Method for monitoring abnormal state of automobile air conditioner controller based on multi-modal data fusion

Abstract

The invention relates to the technical field of automobile electronic intelligent fault diagnosis, and discloses a multi-modal data fusion automobile air conditioner controller abnormal state monitoring method, which comprises the following steps: deploying five types of intelligent sensor nodes at key positions of an automobile air conditioner system to realize hardware-level time synchronization and adaptive sampling; a fast ICA algorithm is combined with sliding window processing to realize signal separation, a self-adaptive wavelet basis selection mechanism is designed to carry out time-frequency analysis, and principal component analysis is used to carry out feature dimension reduction; constructing a lightweight attention fusion network, dynamically adjusting different modal feature weights, and mining feature relevance through a cross-modal interaction module; a hierarchical progressive anomaly detection architecture is designed, a statistical method and a neural network detection result are fused, a time consistency constraint mechanism is established, and multi-time scale anomaly prediction is realized; and fault type identification and severity evaluation are carried out in combination with an expert knowledge base, and a multi-priority protection strategy and an emergency plan are formulated.

Description

Technical Field

[0001] This invention relates to the field of automotive electronic intelligent fault diagnosis technology, and more specifically, to a method for monitoring abnormal states of automotive air conditioning controllers using multimodal data fusion. Background Technology

[0002] As a crucial component of modern automobiles, the reliability of automotive air conditioning systems directly impacts the comfort and safety of passengers. With the increasing intelligence and electrification of automobiles, the complexity of air conditioning controllers is constantly growing, making traditional fault diagnosis methods insufficient to meet the high reliability requirements of modern vehicles.

[0003] Traditional automotive air conditioning fault diagnosis relies primarily on human experience and simple sensor monitoring, using a single temperature or pressure sensor to determine the status and employing fixed thresholds to detect anomalies. This method has significant limitations: firstly, the detection accuracy is low, as a single sensor cannot comprehensively reflect the system status; secondly, the early warning capability is insufficient, only detecting faults when they occur and unable to provide advance warnings; and thirdly, it has poor environmental adaptability, easily generating false alarms under complex operating conditions.

[0004] In recent years, with the development of artificial intelligence and big data technologies, equipment monitoring technologies based on multi-sensor fusion have gradually emerged. However, existing technologies are mainly geared towards industrial equipment or large air conditioning systems, and do not adequately consider the special characteristics of the vehicle environment. Vehicle air conditioning systems are characterized by compact space, complex operating conditions, and variable environments. Especially in extreme application scenarios, existing technologies suffer from technical problems such as signal coupling, feature aliasing, and insufficient real-time performance, making it urgent to develop intelligent monitoring technologies specifically suitable for vehicle environments. Summary of the Invention

[0005] This invention provides a method for monitoring abnormal states of automotive air conditioning controllers using multimodal data fusion, which solves technical problems such as signal coupling, feature aliasing, and insufficient real-time performance in related technologies.

[0006] This invention provides a method for monitoring abnormal states of automotive air conditioning controllers using multimodal data fusion, comprising the following steps: Temperature sensors, pressure sensors, current sensors, vibration sensors, and sound sensors are deployed at key locations in the automotive air conditioning system, and time synchronization is achieved through the vehicle's CAN bus to obtain multimodal synchronized data streams; Data validity checks and outlier removal are performed on the multimodal synchronous data stream. The fast ICA algorithm combined with sliding window processing is used to achieve signal separation. An adaptive wavelet basis selection mechanism is designed for time-frequency analysis. Principal component analysis is used for feature dimensionality reduction, and the dimensionality-reduced feature matrix is ​​output. The dimensionality-reduced feature matrix is ​​standardized and preprocessed, depthwise separable convolution is used, a lightweight attention mechanism is designed to adjust the feature weights of different modalities, the intermodal correlation is calculated through a cross-modal interaction module, and the fused feature vector is output. Based on fused feature vectors, a two-layer progressive detection architecture is adopted. The first layer uses statistical methods for coarse screening, and the second layer uses a lightweight neural network for precise detection. The detection results are fused and calculated to obtain an anomaly score and output an anomaly probability sequence. Based on the identification of fault types using abnormal probability sequences, an expert rule knowledge base is constructed, and data-driven machine learning results are integrated with knowledge-driven expert rules to generate emergency plans according to the fault type and severity.

[0007] In a preferred embodiment, the temperature sensor is deployed at the compressor inlet and outlet, the pressure sensor is installed in the high and low pressure pipelines, the current sensor is connected to the controller power line, the vibration sensor is fixed to the compressor body, the sound sensor is placed at the air outlet, the intelligent sensor node has a built-in microprocessor, the local data cache unit has a capacity of 512KB, the vehicle CAN bus communication rate is 500Kbps, and the time synchronization accuracy is controlled within one millisecond.

[0008] In a preferred embodiment, the data validity check includes data integrity verification, timestamp consistency check, and numerical range reasonableness judgment. Outlier removal adopts the 3σ criterion and quartile method. The fast ICA signal separation adopts a sliding window processing strategy with a window length of one thousand data points and an overlap rate of fifty percent. Blind source separation is achieved by calculating the statistical independence and non-Gaussianity of the signal. The algorithm delay is controlled within one second.

[0009] In a preferred embodiment, the adaptive wavelet basis selection mechanism pre-sets a wavelet basis function library including Daubechies wavelet, Morlet wavelet, Mexican hat wavelet, and Biorthogonal wavelet. It automatically selects the optimal wavelet basis based on the spectral characteristics of the signal and the time-frequency localization requirements. The wavelet decomposition level is set to 5 levels, the reconstruction error threshold is set to 0.01, and the principal component analysis method achieves feature dimensionality reduction by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. The cumulative contribution rate threshold is set to 95%, and the dimensionality after feature dimensionality reduction is controlled within 20% of the original dimensionality.

[0010] In a preferred embodiment, the lightweight attention mechanism groups the input features according to modality type, with each group containing ten to twenty features. The attention weights are calculated using the softmax function. Specifically, the inner product of the attention weight vector and the mean vector of each feature group is used as the input of the exponential function. Then, the exponential values ​​of all feature groups are normalized to ensure that the sum of all weights is one.

[0011] In a preferred embodiment, the anomaly score fusion calculation is obtained by weighted summation of statistical detection score, neural network detection score and time consistency score. Before fusion calculation, different types of scores are standardized preprocessed. Statistical detection score is standardized using Z-score, neural network detection score is normalized to a probability value between zero and one using the softmax function, and time consistency score is normalized to a zero-to-one interval using maximum-minimum normalization.

[0012] In a preferred embodiment, the first layer of the two-layer progressive detection architecture employs Mahalanobi distance detection and a sliding window-based statistical control chart method. The Mahalanobi distance is obtained by calculating the difference between the current feature vector and the mean vector of historical normal samples, and then taking the square root of the quadratic form after transforming the inverse of the covariance matrix. The second layer of the lightweight neural network employs a fully connected neural network with a three-layer structure. The number of nodes in the input layer is equal to the feature dimension, the number of nodes in the hidden layer is twenty to thirty, and the number of nodes in the output layer is two. The smoothing window length of the time consistency constraint mechanism is 10 time steps.

[0013] In a preferred embodiment, the expert rule knowledge base includes three sub-bases: a fault mode base, a diagnostic rule base, and a decision tree base. Fault type identification uses a support vector machine classifier to classify faults into five categories: compressor faults, condenser abnormalities, electronic control module faults, sensor failures, and system leaks. Among them, compressor fault characteristic modes include abnormally increased current, vibration frequency shift, sound spectrum changes, and slow temperature rise. Data-driven and knowledge-driven fusion adopt a weighted average method. The initial value of the fusion weight of data-driven fusion is set to 0.7, and the fusion weight of knowledge-driven fusion is one minus the fusion weight of data-driven fusion.

[0014] In a preferred embodiment, the emergency plan is generated into three levels based on the severity of the fault: minor fault, moderate fault, and severe fault. The minor fault score is 0 to 0.3, indicating that the system can continue to operate but requires attention; the moderate fault score is 0.3 to 0.7, indicating that preventive measures need to be taken to prepare for maintenance; and the severe fault score is 0.7 to 1, indicating that the system needs to be shut down for maintenance or the emergency plan needs to be activated immediately. The fault severity score is calculated by weighted summation of the maximum abnormal probability, the normalized value of the abnormal duration, and the growth trend of the abnormal probability.

[0015] In a preferred embodiment, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of the above-described multimodal data fusion method for monitoring abnormal states of an automotive air conditioning controller.

[0016] The beneficial effects of this invention are as follows: The accuracy and timeliness of anomaly detection have been improved. By constructing a five-modal sensor network and a lightweight signal processing algorithm, comprehensive monitoring of the air conditioner controller's operating status has been achieved. A fast ICA algorithm effectively separates environmental interference from system signals, an adaptive wavelet basis selection mechanism improves feature extraction accuracy, and an attention fusion network dynamically optimizes the weights of multimodal information. The hierarchical, progressive anomaly detection architecture balances real-time performance and accuracy requirements, while a time consistency constraint mechanism avoids drastic fluctuations in detection results. Intelligent monitoring and autonomous decision-making in extreme environments have been achieved. For extreme applications (such as extreme cold), a multi-coupling interference separation algorithm and an environmental adaptive mechanism were designed to effectively address the impact of environmental factors such as low temperature, electromagnetic interference, and air pressure changes on monitoring accuracy. The knowledge-guided decision optimization module integrates data-driven and experience-driven methods to construct a complete fault type identification and severity assessment system. A multi-priority protection strategy ensures that environmental conditions in critical areas are prioritized, and the emergency plan generation function provides detailed operational guidance for on-site maintenance. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for monitoring abnormal states of an automotive air conditioning controller based on multimodal data fusion, according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion, such as... Figure 1 As shown, it includes the following steps: S100 deploys temperature sensors, pressure sensors, current sensors, vibration sensors, and sound sensors at key locations in the automotive air conditioning system, and achieves time synchronization through the vehicle's CAN bus to obtain multimodal synchronized data streams; Input data source; The input data for this step comes from multiple interfaces of the vehicle system; the vehicle CAN bus provides basic vehicle operating parameters, including basic information such as vehicle speed, engine speed, and voltage; the GPS module provides a high-precision time reference signal to ensure system time synchronization; and the environmental monitoring unit provides environmental parameters such as external temperature, humidity, and air pressure.

[0020] Deployment of intelligent sensor nodes; Five types of intelligent sensor nodes are deployed at key locations in the automotive air conditioning system; high-precision temperature sensors are installed at the compressor inlet and outlet to measure refrigerant temperature changes, with a sensor accuracy of ±0.1 degrees Celsius; pressure sensors are installed in the high and low pressure pipelines of the system to monitor refrigerant pressure status, with a measurement range of 0 to 3 MPa; current sensors are installed in the power lines of the air conditioning controller to monitor changes in controller power consumption, with a sampling accuracy at the milliampere level; triaxial vibration acceleration sensors are installed on the compressor and blower bodies to collect mechanical vibration signals, with a frequency response range of 0 to 5 kHz; and sound sensors are installed at the air conditioning vents to capture the sound characteristics of the system operation, with a sound frequency range of 20 Hz to 20 kHz.

[0021] Hardware-level time synchronization mechanism; Each smart sensor node integrates an ARM Cortex-M4 microprocessor and a local data cache unit, achieving hardware-level time synchronization via the vehicle's CAN bus. The specific implementation process is as follows: the main controller periodically broadcasts a time synchronization signal to all sensor nodes, containing a GPS standard timestamp and a synchronization sequence number; upon receiving the synchronization signal, each sensor node records its local timestamp and calculates the time deviation; based on the time deviation, each node adjusts its local clock frequency to achieve a time synchronization accuracy of less than one millisecond.

[0022] Time synchronization calibration is achieved by adding a fixed time offset to the local timestamp, plus the product of a clock drift correction factor and the clock drift amount. The clock drift correction factor is determined statistically based on historical synchronization data and typically ranges from 0.8 to 1.2.

[0023] Adaptive sampling frequency adjustment; To balance data quality and computational resource consumption, an adaptive sampling frequency adjustment mechanism is designed. The system dynamically adjusts the sampling frequency based on the signal rate of change, increasing the sampling frequency when the signal changes drastically and decreasing the sampling frequency when the signal is stable.

[0024] The adaptive sampling frequency is calculated by multiplying the base sampling frequency by one and adding the product of the adaptive coefficient and the signal rate of change. The base sampling frequency is set to 100 Hz, the adaptive coefficient ranges from 0.1 to 0.5, and the signal rate of change is calculated by the difference between adjacent sampling points. When the signal rate of change exceeds a preset threshold, the sampling frequency can reach a maximum of 500 Hz; when the signal is stable, the sampling frequency can be reduced to 50 Hz.

[0025] Output data format; This step outputs a multimodal data stream with a time synchronization accuracy of less than one millisecond. The data format includes four fields: timestamp, sensor identifier, data value, and quality identifier. The timestamp is represented by a 64-bit integer with microsecond precision; the sensor identifier uses an 8-bit code to distinguish different sensor types and installation locations; the data value is represented by a 16-bit or 32-bit floating-point number depending on the sensor type; and the quality identifier contains information such as data validity, signal strength, and noise level, providing a reference for subsequent data processing.

[0026] Optionally, in some embodiments, the more advanced IEEE 1588 precision time protocol is used to replace CAN bus time synchronization, which can further improve the time synchronization accuracy to the nanosecond level, making it particularly suitable for scientific research applications with extremely high time accuracy requirements. Similarly, in some embodiments, an adaptive sampling algorithm based on machine learning is used to train a prediction model through historical data, predict signal change trends in advance, and achieve predictive sampling frequency adjustment. Compared with the traditional responsive adjustment method, the sampling strategy can be adjusted hundreds of milliseconds in advance, but more computing resources are required.

[0027] S200 performs data validity checks and outlier removal on multimodal synchronous data streams, uses a fast ICA algorithm combined with sliding window processing to achieve signal separation, designs an adaptive wavelet basis selection mechanism for time-frequency analysis, and uses principal component analysis to perform feature dimensionality reduction, outputting a dimensionality-reduced feature matrix. Data reception and processing; Receive synchronous data stream Data validity checks and outlier removal are performed. For each sensor channel, statistical characteristics of continuous data points are calculated, including mean, variance, and rate of change. Data points deviating from the statistical characteristics by more than three standard deviations are marked as outliers and interpolated for correction. Cubic spline interpolation is used to ensure data continuity and smoothness.

[0028] Data preprocessing is performed, including converting temperature data from Celsius to Kelvin to ensure all temperature values ​​are positive for subsequent processing; unifying pressure data from megapascals to pascals; and converting current data from milliamperes to amperes. All numerical data are normalized, mapping the numerical range to the interval between zero and one. Specifically, the minimum value for each data type is subtracted from the original value, and then divided by the difference between the maximum and minimum values. For categorized fault status identifiers, one-hot encoding is used to convert them into numerical vectors, with each fault type corresponding to one dimension. In normal conditions, all dimensions are zero, and in abnormal conditions, the corresponding dimension is one.

[0029] Fast ICA signal separation algorithm; To address the problem of multiple coupling interference in extreme environments, a fast ICA algorithm combined with sliding window processing is employed to achieve near real-time signal separation. The mixed signal matrix is ​​represented as follows: ,in Represents the observed signal matrix, Represents a mixture matrix. This represents the source signal matrix. The fast ICA algorithm estimates the separation matrix by maximizing the non-Gaussianity. , making Approximation source signal .

[0030] In the specific implementation process, the observed signal is centered to eliminate the DC component, that is, the average value of all data points is subtracted from each data point; then whitening is performed to eliminate the correlation between signals. The specific method is to multiply the centered signal matrix by the negative 1 / 2 power of the eigenvalue of the covariance matrix and the eigenvector matrix, so that the dimensions of the processed signal are independent of each other and the variance is one.

[0031] Based on the whitened signal, the separation matrix is ​​solved through iterative optimization, and the fast ICA algorithm is used for iterative updating. The specific steps are as follows: calculate the expected value of the result after the whitened signal interacts with the nonlinear function, and then subtract the product of the expected value of the derivative of the nonlinear function and the current separation vector to obtain the new separation vector. The nonlinear function is usually chosen as the hyperbolic tangent function or the cubic function, whose derivatives are one minus the square of the hyperbolic tangent function value or three times the square of the input value, respectively.

[0032] To meet real-time requirements, a sliding window processing strategy is adopted, with a window length of 1,000 data points and an overlap rate of 50%, processing 500 new data points at a time. This sliding window approach ensures separation accuracy while keeping algorithm latency below one second.

[0033] Adaptive wavelet basis selection mechanism; Wavelet transform is used to decompose time-domain signals into time-frequency domains, facilitating the extraction of feature information from different frequency bands. Different types of signals are suited to different wavelet basis functions; therefore, an adaptive wavelet basis selection mechanism is designed. The preset wavelet basis function library includes commonly used wavelet bases such as Daubechies wavelet, Morlet wavelet, Mexican hat wavelet, and Biorthogonal wavelet. Among them, the Daubechies wavelet has compact support properties, making it suitable for processing signals with abrupt changes; the Morlet wavelet has good time-frequency localization properties, making it suitable for analyzing non-stationary signals; the Mexican hat wavelet has symmetry, making it suitable for detecting peaks and edges in signals; and the Biorthogonal wavelet has linear phase properties, making it suitable for applications requiring high reconstruction accuracy.

[0034] For the input signal, the effectiveness of the wavelet basis selection is evaluated by calculating the reconstruction error. Specifically, this involves summing the squared differences between the original signal values ​​and the reconstructed signal values ​​at all sampling points. A smaller reconstruction error indicates that the selected wavelet basis better represents the characteristics of the original signal. By comparing the reconstruction errors of different candidate wavelet bases, the wavelet basis with the smallest error is selected as the optimal wavelet basis.

[0035] The wavelet transform coefficients are calculated by performing an inner product operation between the input signal and the wavelet basis functions that have undergone scaling and time shifting, and then normalizing the result by dividing by the square root of the scaling parameter. This process yields the local features of the signal at different times and frequencies.

[0036] Principal component analysis for dimensionality reduction; To avoid the curse of dimensionality caused by high-dimensional feature spaces, principal component analysis (PCA) is used for feature dimensionality reduction. First, the covariance matrix of the feature matrix is ​​calculated. Specifically, each sample vector is subtracted from its mean vector, and the product of these difference vectors and their transposes is calculated. Finally, normalization is achieved by dividing by the sample size minus one. Eigenvalue decomposition is then performed on the covariance matrix, decomposing it into a product of an eigenvector matrix, an eigenvalue diagonal matrix, and the transpose of the eigenvector matrix. The eigenvalues ​​are sorted by magnitude, and the top principal components with a cumulative contribution rate reaching 95% are selected. The cumulative contribution rate of each feature value is no less than 95%.

[0037] The dimensionality-reduced feature matrix is ​​obtained by subtracting the mean vector from the original feature matrix and then multiplying it by the transpose of the eigenvector matrices corresponding to the first few principal components. Through principal component analysis, the feature dimension can be reduced from hundreds to tens of dimensions, significantly reducing computational complexity.

[0038] Multimodal feature extraction strategy; For sensor data of different modalities, specialized feature extraction strategies are designed. For temperature data, statistical features, trend features, and periodic features are extracted. Statistical features include mean, variance, skewness, and kurtosis. Trend features are characterized by the slope of linear regression to represent the temperature change trend. Periodic features are identified by fast Fourier transform to identify periodic patterns of temperature changes.

[0039] For pressure data, frequency domain features and transient features are extracted: frequency domain features are obtained through power spectral density analysis to obtain the frequency distribution of the pressure signal; transient features are used to identify system start-up and shutdown and load changes by detecting pressure abrupt changes. For current data, harmonic features and power features are extracted: harmonic features are used to identify harmonic components caused by electrical faults through harmonic analysis; power features include electrical parameters such as active power, reactive power, and power factor. For vibration data, time-frequency features and envelope features are extracted: time-frequency features are obtained through short-time Fourier transform to obtain the time-frequency distribution of the vibration signal; envelope features are used to extract the modulation information of the vibration signal through envelope detection, which is used to identify bearing and gear faults. For sound data, spectral features and cepstral features are extracted: spectral features include parameters describing sound characteristics such as spectral centroid, spectral spread, and spectral flux; cepstral features are used to extract the periodicity of sound through cepstral analysis, which is used to identify abnormal sounds from compressors and fans.

[0040] Output feature matrix format; This step outputs the dimensionality-reduced feature matrix. The matrix dimension is the number of samples multiplied by the number of features, typically between fifty and one hundred dimensions. The feature matrix is ​​grouped according to modality type, with each modality containing ten to twenty features to facilitate subsequent attention fusion processing. A feature importance score is also output to guide the initialization of attention weights.

[0041] Optionally, in some embodiments, a hybrid separation algorithm combining independent component analysis and factor analysis is used to further improve signal separation accuracy, which is particularly suitable for complex environments with strong correlation between signals. Similarly, some embodiments use a deep learning-based autoencoder for feature dimensionality reduction, which can better preserve nonlinear feature information and improve the sensitivity of anomaly detection compared to principal component analysis.

[0042] S300 performs standardized preprocessing on the dimensionality-reduced feature matrix, adopts depthwise separable convolution, designs a lightweight attention mechanism to adjust the feature weights of different modalities, calculates the intermodal correlation through a cross-modal interaction module, and outputs a fused feature vector. Input feature preprocessing; Received dimensionality reduction feature matrix Feature standardization is performed to ensure that features of different modalities are within the same numerical range. The standardization method uses Z-score standardization: for each feature dimension, the mean and standard deviation of that dimension on historical data are calculated, and then the current feature value is subtracted from the mean and divided by the standard deviation, so that the mean of the standardized feature is zero and the standard deviation is one, thus eliminating the dimensional differences between different modalities.

[0043] Design of depthwise separable convolutional networks; To reduce the number of network parameters and computational complexity, depthwise separable convolution is used instead of traditional fully connected layers. Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs the convolution operation independently for each input channel. Depthwise convolution produces a depthwise convolution output by convolving the input feature map with a depthwise convolution kernel. Pointwise convolution uses a 1x1 convolution kernel to fuse inter-channel information from the depthwise convolution output, resulting in the final pointwise convolution output. Compared to standard convolution, depthwise separable convolution reduces the number of parameters by approximately eight to ten times and lowers computational complexity by approximately five to eight times.

[0044] Lightweight attention mechanisms; We design a lightweight attention mechanism that dynamically adjusts the importance weights of features from different modalities. Traditional multi-head attention mechanisms are computationally complex; therefore, we employ a simplified scheme combining single-head attention with feature grouping. Input features are grouped according to modality type, with each group containing ten to twenty features.

[0045] For the For each feature group, attention weights are calculated using the softmax function. Specifically, the inner product of the attention weight vector and the mean vector of each feature group is used as input to an exponential function. Then, the exponential values ​​of all feature groups are normalized to ensure that the sum of all weights is one. The attention weight vector is learned through backpropagation, with the optimization objective being to minimize the anomaly detection loss function. The inner product operation is calculated by multiplying each element of the weight vector by the corresponding element of the feature vector, and then summing all the products to obtain a scalar value. This scalar value reflects the importance of the current feature group.

[0046] Cross-modal feature interaction; In addition to intramodal feature fusion, intermodal feature interactions also need to be considered. A cross-modal interaction module is designed to calculate the correlation between different modalities. Cross-modal correlation is measured by calculating the cosine similarity between feature vectors of different modalities. Specifically, the inner product of two feature vectors is divided by the product of their respective vector lengths, yielding a correlation coefficient ranging from negative one to positive one. The larger the absolute value of the correlation coefficient, the stronger the correlation between the two modalities. The vector length is obtained by taking the square root of the sum of the squares of each element in the vector; this value represents the magnitude of the vector. Based on the correlation, intermodal interaction features are constructed: in, Represents cross-modal interaction feature vectors. This represents the element-wise multiplication operation, which multiplies the corresponding elements of two vectors of the same dimension to form a new vector. This indicates the total number of modalities, i.e., the number of different types of sensors; This represents the correlation between the i-th mode and the j-th mode, with a value range of [-1, 1]. This represents the eigenvector of the i-th mode; Let represent the eigenvector of the j-th mode.

[0047] Attention weights are adaptively adjusted. To improve the adaptability of the attention mechanism, an adaptive weight adjustment strategy is designed. Based on the statistical characteristics of the current input features, the calculation method of the attention weights is dynamically adjusted. When the features of a certain modality change significantly, the attention weight of that modality is increased accordingly; when the feature changes are small, the attention weight is decreased.

[0048] Adaptive adjustment of attention weights is achieved by multiplying the original attention weights by one and adding the product of the adaptive adjustment coefficient and the standard deviation of the modal features. Before calculation, the standard deviations of the modal features need to be normalized: the standard deviation of each modal feature is divided by the maximum value of all modal feature standard deviations to ensure that the standard deviation value is within the range of zero to one, avoiding adjustment bias caused by differences in the numerical range of different modal features. The adaptive adjustment coefficient ranges from 0.1 to 0.5, and the normalized standard deviation of the modal feature reflects the relative degree of change of that modal feature. Through adaptive adjustment, the attention mechanism can better adapt to dynamically changing input features.

[0049] Feature fusion calculation; Based on attention weights and cross-modal interaction features, the final fused feature vector is calculated: in, This represents the final fused feature vector. Indicates the total number of modes. The weight coefficients representing the interaction features. This represents the adaptive attention weight for the i-th modality, with a value range of [0, 1]. This represents the eigenvector of the i-th mode; This represents the interaction feature vector between modalities. The fused feature vector has a dimension of thirty to fifty, containing important information about each modality and the interaction information between modalities.

[0050] Network training optimization; The attention fusion network is trained end-to-end, jointly with the subsequent anomaly detection network. The loss function is a weighted combination of cross-entropy loss and attention regularization loss. in, Indicates the total loss. The cross-entropy loss represents the loss for anomaly detection. This represents the attention regularization loss, used to prevent excessive concentration of attention weights. This represents the regularization weight coefficient, which is usually set to 0.01.

[0051] Attention regularization loss is achieved by calculating the negative sum of the products of each modality's attention weight and its logarithm. Specifically, the logarithm of each modality's attention weight is taken, multiplied by that weight, and the sum of all modality products is then negative. This loss function encourages diversity in attention weight distribution and avoids over-reliance on features from any single modality.

[0052] Output fused features; This step outputs the fused feature vector. The feature dimensions range from thirty to fifty, encompassing the essence of multimodal information. It also outputs the attention weight distribution for subsequent interpretability analysis and system debugging.

[0053] Optionally, in some embodiments, the multi-head self-attention mechanism in the Transformer architecture can be used to further improve the feature fusion effect, but the computational complexity will increase accordingly. This is suitable for application scenarios with sufficient computing resources, as well as for using graph neural networks to model the topological relationship between sensors and to achieve more refined feature fusion through graph convolution operations. This is particularly suitable for application scenarios where the spatial distribution of sensors has a clear topological structure.

[0054] S400, based on fused feature vectors, adopts a two-layer progressive detection architecture. The first layer uses statistical methods for coarse screening, and the second layer uses a lightweight neural network for precise detection. The detection results are fused and calculated to obtain an anomaly score and output an anomaly probability sequence. Input feature reception and preprocessing; Receive fused feature vector The feature dimensions are thirty to fifty. First, the quality of the fused features is evaluated by calculating the information entropy and variance of the feature vectors. If the information entropy is too low or the variance is too small, it indicates that the feature information is insufficient, and the parameter settings of the previous steps need to be adjusted.

[0055] Layered detection architecture design; A two-layer progressive detection architecture is adopted. The first layer is a statistical detection layer, which uses simple and fast statistical methods for coarse screening. The second layer is a neural network detection layer, which uses a lightweight deep learning model for accurate detection. This layered design ensures both real-time performance and detection accuracy.

[0056] First-level statistical anomaly detection; The statistical detection layer integrates multiple statistical methods, including distance-based anomaly detection, density-based anomaly detection, and clustering-based anomaly detection. First, the Mahalanobi distance between the current feature vector and historical normal samples is calculated. This distance is obtained by taking the square root of the quadratic form of the difference between the current feature vector and the mean vector of historical normal samples, after transforming the inverse of the covariance matrix. Before calculation, the feature vectors need to be preprocessed: ensuring that the current feature vector and historical normal samples use the same standardization method, typically Z-score standardization to make the feature mean zero and the standard deviation one; the covariance matrix is ​​calculated based on the standardized historical normal samples to eliminate dimensional differences between different feature dimensions. This distance considers the correlation between features and can effectively identify outliers in multidimensional space. When the Mahalanobi distance exceeds a preset threshold, it is initially judged as an anomaly.

[0057] Simultaneously, the local outlier factor is calculated to measure the relative density change of the current sample within its neighborhood. The local outlier factor is obtained by averaging the ratios of the local reachability densities of the sample's k-nearest neighbors to the sample's local reachability density. This factor reflects the degree of density deviation of the sample relative to its neighborhood; a larger value indicates a higher likelihood of an outlier. A local outlier factor greater than 1.5 is considered to indicate that the sample has an anomalous tendency.

[0058] The second layer of the neural network ensures accurate detection; The neural network detection layer employs a lightweight fully connected neural network with a three-layer structure: the input layer has the same number of nodes as the feature dimension, the hidden layer has 20-30 nodes, and the output layer has two nodes, corresponding to the normal and abnormal categories, respectively. To reduce computational complexity, the ReLU function is used as the activation function in the hidden layers. This function has advantages such as simple computation, stable gradient propagation, and the ability to alleviate the vanishing gradient problem, while also introducing sparsity to improve the network's expressive power. The network output is converted into a probability distribution through softmax normalization. The softmax function is implemented as follows: for each element of the input vector, the exponent value of that element is first calculated, and then this exponent value is divided by the sum of the exponent values ​​of all elements, ensuring that the sum of all elements in the output vector is one and that each element is a positive value, thus forming an effective probability distribution. For the anomaly detection task, the network output score of the anomaly category is processed by the softmax function to obtain the anomaly probability.

[0059] Abnormal score fusion calculation; The results of statistical detection and neural network detection are fused. The anomaly score is obtained by weighted summation of the statistical detection score, neural network detection score, and temporal consistency score. Before fusion calculation, different types of scores need to be standardized preprocessed. The statistical detection score is standardized using Z-score, the neural network detection score is normalized to a probability value in the zero-to-one interval using the softmax function, and the temporal consistency score is mapped to the zero-to-one interval using maximum-minimum normalization. Standardization ensures that the scores output by different detection methods have the same numerical range and statistical characteristics, avoiding fusion bias caused by differences in units.

[0060] Time consistency constraint mechanism; To address the consistency issue in multi-timescale prediction, a temporal consistency constraint mechanism is designed. Anomaly detection results are smoothed over time to prevent drastic fluctuations between adjacent time points. The stability of the detection results is evaluated by calculating a temporal consistency score, specifically by statistically analyzing the proportion of adjacent time points within a time window where the difference in anomaly probability is less than a threshold. When the consistency score falls below 0.8, the current detection result is smoothed and corrected using a weighted average method to fuse the current and previous detection results, with a smoothing coefficient of 0.7.

[0061] Multi-timescale prediction is achieved; Based on current anomaly scores and historical trends, anomaly predictions are achieved across three time scales: short-term, medium-term, and long-term. Short-term predictions cover the next one to twenty-four hours and are primarily based on trend analysis of current anomaly scores; medium-term predictions cover the next twenty-four to forty-eight hours and incorporate historical periodic patterns; long-term predictions cover the next forty-eight to seventy-two hours and employ a time series forecasting model.

[0062] The time series forecast uses a simplified ARIMA model with parameters set to ARIMA(2, 1, 2), meaning the autoregression order is two, the differencing order is one, and the moving average order is two. The forecast formula is: in, This indicates that the ARIMA model at time... The predicted value, This represents the constant term of the ARIMA model. and Let represent the autoregressive coefficients for lag one period and lag two period, respectively. and They represent the first The moment and the The observation value at each moment, and These represent the moving average coefficients for lags of one period and two periods, respectively. and They represent the first The moment and the The prediction error at each time point, Indicates the first The random error term at each time step.

[0063] Output anomaly probability sequence; This step outputs a time-consistent sequence of anomaly probabilities. It includes the anomaly score at the current moment and the predicted probability for the next 72 hours. The output format is a time series array, where each element contains a timestamp and an anomaly probability value. The anomaly probability value ranges from zero to one, where zero indicates complete normality and one indicates a definite anomaly.

[0064] Optionally, in some embodiments, an adaptive threshold adjustment mechanism based on deep reinforcement learning is adopted to dynamically adjust the anomaly judgment threshold according to the detection effect, thereby improving detection accuracy and reducing false alarm rate. In addition, an ensemble learning method is adopted to vote and fuse the results of multiple anomaly detection algorithms, including methods such as isolated forest, support vector machine and autoencoder, to further improve the robustness of detection.

[0065] S500 identifies fault types based on abnormal probability sequences, builds an expert rule knowledge base, integrates data-driven machine learning results with knowledge-driven expert rules, and generates emergency plans based on fault type and severity. Input data integration; Receive time consistent anomaly probability sequence At the same time, integrate expert knowledge base The expert knowledge base includes rules and experience related to air conditioning system failures. It contains typical failure modes of air conditioning systems, the mapping relationship between failure types and symptoms, maintenance strategies, and emergency response plans.

[0066] Fault type identification module; Based on anomaly probability sequences and feature vectors, specific fault types are identified. A multi-classifier is designed to categorize possible fault types into five main classes: compressor fault, condenser anomaly, electronic control module fault, sensor failure, and system leakage. Each fault type has a specific feature pattern. Compressor fault characteristic modes: abnormally increased current, vibration frequency deviation, sound spectrum change, slow temperature rise; condenser abnormal characteristic modes: increased high-pressure side pressure, decreased temperature difference, fan current change; electronic control module fault characteristic modes: abnormal control signal, current fluctuation, system response delay; sensor failure characteristic modes: data jump, signal loss, reading deviation; system leakage characteristic modes: pressure drop, weakened cooling effect, extended running time.

[0067] Fault type identification uses a support vector machine classifier: in, Represents the classification decision function. Represents a symbolic function. The kernel function is represented by a radial basis function kernel. Represents the Lagrange multipliers. Indicates training samples Category tags, This represents the input sample to be classified. Indicates the bias term. This represents the number of support vectors. The specific form of the radial basis function kernel is: calculate the square of the Euclidean distance between the input sample and the support vector, then multiply it by the negative kernel parameter, and finally take the exponential function value. This kernel function can map linearly inseparable data to a high-dimensional space to achieve linear separability.

[0068] Fault severity assessment; The severity of the fault is assessed based on the magnitude, duration, and trend of the anomaly probability. The severity scoring formula is as follows: in, The score indicates the severity of the fault. This represents the maximum value of the anomaly probability. The normalized value representing the duration of the anomaly. Indicates an increasing trend in the probability of an anomaly. , , These represent the first, second, and third weight coefficients, respectively, and the sum of the weight coefficients equals one.

[0069] Based on the severity rating, the faults are divided into three levels: minor faults, with a score of 0 to 0.3, the system can continue to operate but requires attention; moderate faults, with a score of 0.3 to 0.7, require preventive measures and preparation for maintenance; and severe faults, with a score of 0.7 to 1, require immediate shutdown for maintenance or activation of the emergency plan.

[0070] Construction of an expert rules knowledge base; Construct a rule knowledge base that incorporates expert experience, with rules represented in IF-THEN format: Rule 1: If the compressor current is greater than 1.2 times the rated current and the vibration frequency deviation exceeds 10%, then the compressor bearings may be worn. Confidence level: 0.85; Rule 2: If the high-pressure side pressure is greater than 3.5 MPa and the condenser outlet temperature is 20 degrees Celsius higher than the ambient temperature, then the condenser has poor heat dissipation. Confidence level: 0.9; Rule 3: If the temperature sensor reading changes at a rate exceeding 5 degrees Celsius per second, then the sensor is faulty. Confidence level: 0.95; The formal representation of the rule is as follows: in, This represents the i-th rule. The conditions are indicated (including specific judgment conditions such as sensor measurements and status parameters). The conclusion is indicated (including diagnostic results such as fault type and severity). This represents the confidence factor, with a value range of [0, 1], reflecting the confidence level of the rule.

[0071] Data-driven and knowledge-driven integrated decision-making; The final decision is formed by fusing data-driven machine learning results with knowledge-driven expert rules. The fusion method employs a weighted average, and the final fused decision is obtained by multiplying the data-driven fusion weight by the data-based machine learning decision result, and then adding the knowledge-driven fusion weight multiplied by the rule-based expert knowledge decision result. Before fusion, both decision results undergo standardization preprocessing: the machine learning decision result is normalized to a probability value between zero and one using the softmax function; the expert knowledge decision result is normalized by maximizing and minimizing the product of the rule matching degree and the confidence factor to ensure the output value is also within the zero-to-one range. Standardization eliminates the dimensional differences between the output results of different decision methods, ensuring the effectiveness of the fusion computation. Specifically, the data-driven fusion weight is dynamically adjusted based on historical validation results, with an initial value set to 0.7, and the knowledge-driven fusion weight is one minus the data-driven fusion weight.

[0072] When data-driven and knowledge-driven results are consistent, the confidence level of the decision is increased; when the results are inconsistent, the confidence level is reduced and a manual verification mechanism is triggered.

[0073] Emergency response plan generation; Based on the type and severity of the fault, a corresponding emergency plan is automatically generated. The plan includes four parts: diagnostic results, risk assessment, handling measures, and maintenance recommendations. Diagnostic results: clearly define the fault type, location, severity, and expected development trend; Risk assessment: analyze the impact of the fault on system functionality, personnel safety, and task execution; Handling measures: formulate immediately actionable emergency response steps based on the current environmental and resource conditions; Maintenance recommendations: provide detailed maintenance guidance, including required tools, spare parts, maintenance procedures, and estimated time.

[0074] Implementation of multi-priority protection strategies; Differentiated protection strategies are developed based on the importance of different areas inside the vehicle. The vehicle interior is divided into three priority zones: Critical zone: the driver's seat and the area where important equipment is placed, with the highest priority and a temperature control accuracy requirement of ±1 degree Celsius; Important zone: the passenger rest area and the general equipment area, with the medium priority and a temperature control accuracy requirement of ±3 degrees Celsius; Ordinary zone: the storage area and the aisle, with the lowest priority and a temperature control accuracy requirement of ±5 degrees Celsius.

[0075] When the air conditioning system malfunctions and reduces its cooling capacity, priority should be given to ensuring the environmental conditions of critical areas. If necessary, the air supply to ordinary areas should be shut off, and the limited cooling capacity should be concentrated in critical and important areas.

[0076] Output the decision results; This step outputs optimized early warning decisions. and detailed repair suggestions Early warning decisions include fault probability, fault type, severity, warning level, and handling recommendations; maintenance recommendations include fault location, possible causes, maintenance steps, required resources, and estimated time.

[0077] Optionally, in some embodiments, a case-based reasoning-based knowledge base expansion mechanism is adopted to automatically add new fault cases and maintenance experience to the knowledge base, enabling continuous learning and updating of the knowledge base; by integrating 3D visualization technology, fault location and maintenance guidance are displayed in the form of 3D models, providing maintenance personnel with more intuitive operation guidance.

[0078] The present invention provides a method for monitoring abnormal states of automotive air conditioning controllers based on multimodal data fusion. Some application examples are as follows: In a northern city, a car is continuously driven in an extreme environment where the temperature drops to below -30 degrees Celsius (including when the engine is running and the car is stationary); the car's air conditioning system needs to maintain the interior temperature between 18 and 22 degrees Celsius. Sensor deployment and configuration: According to the method of the present invention, five types of intelligent sensors are deployed at key locations in the air conditioning system, and the specific configurations are shown in Table 1: Table 1 Smart Sensor Network Configuration Table Table 1 shows the detailed configuration information of the intelligent sensor network. Temperature sensors are deployed at the compressor inlet and outlet to monitor refrigerant temperature changes; pressure sensors are installed on the high and low pressure pipelines to monitor the system pressure status in real time; current sensors are connected to the controller power line to monitor power consumption changes; vibration sensors are fixed to the compressor body to collect mechanical vibration signals; and sound sensors are placed at the air outlet to capture the sound characteristics of the system operation.

[0079] Real-time data acquisition example; During continuous monitoring, the system collects a large amount of data per second; Table 2 shows typical data acquisition results at a certain moment: Table 2 Real-time Data Acquisition Table for Multimodal Sensors Table 2 records the multimodal sensor data of the system under normal operating conditions. Temperature 1 and Temperature 2 represent the compressor inlet and outlet temperatures, respectively; Pressure 1 and Pressure 2 represent the high-pressure and low-pressure pipeline pressures, respectively; Current represents the controller power consumption; Vibration X, Y, and Z represent triaxial vibration acceleration; and Sound represents the system operating noise level. The data shows that the system operates smoothly, and all parameters are within the normal range.

[0080] Key performance indicator verification: Through extensive testing in laboratory and real-world application environments, the key performance indicators of the method of this invention were verified. Table 3 shows a detailed comparison with traditional single-modal detection methods: Table 3 Comparison and Verification Table of Key Performance Indicators Table 3 details the performance comparison results between the method of this invention and the traditional single-modal detection method. Through 1200 anomaly detection tests, the method of this invention outperforms the traditional method in key indicators such as accuracy, false alarm rate, and false negative rate. In particular, the warning time window is extended from 24 hours to 68 hours, providing more time for maintenance decisions.

[0081] Verification of detection effectiveness for different fault types: The detection effectiveness was verified for typical fault types in automotive air conditioning systems. Table 4 shows the detection performance for various fault types. Table 4. Detection Results for Different Fault Types Table 4 summarizes the detection performance for six typical fault types. Compressor bearing wear showed the best detection performance, with an accuracy rate of 96.8% and an average warning time of 72 hours. Sensor drift detection was relatively difficult, but the accuracy rate still reached 89.4%. Overall, the system demonstrated good detection capabilities for various fault types.

[0082] Summary of overall technical effects: As shown in Table 5, the overall technical effects of this invention are summarized as follows: Table 5 Summary of Overall Technical Effects Table 5 comprehensively illustrates the technical effects of this invention in seven aspects: detection performance, early warning capability, system efficiency, environmental adaptability, resource consumption, reliability, and fault coverage. Compared with traditional methods, this invention achieves improvements in all key indicators, providing reliable technical support for monitoring automotive air conditioning systems in extreme environments such as polar regions.

[0083] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion, characterized in that, Includes the following steps: Temperature sensors, pressure sensors, current sensors, vibration sensors, and sound sensors are deployed at key locations in the automotive air conditioning system, and time synchronization is achieved through the vehicle's CAN bus to obtain multimodal synchronized data streams; Data validity checks and outlier removal are performed on the multimodal synchronous data stream. The fast ICA algorithm combined with sliding window processing is used to achieve signal separation. An adaptive wavelet basis selection mechanism is designed for time-frequency analysis. Principal component analysis is used for feature dimensionality reduction, and the dimensionality-reduced feature matrix is ​​output. The dimensionality-reduced feature matrix is ​​standardized and preprocessed, depthwise separable convolution is used, a lightweight attention mechanism is designed to adjust the feature weights of different modalities, the intermodal correlation is calculated through a cross-modal interaction module, and the fused feature vector is output. Based on fused feature vectors, a two-layer progressive detection architecture is adopted. The first layer uses statistical methods for coarse screening, and the second layer uses a lightweight neural network for precise detection. The detection results are fused and calculated to obtain an anomaly score and output an anomaly probability sequence. Based on the identification of fault types using abnormal probability sequences, an expert rule knowledge base is constructed, and data-driven machine learning results are integrated with knowledge-driven expert rules to generate emergency plans according to the fault type and severity.

2. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The temperature sensor is deployed at the compressor inlet and outlet, the pressure sensor is installed in the high and low pressure pipelines, the current sensor is connected to the controller power line, the vibration sensor is fixed to the compressor body, the sound sensor is placed at the air outlet, the intelligent sensor node has a built-in microprocessor, the local data cache unit has a capacity of 512KB, the vehicle CAN bus communication rate is 500Kbps, and the time synchronization accuracy is controlled within one millisecond.

3. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The data validity check includes data integrity verification, timestamp consistency check, and numerical range reasonableness judgment. Outlier removal adopts the 3σ criterion and quartile method. The fast ICA signal separation adopts a sliding window processing strategy with a window length of one thousand data points and an overlap rate of fifty percent. Blind source separation is achieved by calculating the statistical independence and non-Gaussianity of the signal. The algorithm delay is controlled within one second.

4. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The adaptive wavelet basis selection mechanism pre-sets a wavelet basis function library including Daubechies wavelet, Morlet wavelet, Mexican hat wavelet, and Biorthogonal wavelet. It automatically selects the optimal wavelet basis based on the spectral characteristics of the signal and the time-frequency localization requirements. The wavelet decomposition level is set to 5 levels, and the reconstruction error threshold is set to 0.

01. The principal component analysis method achieves feature dimensionality reduction by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. The cumulative contribution rate threshold is set to 95%, and the dimensionality after feature dimensionality reduction is controlled within 20% of the original dimensionality.

5. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The lightweight attention mechanism groups the input features according to modality type, with each group containing ten to twenty features. The attention weights are calculated using the softmax function. Specifically, the inner product of the attention weight vector and the mean vector of each feature group is used as the input of the exponential function. Then, the exponential values ​​of all feature groups are normalized to ensure that the sum of all weights is one.

6. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The anomaly score fusion calculation is obtained by weighted summation of statistical detection score, neural network detection score and time consistency score. Before fusion calculation, different types of scores are standardized preprocessed. Statistical detection score is standardized using Z-score, neural network detection score is normalized to a probability value between zero and one using the softmax function, and time consistency score is normalized to a zero-to-one interval using maximum and minimum value normalization.

7. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The first layer of the two-layer progressive detection architecture uses Mahalanobi distance detection and a sliding window-based statistical control chart method. The Mahalanobi distance is obtained by calculating the difference between the current feature vector and the mean vector of historical normal samples, and then taking the square root of the quadratic form after transforming the inverse of the covariance matrix. The second layer of the lightweight neural network uses a fully connected neural network with a three-layer structure. The number of nodes in the input layer is equal to the feature dimension, the number of nodes in the hidden layer is twenty to thirty, and the number of nodes in the output layer is two. The smoothing window length of the time consistency constraint mechanism is 10 time steps.

8. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The expert rule knowledge base comprises three sub-bases: a fault mode base, a diagnostic rule base, and a decision tree base. Fault type identification uses a support vector machine classifier, classifying faults into five categories: compressor faults, condenser malfunctions, electronic control module faults, sensor failures, and system leaks. Among these, compressor fault characteristic patterns include abnormally increased current, vibration frequency shift, sound spectrum changes, and slow temperature rise. The fusion of data-driven and knowledge-driven approaches uses a weighted average method, with the initial fusion weight of the data-driven approach set to 0.7, and the fusion weight of the knowledge-driven approach being one minus the fusion weight of the data-driven approach.

9. The method for monitoring abnormal states of an automotive air conditioning controller using multimodal data fusion according to claim 1, characterized in that, The emergency response plan is generated in three levels based on the severity of the fault: minor fault, moderate fault, and severe fault. Minor faults are scored from 0 to 0.3, indicating that the system can continue to operate but requires attention. Moderate faults are scored from 0.3 to 0.7, indicating that preventive measures should be taken and maintenance should be prepared. Severe faults are scored from 0.7 to 1, indicating that the system needs to be shut down immediately for maintenance or the emergency response plan should be activated. The severity score is calculated by weighting and summing the maximum value of the abnormal probability, the normalized value of the abnormal duration, and the growth trend of the abnormal probability.

10. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, can run a multimodal data fusion method for monitoring abnormal states of an automotive air conditioning controller as described in any one of claims 1-9.

Images