilation efficiency of the air conditioning system under different operating conditions and reduce the risk of microbial growth.

[0003] However, existing monitoring methods still have certain technical limitations in practical applications. Traditional periodic replacement strategies ignore the differences in actual vehicle operating environments, leading to resource waste in clean environments or filter failure under heavy pollution conditions. While monitoring methods based on a single differential pressure threshold introduce physical detection, their values ​​are easily affected by factors such as air conditioning fan speed, vehicle speed, and air humidity, resulting in significant data fluctuations and noise, making it difficult to reflect the true clogging status of the filter. Furthermore, existing systems mostly use instantaneous state judgment logic, lacking the ability to comprehensively analyze multi-source environmental data and historical aging trends. This makes it impossible to accurately predict the remaining lifespan of the filter or proactively adjust the air conditioning system's operating strategy based on the filter's health status, thus failing to meet the needs of precise maintenance and proactive protection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a health prediction system for automotive air conditioning filters based on multi-source data and deep learning. This system solves the problems of existing automotive air conditioning filter monitoring technologies being greatly affected by environmental noise under complex operating conditions, being unable to accurately predict remaining lifespan, and lacking active protection mechanisms.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a health prediction system for automotive air conditioning filters based on multi-source data and deep learning, which integrates multi-source sensing, edge computing, deep learning prediction and closed-loop control functions.

[0006] The system primarily comprises a multi-source sensor data acquisition module, an edge computing module, a deep learning prediction module, and a closed-loop control module. The multi-source sensor data acquisition module is responsible for collecting environmental data, vehicle status data, and filter differential pressure data from various heterogeneous sensors. The edge computing module is equipped with a Kalman filter algorithm to perform quality control, noise reduction, outlier detection, and multi-source data time alignment on the acquired data, thereby outputting a filtered state vector. The deep learning prediction module deploys a dual-stream neural network model, which includes real-time data stream branches and historical data stream branches, respectively used to receive the filtered state vector. It fuses the features of the two branches through an attention mechanism, ultimately outputting a filter health status score and remaining lifespan. The closed-loop control module compares the output filter health status score with a preset threshold and generates control commands based on the comparison results to dynamically adjust the air conditioning operation mode.

[0007] The system's multi-source sensor data acquisition module is specifically configured at the hardware level. Specifically, the differential pressure sensor monitors the filter differential pressure via an SPI interface at a preset sampling frequency; the environmental sensor group includes PM... 2.5 The laser sensor, VOC sensor, and temperature and humidity sensor are connected via UART interface, I... 2 Data is transmitted using the CAN bus protocol and the single-bus protocol; the vehicle status data acquisition unit collects air conditioning operating parameters and operating time at preset intervals via the CAN bus.

[0008] At the edge computing processing level, this invention employs an improved Kalman filter algorithm, which includes a prediction phase and an update phase. In the prediction phase, the system predicts the state at the next moment based on the current state vector, the system state transition matrix, and the control input. The core innovation of this phase lies in executing a condition-based dynamic adjustment strategy: the system adjusts the dynamic process noise covariance parameter in real time according to the air conditioning operating mode. Specifically, when the air conditioning operating mode is detected to be external circulation, the system determines that external interference has increased, and thus increases the value of the dynamic process noise covariance parameter to improve the algorithm's adaptability to external environmental interference. Subsequently, the adjusted parameter is used to predict the covariance matrix.

[0009] During the update phase, the algorithm calculates the Kalman gain and sets upper and lower saturation limits to prevent system oscillations caused by excessive gain. Simultaneously, it calculates the measurement residual; if the residual exceeds a preset standard deviation multiple threshold, the current measurement value is considered an outlier, triggering a redundant sensor data replacement mechanism. Finally, the state vector is updated based on the Kalman gain and passed to the next module. Furthermore, the edge computing module employs compensation and correction logic in the quality control stage, performing physical characteristic corrections on the monitored values ​​based on environmental parameters, specifically including adjustments for PM2.5 under high humidity conditions. 2.5The readings are compensated for moisture interference, and temperature compensation is initiated for VOC sensor data in high-temperature environments.

[0010] At the deep learning prediction level, the system's real-time data stream branch is responsible for processing short-term features. This branch normalizes the filtered real-time data and performs sliding window sampling to construct time-series data segments. Subsequently, a gated recurrent unit network is used to capture PM. 2.5 The short-term dynamic characteristics of parameters such as concentration are analyzed, with the number of gated recurrent unit network structures set to two to balance computational efficiency and feature extraction capability. Finally, pressure difference and temperature / humidity data are fused through a multi-input fully connected layer to output a real-time feature vector.

[0011] The system's historical data stream branch handles long-term aging characteristics. This branch receives cumulative operating time, historical pollution exposure data, and filter status labels. The system quantifies the cumulative pollution level by calculating a pollution load index, which is obtained by analyzing PM2.5 levels over a statistical period. 2.5 The value is obtained by summing the products of the concentration values ​​and the VOC concentration values. Subsequently, a temporal convolutional network is used to capture the long-term dependency between the cumulative operating time and the filter lifetime, outputting a historical feature vector.

[0012] This invention introduces an attention-gated fusion unit in the feature fusion stage. This attention-gated fusion unit first establishes environmental risk indicators, including PM2.5. 2.5 Risk indicators and differential pressure risk indicators. When real-time PM... 2.5 When the concentration or real-time pressure difference exceeds its respective set threshold, the corresponding risk indicator is triggered. The system calculates dynamic weights based on the environmental risk indicators: when any risk indicator is triggered, indicating a sudden change in the current operating condition or a high risk, the system automatically increases the weight value assigned to the real-time data stream branch, making it higher than the weight value assigned to the historical data stream branch; when no risk indicator is triggered, a balanced weight value is assigned to both branches. Finally, based on the calculated dynamic weights, the real-time feature vector and the historical feature vector are weighted and summed to output the final feature vector.

[0013] The deep learning prediction module also includes an output decision unit, which inputs the final feature vector into the fully connected layer and outputs the filter's health status probability distribution, health status score, and remaining lifetime. When the output filter clogging probability exceeds a preset alarm threshold, the system triggers an emergency decision signal sent to the closed-loop control module.

[0014] At the control level, the closed-loop control module implements an active protection strategy. When the comparison results show that the filter health status score is lower than a preset threshold, a control command is automatically generated to adjust the air conditioning mode. This control command drives the air conditioning system to perform operations such as switching the air conditioning operation mode to internal circulation or reducing the external circulation fan speed, thereby protecting the filter structure and maintaining system operational safety.

[0015] This invention provides a health prediction system for automotive air conditioning filters based on multi-source data and deep learning. It has the following beneficial effects:

[0016] 1. This invention constructs a dual-stream neural network model that includes a real-time data stream branch and a historical data stream branch, and uses an attention-gated fusion unit to dynamically adjust the weights of the two branches according to environmental risk indicators. This enables the simultaneous capture of both the short-term dynamic operating characteristics and long-term physical aging trends of the filter element, thereby solving the technical problem that traditional single prediction models cannot take into account both sudden high-pollution scenarios and long-term gradual clogging processes, and improving the robustness and accuracy of filter element remaining life prediction.

[0017] 2. This invention, by configuring an improved Kalman filter algorithm in the edge computing module, and in particular by introducing a dynamic adjustment strategy based on operating conditions, adjusts the process noise covariance parameter in real time according to the air conditioning internal and external circulation modes. Combined with temperature and humidity physical compensation logic, it achieves the effect of high-precision cleaning and state estimation of multi-source heterogeneous data. This solves the problem of high noise and drift in sensor data caused by external airflow disturbances and changes in ambient temperature and humidity during vehicle operation, ensuring the physical consistency of input data and providing a reliable data foundation for accurate prediction of subsequent models.

[0018] 3. This invention establishes a decision feedback mechanism based on health status scores and blockage probability through a closed-loop control module. When a serious blockage of the filter element or a low score is detected, it can automatically generate an active protection command to switch the internal circulation or reduce the fan speed, thereby achieving real-time intervention and protection of the air conditioning system. This solves the problem that existing technologies can only provide passive alarm prompts but cannot prevent physical damage caused by filter element overload or excessive motor load, and realizes automated preventive maintenance of automotive air conditioning systems. Attached Figure Description

[0019] Figure 1 This is a system module architecture diagram of the present invention;

[0020] Figure 2 This is a diagram of the two-stream neural network structure of the present invention;

[0021] Figure 3 This is a flowchart of the closed-loop control of the present invention. Detailed Implementation

[0022] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a health prediction system for automotive air conditioning filters based on multi-source data and deep learning, comprising: a multi-source sensor data acquisition module 10, an edge computing module 20, a deep learning prediction module 30, and a closed-loop control module 40. These modules work together to realize a complete technical process from physical environment perception, data processing, state prediction to execution feedback.

[0023] The multi-source sensor data acquisition module 10 is used to acquire multi-dimensional operating data of the automotive air conditioning system. The multi-source sensor data acquisition module 10 integrates various heterogeneous sensor devices and communication interfaces at the physical level. Specifically, the multi-source sensor data acquisition module 10 includes differential pressure sensors arranged on the inlet and outlet sides of the air conditioning filter, used to monitor the pressure difference across the filter in real time.

[0024] The multi-source sensor data acquisition module 10 also includes an environmental sensor group, which is installed at a specific location inside the vehicle's air intake or passenger compartment, and includes PM2.5 sensors. 2.5 Laser sensors, VOC sensors, and temperature and humidity sensors are used to collect data on particulate matter concentration, volatile organic compound concentration, and ambient temperature and humidity in the air, respectively. In addition, the multi-source sensor data acquisition module 10 is connected to the vehicle network via the vehicle's CAN bus interface to read vehicle status data, including air conditioning operating mode, fan speed, vehicle speed, and cumulative air conditioning usage time.

[0025] The edge computing module 20 is communicatively connected to the multi-source sensor data acquisition module 10, and is used to receive and preprocess the acquired raw data. The edge computing module 20 is internally configured with a microprocessor or embedded computing unit, running data cleaning algorithms and improved Kalman filtering algorithms. The edge computing module 20 first timestamps the received multi-source heterogeneous data to unify the data's time reference.

[0026] Edge computing module 20 analyzes PM based on ambient temperature and humidity data. 2.5 The laser sensor readings are compensated and corrected to eliminate measurement errors in high humidity environments. The edge computing module 20 uses a pre-set Kalman filter algorithm to iteratively calculate the differential pressure data and environmental parameter data. Through state prediction and measurement update steps, it eliminates sensor noise and random interference, removes abnormal fluctuation values, and generates a smooth and physically consistent filtered state vector.

[0027] The deep learning prediction module 30 is connected to the output of the edge computing module 20, receiving the filtered state vector as input. The deep learning prediction module 30 is equipped with a trained two-stream neural network model. This two-stream neural network model is logically divided into a real-time data stream branch and a historical data stream branch.

[0028] The deep learning prediction module 30 utilizes a real-time data stream branch to process the current differential pressure and environmental parameter sequences, extracting short-term dynamic features reflecting the current operating condition of the filter. Simultaneously, the deep learning prediction module 30 utilizes a historical data stream branch to process cumulative operating time and historical pollution exposure data, extracting long-term trend features reflecting the filter's aging degree. The deep learning prediction module 30 calculates environmental risk indicators through its internal attention-gated fusion unit and dynamically adjusts the fusion weights of the two branch features accordingly, ultimately outputting the filter's health status score, predicted remaining lifespan, and probability of clogging.

[0029] The closed-loop control module 40 is connected to the deep learning prediction module 30 and the vehicle air conditioning actuator. The closed-loop control module 40 executes decision logic based on the prediction results. The closed-loop control module 40 receives the filter health status score and clogging probability output by the deep learning prediction module 30 and compares these values ​​with preset maintenance thresholds and risk thresholds.

[0030] When the filter health status score falls below a preset threshold or the probability of clogging exceeds a warning standard, the closed-loop control module 40 generates a corresponding control command. This control command is sent to the air conditioning controller via the vehicle communication network, driving the air conditioning system actuators to perform actions, including switching between internal and external circulation modes, adjusting fan speed, or limiting the fresh air intake. Simultaneously, the closed-loop control module 40 can send status information to the vehicle display terminal, displaying the current filter health level and maintenance recommendations.

[0031] The multi-source sensor data acquisition module 10 is responsible for performing the system's data sensing tasks. This multi-source sensor data acquisition module 10 integrates various types of sensor hardware and communication interfaces to acquire multi-dimensional data reflecting the operating environment and working status of the automotive air conditioning filter.

[0032] The multi-source sensor data acquisition module 10 is equipped with a differential pressure sensor at the hardware acquisition layer. This differential pressure sensor is configured to communicate with the system via an SPI interface. To continuously monitor the resistance changes across the filter, the differential pressure sensor is set to a sampling frequency of 1Hz. During the acquisition process, the system has a built-in abnormal fluctuation marking logic. When the detected fluctuation of the differential pressure value exceeds ±2Pa, the data at that moment is automatically marked as abnormal for identification in subsequent processing stages.

[0033] The multi-source sensor data acquisition module 10 also includes an environmental sensor group, which integrates PM... 2.5 Laser sensors, VOC sensors, and temperature and humidity sensors. Among them, PM... 2.5The laser sensor transmits data via a UART interface and is set to output concentration values ​​60 times per minute. To smooth instantaneous readings, the sensor or acquisition end is configured to calculate a 10-second moving average based on the output concentration values, using this moving average as the effective PM2.5 concentration value. 2.5 Concentration data. The VOC sensor connects via the I²C protocol, with a data acquisition cycle of once every 30 seconds. Considering the impact of high temperatures on the chemical sensor readings, the VOC sensor is configured to activate a temperature compensation mechanism in high-temperature environments to correct the acquired data. The temperature and humidity sensor connects to the system via a single-bus protocol, with an upload cycle of once every 5 seconds. The uploaded data is the value after relative humidity (RH) and temperature (T) compensation calibration.

[0034] The multi-source sensor data acquisition module 10 acquires vehicle status data via the vehicle CAN bus interface. The vehicle status data acquisition unit is configured to read air conditioning operating parameters and operating duration at 500ms intervals. These air conditioning operating parameters include the air conditioning operating mode and fan speed settings. During all the above acquisition tasks, the multi-source sensor data acquisition module 10 attaches a precise timestamp and corresponding device ID to each acquired data frame, thus providing a spatiotemporal alignment reference for subsequent multi-source data fusion. For SPI, UART, and I... 2 The implementation of the basic physical layer of communication protocols such as C and CAN bus is a well-known technology familiar to those skilled in the art, and will not be described in detail here.

[0035] The edge computing module 20, serving as an intermediate processing unit connecting the perception layer and the decision layer, primarily undertakes tasks such as quality control of multi-source data, signal compensation, and data denoising based on an improved Kalman filter algorithm. The edge computing module 20 executes data processing at cloud or vehicle-side edge nodes to eliminate sensor noise, correct abnormal fluctuations, and extract engineering features, providing reliable input vectors for subsequent deep learning models.

[0036] The edge computing module 20 first performs strict quality control on the data uploaded by the multi-source sensor data acquisition module 10. The edge computing module 20 monitors the communication status of the data, and when it identifies data packets with sensor disconnections or communication timeouts, it discards them as invalid data. The edge computing module 20 performs physical compensation correction on the monitored values ​​based on synchronously acquired environmental parameters. (Regarding PM...) 2.5 The sensor exhibits reading drift characteristics in high humidity environments. When the relative humidity (RH) exceeds 80%, the edge computing module 20 adjusts the PM reading based on the current humidity value. 2.5The readings are compensated for moisture interference. Considering the characteristics of VOC sensors in high-temperature environments, when the monitored ambient temperature exceeds 35℃, the edge computing module 20 initiates a temperature compensation algorithm based on the current temperature value to correct the VOC concentration data. On this basis, the edge computing module 20 further extracts engineering features, calculates the differential pressure change rate, PM2.5 concentration, and other parameters. 2.5 The data includes cumulative data and air conditioning load index, and time alignment of multi-source data.

[0037] The edge computing module 20 runs an improved Kalman filter algorithm for real-time data denoising and outlier detection. The algorithm's execution process specifically includes the following stages:

[0038] During the initialization phase, the edge computing module 20 sets initial parameters based on prior knowledge and sensor characteristics. The system sets a state vector, which includes differential pressure and PM2.5. 2.5 The system uses raw data from concentration and VOC concentration sensors. Simultaneously, based on sensor accuracy and environmental noise, the system presets an initial state covariance matrix, sets the process noise covariance and measurement noise covariance, and designs the system state transition matrix and observation matrix according to the working model of the air conditioning filter.

[0039] During the prediction phase, the edge computing module 20 predicts the state at the next moment based on the current state and updates the covariance matrix. The state prediction follows the formula:

[0040] ;

[0041] In the formula, for The prior state estimate at time t; This is the state transition matrix; for The posterior state estimate at time t; To control the input matrix; for The control input at any time includes data on changes in air conditioning fan speed and ambient temperature.

[0042] Covariance prediction follows the formula:

[0043] ;

[0044] In the formula, The prior error covariance matrix; for The posterior error covariance matrix at time t; This is the transpose of the state transition matrix; Let be the process noise covariance.

[0045] During this process, the edge computing module 20 executes a dynamic adjustment strategy based on operating conditions. The system adjusts the process noise covariance in real time according to the current operating mode of the vehicle's air conditioning. When the air conditioner is in external circulation mode, it determines that environmental interference has increased, and the system automatically increases the process noise covariance. The value is adjusted to adapt to the uncertainties brought about by the external environment.

[0046] During the update phase, the edge computing module 20 uses observed data to correct the predicted values. This phase first calculates the Kalman gain. The calculation formula is as follows:

[0047] ;

[0048] In the formula, Kalman gain; The observation matrix; This is the transpose of the observation matrix; To measure the noise covariance.

[0049] Edge computing module 20 calculates the Kalman gain. Perform saturation limiting and set upper and lower limits for gain. (e.g., 0.9) and (e.g., 0.1), ensure This is to prevent drastic fluctuations in system output.

[0050] Subsequently, the system calculates the measurement residuals and performs outlier detection and state updates based on these residuals. The state update follows the formula:

[0051] ;

[0052] In the formula, for The posterior state estimate at time t; for The actual measured value at that moment; To measure the residual.

[0053] In this step, edge computing module 20 is applied. The criteria are used to remove outliers. If the calculated measurement residual value exceeds... (3 times the standard deviation), the system will display the current measurement value The value is marked as an anomaly, and the redundant sensor data replacement mechanism is triggered or the measurement is discarded directly to prevent abnormal noise from contaminating the state estimation.

[0054] Finally, the edge computing module 20 updates the covariance matrix using the following formula:

[0055] ;

[0056] In the formula, for The posterior error covariance matrix at time t; It is an identity matrix.

[0057] After the above prediction and update steps, the edge computing module 20 outputs a filtered state vector. It is then passed to the subsequent deep learning prediction module 30 for the evaluation of the filter's health status.

[0058] The deep learning prediction module 30 is connected to the output of the edge computing module 20 and is used to evaluate the health status of the filter based on a dual-stream neural network architecture. The deep learning prediction module 30 logically divides the input data stream into a real-time data stream branch and a historical data stream branch, builds models for each branch to capture short-term dynamic features and long-term aging trends, and dynamically fuses the feature outputs of the two branches through an attention mechanism.

[0059] In the real-time data stream branch, the deep learning prediction module 30 first receives the real-time state vector after noise reduction processing by the edge computing module 20. The deep learning prediction module 30 is configured to normalize this real-time data, standardizing the original data to a uniform numerical range [0,1] to eliminate dimensional differences. Subsequently, the deep learning prediction module 30 performs a sliding window sampling operation, with a fixed time window (… Constructing time series data fragments For time series modeling, this branch uses a gated recurrent unit (GRU) network to capture PM. 2.5 The GRU network exhibits short-term dynamic characteristics such as rapid concentration increases. To reduce computational complexity while maintaining performance, a simplified gating structure is adopted, setting the number of gates to two. After GRU processing, the deep learning prediction module 30 fuses multi-source sensor data such as differential pressure, temperature, and humidity through a multi-input fully connected layer, outputting a real-time feature vector. .

[0060] In the historical data stream branch, the deep learning prediction module 30 receives the cumulative working time of the filter, historical pollution exposure data, and filter status labels. The historical pollution exposure data includes PM2.5 levels over a specific past time period (e.g., N days). 2.5 Daily average values, cumulative VOC exposure, and records of extreme temperature and humidity events. This branch is configured to calculate the Pollution Load Index (PLI), using the following formula:

[0061] ;

[0062] In the formula, Pollution load index; The statistical time period; For the first PM within a time unit 2.5 Concentration value; For the first VOC concentration values ​​within a time unit.

[0063] This branch also analyzes PM using moving average or exponential smoothing algorithms. 2.5 This branch analyzes the long-term trend of filter lifespan and standardizes historical data. In terms of neural network architecture, it employs a Temporal Convolutional Network (TCN) to build a long-term trend model, capturing the non-linear dependency between cumulative working time and filter lifetime. It also predicts the filter's remaining lifetime (RUL) and health status level, ultimately outputting a historical feature vector. .

[0064] The deep learning prediction module 30 incorporates an attention-gated fusion unit to dynamically adjust the weights of real-time and historical branches based on environmental risk. This unit first establishes environmental risk indicators based on real-time data. The deep learning prediction module 30 sets PM... 2.5 Concentration threshold (e.g., 150 μg / m 3 and differential pressure threshold (e.g., 400Pa).

[0065] Then the corresponding The risk indicators are as follows:

[0066] ;

[0067] Pressure differential risk indicators:

[0068] ;

[0069] In the formula, For environmental risk indicators (binary values, 1 indicates the presence of risk, 0 indicates no risk); For the current real-time PM 2.5 Concentration value; For PM 2.5 Concentration threshold (e.g.) =150μg / m 3 ); This is the current real-time differential pressure value (the pressure difference across the filter). For differential pressure threshold (e.g.) =400 Pa).

[0070] Based on the aforementioned risk indicators, the attention-gated fusion unit calculates the dynamic weights of the real-time branches according to the following logic. :

[0071] The corresponding dynamic weights can be calculated using the following formula:

[0072] ;

[0073] In the formula, The dynamic weight of the real-time branch; As a pressure differential risk indicator; For PM 2.5 Risk indicators.

[0074] Accordingly, the weights of the historical branches are calculated as follows:

[0075] ;

[0076] In the formula, The weights of the branches in the real-time data stream; The weights of the branches in the historical data stream; For PM 2.5 Risk indicators; This is a pressure differential risk indicator.

[0077] The attention-gated fusion unit uses the calculated weights to perform gated fusion of the feature vectors from the two branches to obtain the final feature vector. The fusion formula is as follows:

[0078] ;

[0079] In the formula, This is the final feature vector after fusion; For real-time feature vectors; This is the historical feature vector.

[0080] The deep learning prediction module 30 will generate the final feature vector. The input is sent to the output decision unit. This output decision unit is configured to output the filter health status probability distribution (including "normal" probability and "blocked" probability), the filter health status score (0-100 points), and the predicted remaining lifetime (RUL). The output decision unit has built-in emergency decision triggering logic; when the calculated blockage probability is greater than 0.7, the output decision unit generates and outputs an emergency decision signal to the closed-loop control module 40. The implementation of the basic network structure of GRU, TCN, and the fully connected layer is a well-known deep learning technology and will not be described in detail here.

[0081] The closed-loop control module 40, as the system's decision-making and execution unit, establishes a communication connection with the deep learning prediction module 30 and the vehicle air conditioning control system (HVAC). The main function of the closed-loop control module 40 is to execute specific control strategies based on the prediction results output by the front-end model to protect the air conditioning system and prompt the user for maintenance.

[0082] The closed-loop control module 40 is configured to receive in real-time filter health status scores, remaining lifetime (RUL) prediction values, and emergency decision signals output by the deep learning prediction module 30. The closed-loop control module 40 has a pre-set filter maintenance threshold, which corresponds to the critical state score when the filter cartridge dust holding capacity reaches its limit or the pressure differential is too large. The closed-loop control module 40 executes comparison logic, comparing the received real-time health status score with the maintenance threshold.

[0083] When the comparison result shows that the real-time health status score is less than the maintenance threshold, or when the closed-loop control module 40 receives an emergency decision signal from the deep learning prediction module 30 (i.e., the congestion probability is greater than 0.7), the closed-loop control module 40 immediately generates an active control command. This active control command is sent to the air conditioning electronic control unit (ECU) via the vehicle CAN bus. For different vehicle operating conditions, the closed-loop control module 40 executes the following specific lower-level control operations:

[0084] When the vehicle is in external air circulation mode and a high risk of filter blockage is detected, the closed-loop control module 40 sends a mode switching command, forcibly switching the air conditioning intake mode from external circulation to internal circulation. This operation aims to cut off the continuous input of external pollution sources, reduce the pressure load on the filter intake side, and prevent physical deformation or damage to the filter element due to excessive pressure difference.

[0085] If the vehicle must remain in external air circulation mode (e.g., when fresh air is needed based on the CO2 concentration inside the vehicle), the closed-loop control module 40 sends a fan speed adjustment command to automatically reduce the speed of the air conditioning blower. By reducing the airflow velocity through the filter, intake resistance is reduced, thereby protecting the filter structure while maintaining minimum ventilation.

[0086] The closed-loop control module 40 is also responsible for data interaction with the human-machine interface (HMI). The closed-loop control module 40 encapsulates the final health status score, remaining lifespan data, and current filter status level (e.g., good, lightly polluted, replacement recommended, severely clogged) into a display message and sends it to the vehicle's central control screen or the accompanying mobile terminal APP. When the aforementioned active control command is triggered, the closed-loop control module 40 will simultaneously send an alarm signal and display a maintenance reminder on the display terminal, advising the user to replace the air conditioning filter in a timely manner. The implementation of the CAN bus communication protocol and the specific drive circuit of the air conditioning ECU are well-known technologies to those skilled in the art and will not be described in detail here.

Claims

1. A health prediction system for automotive air conditioning filters based on multi-source data and deep learning, characterized in that, include: The multi-source sensor data acquisition module is used to acquire environmental data, vehicle status data, and filter differential pressure data through multi-source sensors. An edge computing module is connected to the multi-source sensor data acquisition module and is equipped with a Kalman filter algorithm for quality control, noise reduction, outlier detection, and multi-source data time alignment of the acquired data to obtain a filtered state vector. A deep learning prediction module is connected to the edge computing module and is equipped with a dual-stream neural network model. The dual-stream neural network model includes a real-time data stream branch and a historical data stream branch, which are used to receive the filtered state vector and fuse the features of the two branches through an attention mechanism to output the filter health status score and remaining lifetime. The closed-loop control module, connected to the deep learning prediction module, compares the output filter health status score with a preset threshold and generates control commands based on the comparison results to dynamically adjust the air conditioning operation mode.

2. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The multi-source sensor data acquisition module includes: The differential pressure sensor monitors the filter differential pressure at a preset sampling frequency via an SPI interface. Environmental sensor array, including PM 2.5 Laser sensors, VOC sensors, and temperature and humidity sensors, including PM2.

5. 2.5 The laser sensor outputs concentration values ​​via a UART interface, while the VOC sensor uses I... 2 Data is acquired via the C protocol, and the temperature and humidity sensor uploads data via a single-bus protocol. The vehicle status data acquisition unit collects air conditioning operating parameters and operating time at preset intervals via the CAN bus.

3. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The Kalman filter algorithm specifically includes a prediction phase and an update phase; The prediction phase specifically includes the following steps: Predict the state at the next moment based on the current state vector, the system state transition matrix, and the control input; A dynamic adjustment strategy based on operating conditions is implemented, and the dynamic process noise covariance parameter is adjusted in real time according to the air conditioning operation mode. When the air conditioning operation mode is detected to be external circulation, the value of the dynamic process noise covariance parameter is increased to adapt to external environmental interference. The covariance matrix is ​​predicted using the adjusted dynamic process noise covariance parameter.

4. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 3, characterized in that, The update phase specifically includes the following steps: Calculate the Kalman gain and set upper and lower limits for saturation of the Kalman gain; Calculate the measurement residual. If the measurement residual exceeds a preset standard deviation multiple threshold, mark the current measurement value as an outlier and trigger a redundant sensor data replacement mechanism. The state vector is updated based on Kalman gain and then passed to the deep learning prediction module.

5. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The real-time data stream branch specifically includes the following steps: The filtered real-time data is normalized and sampled using a sliding window to construct time series data segments; Using a gated recurrent unit network to capture PM 2.5 The short-term dynamic characteristics of concentration, the number of gated structures in the gated cyclic unit network is 2; By fusing pressure difference and temperature and humidity data through a multi-input fully connected layer, a real-time feature vector is output.

6. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The historical data stream branch specifically includes the following steps: Receive cumulative working hours, historical pollution exposure data, and filter status labels; The pollution load index is calculated by analyzing PM2.5 levels over a statistical period. 2.5 The result is obtained by summing the products of the concentration values ​​and the VOC concentration values. A temporal convolutional network is used to capture the long-term dependency between cumulative working time and filter lifetime, and output historical feature vectors.

7. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The deep learning prediction module includes an attention-gated fusion unit, which is used for: Establish environmental risk indicators, including PM2.

5. 2.5 Risk indicators and differential pressure risk indicators, when real-time PM 2.5 When the concentration or real-time differential pressure exceeds its respective set threshold, the corresponding risk indicator is triggered; Dynamic weights are calculated based on environmental risk indicators: when any of the aforementioned risk indicators is triggered, the weight value assigned to the real-time data stream branch is increased to be higher than the weight value assigned to the historical data stream branch; when the aforementioned risk indicators are not triggered, a balanced weight value is assigned to the two branches. The real-time feature vector and the historical feature vector are weighted and summed based on the calculated dynamic weights, and the final feature vector is output by fusion.

8. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 1, characterized in that, The deep learning prediction module further includes an output decision unit, which is used for: The final feature vector is then input into the fully connected layer. Output filter health status probability distribution, filter health status score, and remaining lifetime; When the output filter clogging probability exceeds the preset alarm threshold, an emergency decision signal is triggered and sent to the closed-loop control module.

9. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 7, characterized in that, When performing quality control on the collected data, a compensation and correction logic is specifically adopted, which includes the following steps: The monitored values ​​are corrected for physical characteristics based on environmental parameters, specifically for PM2.5 under high humidity conditions. 2.5 The readings are compensated for moisture interference, and temperature compensation is initiated for VOC sensor data in high-temperature environments.

10. The automotive air conditioning filter health prediction system based on multi-source data and deep learning according to claim 7, characterized in that, The generation of control instructions includes executing an active protection strategy, and the specific steps are as follows: When the comparison results show that the filter health status score is lower than the preset threshold, a control command for adjusting the air conditioning mode is automatically generated. The operations performed by the control commands driving the air conditioning system include: switching the air conditioning operating mode to internal circulation, or reducing the external circulation fan speed.

Images