On-board intelligent diagnosis method and system integrating cloud-edge computing
By adopting cloud-edge computing method in the vehicle intelligent diagnostic system, the implicit representation of the first sensing type data is mined and projected into the domain of the second sensing type, and the vehicle diagnosis network is optimized, which solves the high-precision sensing data adjustment Problems improve diagnostic accuracy and flexibility.
Patent Information
- Application Number
- CN202411170894.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-26
AI Technical Summary
Existing vehicle intelligent diagnostic technology faces great difficulty when adjusting sensor data with high accuracy and high information representation capabilities, and sensor data with relatively easy access but weak information representation capabilities cannot be effectively utilized, resulting in difficulty in tuning the diagnostic network.
Using the method of fusion cloud edge computing, the first sensing type data is obtained for implicit representation mining, and it is projected to the implicit representation domain of the second sensing type, and the configuration parameters of the vehicle diagnostic network are optimized to reduce adjustment. Difficulty and improve diagnostic accuracy.
It effectively reduces the difficulty of tuning the vehicle diagnostic network, improves the diagnostic accuracy of sensing data for high information characterization capabilities, reduces data acquisition costs, and enhances the flexibility and accuracy of the vehicle's intelligent diagnostic system.
Smart Images

Figure CN118711276B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing, and more specifically, to an on-board intelligent diagnosis method and system integrating cloud-edge computing. Background Art
[0002] With the rapid development of the automobile industry, the requirements for vehicle safety and reliability are increasing, and the rapid and accurate diagnosis of vehicle faults has become an important research direction in the automobile industry. In the existing vehicle intelligent diagnosis technology, a vehicle diagnosis network is usually pre-trained in the cloud and then deployed on the vehicle side (i.e., the edge side). The performance of the vehicle diagnosis network directly determines the accuracy and reliability of the diagnosis results. However, there are many challenges in building and tuning the vehicle diagnosis network. In particular, for some high-precision and high-information representation capability sensor information (such as engine speed, oil pressure, temperature and other second sensor type data), although these data are crucial for fault diagnosis, they are difficult to collect and costly, and also put forward higher requirements for the tuning of the vehicle diagnosis network. In contrast, some relatively easy-to-obtain and low-cost sensor information (such as engine vibration signals, sound signals and other first sensor type data) have weak information representation capabilities, but can also reflect the operating status of the vehicle to a certain extent. In order to overcome the above technical difficulties, researchers continue to explore new technical solutions to reduce the difficulty of tuning the vehicle diagnosis network while ensuring the accuracy of diagnosis. Summary of the invention
[0003] The purpose of the present invention is to provide an on-board intelligent diagnosis method and system integrating cloud-edge computing. The embodiment of the present application is implemented as follows:
[0004] In a first aspect, an embodiment of the present application provides an on-board intelligent diagnosis method integrating cloud-edge computing, which is applied to a cloud server, and the method includes: obtaining first example sensor data belonging to a first sensor type, the first example sensor data carrying first prior information, and the first prior information is used to indicate an actual diagnostic result of the first example sensor data; performing implicit representation mining on the first example sensor data based on a first implicit representation mining network to obtain a first implicit representation array, the first implicit representation mining network is used to project the sensor information corresponding to the first sensor type to a target implicit representation domain, and the target implicit representation domain is a sensor information belonging to a second sensor type. The implicit representation domain of the implicit representation array corresponding to the information is located, the first sensor type and the second sensor type are different data types, and the information representation capability of the sensor information corresponding to the first sensor type is less than the information representation capability of the sensor information corresponding to the second sensor type; according to the first implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result; according to the error between the inference diagnosis result and the first prior information, the network configuration parameters of the vehicle diagnostic network to be calibrated are optimized to obtain a vehicle diagnostic network, and the vehicle diagnostic network is used to diagnose the sensor information corresponding to the second sensor type.
[0005] On the other hand, an embodiment of the present application provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.
[0006] The technical effects of the present application include at least: in the embodiment of the present application, because the information representation capability of the sensor information corresponding to the second sensor type is strong, but its collection is difficult, the adjustment of the corresponding vehicle diagnostic network is difficult. Then, in order to alleviate the difficulty of adjusting the vehicle diagnostic network, the vehicle diagnostic network is no longer adjusted according to the example sensor data corresponding to the second sensor type, but instead, the vehicle diagnostic network is adjusted according to the example sensor data corresponding to the first sensor type. Among them, the information representation capability of the sensor information corresponding to the first sensor type is less than the information representation capability of the sensor information corresponding to the second sensor type. In other words, the collection of the example sensor data corresponding to the first sensor type is relatively simple. During calibration, first example sensor data including first prior information is obtained, and implicit representation mining is performed on the first example sensor data based on a first implicit representation mining network to obtain a first implicit representation array. The first implicit representation mining network can project sensor information corresponding to the first sensor type to a target implicit representation domain, and the target implicit representation domain is an implicit representation domain where the implicit representation array corresponding to the sensor information of the second sensor type is located. In other words, the first implicit representation mining network can project sensor information corresponding to the first sensor type and sensor information corresponding to the second sensor type to one implicit representation domain. According to the first implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result. In other words, the calibration of the vehicle diagnostic network to be calibrated can process the implicit representation array located in the target implicit representation domain to optimize the network configuration parameters of the vehicle diagnostic network to be calibrated according to the error between the inference diagnosis result and the first prior information, so that the error between the inference diagnosis result and the first prior information is gradually reduced to obtain the vehicle diagnostic network. Because the vehicle diagnostic network can parse the implicit representation array located in the target implicit representation domain, the vehicle diagnostic network can also process the sensor information corresponding to the second sensor type, thereby obtaining the fault diagnosis result of the sensor information corresponding to the second sensor type. Therefore, without obtaining the sample sensor data corresponding to the second sensor type, the vehicle diagnostic network can also be debugged based on the sample sensor data of other types with weaker information representation capabilities, thereby reducing the difficulty of adjusting the vehicle diagnostic network. BRIEF DESCRIPTION OF THE DRAWINGS
[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.
[0008] Figure 1 This is a flow chart of an on-board intelligent diagnosis method integrating cloud-edge computing provided in an embodiment of the present application.
[0009] Figure 2 It is a schematic diagram of the composition of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION
[0010] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method section of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application. The execution subject of the vehicle-mounted intelligent diagnosis method integrating cloud-edge computing in the embodiments of the present application is a computer system, including but not limited to servers, personal computers, laptops, tablet computers, etc. Figure 1 As shown, the method includes:
[0011] Step S10: Acquire first example sensor data belonging to a first sensor type, where the first example sensor data carries first priori information, and the first priori information is used to indicate an actual diagnosis result of the first example sensor data.
[0012] In one implementation scenario, assume that an on-board intelligent diagnostic system is being developed, which aims to predict and diagnose potential mechanical failures by analyzing the real-time sensor data of the vehicle. In this scenario, the vehicle is equipped with a variety of sensors to monitor engine vibration, sound, temperature, oil pressure and other parameters. These sensor data are valuable input information for the diagnostic system.
[0013] In this scenario, the first sensing type can be defined as sensor data that is relatively easy to obtain and has a low cost, such as engine vibration data and sound data. Although the amount of information in these data is relatively small, they are sufficient to provide some preliminary signs of faults. The computer system collects the raw data of the engine vibration and sound sensors in real time or offline as the first example sensing data through the vehicle network or directly connected to the sensor interface. These data are stored in the form of time series, and each time point contains multiple characteristic values, such as vibration amplitude, frequency distribution, sound intensity, etc. For each set of first example sensing data, a clear label or result is required to indicate the actual fault state or diagnosis result corresponding to this set of data, which is the first prior information. For example, for a set of vibration and sound data, the prior information may be "engine bearing wear" or "no fault". This prior information usually comes from professional fault diagnosis personnel or historical fault records.
[0014] Specifically, the computer system reads the vibration and sound data of the engine through a network interface or directly from the sensor. These data are captured in a digital form and stored as time series data. In order to obtain the first prior information, the computer system accesses a database containing historical fault records and expert diagnosis results. For each set of first example sensor data collected, the system will search for the corresponding diagnosis result in the database and use it as the prior information of the set of data. Before the sensor data and prior information are used for training, the system can also perform some preprocessing steps, such as data cleaning (removing noise or outliers), feature extraction (calculating statistics or conversion from raw data), etc., to improve the efficiency and accuracy of the subsequent training process. The preprocessed first example sensor data and its corresponding prior information are stored in the training data set for use in subsequent steps (such as step S20). Through such steps, the computer system can collect a large amount of first example sensor data with clear diagnosis results, providing a solid foundation for subsequent vehicle diagnosis network training.
[0015] Step S20: Based on the first implicit representation mining network, the first example sensor data is implicitly mined to obtain a first implicit representation array. The first implicit representation mining network is used to project the sensor information corresponding to the first sensor type to the target implicit representation domain. The target implicit representation domain is the implicit representation domain where the implicit representation array corresponding to the sensor information belonging to the second sensor type is located. The first sensor type and the second sensor type are different data types. The information representation capability of the sensor information corresponding to the first sensor type is less than the information representation capability of the sensor information corresponding to the second sensor type.
[0016] In an embodiment of the present application, step S20 converts first sensor type data (such as engine vibration data) that is relatively easy to obtain but has weak information representation capabilities into a higher-level implicit representation domain, so that it can simulate or approach the representation form of second sensor type data (such as detailed engine operating parameters) that has stronger information representation capabilities but is more difficult to collect.
[0017] In step S20, the computer system runs the first implicit representation mining network, processes the first example sensor data, and generates a first implicit representation array. The first implicit representation mining network is a complex machine learning model, which can be a neural network based on deep learning, such as a convolutional neural network (CNN), an autoencoder (Autoencoder) or a variational autoencoder (VAE), etc. These network structures are particularly suitable for extracting high-order features of data and mapping them to a new implicit representation domain.
[0018] In the embodiment of the present application, the first sensor type data may include vibration signals of the engine, which are collected by sensors such as accelerometers. Although these vibration signals can provide certain fault information, their information representation ability is relatively weak compared to detailed engine operating parameters (such as oil pressure, speed, temperature, etc., i.e., the second sensor type data).
[0019] When the computer system executes step S20, it first receives the first example sensor data (such as the vibration signal of the engine) from step S10. Then, these data are input into the first implicit representation mining network. Inside the network, the input data is feature extracted and abstracted through multiple layers of nonlinear transformations (such as convolutional layers, pooling layers, fully connected layers, etc.). In particular, these transformations are intended to capture key features in the vibration signal, which are essential for diagnosing engine faults but may exist in the second sensor type data in different forms. Finally, the first implicit representation mining network outputs a first implicit representation array, which can be regarded as a projection of the vibration signal in the target implicit representation domain. This target implicit representation domain is the implicit feature space of the second sensor type data (such as engine operating parameters). Although the two data types have different original forms, they share similar feature representations in this implicit representation domain. For example, assume that the input first example sensor data (vibration signal) is sampled to form a feature matrix, in which each row represents the vibration data at a time point, including multiple features such as amplitude and frequency. For example:
[0020] ;
[0021] After being processed by the first implicit representation mining network, the output may be a new feature vector (i.e., the first implicit representation array), which captures more critical and abstract features of the original vibration signal for fault diagnosis. For example:
[0022] ;
[0023] Each feature is obtained after a complex transformation of the original vibration signal. They may represent a specific mode or state of engine operation, which is not easy to directly observe in the original vibration signal, but is highly correlated with the features in the second sensor type data (such as detailed engine operating parameters) in the target implicit representation domain. Through such an implicit representation mining process, the computer system can convert the first sensor type data with weak information representation ability into a richer and more useful representation in a higher-level implicit representation domain, providing strong support for the subsequent vehicle diagnosis network training.
[0024] Step S30: performing fault diagnosis based on the vehicle diagnostic network to be calibrated according to the first implicit representation array to obtain an inference diagnosis result.
[0025] In step S30 of the embodiment of the present application, the computer system uses the first implicit representation array obtained by processing the first implicit representation mining network to evaluate and optimize the performance of the vehicle diagnostic network to be adjusted. The vehicle diagnostic network to be adjusted is a machine learning model that has been preliminarily constructed, which is used to predict possible faults of the vehicle based on the input sensor data (the first implicit representation array in this scenario). This model can be built based on a neural network (such as a convolutional neural network CNN, a recurrent neural network RNN or its variant LSTM / GRU), or other types of machine learning algorithms, such as support vector machines (SVM), random forests, etc. The specific choice depends on the characteristics of the data and the complexity of the diagnostic task.
[0026] In step S30, the computer system obtains the first implicit representation array from the previous step S20. These arrays serve as inputs to the vehicle diagnostic network to be adjusted, and the network generates corresponding inference diagnosis results by processing and calculating the input data. The inference diagnosis result is the predicted probability or label of one or more fault categories, indicating the vehicle state corresponding to the input data. For example, for engine fault diagnosis, the inference diagnosis result may be one of the labels such as "normal", "bearing wear", "low oil pressure", etc., or a distribution vector containing the probabilities of these fault categories. For example, assume that the first implicit representation array is a vector containing multiple features. These features may represent a certain pattern or anomaly in the engine vibration signal, and after being extracted and converted by the first implicit representation mining network, they are closer to the essential characteristics of the vehicle fault. After receiving this vector, the vehicle diagnostic network to be adjusted performs weighted summation, activation function processing and other operations on these features through its internal computing units (such as neurons of a neural network), and finally outputs the inference diagnosis result. For example, the output may be a probability distribution vector:
[0027] ;
[0028] Among them, each element represents the probability that the input data belongs to the corresponding fault category.
[0029] The computer system not only outputs the inference diagnosis results, but also compares these results with the known first prior information (i.e., the true fault label) to calculate the error or loss. Common loss functions include cross-entropy loss (Cross-Entropy Loss), mean squared error (MSE), etc. By minimizing this loss function, the computer system can adjust the parameters of the vehicle diagnostic network to be adjusted, optimize its performance, and make the inference diagnosis results more accurate. This process is usually implemented through the backpropagation algorithm, which updates the network weights according to the gradient of the loss function.
[0030] Step S40: Optimizing network configuration parameters of the vehicle diagnostic network to be calibrated based on the error between the inference diagnosis result and the first priori information to obtain a vehicle diagnostic network, where the vehicle diagnostic network is used to diagnose the sensor information corresponding to the second sensor type.
[0031] In an embodiment of the present application, by comparing the error between the inference diagnosis result and the known first prior information (i.e., the actual fault label), the configuration parameters of the vehicle diagnostic network to be adjusted are optimized until a satisfactory performance level is achieved, and finally a vehicle diagnostic network capable of processing the second sensor type sensor information is generated.
[0032] First, the computer system calculates the error between the inference diagnosis result and the first prior information. This is usually achieved by defining a loss function, which quantifies the difference between the model prediction and the true label. In classification tasks, commonly used loss functions include cross-entropy loss, negative log-likelihood loss, etc. After obtaining the error, the computer system uses optimization algorithms (such as gradient descent method, Adam optimizer, etc.) to adjust the configuration parameters (such as weights and biases) of the vehicle diagnostic network to be adjusted to minimize the loss function. The optimization algorithm updates the parameter value according to the gradient of the loss function with respect to the network parameters, gradually approaching the global optimal solution.
[0033] After each parameter update, the computer system re-evaluates the performance of the model. This is usually done by calculating the error on an independent validation set to ensure that the model has good generalization ability. If the error is still large, the system will continue to iterate steps S30 and S40 until the error is reduced to an acceptable level or a predetermined number of iterations is reached.
[0034] After multiple iterations of optimization, the performance of the vehicle diagnostic network to be tuned gradually improved, and eventually became a vehicle diagnostic network that can accurately diagnose vehicle faults. This network can not only effectively diagnose the sensor data of the first sensor type (although this is achieved through implicit representation mining), but also because it has learned the fault feature representation related to the second sensor type data, it also has the ability to diagnose the sensor information of the second sensor type.
[0035] Through the training process of steps S10 to S30, the computer system has been able to generate inference diagnosis results using the first sensor type data. In step S40, the computer system optimizes the vehicle diagnostic network to be adjusted by comparing the errors between these inference diagnosis results and the real fault labels (i.e., the first prior information). After multiple iterations of optimization, a vehicle diagnostic network with excellent performance is obtained. This network can not only process the first sensor type data such as engine vibration and sound signals to diagnose faults (although these data have weak information representation capabilities in their original form), but also because it has learned the fault feature representation related to the second sensor type data such as oil pressure and speed, it can also directly diagnose the second sensor type data accurately. This makes the on-board intelligent diagnostic system more flexible and powerful in practical applications. Based on the aforementioned steps, the embodiment of the present application uses the first implicit representation mining network to process the first example sensor data and project it into the target implicit representation domain, which is associated with the sensor information of the second sensor type. Through implicit representation mining, the data of the first sensor type is converted into a form that expresses fault characteristics at a higher level, which are highly correlated with the characteristics in the target domain (the implicit representation domain of the second sensor type data). This solves the problem that the first sensor type data has weak information representation ability in its original form, and makes it possible to use this information to train the vehicle diagnosis network in the future.
[0036] The first implicit representation array obtained by implicit representation mining is input into the vehicle diagnostic network to be adjusted, and fault diagnosis is performed to obtain the inference diagnosis result. The performance of the vehicle diagnostic network to be adjusted in processing the converted first sensor type data is verified. By comparing the inference diagnosis result with the first prior information, the diagnostic ability of the network can be preliminarily evaluated. According to the error between the inference diagnosis result and the first prior information, the configuration parameters of the vehicle diagnostic network to be adjusted are adjusted by the optimization algorithm to minimize the error. Through the continuous iterative optimization process, the performance of the vehicle diagnostic network to be adjusted is gradually improved, and finally converges to a stable and accurate model state. This model can not only accurately process the data of the first sensor type (after implicit representation mining), but also indirectly has the ability to process the second sensor type data because it is associated with the characteristics of the second sensor type data during the implicit representation mining process. This ability makes the trained vehicle diagnostic network more flexible and powerful in practical applications, and can adapt to different types of sensor data input, improving the accuracy and comprehensiveness of fault diagnosis.
[0037] In summary, by collecting the first sensor type data with real labels and converting it into the implicit representation domain associated with the second sensor type data using implicit representation mining technology, and then training and optimizing the vehicle diagnostic network, we finally get an on-board intelligent diagnostic model that can handle multiple sensor data types with high accuracy and comprehensiveness.
[0038] As an implementation mode, the method further includes an application process of a vehicle diagnostic network, specifically including:
[0039] Step S50: deploying the vehicle diagnostic network on the vehicle terminal, and acquiring to-be-diagnosed sensor data corresponding to the second sensor type based on the vehicle terminal;
[0040] Step S60: extracting features from the sensor data to be diagnosed based on the second implicit representation mining network to obtain a second implicit representation array, wherein the second implicit representation mining network is used to project sensor information corresponding to the second sensor type into a target implicit representation domain;
[0041] Step S70: performing fault diagnosis based on the vehicle diagnostic network according to the second implicit representation array to obtain a fault diagnosis result of the sensor data to be diagnosed.
[0042] In step S50, the computer system is connected to the sensor interface through the vehicle network or directly to obtain sensor data corresponding to the second sensor type on the vehicle in real time or on demand. Such data usually has a higher information representation capability and is crucial for accurate diagnosis of vehicle faults.
[0043] For example, suppose the vehicle is equipped with a variety of high-precision sensors for monitoring various operating parameters of the engine, such as oil pressure sensor, speed sensor, temperature sensor, etc. These sensors constitute the data source of the second sensor type. When the vehicle is running, the computer system periodically or triggers to obtain real-time data from these sensors as sensor data to be diagnosed.
[0044] In step S60, the computer system processes the sensor data to be diagnosed using the pre-trained second implicit representation mining network, and converts it from the original representation domain to the target implicit representation domain. The second implicit representation mining network is a complex machine learning model, which may be a neural network based on deep learning, such as an autoencoder or a variational autoencoder (VAE). The network is trained to capture the intrinsic features of the second sensor type data and map it to the same implicit representation domain as the first sensor type data after the first implicit representation mining. The sensor data to be diagnosed is input into the second implicit representation mining network, and feature extraction and abstraction are performed through the computing units (such as convolutional layers, pooling layers, fully connected layers, etc.) inside the network, and finally the second implicit representation array is output. This array contains the key features of the data to be diagnosed in the target implicit representation domain, which are crucial for subsequent fault diagnosis. Assume that the output is a vector containing multiple features, such as oil pressure features, speed features, temperature features, etc., which are represented in the form of numerical values after being processed and converted by the network to form the second implicit representation array.
[0045] Step S70 performs fault diagnosis based on the vehicle diagnostic network according to the second implicit representation array. The computer system uses the second implicit representation array as input and passes it to the trained vehicle diagnostic network. The network uses the learned mapping relationship between fault features and fault categories to classify or regress the data to be diagnosed and output the fault diagnosis results. The vehicle diagnostic network is a machine learning model trained in steps S10 to S40, which can accurately identify different types of vehicle faults. The network may be a multi-classification neural network, each output node corresponds to a fault category, and the output value represents the probability that the input data belongs to this category. The second implicit representation array is input into the vehicle diagnostic network, and the network obtains the predicted probability of each fault category through calculation, and determines the final fault category according to certain decision rules (such as the maximum probability principle).
[0046] Assume that the network outputs a probability distribution vector, which indicates the probability that the input data belongs to different fault categories. For example, the output vector may be [0.1, 0.8, 0.1], which means that the input data has an 80% probability of belonging to the "abnormal oil pressure" fault category, while the probability of belonging to other fault categories is lower. Therefore, the fault diagnosis result finally output by the computer system is "abnormal oil pressure".
[0047] As an implementation mode, the first implicit representation mining network includes a first sensor type implicit representation mining network and a second sensor type implicit representation mining network, the first sensor type implicit representation mining network is used to project sensor information corresponding to the first sensor type to a target implicit representation domain, and the second sensor type implicit representation mining network is used to project sensor information corresponding to the second sensor type to a target implicit representation domain. Based on this, step S20, based on the first implicit representation mining network, performs implicit representation mining on the first example sensor data to obtain a first implicit representation array, including:
[0048] Performing implicit representation mining on the first example sensor data based on the first sensor type implicit representation mining network to obtain a first implicit representation array;
[0049] The method also includes:
[0050] Step S50: deploying the vehicle diagnostic network on the vehicle terminal, and acquiring to-be-diagnosed sensor data corresponding to the second sensor type based on the vehicle terminal;
[0051] Step S60: performing implicit representation mining on the sensor data to be diagnosed based on the second sensor type implicit representation mining network to obtain a second implicit representation array;
[0052] Step S70: performing fault diagnosis based on the vehicle diagnostic network according to the second implicit representation array to obtain a fault diagnosis result of the sensor data to be diagnosed.
[0053] The first sensing type implicit representation mining network is a complex machine learning model, usually built on a deep learning framework. The network is specifically designed to process first sensing type data (such as engine vibration signals, sound signals, etc.), which are relatively weak in information representation capabilities but easy to obtain and low-cost. The network may use a convolutional neural network (CNN), an autoencoder, or its variants as the core architecture. By stacking multiple convolutional layers, pooling layers, and nonlinear activation layers, the network can automatically extract high-level feature representations from the raw sensor data. These features reduce the dimensionality and redundancy of the data while maintaining the key information of the data, providing convenience for subsequent processing.
[0054] In step S20, the computer system inputs the first example sensor data into the first sensor type implicit representation mining network. The network processes the input data layer by layer through the forward propagation process and finally outputs the first implicit representation array. This array contains the key features of the input data in the target implicit representation domain, which are crucial for subsequent fault diagnosis tasks.
[0055] Assuming that the first example sensor data is the vibration signal of the engine, after being processed by the first sensor type implicit representation mining network, the output first implicit representation array may contain vibration frequency features, amplitude features, waveform complexity features, etc. These features are represented in the form of numerical vectors, such as [0.5, 120, 0.75] T , where each element corresponds to a specific vibrational feature.
[0056] Similar to the first sensor type implicit representation mining network, the second sensor type implicit representation mining network is also a complex model based on deep learning. However, it is specifically used to process the data of the second sensor type (such as high-precision sensor data such as oil pressure, speed, temperature, etc.), which has stronger information representation capabilities and is crucial for accurate diagnosis of vehicle faults. In step S60, when the computer system obtains the sensor data to be diagnosed corresponding to the second sensor type, it inputs these data into the second sensor type implicit representation mining network. Through a similar processing flow, the network outputs a second implicit representation array, which also contains the key features of the input data in the target implicit representation domain. Assuming that the sensor data to be diagnosed includes three parameters: oil pressure, speed, and temperature, after being processed by the second sensor type implicit representation mining network, the output second implicit representation array may contain features such as abnormal oil pressure, speed fluctuation, and excessive temperature. These features are also represented in the form of numerical vectors, such as [0.2, 0.05, 0.8]T, where each element corresponds to a specific operating parameter feature.
[0057] The process of steps S50 to S70 can refer to the above description and will not be repeated here. Technical Effects and Advantages
[0058] By introducing the first sensor type implicit representation mining network and the second sensor type implicit representation mining network, the two implicit representation mining networks map different types of data into a unified target implicit representation domain, making data from different sources more consistent and comparable in feature representation. This helps the vehicle diagnosis network better understand and identify various fault features. Since the vehicle diagnosis network is trained on data that has been mined through implicit representation, it can process and identify the feature representations of different sensor type data in the implicit representation domain. This capability makes the model more flexible and powerful in practical applications and can adapt to different types of sensor data input. Through the refined feature extraction and conversion process, the implicit representation mining network can capture more critical and subtle fault features in the data. These features provide the vehicle diagnosis network with richer and more accurate input information, thereby improving the accuracy of fault diagnosis. The first sensor type data is usually easy to obtain and low-cost, while the second sensor type data has strong information representation capabilities but high acquisition costs. By training the vehicle diagnosis network with the first sensor type data and processing the second sensor type data with the help of the implicit representation mining network, the system can reduce data acquisition costs while ensuring diagnostic accuracy. In summary, the introduction of the first sensor type implicit representation mining network and the second sensor type implicit representation mining network is an efficient and practical data processing strategy for the vehicle intelligent diagnosis system. It not only improves the consistency and comparability of data, but also enhances the generalization ability and diagnostic accuracy of the model, providing more reliable and accurate technical support for vehicle fault diagnosis.
[0059] As an implementation manner, the method further includes a training process of a first sensing type implicit representation mining network and a second sensing type implicit representation mining network, specifically including:
[0060] Step S1: Acquire multiple example sensor data tuples, the example sensor data tuples including a first component example sensor data corresponding to a first sensor type and a second component example sensor data corresponding to a second sensor type, the first component example sensor data and the second component example sensor data corresponding to an example sensor data tuple monitor the same fault type.
[0061] In step S1, the computer system obtains a plurality of example sensor data tuples. These tuples consist of two parts: a first component example sensor data and a second component example sensor data.
[0062] The first component example sensor data corresponds to the first sensor type, which is usually data collected by some relatively easy-to-obtain and low-cost sensors. Taking the engine as an example, the data of the first sensor type may include the vibration signal, sound signal, etc. of the engine. Although the information representation ability of these data is relatively weak, they are sufficient to provide some preliminary signs of faults.
[0063] The second component example sensor data corresponds to the second sensor type, which is usually data collected by some high-precision, high-cost sensors. These data have stronger information representation capabilities and can more accurately reflect the operating status and potential failures of the engine. Taking the engine as an example, the data of the second sensor type may include key operating parameters such as oil pressure, speed, and temperature.
[0064] The first component example sensor data and the second component example sensor data in each example sensor data tuple are monitored for the same fault type, which means that during the data collection process, it is necessary to ensure the consistency of the two sets of data in time and space so that they can correspond to each other and jointly describe the same fault phenomenon.
[0065] In order to ensure the consistency and correspondence of the data, the computer system can perform synchronization operations during the data collection stage. For example, it can synchronize the data collection process of different sensors through a trigger signal or a timestamp to ensure that each pair of the first component example sensor data and the second component example sensor data are collected under the same conditions.
[0066] Step S1 provides basic data support for the subsequent implicit representation mining network training by collecting multiple example sensor data tuples. These data tuples ensure the correlation and consistency between the first sensor type data and the second sensor type data during the training process, so that the network can learn how to map different types of data into a unified target implicit representation domain.
[0067] Step S2: for a target example sensor data tuple in a plurality of example sensor data tuples, based on an initial first sensor type implicit representation mining network, perform implicit representation mining on a first component example sensor data in the target example sensor data tuple to obtain a first implicit representation subarray.
[0068] In step S2, the computer system selects a target example sensor data tuple from a plurality of example sensor data tuples. The tuple consists of a first component example sensor data and a second component example sensor data, and both monitor the same fault type.
[0069] Next, the computer system processes the first component of the target example sensor data binary using the initial first sensor type implicit representation mining network. This network is a pre-designed machine learning model, usually built based on a deep learning framework, such as a convolutional neural network (CNN), an autoencoder, or a variational autoencoder (VAE).
[0070] Taking the autoencoder as an example, the initial first sensor type implicit representation mining network may include two parts: an encoder and a decoder. The encoder is responsible for compressing the input first component example sensor data (such as a vibration signal or a sound signal) into a low-dimensional implicit representation, namely the first sub-implicit representation array. The decoder attempts to reconstruct the original input data from this implicit representation, although in actual applications the decoder part may not be directly used for subsequent diagnostic tasks.
[0071] Assume that the first component of the input sensor data is the vibration signal of the engine under specific working conditions. This signal can be represented as a series of time series data or spectrum features. When these data are input into the initial first sensor type implicit representation mining network, the network extracts the features of the data layer by layer through multiple layers of nonlinear transformations (such as convolution, pooling, activation, etc.).
[0072] During the encoding process, the network gradually maps the high-dimensional feature space of the input data to a low-dimensional implicit representation space to form the first sub-implicit representation array. Each element in this array represents a feature dimension of the input data in the implicit representation domain. These features have higher information content and discrimination for subsequent fault diagnosis tasks.
[0073] Through step S2, the computer system uses the initial first sensor type implicit representation mining network to effectively extract and transform the first component example sensor data, and generates a first sub-implicit representation array containing key fault features. This process provides important data support for subsequent network training and fault diagnosis tasks, allowing the system to more accurately identify potential engine faults.
[0074] Step S3: Based on the initial second sensor type implicit representation mining network, the second component example sensor data in the target example sensor data tuple is subjected to implicit representation mining to obtain a second implicit representation subarray.
[0075] In step S3, the computer system processes the second component example sensor data in the target example sensor data tuple using the initial second sensor type implicit representation mining network. Similar to step S2, the network here is also a pre-designed machine learning model, but it is specifically optimized for the second sensor type data.
[0076] The specific implementation of the initial second sensor type implicit representation mining network may vary depending on the data type and diagnostic requirements, but it is usually built based on a deep learning framework. For example, for multi-dimensional time series data (such as records of oil pressure and speed changes over time), the network may use a recurrent neural network (RNN) variant such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to capture the temporal dependencies in the data.
[0077] When the second component example sensor data (such as oil pressure and speed data) is input into the initial second sensor type implicit representation mining network, the network extracts the intrinsic features of the data through multiple layers of nonlinear transformation. These features may include data change trends, periodic patterns, abnormal fluctuations, etc., which are of great significance for diagnosing engine faults.
[0078] After network processing, the second component example sensor data is converted into a low-dimensional second sub-implicit representation array. Each element in this array represents a feature dimension of the input data in the implicit representation domain, and these features will play a key role in the subsequent diagnosis task.
[0079] Through step S3, the computer system successfully converts the data of the second sensor type (such as oil pressure, speed and other data collected by high-precision sensors) into a low-dimensional feature array in the implicit representation domain. This process not only reduces the dimension and redundancy of the data, but also retains the key information in the data that is crucial for fault diagnosis. This provides strong data support for subsequent vehicle diagnostic network training and real-time fault diagnosis. At the same time, since the processing processes of the two implicit representation mining networks (the first sensor type and the second sensor type) are independent of each other but have the same goal (i.e., mapping different types of data into a unified implicit representation domain), there is a certain complementarity between them, which jointly improves the performance and accuracy of the entire intelligent diagnosis system.
[0080] Step S4: Determine a plurality of example sensor data pairs as target example sensor data pairs respectively, and obtain a plurality of first sub-implicit representation arrays and a plurality of second sub-implicit representation arrays.
[0081] In step S4, the computer system initializes an empty data set (or data structure) for storing all generated sub-implicit representation arrays. Then, it traverses all example sensor data tuples. For each set of example sensor data tuples, the computer system inputs its first component example sensor data (first sensor type) and second component example sensor data (second sensor type) into the corresponding implicit representation mining network. For the first sensor type data, the initial first sensor type implicit representation mining network is used for processing to generate a first sub-implicit representation array. For the second sensor type data, the initial second sensor type implicit representation mining network is used for processing to generate a second sub-implicit representation array.
[0082] The generated first and second implicit sub-representation arrays are associated with the corresponding example sensor data tuples (via index, label, or other identification information) and added to the previously initialized dataset. Similarly, for subsequent tuples, the above process is repeated until all tuples are processed. Finally, the computer system will obtain a dataset containing 200 implicit sub-representation arrays (two sub-arrays for each tuple), which will be used for subsequent network optimization and training processes.
[0083] Step S4 batch processes all example sensor data tuples and uses two independent implicit representation mining networks to generate the corresponding first and second implicit representation arrays, respectively, providing rich data support for subsequent model training and optimization. This process ensures the consistency and comparability of the data, allowing the model to learn the common features of different types of data in the implicit representation domain, thereby improving the accuracy and generalization ability of the diagnostic system.
[0084] Step S5: Optimize the network configuration parameters of the initial first sensor type implicit representation mining network and the network configuration parameters of the initial second sensor type implicit representation mining network along the direction of maximizing the commonality of the target sub-implicit representation array binary group and minimizing the commonality of the remaining sub-implicit representation array binary groups as the adjustment target, and obtain the first sensor type implicit representation mining network and the second sensor type implicit representation mining network. The first sub-implicit representation array and the second sub-implicit representation array included in the target sub-implicit representation array binary group are obtained based on the same example sensor data binary group, and the first sub-implicit representation array and the second sub-implicit representation array included in the remaining sub-implicit representation array binary groups are not obtained based on the same example sensor data binary group.
[0085] In step S5, the computer system adopts an iterative optimization strategy based on the idea of contrastive learning or metric learning, which aims to maximize the degree of commonality (similarity) between implicit representation arrays generated by the same fault type data, and minimize the degree of commonality (similarity) between implicit representation arrays generated by different fault type data.
[0086] For each pair of target sub-implicit representation array tuples (i.e., the first sub-implicit representation array and the second sub-implicit representation array) generated by the same example sensor data tuple, the computer system adjusts the network parameters so that the distance between the two arrays in the implicit representation domain is as close as possible. This is usually achieved by minimizing a certain distance metric between them (such as Euclidean distance, cosine similarity, etc.). For other non-target sub-implicit representation array tuples (i.e., array pairs not generated by the same example sensor data tuple), the computer system strives to increase the distance between them to reduce the confusion of different fault type data in the implicit representation domain.
[0087] In order to achieve the above optimization goals, the computer system can use optimization algorithms such as gradient descent and Adam optimizer to iteratively update network parameters. In each iteration, the system calculates the distance between all pairs of sub-implicit representation arrays under the current parameters, and then calculates the gradient according to the loss function (such as contrast loss, triple loss, etc.), and uses these gradients to update the network parameters. Taking contrast loss as an example, its calculation formula can be expressed as:
[0088] ;
[0089] Where N is the number of sample pairs, y i is the sample pair label (1 for identical binary pairs and 0 for different binary pairs), d i is the distance measure between the i-th sample pair, and m is the margin. For the target sub-array implicit representation, y i = 1, the system tries to minimize the distance d between them i ; For the remaining sub-implicit representation array tuples, y i =0, the system tries to ensure that the distance between them is greater than m, otherwise a loss will occur.
[0090] During the optimization process, the computer system continuously adjusts the weights and biases of the two implicit representation mining networks according to the current loss value until the loss function converges to a smaller value. At this point, the two networks can effectively map different types of data into a unified implicit representation domain, and the representation in this domain can accurately reflect the fault characteristics of the data.
[0091] Through the optimization process of step S5, the computer system successfully obtains the optimized first sensor type implicit representation mining network and the second sensor type implicit representation mining network. These two networks can generate implicit representation arrays with high consistency and discrimination, providing high-quality input data for the subsequent vehicle diagnosis network. This process not only improves the performance of the entire intelligent diagnosis system, but also enhances its adaptability and generalization ability to different types of data.
[0092] As an implementation mode, if the sensing information corresponding to the first sensing type is device environment data, and the sensing information corresponding to the second sensing type is device operation data, then the method further includes a process of determining the first prior information, specifically including:
[0093] Step S101: Obtain K second prior information and a first description instruction sample including a prior information reserved position, wherein the second prior information is used to indicate a fault diagnosis result corresponding to the sensor information of the second sensor type, and the first description instruction sample is used to guide a large-scale industrial model to generate device environment data corresponding to multiple device operation data groups, and multiple device operation data groups corresponding to the device environment data corresponding to the multiple device operation data groups are composed of target device operation data associated with the prior information corresponding to the prior information reserved position, and K≥1.
[0094] Second a priori information refers to the fault diagnosis results associated with the second sensor type (such as engine speed, oil pressure, oil temperature and other device operating data) that have been confirmed by professional diagnostic means. This information is accurate and reliable, and can directly reflect the current health status or fault type of the vehicle or equipment. For example, suppose there are the following second a priori information: "The engine speed is abnormal, there may be bearing wear", "The oil pressure is too low, which may cause insufficient lubrication", "The oil temperature is too high, and the cooling system needs to be checked".
[0095] The first description instruction sample is a predefined template or framework that contains specific instructions for guiding large-scale industrial models to generate data. This sample reserves one or more positions (which can be understood as "label slots") for subsequent filling of second prior information to generate device environment data associated with a specific fault. For example, assume that the first description instruction sample is as follows: "Generate engine vibration data associated with [fault type] to simulate the vibration characteristics generated by the engine under [fault type] conditions." Among them, "[fault type]" is a reserved position for prior information, waiting to be filled with specific second prior information later.
[0096] In the actual application scenario of the on-board intelligent diagnosis system, the computer system will obtain multiple second prior information from the fault diagnosis record or expert knowledge base. Then, it will use this prior information to fill in the reserved positions in the first description instruction sample to generate a specific first description instruction.
[0097] Specifically, the computer system retrieves K (K≥1) second prior information from the database, as shown in the above example. For each second prior information, the computer system fills it into the reserved position of the first description instruction sample to generate a specific first description instruction. For example, for the prior information "the engine speed is abnormal, and there may be bearing wear", the semantics of the generated first description instruction may be: generate engine vibration data associated with "the engine speed is abnormal, and there may be bearing wear" to simulate the vibration characteristics generated by the engine under this fault condition. Repeat the above process until all second prior information is converted into corresponding first description instructions.
[0098] These first description instructions filled with specific fault information are then input into the large-scale industrial model. This model is similar to the large language model, but is specifically designed to generate industrial-related data. Based on the instructions received, it simulates and generates device environment data associated with specific faults (such as engine vibration data, sound data, etc.). These data will be used as training samples for subsequent training and optimization of vehicle diagnostic networks. The large-scale industrial model is an existing advanced machine learning model that focuses on generating and processing relevant data in the industrial field. The large-scale industrial model uses advanced technologies such as deep learning and reinforcement learning to simulate and optimize industrial production processes to help companies improve production efficiency, reduce costs and improve product quality. It can process massive amounts of industrial data, including production data, equipment status data, supply chain data, etc. Through the analysis and learning of historical data, the model can achieve high-precision predictions of future production conditions, equipment failures, market demand, etc., and provide strong support for corporate decision-making. The model can be continuously optimized and adaptively adjusted based on new data input and feedback to adapt to the ever-changing industrial production environment. When building a large-scale industrial model, you can first collect data from equipment, sensors, etc., and clean, transform and normalize them to meet the needs of model training. The model structure is built based on deep learning frameworks (such as TensorFlow, PyTorch, etc.), and a large amount of labeled data is used for training. The model parameters are optimized through the back propagation algorithm. Useful features are extracted from the original data, and redundant and irrelevant features are removed through feature selection methods to improve the generalization ability of the model. The performance of the model is evaluated through methods such as cross-validation, and the optimal model is selected for deployment. After deployment, the model will receive new data input in real time and output prediction results or decision recommendations. With the continuous development of Industry 4.0 and intelligent manufacturing, large-scale industrial models will show the following trends: The model will have stronger self-learning and adaptive capabilities, and can continuously optimize its performance based on new data and feedback. It will no longer be limited to applications in a single field, but will realize cross-field data fusion and sharing, providing enterprises with more comprehensive solutions. With the improvement of computing power and the optimization of algorithms, the prediction and decision-making speed of the model will be faster and more accurate, meeting the needs of application scenarios with high real-time requirements.
[0099] Step S102: fill K pieces of second priori information into the reserved positions of the priori information in the first description instruction sample to obtain the first description instruction.
[0100] Step S102 combines the abstract fault diagnosis knowledge (ie, the second prior information) with the data generation template (ie, the first description instruction sample) to provide specific guidance for generating device environment data (eg, engine vibration data) associated with a specific fault.
[0101] Assume that the following three second prior information are obtained through step S101, which respectively indicate different fault diagnosis results: second prior information 1: "The engine speed is unstable, suspected ignition system failure"; second prior information 2: "The oil pressure fluctuation is abnormal, and the oil pump status needs to be checked"; second prior information 3: "The engine vibration is abnormal, and there may be bearing wear."
[0102] The first description instruction example mentioned in step S101 is a predefined template, which contains a reserved position for prior information (such as "[fault type]"), which is used to fill in specific second prior information later. For example, the example may be as follows: Generate engine vibration data associated with [fault type] to simulate the vibration characteristics under this fault condition.
[0103] In step S102, the computer system will process each second prior information one by one and fill it into the prior information reserved position of the first description instruction sample. The specific process is as follows: Process the first second prior information: fill "unstable engine speed, suspected ignition system failure" into the sample to obtain the first first description instruction: generate engine vibration data associated with "unstable engine speed, suspected ignition system failure" to simulate the vibration characteristics under this fault condition. Process the second second prior information: fill "abnormal oil pressure fluctuation, need to check the oil pump status" into the sample to obtain the second first description instruction: generate engine vibration data associated with "abnormal oil pressure fluctuation, need to check the oil pump status" to simulate the vibration characteristics under this fault condition. Process the third second prior information: fill "abnormal engine vibration, possible bearing wear" into the sample to obtain the third first description instruction: generate engine vibration data associated with "abnormal engine vibration, possible bearing wear" to simulate the vibration characteristics under this fault condition. These first description instructions filled with specific fault information will be input into the large-scale industrial model. Based on these instructions, the model can generate device environmental data such as engine vibration data associated with specific faults. This data is then used as training samples to train the vehicle diagnostic network to recognize and diagnose different types of engine faults.
[0104] Step S103: according to the first description instruction, first example sensor data is generated based on the large-scale industrial model, where the first example sensor data is device environment data corresponding to the device operation data group associated with K second prior information.
[0105] In step S103, the computer system performs the data generation process according to the following steps: the first description instruction generated in step S102 is passed as input to the large-scale industrial model. These instructions contain a detailed description of the fault type, such as "generate engine vibration data associated with 'unstable engine speed, suspected ignition system failure'". After receiving the instruction, the large-scale industrial model first parses the fault type information in the instruction. Then, it uses the internally learned data distribution and generation rules to simulate and generate device environment data corresponding to the fault type. This process may involve complex calculations and optimizations, such as gradually transforming random noise vectors through multi-layer neural networks so that they gradually approach the characteristic distribution of target data. Finally, the model outputs the generated first example sensor data. These data are device environment data closely related to the second prior information (i.e., the fault type), such as the vibration waveform and sound signal of the engine under a specific fault.
[0106] Step S104: taking the K second priori information as the first priori information of the first example sensing data.
[0107] Step S104 ensures that a clear association is established between the generated first example sensor data (such as engine vibration data) and its corresponding fault diagnosis result (ie, the second prior information), thereby providing accurate label information for subsequent vehicle diagnosis network training.
[0108] In the on-board intelligent diagnosis system, various sensors of the vehicle (such as vibration sensors, sound sensors, etc.) will collect device environment data in real time or regularly, and these data are crucial for diagnosing vehicle faults. However, it is often very difficult to extract effective fault features directly from these raw data. Therefore, by generating the first example sensor data associated with a specific fault and giving it an accurate fault diagnosis result (i.e., the first prior information), the subsequent model training process can be greatly simplified. In step S104, the computer system performs the following operations: for each set of first example sensor data (such as vibration waveform data associated with "engine bearing wear"), the computer system will find the corresponding second prior information (i.e., the fault diagnosis result of "engine bearing wear"). The computer system uses the found second prior information as the first prior information corresponding to the first example sensor data, that is, it gives it an accurate fault diagnosis label. Finally, the computer system stores these labeled first example sensor data in the training data set for subsequent training of the vehicle diagnosis network.
[0109] As an implementation mode, if the first description instruction example is also used to guide the generation of large-scale industrial models and the remaining prior information associated with the target device operation data, then step S104, using K second prior information as the first prior information of the first example sensor data, includes: using K second prior information and the remaining prior information as the first prior information of the first example sensor data.
[0110] In the on-board intelligent diagnosis system that integrates cloud-edge computing, the capabilities of large-scale industrial models are fully utilized to generate diverse training data. This data is not limited to device environmental data (such as engine vibration data), but can also include other prior information directly related to device operating data (such as engine speed, oil pressure, etc.). Together, this information constitutes a comprehensive description of vehicle faults, which helps to improve the accuracy and robustness of the diagnosis model.
[0111] When the first description instruction sample has the ability to generate multiple prior information, the execution process of step S104 is as follows: For each first description instruction, the large-scale industrial model not only generates the first example sensor data associated with a specific fault (such as a vibration waveform), but also generates the remaining prior information related to the fault, such as the probability of the fault occurring, the severity of the fault impact, etc. The computer system matches the generated first example sensor data with all relevant prior information (including K second prior information and the remaining prior information). Here, the K second prior information specifically refers to those that directly indicate the fault diagnosis results, while the remaining prior information may be a further description of the fault characteristics, the scope of influence, etc. Different from the traditional method of only using K second prior information as the first prior information, in this implementation method, the computer system uses the K second prior information and all the remaining generated prior information as the first prior information of the first example sensor data. This means that each first example sensor data is accompanied by a rich label set, which contains the multi-dimensional information required for fault diagnosis. Finally, the computer system stores these first example sensor data with rich labels for subsequent vehicle diagnosis network training. During training, the network learns how to extract features from device environmental data and predict a complete fault description based on these features (including fault diagnosis results and other prior information). For example, assume that the first description instruction example guides the model to generate the following data: First example sensor data: Vibration waveform associated with "engine bearing wear". Second prior information: "Engine bearing wear". Other prior information: Fault impact level is "moderate", and the estimated remaining service life is reduced by 30%.
[0112] In step S104, the computer system combines this information as the first prior information of the first example sensor data. Therefore, the complete label set of the data includes: "engine bearing wear", "fault impact level: medium" and "estimated remaining service life reduced by 30%".
[0113] As an implementation manner, step S30, performing fault diagnosis based on the vehicle diagnostic network to be calibrated according to the first implicit representation array to obtain an inference diagnosis result, may specifically include:
[0114] Step S31: According to the first implicit representation array, a transformation is performed based on a synchronous network to obtain a first synchronous implicit representation array, wherein the synchronous network is used to transform the implicit representation array from a target implicit representation domain to a network implicit representation domain, and the network implicit representation domain is an implicit representation domain that can be processed by a vehicle diagnostic network.
[0115] After the first sensor type data (such as engine vibration signals, sound signals, etc.) is processed by the first implicit representation mining network, a first implicit representation array is generated. This array is in the target implicit representation domain, which captures the key features of the original data, but these features do not directly meet the processing requirements of the vehicle diagnostic network. Therefore, it is necessary to transform these features from the target implicit representation domain to the network implicit representation domain that the vehicle diagnostic network can understand and process through the synchronization network.
[0116] The synchronous network is a machine learning model, which can be a neural network built based on a deep learning framework. This network is specially trained to learn the mapping relationship from the target implicit representation domain to the network implicit representation domain. It may contain multiple fully connected layers, convolutional layers, activation functions and other components. Through the combination and stacking of these components, nonlinear transformation and feature reorganization of input data are achieved. In step S31, the computer system passes the first implicit representation array as input to the synchronous network. After receiving this array, the synchronous network processes each element (or feature) in the array through a series of computing units (such as neurons, convolution kernels, etc.) inside it. This processing process may include weighted summation, nonlinear activation, feature dimensionality reduction or dimensionality increase, etc., aiming to convert the original features into a new feature set that better meets the needs of the vehicle diagnostic network.
[0117] After being transformed by the synchronous network, the first implicit representation array is converted to the first synchronized implicit representation array. This new array is in the network implicit representation domain, and each element of it represents a new feature representation of the original data in the network space. These new features are more compact and efficient, and match the internal structure of the vehicle diagnostic network, enabling the network to more accurately identify fault modes in the data. For example, suppose the first implicit representation array is z=[z1,z2,z3], where z1, z2, and z3 represent the frequency, amplitude, and waveform complexity of the vibration signal, respectively. When this array is input into the synchronous network, the network may generate the first synchronized implicit representation array through the following calculation process (simplified representation):
[0118] ;
[0119] Among them, W is the weight matrix, b is the bias vector, and f is the nonlinear activation function (such as ReLU, Sigmoid, etc.). After this calculation process, the output first synchronization implicit representation array z can be changed to [zsync, 1zsync, 2], where zsync, 1zsync, 2 are the new features after transformation, which more concisely and effectively describe the fault characteristics of the original data.
[0120] Step S32: performing fault diagnosis based on the vehicle diagnostic network to be calibrated according to the first synchronous implicit representation array to obtain an inference diagnosis result.
[0121] Vehicle sensors continuously collect various types of data (such as engine vibration signals, sound signals, speed, oil pressure, etc.). After a series of preprocessing and feature extraction steps, these data are converted into a form suitable for vehicle diagnostic network processing. In step S31, the first implicit representation array has been transformed by the synchronous network to obtain the first synchronous implicit representation array, which is located in the network implicit representation domain that can be understood and processed by the vehicle diagnostic network.
[0122] The vehicle diagnostic network to be adjusted is a complex machine learning model, usually built based on a deep learning framework. It may be a multi-layer neural network, containing various types of layer structures such as convolutional layers, pooling layers, and fully connected layers. These layers are stacked and combined to form a powerful ability to extract features, classify, or regress input data. During the training process, the vehicle diagnostic network to be adjusted has learned the mapping relationship from a specific type of input data (in this case, a synchronous implicit representation array) to the output diagnostic result. This mapping relationship is learned based on a large amount of labeled training data and can accurately reflect the intrinsic connection between the input data and the fault type.
[0123] In step S32, the computer system passes the first synchronous implicit representation array as input to the vehicle diagnostic network to be adjusted. After receiving the input data, the network calculates layer by layer through the forward propagation process and finally outputs the inference diagnosis result. This inference diagnosis result may be a probability distribution vector, in which each element corresponds to a possible fault type, and the value of the element represents the probability that the input data belongs to the fault type.
[0124] The method further includes a network application process based on step S31 and step S32, specifically including:
[0125] Step S50: Acquire the to-be-diagnosed sensing data corresponding to the second sensing type;
[0126] Step S60': performing implicit representation mining on the sensor data to be diagnosed based on the first implicit representation mining network to obtain a second implicit representation array;
[0127] Step S70': performing transformation based on the synchronous network according to the second implicit representation array to obtain a second synchronous implicit representation array;
[0128] Step S80': performing fault diagnosis based on the vehicle diagnosis network according to the second synchronous implicit representation array to obtain a fault diagnosis result of the sensor data to be diagnosed.
[0129] In the embodiment of the present application, steps S50 to S80' constitute a complete real-time fault diagnosis process, which fully utilizes the capabilities of the implicit representation mining network, synchronization network and vehicle diagnosis network mentioned above to achieve efficient and accurate diagnosis of the second sensor type data.
[0130] For step S50, reference may be made to the description of the aforementioned step S50, which will not be repeated here.
[0131] In step S60', after the computer system obtains the sensor data to be diagnosed, it inputs the data into the first implicit representation mining network. The network processes the input data through its internal complex computing units (such as convolutional layers, pooling layers, fully connected layers, etc.) to extract high-level feature representations. These features are encoded in the form of vectors or arrays in the implicit representation domain, that is, the second implicit representation array. For example, assuming that the sensor data to be diagnosed includes three parameters: engine speed, oil pressure and oil temperature, these data are organized into a feature vector x=[x rpm ,x pressure ,x temp ], where x rpm Indicates the speed value, x pressure Indicates the oil pressure value, x tempRepresents the oil temperature value. When this vector is input into the first implicit representation mining network, the network generates the second implicit representation array z′=[z1′,z2′,…z m ′], where m is the dimension of the implicit representation.
[0132] In step S70', the synchronization network plays the same role as in step S31, that is, transforming the data from the implicit representation domain to the network implicit representation domain so that the vehicle diagnostic network can process it. Since the second implicit representation array and the first implicit representation array are both in the same implicit representation domain (although they may be generated by different types of sensor data), the synchronization network can perform the same transformation processing on them.
[0133] The computer system inputs the second implicit representation array z′ into the synchronous network. The synchronous network transforms the input array through its internal learning parameters (such as weights and biases) to generate the second synchronous implicit representation array z sync ′=z sync,1 ′z sync,2 '...z sync,n ′].
[0134] This transformation process not only retains the key feature information of the input data, but also makes it more consistent with the input requirements of the vehicle diagnosis network. By synchronizing the network in this processing step, the computer system ensures that both the data in the training process and the data in the actual application can be processed by the vehicle diagnosis network in a unified form.
[0135] In step S80', when the computer system inputs the second synchronous implicit representation array zsync' into the vehicle diagnosis network, the network calculates the output value of each neuron layer by layer through the forward propagation process. In the output layer, the network gives the predicted probability or label of one or more fault categories based on the calculated feature representation.
[0136] For example, if the vehicle diagnosis network is a multi-classification network and is trained to identify five common engine faults (such as bearing wear, abnormal oil pressure, ignition failure, etc.), the output may be a probability distribution vector containing five elements, such as [p1, p2, p3, p4, p5], where each element pi represents the probability that the input data belongs to the i-th fault category.
[0137] After receiving the output of the vehicle diagnostic network, the computer system determines the final fault diagnosis result based on preset decision rules (such as the maximum probability principle). It can then present this result to the driver or maintenance personnel in a visual form, or send it to a remote server through the vehicle network for further analysis or recording. In the on-board intelligent diagnosis method that integrates cloud-edge computing, this diagnostic result may be uploaded to the cloud server in real time for storage and analysis. The cloud server can use big data analysis and machine learning algorithms to summarize and analyze the diagnostic results from multiple vehicles to discover potential failure modes and trends. At the same time, the computer system on the edge device issues an alarm to the driver or provides maintenance suggestions in real time based on the diagnostic results.
[0138] As an implementation mode, the method further includes a training process of a synchronization network, specifically including:
[0139] Step SA1: Acquire second example sensing data corresponding to the first sensing type, where the second example sensing data carries third priori information.
[0140] The second example sensor data, as the input data required for training the synchronization network, needs to be representative and diverse to cover as many fault modes and operating conditions as possible. For example, in an engine fault diagnosis scenario, the second example sensor data may include vibration signal samples under different speeds, loads, and temperatures.
[0141] During the data acquisition process, the computer system determines which data sources contain the required first sensor type data. This may involve operations such as interfacing with the vehicle sensor network, database queries, or remote API calls. The second example sensor data that meets the requirements is screened out from the selected data source. The screening conditions may include the timestamp of the data, the acquisition conditions (such as engine speed range, load level, etc.), and the known fault diagnosis results (i.e., the third prior information). For data samples that do not directly carry the third prior information, the computer system can rely on expert knowledge or historical fault records for manual labeling. The labeling process includes assigning one or more fault category labels to each data sample to indicate its corresponding fault type. The screened and labeled second example sensor data is stored in an appropriate location for subsequent processing and analysis.
[0142] For example, suppose that a synchronous network is being trained to diagnose engine bearing wear. In this scenario, the first sensor type data is the vibration signal of the engine. The computer system captures a set of vibration signal samples from the on-board sensor network as the second example sensor data. These samples cover the engine operating conditions under different speeds and load conditions, and each sample has its corresponding fault diagnosis result (i.e., whether there is bearing wear and the degree of wear). Step SA1 provides a solid foundation for the training of the synchronous network by retrieving, screening and annotating the second example sensor data corresponding to the first sensor type from various data sources. These data are not only representative and diverse, but also carry rich fault diagnosis information (third prior information), which helps to train a more accurate and robust synchronous network model.
[0143] Step SA2: According to the second example sensor data, perform implicit representation mining on the second example sensor data based on the first implicit representation mining network to obtain a third implicit representation array.
[0144] In the actual application of on-board intelligent diagnosis, the second example sensor data usually comes from the on-board sensor network. These data reflect the changes in various physical quantities of the vehicle under different operating conditions, such as vibration, sound, temperature, pressure, etc. Although these data come directly from the sensor, they often contain a lot of redundant information and noise, and may not work well when used directly for fault diagnosis. Therefore, it is necessary to preprocess and extract features from these raw data through the implicit representation mining network.
[0145] The first implicit representation mining network is a specially trained machine learning model that can map the input raw sensor data into a higher-level implicit representation domain. In this domain, the key features of the data are effectively extracted and represented, while redundant information and noise are greatly reduced. This implicit representation not only retains the important information in the original data, but also makes subsequent data processing and fault diagnosis more efficient and accurate.
[0146] In step SA2, the computer system performs the following operations: passing the second example sensor data as input to the first implicit representation mining network. These data may be a series of time series data, image data, or other types of sensor data, depending on the type and configuration of the vehicle sensor. The first implicit representation mining network processes the input data through its internal complex computing units (such as convolutional layers, pooling layers, fully connected layers, etc.). These computing units extract key features from the data through operations such as weighted summation and nonlinear activation, and encode them into the form of implicit representation arrays. After feature extraction, the first implicit representation mining network outputs a third implicit representation array. This array is located in the implicit representation domain, which captures the core features and structural information of the input data, and provides a more compact and effective data representation for subsequent data processing and fault diagnosis.
[0147] For example, suppose the second example sensor data is a vibration signal of an engine, which is represented as a one-dimensional time series array x=[x1,x2,…,xn], where each element represents the vibration amplitude at a different time point. When this array is input into the first implicit representation mining network, the network extracts key features (such as frequency components, amplitude changes, etc.) in the vibration signal through a series of convolution and pooling operations, and generates a third implicit representation array z=[z1,z2,…,zm], where m may be smaller than n, because the implicit representation array usually has a lower dimension and higher information density.
[0148] Step SA2 successfully converts the original sensor data into a more compact and effective implicit representation array by using the first implicit representation mining network to perform implicit representation mining on the second example sensor data. This conversion process not only retains the key feature information in the data, but also facilitates subsequent data processing and fault diagnosis.
[0149] Step SA3: According to the third implicit representation array, transformation is performed based on the initial synchronization network to obtain an inference synchronization implicit representation array.
[0150] The third implicit representation array is obtained by performing feature extraction and implicit representation mining on the second example sensor data through the first implicit representation mining network. These arrays are in the implicit representation domain. Although they capture the key features of the original sensor data, they are not directly applicable to the input requirements of the vehicle diagnosis network. Therefore, it is necessary to transform these implicit representation arrays into the network implicit representation domain through the initial synchronization network to generate the inference synchronization implicit representation array that the vehicle diagnosis network can process.
[0151] The initial synchronization network is a machine learning model that has not been fully trained. Its goal is to map the input data from the implicit representation domain to the network implicit representation domain. This mapping process requires learning the complex relationship between the input data and the output data to ensure that the transformed data can retain the key information of the original data and meet the input requirements of the vehicle diagnosis network.
[0152] In the on-board intelligent diagnosis method that integrates cloud-edge computing, the initial synchronization network may be deployed on the edge device to quickly respond to real-time diagnosis needs. At the same time, it can also interact with the data center in the cloud to optimize the model performance by utilizing the rich computing resources and big data support in the cloud.
[0153] In step SA3, the computer system performs the following operations: passing the third implicit representation array as input to the initial synchronization network. These arrays may contain multiple feature dimensions, each of which represents a specific aspect of the original sensor data in the implicit representation domain. After receiving the input array, the initial synchronization network transforms the array through its internal multi-layer neural network structure (such as fully connected layers, convolutional layers, etc.). This transformation process may include multiple operations such as feature reorganization, dimensionality reduction, dimensionality increase, nonlinear activation, etc., with the purpose of mapping the input data to a new representation space, namely the network implicit representation domain. After the network transformation, the initial synchronization network outputs the inference synchronization implicit representation array. This array is located in the network implicit representation domain, which retains the core features of the input data and is more in line with the input requirements of the vehicle diagnostic network.
[0154] Step SA3 transforms the third implicit representation array through the initial synchronization network, successfully mapping the data from the implicit representation domain to the network implicit representation domain. This transformation process not only retains the key feature information of the original data, but also ensures that the transformed data can be efficiently processed by the vehicle diagnosis network.
[0155] Step SA4: According to the inference synchronous implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain a fault diagnosis result.
[0156] The vehicle diagnostic network to be calibrated is a machine learning model that has been preliminarily trained but has not yet reached optimal performance. The model is capable of receiving input data in a specific format (such as an implicit representation array) and outputting corresponding fault diagnosis results. In step SA4, the inference synchronous implicit representation array is passed as input data to the vehicle diagnostic network to be calibrated to test its performance when processing synchronous network output data.
[0157] In step SA4, the computer system performs the following operations: passing the inference synchronous implicit representation array as input to the vehicle diagnostic network to be adjusted. This array is obtained by transforming the third implicit representation array through the initial synchronous network in step SA3. It is located in the network implicit representation domain and meets the input requirements of the vehicle diagnostic network to be adjusted. After receiving the input array, the vehicle diagnostic network to be adjusted processes the array through its internal complex calculation logic (which may include multi-layer neural networks, convolutional layers, pooling layers, etc.). The network uses the learned mapping relationship to map the input data to the fault diagnosis result space and outputs the predicted probability or label of one or more fault categories. The computer system receives the output of the vehicle diagnostic network to be adjusted, that is, the fault diagnosis result. This result may be a probability distribution vector (indicating the probability that the input data belongs to different fault categories) or a specific fault category label.
[0158] For example, suppose the vehicle diagnostic network to be tuned is a multi-classification neural network used to identify five common engine fault types (such as bearing wear, oil pressure abnormality, ignition fault, etc.). When the inference is synchronized, the implicit representation array z sync When it is input into the network, the network outputs a probability distribution vector after forward propagation calculation, such as [0.05, 0.85, 0.03, 0.02, 0.05]. Based on this vector, the computer system can determine that the most likely fault category is "abnormal oil pressure" (assuming that the corresponding probability is the highest) and output it as the fault diagnosis result.
[0159] After step SA4, the computer system can evaluate the performance of the vehicle diagnostic network to be adjusted based on the error between the fault diagnosis result and the known true fault label (i.e., the third prior information), and optimize the parameters of the synchronization network accordingly. This optimization process may involve optimization techniques such as back propagation algorithm and gradient descent method, aiming to minimize the error and improve the mapping ability of the synchronization network.
[0160] Step SA5: According to the error between the fault diagnosis result and the third priori information, the network configuration parameters of the vehicle diagnosis network to be adjusted are fixed, the network configuration parameters of the initial synchronization network are optimized, and the synchronization network is obtained.
[0161] The third prior information represents the known real fault condition, while the fault diagnosis result is the prediction result generated by the vehicle diagnosis network to be adjusted according to the inference synchronous implicit representation array. In this context, step SA5 optimizes the performance of the synchronous network by comparing the error between the two.
[0162] In step SA5, the computer system performs the following operations: First, the computer system calculates the error between the fault diagnosis result and the third prior information. This error can be quantified by defining a loss function, which measures the difference between the model prediction and the true label. Common loss functions include cross-entropy loss (Cross-Entropy Loss), mean squared error (MSE), etc. Before performing parameter optimization, the computer system fixes the network configuration parameters of the vehicle diagnostic network to be adjusted. This is because in the current optimization process, the focus is on the performance improvement of the synchronization network, not the vehicle diagnostic network to be adjusted. Therefore, the adjustment of the vehicle diagnostic network to be adjusted is not considered for the time being. Next, the computer system uses an optimization algorithm (such as gradient descent method, Adam optimizer, etc.) to adjust the configuration parameters (such as weights and biases) of the initial synchronization network to minimize the error function. The optimization algorithm updates the parameter value according to the gradient of the error function with respect to the network parameters, gradually approaching the global optimal solution. After each parameter update, the computer system re-evaluates the performance of the synchronization network. This is usually done by calculating the error on the validation set to ensure that the model has good generalization ability. If the error is still large, the process of error calculation, parameter optimization and model evaluation will continue to be iterated until the error is reduced to an acceptable level or the predetermined number of iterations is reached. For example, assume that the vehicle diagnostic network to be adjusted is already able to output relatively accurate fault diagnosis results, but the initial synchronization network has a certain error when transforming the third implicit representation array into the inference synchronization implicit representation array. In step SA5, the computer system calculates the error between the fault diagnosis result and the third prior information and finds that the mapping ability of the synchronization network is weak in certain feature dimensions. Therefore, the system uses optimization algorithms such as gradient descent to adjust the weights and biases of the synchronization network to enhance the mapping ability of these feature dimensions. After multiple iterative optimizations, the performance of the synchronization network has been significantly improved and the error has been significantly reduced.
[0163] By comparing the error between the fault diagnosis result and the third prior information, the network configuration parameters of the vehicle diagnostic network to be adjusted are fixed, and the network configuration parameters of the initial synchronization network are optimized, and finally the optimized synchronization network is obtained. This process not only improves the data transformation capability of the synchronization network, but also provides more accurate and reliable input data for subsequent on-board intelligent diagnosis. In the on-board intelligent diagnosis method integrating cloud-edge computing, this step is an indispensable part of achieving efficient and accurate fault diagnosis.
[0164] As an embodiment, the method further includes:
[0165] Step Sa1: obtaining a second description instruction sample and a synchronization command, wherein the second description instruction sample includes an implicit representation array reserved position and a synchronization command reserved position, and the synchronization command is used to guide the vehicle diagnosis network to be adjusted to generate a fault diagnosis result;
[0166] Step Sa2: filling the inference synchronization implicit representation array into the reserved position of the implicit representation array in the second description instruction sample, and filling the synchronization command into the reserved position of the synchronization command in the second description instruction sample, to obtain the second description instruction;
[0167] Based on steps Sa1~Sa2, step SA4, according to the inference synchronous implicit representation array, performs fault diagnosis based on the vehicle diagnostic network to be calibrated to obtain a fault diagnosis result, which may include: according to the second description instruction, performs fault diagnosis based on the vehicle diagnostic network to be calibrated to obtain a fault diagnosis result.
[0168] The second description instruction sample is a template or framework for guiding how to combine data and instructions to generate a complete diagnostic request. It contains specific placeholders or reserved positions for subsequent filling of specific data and commands. In this embodiment, the second description instruction sample includes at least two reserved positions:
[0169] Implicit representation array reserved position: used to store the implicit representation array to be processed, here is the inference synchronization implicit representation array. This array is the result of data preprocessing and feature extraction, and contains key information for fault diagnosis.
[0170] Synchronous command reserved position: used to store synchronous commands. These commands instruct the vehicle diagnostic network to be adjusted on how to process the input implicit representation array and generate corresponding fault diagnosis results.
[0171] Synchronous commands are a set of predefined instructions or operations that are designed to be compatible with the vehicle diagnostic network to be tuned and can trigger the network to execute specific diagnostic logic. These commands may include selecting a specific diagnostic mode, enabling or disabling certain feature processing modules, adjusting diagnostic thresholds, etc.
[0172] Assume that you are developing an onboard intelligent diagnostic system that can identify multiple engine faults. The second description instruction example may be a JSON format template as follows:
[0173] json
[0174] {
[0175] "dataPlaceholder": [], / / implicitly represents the reserved position in the array
[0176] "command": "" / / Reserved position for synchronization commands
[0177] }
[0178] In actual applications, this template will be filled with specific inference synchronization implicit representation arrays and synchronization commands to form a complete diagnostic request.
[0179] In step Sa2, the computer system fills the inference synchronization implicit representation array generated in step SA3 into the reserved position of the implicit representation array of the second description instruction sample, and fills the predefined synchronization command into the reserved position of the synchronization command. This process is actually building a complete and executable diagnostic request. For example, continuing the above example of engine fault identification, assume that the inference synchronization implicit representation array is z.
[0180] sync=[0.2,−0.1,0.3] T , and the synchronization command is "diagnose_engine_faults". The second description instruction after filling may look like this:
[0181] json
[0182] {
[0183] "dataPlaceholder": [0.2, -0.1, 0.3],
[0184] "command": "diagnose_engine_faults"
[0185] }
[0186] This command now contains all the information needed to calibrate the vehicle diagnostic network to perform fault diagnosis.
[0187] In the traditional step SA4, fault diagnosis is performed directly based on the inference synchronous implicit representation array. However, in this derivative implementation, step SA4 is optimized by introducing the concept of second description instruction sample and synchronization command. Specifically, the optimized step SA4 no longer relies solely on the inference synchronous implicit representation array, but performs fault diagnosis based on the filled second description instruction. The computer system passes this instruction as input to the vehicle diagnostic network to be adjusted. The network executes the corresponding diagnostic logic according to the data and commands in the instruction, and finally outputs the fault diagnosis result.
[0188] Back to the example of engine fault identification, when the computer system passes the filled second description instruction to the vehicle diagnostic network to be adjusted, the network first parses the data and commands in the instruction. Then, it selects the appropriate diagnostic mode and parameters according to the synchronization command "diagnose_engine_faults". Then, the network uses the inference synchronization implicit representation array zsync as input and generates fault diagnosis results through a complex internal calculation process (which may include forward propagation of a multi-layer neural network).
[0189] Assume that the result of the network output is a probability distribution vector, which represents the probability that the input data belongs to different fault categories. For example, the output vector may be [0.05, 0.1, 0.8, 0.03, 0.02]T, corresponding to the five categories of "normal", "slight wear", "severe wear", "overheating" and "other faults". Based on this output, the computer system can determine that the most likely fault category is "severe wear". By introducing the concepts of second description instruction samples and synchronization commands, this derivative implementation brings the following advantages: it is possible to dynamically construct diagnostic requests according to different diagnostic requirements without modifying the structure or parameters of the vehicle diagnostic network to be adjusted. By adding new synchronization commands or modifying the second description instruction samples, the system can easily support more diagnostic scenarios and fault types. When the interface of the vehicle diagnostic network to be adjusted changes (such as adding new input features, modifying the command format, etc.), the system administrator only needs to update the second description instruction samples and synchronization commands without large-scale reconstruction of the entire system. Passing data and commands to the vehicle diagnostic network to be adjusted in the form of instructions makes the entire diagnostic process clearer and controllable. This helps debugging and troubleshooting, and improves the testability of the system.
[0190] By introducing the concept of second description instruction sample and synchronization command, the data processing and fault diagnosis process of the on-board intelligent diagnosis system are optimized. It not only improves the flexibility and scalability of the system, but also enhances the maintainability and testability of the system.
[0191] As an implementation manner, step S40, based on the error between the inference diagnosis result and the first priori information, optimizing the network configuration parameters of the vehicle diagnostic network to be adjusted to obtain the vehicle diagnostic network, may specifically include:
[0192] Step S41: Optimizing the network configuration parameters of the vehicle diagnostic network to be calibrated based on the error between the inference diagnosis result and the first priori information to obtain the vehicle diagnostic network, and optimizing the network configuration parameters of the synchronization network based on the error between the inference diagnosis result and the first priori information to obtain an updated synchronization network.
[0193] The synchronization network is responsible for converting the implicit representation of the first sensor type data (obtained after processing by the first implicit representation mining network) into a format that can be efficiently processed by the vehicle diagnosis network. At the same time, the vehicle diagnosis network to be adjusted performs fault diagnosis based on these converted data. In order to improve the performance of the entire system, it is necessary to optimize the parameters of these two networks at the same time to ensure that they can work together to achieve higher diagnostic accuracy.
[0194] In step S40, the computer system calculates the error between the inference diagnosis result and the first prior information. This error reflects the performance of the current model (including the vehicle diagnostic network to be adjusted and the synchronization network) on the diagnostic task. Next, the system adopts a joint optimization strategy to adjust the parameters of the two networks at the same time to minimize this error. The computer system calculates the error by comparing the inference diagnosis result and the first prior information. This error can be measured by a variety of indicators, such as cross entropy loss, mean square error, etc. After obtaining the error, the computer system uses the back propagation algorithm and gradient descent method (or other optimizers, such as Adam) to simultaneously optimize the parameters of the vehicle diagnostic network to be adjusted and the synchronization network. The optimization process aims to find a set of parameters that minimizes the error function.
[0195] Vehicle diagnosis network to be tuned: By adjusting the weights and biases in its network layers, the network can more accurately learn fault features from the implicit representation output by the synchronous network and output correct fault diagnosis results.
[0196] Synchronize the network: By adjusting the connection weights and activation function parameters of its internal neurons, the network is able to generate an implicit representation array that is more suitable for the vehicle diagnostic network to be tuned. This step ensures the effective conversion of data between the implicit representation domain and the network implicit representation domain.
[0197] The optimization process is an iterative process. In each iteration, the computer system calculates the gradient based on the error value under the current parameters and updates the parameter value in the opposite direction of the gradient. This process is repeated until the error is reduced to an acceptable level or a predetermined number of iterations is reached.
[0198] After the optimization is complete, the computer system evaluates the updated vehicle diagnostic network and the synchronization network. This is usually done by calculating the error on an independent validation set to ensure that the model has good generalization ability.
[0199] For example, suppose that during the training process, the inference diagnosis result predicts that the probability of a sample being "bearing wear" is 0.8, and the first prior information indicates that the sample does belong to the "bearing wear" category. However, due to poor parameter settings of the current model, this prediction result may not be accurate enough. By calculating the error between the inference diagnosis result and the first prior information (such as cross entropy loss), the computer system identifies this problem and starts the optimization process.
[0200] During the optimization process, the computer system simultaneously adjusted the parameters of the vehicle diagnostic network and synchronization network to be adjusted. After multiple iterations, the model gradually learned how to extract key fault features from the implicit representation output by the synchronization network and output more accurate diagnostic results. Finally, when the error is reduced to a satisfactory level, the optimization process ends and the computer system obtains the updated vehicle diagnostic network and synchronization network.
[0201] Through this implementation in step S40, the computer system can simultaneously optimize the parameters of the vehicle diagnostic network to be adjusted and the synchronization network, thereby improving the performance of the entire on-board intelligent diagnostic system. This joint optimization strategy ensures that data can be processed efficiently and accurately in both representation conversion and fault diagnosis, providing a more reliable diagnostic service for the vehicle.
[0202] As an implementation mode, if the sensing information corresponding to the first sensing type is device environment data and the sensing information corresponding to the second sensing type is device operation data, then the second example sensing data includes device environment data corresponding to multiple device operation data groups, and the third prior information is global device environment data of the device operation data corresponding to the second example sensing data or the device environment data corresponding to multiple device operation data groups.
[0203] When the first sensor type corresponds to device environment data (such as engine vibration signals, sound signals, etc.), and the second sensor type corresponds to device operation data (such as high-precision parameters such as speed, oil pressure, temperature, etc.), an innovative implementation method takes into account the comprehensiveness and globality of the data. This approach specifically focuses on how to more effectively use the second sensor type data (device operation data) to enrich and enhance the understanding of the first sensor type data (device environment data), thereby improving the performance of the vehicle diagnostic network.
[0204] In a vehicle environment, device environmental data (such as vibration signals) often provide indirect clues about the health status of the device, but these data themselves may not be sufficient to accurately diagnose all types of faults. In contrast, device operating data (such as speed, oil pressure, etc.) usually reflects the operating status of the device more directly, but the acquisition cost is high and the processing is complex. This implementation aims to build a more comprehensive and accurate diagnostic model by combining these two types of data.
[0205] In this embodiment, the second example sensor data is not limited to a single device operation data point or time series, but includes multiple device operation data groups and their corresponding device environment data. Each device operation data group may represent the operating state of the vehicle under different working conditions, different loads or different time points. This design makes the second example sensor data rich in diversity and representativeness, and can more comprehensively cover possible failure modes and operating conditions.
[0206] Different from the traditional method of associating only a single data point with its corresponding label (i.e., prior information), the third prior information in this embodiment is defined as data related to the global state of the second example sensor data or its corresponding device operation data. Specifically, the "global device environment data" here refers to data or features extracted from multiple device operation data groups that can reflect the overall operation state or health trend of the equipment. These data or features may be obtained through methods such as statistical analysis, signal processing, or machine learning, and are used to provide more comprehensive contextual information for the second example sensor data. By combining multiple device operation data groups and their corresponding device environment data, the system can learn the changing patterns of device environment data under different operating conditions, thereby more accurately identifying potential failure modes. The introduction of global device environment data provides the system with a broader perspective to examine the operating state of the device. This helps the system capture those signs of failure or trends that are difficult to detect through a single data point.
[0207] During the training process, using this comprehensive data with a global perspective as prior information can guide the vehicle diagnosis network to learn more robust and generalized feature representations, thereby improving its diagnostic accuracy and reliability in practical applications.
[0208] As an embodiment, the method further includes:
[0209] Step SB1: obtaining core data of the first example sensor data;
[0210] Step SB2: Perform implicit representation mining on the core data to obtain a core implicit representation array;
[0211] Based on the above steps SB1 to SB2, step S30, according to the first implicit representation array, based on the vehicle diagnostic network to be calibrated, fault diagnosis is performed to obtain the inference diagnosis result, which may include: according to the first implicit representation array and the core implicit representation array, based on the vehicle diagnostic network to be calibrated, fault diagnosis is performed to obtain the inference diagnosis result.
[0212] In the on-board intelligent diagnosis system, the first type of sensor data (such as engine vibration signals, sound signals and other device environment data) usually contains a lot of information, but not all data are equally important for fault diagnosis. Core data refers to those data fragments or features that are most critical and most discriminating for fault diagnosis. These data may correspond to specific fault modes, operating conditions or abnormal phenomena.
[0213] The computer system performs a preliminary analysis on the first example sensor data to identify which data segments or features are most likely to be related to the fault. This may involve statistical analysis, signal processing or machine learning techniques, such as feature selection, principal component analysis, etc. Based on the analysis results, the computer system extracts core data from the first example sensor data. These core data may be a series of key time points, frequency components, waveform features, or other forms of feature vectors. For example, assuming that the first example sensor data is a vibration signal of an engine, the computer system finds through analysis that within a specific speed range, a certain frequency component of the vibration signal is significantly enhanced, and this phenomenon is highly correlated with bearing wear faults. Therefore, the computer system extracts this frequency component and its corresponding time point as core data. After extracting the core data, the computer system further processes the core data using the first implicit representation mining network (or a specially designed core data implicit representation mining model), and converts it into a core implicit representation array in the implicit representation domain through operations such as feature extraction, dimensionality reduction, and encoding. The core data implicit representation mining model can be a neural network built on a deep learning framework, such as autoencoder, variational autoencoder (VAE) or deep belief network (DBN), etc. These models can map the core data into a low-dimensional and information-rich implicit representation space by learning its intrinsic structure and feature representation.
[0214] After mining the implicit representation, the computer system obtains the core implicit representation array, which not only retains the key feature information of the core data, but also removes redundancy and noise through nonlinear transformation and encoding, making the subsequent diagnosis process more efficient and accurate.
[0215] For example, continuing the vibration signal example above, the computer system inputs the extracted core data (specific frequency components and their corresponding time points) into the core data implicit representation mining model. The model processes the core data through a series of computing units (such as convolutional layers, pooling layers, fully connected layers, etc.), and finally outputs the core implicit representation array. This array may be a vector containing multiple feature dimensions, each of which represents a specific aspect of the core data in the implicit representation domain.
[0216] In the optimized step S30 , the computer system no longer relies solely on the first implicit representation array for fault diagnosis, but combines the first implicit representation array with the core implicit representation array as inputs to the vehicle diagnostic network to be calibrated.
[0217] The computer system fuses the first implicit representation array and the core implicit representation array. This fusion can be a simple concatenation operation (i.e. directly connecting the two arrays into a longer array) or a more complex feature-level fusion (such as interacting and combining the features of the two arrays through attention mechanisms, feature crossover, etc.).
[0218] The fused data is input into the vehicle diagnostic network to be calibrated. The network processes the input data through its internal complex computational logic and finally outputs the inference diagnosis result. Since the core implicit representation array contains more critical and discriminative feature information, the fused data can improve the diagnostic accuracy of the vehicle diagnostic network to be calibrated.
[0219] For example, assume that the vehicle diagnostic network to be adjusted is a multi-classification neural network used to identify different types of engine faults. In the optimized step S30, the computer system concatenates the first implicit representation array and the core implicit representation array to generate a new input array. This array is then input into the vehicle diagnostic network to be adjusted, and the network outputs a probability distribution vector through forward propagation calculation, indicating the probability that the input data belongs to different fault categories. Based on this vector, the computer system can determine the final fault diagnosis result.
[0220] By introducing the concept of core data and mining its implicit representation, this derivative implementation not only enriches the input data of the vehicle diagnostic network to be adjusted, but also improves the diagnostic accuracy of the network through data fusion. In the on-board intelligent diagnosis scenario that integrates cloud-edge computing, this strategy can more effectively utilize the computing resources of the cloud and edge to achieve real-time, accurate and efficient fault diagnosis services. At the same time, since the extraction of core data and the implicit representation mining process can be completed in advance in the cloud, the computing burden on the edge is reduced, and the overall performance and response speed of the system are also improved.
[0221] As an embodiment, the method further includes:
[0222] Step SC1: obtaining a third description instruction sample and a detail optimization command, wherein the third description instruction sample includes an implicit representation array reserved position and a detail optimization command reserved position, and the detail optimization command is used to guide the vehicle diagnosis network to be adjusted to generate an inference diagnosis result;
[0223] Step SC2: Fill the first implicit representation array into the reserved position of the implicit representation array in the third description instruction sample, and fill the detail optimization command into the reserved position of the detail optimization command in the third description instruction sample to obtain the third description instruction;
[0224] Based on the above steps SC1 to SC2, step S30, according to the first implicit representation array, based on the vehicle diagnostic network to be adjusted, performs fault diagnosis to obtain the inference diagnosis result, which may include: according to the third description instruction, based on the vehicle diagnostic network to be adjusted, performs fault diagnosis to obtain the inference diagnosis result. The third description instruction sample is a structured template for guiding how to construct a complete instruction including the implicit representation array and the detail optimization command. It contains at least two reserved positions:
[0225] Implicit representation array reserved position: used to subsequently fill the first implicit representation array after implicit representation mining processing.
[0226] Detail optimization command reserved position: used to fill in commands to guide the vehicle diagnostic network to be tuned for detail optimization. These commands may involve adjusting a specific parameter of the diagnostic logic, enabling or disabling certain feature processing modules, etc.
[0227] Detailed optimization commands are a set of optimization instructions for the vehicle diagnostic network to be tuned, which can guide the network to adopt a more refined processing strategy when generating inference diagnosis results. These commands are designed according to specific diagnostic requirements and optimization goals to improve the accuracy and efficiency of diagnosis.
[0228] In step SC2, the computer system fills the first implicit representation array generated in step S20 into the reserved position of the implicit representation array of the third description instruction sample, and fills the predefined detail optimization command into the reserved position of the detail optimization command. This process actually constructs a complete diagnosis request containing specific data and optimization instructions.
[0229] For example, assume that the third description instruction example is a JSON format template as follows:
[0230] json
[0231] {
[0232] "implicitArrayPlaceholder": [], / / Implicit array placeholder
[0233] "detailOptimizationCommand": "" / / Detail optimization command reserved position
[0234] }
[0235] During the filling process, assume that the first implicit representation array is z=[z1,z2,…,zn]T and the detailed optimization command is "adjust_threshold_for_sensitivity". The third description instruction after filling may be as follows:
[0236] json
[0237] {
[0238] "implicitArrayPlaceholder": [z_1, z_2, ..., z_n],
[0239] "detailOptimizationCommand": "adjust_threshold_for_sensitivity"
[0240] }
[0241] In the optimized step S30, the computer system no longer performs fault diagnosis based solely on the first implicit representation array, but instead passes the third description instruction as input to the vehicle diagnostic network to be calibrated. After receiving the instruction, the network first parses the implicit representation array and the detail optimization command, and then generates an inference diagnosis result based on this information and the knowledge learned within the network.
[0242] By introducing the third description instruction sample and detail optimization command, this implementation enables the vehicle diagnostic network to be adjusted to receive more specific and refined guidance when performing fault diagnosis. This not only improves the accuracy and efficiency of diagnosis, but also enhances the flexibility and scalability of the system.
[0243] As an implementation manner, the third description instruction sample further includes a core data reserved location, and the method further includes:
[0244] Step SD1: Acquire core data of the first example sensor data;
[0245] Step SD2: Perform implicit representation mining on the core data to obtain a core implicit representation array;
[0246] Based on the above steps SD1 to SD2, step SC2, filling the first implicit representation array into the reserved position of the implicit representation array in the third description instruction sample, filling the detail optimization command into the reserved position of the detail optimization command in the third description instruction sample, to obtain the third description instruction, may include: filling the first implicit representation array into the reserved position of the implicit representation array in the third description instruction sample, filling the detail optimization command into the reserved position of the detail optimization command in the third description instruction sample, filling the core implicit representation array into the core data reserved position in the third description instruction sample, to obtain the third description instruction.
[0247] In the original third description instruction sample, in addition to the reserved positions for implicit representation arrays and detailed optimization command reserved positions, a new core data reserved position is added. This reserved position is used to store the core implicit representation array after implicit representation mining, so as to provide richer data support for the vehicle diagnostic network to be adjusted during the fault diagnosis process.
[0248] In step SD1, the computer system retrieves and obtains the core data of the first example sensor data (such as the vibration signal, sound signal, etc. of the engine) from the storage medium. Core data refers to those data segments or features that are most critical and distinguishing for fault diagnosis, which may correspond to specific fault modes, abnormal phenomena or critical operating conditions. For example, assuming that the first example sensor data is a segment of an engine vibration signal containing multiple vibration frequency components, the computer system finds through analysis that the amplitude of the vibration signal within a specific frequency range (such as the characteristic frequency related to bearing wear) is abnormally increased, and this phenomenon occurs in multiple fault samples. Therefore, the vibration signal within this specific frequency range is identified as core data.
[0249] In step SD2, the computer system further processes the core data using the first implicit representation mining network (or a specially designed core data implicit representation mining model). Through operations such as feature extraction, dimensionality reduction, and encoding, the core data is mapped to the implicit representation domain to generate a core implicit representation array. This array not only retains the key feature information of the core data, but also removes redundancy and noise through nonlinear transformation and encoding, making the subsequent diagnosis process more efficient and accurate.
[0250] After obtaining the first implicit representation array, the detail optimization command and the core implicit representation array, the computer system constructs the third description instruction according to the following steps: filling the first implicit representation array into the implicit representation array reserved position in the third description instruction sample.
[0251] Fill the detail optimization command into the detail optimization command reserved position in the third description instruction sample.
[0252] Fill the core implicit representation array into the core data reserved position in the third description instruction sample.
[0253] After the above filling process, the computer system generates a third description instruction containing complete information. This instruction not only contains the data to be processed (the first implicit representation array) and the diagnostic optimization instruction (detailed optimization command), but also provides an additional implicit representation of the core data (core implicit representation array), providing more comprehensive and in-depth data support for the vehicle diagnostic network to be adjusted.
[0254] In the on-board intelligent diagnostic system, when a bearing wear fault occurs in the vehicle engine, the computer system obtains the vibration signal of the engine through the sensor as the first example sensor data. Subsequently, the system identifies the vibration signal within a specific frequency range related to bearing wear as the core data, and performs implicit representation mining on it to obtain the core implicit representation array. At the same time, the system generates detailed optimization commands based on diagnostic requirements, such as adjusting the diagnostic threshold to increase sensitivity to bearing wear faults. Finally, the system combines the first implicit representation array, the detailed optimization command, and the core implicit representation array into a third description instruction, and passes it to the vehicle diagnostic network to be adjusted for fault diagnosis. After receiving the instruction, the network generates an inference diagnosis result based on the data and optimization instructions therein, and provides maintenance personnel with accurate fault location and repair suggestions.
[0255] By introducing the concepts of core data reserved locations and core implicit representation arrays, this derived implementation further improves the accuracy and reliability of the on-board intelligent diagnostic system and provides strong guarantees for the safe operation of the vehicle.
[0256] The present application embodiment provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.
[0257] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-described method is implemented.
Claims
1. A vehicle-mounted intelligent diagnosis method integrating cloud-edge computing, characterized in that: Applied to a cloud server, the method comprises: Acquire first example sensor data belonging to a first sensor type, wherein the first example sensor data carries first a priori information, and the first a priori information is used to indicate an actual diagnosis result of the first example sensor data; Performing implicit representation mining on the first example sensor data based on a first implicit representation mining network to obtain a first implicit representation array, wherein the first implicit representation mining network is used to project sensor information corresponding to the first sensor type to a target implicit representation domain, wherein the target implicit representation domain is an implicit representation domain where an implicit representation array corresponding to sensor information belonging to a second sensor type is located, wherein the first sensor type and the second sensor type are different data types, and an information representation capability of the sensor information corresponding to the first sensor type is less than an information representation capability of the sensor information corresponding to the second sensor type; According to the first implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result; Optimizing network configuration parameters of the vehicle diagnostic network to be calibrated according to an error between the inference diagnosis result and the first priori information to obtain a vehicle diagnostic network, wherein the vehicle diagnostic network is used to diagnose sensor information corresponding to the second sensor type; The method further comprises: Deploying the vehicle diagnostic network on a vehicle terminal; Acquiring to-be-diagnosed sensor data corresponding to the second sensor type based on the vehicle-mounted terminal; Performing feature extraction on the sensor data to be diagnosed based on a second implicit representation mining network to obtain a second implicit representation array, wherein the second implicit representation mining network is used to project the sensor information corresponding to the second sensor type into the target implicit representation domain; According to the second implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to obtain a fault diagnosis result of the sensor data to be diagnosed.
2. The method according to claim 1, characterized in that The first implicit representation mining network includes a first sensor type implicit representation mining network and a second sensor type implicit representation mining network, the first sensor type implicit representation mining network is used to project sensor information corresponding to the first sensor type to the target implicit representation domain, and the second sensor type implicit representation mining network is used to project sensor information corresponding to the second sensor type to the target implicit representation domain; The step of performing implicit representation mining on the first example sensor data based on a first implicit representation mining network to obtain a first implicit representation array includes: Performing implicit representation mining on the first example sensor data based on the first sensor type implicit representation mining network to obtain a first implicit representation array; The method further comprises: Acquiring to-be-diagnosed sensor data corresponding to the second sensor type; Performing implicit representation mining on the sensor data to be diagnosed based on the second sensor type implicit representation mining network to obtain a second implicit representation array; According to the second implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to obtain a fault diagnosis result of the sensor data to be diagnosed.
3. The method according to claim 2, characterized in that The method further comprises: Acquire a plurality of example sensor data tuples, wherein the example sensor data tuples include a first component example sensor data corresponding to the first sensor type and a second component example sensor data corresponding to the second sensor type, wherein the first component example sensor data and the second component example sensor data corresponding to one example sensor data tuple monitor the same fault type; For a target example sensor data tuple in the plurality of example sensor data tuples, performing implicit representation mining on a first component example sensor data in the target example sensor data tuple based on an initial first sensor type implicit representation mining network to obtain a first implicit representation subarray; Based on the initial second sensor type implicit representation mining network, the second component example sensor data in the target example sensor data tuple is subjected to implicit representation mining to obtain a second implicit representation subarray; Determine the plurality of example sensor data pairs as the target example sensor data pairs respectively, and obtain a plurality of the first implicit sub-arrays and a plurality of the second implicit sub-arrays; The network configuration parameters of the initial first sensor type implicit representation mining network and the network configuration parameters of the initial second sensor type implicit representation mining network are optimized along the direction of maximizing the commonality of the target sub-implicit representation array binary group and minimizing the commonality of the remaining sub-implicit representation array binary groups as the tuning target, and the first sensor type implicit representation mining network and the second sensor type implicit representation mining network are obtained. The first sub-implicit representation array and the second sub-implicit representation array included in the target sub-implicit representation array binary group are obtained based on the same example sensor data binary group, and the first sub-implicit representation array and the second sub-implicit representation array included in the remaining sub-implicit representation array binary groups are not obtained based on the same example sensor data binary group.
4. The method according to claim 1, characterized in that: If the sensing information corresponding to the first sensing type is device environment data, and the sensing information corresponding to the second sensing type is device operation data, the method further includes: Acquire K second prior information and a first description instruction sample including a prior information reserved position, wherein the second prior information is used to indicate a fault diagnosis result corresponding to the sensor information of the second sensor type, and the first description instruction sample is used to guide a large-scale industrial model to generate device environment data corresponding to a plurality of device operation data groups, wherein a plurality of device operation data groups corresponding to the device environment data corresponding to the plurality of device operation data groups constitute target device operation data associated with the prior information corresponding to the prior information reserved position, and K≥1; Fill the K second priori information into the priori information reserved position in the first description instruction sample to obtain a first description instruction; According to the first description instruction, the first example sensor data is generated based on the large-scale industrial model, where the first example sensor data is device environment data corresponding to the device operation data group associated with the K second prior information; Using the K second priori information as the first priori information of the first example sensor data; If the first description instruction sample is also used to guide the large-scale industrial model to generate the remaining prior information associated with the target device operation data, then taking the K second prior information as the first prior information of the first example sensor data includes: taking the K second prior information and the remaining prior information as the first prior information of the first example sensor data.
5. The method according to claim 1, characterized in that The method of performing fault diagnosis based on the first implicit representation array and the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result includes: According to the first implicit representation array, transforming based on a synchronous network to obtain a first synchronous implicit representation array, the synchronous network is used to transform the implicit representation array from the target implicit representation domain to a network implicit representation domain, the network implicit representation domain is an implicit representation domain that can be processed by the vehicle diagnostic network; According to the first synchronous implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result; The method further comprises: Acquiring to-be-diagnosed sensor data corresponding to the second sensor type; Performing implicit representation mining on the sensor data to be diagnosed based on the first implicit representation mining network to obtain a second implicit representation array; According to the second implicit representation array, transforming based on the synchronization network to obtain a second synchronous implicit representation array; According to the second synchronous implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to obtain a fault diagnosis result of the sensor data to be diagnosed.
6. The method according to claim 5, characterized in that The method further comprises: Acquire second example sensing data corresponding to the first sensing type, wherein the second example sensing data carries third prior information; According to the second example sensor data, performing implicit representation mining on the second example sensor data based on the first implicit representation mining network to obtain a third implicit representation array; According to the third implicit representation array, transforming based on the initial synchronization network to obtain an inference synchronization implicit representation array; According to the inference synchronous implicit representation array, fault diagnosis is performed based on the vehicle diagnostic network to be adjusted to obtain a fault diagnosis result; According to the error between the fault diagnosis result and the third priori information, the network configuration parameters of the vehicle diagnostic network to be adjusted are fixed, the network configuration parameters of the initial synchronization network are optimized, and the synchronization network is obtained.
7. The method according to claim 6, characterized in that The method further comprises: Obtaining a second description instruction sample and a synchronization command, wherein the second description instruction sample includes an implicit representation array reserved position and a synchronization command reserved position, and the synchronization command is used to guide the vehicle diagnostic network to be adjusted to generate the fault diagnosis result; Fill the inference synchronization implicit representation array into the reserved position of the implicit representation array in the second description instruction sample, and fill the synchronization command into the reserved position of the synchronization command in the second description instruction sample to obtain a second description instruction; The step of performing fault diagnosis based on the inference synchronous implicit representation array and the vehicle diagnostic network to be calibrated to obtain a fault diagnosis result includes: According to the second description instruction, performing fault diagnosis based on the vehicle diagnostic network to be calibrated to obtain the fault diagnosis result; The step of optimizing the network configuration parameters of the vehicle diagnostic network to be calibrated based on the error between the inference diagnosis result and the first priori information to obtain the vehicle diagnostic network comprises: According to the error between the inference diagnosis result and the first a priori information, the network configuration parameters of the vehicle diagnostic network to be adjusted are optimized to obtain a vehicle diagnostic network, and according to the error between the inference diagnosis result and the first a priori information, the network configuration parameters of the synchronization network are optimized to obtain an updated synchronization network; If the sensing information corresponding to the first sensing type is device environment data, and the sensing information corresponding to the second sensing type is device operation data, then the second example sensing data includes device environment data corresponding to multiple device operation data groups, and the third prior information is global device environment data of the device operation data corresponding to the second example sensing data or the device environment data corresponding to the multiple device operation data groups.
8. The method according to claim 1, characterized in that The method further comprises: Acquire core data of the first example sensor data; Performing implicit representation mining on the core data to obtain a core implicit representation array; At this time, the fault diagnosis is performed based on the first implicit representation array and the vehicle diagnostic network to be calibrated to obtain the inference diagnosis result, including: Performing fault diagnosis based on the vehicle diagnostic network to be calibrated according to the first implicit representation array and the core implicit representation array to obtain the inference diagnosis result; Alternatively, the method further comprises: Obtaining a third description instruction sample and a detail optimization command, wherein the third description instruction sample includes an implicit representation array reserved position and a detail optimization command reserved position, and the detail optimization command is used to guide the vehicle diagnostic network to be adjusted to generate the inference diagnosis result; Filling the first implicit representation array into the reserved position of the implicit representation array in the third description instruction sample, and filling the detail optimization command into the reserved position of the detail optimization command in the third description instruction sample, to obtain a third description instruction; At this time, the fault diagnosis is performed based on the first implicit representation array and the vehicle diagnostic network to be calibrated to obtain the inference diagnosis result, including: According to the third description instruction, fault diagnosis is performed based on the vehicle diagnostic network to be calibrated to obtain an inference diagnosis result.
9. The method according to claim 8, characterized in that The third description instruction sample also includes a core data reserved location, and the method further includes: Acquire core data of the first example sensor data; Performing implicit representation mining on the core data to obtain a core implicit representation array; The step of filling the first implicit representation array into the reserved position of the implicit representation array in the third description instruction example, and filling the detail optimization command into the reserved position of the detail optimization command in the third description instruction example, to obtain the third description instruction, comprises: Fill the first implicit representation array into the reserved position of the implicit representation array in the third description instruction sample, fill the detail optimization command into the reserved position of the detail optimization command in the third description instruction sample, fill the core implicit representation array into the reserved position of the core data in the third description instruction sample, and obtain the third description instruction.
10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.
