Automobile air conditioner heat load prediction method based on multi-source thermal environment data fusion

Abstract

The present application relates to a vehicle air conditioner heat load prediction method based on multi-source thermal environment data fusion, through unified time synchronization collection, standardized processing, outlier rejection, semantic mapping and logical rule verification, the automatic identification of key thermal physical state variables such as cabin, evaporator and battery and the weight freezing of abnormal channel are realized. Combined with the inversion of heat transfer equation and the modeling of probability density function, the abnormal data is resampled and replaced, the corrected multi-source data stream is input into the prediction model, the real-time net heat load power prediction value is output, and the online adaptive optimization of the rule base is realized combined with feedback. The method improves the data quality and prediction accuracy of thermal management, and enhances the system robustness.

Description

Technical Field

[0001] This invention relates to the field of vehicle thermal management and multi-source sensor data fusion technology, and in particular to a method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion. Background Technology

[0002] The development of intelligent automotive thermal management systems has driven innovation in real-time prediction technology for vehicle air conditioning heat load. Mainstream solutions often employ methods based on multi-source sensor data fusion, time-series modeling, and deep learning to improve prediction accuracy and control response. Among these, the acquisition, fusion, and analysis of multi-scale thermal environment data have become a key focus in the industry. For example, existing solutions commonly use Kalman filtering, wavelet decomposition, attention mechanisms, and signal-level weighting to process heterogeneous temperature, pressure, and electrical data to suppress noise, compensate for sampling errors, and improve model robustness. Meanwhile, in recent years, data-driven temperature control optimization and dynamic heat load assessment have been widely applied to energy consumption management during vehicle parking, initial driving, and complex operating conditions.

[0003] However, existing multi-source data fusion technologies suffer from several prominent problems when dealing with multi-scale, heterogeneous data. First, conventional filtering and weighting methods, when fusing sensor data of different granularities and precisions, cannot adequately identify noise disturbances from low-quality sensors. This leads to noise amplification during the fusion process, not only contaminating key feature representations but also potentially causing model predictions to deviate significantly from actual thermal conditions. The root cause lies in the fact that traditional signal processing paradigms focus only on numerical statistical consistency, failing to determine the validity of sensor data at the thermophysical semantic level and whether it conforms to fundamental engineering logic such as heat transfer, energy conservation, and thermal inertia.

[0004] Therefore, current technology urgently needs a novel multi-source data fusion method that can assess data quality from the perspective of thermophysical semantic consistency and dynamically suppress noise amplification effects. This method can utilize the professional engineering knowledge of vehicle thermal management systems to enable the prediction model to receive only high-confidence state variables at the semantic level during time-series extrapolation, effectively improving the robustness and generalization ability of the model under complex operating conditions. Summary of the Invention

[0005] This application provides a method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion, aiming to solve the problem mentioned in the background art that the traditional signal processing paradigm only focuses on numerical statistical consistency and has difficulty in judging the validity of sensor data at the thermophysical semantic level.

[0006] The automotive air conditioning heat load prediction method based on multi-source thermal environment data fusion provided in this application specifically includes: S1: Obtain the original multi-scale thermal environment dataset during the initial stage of vehicle startup or under steady-state parking conditions; S2: Based on the original dataset of the multi-scale thermal environment, a lightweight thermal state encoder trained offline is used to perform mapping transformation processing to transform the original sensor readings into a set of thermal physical semantic state variables. S3: Call the preset thermal logic rule base to perform a consistency check operation on the set of thermal physical semantic state variables, identify semantic conflict markers that violate thermal physical common sense by comparing the heat transfer equation constraint relationship between variables, and generate a thermal semantic consistency check result containing abnormal channel identifiers. S4: For the abnormal channels marked in the thermal semantic consistency verification results, freeze the contribution weight of the abnormal channels in the current time window and solve the theoretically required reading range in reverse according to the thermal physical semantic state variables of the remaining reliable channels, thereby constructing a narrowband probability density function. S5: The original data of the abnormal channel is resampled and replaced using the narrowband probability density function to eliminate the noise amplification effect and output the multi-source fused data stream after semantic consistency correction; S6: Input the multi-source fusion data stream into the time-series prediction model, perform net heat load power extrapolation calculation, and directly output the real-time heat load prediction value adapted to the control requirements of the automotive air conditioning system; S7: Based on the deviation between the real-time heat load prediction value and the actual operation feedback of the vehicle thermal management system, dynamically update the engineering knowledge accumulation parameters in the thermal logic rule base to complete the online iterative optimization of the thermophysical semantic consistency constraint conditions.

[0007] The automotive air conditioning heat load prediction method based on multi-source thermal environment data fusion provided in this application has the following beneficial effects: By introducing a state coding mechanism at the thermophysical semantic level, this application effectively overcomes the noise sensitivity and feature contamination problems caused by the direct dependence on raw sensor signals in traditional data-driven methods for predicting automotive air conditioning heat load. This scheme innovatively constructs a lightweight thermal state encoder, mapping raw data into intermediate variables with clear thermodynamic interpretations, such as "cabin air thermal inertia index" and "evaporator condensation tendency factor," enabling the data fusion process to leap from low-level signal processing to high-level physical behavior understanding, significantly improving the consistency and interpretability of multi-source information representation. Based on this, semantic consistency verification is performed using a pre-built thermal logic rule base, which can proactively identify variable combinations that violate heat transfer laws and accurately locate potential data anomaly channels. This achieves a paradigm shift from "passive filtering" to "active diagnosis," significantly reducing the false positive and false negative rates, and providing a high-fidelity, physically consistent input foundation for subsequent prediction models. Attached Figure Description

[0008] Figure 1 This is the main flowchart of a method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion.

[0009] Figure 2 This is a sub-flowchart of a method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion.

[0010] Figure 3 This is another sub-flowchart of the automotive air conditioning heat load prediction method based on multi-source thermal environment data fusion. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1 As shown, this application provides a method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion, specifically including: S1: Obtain the original multi-scale thermal environment dataset during the initial stage of vehicle startup or under steady-state parking conditions; S2: Based on the original dataset of the multi-scale thermal environment, a lightweight thermal state encoder trained offline is used to perform mapping transformation processing to transform the original sensor readings into a set of thermal physical semantic state variables. S3: Call the preset thermal logic rule base to perform a consistency check operation on the set of thermal physical semantic state variables, identify semantic conflict markers that violate thermal physical common sense by comparing the heat transfer equation constraint relationship between variables, and generate a thermal semantic consistency check result containing abnormal channel identifiers. S4: For the abnormal channels marked in the thermal semantic consistency verification results, freeze the contribution weight of the abnormal channels in the current time window and solve the theoretically required reading range in reverse according to the thermal physical semantic state variables of the remaining reliable channels, thereby constructing a narrowband probability density function. S5: The original data of the abnormal channel is resampled and replaced using the narrowband probability density function to eliminate the noise amplification effect and output the multi-source fused data stream after semantic consistency correction; S6: Input the multi-source fusion data stream into the time-series prediction model, perform net heat load power extrapolation calculation, and directly output the real-time heat load prediction value adapted to the control requirements of the automotive air conditioning system; S7: Based on the deviation between the real-time heat load prediction value and the actual operation feedback of the vehicle thermal management system, dynamically update the engineering knowledge accumulation parameters in the thermal logic rule base to complete the online iterative optimization of the thermophysical semantic consistency constraint conditions.

[0014] like Figure 2 As shown, step S1 involves acquiring a multi-scale raw dataset of the thermal environment during the initial stage of vehicle startup or under steady-state parking conditions. Specifically, this includes: S1.1: Acquire the trigger signal during the initial stage of vehicle startup or under steady-state parking conditions, and perform time reference synchronization calibration processing on the cabin infrared temperature measurement array, evaporator surface thermocouple group, compressor pressure sensor, battery module voltage acquisition unit and vehicle CAN bus interface based on the trigger signal to generate a multi-source sensor synchronous acquisition instruction set with a unified timestamp.

[0015] The system monitors the vehicle's ignition switch status signal and the parking brake sensor level. When the ignition signal changes from OFF to ON or the parking brake remains active for more than a preset time limit, the system determines that the trigger condition has been met and generates a data acquisition start command.

[0016] High-precision time synchronization protocol messages are broadcast to the cabin infrared temperature measurement array, evaporator surface thermocouple group, compressor pressure sensor and battery module voltage acquisition unit, and the internal timer phase of each distributed sensor is aligned with the high-stability crystal oscillator clock source of the vehicle main control chip as a reference.

[0017] Calculate the compensation values ​​for the hardware response delay of each sensor and the transmission delay of the communication bus, and establish a time offset mapping table that includes the differences in infrared array frame rate, thermocouple sampling period, pressure sensor switching time and CAN bus baud rate.

[0018] Based on the time offset mapping table, microsecond-level timestamp correction is performed on the reception time of each sensor data packet to eliminate phase misalignment caused by asynchronous acquisition and generate a multi-source sensor synchronous acquisition instruction set with a unified global time reference.

[0019] Through the above-mentioned time reference synchronization calibration process, the trigger signal of the previous step is transformed into a synchronous acquisition instruction set with strict time sequence consistency, realizing the accurate alignment of multi-source heterogeneous data in the time dimension, laying a solid foundation for eliminating time domain noise interference in subsequent multi-scale thermal environment data fusion.

[0020] For example, when the vehicle ignition switch is turned on and the parking brake is activated, the system uses the 10MHz crystal oscillator of the main control chip as a reference to send synchronization messages to an infrared array with a sampling frequency of 50Hz, a thermocouple group with a sampling frequency of 100Hz, and a pressure sensor with a sampling frequency of 200Hz. The measured communication delay is 5ms for the infrared array, 2ms for the thermocouple group, and 1ms for the pressure sensor. The system constructs a time offset mapping table, subtracting 5ms from the timestamp of the infrared data, 2ms from the timestamp of the thermocouple data, and 1ms from the timestamp of the pressure data. After correction, the timestamp deviation of all sensor data is controlled within ±0.1ms, generating a synchronous acquisition instruction set with a unified time reference, which significantly improves the timing consistency of multi-source data fusion.

[0021] S1.2: Based on the multi-source sensor synchronous acquisition instruction set, drive the cabin infrared temperature measurement array to acquire the spatial temperature distribution sequence, drive the evaporator surface thermocouple group to acquire the surface temperature gradient sequence, drive the compressor pressure sensor to acquire the intake and exhaust pressure difference sequence, drive the battery module voltage acquisition unit to acquire the terminal voltage fluctuation sequence, and drive the vehicle CAN bus interface to read the vehicle speed and air conditioning damper opening control instructions, so as to generate a raw data stream set containing five types of heterogeneous physical quantities.

[0022] Receive the multi-source sensor synchronous acquisition instruction set with a unified timestamp generated by S1.1, and parse the channel identifier and sampling trigger timing parameters in the instruction set.

[0023] The underlying driver interface of the cockpit infrared temperature measurement array is activated based on the channel identifier, the array scanning frequency is configured to 10Hz, non-contact space temperature field data acquisition is performed, and a two-dimensional matrix sequence containing the front, middle and rear areas of the cockpit and the body surface temperature of the occupants is generated.

[0024] The thermocouple array on the surface of the synchronously driven evaporator reads the thermoelectric potential signals of key nodes at the root and middle of the fins at a high-frequency sampling rate of 50Hz. After cold junction compensation calculation, it is converted into a surface temperature gradient sequence to capture the transient heat flow changes during the phase change process.

[0025] The analog-to-digital conversion module of the pressure sensors on the high-pressure and low-pressure sides of the compressor is triggered to collect the refrigerant suction and discharge pressure difference data in real time. The data is then mapped to the corresponding saturation temperature deviation sequence through a lookup table method, reflecting the thermodynamic work state of the refrigeration cycle.

[0026] The CAN messages of the battery module voltage acquisition unit are read in parallel, the single cell voltage fluctuation sequence is extracted, and the battery Joule heat production rate is estimated by combining the internal resistance model to form the original electrical signal flow characterizing the thermal diffusion characteristics of the power battery.

[0027] Subscribe to vehicle speed, motor torque, and air conditioning damper opening control command frames from the vehicle CAN bus gateway, parse the physical quantity values ​​and add the same timestamp label, and construct a sequence of control variables that reflect the vehicle's driving conditions and the state of the air conditioning actuator.

[0028] The raw data streams of the above five types of heterogeneous physical quantities are aligned in memory buffers according to timestamp indexes to form a set of raw data streams containing spatial temperature fields, surface temperature gradients, pressure difference sequences, voltage fluctuations, and control commands.

[0029] By using a multi-channel parallel drive and timestamp binding processing method, the synchronous acquisition command is transformed into five types of heterogeneous raw data streams with spatiotemporal correlation, realizing the complete capture and preliminary structuring of multi-source thermal environment data, and providing a full-dimensional input basis for subsequent frequency normalization.

[0030] For example, under steady-state conditions during the initial vehicle startup phase, the cabin infrared temperature measurement array outputs a 32x32 pixel temperature matrix with an average temperature of 24.5℃; the evaporator thermocouple group collects temperature data at 500 points / second with a gradient change rate of 0.2℃ / s; the compressor intake and exhaust pressure difference remains stable at 1.2MPa; the battery module voltage fluctuation is less than 5mV; the CAN bus vehicle speed is 0km / h, and the damper opening is 30%. The system uniformly timestamps these data of different frequencies and dimensions at the microsecond level and stores them in a circular buffer to ensure the integrity and timeliness of subsequent data processing.

[0031] S1.3: Normalize the sampling frequency and standardize the dimensions of various types of data in the original data stream set. Use linear interpolation to map the cabin infrared temperature field distribution sequence, evaporator surface temperature gradient sequence, compressor intake and exhaust pressure difference sequence, battery module terminal voltage fluctuation sequence, and vehicle speed and air conditioning damper opening control commands at different sampling frequencies to a unified time resolution grid to generate time-aligned standardized multi-source data blocks.

[0032] It receives the raw data stream set containing five types of heterogeneous physical quantities generated by S1.2, analyzes the sampling timestamps and data frame structures of each sensor channel, and identifies the inherent sampling frequency differences of the cabin infrared temperature measurement array, evaporator surface thermocouple group, compressor pressure sensor, battery module voltage acquisition unit and CAN bus interface.

[0033] A unified high-resolution time reference grid is constructed, with a target sampling period of 10 milliseconds, to generate an equally spaced time point sequence covering the current data window, which serves as a common reference axis for multi-source data resampling.

[0034] For low-frequency battery module voltage fluctuation sequences and CAN bus vehicle speed commands, a linear interpolation method is used to calculate the values ​​between adjacent original sampling points at the target time point. The time gap is filled by proportional allocation to achieve time domain refinement of low-frequency signals.

[0035] For the high-frequency fluctuating cabin infrared temperature field distribution sequence and evaporator surface temperature gradient sequence, downsampling processing is performed to select the original sampled value closest to the target time point or extract representative values ​​through local mean filtering, so as to match a unified time resolution and suppress high-frequency aliasing noise.

[0036] Phase alignment correction is performed on the compressor intake and exhaust pressure difference sequence. Based on the transmission delay parameter obtained by S1.1 calibration, time shift compensation is performed on the interpolated data to eliminate dynamic hysteresis error caused by different sensor response speeds.

[0037] Perform dimensional standardization processing, read the engineering unit definitions and ranges of various physical quantities, and apply the minimum-maximum normalization formula to map the data to the [0,1] interval:

[0038] in, For the original physical quantity readings, and These are the lower and upper limits of the physical limits determined during the calibration phase of the sensor, respectively. This is a dimensionless standardized value.

[0039] The five types of data streams, which have been time-aligned and dimensionally standardized, are reorganized into a matrix based on timestamp indexes to form a two-dimensional data matrix with rows corresponding to time steps and columns corresponding to sensor channels.

[0040] By using linear interpolation and normalization, multi-scale heterogeneous raw data is transformed into standardized multi-source data blocks that are time-synchronized and have unified dimensions. This eliminates feature fusion bias caused by asynchronous sampling and differences in magnitude, providing a consistent input basis for subsequent hot state coding.

[0041] S1.4: Based on the time-aligned standardized multi-source data blocks, outlier removal and missing data marking are performed. Noise points exceeding physical limits are identified and filtered through sliding window statistical tests. At the same time, invalid placeholders are filled for data gaps caused by brief communication interruptions to generate a clean multi-source data stream that has been pre-screened for quality.

[0042] Receive time-aligned standardized multi-source data blocks, which include the cockpit infrared temperature field, evaporator surface temperature gradient, compressor intake and exhaust pressure difference, battery module terminal voltage, and synchronization sequence of CAN bus control commands.

[0043] Sliding time windows are constructed for data streams from various sensors. The window length is set to a fixed number of sampling points to adapt to the thermal response characteristics of each physical quantity. For example, 10 points are taken for the infrared temperature field and 50 points are taken for the pressure signal.

[0044] Local statistical characteristics of the data, including mean, standard deviation, and extreme values, are calculated within a sliding window and used as benchmark references for subsequent anomaly detection.

[0045] Based on the safe operating boundaries of each physical quantity defined in the vehicle thermal management system engineering knowledge base, a dynamic threshold range is set, which is composed of the theoretical limit value and the allowable fluctuation deviation.

[0046] The real-time sampling points within the sliding window are compared with the dynamic threshold range to identify outliers that exceed the upper and lower limits. These outliers are then marked as hard fault noise and removed.

[0047] For data points that are not identified as hard faults but deviate from the local mean by more than three standard deviations, they are marked as soft noise interference, their original index positions are retained but they are placed in a state to be corrected.

[0048] Detect consecutive missing bits in the data stream and determine whether the duration of the missing bit is less than the preset communication interruption tolerance threshold. If so, mark it as a temporary missing bit.

[0049] For data vacancies marked as transiently missing, fill them with specific invalid placeholders. These placeholders are distinct from the normal value of zero and are used to indicate to subsequent modules to skip the calculation at that position.

[0050] For data points marked as soft noise interference, a weighted average of the nearest valid data is used for initial smoothing to generate temporary substitute values ​​to maintain the continuity of the data stream.

[0051] The data from each channel, after being processed by hard fault removal, soft noise smoothing, and missing bit marking, are integrated while maintaining the original timestamp alignment.

[0052] Through the aforementioned sliding window statistical test and hierarchical processing mechanism, the time-aligned standardized data from the previous step is transformed into a clean multi-source data stream that has undergone quality pre-screening, thereby effectively suppressing non-physical noise in multi-source heterogeneous data and providing preliminary assurance of data integrity.

[0053] S1.5: The clean multi-source data stream that has undergone quality pre-screening is aggregated and recombined according to a preset data structure encapsulation format. The cabin infrared temperature field distribution sequence, evaporator surface temperature gradient sequence, compressor suction and exhaust pressure difference sequence, battery module terminal voltage fluctuation sequence, and vehicle speed and air conditioning damper opening control commands are bound into a single data object to form a multi-scale thermal environment raw dataset to be processed.

[0054] like Figure 3As shown, step S2 involves: based on the original multi-scale thermal environment dataset, performing a mapping transformation process using an offline-trained lightweight thermal state encoder to convert the original sensor readings into a set of thermal physical semantic state variables. Specifically, this includes: S2.1: Based on the original multi-scale thermal environment dataset, perform time window alignment and dimension normalization on the cabin infrared temperature field distribution sequence, evaporator surface temperature gradient sequence, and compressor suction and exhaust pressure difference sequence to eliminate phase deviation caused by different sampling frequencies and generate a standardized multi-source sensor signal matrix.

[0055] The original sampling points of the three types of heterogeneous data streams within a preset sliding time window are extracted. Based on the unified time resolution grid generated in S1.3, cubic spline interpolation is used to resample the low-frequency sampled infrared temperature field data to eliminate phase lag bias caused by sensor response delay. Spatial mean aggregation is performed on the resampled cabin infrared temperature field sequence to reduce the two-dimensional temperature matrix to a one-dimensional cabin average temperature time-series vector, preserving the macroscopic features representing the overall thermal potential energy. For the evaporator surface temperature gradient sequence, the temperature difference ratio between adjacent thermocouple nodes is calculated to construct a relative gradient index reflecting local heat transfer efficiency, eliminating the interference of absolute temperature reference drift on feature extraction. The compressor suction and discharge pressure difference sequence is read, and sliding median filtering is applied to remove high-frequency mechanical vibration noise, extracting the pressure fluctuation envelope as a proxy variable for refrigerant flow rate changes. Min-Max normalization is performed on each physical quantity, mapping the cabin average temperature, evaporator relative gradient, and compressor pressure envelope to the [0,1] interval to eliminate the inhibitory effect of dimensional differences on subsequent neural network weight updates. Z-score standardization further corrects the skewness of the distribution of each variable, generating a standard normal distribution data matrix with a mean of 0 and a variance of 1. Through the above-mentioned time alignment, feature dimensionality reduction, and double standardization processing, the multi-source heterogeneous raw data from the previous step is transformed into a standardized multi-source sensor signal matrix with a unified spatiotemporal reference and statistical characteristics, achieving the expected technical effect of eliminating phase bias and improving the model convergence stability.

[0056] S2.2: Using the feature extraction layer in the pre-built lightweight thermal state encoder, perform multidimensional convolution operations and nonlinear activation processing on the standardized multi-source sensing signal matrix to extract deep spatiotemporal feature vectors characterizing transient thermal response properties.

[0057] It receives a standardized multi-source sensor signal matrix, which contains time-aligned cabin infrared temperature field, evaporator surface temperature gradient, and compressor intake and exhaust pressure difference sequence data.

[0058] By utilizing the first layer of one-dimensional convolutional kernels in a lightweight thermal encoder, local feature scanning is performed on the input signal along the time axis to extract the rate of change and trend components of each physical quantity within a short time window, thereby generating a primary time-series feature map.

[0059] A second-layer two-dimensional convolution kernel group is used to perform neighborhood correlation analysis on the cabin infrared temperature measurement array data in the spatial dimension, capture the spatial distribution non-uniformity of the temperature field and its dynamic diffusion pattern, and form a spatial topological feature map.

[0060] The primary temporal feature map and the spatial topological feature map are spliced ​​and fused along the channel dimension to construct a multidimensional joint feature tensor, so as to preserve the spatiotemporal coupling information.

[0061] By performing element-wise mapping on the joint feature tensor using the nonlinear activation function ReLU, negative responses are suppressed and nonlinear transformation capabilities are introduced, enhancing the model's potential to fit complex thermal dynamic relationships.

[0062] A batch normalization layer is applied to standardize the activated feature data, eliminating gradient oscillations caused by differences in the dimensions of different sensors, accelerating network convergence and improving feature stability.

[0063] By using the Dropout layer to randomly mask some neuron connections with a preset probability, overfitting is prevented and the robustness of feature extraction is improved, resulting in a deep spatiotemporal feature vector that characterizes transient thermal response.

[0064] By using multidimensional convolution operations and nonlinear activation processing, the standardized sensing signals from the previous step are transformed into deep spatiotemporal feature vectors, realizing the transformation from raw data to abstract features with high thermodynamic correlation, and providing high signal-to-noise ratio input for subsequent semantic variable mapping.

[0065] S2.3: Based on the deep spatiotemporal feature vector, the pre-trained thermodynamic mapping head in the lightweight thermal state encoder is called to perform nonlinear regression transformation processing, mapping the abstract feature space coordinates into an intermediate thermodynamic parameter set with a clear physical definition.

[0066] Receive the deep spatiotemporal feature vector output by S2.2, which encodes the transient thermal response pattern and spatial distribution correlation information of multi-source sensor data.

[0067] The pre-trained thermodynamic mapping head in the lightweight thermal state encoder is invoked. This mapping head consists of a multi-layer fully connected neural network and aims to establish a nonlinear mapping relationship from the abstract feature space to the specific thermophysical parameter space.

[0068] The forward propagation calculation is performed, and the deep spatiotemporal feature vector is input into the first hidden layer of the mapping head. Through weighted summation and activation function processing, the intermediate latent variables characterizing the thermal conduction rate and thermal capacity are extracted.

[0069] Based on the first law of thermodynamics and the basic principles of heat transfer, a constraint loss function is constructed to force the model to learn parameter transformation logic that conforms to physical laws during the training phase, thereby ensuring the physical interpretability of the output variables.

[0070] The intermediate thermodynamic parameter set is calculated using the following nonlinear regression formula:

[0071] in, For deep spatiotemporal feature vectors, For intermediate thermodynamic parameter set, This is the nonlinear transformation function of the thermodynamic mapping head.

[0072] Further expanding the internal calculations of the mapping function, for the cabin air thermal inertia index, the ratio of heat capacity to heat transfer coefficient is extracted from the features using the following formula:

[0073] in, The thermal inertia index of the cabin air. , For the weight vector, , This is a bias term.

[0074] For the evaporator condensation tendency factor, the feature is mapped to the interval of 0 to 1 using the sigmoid activation function, characterizing the probability trend of the surface temperature being lower than the dew point temperature:

[0075] in, This is the evaporator condensation tendency factor. For the weight vector, This is a bias term.

[0076] For the battery thermal diffusion hindrance coefficient, the hyperbolic tangent function is used to limit the output range, reflecting the effect of the internal thermal resistance of the battery module on the delay of temperature changes:

[0077] in, This is the battery thermal diffusion hindrance coefficient. For the weight vector, This is a bias term.

[0078] The calculated sub-parameters are combined to form an intermediate thermodynamic parameter set that includes the cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient.

[0079] By using the nonlinear regression transformation of the thermodynamic mapping head, the abstract deep spatiotemporal feature vectors from the previous step are transformed into an intermediate thermodynamic parameter set with clear physical definitions, realizing a precise mapping from data-driven features to physical semantic variables, and providing thermodynamically meaningful quantitative indicators for subsequent consistency verification.

[0080] S2.4: Based on the definition rules in the vehicle thermal management system engineering knowledge base, perform semantic tag association and unit verification processing on the intermediate thermodynamic parameter set to generate a set of thermophysical semantic state variables including cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient.

[0081] The system receives an intermediate thermodynamic parameter set output from a lightweight thermal state encoder. This parameter set contains numerical vectors mapped from an abstract feature space. It then calls upon the semantic definition rule table in the vehicle thermal management system engineering knowledge base to retrieve variable metadata matching the current vehicle platform and air conditioning configuration, including physical dimensions, value ranges, and thermodynamic correlation attributes.

[0082] Dimensionality checks were performed on the intermediate thermodynamic parameters, and the dimensionless neural network output values ​​were converted into engineering units with definite physical meaning using a unit conversion matrix. The physical dimension conversion of the cabin air thermal inertia index was calculated using the following formula:

[0083] in, The thermal inertia index of the cabin air. These are the raw feature values ​​output by the encoder. The linear scaling factor determined during the pre-training phase. This is the zero-point offset.

[0084] Based on the first law of thermodynamics and the fundamental principles of heat transfer, the physical rationality boundaries of the reduced parameters were checked. The effective range for the cabin air thermal inertia index was set to [0.5, 5.0] kJ / K, the effective range for the evaporator condensation tendency factor was [0, 1], and the effective range for the battery thermal diffusion hindrance coefficient was [0.1, 2.0] s. 0.5 Remove abnormal parameter values ​​that exceed the above physical limits and mark the corresponding potential faults in the sensor channels.

[0085] Semantic tag association processing is performed, binding the validated numerical parameters with predefined semantic tags. The cabin air thermal inertia index is associated with the tag "Cabin_Thermal_Inertia", representing the overall state of the crew cabin's heat capacity and heat exchange efficiency; the evaporator condensation tendency factor is associated with the tag "Evap_Dew_Tendency", representing the evaporator surface humidity saturation level and latent heat load risk; and the battery thermal diffusion hindrance coefficient is associated with the tag "Battery_Diffusion_Lag", representing the delay characteristics of heat transfer inside the battery pack.

[0086] Construct a set of thermophysical semantic state variables containing the three labeled variables mentioned above, ensuring that each variable has a traceable physical definition and unit identifier. This set serves as the standard input object for subsequent consistency verification of the thermal logic rule base.

[0087] Through chained processing of dimensional reduction, boundary verification, and semantic label binding, abstract neural network features are transformed into a set of state variables with clear thermodynamic meaning and in line with physical common sense. This achieves a precise mapping from data-driven features to physically interpretable semantics, providing a reliable semantic benchmark for subsequent data noise suppression based on thermophysical laws.

[0088] Step S3: A pre-set thermal logic rule base is invoked to perform a consistency check operation on the set of thermal physical semantic state variables. By comparing the constraint relationships of the heat transfer equations between variables, semantic conflict markers that violate common thermal physics are identified, thereby generating a thermal semantic consistency check result containing abnormal channel identifiers. Specifically, this includes: S3.1: Obtain the cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion resistance coefficient from the set of thermophysical semantic state variables. Construct a thermal logic rule base based on the engineering knowledge of the whole vehicle thermal management system, and transform the constraint relationship of the heat transfer equation into executable Boolean judgment logic and numerical threshold range to form a rule judgment benchmark with thermophysical interpretability.

[0089] The system receives cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient from the set of thermophysical semantic state variables as input data sources for rule determination.

[0090] The system accesses the vehicle thermal management system engineering knowledge base to retrieve the thermal property parameters boundary and thermal response characteristic curves under typical operating conditions for the corresponding vehicle platform.

[0091] Based on Fourier's law of heat conduction and Newton's cooling formula, a set of differential equations constrained model is constructed to describe the thermal coupling relationship between the cabin, evaporator, and battery.

[0092] The continuous heat transfer equation is discretized into a difference inequality for a specific time step, and the threshold range of numerical linkage between semantic variables is established.

[0093] Boolean logic algebra is used to transform the threshold range into executable judgment statements, and conflict flag triggering conditions are defined when the combination of variables deviates from the theoretical range.

[0094] A dynamic weight allocation matrix is ​​established to adaptively adjust the confidence priority of each thermal logic rule based on external disturbance factors such as current vehicle speed and ambient temperature.

[0095] Through the above processing method, the abstract thermodynamic principle is transformed into a rule-based judgment benchmark with thermophysical interpretability, thereby realizing the logical standardization of semantic consistency verification of multi-source data.

[0096] For example, the threshold for cabin air thermal inertia index is set to 0.85 to 1.15, and the threshold for evaporator condensation tendency factor is set to 0.2 to 0.8. When an inertia index of 1.1 and a condensation factor of 0.9 are detected, based on the conflict logic in the rule base where an inertia index > 1.05 and a condensation factor > 0.85, it is determined that the sensor data violates the thermal hysteresis law, and an anomaly marker is generated. This mechanism significantly improves the accuracy of noise identification and effectively suppresses the interference of false signals on the prediction model.

[0097] S3.2: Based on the rule judgment benchmark, perform transient response matching degree calculation on the cabin air thermal inertia index and evaporator condensation tendency factor. By comparing the theoretical change rate of temperature gradient under high thermal inertia state with the actual observed value, identify whether the high frequency oscillation characteristics violate the thermal capacity hysteresis law, so as to generate a preliminary consistency judgment label for the cabin-evaporator coupling channel.

[0098] Extract the cabin air thermal inertia index and evaporator condensation tendency factor sequence within the current time window as the input data object for transient response matching degree calculation.

[0099] Perform a first-order difference operation on the cabin air thermal inertia index sequence, calculate the rate of change between adjacent sampling points, and construct an instantaneous thermal inertia gradient vector characterizing the dynamic evolution of the cabin thermal state.

[0100] Based on the physical law of thermal capacity hysteresis, a theoretical threshold range for the temperature gradient under high thermal inertia is set, which is defined by the cabin equivalent heat capacity parameter and the minimum heat transfer coefficient.

[0101] A sliding window mechanism was used to extract high-frequency oscillation features from the evaporator condensation tendency factor sequence, and the standard deviation and peak frequency of the data within the window were calculated to quantify the intensity of fluctuations in the surface temperature field.

[0102] A transient response mismatch evaluation index is constructed, and the semantic conflict score of the cabin-evaporator coupling channel is calculated using the following formula:

[0103] in, Score semantic conflicts. For instantaneous thermal inertia gradient, As a theoretical reference gradient, The standard deviation of evaporator temperature fluctuation. This is the high-frequency oscillation frequency.

[0104] When the semantic conflict score exceeds the preset confidence threshold of 1.5, it is determined that there is a logical paradox between the high thermal inertia state and the violent temperature oscillation under the current operating conditions, and it is marked as violating the thermal capacity hysteresis law.

[0105] Generate preliminary consistency labels containing exception type codes and conflict intensity values ​​to identify the data credibility status of the cockpit-evaporator coupling channel.

[0106] Through the above chain-like derivation process, multi-source heterogeneous thermophysical semantic variables are transformed into quantified semantic conflict indicators, enabling effective identification and isolation of pseudo-signals caused by sensor noise.

[0107] S3.3: Using the preliminary consistency discrimination label combined with the battery thermal diffusion hindrance coefficient, perform cross-domain heat flow continuity verification. Based on the theoretical heat generation rate derived from the motor torque fluctuation and battery module terminal voltage fluctuation sequence, verify whether the temperature rise delay effect characterized by the battery thermal diffusion hindrance coefficient conforms to the law of energy conservation, so as to output the cross-validation conflict flag between multiple source variables.

[0108] Receive the initial consistency discrimination tag of the cockpit-evaporator coupling channel generated by S3.2, and simultaneously acquire the battery thermal diffusion hindrance coefficient, motor torque fluctuation sequence and battery module terminal voltage fluctuation sequence within the current time window.

[0109] Based on the motor torque fluctuation sequence and battery internal resistance characteristics, the instantaneous theoretical heat generation power inside the battery pack is calculated using Joule's law, and a heat source intensity benchmark is constructed.

[0110] By combining the battery thermal diffusion hindrance coefficient, a first-order thermal response delay model is established, and the theoretical rate of change of battery surface temperature driven by theoretical heat generation power is derived.

[0111] The derived theoretical rate of temperature change is compared with the actual temperature gradient of the battery surface measured by the sensor, and the absolute value of the residual between the two is calculated.

[0112] A dynamic threshold range is set. When the absolute value of the residual exceeds the upper limit of the range, the energy conservation logic is determined to be broken and marked as a cross-domain heat flow continuity conflict.

[0113] If the initial consistency judgment label has indicated an anomaly on the cockpit side, and the current battery side is also marked as conflicting, then a cross-validation conflict label is generated between the multi-source variables, indicating that there is a risk that both the cockpit and battery dual-domain data violate the common sense of thermophysics.

[0114] By using the above-mentioned cross-domain heat flow continuity verification processing method, the local discrimination result of the previous step is transformed into a cross-validation conflict marker between multiple source variables containing energy conservation verification information. This achieves deep physical logic closed-loop verification of the consistency of multi-source sensor data under complex working conditions, significantly improving the robustness of anomaly detection.

[0115] S3.4: For the abnormal combinations that violate thermophysical common sense indicated in the cross-validation conflict markers, initiate the semantic conflict aggregation analysis program to map the local contradiction points scattered in the cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion resistance coefficient into specific sensor channel failure modes, so as to determine the list of abnormal channel identifiers to be frozen for contribution weight.

[0116] The system receives cross-validation conflict markers from multi-source variables output by S3.3 and analyzes the local conflict point data, including cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient. It performs spatiotemporal correlation mapping on the analyzed local conflict points, establishing a correspondence matrix between semantic conflict points and specific hardware channels based on the sensor's physical installation location and heat transfer path topology. Based on this matrix, it calculates the contribution weight of each conflict point in the thermodynamic causal chain, identifying the dominant anomaly source causing the semantic violation. Using a Bayesian inference model and a historical failure mode library, it classifies the dominant anomaly source into failure modes, distinguishing between sensor drift, signal noise, and communication packet loss. Based on the failure mode classification results, it generates a list of anomaly channel identifiers with contribution weights to be frozen, clearly labeling the ID of each anomaly channel and its corresponding confidence degradation parameters. Through this processing, the scattered semantic conflict markers are transformed into specific sensor channel failure modes and a freeze command list, achieving precise localization from logical anomalies to physical channels, providing accurate operational targets for subsequent adaptive inversion compensation.

[0117] S3.5: Based on the abnormal channel identifier list, integrate the data status of all unmarked conflicting trusted channels, encapsulate the hot semantic consistency verification result containing the abnormal channel identifier and the corresponding conflict type description, and use it as a direct input instruction to drive the subsequent adaptive inversion compensation mechanism to construct the narrowband probability density function.

[0118] Step S4: For the abnormal channels marked in the thermal semantic consistency verification results, freeze the contribution weight of the abnormal channels in the current time window and solve the theoretically required reading range in reverse based on the thermal physical semantic state variables of the remaining reliable channels, thereby constructing a narrowband probability density function. Specifically, this includes: S4.1: Based on the abnormal channel identifier in the thermal semantic consistency verification result, the contribution weight of the abnormal channel in the current time window is frozen to generate a set of thermal physical semantic state variables under weight locking state, so as to prevent noise data from interfering with the overall thermal load simulation in subsequent calculations.

[0119] Receive the list of abnormal channel identifiers from the thermal semantic consistency verification result, parse the sensor ID, conflict type and corresponding time window index contained therein, and establish a mapping index table for the channels to be processed.

[0120] Traverse the set of thermophysical semantic state variables within the current time window, retrieve the state variable node that matches the abnormal channel identifier, and locate its specific coordinate position in the multidimensional feature space.

[0121] The contribution weight parameter of the state variable corresponding to the matched abnormal channel is forcibly set to zero, cutting off the signal transmission path of the channel data to the subsequent heat load simulation module.

[0122] Retain the state variables of the reliable channels that are not marked as anomalous and their original weight coefficients, and construct a dimension-reduced weight matrix that contains only effective thermodynamic information.

[0123] Based on the dimensionality-reduced weight matrix and the state variables of the remaining reliable channels, a set of thermophysical semantic state variables under the weight-locked state is reorganized to ensure that the set does not contain noise interference terms.

[0124] By freezing the weights of abnormal channels and reorganizing the set of trusted variables, the verification results of the previous step are transformed into pure thermodynamic state inputs, achieving the expected technical effects of noise isolation and improved computational stability.

[0125] S4.2: Utilize the data of the remaining reliable channels in the set of thermophysical semantic state variables under the weighted locking state, and combine them with the pre-set heat transfer equation constraint relationship to perform inverse solution operation, so as to calculate the theoretical reading range of the abnormal channel under the current operating condition and establish the data boundary that conforms to thermodynamic logic.

[0126] Receive the set of thermophysical semantic state variables under weighted locking state, and extract the reliable channel data that has not been marked as abnormal, including the effective subset of cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient.

[0127] Based on the pre-defined heat transfer equation constraints, an inversion solution model is constructed with known quantities for reliable channels and unknown quantities for abnormal channel theoretical readings, thus establishing thermodynamic equilibrium boundary conditions.

[0128] To address the anomalies in the cabin-evaporator coupling channel, a transient thermal balance equation is established using the law of conservation of energy. The cabin air thermal inertia index is correlated with the rate of change of the evaporator surface temperature gradient, and the theoretical surface temperature range of the evaporator is derived.

[0129] In response to the anomalies in the battery thermal management channel, the theoretical heat generation rate derived from the motor torque fluctuation was combined with the inverse solution of the thermal diffusion equation to determine the theoretical surface temperature of the battery module, and to verify whether the temperature rise delay effect conforms to the law of conservation of energy.

[0130] Through the above reverse calculation, the theoretical reading range of the abnormal channel under the current operating conditions is calculated, forming a data boundary constraint that conforms to thermodynamic logic.

[0131] By using thermodynamic inversion derivation, the reliable state variables from the previous step are transformed into the theoretical reading confidence intervals of the abnormal channel, thereby achieving the expected technical effect of physical and logical correction and boundary establishment of the noise data.

[0132] S4.3: Based on the calculated theoretical reading range, the Gaussian kernel function estimation method is used to perform probability distribution modeling to construct a narrow-band probability density function centered at the theoretical value and with controlled variance, forming a probability distribution model to guide the correction of outlier data.

[0133] The theoretically required reading range of the abnormal channel, received from the S4.2 output, including the center value and boundary thresholds, serves as the initial constraint for probability distribution modeling. The set of thermophysical semantic state variables for the remaining reliable channels is extracted, and the statistical variance of each variable within the current time window is calculated to quantify the instantaneous fluctuation amplitude of the thermal environment state. A Gaussian kernel function is constructed based on the theoretical center value, mapping the half-width of the theoretical reading confidence interval to the standard deviation parameter, thus establishing the dispersion of the probability density function. A kernel density estimation method is used to smooth the theoretically required reading range, eliminating the step effect caused by discrete boundaries and generating a continuously differentiable probability density curve. Based on the inertial characteristics of the thermodynamic process, the bandwidth parameter of the Gaussian kernel function is dynamically adjusted to ensure that the narrowband distribution covers reasonable fluctuations while suppressing extreme outliers. The center value and controlled variance are integrated to instantiate a narrowband probability density function model, forming a mathematical expression describing the true distribution characteristics of the abnormal channel data. Through Gaussian kernel function estimation, the theoretical reading range from the previous step is transformed into a narrowband probability density function with statistical properties, realizing probability distribution modeling for abnormal data correction and providing accurate guidance for subsequent resampling.

[0134] Step S5: The original data from the abnormal channel is resampled and replaced using the narrowband probability density function to eliminate noise amplification and output a multi-source fused data stream corrected for semantic consistency. Specifically, this includes: S5.1: Based on the abnormal channel identifier marked in the thermal semantic consistency verification result, extract the original sensor reading sequence within the corresponding time window as the input object to be corrected, and freeze the contribution weight parameter of the abnormal channel in the current calculation cycle to generate a set of isolated abnormal data streams.

[0135] Receive the list of abnormal channel identifiers from the thermal semantic consistency verification result, parse the list to determine the sensor data channel indexes that need to be isolated within the current calculation cycle.

[0136] Based on the abnormal channel index, the original sensor reading sequence within the corresponding time window is extracted from the time-aligned standardized multi-source data block to construct the original data subset to be corrected.

[0137] Read the weight configuration matrix of each input channel in the preset heat load prediction model and locate the weight parameter position corresponding to the abnormal channel index.

[0138] The located weight parameters are forcibly set to zero or a very small constant value, and a weight freeze operation is performed to cut off the gradient propagation path of abnormal data in subsequent feature extraction layers.

[0139] The state markers after freezing the weights are structurally encapsulated with the extracted subset of original data to generate a set of isolated abnormal data streams containing data isolation identifiers.

[0140] By using weight freezing and data isolation processing, the verification results of the previous step are transformed into physically decoupled abnormal data objects, achieving the expected technical effect of zero interference from noise signals in the heat load extrapolation process.

[0141] For example, in the case where sensor number 3 in the cockpit infrared temperature measurement array is marked as an abnormal channel, the system extracts 10 temperature sampling points of this sensor within the current 200ms time window. The weight value of 0.15 corresponding to channel 3 in the input layer weight matrix of the prediction model is read, modified to 0.0, and locked. The data from these 10 sampling points, along with the weight lock status, is packaged to generate an isolated abnormal data stream set. This ensures that subsequent steps in S5.2 only utilize data from other reliable channels for inversion derivation, completely eliminating the contamination of thermal inertia index calculation by the sensor's high-frequency noise.

[0142] S5.2: Using the set of thermophysical semantic state variables of the remaining reliable channels as boundary constraints, the thermodynamic inversion derivation method is executed to solve the expected reading range of the abnormal channels under the theoretical thermal equilibrium state in reverse order, so as to construct the theoretical reading confidence interval.

[0143] The set of thermophysical semantic state variables under the weighted locking state is received, and the cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient corresponding to the remaining trusted channels are extracted as boundary constraints.

[0144] The system calls upon a pre-defined library of heat transfer equation constraints and matches the corresponding thermodynamic inversion model based on the physical properties of the abnormal channel to determine the functional mapping relationship between the input variables and the output target of the inversion calculation.

[0145] Based on the law of conservation of energy and Fourier's law of heat conduction, a theoretical reading inversion formula for abnormal channels is constructed, and the state variables of the reliable channels are substituted into the formula for numerical solution.

[0146] The following formula is used to calculate the expected center value of the abnormal channel under theoretical thermal equilibrium:

[0147] in, This is the center value of the theoretical reading for the abnormal channel. The specific heat capacity of air, For cabin air quality, The average cabin temperature. The surface temperature of the evaporator. The convective heat transfer coefficient is... This refers to the heat exchange area.

[0148] By combining the statistical characteristics of historical operating data, the variance parameter of theoretical readings is calculated, the uncertainty range of the inversion results is evaluated, and data boundaries that conform to thermodynamic logic are established.

[0149] Based on the central value and variance parameters, a theoretical reading confidence interval is constructed to define the range of numerical distributions of abnormal channels that conform to physical laws within the current time window.

[0150] By using thermodynamic inversion derivation, the weighted locking state of the previous step is transformed into theoretical reading confidence interval data, thereby achieving the expected technical effect of accurately defining the range of abnormal channel data repair.

[0151] S5.3: Based on the center value and variance characteristics of the theoretical reading confidence interval, construct a narrow-band probability density function model with a Gaussian distribution shape, and map the theoretical reading confidence interval to probability distribution parameters to generate a probability density distribution model for data reconstruction.

[0152] Receive the theoretical reading confidence interval output from step S5.2, extract its center value as the mean parameter of the Gaussian distribution, and extract the interval half-width combined with the historical noise variance as the standard deviation parameter.

[0153] A Gaussian probability density function model with the theoretical center value as the expectation is constructed to ensure that the alternative data maintains thermodynamic continuity.

[0154] The upper and lower boundaries of the theoretical reading confidence interval are mapped to the cutoff threshold of the probability density function, eliminating the low-probability tail region that falls outside the confidence interval, thus forming a compact Gaussian distribution.

[0155] The generated Gaussian probability density function is normalized to ensure that the sum of the probability integrals within the cutoff interval is 1, thus guaranteeing the statistical consistency of the resampled data.

[0156] Through the Gaussian distribution modeling process described above, the theoretical reading range of the previous step is transformed into a narrow-band probability density distribution model with clear probability characteristics, thereby achieving precise constraints on the reconstruction space of abnormal channel data and effectively eliminating the noise amplification effect in multi-scale fusion.

[0157] S5.4: Apply the Monte Carlo random sampling method to perform multiple iterative sampling operations on the probability density distribution model, extract alternative numerical samples that conform to thermophysical logic from the probability density distribution model, and generate a resampled alternative data sequence.

[0158] The narrowband probability density function model parameters generated by S5.3 are received, including the center value and variance characteristics of the theoretical reading confidence interval, as the distribution benchmark for Monte Carlo random sampling.

[0159] Set the number of sampling iterations N. Based on the real-time constraints of the vehicle thermal management control cycle, configure N as a fixed threshold that adapts to the processor's computing power to ensure that data reconstruction is completed within milliseconds.

[0160] Initialize the pseudo-random number generator to generate a uniformly distributed random sequence based on the current system timestamp seed, ensuring the statistical independence and reproducibility of each resampling process.

[0161] By employing the inverse transformation sampling method or the accept-rejection sampling method, uniformly distributed random numbers are mapped to a narrow-band Gaussian probability density function space, and N independent sampling operations are performed.

[0162] Physical boundary truncation is performed on the generated N alternative numerical samples to remove outliers that exceed the sensor's range or violate thermodynamic limits, ensuring the physical validity of the sample set.

[0163] Calculate the arithmetic mean or median of the effective sample set as the final resampling replacement value for the abnormal channels within the time window to suppress the uncertainty caused by single random fluctuations.

[0164] The resampled replacement values ​​of each time window are arranged in chronological order to construct a continuous resampled replacement data sequence, filling the gaps in the original anomalous data stream.

[0165] By using Monte Carlo random sampling and statistical aggregation, the probability distribution model from the previous step is transformed into a deterministic alternative data sequence that conforms to thermophysical logic, thereby suppressing noise amplification and restoring data quality.

[0166] S5.5: Replace the resampled alternative data sequence with the corresponding position in the isolated abnormal data stream set, release the weight freeze state of the abnormal channel and re-integrate all channel data to output the multi-source fused data stream after semantic consistency correction.

[0167] The system receives the resampled replacement data sequence and the original data stream of isolated abnormal channels, locating the start and end indices of the abnormal time windows within the overall data frame. A null-filling matrix with the same dimensions as the original data stream is constructed, mapping the resampled replacement data sequence to the corresponding index positions in the filling matrix according to timestamps. A data mask unmasking operation is performed, removing the weight freeze flags of the abnormal channels and restoring their normal contribution rights in the fusion calculation. A multi-source data aggregation engine is invoked to perform tensor concatenation between the filled abnormal channel data and reliable channel data such as the unaffected cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient. Consistency and integrity checks are performed on the concatenated multi-dimensional data to ensure that the alignment accuracy of each channel's data on the time axis meets microsecond-level synchronization requirements. Through the above data replacement and recombination processes, the locally corrected discrete data is transformed into a structurally complete and semantically consistent corrected multi-source fusion data stream, achieving high-fidelity input variable output after noise suppression, providing a clean data foundation for subsequent heat load prediction models.

[0168] Step S6: Input the multi-source fused data stream into the time-series prediction model, perform net heat load power extrapolation calculation, and directly output a real-time heat load prediction value adapted to the control requirements of the automotive air conditioning system. Specifically, this includes: S6.1: Obtain the cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient from the multi-source fusion data stream after semantic consistency correction. Perform time-series alignment and normalization processing on the set of thermophysical semantic state variables based on the sliding time window mechanism to generate a standardized time-series feature vector sequence.

[0169] The system receives a multi-source fused data stream after semantic consistency correction and extracts the cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient as basic input variables.

[0170] Based on a preset sliding time window mechanism, a time-series buffer queue of length N is constructed. The current time t and the thermophysical semantic state variables of the previous N-1 sampling periods are stored in the queue in sequence to form a multi-dimensional time-series data matrix.

[0171] Zero-mean standardization is performed on the multidimensional time series data matrix to calculate the mean and standard deviation of each semantic variable within the time window and normalize their dimensions.

[0172] For the normalized sequence data, a linear interpolation method is used to align the tiny timestamp discrepancies caused by transmission delays from different sensors, ensuring that the data from the cabin, evaporator, and battery domains are strictly synchronized on the same time grid.

[0173] The aligned multivariate sequences are spliced ​​along the channel dimension to construct a high-dimensional tensor structure containing the historical thermal state evolution trajectory, which serves as the standardized input for the time series prediction model.

[0174] Through the above-mentioned temporal alignment and normalization processing, discrete thermophysical semantic state variables with different dimensions are transformed into a standard temporal feature vector sequence that can be directly read by the model, eliminating the influence of data scale differences on gradient convergence and realizing the structuring and standardization of the input data of the prediction model.

[0175] S6.2: Receive the standardized time-series feature vector sequence, and use the gated recurrent unit network in the preset lightweight time-series prediction model to perform deep time-series dependency mining processing to extract hidden state feature representations containing historical heat accumulation effects and transient disturbance trends.

[0176] The system receives a standardized temporal feature vector sequence output from S6.1. This sequence contains time-aligned data for cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient. The temporal feature vector sequence is then input into a pre-defined lightweight gated recurrent unit (GRU) network input layer, where a linear transformation is performed to match the hidden layer dimension. An update gate mechanism is used to calculate the correlation weights between the current input features and the previous hidden state.

[0177] Based on the update gate output, the reset gate mechanism is used to filter redundant information in the historical heat accumulation effect, suppress noise interference in long-term dependence, and calculate the reset gate control signal.

[0178] By combining the reset gate signal with the current input, the candidate hidden state is calculated. This state characterizes the transient thermal disturbance trend and short-term thermal response characteristics after filtering.

[0179] The hidden state of the previous time step and the candidate hidden state of the current time step are weighted and fused using the update gate weights to generate the final hidden state feature representation of the current time step.

[0180] By iterating through all time steps within the sliding time window, the above gating operation is performed to construct a high-dimensional hidden layer state sequence containing the complete thermal history evolution trajectory.

[0181] By using deep temporal dependency mining of the GRU network, the standardized features from the previous step are transformed into hidden state features that contain dynamic thermophysical laws, thus achieving accurate capture of heat load change trends under complex operating conditions and enhanced noise robustness.

[0182] S6.3: Based on the hidden layer state feature representation, a nonlinear transformation process is performed through a fully connected regression mapping layer to map the high-dimensional hidden layer state feature representation into an estimated instantaneous heat power demand with clear physical dimensions, so as to form a preliminary net heat load power projection result.

[0183] The hidden layer state feature representation, which includes historical heat accumulation effect and transient perturbation trend, received from the output of S6.2, is used as the input vector of the fully connected regression mapping layer.

[0184] The hidden layer state feature representation is subjected to dimensionality reshaping processing, which flattens it into a one-dimensional feature tensor to adapt to the matrix operation requirements of the fully connected layer.

[0185] A fully connected network architecture with three layers of hidden neurons is constructed. The first hidden layer maps the high-dimensional hidden layer features to the intermediate semantic space through linear transformation. The activation function is a modified linear unit to introduce non-linear expressive power.

[0186] The second hidden layer further compresses the feature dimensions, extracts the core thermodynamic factors that are strongly correlated with net heat load power, and eliminates redundant information interference.

[0187] The output layer adopts a single-node linear regression structure, which maps the compressed feature vector into a scalar form of instantaneous heat power demand estimate.

[0188] During the mapping process, a physical constraint loss term is introduced to ensure that the dimensions of the output value are consistent with the units of cooling / heating power of the automotive air conditioning system.

[0189] Through nonlinear transformation processing of the fully connected regression mapping layer, the high-dimensional hidden layer state feature representation of the previous step is transformed into an instantaneous heat power demand estimate with clear physical dimensions, forming a preliminary net heat load power extrapolation result, and realizing a precise mapping from abstract spatiotemporal features to specific control indicators.

[0190] S6.4: Based on the preliminary net heat load power projection results, dynamic boundary constraint verification processing is performed in conjunction with the current air conditioning damper opening control command and the compressor intake and exhaust pressure difference sequence to eliminate abnormal projection values ​​that exceed the physical limits of the vehicle thermal management system and generate corrected net heat load power projection results.

[0191] The system receives the preliminary net heat load power projection results from S6.3 and uses them as the initial heat power estimate to be verified. Simultaneously, it acquires the current air conditioning system damper opening control command sequence and parses out the specific opening percentage values ​​of the main air duct, foot air duct, and defrost air duct. It reads the compressor suction and discharge pressure difference sequence and extracts the average pressure difference value and pressure difference fluctuation standard deviation within the current time window. Based on the compressor performance curves in the vehicle thermal management system engineering manual, it constructs a maximum cooling power boundary constraint model.

[0192] Based on the damper opening command, the effective flow area ratio of each air duct is calculated, and the theoretical maximum cooling power is corrected to obtain the upper limit of the actual achievable heat load regulation of the system. The preliminary net heat load power projection result is compared with the upper limit of the actual achievable heat load regulation of the system. If the projection result exceeds the regulation limit, it is judged as an abnormal projection value that is physically unreachable. For abnormal projection values, saturation truncation is performed to forcibly limit them within the regulation limit range. If the projection result is lower than the minimum sustaining power threshold of the system, it is judged as an underdriven anomaly, and it is raised to the minimum sustaining power threshold of 0.5kW. For normal projection values ​​between the upper and lower limits, their original values ​​are retained unchanged. The power values ​​after boundary constraint verification are integrated to generate the corrected net heat load power projection result.

[0193] By using dynamic boundary constraint verification, the preliminary deduction results from the previous step are transformed into corrected net heat load power deduction results that conform to the physical limits of the vehicle thermal management system, thereby achieving the expected technical effect of eliminating non-physical prediction biases caused by model overfitting or sensor transient noise.

[0194] S6.5: Output the corrected net heat load power projection result as a real-time heat load prediction value to adapt to the control requirements of the automotive air conditioning system, and send the real-time heat load prediction value to the vehicle thermal management controller to drive the air conditioning actuator to operate.

[0195] The system receives the corrected net heat load power projection result from S6.4 and encapsulates it into a standard data frame format conforming to the vehicle controller communication protocol. Based on the automotive air conditioning system control cycle requirements, the net heat load power projection result is timestamped to ensure strict synchronization between the predicted value and the current vehicle operating status. The timestamped net heat load power data is sent to the vehicle thermal management controller domain via the vehicle CAN bus or Ethernet communication interface. The vehicle thermal management controller parses the received data frame and extracts the real-time heat load prediction value as the core parameter of the feedforward control command. Based on the real-time heat load prediction value, the controller calculates the target speed command using the compressor performance mapping table and simultaneously calculates the target opening command of the electronic expansion valve based on the evaporator heat exchange efficiency model. The generated target speed command and target opening command are distributed to the air conditioning compressor driver and expansion valve actuator, driving the actuators to quickly respond to changes in the cabin heat load. Through the aforementioned data encapsulation, synchronization marking, communication transmission, and actuator-driven processing methods, the revised deduction results from the previous step are transformed into specific air conditioning actuator control commands, enabling the automotive air conditioning system to respond accurately and quickly to the dynamic thermal environment, significantly improving the thermal comfort of the passenger compartment and the system's energy efficiency ratio.

[0196] Step S7: Based on the deviation between the real-time predicted heat load value and the actual operation feedback of the vehicle thermal management system, dynamically update the engineering knowledge accumulation parameters in the thermal logic rule base to complete the online iterative optimization of the thermophysical semantic consistency constraints. Specifically, this includes: S7.1: Obtain the deviation sequence formed by the real-time heat load prediction value and the actual operation feedback of the vehicle thermal management system, and use the sliding window statistical method to perform time series feature extraction processing on the deviation sequence to generate a deviation feature vector containing the mean drift and variance volatility.

[0197] The system receives the real-time heat load prediction sequence output by S6 and simultaneously collects the actual cooling / heating power data and measured cabin temperature data fed back by the actuators of the vehicle's thermal management system. It calculates the instantaneous deviation between the predicted and measured values, constructs the original deviation sequence on a continuous time axis, and uses this as the basic data source characterizing the model's prediction accuracy.

[0198] By setting a fixed-length sliding time window, the original deviation sequence is divided into multiple overlapping data subsets to capture the local statistical characteristics of deviation changes over time and eliminate the interference of noise at a single moment on the overall trend judgment.

[0199] The mean of the deviation data within each time window is calculated to obtain the average deviation level of that window period, which reflects the degree of systematic shift of the prediction model in the short term, i.e., the mean drift.

[0200] Variance calculation is performed on the deviation data within the same window to quantify the dispersion of the deviation data and reflect the stability of the fluctuation of the prediction result relative to the true value, i.e., variance volatility.

[0201] The mean drift and variance volatility are combined to form a two-dimensional bias feature vector, which fully describes the central trend and distribution breadth of the prediction error in the current time period.

[0202] By using the sliding window statistical method, the discrete instantaneous deviation sequence is transformed into a deviation feature vector containing mean drift and variance volatility, thereby realizing a quantitative characterization of the short-term performance fluctuation of the prediction model and providing a stable input basis for subsequent reverse attribution analysis.

[0203] S7.2: Based on the aforementioned deviation feature vector, the thermal sensitivity mapping matrix is ​​invoked to perform reverse attribution analysis, mapping the deviation feature vector into semantic variable confidence correction coefficients for cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient.

[0204] Receive the deviation feature vector generated by S7.1, which includes mean drift and variance volatility, as the input benchmark for inverse attribution analysis.

[0205] A thermal sensitivity mapping matrix is ​​constructed, which characterizes the sensitivity relationship between the partial derivatives of the cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient and the net heat load prediction value, reflecting the contribution weight of small perturbations of each semantic variable to the final prediction result.

[0206] Perform matrix multiplication on the bias feature vector and the thermal sensitivity mapping matrix to quantify the responsibility of each semantic variable in generating the current prediction bias and generate the initial attribution weight vector.

[0207] The initial attribution weight vector is normalized to eliminate the influence of dimensional differences, ensuring that the sum of the contributions of each variable is 1, thus forming a standardized relative sensitivity distribution.

[0208] By combining the variance volatility in the bias feature vector, an uncertainty penalty factor is introduced to attenuate the confidence of semantic variables corresponding to high volatility, thereby suppressing noise-driven spurious attributions.

[0209] For the cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient, respectively, the confidence correction coefficients of the three dimensions are calculated independently by substituting their respective attribution weight deviation values ​​and variance volatility.

[0210] Through the aforementioned reverse attribution analysis and nonlinear mapping processing, the macroscopic prediction deviation from the previous step is transformed into a microscopic confidence correction coefficient for specific thermophysical semantic variables, thereby achieving accurate location and quantitative evaluation of the prediction error source and providing a precise basis for the dynamic adjustment of the rule base threshold in the future.

[0211] S7.3: Using the confidence correction coefficient of the semantic variable, perform adaptive relaxation or tightening adjustment on the threshold of the constraint relationship of the heat transfer equation stored in the preset thermal logic rule base to generate updated thermophysical semantic consistency constraints with dynamic adaptability.

[0212] The semantic variable confidence correction coefficients for cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient output by S7.2 are received as the input basis for threshold adjustment.

[0213] The threshold values ​​of the initial heat transfer equation constraint relationship corresponding to each semantic variable are read from the pre-set thermal logic rule base, including the upper limit of temperature change rate, the tolerance range of pressure difference fluctuation, and the baseline value of thermal diffusion delay time window.

[0214] A nonlinear mapping function based on confidence correction coefficients is constructed, and the correction coefficients are transformed into threshold scaling factors to ensure that high-confidence variables correspond to tighter constraint boundaries, while low-confidence variables correspond to moderately relaxed constraint boundaries.

[0215] The updated constraint thresholds are calculated using the following formula:

[0216] in, For the updated constraint threshold, As the initial baseline threshold, To adjust the sensitivity coefficient, The confidence adjustment coefficient for semantic variables.

[0217] Adaptive adjustment is performed on the threshold of the rate of change of temperature gradient corresponding to the cabin air thermal inertia index. When the confidence correction coefficient indicates that the variable has a risk of drift, the upper limit of the threshold is increased to accommodate normal transient thermal response fluctuations and avoid misjudgment as semantic conflict.

[0218] Dynamic tightening is applied to the surface temperature and dew point temperature difference threshold corresponding to the evaporator condensation tendency factor. When the confidence correction coefficient indicates that the data quality of this variable is extremely high, the allowable deviation range is reduced to enhance the sensitivity detection of minor condensation risks.

[0219] Linear interpolation correction is performed on the temperature rise delay time window threshold corresponding to the battery thermal diffusion hindrance coefficient. Based on the confidence correction coefficient, a new effective range is determined between the minimum delay lower limit and the maximum delay upper limit to adapt to the thermal characteristic changes caused by battery aging.

[0220] The calculated threshold values ​​of each variable are re-encapsulated into the data structure of the thermal logic rule base, and the version iteration identifier is marked to form updated thermal physical semantic consistency constraints with dynamic adaptability.

[0221] Through the above adaptive relaxation or tightening adjustment process, the confidence correction coefficient of the previous step is transformed into a dynamically optimized threshold parameter, realizing the real-time tracking and precise constraint of the thermal logic rule base on the changes in vehicle thermal state, and significantly improving the robustness of semantic consistency verification.

[0222] S7.4: Based on the updated thermophysical semantic consistency constraints, reconstruct the set of engineering knowledge sedimentation parameters in the thermal logic rule base, replace the old parameters through a version overwrite mechanism, so as to complete the online iterative optimization of the thermophysical semantic consistency constraints and output the final updated thermal logic rule base.

[0223] Receive the updated thermophysical semantic consistency constraints generated by S7.3, which have dynamic adaptability and include the adjusted upper limit of temperature change rate, the tolerance range of pressure fluctuation, and the threshold of thermal diffusion delay time window.

[0224] The updated constraint data structure is analyzed to extract the latest threshold parameter set corresponding to the cabin air thermal inertia index, evaporator condensation tendency factor, and battery thermal diffusion hindrance coefficient.

[0225] Read the current version parameter set in the preset hot logic rule base and identify the storage address index and version number identifier of each semantic variable in the rule base.

[0226] Perform an atomic write operation to overwrite the latest threshold parameter set extracted into the storage address corresponding to the rule base, ensuring the integrity and consistency of the data update process.

[0227] Generate a new version iteration identifier, which consists of a timestamp sequence number and a checksum, and is used to mark the uniqueness and validity of this parameter update.

[0228] The new version identifier is associated with the updated parameter set to form a complete thermal logic rule base engineering knowledge accumulation parameter object.

[0229] The broadcast rule base update completion signal notifies the thermal semantic consistency verification module to load the latest version of parameters, ensuring that subsequent verification processes are executed based on the latest thermal physical constraint boundaries.

[0230] By replacing old parameters through a version overwrite mechanism, the dynamically optimized threshold parameters are solidified into the baseline rules for system operation, enabling online iterative optimization of thermophysical semantic consistency constraints and outputting the final updated thermological rule library.

[0231] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0232] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0233] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion, characterized in that, Specifically, it includes: S1: Obtain the original multi-scale thermal environment dataset during the initial stage of vehicle startup or under steady-state parking conditions; S2: Based on the original dataset of the multi-scale thermal environment, a lightweight thermal state encoder trained offline is used to perform mapping transformation processing to transform the original sensor readings into a set of thermal physical semantic state variables. S3: Call the preset thermal logic rule base to perform a consistency check operation on the set of thermal physical semantic state variables, identify semantic conflict markers that violate thermal physical common sense by comparing the heat transfer equation constraint relationship between variables, and generate a thermal semantic consistency check result containing abnormal channel identifiers. S4: For the abnormal channels marked in the thermal semantic consistency verification results, freeze the contribution weight of the abnormal channels in the current time window and solve the theoretically required reading range in reverse according to the thermal physical semantic state variables of the remaining reliable channels, thereby constructing a narrowband probability density function. S5: The original data of the abnormal channel is resampled and replaced using the narrowband probability density function to eliminate the noise amplification effect and output the multi-source fused data stream after semantic consistency correction; S6: Input the multi-source fusion data stream into the time-series prediction model, perform net heat load power extrapolation calculation, and directly output the real-time heat load prediction value that adapts to the control requirements of the automotive air conditioning system.

2. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 1, characterized in that, S6 is followed by: S7: Based on the deviation between the real-time heat load prediction value and the actual operation feedback of the vehicle thermal management system, dynamically update the engineering knowledge accumulation parameters in the thermal logic rule base to complete the online iterative optimization of the thermophysical semantic consistency constraint conditions.

3. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 1, characterized in that, The multi-source heterogeneous raw data stream includes the cabin infrared temperature field distribution sequence, the evaporator surface temperature gradient sequence, the compressor suction and exhaust pressure difference sequence, the battery module terminal voltage fluctuation sequence, and the CAN bus vehicle speed and air conditioning damper opening control commands.

4. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 1, characterized in that, The set of thermophysical semantic state variables includes the cabin air thermal inertia index, the evaporator condensation tendency factor, and the battery thermal diffusion hindrance coefficient.

5. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 1, characterized in that, The time-series prediction model is a gated recurrent unit network.

6. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 4, characterized in that, S3 specifically includes: The cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient are obtained from the set of thermophysical semantic state variables. Based on the engineering knowledge of the whole vehicle thermal management system, a thermal logic rule base is constructed, and the constraint relationship of the heat transfer equation is transformed into executable Boolean judgment logic and numerical threshold range to form a rule judgment benchmark with thermophysical interpretability. Based on the aforementioned rule-based judgment criteria, a transient response matching degree calculation is performed on the cabin air thermal inertia index and the evaporator condensation tendency factor to generate a preliminary consistency discrimination label for the cabin-evaporator coupling channel. Using the preliminary consistency discrimination label in combination with the battery thermal diffusion hindrance coefficient, cross-domain heat flow continuity verification is performed. Based on the theoretical heat generation rate derived from the motor torque fluctuation and battery module terminal voltage fluctuation sequence, the temperature rise delay effect characterized by the battery thermal diffusion hindrance coefficient is verified to conform to the law of energy conservation, so as to output the cross-validation conflict flag between multiple source variables. For the abnormal combinations that violate thermophysical common sense indicated in the cross-validation conflict markers, a semantic conflict aggregation analysis program is initiated to map the local contradictions scattered in the cabin air thermal inertia index, evaporator condensation tendency factor and battery thermal diffusion hindrance coefficient into specific sensor channel failure modes, so as to determine the list of abnormal channel identifiers to be frozen for contribution weight. Based on the abnormal channel identifier list, all unmarked conflicting trusted channel data states are integrated and encapsulated into a hot semantic consistency verification result.

7. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 6, characterized in that, The transient response matching degree calculation of the cabin air thermal inertia index and evaporator condensation tendency factor based on the rule judgment benchmark is specifically performed by comparing the theoretical rate of change of temperature gradient under high thermal inertia state with the actual observed value to identify whether the high-frequency oscillation characteristics violate the thermal capacity hysteresis law.

8. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 6, characterized in that, The hot semantic consistency verification result includes an abnormal channel identifier and a description of the corresponding conflict type.

9. The method for predicting automotive air conditioning heat load based on multi-source thermal environment data fusion according to claim 1, characterized in that, S4 specifically includes: Based on the abnormal channel identifier in the thermal semantic consistency verification result, the contribution weight of the abnormal channel in the current time window is frozen to generate a set of thermal physical semantic state variables under weight locking state. By utilizing the data of the remaining reliable channels in the set of thermophysical semantic state variables under the weight-locked state, and combining the pre-set heat transfer equation constraint relationship, a reverse solution operation is performed to calculate the theoretical reading range of the abnormal channel under the current operating condition, and to establish the data boundary that conforms to thermodynamic logic. Based on the calculated theoretical reading range, the Gaussian kernel function estimation method is used to perform probability distribution modeling to construct a narrow-band probability density function centered at the theoretical value and with controlled variance.

Images