A self-learning and self-adaptive wiper control method and system based on driving behavior

Abstract

The application discloses a self-learning and self-adaptive wiper control method and system based on driving behavior, which comprises the following steps: acquiring environmental characteristic data and calculating a basic wiping instruction; when a manual intervention signal of the wiper is detected, acquiring environmental characteristic data and time-domain fluctuation characteristic data in a preset time period before intervention occurs as an environmental characteristic snapshot; establishing a corresponding relationship between an environmental characteristic interval and a wiping intensity correction amount; and according to real-time environmental characteristic data, matching a corresponding wiping intensity correction amount, dynamically correcting the basic wiping instruction, and outputting a final execution instruction. The manual intervention action of the driver on the wiper handle is regarded as a feedback signal of the control error, the driver identity is distinguished through tactile feature recognition, a continuous wiping intensity correction amount is outputted and is smoothly superimposed with the basic instruction, the driver's comfort threshold is actively learned and fitted, and the frequency of repeated intervention under the same scene is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of vehicle control technology, and in particular to a self-learning adaptive windshield wiper control method and system based on driving behavior. Background Technology

[0002] Currently, most mainstream cars are equipped with automatic windshield wipers based on optical sensors. The basic principle is as follows: the sensor emits infrared light and receives reflected light to calculate the real-time rainfall; the controller compares the rainfall value with a preset threshold using fuzzy logic, and compensates for it by combining the vehicle speed signal. Finally, it outputs the PWM duty cycle through a lookup table method or PID calculation to drive the wiper motor in an open loop.

[0003] The aforementioned existing technologies suffer from a mismatch between the sensor's measurement target and the driver's physiological perception target. Optical sensors detect the rate of change in luminous flux on the glass surface, while the driver's judgment of whether to wipe the windshield is based on the overall clarity of the windshield's view. For example, when a continuous, uniform water film forms on the windshield, an optical "lens effect" occurs, resulting in minimal change in luminous flux. The system may misinterpret this as dryness, even though the driver's vision is actually severely blurred. To address this issue of insufficient sensitivity in AUTO mode adjustment and a significant discrepancy between system output and driver expectations, those skilled in the art typically compensate for the sluggishness of automatic mode adjustment by allowing and relying on frequent manual intervention by the driver using the wiper lever to trigger a single wipe.

[0004] However, the above methods lack a feedback learning mechanism, directly discarding driver's manual intervention data, causing control deviations to recur in specific scenarios, and trapping the user experience in a cycle of repeated corrections. Furthermore, the factory-calibrated fixed control strategy is based solely on group averages and cannot accommodate individual preferences for either a clean field of vision or quietness, resulting in the system consistently failing to achieve personalized adaptive control tailored to individual driver differences. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a self-learning adaptive windshield wiper control method and system based on driving behavior.

[0006] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0007] The present invention adopts the following technical solution: In a first aspect, the present invention provides a self-learning adaptive windshield wiper control method based on driving behavior, comprising the following steps: Based on the acquired environmental feature data, basic scraping instructions are calculated according to a preset basic mapping function. When a manual intervention signal from the windshield wipers is detected, environmental characteristic data and time-domain fluctuation characteristic data within a preset time period before the intervention occur are acquired as an environmental characteristic snapshot. The environmental feature snapshots are associated with the intervention type of the manual intervention signal and stored in the historical behavior database, and a correspondence between the environmental feature intervals and the brushing intensity correction amount is established. In automatic control mode, the corresponding scraping intensity correction amount is matched according to real-time environmental characteristic data, and the basic scraping command is dynamically corrected before the final execution command is output.

[0008] The environmental characteristic data includes: optical rainfall signal, vehicle speed signal, and ambient light intensity signal; the time-domain fluctuation characteristic data includes the average rate of change of optical rainfall signal within a preset time period before the intervention, the number of times optical rainfall signal exceeds a preset threshold, and the extreme value difference of optical rainfall signal; the manual intervention signal refers to the physical operation signal actively triggered by the driver through the human-machine interaction device in automatic control mode to change the current automatic operation state of the windshield wipers, including acceleration intervention signal and deceleration intervention signal.

[0009] The process of establishing the correspondence between environmental feature intervals and wiper intensity correction amounts includes: training historical intervention data in the historical behavior database using a deep learning model to generate a wiper control prediction model; the wiper control prediction model outputs the expected wiper intensity correction amount based on real-time environmental feature data; in automatic control mode, the real-time environmental feature data is input into the wiper control prediction model to obtain personalized correction amounts, and the final execution instructions are generated by combining the basic wiper instructions.

[0010] The self-learning adaptive wiper control method based on driving behavior further includes: real-time monitoring of the driver's manual wiper intervention signal; when the manual intervention signal is detected, the manual intervention signal is executed first, and the environmental feature snapshot is acquired simultaneously.

[0011] Before storing the environmental feature snapshot in the historical behavior database in association with the intervention type of the manual intervention signal, the method further includes: collecting the force, duration, and frequency of the wiper lever's movement during manual intervention to generate a tactile feature data set; comparing the tactile feature data set with a pre-calibrated driver tactile feature database to identify the current operator's driver ID; and storing the environmental feature snapshot in the historical behavior database corresponding to the ID only when a specific driver ID is identified.

[0012] The self-learning adaptive wiper control method based on driving behavior further includes: filtering abnormal states of the environmental feature snapshot; the filtering of abnormal states of the environmental feature snapshot includes: extracting the optical rainfall signal, ambient light intensity signal, and the number of times the optical rainfall signal exceeds a preset threshold from the environmental feature snapshot; if the optical rainfall signal is lower than a preset water film threshold, and the number of times the optical rainfall signal exceeds the preset threshold is greater than a preset flicker threshold, and the ambient light intensity signal is lower than a preset dark light threshold, then it is determined to be a water film lens effect scenario; when an acceleration intervention signal is received, the highest wiper intensity correction amount covering the current rainfall value range is directly generated, and the environmental feature snapshot corresponding to the water film lens effect scenario is marked as a high-weight negative feedback label, and the learning weight coefficient in the historical behavior database is increased.

[0013] The self-learning adaptive wiper control method based on driving behavior further includes: when the optical rainfall signal exceeds a preset rainstorm threshold, or when the average rate of change of the optical rainfall signal exceeds a preset mutation rate threshold, locking the adjustment direction of the wiper intensity correction amount, and only allowing acceleration correction on the basic wiper command.

[0014] The self-learning adaptive wiper control method based on driving behavior further includes: when alternating acceleration intervention signals and deceleration intervention signals are continuously received within a preset time in the same environmental feature range, and the frequency exceeds a preset threshold, it is determined that the system calibration is inaccurate. Then, the update of the wiper intensity correction amount in the current environmental feature range is paused, and a continuous wiper action at a fixed speed is output for a preset time.

[0015] The self-learning adaptive wiper control method based on driving behavior further includes: when the average rate of change of the optical rainfall signal is greater than a preset abrupt change threshold, marking the current manual intervention signal as an emergency request and reducing the update weight of the current manual intervention signal when calculating the wiper intensity correction; when the extreme value difference of the optical rainfall signal is greater than a preset non-uniformity threshold, limiting the rate at which the currently calculated wiper intensity correction is interpolated and transmitted to adjacent rainfall intervals.

[0016] Secondly, the present invention also provides a self-learning adaptive windshield wiper control system based on driving behavior, comprising: The instruction generation module is used to calculate basic scraping instructions based on the acquired environmental feature data and a preset basic mapping function. The intervention monitoring module is used to acquire environmental characteristic data and time domain fluctuation characteristic data within a preset time period before the intervention occurs as an environmental characteristic snapshot when a manual intervention signal from the windshield wipers is detected. The storage establishment module is used to associate the environmental feature snapshot with the intervention type of the manual intervention signal and store it in the historical behavior database, and to establish the correspondence between the environmental feature interval and the brushing intensity correction amount; The correction control module is used to dynamically correct the basic scraping command by matching the corresponding scraping intensity correction amount according to the real-time environmental characteristic data in automatic control mode, and then output the final execution command.

[0017] The beneficial effects of this invention are as follows: 1. Treat the driver's manual intervention on the wiper lever as a feedback signal of control error. Identify the driver's identity through tactile features, bind the feedback signal to a real-time snapshot of environmental features, output continuous wiping intensity correction and smoothly superimpose it with the basic command, actively learn and conform to the driver's comfort threshold, and significantly reduce the frequency of repeated intervention in the same scenario. 2. By combining absolute environmental quantities with temporal fluctuation characteristics, a gridded environmental characteristic range is constructed, enabling the output personalized correction quantity to accurately match local weather changes and vehicle driving conditions, achieving high-precision dynamic brush control, and constructing safety limit and degradation strategies to ensure that personalized control absolutely obeys driving safety. Attached Figure Description

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

[0019] Figure 1 This is a flowchart illustrating a self-learning adaptive windshield wiper control method based on driving behavior according to the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown in some illustrative embodiments, a self-learning adaptive windshield wiper control method based on driving behavior is provided, including the following steps: 101: After the vehicle is started or the power is turned on, the body controller or the dedicated wiper controller completes the system initialization, the wiper system is powered on, and the controller obtains the current physical position status of the wiper combination switch in real time through IO port level reading or CAN / LIN bus communication, that is, it collects the wiper lever position signal.

[0022] 102: Determine if the wiper lever is in AUTO mode. If yes, enter the closed-loop control process and execute step 104; otherwise, proceed to step 103.

[0023] The control logic within the controller performs calibration and comparison on the acquired gear position signals. If it identifies a specific resistance value of the lever in the AUTO position or a CAN message signal, it sets the self-learning enable flag and starts a background timed task to enter the closed-loop control process.

[0024] 103: Exit the closed-loop control process.

[0025] By clearing temporary environmental snapshot variables from memory and suspending self-learning and prediction model computation tasks, the wiper system degenerates into a traditional direct on / off control mode. This saves controller computing resources, and in non-automatic mode, the system does not need to perform complex data acquisition and model inference, reducing system power consumption and load.

[0026] 104: When the wiper system is in automatic control mode, acquire environmental feature data, and calculate basic wiping instructions based on the acquired environmental feature data and a preset basic mapping function.

[0027] The environmental characteristic data includes: optical rainfall signal, vehicle speed signal, and ambient light intensity signal.

[0028] The controller continuously reads the luminous flux conversion value from the optical rain sensor, the vehicle speed signal sent by the vehicle's ESC / instrument panel, and the ambient light intensity signal output by the photosensor via A / D sampling or bus communication. Using this data as input, it substitutes it into a pre-calibrated baseline MAP based on the group average to calculate the appropriate baseline PWM duty cycle or baseline wiping frequency. This maintains the system's basic safe operating capability. The traditionally calibrated baseline commands serve as the benchmark for subsequent personalized adjustments, ensuring that even with insufficient self-learning data, the wiper system can still provide industry-standard automated wiping.

[0029] 105: Real-time monitoring of the driver's manual wiper intervention signal. When a manual wiper intervention signal is detected, the current automatic output is immediately interrupted, and the manual intervention signal is executed first.

[0030] Manual intervention signals refer to physical operation signals actively triggered by the driver through the human-machine interface in automatic control mode to change the current automatic operation state of the windshield wipers. These include acceleration intervention signals and deceleration intervention signals. Acceleration intervention signals are generated when the driver triggers a single wipe command via the wiper lever, manually switches the wiper speed to a higher setting, or increases the sensitivity setting via the sensitivity adjustment knob. Deceleration intervention signals are generated when the driver decreases the sensitivity setting via the sensitivity adjustment knob or manually switches the wiper speed to a lower setting.

[0031] The system employs hardware interrupts or a high-priority polling mechanism to monitor switch transitions in AUTO mode. Upon detecting acceleration interventions (such as a single wipe, shifting to a higher gear, or increasing the sensitivity knob) or deceleration interventions (such as shifting to a lower gear or decreasing the sensitivity knob), the output of the basic mapping function is immediately masked, directly converting the manual command into a motor drive signal. Simultaneously, the intervention signals are categorized and labeled. This ensures the driver maintains absolute control in emergency situations, such as when sluggish automatic wipers impair visibility. Furthermore, the positive and negative polarities of the intervention signals provide clear directional labels for subsequent self-learning, indicating whether the system should accelerate or decelerate.

[0032] 106: When a manual intervention signal for the windshield wipers is detected and executed, environmental characteristic data and time-domain fluctuation characteristic data within a preset time period before the intervention occur are acquired as environmental characteristic snapshots.

[0033] The temporal fluctuation characteristic data includes: the average rate of change of optical rainfall signal within a preset time period before the intervention, the number of times the optical rainfall signal exceeds a preset threshold, and the extreme value difference of the optical rainfall signal.

[0034] The amount of rainfall determines the base range in which the system operates, while the temporal fluctuation characteristics determine the physical scenario in which the system is currently located. Under different physical scenarios, when a driver makes the same manual intervention, the underlying intention may be different. Therefore, the method and weight for calculating personalized correction amounts must also be different.

[0035] If the average rate of change of the optical rain gauge signal is extremely high, such as when a sudden downpour hits the windshield, the driver's manual intervention triggering a single wiper is likely due to the suddenness of the downpour and the automatic wipers' delayed response. This intervention is for emergency safety, not a driver's nitpicking of the wiper's normal sensitivity. If a high rate of change is detected, the learning weight of this intervention data should be reduced when calculating the wiper intensity correction to avoid setting the wiper speed excessively high across the entire environmental characteristic range due to a single, isolated downpour. Conversely, if the rate of change is extremely low, driver intervention in this case reflects a sensitivity preference and should be given a high weight, significantly updating the offset coefficient for that environmental characteristic range.

[0036] When a continuous water film forms on glass, it creates a lensing effect, causing optical sensors to misinterpret it as dry, even though the human eye is already blurred. In this situation, the sensor's signal characteristics often show a low overall rainfall value, such as 30%, but due to the water film's flow and surface tension, light reflection causes constant flickering, resulting in extremely high scintillation frequency. If low absolute rainfall and high fluctuation frequency are detected, it can be determined that the system is currently in a water film lensing effect scenario. In this scenario, once the driver triggers acceleration intervention, instead of simply adding a personalized correction to the current rainfall value, a strong positive personalized correction is directly generated, forcing the system to overtake the baseline curve in advance when encountering the same characteristic again.

[0037] The extreme value difference reflects the unevenness of rainfall. A small extreme value difference indicates that the rain is relatively even; a large extreme value difference indicates showers or uneven dripping, such as water splashed up by the car in front. If the system learns the driver's intervention actions in scenarios with large extreme value differences, the generated personalized correction values ​​must use conservative linear interpolation when applied to adjacent rainfall intervals to prevent sudden speed changes in adjacent intervals from causing dry wiping or abnormal noise from the wipers. However, personalized correction values ​​learned in scenarios with small extreme value differences can be applied in a wide-ranging, stepped manner, with the extreme value difference determining how the learned personalized correction values ​​smoothly transition to surrounding intervals.

[0038] 107: Identifying the effectiveness of interventions. This specifically includes: First, the force, duration, and frequency of the wiper lever's movement during manual intervention are collected to generate a tactile feature data set. Then, the tactile feature data set is compared with a pre-calibrated driver tactile feature database to identify the current operator's driver ID. Only when a specific driver ID is identified is step 108 executed, which associates the environmental feature snapshot with the intervention type of the manual intervention signal and stores it in the historical behavior database corresponding to that ID.

[0039] The force of the action reflects the driver's decisiveness. For example, some people habitually slap the lever hard, while others prefer a gentle push. This parameter can be obtained indirectly through a pressure sensor at the bottom of the lever or by analyzing the rate of change of current / voltage at the moment the switch is turned on and off. Duration refers to the time it takes for the lever to return to its original position after deviating from its original position. Some people prefer a very short press, such as 0.1 seconds, while others prefer a slight pause, such as 0.5 seconds. Frequency of operation refers to the number of consecutive interventions within a very short time window, reflecting the driver's impatience or dissatisfaction with the current situation.

[0040] The system pre-defines a database of tactile features of legitimate drivers, such as vehicle owners and their families, with each ID corresponding to a feature vector center. When a set of tactile feature data for a single operation is collected in real time, a classification algorithm is used to calculate the similarity between the current data set and the ID templates in the database. If the similarity to a particular ID is greater than a preset safety confidence threshold, the current operator is identified as that ID; otherwise, it is determined to be either an unfamiliar driver or a random accidental touch.

[0041] Within the same household, different drivers often have opposite preferences for windshield wiper sensitivity. Indiscriminately storing all drivers' intervention data in a single database would lead to serious logical conflicts during model training, causing oscillating model predictions and a poor user experience. This step identifies drivers by tactile features, mapping the data to a dedicated database to ensure the self-learning model follows the current driver's preferences, achieving personalized isolation. Furthermore, in actual driving, drivers may unintentionally touch the lever, or a friend who doesn't frequently drive the car may drive it for them. These occasional actions do not represent the owner's true preferences. Because unfamiliar actions cannot be matched with existing IDs, they are directly deemed invalid and blocked, not written to the historical behavior database. This effectively prevents noisy data from contaminating the learning samples, making the personalized corrections output by the model during training more accurate and stable.

[0042] 108: Associate environmental feature snapshots with the intervention type of manual intervention signals and store them in the historical behavior database.

[0043] A complete historical behavior data set comprises three parts: environmental background, i.e., a snapshot of environmental characteristics obtained in step 106, such as absolute environmental quantities and temporal fluctuation characteristics within 5 seconds prior to intervention, including optical rainfall values, vehicle speed, illuminance change rate, fluctuation frequency, and extreme value differences; behavior tag, i.e., the intervention type obtained in step 105, such as positive acceleration intervention or negative deceleration intervention; and identity index, i.e., the driver's identity ID identified in step 107. The system associates and binds these three parts according to a specific data structure to generate an event log. The packaged data record is written to the vehicle controller's built-in non-volatile memory, and based on the identity ID identified in step 107, the record is addressed and written to the storage block corresponding to that ID.

[0044] This step provides supervised learning samples for the model by binding and storing environmental feature snapshots with explicit intervention type labels. The model analyzes the accumulated data to identify potential patterns in driver acceleration / deceleration under various combinations of environmental features. Recording each driver correction action allows the system to anticipate and predict future scenarios with similar environmental features. Storing these actions by associating them with the driver's identity ID not only avoids cross-contamination of different driver preferences but also ensures that when the same driver drives the vehicle again, the system can access the corresponding historical database for model inference, guaranteeing the stability and continuity of personalized corrections.

[0045] 109: Establish the correspondence between environmental characteristic ranges and brushing intensity correction amounts.

[0046] Specifically, a deep learning model is used to train historical intervention data in a historical behavior database to generate a wiper control prediction model. The wiper control prediction model is configured to output the expected wiper intensity correction based on real-time environmental feature data.

[0047] An environmental feature interval refers to several gridded data intervals obtained after discretizing the collected continuous environmental feature data according to a preset threshold. Each interval represents a specific physical scenario. The model generates and associates the corresponding brush intensity correction amount for that interval by statistically analyzing which interval the driver's manual intervention event falls into, thereby achieving precise and personalized compensation at the interval level.

[0048] Taking optical rainfall signals as an example, the continuous normalized values ​​output by the sensor, such as 0 to 100, can be divided into continuous numerical ranges such as light rain, moderate rain, heavy rain, and torrential rain; vehicle speed signals can be divided into speed ranges such as stationary driving, urban driving, and highway driving; ambient light intensity signals can be divided into light intensity ranges such as nighttime, cloudy or tunnel, and daytime; and time-domain fluctuation frequency is divided into frequency ranges such as steady-state rain, intermittent rain, and splashing or water film effect. When a set of real-time environmental feature snapshots is collected, its various indicators are classified into a specific block in the above-mentioned interval grid. For example, when the detected optical rainfall value is 15, it falls into the light rain range; when the vehicle speed is 95 km / h, it falls into the highway range; when the illuminance is 20 lux, it falls into the nighttime range; and when the time-domain fluctuation frequency is 7 times / second, it falls into the high-frequency range, this set of data is clearly defined as a specific environmental feature range. This specific interval corresponds to a complex scenario of high-speed driving at night and the formation of a water film lens effect on the glass. At this time, the driver's acceleration intervention behavior in this interval is bound to this interval, and a strong positive brush intensity correction amount is generated and associated for this interval after model training.

[0049] 110: In automatic control mode, the corresponding wiping intensity correction amount is matched according to real-time environmental feature data. The basic wiping command is dynamically corrected and the final execution command is output. That is, the real-time environmental feature data is input into the wiper control prediction model to obtain the personalized correction amount, and the final execution command is generated by combining the basic wiping command.

[0050] The large number of records accumulated in the historical behavior database are transformed into supervised learning samples. Environmental feature snapshots serve as feature input vectors, while the polarity of manual intervention is used as labels. Deep learning algorithms, such as Multilayer Perceptrons (MLPs) or Regressive Neural Networks, are used for offline or incremental training within the controller to fit a nonlinear mapping function between multidimensional environmental features and wiper intensity correction, ultimately solidifying it into a wiper control prediction model. In the AUTO mode of actual vehicle operation, the model is in real-time inference mode. The controller inputs the real-time environmental feature data collected at the current moment into the prediction model. After forward propagation calculation, the model outputs a continuous personalized correction, such as a +0.15 PWM duty cycle compensation value or wiper frequency multiplier. Subsequently, a control fusion algorithm is executed, linearly superimposing or multiplying the personalized correction with the basic wiper command calculated in step 104 to generate the final execution command output to the wiper motor drive circuit.

[0051] Traditional control curves are static data fixed at the factory based on group averages. In contrast, the model trained in this step is a dynamic system that learns from an individual driver's historical intervention data to understand their implicit preferences. This allows it to automatically output corrections in real-time that align with the driver's psychological expectations, completely solving the problem of factory-fixed strategies failing to cover diverse user groups. Because deep learning models can handle multi-dimensional nonlinear features, they can capture anomalous combinations of low optical rainfall values ​​and high temporal fluctuation frequencies—a characteristic of the water film lensing effect. After repeatedly learning from drivers' acceleration interventions in such scenarios in historical data, the model proactively outputs strong positive corrections during real-time inference, forcibly increasing the wiper speed. Furthermore, compared to traditional systems that abruptly switch wiper levels when rainfall exceeds limits, this solution uses a fusion of basic commands and continuous personalized corrections, resulting in a smooth and gradual adjustment of the final command. The correction amount of the model output can be continuously adjusted in small increments, which effectively avoids the mechanical jerking and noise caused by frequent start-stop of the wipers or sudden speed changes, and greatly improves the overall sense of luxury and driving comfort of the vehicle.

[0052] 111: Filter out abnormal states from environmental feature snapshots.

[0053] The specific process includes: extracting the optical rainfall signal, ambient light intensity signal, and the number of times the optical rainfall signal exceeds a preset threshold from the environmental feature snapshot; if the optical rainfall signal is lower than a preset water film threshold, and the number of times the optical rainfall signal exceeds the preset threshold is greater than a preset flicker threshold, and the light intensity signal is lower than a preset dark light threshold, then it is determined to be a water film lens effect scene; when an acceleration intervention signal is received, the highest brushing intensity correction amount covering the current rainfall value range is directly generated, and the environmental feature snapshot corresponding to the water film lens effect scene is marked as a high-weight negative feedback label, and the learning weight coefficient in the historical behavior database is increased.

[0054] When a continuous, uniform water film forms on the windshield surface, infrared light undergoes total internal reflection at the interface between the glass, water film, and air. The intensity of the light reflected back to the receiver is actually close to the baseline value under dry conditions. At this point, the normalized rainfall value output by the sensor will be abnormally low, falling below the system's preset water film threshold. Based on this, the confidence level of the sensor's current rainfall value is determined to be low. Although the continuous water film lowers the average rainfall value, the intensity of reflected light received by the sensor will fluctuate frequently during the dynamic process of water film formation, rupture, and reorganization. By statistically analyzing the number of times the rainfall signal crosses a certain intermediate threshold within a preset time period before intervention, if this number exceeds a preset flicker threshold, it indicates that the glass surface is not in a steady-state dry state but is undergoing drastic liquid film dynamic changes. In low-light environments such as at night or in tunnels, the water film and water droplets on the glass will produce strong refraction and glare interference. Even an extremely thin water film can significantly worsen the driver's visual acuity. However, under strong daylight, the same water film has a relatively weaker impact on vision. Therefore, illuminance is introduced as a scene correction condition. Only when the illuminance is lower than the dark light threshold is it determined that the threat of the water film to visual perception in that scene has reached a level that requires immediate intervention.

[0055] When a driver's acceleration intervention signal is received in the above scenario, instead of gradually approaching the target value through the conventional incremental learning path, the system directly generates the highest wiper intensity correction amount covering the current rainfall range. The water film lens effect is a known physical blind spot for sensors. In this scenario, the sensor output deviates from the actual rainfall. The system directly raises the correction amount for that rainfall range to the highest level, effectively covering the entire range with a strong offset coefficient. This approach bypasses the misleading sensor signal and uses the driver's emergency intervention for forced calibration, ensuring that the system can instantly output sufficient wiper intensity in subsequent identical scenarios, compensating for the sensor's perception deficiencies.

[0056] Water film effects are low-frequency events in regular rainy weather. If trained with conventional weights, the model needs to accumulate a large number of samples to learn the mapping relationship for this scene. By increasing the learning weight coefficients, for example, by multiplying the weight of the loss function for this type of sample, the model can quickly establish a strong response to this scene with a small number of samples. High-weight labels ensure that the correction instructions for water film effects have a high priority in model parameter updates, avoiding averaging by regular samples, thus maintaining the stability of strong offset coverage in this scene. Labeling as negative feedback allows the model to continuously strengthen its sensitivity to this scene in subsequent training. Even if the physical characteristics of the sensor drift due to factors such as glass aging or film replacement, the model can still maintain its ability to recognize water film effects through repeated learning from high-weight samples.

[0057] 112: One-way restriction for driving safety, specifically including: when the optical rain gauge signal exceeds a preset heavy rain threshold, or when the average rate of change of the optical rain gauge signal exceeds a preset abrupt change rate threshold, the adjustment direction of the wiper intensity correction is locked, allowing only acceleration correction on the basic wiper command. That is, when any of the above conditions are met, after the model outputs the personalized correction amount, the sign bit of the personalized correction amount is forcibly locked, with the lower limit set to 0, meaning negative deceleration correction is not allowed. At this time, regardless of how strong the driver's deceleration preference is in the historical behavior database, any negative offset coefficient output by the model will be set to zero, allowing only positive acceleration correction to be superimposed on the basic wiper command.

[0058] When the absolute rainfall reaches the level of a rainstorm, the thickness of the water film and the density of water droplets on the windshield surface are extremely high, and the light transmittance drops sharply. By monitoring the absolute rainfall value, this persistent high-risk physical environment can be identified. The average rate of change of the optical rainfall signal exceeds a preset abrupt change rate threshold. For example, if a vehicle suddenly enters a localized rainstorm area, or if a large splash of water from the tires of a nearby vehicle instantly covers the windshield, although the instantaneous absolute rainfall value may not yet reach the rainstorm threshold, its upward slope is extremely high.

[0059] If the driver has a habit of slowing down, the model will still output a negative correction value in heavy rain or sudden changes in rainfall, causing a decrease in wiper frequency and hindering the recovery of visibility. Step 112 uses amplitude limiting and locking to adjust the direction, thus blocking the negative correction value and ensuring that the control output in extreme environments does not violate driving safety principles.

[0060] Moreover, when environmental features reach the trigger threshold, the system directly intervenes in the output command, eliminating the comparison and calculation process of the model, thereby eliminating the time consumption of algorithm processing and meeting the real-time requirements of field of view recovery in extreme scenarios.

[0061] 113: If alternating acceleration and deceleration intervention signals are received continuously within a preset time in the same environmental characteristic range, and the frequency exceeds a preset threshold, it is determined that the system calibration is inaccurate. Then, the update of the scraping intensity correction amount in the current environmental characteristic range is paused, and a continuous scraping action at a fixed speed is output for a preset time.

[0062] Within the same environmental feature range, the correction value of the model output should tend to a steady state, and the driver's intervention behavior should show a unidirectional convergence trend. If alternating acceleration and deceleration intervention signals are continuously received within a preset time window, and the frequency exceeds a preset threshold, it indicates that there is a persistent bidirectional deviation between the system control output and the driver's expectations. This phenomenon usually stems from two types of anomalies: first, the optical sensor output is affected by high-frequency noise generated by interference in a specific frequency band; second, the mapping function of the current environmental feature range drifts. Based on this, the system determines that there is a calibration misalignment in the current range. If the model continues to learn contradictory data in this state, it will lead to weight divergence and local overfitting. After detecting the calibration misalignment flag, the process of writing to the historical behavior database and updating the correction value of the current environmental feature range are interrupted.

[0063] By pausing the update of correction values, contradictory intervention data can be prevented from entering the training sample set, thus maintaining the validity of the historical behavior database. Outputting continuous actions with a fixed wiping frequency achieves steady-state operation for a preset time and eliminates transient signal interference, eliminating the output uncertainty of the self-learning algorithm under abnormal data and ensuring that the driver obtains a stable visual loop during system anomalies.

[0064] 114: Isolate local abnormal data, specifically including: when the average rate of change of the optical rainfall signal is greater than the preset abrupt change threshold, mark the current manual intervention signal as an emergency request and reduce the update weight of the current manual intervention signal when calculating the brush intensity correction; when the extreme value difference of the optical rainfall signal is greater than the preset non-uniformity threshold, limit the rate at which the currently calculated brush intensity correction is interpolated and transmitted to adjacent rainfall intervals.

[0065] An average rate of change of optical rainfall signals exceeding a preset abrupt change threshold indicates a sudden and dramatic change in the external environment, such as nearby vehicles splashing water that instantly covers the windshield or driving into a localized rainstorm area. In this extremely short time, the acceleration intervention signal triggered by the driver is a transient emergency request, not a steady-state driving habit based on visual comfort evaluation. Using such high-frequency, large-amplitude intervention signals as regular feature samples input into the model will cause the gradient descent algorithm to overfit at local extrema. An extreme value difference in optical rainfall signals exceeding a preset non-uniformity threshold indicates a highly discrete state of moisture distribution within the sensor's field of view, such as intermittent rainfall or localized water flow. In this state, the wiper intensity correction calculated for the current environmental feature interval only represents a local solution for this transient, unsteady scenario, lacking physical representativeness for generalization to adjacent steady-state rainfall intervals.

[0066] After identifying abrupt changes in rainfall during the data preprocessing stage, the current manual intervention signal is marked as an emergency request category. A penalty factor is applied when constructing the loss function or feature matrix, such as multiplying the learning rate or sample weight coefficient by a decay coefficient of 0.1. This operation significantly reduces the contribution of this sample to model parameter updates during the backpropagation stage. For non-uniform scenarios where extreme value differences exceed limits, the system introduces a rate limiter when performing correction space mapping. When the correction amount in the current interval needs to be smoothly interpolated or weighted to adjacent rainfall intervals, the propagation step size is reduced to limit the propagation slope of the correction amount across intervals.

[0067] By reducing weights, the system distinguishes between steady-state habitual preferences and transient risk avoidance demands, preventing the model from deviating from the driver's true comfort threshold under normal rainfall conditions due to learning extreme intervention actions in occasional emergency situations, thus ensuring the steady-state stability of the model's output correction. Extreme value difference constraints prevent the rapid spread of local abnormal corrections in non-uniform rainfall scenarios to adjacent steady-state intervals, avoiding drift in control over a large area of ​​rainfall intervals caused by a single transient disturbance, and maintaining the smoothness and logical continuity of the transition between corrections in adjacent intervals.

[0068] This invention provides a self-learning adaptive windshield wiper control system based on driving behavior, comprising: The instruction generation module is used to calculate basic scraping instructions based on the acquired environmental feature data and a preset basic mapping function.

[0069] The intervention monitoring module is used to acquire environmental characteristic data and time domain fluctuation characteristic data within a preset time period before the intervention occurs as an environmental characteristic snapshot when a manual intervention signal from the windshield wipers is detected.

[0070] The storage establishment module is used to associate environmental feature snapshots with the intervention type of manual intervention signals and store them in the historical behavior database, and to establish the correspondence between environmental feature intervals and brushing intensity correction amounts.

[0071] The correction control module is used in automatic control mode to match the corresponding brushing intensity correction amount according to real-time environmental characteristic data, dynamically correct the basic brushing command, and output the final execution command.

[0072] The priority execution module is used to monitor the driver's manual wiper intervention signals in real time. When a manual intervention signal is detected, the manual intervention signal is executed first, and an environmental feature snapshot is obtained at the same time.

[0073] The classification and verification module is used to collect the force, duration and frequency of the wiper lever action during manual intervention, generate a tactile feature data set, compare the tactile feature data set with a pre-calibrated driver tactile feature database, identify the driver's identity ID of the current operator, and only when a specific driver identity ID is identified, associate the environmental feature snapshot with the intervention type of the manual intervention signal and store it in the historical behavior database corresponding to that ID.

[0074] The first processing module is used to lock the adjustment direction of the brush intensity correction amount when the optical rainfall signal exceeds the preset rainstorm threshold, or when the average rate of change of the optical rainfall signal exceeds the preset abrupt change rate threshold, and only allow acceleration correction on the basic brushing command.

[0075] The second processing module is used to determine that the system calibration is inaccurate when alternating acceleration intervention signals and deceleration intervention signals are continuously received within a preset time in the same environmental feature range, and the frequency exceeds a preset threshold. Then, it pauses the update of the scraping intensity correction amount in the current environmental feature range and outputs a continuous scraping action at a fixed speed for a preset time.

[0076] The third processing module is used to mark the current manual intervention signal as an emergency request and reduce the update weight of the current manual intervention signal when calculating the brush intensity correction when the average rate of change of the optical rainfall signal is greater than the preset abrupt change threshold; when the extreme value difference of the optical rainfall signal is greater than the preset non-uniformity threshold, it limits the rate at which the currently calculated brush intensity correction is interpolated and transmitted to adjacent rainfall intervals.

[0077] The abnormal state filtering module is used to extract the optical rainfall signal, ambient light intensity signal, and the number of times the optical rainfall signal exceeds a preset threshold from the environmental feature snapshot. If the optical rainfall signal is lower than the preset water film threshold, and the number of times the optical rainfall signal exceeds the preset threshold is greater than the preset flicker threshold, and the light intensity signal is lower than the preset dark light threshold, then it is determined to be a water film lens effect scene. When an acceleration intervention signal is received, the highest brushing intensity correction amount covering the current rainfall value range is directly generated. At the same time, the environmental feature snapshot corresponding to the water film lens effect scene is marked as a high-weight negative feedback label, and the learning weight coefficient in the historical behavior database is increased.

[0078] The environmental characteristic data includes: optical rainfall signal, vehicle speed signal, and ambient light intensity signal; the temporal fluctuation characteristic data includes the average rate of change of optical rainfall signal within a preset time period before the intervention, the number of times the optical rainfall signal exceeds a preset threshold, and the extreme value difference of the optical rainfall signal; the manual intervention signal refers to the physical operation signal that the driver actively triggers through the human-machine interaction device to change the current automatic operation state of the windshield wipers in automatic control mode, including acceleration intervention signal and deceleration intervention signal.

[0079] The storage creation module includes: The model building unit is used to train a deep learning model on historical intervention data in the historical behavior database to generate a wiper control prediction model. The wiper control prediction model outputs the expected wiper intensity correction amount based on real-time environmental feature data. The instruction output unit is used to input real-time environmental feature data into the wiper control prediction model in automatic control mode to obtain personalized correction values, and combine them with basic wiping instructions to generate the final execution instruction.

[0080] This invention achieves personalized adaptive control by collecting tactile feature data from the wiper lever to identify the current driver's ID. It then maps environmental snapshots and intervention types to an independent historical behavior database based on the ID. This prevents cross-contamination of feature data during multi-driver co-driving situations, transforming static factory calibration into a dynamic nonlinear mapping based on individual steady-state preferences. It outputs continuous wiper intensity corrections that are smoothly integrated with basic commands, achieving adaptive control and eliminating the repetitive correction loops caused by the lack of memory in traditional systems.

[0081] A criterion for the water film lens effect is established. When a sensor reflection failure scenario is identified, the highest wiper intensity correction amount covering the current rainfall range is directly generated, and this scenario is marked as a high-weight negative feedback label. Bypassing unreliable sensor signals, forced calibration is performed through driver intervention, accelerating the model's convergence in sensor blind zone scenarios and effectively avoiding the dilution of abnormal correction effects by conventional samples, thereby solving the vision deprivation problem under the water film lens effect. In extreme conditions such as continuous heavy rain or sudden water spray, only positive acceleration correction and basic commands are allowed to be superimposed, avoiding conflicts between self-learning preferences and driving safety principles, eliminating algorithm time consumption, ensuring that the wiper motor executes acceleration actions in a timely manner, and meeting the real-time vision recovery requirements in extreme scenarios.

[0082] For high-frequency alternating bidirectional interventions within the same interval, which are identified as calibration inaccuracies, the model update is frozen and a fixed-frequency continuous scraping is output. Simultaneously, for emergency interventions triggered by sudden changes in rainfall, data-level weighting is implemented. Furthermore, for non-uniform scenarios with excessive extreme value differences, the interpolation propagation rate of corrections to adjacent intervals is limited. These three methods work together to isolate transient risk-avoidance behaviors and local non-steady-state noise from contaminating the long-term preference model, preventing local anomalous data from affecting the global calibration domain, thereby ensuring the convergence stability of model parameters.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A self-learning adaptive windshield wiper control method based on driving behavior, characterized in that, Includes the following steps: Based on the acquired environmental feature data, basic scraping instructions are calculated according to a preset basic mapping function. When a manual intervention signal from the windshield wipers is detected, environmental characteristic data and time-domain fluctuation characteristic data within a preset time period before the intervention occur are acquired as an environmental characteristic snapshot. The environmental feature snapshots are associated with the intervention type of the manual intervention signal and stored in the historical behavior database, and a correspondence between the environmental feature intervals and the brushing intensity correction amount is established. In automatic control mode, the corresponding scraping intensity correction amount is matched according to real-time environmental characteristic data, and the basic scraping command is dynamically corrected before the final execution command is output.

2. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 1, characterized in that, The environmental characteristic data includes: optical rainfall signal, vehicle speed signal, and ambient light intensity signal; The time-domain fluctuation characteristic data includes the average rate of change of optical rainfall signal within a preset time period before the intervention, the number of times the optical rainfall signal exceeds a preset threshold, and the extreme value difference of the optical rainfall signal. The manual intervention signal refers to the physical operation signal that the driver actively triggers through the human-machine interaction device to change the current automatic operation state of the windshield wipers in automatic control mode, including acceleration intervention signal and deceleration intervention signal.

3. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 2, characterized in that, The process of establishing the correspondence between environmental characteristic intervals and brushing intensity correction amounts includes: A deep learning model is used to train historical intervention data in the historical behavior database to generate a wiper control prediction model. The wiper control prediction model outputs the expected wiper intensity correction amount based on real-time environmental feature data. In automatic control mode, real-time environmental feature data is input into the wiper control prediction model to obtain personalized correction values, which are then combined with the basic wiping instructions to generate the final execution instructions.

4. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 3, characterized in that, Also includes: The system monitors the driver's manual wiper intervention signal in real time. When the manual intervention signal is detected, the manual intervention signal is executed first, and a snapshot of the environmental features is acquired at the same time.

5. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 4, characterized in that, Before storing the environmental feature snapshot associated with the intervention type of the manual intervention signal in the historical behavior database, the method further includes: Collect data on the force, duration, and frequency of manual intervention of the wiper lever to generate a set of tactile feature data. The tactile feature data set is compared with a pre-calibrated driver tactile feature database to identify the current operator's driver identity ID. Only when a specific driver identity ID is identified, the environmental feature snapshot is associated with the intervention type of the manual intervention signal and stored in the historical behavior database corresponding to that ID.

6. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 5, characterized in that, Also includes: The environmental feature snapshot is subjected to abnormal state filtering; the abnormal state filtering of the environmental feature snapshot includes: Extract the optical rainfall signal, ambient light intensity signal, and the number of times the optical rainfall signal exceeds a preset threshold from the environmental feature snapshot; If the optical rainfall signal is lower than a preset water film threshold, and the number of times the optical rainfall signal exceeds the preset threshold is greater than a preset flicker threshold, and the illuminance signal is lower than a preset dark light threshold, then it is determined to be a water film lens effect scene. When an accelerated intervention signal is received, the highest brush intensity correction amount covering the current rainfall range is directly generated. At the same time, the environmental feature snapshot corresponding to the water film lens effect scenario is marked as a high-weight negative feedback label, and the learning weight coefficient in the historical behavior database is increased.

7. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 6, characterized in that, Also includes: When the optical rainfall signal exceeds a preset rainstorm threshold, or when the average rate of change of the optical rainfall signal exceeds a preset abrupt change rate threshold, the adjustment direction of the brush intensity correction is locked, and acceleration correction is only allowed on the basic brush command.

8. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 7, characterized in that, Also includes: If alternating acceleration and deceleration intervention signals are received continuously within a preset time within the same environmental characteristic range, and the frequency exceeds a preset threshold, it is determined that the system calibration is inaccurate. In this case, the update of the scraping intensity correction amount for the current environmental characteristic range is paused, and a continuous scraping action at a fixed speed is output for a preset time.

9. The self-learning adaptive windshield wiper control method based on driving behavior according to claim 8, characterized in that, Also includes: When the average rate of change of the optical rainfall signal is greater than a preset abrupt change threshold, the current manual intervention signal is marked as an emergency request, and the update weight of the current manual intervention signal in calculating the brush intensity correction is reduced; when the extreme value difference of the optical rainfall signal is greater than a preset non-uniformity threshold, the rate at which the currently calculated brush intensity correction is interpolated and transmitted to adjacent rainfall intervals is restricted.

10. A self-learning adaptive windshield wiper control system based on driving behavior, characterized in that, include: The instruction generation module is used to calculate basic scraping instructions based on the acquired environmental feature data and a preset basic mapping function. The intervention monitoring module is used to acquire environmental characteristic data and time domain fluctuation characteristic data within a preset time period before the intervention occurs as an environmental characteristic snapshot when a manual intervention signal of the windshield wipers is detected. The storage establishment module is used to associate the environmental feature snapshot with the intervention type of the manual intervention signal and store it in the historical behavior database, and to establish the correspondence between the environmental feature interval and the brushing intensity correction amount; The correction control module is used to dynamically correct the basic scraping command by matching the corresponding scraping intensity correction amount according to the real-time environmental characteristic data in automatic control mode, and then output the final execution command.

Images