Automobile air conditioner controller intelligent regulation and control method based on self-adaptive neural network

Abstract

The invention relates to the technical field of intelligent control of automobile air conditioners, and discloses an intelligent regulation and control method of an automobile air conditioner controller based on a self-adaptive neural network, which realizes comprehensive perception of environment information inside and outside an automobile through distributed multi-source data intelligent fusion processing. Establishing a data quality evaluation model and an intelligent data interpolation algorithm to ensure data reliability; constructing a lightweight dual-path adaptive neural network model by adopting a knowledge distillation technology; establishing a hierarchical multi-objective optimization control architecture which comprises a three-layer optimization structure of a strategic layer, a tactical layer and an operation layer, and realizing organic combination of long-term planning, medium-term prediction and real-time control; the method further comprises key technologies such as multi-level data missing fault-tolerant processing, actuator self-adaptive compensation control, system-level fault self-healing and degradation control and the like, and stable and reliable intelligent regulation and control can be achieved under the extreme environment condition. According to the invention, the control precision, the energy consumption optimization and the user experience in extreme environments such as polar regions, plateaus and deserts are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of automobile air conditioners, and more particularly to an intelligent control method for an automobile air conditioner controller based on an adaptive neural network. BACKGROUND

[0002] With the development of the automobile industry towards intelligence and electrification, the automobile air conditioner control system, as an important part of the driving and riding experience and energy consumption management, its intelligent degree of control strategy has increasingly become a key indicator to measure the advanced nature of automobile technology. Traditional automobile air conditioner control technology mainly relies on PID control, fuzzy control and preset logic control methods. Although these methods can provide basic temperature regulation function in normal environment, they have obvious technical limitations when facing extreme environmental conditions. Especially in special application scenarios such as polar expedition, plateau transportation, desert operation, the complexity and extremity of environmental conditions pose unprecedented technical challenges to the air conditioner control system.

[0003] The existing PID control method has problems of fixed parameters, poor adaptability to nonlinear systems, insufficient processing capacity for multivariable coupling, etc., and it is difficult to cope with the huge load changes in extreme environments. Although the fuzzy control method has certain intelligent characteristics, the rule base is static, the dependence on expert experience is strong, and the adaptability to complex environment is poor, which makes its control effect not good under extreme conditions. In recent years, with the development of artificial intelligence technology, neural network control method has begun to be applied in the field of automobile air conditioner control, but the existing neural network control technology is mainly designed for conventional passenger car scenarios, and lacks special consideration for extreme environments.

[0004] The existing technology also has environmental adaptability problems, mainly manifested as insufficient processing capacity for special environmental conditions such as extreme temperature difference, low air pressure and high humidity, system response lag when environmental conditions change rapidly, and difficulty in simultaneously processing complex coupling relationship of multiple environmental factors. In addition, the lack of individualized control capability is also an important defect of the existing technology. The existing system lacks effective user behavior pattern learning mechanism, and it is difficult to meet the different needs of multiple passengers in the vehicle, and the influence of individual differences on thermal comfort is not fully considered. Therefore, it is urgent to develop a new type of automobile air conditioner control technology that can adapt to extreme environmental conditions, has intelligent learning ability and realizes individualized control. SUMMARY

[0005] The present application provides an intelligent control method for an automobile air conditioner controller based on an adaptive neural network, which solves the complex technical problems of environmental adaptability in related technologies.

[0006] The present application provides an intelligent control method for an automobile air conditioner controller based on an adaptive neural network, which includes the following steps: Multi-source data from inside and outside the vehicle is collected by a sensor array, and data fusion is performed using a timestamp synchronization algorithm. A data quality assessment model is established and multi-source data repair is performed based on an LSTM network to output a fused data vector. A dual-path adaptive neural network is constructed based on fused data vectors. Knowledge distillation is used to achieve lightweighting. The weight coefficients are dynamically adjusted through an environmental change detector to achieve model quantization compression. Based on the fusion of data vectors and user-defined comfort targets, a hierarchical multi-objective optimization control is implemented, establishing a three-layer architecture of strategy, tactics, and operation, and designing an adaptive adjustment mechanism for target weights to output the optimal air conditioning control command. A three-level sensor fault diagnosis and fault tolerance mechanism is established, and a virtual sensor algorithm is used to estimate the output value of faulty sensors. An actuator adaptive compensation control is adopted, which estimates the actuator characteristic parameters by online identification model and designs a feedforward compensation algorithm to eliminate actuator characteristic changes and achieve multi-actuator coordination. To achieve fault self-healing and degradation control, an expert system is established to detect faults, and the system operation is divided into four modes: normal, energy-saving, safe, and emergency, based on the degree of fault impact. Implement intelligent power management, establish an energy consumption model, optimize power allocation based on Markov decision processes, and realize dynamic power limiting and hibernation mechanisms.

[0007] In a preferred embodiment, the sensor array includes a temperature sensor, a humidity sensor, a barometric pressure sensor, an infrared thermal imaging sensor, a wind speed sensor, and a personnel detection sensor. Based on the sensing frequency, it is divided into a high-frequency group, a mid-frequency group, and a low-frequency group. The high-frequency group includes the infrared thermal imaging sensor and the wind speed sensor, with a sampling frequency of 100Hz; the mid-frequency group includes the temperature sensor and the humidity sensor, with a sampling frequency of 10Hz; and the low-frequency group includes the barometric pressure sensor and the personnel detection sensor, with a sampling frequency of 1Hz. The data quality assessment model uses a weighted combination of information entropy and signal-to-noise ratio for calculation. Data with a quality index below 0.7 is marked as low-quality data. Low-quality data is repaired using a time-series interpolation method with an LSTM network.

[0008] In a preferred embodiment, the knowledge distillation technique in the dual-path adaptive neural network includes a teacher network training phase and a student network distillation phase. The teacher network comprises a three-layer LSTM structure with 128 neurons in the first layer, 256 neurons in the second layer, and 128 neurons in the third layer. The student network comprises 64 LSTM units and 32 fully connected layer neurons. The distillation loss function is a weighted combination of the cross-entropy loss between the student network output and the true label and the KL divergence between the student network output and the teacher network output.

[0009] In a preferred embodiment, the dynamic adjustment of the weight coefficient by the environmental change detector comprises: dynamically adjusting the weight coefficient of the main path LSTM network and the auxiliary path CNN network according to the environmental change degree index, the main path weight basic value is 0.8, and the weight is adjusted according to the environmental change degree by a sigmoid function; the auxiliary path weight is complementary to the main path weight, and the sum of the weights of the two is always 1; and the environmental change degree index is obtained by calculating the deviation degree of the current environmental parameter and the historical mean value.

[0010] In a preferred embodiment, the hierarchical multi-objective optimization control comprises three levels of strategic layer, tactical layer and operation layer, the strategic layer solves task priority allocation and energy consumption budget planning, the time scale is hour level, the tactical layer makes medium-term prediction and control decision, the time scale is minute level, and the operation layer performs real-time control execution, the time scale is second level, the coordination and optimization between each level are realized through target decomposition and feedback mechanism, and the Pareto optimal solution set is used for multi-objective trade-off.

[0011] In a preferred embodiment, in the three-level sensor fault diagnosis and fault-tolerant mechanism, the key-level sensors include an in-vehicle temperature sensor and a personnel detection sensor, the important-level sensors include a humidity sensor and a wind speed sensor, and the general-level sensors include an air pressure sensor and an illumination sensor; the virtual sensor algorithm estimates the output value of the faulty sensor by combining a physical model and a data-driven model, the physical model establishes a temperature transfer equation based on the law of thermodynamics, and the data-driven model adopts a support vector regression algorithm.

[0012] In a preferred embodiment, in the adaptive compensation control of the actuator, the online identification model adopts a recursive least squares method to estimate the dynamic characteristic parameters of the actuator in real time, and a forgetting factor is set to 0.95; the feedforward compensation algorithm designs a compensator according to the identified actuator characteristic parameters to offset the influence of actuator characteristic changes on control performance, and the multi-actuator coordination adopts a master-slave control structure to take the compressor as a master actuator responsible for temperature regulation, and takes the fan and the damper as slave actuators responsible for auxiliary regulation and coordination.

[0013] In a preferred embodiment, in the fault self-healing and degradation control, a fault diagnosis expert system integrates rule-based reasoning and case-based reasoning, the rule-based reasoning includes more than 200 diagnosis rules in a fault diagnosis rule base established based on expert knowledge, the case-based reasoning finds the most similar case in a historical fault case library through similarity matching, and the fault influence degree is quantified from three dimensions of function influence, safety influence and task influence.

[0014] In a preferred embodiment, the energy consumption model comprehensively considers the energy consumption characteristics of the power system, air conditioning system, lighting system, communication system and each subsystem; the power distribution strategy is modeled as a Markov decision process, and the power distribution decision is optimized by setting the state space, action space and reward function.

[0015] In a preferred embodiment, a computer readable storage medium is used to store computer readable instructions, which can run the above-mentioned adaptive neural network-based intelligent control method of an automobile air conditioner controller when read by a computer.

[0016] The beneficial effects of the present application are: By establishing a distributed multi-source data intelligent fusion processing mechanism and a multi-level data missing fault tolerance processing technology, the problem that the traditional control method is prone to failure in extreme environments is effectively solved. The design of the lightweight dual-path adaptive neural network enables the system to simultaneously consider long-term environmental change trend prediction and rapid response to sudden environmental changes, and realizes intelligent control strategy switching through an adaptive weight fusion mechanism, thereby improving the adaptability and control stability of the system in complex and variable environments. Through hierarchical multi-objective optimization control and intelligent power management technology based on energy consumption perception, the system energy consumption is reduced under the premise of ensuring control performance; the combination of long-term energy consumption budget planning at the strategic level, medium-term environmental change prediction at the tactical level and real-time optimization control at the operational level realizes full-time energy consumption optimization management. The intelligent power management system dynamically allocates power to each subsystem according to the task priority and remaining energy state, maximizes energy utilization efficiency under the premise of ensuring key functions. The multi-level degradation control strategy ensures that basic life support functions can still be maintained when energy is insufficient, providing reliable energy consumption support for long-term operation in extreme environments. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of an adaptive neural network-based intelligent control method of an automobile air conditioner controller of the present application. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described in some examples can be combined in other examples.

[0019] The application discloses an adaptive neural network-based intelligent regulation method for a vehicle air conditioner controller, as shown in the formula (I): Figure 1 The method comprises the following steps: S100, collecting multi-source data inside and outside the vehicle through a sensor array, performing data fusion by using a timestamp synchronization algorithm, establishing a data quality evaluation model, and repairing the multi-source data based on an LSTM network to output a fusion data vector; The input of this step is multi-source data collected by various sensors inside and outside the vehicle, including temperature sensor data , humidity sensor data , air pressure sensor data , infrared thermal imaging data , personnel detection sensor data and other multi-source heterogeneous data.

[0020] All sensor data are standardized and pretreated; the maximum and minimum value standardization method is used for temperature data, the original temperature data are mapped to the range of 0 to 1, wherein the minimum value corresponds to the lowest temperature of an extreme environment, -50℃, and the maximum value corresponds to the highest temperature, 60℃; the humidity data are converted into 0 to 1 values by dividing by 100, that is, the percentage humidity is converted into 0 to 1 values; the air pressure data are standardized based on the standard air pressure at sea level, wherein the standard air pressure at sea level is 101.325 kPa, and the maximum and minimum air pressures correspond to the extreme values at different altitudes; the one-hot encoding method is used for classification data, the protective equipment types include three types of light protective clothing, medium protective clothing and heavy protective clothing, and are encoded as [1, 0, 0], [0, 1, 0] and [0, 0, 1] respectively; the statistical-based anomaly detection method is used, the 3σ criterion is used to identify abnormal data, when the deviation of new data from the historical mean value exceeds 3 times the standard deviation, the abnormal data are repaired by using the median replacement method.

[0021] A distributed data preprocessing architecture is established; according to the sensor sampling frequency, the sensors are divided into three categories: the high-frequency group includes an infrared thermal imaging sensor (30 Hz) and a wind speed sensor (10 Hz), the medium-frequency group includes a temperature sensor (1 Hz) and a humidity sensor (1 Hz), and the low-frequency group includes an air pressure sensor (0.1 Hz) and a personnel detection sensor (0.2 Hz). The high-frequency sensor data are pretreated in the in-vehicle edge computing node, including digital filtering, feature extraction and preliminary fusion. The medium-frequency and low-frequency sensor data are transmitted to the central processor for comprehensive fusion processing.

[0022] An adaptive timestamp synchronization algorithm based on the Network Time Protocol is adopted. The algorithm establishes a correspondence between the local time and the network standard time, calculates the time offset of each sensor, and corrects the timestamps of all sensor data. The time synchronization accuracy reaches the microsecond level, ensuring the time consistency of multi-source data. Specifically, by comparing the sensor local timestamp with the standard network time, the time offset of each sensor is calculated, and then the timestamp of the sensor data is corrected accordingly.

[0023] A data quality evaluation model is established. The model uses a weighted combination of information entropy and signal-to-noise ratio to evaluate data quality. The specific calculation method is: multiply the data information entropy by the weight coefficient 0.6, multiply the data signal-to-noise ratio by the weight coefficient 0.4, and then add them to get the data quality evaluation index. Among them, information entropy reflects the uncertainty of data, signal-to-noise ratio reflects the clarity of data, and weight coefficients 0.6 and 0.4 represent the importance of information entropy and signal-to-noise ratio in quality evaluation, respectively.

[0024] Information entropy is calculated by discretizing continuous data into 256 intervals and then calculating the probability distribution of each interval. The signal-to-noise ratio is obtained by Fourier transform analysis of the signal spectrum, which is the ratio of signal power to noise power. The weight coefficient is determined by statistical analysis of a large amount of experimental data. When the data quality index is lower than the threshold value 0.7, it is marked as low-quality data and needs to be repaired or replaced.

[0025] An intelligent data interpolation algorithm is implemented; when data loss or poor quality is detected, a time series interpolation method based on the Long Short-Term Memory (LSTM) network is used for data repair. This method uses the time series characteristics of historical data to establish a prediction model of data changes, and predicts the missing data points by learning the time series rules of the data. Specifically, for the missing temperature data , the historical data to of the previous 30 minutes are used as input, and is obtained as the estimated value of the missing data through the LSTM network.

[0026] A sensor fault self-diagnosis mechanism is established; data consistency test and physical constraint check are used to detect sensor faults. Data consistency test compares the measurement values of the same type of sensors to find abnormalities. When the measurement value of a sensor differs from other sensors of the same type by more than a pre-set threshold, it is determined that the sensor may have a fault. Physical constraint check verifies whether the sensor data conforms to the physical law to judge the rationality of the data, such as the physical constraint that the temperature change rate should not exceed 5℃ per minute.

[0027] The output of this step is the fused data vector The fusion data vector contains reliable environmental information after time synchronization, quality evaluation, data repair, etc., and provides high-quality input data for subsequent neural network processing.

[0028] Optionally, in some embodiments, a data integrity verification mechanism based on blockchain technology is adopted to record the hash value of sensor data in a distributed ledger, ensuring the integrity and tamper resistance of data during transmission and processing. This mechanism can effectively prevent malicious attacks and data tampering, improving the security of the system; similarly, in some embodiments, federated learning technology is used to realize data sharing and collaborative learning among multiple vehicles. Each vehicle trains a neural network model locally and only shares model parameters without sharing raw data, which protects data privacy and improves model performance by utilizing the data of multiple vehicles S200, based on the fusion data vector, a double-path adaptive neural network is constructed, a knowledge distillation technique is used to realize lightweight, and a weight coefficient is dynamically adjusted through an environmental change detector to realize model quantization compression; The input of this step is the fusion data vector and historical environmental data (i.e. historical multi-source data), the goal is to build a lightweight neural network model that can operate stably in extreme environments and perform hierarchical multi-objective optimization control.

[0029] A knowledge distillation technique is used to construct a lightweight network architecture; this technique includes two stages: teacher network training stage and student network distillation stage; in the teacher network training stage, historical environmental data and corresponding optimal control strategies are collected as training samples, the data set contains 10,000 sample points, covering various environmental conditions; a teacher network containing three layers of LSTM structure is constructed, the first layer has 128 neurons, the second layer has 256 neurons, and the third layer has 128 neurons, using activation function; the training process uses Adam optimizer, the learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 200 rounds, and the loss function uses mean square error, and the prediction accuracy is obtained on the validation set after training; in the student network distillation stage, a lightweight network containing 64 LSTM units and 32 fully connected layer neurons is constructed, and the calculation method of the knowledge distillation loss function is: the cross-entropy loss of the student network prediction result and the true label is multiplied by the weight coefficient , plus the KL divergence loss of the student network prediction result and the teacher network prediction result multiplied by the weight coefficient , to get the total distillation loss; the KL divergence is calculated by the relative entropy method between probability distributions, which measures the difference between the teacher network and the student network by comparing their output probability distributions.

[0030] The dual-path network structure is designed; the main path adopts the long short-term memory (LSTM) network structure, which is specially used to process the change rule of the time sequence environment data; the path includes an input layer, an LSTM hidden layer, a full connection layer and an output layer. The input layer receives a 6-dimensional environment feature vector, the LSTM hidden layer contains 64 neurons, and the sliding window mechanism is used to process the time sequence data with a length of 30 minutes. The auxiliary path adopts a one-dimensional convolutional neural network structure, which is specially used to process sudden environmental changes and abnormal situations. The path includes a convolutional layer, an attention mechanism layer, a pooling layer and a full connection layer. The convolutional layer contains 32 convolutional kernels with a kernel size of 3 and a step size of 1. The attention mechanism layer adopts a self-attention mechanism to calculate the importance weight of the input features.

[0031] An adaptive weight fusion mechanism is established; the weight coefficients of the main path and the auxiliary path are dynamically adjusted through the environment change detector; the environment change detector judges the degree of environmental change based on data variance analysis; when the environmental change is small, the weight of the main path is set to 0.8, and the weight of the auxiliary path is set to 0.2; when a sudden environmental change is detected, the weight of the main path is reduced to 0.4, and the weight of the auxiliary path is increased to 0.6, to improve the response ability to sudden changes. The first adaptive weight is calculated as follows: 0.4 multiplied by the value of the sigmoid function, where the input of the sigmoid function is the difference between the variance and the threshold. The second adaptive weight β is calculated as follows: 1 minus the value of the first adaptive weight . This design ensures that the sum of the two weights is always 1, achieving dynamic balance.

[0032] A model quantization compression technology is implemented; the 32-bit floating-point number parameters in the neural network are quantized to 8-bit integers, reducing the storage space and computational complexity of the model; the quantization process adopts a linear quantization method, and the calculation method of the weight quantization is as follows: subtract the minimum value from the floating-point weight, then normalize it by dividing the difference between the maximum value and the minimum value, multiply it by 255 and round it to get an 8-bit integer weight. This quantization method compresses the 32-bit floating-point number to an 8-bit integer, greatly reducing the storage space and computational complexity.

[0033] An incremental learning mechanism is established; an elastic weight consolidation algorithm is used to learn new environmental patterns while avoiding forgetting learned knowledge. The core idea of the algorithm is to assign larger penalty weights to important parameters and smaller penalty weights to unimportant parameters. The importance weight is calculated through the Fisher information matrix, and the loss function of incremental learning is as follows: wherein, is the total loss function, is the loss function of the new task, and the new task loss (the task loss in the new environment, calculated by the mean square error). is the regularization coefficient (set to 100), is the Fisher information matrix value of the i-th parameter, is the current parameter value, is the parameter value in the original task. The calculation method of the Fisher information matrix is: where, is the Fisher information matrix value of the i-th parameter, is the mathematical expectation operation, is the conditional probability of the output given the input and the parameter , is the model parameter vector, is the current parameter value, is the partial derivative of the log-likelihood function with respect to the parameter , is the model input data, is the model output data. The Fisher information matrix is estimated by calculating the second moment of the parameter gradient on the original task data. The parameter importance evaluation process is: calculate the gradient of each parameter on the original task data; square the gradient and take the expectation to get the Fisher information value; the larger the Fisher information value, the more important the parameter, and the greater the penalty weight; this mechanism can effectively learn new environmental patterns while maintaining the original knowledge, achieving continuous accumulation of knowledge.

[0034] The output of this step is a lightweight dual-path neural network model with 64 LSTM units and 32 convolutional kernels, with a model size of only 2.3 MB and an inference time of less than 10 milliseconds, capable of real-time operation on a vehicle embedded system.

[0035] Optionally, in some embodiments, a neural architecture search technique is used to automatically design the optimal network structure. This technique uses a reinforcement learning algorithm to search for the optimal network architecture combination in a predefined search space, and can automatically generate customized network structures according to specific hardware platforms and performance requirements; similarly, in some embodiments, a multi-task learning strategy is used to simultaneously predict multiple target variables such as temperature, humidity, and comfort. By sharing the network's underlying feature extraction part and designing a dedicated output branch for each task, the model's generalization ability can be improved and the number of parameters can be reduced.

[0036]

[0037] ​​S300, performing hierarchical multi-objective optimization control based on the fusion data vector and the user-set comfort target, establishing a three-layer architecture of strategy, tactics and operation, and designing a target weight adaptive adjustment mechanism to output optimal air conditioning control instructions; The inputs of this part are the fusion data vector and the user-set comfort target, energy consumption limit and other control parameters, and the goal is to generate optimal air conditioning control instructions.

[0038] A three-layer optimization architecture is established. The strategic layer is responsible for long-term task planning and global target optimization, with an optimization time span of 4-6 hours, mainly solving problems such as task priority allocation and energy consumption budget planning. The tactical layer is responsible for medium-term control strategy optimization, with an optimization time span of 30 minutes to 2 hours, mainly solving problems such as environmental change adaptation and system coordinated control. The operation layer is responsible for real-time tracking control, with a control period of 1 second, mainly realizing accurate tracking of set values. The three layers communicate through hierarchical interfaces: the strategic layer transmits optimization targets and constraint conditions to the tactical layer, and the tactical layer transmits control set values and control parameters to the operation layer.

[0039] A multi-objective evolutionary algorithm is used in the strategic layer to solve long-term optimization problems. The optimization targets of this layer include minimizing total energy consumption, maximizing average comfort, and maximizing equipment life. Optimization variables include temperature set values, air speed set values, and system operation modes in each time period. The constraint conditions of the multi-objective optimization problem include: the temperature constraint is that the indoor set temperature must be between 16°C and 30°C, the air speed constraint is that the fan speed must be between 10% and 100% of the rated speed, the compressor constraint is that the compressor power must be between 0 and the maximum power, the energy consumption constraint is that the total power consumption must not exceed 90% of the available power, the comfort constraint is that the predicted comfort index must not be less than 7 points, the response time constraint is that the control response time must not exceed 30 seconds, and the equipment protection constraint is that all equipment operation parameters must be within the safety range.

[0040] Multi-objective optimization is achieved by simultaneously minimizing total energy consumption, maximizing average comfort, and maximizing equipment life. The constraint conditions ensure that the set temperature is between 16°C and 30°C, the fan speed is between 10% and 100% of the rated speed, and the total power consumption does not exceed 90% of the available power. The NSGA-II algorithm is used for solution, with a population size of 50, an evolution number of 100, a crossover probability of 0.9, and a mutation probability of 0.1. The algorithm outputs a set of Pareto optimal solutions, and the decision maker can select the appropriate solution according to the actual situation.

[0041] Model predictive control algorithm is adopted in the tactical layer to handle the medium-term optimization. The dynamic model of the vehicle interior environment is established, which includes thermodynamic equations, air conditioning system dynamics equations and human thermal load model. The temperature change of the environment is described by the relationship between refrigeration capacity, human and equipment heat load, vehicle body heat loss, and air density, specific heat capacity and effective volume of the vehicle interior. The prediction horizon is set to 20 minutes, the control horizon is set to 5 minutes, and the sampling period is set to 1 minute. The model predictive control realizes the optimal control by minimizing the weighted sum of the square of the difference between the reference temperature and the actual temperature, the square of the control increment and the square of the control input.

[0042] An improved PID controller is adopted in the operational layer for real-time tracking control. The controller adjusts the parameters adaptively according to the nonlinear characteristics of the air conditioning system. The output of the PID controller is calculated by the linear combination of the proportional term, the integral term and the derivative term, where the proportional term is proportional to the current error, the integral term is proportional to the integral of the error, and the derivative term is proportional to the rate of change of the error. The parameter adaptive adjustment uses fuzzy logic method to dynamically adjust the PID parameters according to the error size and the error change rate. When the error is large, the proportional coefficient is increased to improve the response speed; when the error is small, the integral coefficient is increased to eliminate the steady-state error; when the system oscillates, the derivative coefficient is increased to suppress the overshoot.

[0043] A target weight adaptive adjustment mechanism is designed. The mechanism dynamically adjusts the weights of each target in multi-objective optimization according to the current task priority, environmental conditions and user preferences. The task priority is determined by the expert knowledge base, including life safety, equipment protection, work efficiency and comfort experience, with weights of 0.4, 0.3, 0.2 and 0.1 respectively. The environmental conditions are evaluated by the environmental severity index, and the higher the severity, the greater the weights of safety and protection. User preferences are learned from historical operation data, reflecting user preferences for comfort and energy consumption. The comfort weight is adjusted by multiplying the basic comfort weight by the user preference factor and the environmental severity factor, and the energy consumption weight is adjusted by multiplying the basic energy consumption weight by the corresponding factor.

[0044] The output of this step is the optimal control instruction vector , which includes: compressor control instructions, including compressor start-stop state, compressor speed and compressor working mode; fan control instructions, including fan start-stop state, fan speed and wind direction setting; damper control instructions, including internal-external circulation damper position, mixed damper position and air outlet selection; The following performance monitoring indicators are also output: control accuracy indicators, reflecting the accuracy of temperature control, including temperature deviation, adjustment time and overshoot; energy consumption indicators, including compressor power, fan power and total power consumption; comfort indicators, including PMV value (-3 to +3), temperature uniformity and airflow perception; system response indicators, including control delay, adjustment time and settling time.

[0045] S400, establish a three-level sensor fault diagnosis and fault-tolerant mechanism, and estimate the output value of the faulty sensor by using a virtual sensor algorithm; The input S100 has missing or faulty sensor data, and the S200 constructed double-path adaptive neural network, the goal is to establish a perfect fault-tolerant mechanism to ensure that the system can still run normally in the case of sensor failure.

[0046] Establish a sensor importance classification system. According to the influence of the sensor on the control performance, the sensor is divided into three levels. Key sensors include in-vehicle temperature sensors and personnel detection sensors, the failure of these sensors will directly affect the basic functions of the control system. Important sensors include humidity sensors, barometric pressure sensors, and wind speed sensors, the failure of these sensors will affect the control accuracy but will not cause system failure. Auxiliary sensors include light sensors and noise sensors, these sensors are mainly used to optimize the control strategy. Different levels of sensor failure handling strategies are different: key sensor failure starts emergency handling mode, important sensor failure starts degraded control mode, auxiliary sensor failure only records fault information.

[0047] Design a virtual sensor algorithm based on physical constraints. This algorithm uses known physical relationships to estimate missing sensor data. For example, when the humidity sensor fails, the relative humidity is estimated by psychrometric equation using temperature and barometric pressure data. The estimation method of relative humidity is: divide the actual water vapor pressure by the saturated water vapor pressure, and then multiply by 100% to get the relative humidity value in percentage form. Among them, the actual water vapor pressure reflects the actual amount of water vapor in the air, and the saturated water vapor pressure reflects the maximum amount of water vapor that the air can hold at the current temperature.

[0048] The saturated water vapor pressure is calculated by the Magnus formula, and the actual water vapor pressure is calculated by combining the ideal gas law with temperature and barometric pressure data. This algorithm also includes a data rationality verification mechanism, which verifies the rationality of the estimated data through physical constraint conditions, such as the relative humidity must be between 0% and 100%, and the temperature change rate should not exceed 5℃ per minute, etc.

[0049] Implement a multi-sensor information complementary strategy. This strategy is based on the correlation model between sensors to estimate the data of the faulty sensor. The establishment process of the sensor correlation matrix is: collect at least 1000 hours of historical operation data, calculate the Pearson correlation coefficient between each sensor data, and form a correlation matrix. When the sensor fails, the data of the sensor with the highest correlation is used for weighted estimation: The estimation method of sensor data is: multiplying the measurement values of each sensor participating in estimation by the corresponding weight coefficient respectively, and then adding all the product results to obtain the estimated value. The weight coefficient reflects the importance of each sensor in the estimation process. The calculation method of the weight coefficient is: taking the correlation coefficient between the target sensor and a certain reference sensor as the numerator, and taking the sum of the correlation coefficients between the target sensor and all sensors as the denominator, and dividing the two to obtain the weight coefficient of the reference sensor. The correlation coefficient reflects the similarity of data changes between sensors. The correlation coefficient ranges from -1 to 1, and the greater the absolute value, the stronger the correlation. The strategy also includes a confidence evaluation mechanism to evaluate the reliability of the estimated data according to the data quality and correlation strength of the related sensors.

[0050] A degradation control strategy is established. The strategy automatically adjusts the complexity of the control algorithm according to the number and type of available sensors. When more than 30% of the sensors fail, the system automatically switches to a simplified control mode, using only temperature sensors for basic temperature control. When more than 50% of the sensors fail, the system switches to a safety mode, using a pre-set conservative control strategy to ensure personnel safety. The switching of different control modes is based on the importance weighted score of the failed sensors: The calculation method of the failure score is: multiplying the importance weight of each failed sensor by the corresponding failure indication value, and then adding all the product results to obtain the total failure score. The score reflects the influence of the current failure state on the overall performance of the system. The importance weight is determined according to the influence of the sensor on the control performance, and the weight of the key sensor is higher, and the weight of the auxiliary sensor is lower.

[0051] An online calibration algorithm for failed sensors is designed. The algorithm uses the information of redundant sensors to recalibrate the failed sensors. The algorithm includes three steps of fault detection, fault location and parameter correction. Fault detection determines whether the sensor is faulty by statistical hypothesis testing, and the calculation method of the test statistic is: taking the difference between the current measurement value of the sensor and its historical mean value as the numerator, and taking the historical standard deviation divided by the square root of the sample number as the denominator, and dividing the two to obtain the test statistic. The historical mean value reflects the typical value of the sensor when it is working normally, the historical standard deviation reflects the dispersion of the data, and the square root of the sample number is used for standardization.

[0052] Fault location determines the fault type through residual analysis, including offset fault, gain fault and drift fault. Parameter correction re-estimates the calibration parameters of the sensor through least squares method.

[0053] The output of this step is a complete environmental state estimation vector: , including temperature, humidity, air pressure, wind speed, light and noise, etc. The complete environmental state estimation vector can provide reliable environmental information even in the case of partial sensor failure. At the same time, the sensor health index is output , wherein represents the sensor The health degree at time , the value range is 0 to 1, 1 represents complete health, 0 represents complete failure, and n represents the number of sensors.

[0054] Optionally, in some embodiments, a virtual sensor system based on digital twin technology is adopted. The system establishes a digital model of the vehicle and the environment, and provides virtual sensor data through simulation calculation as a supplement and verification of physical sensors. The digital twin model can synchronize the state of the physical system in real time, and provide high-precision virtual sensor data; Similarly, in some embodiments, a distributed fault diagnosis algorithm based on edge artificial intelligence is adopted, and a distributed optimization algorithm is adopted to realize multi-region collaborative control. The distributed fault diagnosis algorithm based on edge artificial intelligence deploys lightweight fault diagnosis models on multiple edge computing nodes of the vehicle, which can quickly detect and locate sensor faults, and improve the accuracy and real-time performance of fault diagnosis through the collaborative work between nodes; The distributed optimization algorithm divides the in-vehicle space into multiple independent control regions, each region is configured with an independent optimization controller, and the global optimal control is realized through information exchange and negotiation between regions. This method can better meet the individual needs of different regions.

[0055] S500, an actuator adaptive compensation control is adopted, the actuator characteristic parameters are estimated through online identification model, and a feedforward compensation algorithm is designed to eliminate the actuator characteristic changes, so as to realize multi-actuator coordination; The input is a hierarchical optimization control instruction And an actuator performance monitoring index, the target is to compensate the influence of actuator characteristic changes on control performance, and to ensure the accurate execution of control instruction.

[0056] An online identification model of actuator characteristics is established. The online identification of actuator characteristics adopts the recursive least squares method, which is a classical algorithm in control theory. The algorithm tracks the changes of actuator characteristics by updating parameter estimates in real time. The core idea of the algorithm is to use new input and output data to correct parameter estimates. In specific implementation, a forgetting factor mechanism is adopted, so that the algorithm has good tracking ability to parameter changes. Taking a compressor as an example, its dynamic characteristics are described by a first-order inertia link plus pure lag: , wherein is the transfer function of the actuator, which describes the dynamic characteristics of the actuator, is the static gain, is the time constant, is the pure time delay, is the complex variable of Laplace transform, which is used to convert time-domain signals to frequency domain.

[0057] An adaptive feedforward compensation algorithm is designed. The design of adaptive feedforward compensation is based on the principle of inverse model, that is, a compensator opposite to the actuator transfer function is designed to offset the influence of actuator characteristics. The design of compensator takes into account the time constant and pure time delay characteristics of the actuator, and ensures the realizability of the compensator through reasonable approximation processing. The algorithm designs a compensator according to the identified actuator characteristic parameters to offset the influence of actuator characteristic changes on control performance. The control input after compensation is: The calculation method of feedforward control output is to multiply the transfer function of feedforward controller with reference input signal to obtain the control output of compressor. The feedforward controller is designed according to the inverse model of the system, which can pre-compensate the influence of known disturbance; the algorithm also contains an adaptive adjustment mechanism of compensator parameters, which updates the parameters of compensator synchronously when the actuator characteristics change.

[0058] A multi-actuator coordinated control strategy is implemented. This strategy adopts a master-slave control structure, taking the compressor as the master actuator and the fan and damper as the slave actuators. The master actuator is responsible for the main temperature regulation, and the slave actuators are responsible for auxiliary regulation and coordination. The core of coordinated control is to establish the coupling model between actuators to describe the comprehensive influence of each actuator output on system performance. The coupling matrix is: where, represents the transfer function of the i-th output variable to the j-th input variable. Each element in the matrix is a transfer function in Laplace domain, which describes the dynamic response characteristics of input signal after passing through the system, is the complex variable of Laplace transform; The relationship between output variables and input variables is established through the transfer function matrix, that is, the output vector is equal to the product of the transfer function matrix and the input vector. The output vector contains three components of output temperature, output relative humidity and output air speed, and the input vector contains three components of compressor control input, fan control input and damper control input. The control allocation scheme of each actuator is obtained through matrix operation: the input vector is equal to the product of the inverse matrix of the transfer function matrix and the desired output vector. In order to avoid numerical problems of matrix inversion, singular value decomposition method is adopted: the transfer function matrix is decomposed into the product of three matrices, which contain two orthogonal matrices and a diagonal matrix, and then the inverse matrix is obtained by recombining these matrices.

[0059] A mechanism for handling actuator saturation is established. In extreme environments, actuators can reach their physical limits and become saturated. This mechanism includes three parts: saturation detection, control allocation, and anti-windup compensation. Saturation detection determines the saturation state by monitoring whether the actuator output reaches the upper or lower limit. When saturation is detected, the control allocation algorithm redistributes the control tasks of each actuator, transferring the load of the saturated actuator to other actuators. Anti-windup compensation prevents integral saturation by modifying the integrator input: when the actuator is saturated, the modified integrator input is obtained by subtracting the product of the anti-windup gain and the saturation error from the original integrator input. The saturation error is the difference between the actual control output after saturation and the expected control output, and the anti-windup gain is used to adjust the strength of the modification.

[0060] An actuator fault detection and isolation algorithm is designed. This algorithm detects actuator faults based on residual generation and threshold decision principles. Residual generation calculates the residual by comparing the expected output of the actuator with the actual output: the residual value at a certain time is obtained by subtracting the expected output from the actual output at that time. The residual reflects the degree of deviation between the actual performance and the expected performance of the actuator. The statistical characteristics of the residual, including mean, variance, and peak, are calculated using a sliding window statistical method. The fault detection threshold is determined through statistical analysis of historical data, and when the statistical characteristics of the residual exceed the threshold, a fault is detected. Fault isolation determines the fault type and location by analyzing the patterns of different residuals. The algorithm also includes fault severity assessment, which classifies faults into three levels: slight, moderate, and severe, based on the degree of impact on system performance.

[0061] The output of this step is the compensated actuator control signal This control signal contains three parts: compressor compensation control signal, fan compensation control signal, and damper compensation control signal. These signals have been optimized through characteristic compensation, coordinated control, and saturation handling, so that good control performance can be maintained even when the characteristics of the actuators change; The health status information of the actuators includes three indicators: compressor health status, fan health status, and damper health status, which are used to monitor the operation of each actuator in real time.

[0062] Optionally, in some embodiments, a robust compensation algorithm based on sliding mode control is used. This algorithm designs a sliding surface and a reaching law to make the system maintain good robustness under actuator parameter changes and external disturbances. Sliding mode control is not sensitive to parameter changes and disturbances, making it particularly suitable for control applications in extreme environments.

[0063] Similarly, in some embodiments, the proactive maintenance of the actuators is implemented using predictive maintenance techniques. This technique monitors the operating parameters of the actuators, such as vibration, temperature, current, etc., to establish a device degradation model, predict the remaining useful life of the actuators, and perform preventive maintenance before failure occurs.

[0064] S600, implement fault self-healing and degradation control, establish an expert system to detect faults, and divide system operation into four modes of normal, energy saving, safety, and emergency according to the degree of influence of the faults; The input of this step is the health status information of each component of the system, and the goal is to achieve automatic fault handling and system function degradation when the system fails, and to ensure the continuous availability of critical functions.

[0065] A system-level fault diagnosis expert system is established. This system integrates rule-based reasoning and case-based reasoning, and can accurately diagnose complex system-level faults. Rule-based reasoning is based on a fault diagnosis rule base established by expert knowledge, which contains the mapping relationship between fault symptoms, fault causes and diagnosis conclusions. The rule base contains more than 200 diagnostic rules, covering sensor faults, actuator faults, controller faults and communication faults, etc. Case-based reasoning finds the most similar case to the current fault symptoms in the historical fault case library through similarity matching, and learns from the handling method of the historical case. The similarity calculation uses weighted Euclidean distance: The calculation method of case similarity is as follows: for each feature dimension, calculate the difference between the feature values of the two cases and square it, then multiply by the corresponding weight coefficient, sum the weighted squared differences of all dimensions, add 1, and finally take the reciprocal to get the similarity value.

[0066] To ensure the comparability of different types of features, each feature needs to be standardized before calculation: numerical features (such as temperature, pressure, current, etc.) are standardized by Z-score, categorical features (such as fault type, device model, etc.) are converted to numerical vectors by one-hot encoding, and time features are normalized by timestamp. The weight coefficient is set according to the importance of the feature, with key features weighted at 0.3-0.5 and general features weighted at 0.1-0.2, and the sum of all weights equal to 1. The results of the two reasoning methods are integrated by evidence fusion algorithm to get the final fault diagnosis conclusion.

[0067] A fault impact assessment model is designed. The model quantifies the impact of different faults on system performance, providing a basis for the selection of fault handling strategies. Impact assessment is conducted from three dimensions: functional impact, safety impact, and task impact. Functional impact assesses the impact of faults on the basic functions of the system, including temperature control, humidity control, and air volume control. Safety impact assesses the threat of faults to personnel safety, using a risk matrix method: multiplying the probability of fault occurrence and the severity of fault consequences to get the risk level value. Fault probability is determined through historical data statistics and reliability analysis, and consequence severity is evaluated according to the impact of faults on system function, safety, and task completion; among them, fault probability represents the possibility of fault occurrence (value range 0-1), and consequence severity represents the loss degree caused by fault (classified as 1-5 levels). Task impact assesses the impact of faults on task execution, including task delay, task quality decline, and task failure. The comprehensive impact score is obtained by weighted summation: The calculation method of total impact score is: multiply the functional impact score by the corresponding weight coefficient, multiply the safety impact score by the corresponding weight coefficient, multiply the task impact score by the corresponding weight coefficient, and then add the three weighted scores to get the total impact score.

[0068] To ensure the comparability of different impact dimensions, each impact score needs to be normalized before calculation: functional impact score is normalized by dividing the functional loss degree by the total number of system functions, with a value range of 0 to 1; safety impact score is normalized by dividing the risk level by the highest risk level, with a value range of 0 to 1; task impact score is normalized by dividing the task completion loss by the total task completion, with a value range of 0 to 1. The weight coefficient is set based on the safety priority principle, with the highest safety impact weight of 0.5, the functional impact weight of 0.3, and the task impact weight of 0.2, ensuring that the system prioritizes safety in the event of a fault.

[0069] An automatic fault isolation mechanism is implemented. This mechanism isolates the fault component immediately after detecting the fault to prevent the fault from spreading and affecting the entire system. Fault isolation includes physical isolation and logical isolation. Physical isolation is achieved by cutting off the power supply or communication connection of the fault component, suitable for hardware faults. Logical isolation disables the function of the fault component by modifying the software configuration, suitable for software faults. Isolation decisions are based on fault impact assessment results: when the fault impact score exceeds 0.7, immediate isolation is performed; when the fault impact score is between 0.4 and 0.7, controlled isolation is performed; when the fault impact score is less than 0.4, no isolation is performed but monitoring is strengthened. Isolation operations are implemented through a fault isolation matrix, which defines the isolation operations corresponding to different fault types.

[0070] A multi-level degradation control strategy is established. The strategy divides the system operation into four modes: normal mode, energy-saving mode, safety mode, and emergency mode, according to the available resources and the severity of the fault. In normal mode, the system operates with full functionality, providing optimal comfort and control accuracy. In energy-saving mode, the system reduces the performance of non-critical functions, prioritizes core functions, and reduces energy consumption by 30%. In safety mode, the system only maintains basic temperature control functions to ensure personnel safety, reducing energy consumption by 60%. In emergency mode, the system switches to a pre-set fixed control strategy, providing only the minimum life support, reducing energy consumption by 80%. The decision conditions for mode switching are based on the number of available sensors, the number of available actuators, and the severity of the fault: The calculation method of availability is to divide the number of currently available components by the total number of system components, then multiply by 100% to get the availability in percentage form. This indicator reflects the overall health and service capacity of the system in the current state.

[0071] where availability represents the percentage of available system components. Switch to safety mode when the number of available sensors is less than 60% or there is a serious fault, and switch to emergency mode when the number of available sensors is less than 30% or there is a dangerous fault.

[0072] Design a fault recovery strategy. This strategy realizes the smooth recovery of system performance after fault repair, avoiding system oscillation or instability during the recovery process. Fault recovery includes two stages: function recovery and performance recovery. The function recovery stage re-enables the repaired components and restores the corresponding system functions. The performance recovery stage gradually improves system performance and eventually recovers to the level before the fault. The recovery process uses a gradual strategy, with each recovery step lasting 30 seconds to ensure that the system reaches a stable state before proceeding to the next step. The recovery progress evaluation formula is: The calculation method of recovery progress is to divide the number of recovered functions by the total number of functions, then multiply by 100% to get the recovery progress percentage. This percentage reflects the degree of completion of system function recovery. During the recovery process, the system performance indicators are continuously monitored, and if any abnormalities are found, the recovery is immediately stopped and the system is returned to a safe state.

[0073] The output of this step is the system operation mode instruction, which takes the values of normal, energy-saving, safety, or emergency, indicating the operation mode that the current system should adopt. At the same time, the fault handling status report is output, including fault diagnosis results, fault impact assessment, isolation operation records, and recovery progress, etc.

[0074] Optionally, in some embodiments, a digital twin-based fault prediction technology is adopted. This technology establishes a digital model of the system, predicts possible faults through simulation analysis, and takes preventive measures in advance. The digital twin model can synchronize the state of the physical system in real time, provide accurate fault prediction and system performance evaluation; similarly, in some embodiments, a swarm intelligence algorithm is used to optimize the fault handling strategy. This algorithm models the selection of fault handling strategies as an optimization problem, and finds the optimal fault handling scheme through intelligent optimization methods such as particle swarm algorithm and ant colony algorithm, which can find better handling strategies in complex fault scenarios.

[0075] S700, intelligent power management is performed, an energy consumption model is established, power distribution is optimized based on Markov decision process, and dynamic power limiting and hibernation mechanism are implemented; The input of this step is vehicle energy state information and task priority setting, and the goal is to achieve optimal allocation and management of energy consumption under the premise of ensuring key functions.

[0076] A vehicle energy consumption model is established. This model takes into account the energy consumption characteristics of the power system, air conditioning system, lighting system, communication system and other subsystems. The power system energy consumption model is based on vehicle rolling resistance, air resistance and slope resistance. The calculation method of power system power is: add the rolling resistance, air resistance and slope resistance to get the total resistance, then multiply by the vehicle speed to get the resistance power, and finally divide by the transmission efficiency to get the power required by the power system. Among them, the rolling resistance comes from the friction between the tire and the ground, the air resistance comes from the wind resistance when the vehicle is running, and the slope resistance comes from the gravity component when the vehicle climbs a slope.

[0077] The calculation method of air conditioning system power is: divide the refrigerating capacity by the refrigeration coefficient to get the compressor power, then add the fan power and control system power to get the total power of the air conditioning system. Among them, the refrigerating capacity reflects the refrigerating capacity of the air conditioning system, and the refrigeration coefficient reflects the refrigeration efficiency.

[0078] The calculation method of other subsystem power is: multiply the rated power by the load rate to get the actual power. Among them, the rated power is the maximum power of the subsystem, and the load rate reflects the actual working intensity of the subsystem.

[0079] A power distribution strategy based on Markov decision process is designed. This strategy models the power distribution problem as a Markov decision process, the state space includes remaining power, environmental conditions, task requirements, etc., the action space includes the power distribution proportion of each subsystem, and the reward function considers task completion, energy efficiency and system stability. State transition probability is obtained by statistical analysis of historical data: The calculation method of state transition probability is to divide the number of times of transition from a certain state to a target state by the total number of times of executing the action in that state. Among them, the transition times are counted by traversing the historical trajectory data, and the total number is the total number of records of executing the action in that state. The counting process uses a sliding window mechanism, only considering the data of the last 1000 decision cycles, to ensure that the probability estimation reflects the current characteristics of the system. The value function is solved by Bellman equation: The calculation method of value function is to calculate the expected return of each possible action in the current state, and then select the value corresponding to the action with the maximum expected return as the value of the state. The calculation of expected return includes immediate reward and discounted value of future reward, where future reward is obtained by weighted sum through state transition probability. The discount factor is set to 0.9 to balance the importance of immediate reward and long-term reward. The argmax function in optimal policy selection represents selecting the action with the maximum Q value among all possible actions. The specific implementation process is to calculate the Q values corresponding to all possible actions for a given state s, then compare the sizes of these Q values, and select the action with the maximum Q value as the optimal action. When there are multiple maximum Q values, a random selection strategy is used to avoid decision bias.

[0080] A dynamic power limiting mechanism is implemented. This mechanism dynamically adjusts the power upper limit of each subsystem according to the remaining energy and task demand, ensuring the energy supply of critical functions. The power limiting strategy adopts a hierarchical control principle, dividing electrical equipment into four levels according to importance: life support level (respiratory system, communication equipment), safety guarantee level (lighting, alarm system), task guarantee level (air conditioning system, navigation system), and comfort guarantee level (entertainment system, auxiliary lighting). When the remaining power is less than 80%, limit the power of comfort guarantee level devices; when the remaining power is less than 60%, limit the power of task guarantee level devices; when the remaining power is less than 40%, limit the power of safety guarantee level devices; when the remaining power is less than 20%, only guarantee the operation of life support level devices. The power limiting algorithm is: The calculation method of power limiting value is to multiply the rated power of the device by the corresponding limiting factor to get the power limiting value of the device at the current time. Among them, the limiting factor is a function of the remaining power, which determines the different limiting degrees according to different levels of remaining power.

[0081] An energy consumption prediction model is established. This model estimates future energy consumption demand based on historical data and environmental prediction, providing forward-looking information for power allocation. Energy consumption prediction uses multivariate time series analysis method to establish the relationship model between energy consumption and environmental temperature, vehicle speed, load weight, etc.: The calculation method of energy consumption prediction is: adding the constant term, the autoregressive term of historical energy consumption and the external influencing factor term, and then adding the prediction error term to obtain the predicted energy consumption at the future time. Among them, the autoregressive term is obtained by multiplying the historical energy consumption values by the corresponding autoregressive coefficients and then summing them up, and the external influencing factor term is obtained by multiplying the external factor values by the corresponding regression coefficients and then summing them up. The prediction time domain, the autoregressive order and the number of external factors are determined according to the specific application scenario. The model parameters are estimated by the least squares method, and the prediction accuracy is evaluated by cross-validation. The prediction model also includes uncertainty quantification, which describes the confidence interval of the prediction result through probability distribution.

[0082] An intelligent sleep strategy is designed. This strategy automatically enters a low-power mode when the system is idle, reducing unnecessary energy consumption. The sleep strategy contains two levels of system-level sleep and component-level sleep. System-level sleep is started when the vehicle is parked for a long time and there is no human activity, and unnecessary system functions are turned off, reducing power consumption by 90%. Component-level sleep is started when a single component is idle, switching the component to a low-power mode, reducing power consumption by 60%-80%. Sleep decisions are based on activity detection and demand prediction: when no human activity is detected for 30 consecutive minutes and no scheduled tasks are scheduled for the next 2 hours, system-level sleep is started; when a single component is not used for 5 consecutive minutes and no usage demand is expected for the next 30 minutes, component-level sleep is started. The sleep decision function is: The calculation method of the sleep decision value is: multiplying the activity detection score, the demand prediction score and the energy state score by the corresponding weight coefficients respectively, and then adding the three product results to obtain the sleep decision value. This decision value considers three key factors: system activity state, future demand and energy status. To ensure the comparability of different score indicators, each score needs to be normalized before calculation: activity detection is normalized by dividing the detected activity intensity value by the historical maximum activity intensity value, with a value range of 0 to 1; demand prediction is normalized by dividing the predicted demand value by the maximum processing capacity of the system, with a value range of 0 to 1; energy status is normalized by dividing the current remaining power by the total capacity of the battery, with a value range of 0 to 1. The weight coefficients are set to 0.4, 0.3 and 0.3 respectively to ensure the comprehensive consideration of the three factors. The wake-up mechanism is realized through sensor monitoring and timer to ensure that the system functions can be quickly restored when needed.

[0083] The output of this step is a power allocation scheme which specifies the power allocation of each subsystem at the current time. At the same time, the energy consumption prediction report is output, which includes the energy consumption prediction curve, the uncertainty interval and the key influencing factor analysis for the next 24 hours.

[0084] Optionally, in some embodiments, a hybrid power system that combines multiple energy sources such as solar panels and fuel cells is employed. The system optimizes the sequence and proportion of multiple energy sources through intelligent energy management algorithms, improving energy utilization efficiency and extending the range. The energy management algorithm takes into account the characteristics, cost and environmental impact of various energy sources; similarly, in some embodiments, a blockchain-based energy trading mechanism is employed. In the multi-vehicle collaborative work scenario, vehicles can trade energy through smart contracts, achieving optimized allocation and mutual supply of energy. This mechanism can improve overall energy utilization efficiency and enhance system resilience.

[0085] In one embodiment of the present application, an application example of an intelligent control method for an automobile air conditioner controller based on an adaptive neural network is provided. Application verification was carried out in high-temperature, high-cold and high-altitude areas, and the system control performance under different environmental conditions was compared as shown in Table 1. Table 1: Comparison of system control performance under different environmental conditions Table 1 shows the performance improvement of the method of the present application compared with the traditional control method under different extreme environmental conditions. The data is derived from a 6-month field test, with at least 100 hours of continuous testing under each environmental condition; The system fault tolerance performance test results are shown in Table 2. Table 2: System fault tolerance performance test results Table 2 shows the fault tolerance performance of the system under different fault scenarios. Fault detection time refers to the time interval from fault occurrence to system detection of fault, system recovery time refers to the time interval from fault detection to system recovery and stable operation, and function retention rate refers to the proportion of functions that the system can still maintain during fault; The energy consumption optimization effect comparison analysis is shown in Table 3. Table 3: Energy consumption optimization effect comparison analysis Table 3 shows the energy consumption optimization effect of the present application under different application scenarios. The data is derived from long-term tracking tests, with at least 30 days of operation data collected for each scenario. Range improvement refers to the extension of working time under the same battery capacity; The technical solution implementation effect verification data is shown in Table 4. Table 4 compares the design targets and actual implementation effects of the technical solution, verifying the effectiveness and practicality of the technical solution. All core technical indicators meet or exceed the design targets, proving the advanced nature and practical value of the present application.

[0086] The above describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, which is only illustrative but not restrictive. Those skilled in the art can make more equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. An intelligent control method for an automobile air conditioner controller based on an adaptive neural network, characterized in that, The method comprises the following steps: Collecting multi-source data inside and outside the vehicle through a sensor array, performing data fusion using a timestamp synchronization algorithm, establishing a data quality evaluation model, repairing multi-source data based on an LSTM network, and outputting a fusion data vector; Based on the fusion data vector, a double-path adaptive neural network is constructed, a knowledge distillation technique is used to achieve lightweight, and the weight coefficients are dynamically adjusted through an environment change detector to achieve model quantization compression; Based on the fusion data vector and the user-set comfort target, hierarchical multi-objective optimization control is performed, a three-layer architecture of strategy, tactics, and operation is established, a target weight adaptive adjustment mechanism is designed, and the optimal air conditioning control instruction is output; A three-level sensor fault diagnosis and fault tolerance mechanism is established, and a virtual sensor algorithm is used to estimate the output value of the faulty sensor; An adaptive compensator is used to estimate the actuator characteristic parameters through online identification model, and a feedforward compensation algorithm is designed to eliminate the actuator characteristic changes, realizing multi-actuator coordination; Fault self-healing and degradation control are realized, an expert system is established to detect faults, and the system operation is divided into four modes: normal, energy saving, safety, and emergency according to the fault influence degree; Intelligent power management is performed, an energy consumption model is established, power distribution is optimized based on Markov decision process, and dynamic power limitation and hibernation mechanism are realized. 2.The intelligent regulation and control method of an automobile air conditioner controller based on an adaptive neural network according to claim 1, characterized in that, The sensor array includes temperature sensors, humidity sensors, air pressure sensors, infrared thermal imaging sensors, wind speed sensors, and personnel detection sensors, which are divided into high-frequency, medium-frequency, and low-frequency groups according to the sampling frequency. The high-frequency group includes infrared thermal imaging sensors and wind speed sensors, with a sampling frequency of 100 Hz. The medium-frequency group includes temperature sensors and humidity sensors, with a sampling frequency of 10 Hz. The low-frequency group includes air pressure sensors and personnel detection sensors, with a sampling frequency of 1 Hz. The data quality evaluation model uses a weighted combination of information entropy and signal-to-noise ratio to calculate the data quality index. When the data quality index is less than 0.7, it is marked as low-quality data. The low-quality data is repaired using the time series interpolation method of the LSTM network. 3.The intelligent regulation and control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, The knowledge distillation technique in the double-path adaptive neural network includes a teacher network training phase and a student network distillation phase. The teacher network includes a three-layer LSTM structure, with 128 neurons in the first layer, 256 neurons in the second layer, and 128 neurons in the third layer. The student network includes 64 LSTM units and 32 fully connected layer neurons. The distillation loss function is a weighted combination of the cross-entropy loss of the student network output and the true label and the KL divergence of the student network output and the teacher network output.

4. The intelligent control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, The dynamic adjustment of the weight coefficients through the environment change detector includes dynamically adjusting the weight coefficients of the main path LSTM network and the auxiliary path CNN network according to the environment change degree index. The main path weight basic value is 0.8, which is adjusted by the sigmoid function according to the environment change degree. The auxiliary path weight is complementary to the main path weight, and the sum of the two weights is always 1. The environment change degree index is obtained by calculating the deviation degree of the current environment parameters and the historical mean.

5. The intelligent control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, The hierarchical multi-objective optimization control includes three levels of strategy layer, tactical layer and operation layer, the strategy layer solves task priority allocation and energy consumption budget planning, the time scale is hour level, the tactical layer makes medium-term prediction and control decision, the time scale is minute level, the operation layer performs real-time control execution, the time scale is second level, the coordination optimization between each level is realized through target decomposition and feedback mechanism, and the Pareto optimal solution set is used for multi-objective trade-off.

6. The intelligent regulation and control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, In the three-level sensor fault diagnosis and fault-tolerant mechanism, the key level sensor includes an indoor temperature sensor and a personnel detection sensor, the important level sensor includes a humidity sensor and a wind speed sensor, and the general level sensor includes an air pressure sensor and an illumination sensor; the virtual sensor algorithm estimates the output value of the faulty sensor by combining a physical model and a data-driven model, the temperature transfer equation is established based on the thermodynamic law, and the support vector regression algorithm is used in the data-driven model.

7. The intelligent control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, In the adaptive compensation control of the actuator, the online identification model uses the recursive least squares method to estimate the dynamic characteristic parameters of the actuator in real time, and the forgetting factor is set to 0.95; the feedforward compensation algorithm designs a compensator according to the identified actuator characteristic parameters to offset the influence of actuator characteristic changes on control performance, and the multi-actuator coordination adopts a master-slave control structure to take the compressor as the master actuator responsible for temperature regulation, and takes the fan and the damper as the slave actuators responsible for auxiliary regulation and coordination. 8.The intelligent regulation and control method of an automobile air conditioner controller based on an adaptive neural network according to claim 1, characterized in that, In the fault self-healing and degradation control, the fault diagnosis expert system integrates rule-based reasoning and case-based reasoning, the rule-based reasoning includes more than 200 diagnostic rules based on the fault diagnosis rule base established by expert knowledge, the case-based reasoning finds the most similar case in the historical fault case library through similarity matching, and the fault influence degree is quantified from three dimensions of function influence, safety influence and task influence. 9.The intelligent control method of the automobile air conditioner controller based on the adaptive neural network according to claim 1, characterized in that, The energy consumption model comprehensively considers the energy consumption characteristics of the power system, the air conditioning system, the lighting system, the communication system and each subsystem; the power distribution strategy is modeled as a Markov decision process, and the power distribution decision is optimized by setting the state space, the action space and the reward function.

10. A computer-readable storage medium, characterized in that, The computer readable instructions can run the intelligent control method of the automobile air conditioner controller based on the adaptive neural network when the computer readable instructions are read by the computer. The computer readable instructions can run the intelligent control method of the automobile air conditioner controller based on the adaptive neural network when the computer readable instructions are read by the computer.

Images