In-vehicle computing system and vehicle

By introducing a pre-fault prediction module and a scheduling module into the vehicle computing system, combined with a biomimetic synaptic structure and a dynamic fault-tolerant mechanism, proactive prediction and rapid response to neuromorphic chip faults are achieved, solving the problem of long fault response time and improving the stability and response speed of the autonomous driving system.

WO2026237569A1PCT designated stage Publication Date: 2026-11-19CHERY AUTOMOBILE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2026/086975
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-10-31
Filing Date
2026-03-30
Publication Date
2026-11-19

Smart Images

  • Figure CN2026086975_19112026_PF_FP_ABST
    Figure CN2026086975_19112026_PF_FP_ABST
Patent Text Reader

Abstract

The present application relates to an in-vehicle computing system and a vehicle. The system comprises: at least one main computing module, which is configured to execute a target computing task; a pre-fault prediction module, which comprises a pre-fault prediction model, and is configured to predict a fault time interval of the main computing module on the basis of the pre-fault prediction model; and a scheduling module, which is configured to trigger a corresponding fault tolerance strategy on the basis of the fault time interval, and execute the fault tolerance strategy.
Need to check novelty before this filing date? Find Prior Art

Description

In-vehicle computing systems and vehicles

[0001] This application claims priority to Chinese Patent Application No. 202511578094.5, filed on October 31, 2025, entitled "In-vehicle Computing System and Vehicle", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of vehicle technology, and in particular to an in-vehicle computing system and a vehicle. Background Technology

[0003] As autonomous driving advances to Level 4 and above, the demands for real-time perception, multi-sensor fusion, and dynamic decision-making in vehicle computing have increased significantly. Neuromorphic chips, with their brain-like architecture, low power consumption, and real-time processing advantages, have been widely adopted. Since autonomous driving is directly related to driver and passenger safety, and neuromorphic chips are the core for processing multi-source sensor data, their stability and reliability directly affect the accuracy of decision-making. Failures could lead to serious accidents such as collisions and loss of control. Therefore, accurate and timely fault monitoring of in-vehicle neuromorphic chips is essential.

[0004] However, fault detection in neural chips in related technologies mainly relies on passive detection mechanisms such as hardware CRC (Cyclic Redundancy Check), which means that the verification and processing of anomalies can only be performed when erroneous data is generated during a fault. This results in a long fault response time, which cannot meet the requirements of fault response speed in emergency scenarios of autonomous driving. Summary of the Invention

[0005] This application provides an in-vehicle computing system and a vehicle to solve problems such as long fault response time in related technologies.

[0006] The first aspect of this application provides an in-vehicle computing system, including: at least one main computing module for executing a target computing task; a pre-fault prediction module, which includes a pre-fault prediction model and predicts the fault time interval of the main computing module based on the pre-fault prediction model; and a scheduling module for triggering a corresponding fault tolerance strategy according to the fault time interval and executing the fault tolerance strategy.

[0007] Optionally, in one embodiment of this application, the main computing module is composed of multiple biomimetic synapses, wherein the biomimetic synapse includes: a substrate; a graphene bottom gate disposed on the substrate, and a top gate covering the graphene bottom gate, wherein a biosynaptic interface is disposed on the outer side of the top gate for receiving input voltage from an external sensor; a molybdenum disulfide channel layer disposed on the top gate, and an ion gel layer covering the molybdenum disulfide channel layer.

[0008] Optionally, in one embodiment of this application, the top gate and the graphene bottom gate form a dual-gate control structure to simulate calcium ion channels and adjust the conductivity of the molybdenum disulfide channel layer according to the synaptic dynamics equation; the molybdenum disulfide channel layer and the ion gel layer constitute a heterojunction to adjust the weights through changes in conductivity and realize the analog domain multiplication and addition operation of the input voltage pulse and the weights.

[0009] Optionally, in one embodiment of this application, the synaptic dynamics equation is: ΔW=γ·[Ca 2+ ]·(V_pre-V_post);

[0010] Where ΔW is the weight change, γ is the calcium ion diffusion coefficient, and Ca 2+ V_pre is the calcium ion concentration, V_post is the input voltage, and V_post is the output voltage after biomimetic synapse processing.

[0011] Optionally, in one embodiment of this application, the bionic synapse is further used to: adjust the short-term enhancement factor based on the computational load of the target computing task.

[0012] Optionally, in one embodiment of this application, the pre-fault prediction module is further configured to: obtain the equivalent series resistance value, thermal noise spectrum and temperature change rate of the electrolytic capacitor of the main calculation module; input the equivalent series resistance value, thermal noise spectrum and temperature change rate of the electrolytic capacitor into the pre-fault prediction model, and the pre-fault prediction model outputs the predicted fault time interval.

[0013] Optionally, in one embodiment of this application, the in-vehicle computing system further includes: at least one backup computing module, wherein the backup computing module is composed of a plurality of bionic synapses.

[0014] Optionally, in one embodiment of this application, the scheduling module is further configured to: if the fault time interval is greater than a first preset duration, then the fault tolerance strategy is the first strategy; if the fault time interval is greater than a second preset duration and less than or equal to the first preset duration, then the fault tolerance strategy is the second strategy; if the fault time interval is less than or equal to the second preset duration, then the fault tolerance strategy is the third strategy.

[0015] Optionally, in one embodiment of this application, the first strategy is to save the bionic synaptic state difference at target intervals; the second strategy is to copy the target computing task to the backup computing module, and the main computing module and the backup computing module execute the target computing task synchronously; the third strategy is to use the target hardware hot-swap protocol to switch the hardware interface connected to the main computing module to the backup computing module, wherein the backup computing module executes the target computing task and the main computing module is disabled.

[0016] A second aspect of this application provides a vehicle including an on-board computing system as described in the above embodiments.

[0017] Therefore, this application has at least the following beneficial effects:

[0018] This application provides an in-vehicle computing system comprising a main computing module, a fault prediction module, and a scheduling module. The fault prediction module can predict the fault interval of the main computing module based on a fault prediction model. The scheduling module triggers corresponding fault-tolerant strategies based on the fault interval, proactively implementing fault prevention and fault-tolerant processing, reducing fault response time, improving fault handling efficiency, and avoiding task interruption caused by sudden failure of the main computing module. This solves the technical problems of long fault response times in related technologies.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0021] Figure 1 is a schematic diagram of an in-vehicle computing system provided according to an embodiment of this application;

[0022] Figure 2 is a schematic diagram of the structure of the bionic synapse provided according to an embodiment of this application;

[0023] Figure 3 is a flowchart of the pre-fault prediction provided according to an embodiment of this application;

[0024] Figure 4 is a timing diagram of a dynamic fault-tolerant mechanism provided according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0026] Before describing the solution of this application, let me first introduce the relevant technologies and existing problems of this application.

[0027] 1. Energy efficiency bottleneck.

[0028] Neuromorphic chips in related technologies (such as Intel Loihi) still consume more than 5W per core, which cannot meet the energy efficiency requirements of automotive sensor fusion scenarios.

[0029] 2. Passive fault-tolerant delay.

[0030] In related technologies, fault detection relies on hardware CRC check, and the fault response time is >10ms (ASIL-D requires ≤1ms);

[0031] 3. Insufficient utilization of biological characteristics.

[0032] The related technologies do not simulate the STP (Short-Term Plasticity) characteristics of biological synapses, and therefore cannot dynamically adapt to sudden computational loads.

[0033] Furthermore, the relevant technologies lack hardware-level simulation of biological neural calcium ion channel mechanisms; no pre-fault prediction model has been established; and the coordinated control of dynamic voltage regulation and synaptic weight update is missing.

[0034] To this end, this application provides an in-vehicle computing system, including a main computing module, a pre-fault prediction module, and a scheduling module. The pre-fault prediction module can predict the fault time interval of the main computing module based on the pre-fault prediction model. The scheduling module triggers the corresponding fault tolerance strategy based on the fault time interval, actively realizes fault prevention and fault tolerance processing, reduces fault response time, improves fault handling efficiency, and avoids task interruption caused by sudden failure of the main computing module.

[0035] Specifically, Figure 1 is a schematic diagram of an in-vehicle computing system provided in an embodiment of this application.

[0036] As shown in Figure 1, the vehicle-mounted computing system 10 includes at least one main computing module 11, a pre-fault prediction module 12, and a scheduling module 13.

[0037] The main computing module 11 is used to execute the target computing task; the pre-fault prediction module 12 includes a pre-fault prediction model, which predicts the fault time interval of the main computing module 11 based on the pre-fault prediction model; the scheduling module 13 is used to trigger the corresponding fault tolerance strategy according to the fault time interval and execute the fault tolerance strategy.

[0038] The main computing module in this embodiment is the core computing unit of the vehicle computing system and the main carrier for performing target computing tasks. The target computing tasks can be sensor data fusion, target tracking, path planning, etc. in autonomous driving scenarios. The pre-fault prediction model can be an LSTM (Long Short-Term Memory) prediction model. The fault time interval is the time difference from the current moment to the moment when the main computing module predicts the occurrence of the fault.

[0039] It is understood that the embodiments of this application construct an in-vehicle computing system 10, including a main computing module 11, a pre-fault prediction module 12 and a scheduling module 13. The pre-fault prediction module 12 can predict the fault time interval of the main computing module based on the pre-fault prediction model. The scheduling module 13 triggers the corresponding fault tolerance strategy based on the fault time interval, actively realizes fault prevention and performs fault tolerance processing, reduces fault response time, improves fault handling efficiency, and avoids task interruption caused by sudden failure of the main computing module 11.

[0040] Furthermore, in one embodiment of this application, the main computing module 11 is composed of a plurality of biomimetic synapses, wherein the biomimetic synapse includes: a substrate; a graphene bottom gate disposed on the substrate, and a top gate covering the graphene bottom gate, wherein a biosynaptic interface is disposed on the outer side of the top gate for receiving input voltage from an external sensor; a molybdenum disulfide channel layer disposed on the top gate, and an ion gel layer covering the molybdenum disulfide channel layer.

[0041] It is understood that the main computing module 11 in this embodiment of the application is composed of multiple biomimetic synapses. The specific structure of the biomimetic synapse is shown in Figure 2, including: from bottom to top: a substrate, such as an SiO2 substrate; a graphene bottom gate disposed on the substrate, and a top gate covering the graphene bottom gate, wherein a biosynaptic interface is disposed on the outer side of the top gate, i.e., side 1 of the top gate shown in Figure 2, and the biosynaptic interface is used to receive the input voltage of an external sensor; a molybdenum disulfide (MoS2) channel layer disposed on the top gate, and an ion gel layer covering the molybdenum disulfide channel layer.

[0042] The biomimetic synapse in this application has a synapse density of 10⁹ cells / cm². 2 or not less than 109 cells / cm 2 The weight update delay is ≤5ns, and the single synapse energy consumption is 0.05fJ / bit, or not less than 0.05fJ / bit.

[0043] This application uses biomimetic synapses as computing units and adopts a MoS2 / ion gel heterojunction structure. This material has extremely high charge retention capability and extremely low leakage current, which fundamentally reduces static power consumption.

[0044] Furthermore, in one embodiment of this application, the top gate and the graphene bottom gate form a dual-gate control structure to simulate calcium ion channels and adjust the conductivity of the molybdenum disulfide channel layer according to the synaptic dynamics equation; the molybdenum disulfide channel layer and the ion gel layer constitute a heterojunction to adjust the weights through changes in conductivity and realize the analog domain multiplication and addition operation of the input voltage pulse and the weights.

[0045] The dual-gate control structure is a synergistic control structure composed of a top gate and a graphene bottom gate. By applying different voltages to the top and bottom gates, it simulates the voltage-gated characteristics of calcium ion channels in biological nerves, thereby achieving precise and dynamic adjustment of the conductivity of the MoS2 channel layer. The simulation of calcium ion channels is achieved by adjusting the voltage of the dual-gate control structure, i.e., controlling the gate voltage of the dual-gate control structure, to simulate the transmembrane flow of calcium ions in biological nerve cells. The calcium ion concentration is simulated by the gate voltage, and then the plasticity of the synapse (weight change) is realized through gate voltage control, thus restoring the signal transmission mechanism of biological synapses. The heterojunction is a heterostructure composed of a MoS2 channel layer (semiconductor) and an ion gel layer (electrolyte). This heterostructure combines the tunable conductivity of MoS2 with the low leakage current characteristics of ion gel, and is the core structure for realizing low power consumption, high stability weight adjustment and analog domain operation. The analog domain multiply-add operation is to perform multiplication using the input voltage pulse (analog signal) and the conductivity value (weight) of the MoS2 channel layer, and to perform addition by superimposing the output currents of multiple synapses, ultimately outputting a current signal.

[0046] The dual-gate control structure of this application simulates calcium ion channels and, combined with the weight adjustment capability of heterojunctions, enables the biomimetic synapse to possess plasticity similar to that of biological nerves. It can dynamically adjust the weights according to the input signal strength (voltage pulse amplitude) and frequency, i.e., the voltage pulse amplitude and frequency applied to the dual-gate control structure, i.e., the dynamic changes in the gate voltage. It can adapt to sudden computing loads (such as the need for rapid decision-making when obstacles suddenly appear in autonomous driving), increasing computing power by 3-5 times within 100μs. Moreover, the analog domain multiplication and addition operations do not require analog-to-digital conversion and clock synchronization in traditional digital computing, reducing conversion losses and synchronization overhead. At the same time, the synaptic dynamics equation guides the weight adjustment, making the change in conductivity (weight) quantitatively controllable and avoiding excessive weight drift.

[0047] Furthermore, in one embodiment of this application, the synaptic dynamics equation is: ΔW=γ·[Ca 2+ ]·(V_pre-V_post);

[0048] Where ΔW is the weight change, γ is the calcium ion diffusion coefficient, and Ca 2+ V_pre represents the calcium ion concentration, simulated by the gate voltage, V_post represents the input voltage, and V_post represents the output voltage after biomimetic synapse processing.

[0049] Furthermore, in one embodiment of this application, the bionic synapse is further used to: adjust the short-term enhancement factor based on the computational load of the target computing task.

[0050] Among them, the short-term enhancement factor is a parameter that simulates the short-term plasticity of biological synapses. It is used to adjust the instantaneous computing power of bionic synapses. The larger the value of the short-term enhancement factor, the higher the signal transmission efficiency of the synapse per unit time and the stronger the computational load that can be processed.

[0051] It is understood that the bionic synapse in this application embodiment can adjust the short-term enhancement factor based on the computational load of the target computing task to enhance the computational stability of the vehicle computing system under sudden loads. For example, when sudden target computing tasks such as obstacle avoidance and emergency braking occur in an autonomous driving scenario, the vehicle computing system can increase the value of the short-term enhancement factor in real time, quickly improve the signal transmission efficiency and instantaneous computing capability of the bionic synapse, significantly improve the system's computing power response speed in a short period of time, ensure the stable execution of high real-time target computing tasks, and avoid task blocking or response delay problems caused by sudden increases in computational load.

[0052] Specifically, the biomimetic synapse in this application embodiment includes a material layer, a structural layer, and a circuit layer. The material layer uses a two-dimensional MoS2 and ion gel heterojunction. Compared with the ion migration energy consumption (>1fJ) of traditional CMOS (Complementary Metal Oxide Semiconductor), the ion migration energy consumption of the two-dimensional MoS2 and ion gel heterojunction in this application embodiment can reach ≤0.05fJ / bit. The structural layer is based on a dual-gate transistor simulating a calcium ion channel, with a weight update speed of ≤5ns, which is about 20 times faster than related technologies. The circuit layer is an STP characteristic simulation circuit. The short-term synaptic enhancement factor β can be set to, for example, 0.8 and can be dynamically adjusted. This application does not impose any restrictions on the value of the short-term synaptic enhancement factor β.

[0053] Furthermore, in one embodiment of this application, the pre-fault prediction module 12 is further configured to: obtain the equivalent series resistance value of the electrolytic capacitor, the thermal noise spectrum, and the temperature change rate of the main calculation module 11; input the equivalent series resistance value of the electrolytic capacitor, the thermal noise spectrum, and the temperature change rate into the pre-fault prediction model, and the pre-fault prediction model outputs the predicted fault time interval.

[0054] It is understood that the embodiments of this application can monitor the main computing module 11, namely the equivalent series resistance value, thermal noise spectrum and temperature change rate of the electrolytic capacitor of the bionic synapse, and input the equivalent series resistance value, thermal noise spectrum and temperature change rate of the electrolytic capacitor into the pre-fault prediction model. The pre-fault prediction model outputs the predicted fault time interval, thereby realizing the early prediction of faults and completing the fault prediction before the hardware completely fails, reserving sufficient time for the scheduling module to trigger the fault tolerance strategy.

[0055] The formula for calculating the fault time interval in this application embodiment can be: T_fault=f(ESR,dT / dt,Noise_PSD);

[0056] Where T_fault is the fault time interval, ESR is the equivalent series resistance of the electrolytic capacitor, dT / dt is the temperature change rate, and Noise_PSD is the thermal noise spectrum.

[0057] Furthermore, in one embodiment of this application, the in-vehicle computing system 10 further includes at least one backup computing module, wherein the backup computing module is composed of a plurality of bionic synapses.

[0058] It is understood that the vehicle computing system 10 in this application embodiment also includes a backup computing module. The backup computing module has the same structure as the main computing module, so that when the main computing module fails, the backup computing module is used to reduce the impact of the failure and improve the reliability of the vehicle computing system.

[0059] Furthermore, in one embodiment of this application, the scheduling module 13 is further configured to: if the fault time interval is greater than a first preset duration, then the fault tolerance strategy is the first strategy; if the fault time interval is greater than a second preset duration and less than or equal to the first preset duration, then the fault tolerance strategy is the second strategy; if the fault time interval is less than or equal to the second preset duration, then the fault tolerance strategy is the third strategy.

[0060] The first preset duration and the second preset duration can be set according to specific circumstances, without any specific limitations. For example, the first preset duration can be set to 5ms or 6ms, and the second preset duration can be set to 1ms or 2ms.

[0061] It is understood that the embodiments of this application can determine different fault tolerance strategies based on the predicted fault time interval, match differentiated strategies for different risk levels, and balance the problems of excessive fault tolerance and insufficient fault tolerance.

[0062] Furthermore, in one embodiment of this application, the first strategy is to save the bionic synaptic state difference at target intervals; the second strategy is to copy the target computing task to the backup computing module, and the main computing module 11 and the backup computing module execute the target computing task synchronously; the third strategy is to use the target hardware hot-swap protocol to switch the hardware interface connected to the main computing module 11 to the backup computing module, wherein the backup computing module executes the target computing task and the main computing module 11 is disabled.

[0063] The target duration can be set according to specific circumstances, without any specific limitation, such as setting it to 1ms or 2ms; the state difference is divided into the part of the total change between the current bionic synapse state and the last saved state; the target hardware hot-swap protocol can be PCIe Gen5 (peripheral component interconnect generation 5).

[0064] It is understood that in this embodiment of the application, when the fault time interval is greater than the first preset time, the bionic synaptic state difference is saved at a target time interval, which can significantly reduce storage overhead by 70%. If the main computing module fails later, the most recent state difference can be read from the cache and combined with historical data to quickly restore the synaptic state.

[0065] When the fault interval is greater than the second preset duration and less than or equal to the first preset duration, task cloning can be performed to copy the target computing task to the backup computing module. The main computing module and the backup computing module run in parallel, executing the target computing task at the same time, ensuring that the backup computing module can take over immediately without waiting for task migration or status recovery when the main computing module fails.

[0066] When the fault interval is less than or equal to the second preset duration, hardware switching can be completed quickly through PCIe Gen5, immediately stopping the main computing module, cutting off the signal connection with sensors and actuators, with a switching delay of ≤100ns, and the backup computing module executes the target computing task.

[0067] This application enables hardware switching to be completed within 82ns using PCIe Gen5. Based on the atomic operation protocol of PCIe Gen5, it can avoid the scheduling overhead of traditional operating systems. Furthermore, it adopts a pre-synchronization mechanism to ensure that the backup NPU (Neural Processing Unit) (i.e., backup computing module) is always synchronized with the main NPU (i.e., main computing module). During switching, only the difference data needs to be transmitted. At the same time, charge preservation technology is used to ensure that the synaptic state is not lost during the switching process.

[0068] The embodiments of this application can also achieve energy efficiency compensation and reduce energy efficiency fluctuations through adaptive adjustment of synaptic voltage.

[0069] Furthermore, it should be noted that, in addition to using the LSTM prediction model to output the fault time interval to trigger the corresponding fault tolerance strategy, the embodiments of this application can also use the probability output by the fault probability model to trigger the corresponding fault tolerance strategy. For example, the resistance change rate, thermal noise spectrum and temperature change rate are input into the fault probability model, the fault probability model outputs the corresponding fault probability, and the corresponding fault tolerance strategy is matched based on the fault probability. For example, if the fault probability is greater than the first preset probability, the fault tolerance strategy is the first strategy; if the fault probability is greater than the second preset probability and less than or equal to the first preset probability, the fault tolerance strategy is the second strategy; if the fault probability is less than or equal to the second preset probability, the fault tolerance strategy is the third strategy.

[0070] The failure probability model can be: P_fault=1 / (1+exp(-z));

[0071] Where P_fault is the failure probability, z is the log odds of the linear regression output, z = ∑(w_i·x_i) + b, w_i and b are model parameters, w_i is the weight, b is the bias term, and x_i is the input feature, x = [dR / dt, f_c, d 2 T / dt 2 ], dR / dt is the rate of change of resistance, dT / dt is the rate of change of temperature, and f_c is the characteristic frequency of the thermal noise spectrum.

[0072] Specifically, the dynamic fault tolerance strategy of this application embodiment is shown in Table 1.

[0073] Table 1

[0074] Specifically, the implementation of the vehicle computing system in this application mainly includes three parts: biomimetic synapses as computing units, a pre-fault prediction algorithm, and a dynamic fault tolerance mechanism.

[0075] I. Bionic synapse.

[0076] The specific structure of the biomimetic synapse is shown in Figure 2. It adopts a MoS2 / ion gel heterojunction structure. This material has extremely high charge retention capability and extremely low leakage current, which fundamentally reduces static power consumption. It simulates the voltage gating characteristics of biological nerve calcium ion channels and achieves precise charge injection control through dual-gate regulation, avoiding the switching power loss of traditional CMOS devices. Utilizing the STP characteristic, it automatically enters the subthreshold working state when there is no computing task, reducing power consumption to the nW level.

[0077] Each bionic synapse is equivalent to a MAC (multiply-accumulate) unit in a traditional AI (artificial intelligence) accelerator, but it has the ability to dynamically adjust weights. The weights are represented by the conductance of the MoS2 channel. The input voltage pulse and the weights are multiplied and added in the analog domain, and the output current signal is generated. Event-driven calculation is used, and the synapse is activated only when the input signal arrives, thus avoiding the power loss caused by clock synchronization.

[0078] Furthermore, this application can dynamically adjust the weights and enhancement factors of the biomimetic synapse, achieving adaptability to the biological nervous system. The weights determine the strength of information transmission, corresponding to long-term memory and learning ability, while the enhancement factors simulate calcium ion-dependent short-term plasticity, enabling instantaneous computational power enhancement under burst loads. When the system detects a sudden computational demand, it simultaneously increases the enhancement factor and adjusts the weights, increasing computational power by approximately 3-5 times within about 100 μs.

[0079] The dynamic equation of the synapse is as follows: ΔW=γ·[Ca 2+ ]·(V_pre-V_post);

[0080] Where ΔW is the weight change, γ is the calcium ion diffusion coefficient, which determines the weight update rate, and γ = 0.03. 2+ The calcium ion concentration is determined by the gate voltage V. ecc The simulation controls the plasticity state of the synapse, where V_pre is the input voltage and V_post is the output voltage after the bionic synapse processing.

[0081] II. The overall process of the PFP (Pre-Fault Prediction) algorithm is shown in Figure 3.

[0082] The goal of the fault prediction algorithm in this application is to predict faults more than 10ms before complete hardware failure, thus providing a time window for fault-tolerant switching.

[0083] 2.1 Fault feature extraction.

[0084] The electrolytic capacitance (equivalent series resistance) value (threshold > 50 mΩ), thermal noise spectrum (1 / f noise inflection point detection), and temperature change rate of the biomimetic synapse were monitored online. Among them, the increase of ESR value indicates ion gel aging, and the 1 / f noise inflection point indicates the generation of MoS2 channel defects.

[0085] 2.2 Prediction Model.

[0086] Fault time prediction based on LSTM prediction model: T_fault=f(ESR,dT / dt,Noise_PSD), prediction accuracy>92%, false alarm rate<0.1%.

[0087] Specifically, the LSTM prediction model takes multiple time-series monitoring data as input, including ESR, temperature change rate, and 128 frequency domain features of thermal noise spectrum, and outputs the predicted time interval Δt from the current moment to the occurrence of the fault, with an accuracy of milliseconds.

[0088] III. Dynamic fault tolerance mechanism, as shown in Table 1 and Figure 4. Figure 4 is a timing diagram of dynamic fault tolerance scheduling, which realizes a 100ns-level switching process.

[0089] The main advantages of this application compared to NVIDIA Orin in related technologies are shown in Table 2.

[0090] Table 2

[0091] In summary, this application can break through the power consumption barrier of neuromorphic computing (target ≤0.5μW / synapse, related technologies ≥2μW); achieve pre-fault detection (predicting faults 10ms in advance) and fault-tolerant switching latency ≤100ns; and improve computing stability under burst loads (task completion rate >99.999%).

[0092] The implementation process of the vehicle-mounted computing system of this application embodiment is described below through specific embodiments.

[0093] Example 1: Bionic synaptic computation.

[0094] 1. Input: 4D millimeter-wave radar point cloud (100,000 points / frame);

[0095] 2. Processing: The synaptic array performs 3D convolution (dynamically allocating weights using STP characteristics);

[0096] 3. Output: Target tracking trajectory (delay ≤ 800μs).

[0097] Example 2: Pre-fault response.

[0098] 1. Detection: The capacitor ESR value rose to 55mΩ (predicted fault countdown 8.2ms);

[0099] 2. Migration: Clone the scheduled task to the backup NPU (takes 75ns);

[0100] 3. Recovery: The faulty node enters ion rebalancing mode (power consumption drops to 0.1μW).

[0101] The vehicle-mounted computing system proposed in the embodiments of this application includes a main computing module, a pre-fault prediction module, and a scheduling module. The pre-fault prediction module can predict the fault time interval of the main computing module based on the pre-fault prediction model. The scheduling module triggers the corresponding fault tolerance strategy based on the fault time interval, actively realizes fault prevention and performs fault tolerance processing, reduces fault response time, improves fault handling efficiency, and avoids task interruption caused by sudden failure of the main computing module.

[0102] Embodiments of this application also provide a vehicle including the on-board computing system described above.

[0103] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0104] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0105] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0106] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0107] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. An in-vehicle computing system, wherein, include: At least one main computing module is used to execute the target computing task; A fault prediction module, comprising a fault prediction model, for predicting the fault time interval of the main computing module based on the fault prediction model; The scheduling module is used to trigger the corresponding fault tolerance strategy according to the fault time interval and execute the fault tolerance strategy.

2. The in-vehicle computing system of claim 1, wherein, The main computing module is composed of multiple bionic synapses, wherein the bionic synapses include: Base; A graphene bottom grid is disposed on the substrate, and a top grid is covered on the graphene bottom grid, wherein a biosynaptic interface is disposed on the outer side of the top grid, and the biosynaptic interface is used to receive the input voltage of an external sensor. A molybdenum disulfide channel layer disposed on the top grid, and an ionogel layer covering the molybdenum disulfide channel layer.

3. The in-vehicle computing system of claim 2, wherein, The top gate and the graphene bottom gate form a dual-gate control structure to simulate calcium ion channels and adjust the conductivity of the molybdenum disulfide channel layer according to the synaptic dynamics equation; the molybdenum disulfide channel layer and the ion gel layer form a heterojunction to adjust the weights by changing the conductivity and realize the analog domain multiplication and addition operation of the input voltage pulse and the weights.

4. The in-vehicle computing system of claim 3, wherein, The synaptic kinetics equation is: ΔW = γ - [Ca 2+ ] - (V_pre - V_post); wherein, AW is the weight change amount, γ is the calcium ion diffusion coefficient, Ca 2+ is the calcium ion concentration, V_pre is the input voltage, and V_post is the output voltage after the processing of the biomimetic synapse.

5. The in-vehicle computing system of claim 2, wherein, The bionic synapse is further used to: adjust the short-term enhancement factor based on the computational load of the target computing task.

6. The in-vehicle computing system of claim 2, wherein, The fault prediction module is further used for: Obtain the equivalent series resistance value of the electrolytic capacitor, the thermal noise spectrum, and the temperature change rate of the main computing module; The equivalent series resistance of the electrolytic capacitor, the thermal noise spectrum, and the temperature change rate are input into the pre-fault prediction model, and the pre-fault prediction model outputs the predicted fault time interval.

7. The in-vehicle computing system of claim 1, wherein, The in-vehicle computing system further includes at least one backup computing module, wherein the backup computing module is composed of multiple bionic synapses.

8. The in-vehicle computing system of claim 7, wherein, The scheduling module is further used for: If the fault time interval is greater than the first preset duration, then the fault tolerance strategy is the first strategy; If the fault time interval is greater than the second preset duration and less than or equal to the first preset duration, then the fault tolerance strategy is the second strategy. If the fault time interval is less than or equal to the second preset duration, then the fault tolerance strategy is the third strategy.

9. The in-vehicle computing system of claim 8, wherein, The first strategy is to save the biomimetic synaptic state difference at intervals of the target duration; The second strategy is to copy the target computing task to the backup computing module, and the main computing module and the backup computing module execute the target computing task synchronously. The third strategy involves using a target hardware hot-swappable protocol to switch the hardware interface connected to the main computing module to the backup computing module, wherein the backup computing module executes the target computing task, and the main computing module is disabled.

10. A vehicle characterized by comprising: Including the in-vehicle computing system as described in any one of claims 1-9.