System architecture of intelligent terminal

Abstract

The invention discloses a system architecture of an intelligent terminal, and belongs to the field of artificial intelligence. The intelligent terminal system architecture is composed of a neural network processing unit, a central processing unit, a sensing module, a knowledge base input module, a feature memory bank, a local memory bank, an instruction memory, an input and output module, a control module, an execution result storage module and the like. The intelligent terminal system architecture is more intelligent, has the functions of automatic code generation, multi-level memory implementation, multi-dimensional sensing data input, knowledge base data input and diversified reasoning, and can better interact with human beings. The local memory in the system architecture of the intelligent terminal is stored in an internal private format, so that the privacy of the memory and the security of the intelligent terminal are improved. The system architecture can be widely applied to the field of various intelligent terminals such as robots and intelligent automobiles. The method is novel, high in operability, wide in application range and great in application significance.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a system architecture of an intelligent terminal. Background Art

[0002] With the rapid development of science and technology, artificial intelligence has brought about a technological revolution. Artificial intelligence is being used more and more widely in various devices and systems, from mobile phones to cars, from medical equipment to industrial automation equipment, from homes to companion robots, etc. Its influence is everywhere.

[0003] In the field of self-driving cars, artificial intelligence can perceive and analyze the environment around the vehicle in real time, providing precise navigation and control for the vehicle's driving. In the field of industrial automation, industrial robots can more accurately identify and process objects through deep learning, improving production efficiency. In the field of smart homes, through artificial intelligence technology, home appliances can better understand and predict user behavior and needs, providing users with a more convenient, comfortable and personalized living environment. In the field of network recommendation systems, artificial intelligence provides users with more accurate and personalized recommendations through deep learning of user behavior, content features and other information.

[0004] However, current smart terminals generally have the following major deficiencies: lack of multi-level memory function, and the inability to use memory for reasoning when making decisions and controlling. Smart terminals are not yet capable of personalization and autonomous learning, especially they cannot automatically generate operating code for upgrade and optimization through autonomous learning. Smart terminals have limited types of multimedia input information, and insufficient use of the already existing huge multimedia knowledge base data. Summary of the invention

[0005] The present application provides a system architecture of an intelligent terminal. The system architecture of the intelligent terminal realizes automatic code generation, multi-level memory, and diversified reasoning, accepts input knowledge base and multi-modal sensor input information, and can be used to solve the technical problems of the lack of multi-level memory, the inability to use memory for reasoning, and the inability to automatically generate running code in the intelligent terminal.

[0006] The present application provides an intelligent terminal system architecture, including:

[0007] Neural network processing unit, central processing unit, sensor module [1:n], knowledge base input module, feature memory, local memory, instruction memory, input and output module, control module, execution result storage module, etc.

[0008] Wherein, the neural network processing unit pre-processes, deduce, predict and make decisions on the multi-dimensional sensor data from the sensor module [1:n], the feature memory from the feature memory bank, the local memory of the local memory bank and the operation output from the central processing unit, generates control decisions and goes to the control module; at the same time, pre-processes, predicts, makes decisions, reconstructs and extracts features on the multi-dimensional sensor data from the sensor module [1:n], the knowledge base data from the knowledge base input module and the operation output data from the central processing unit, generates local memory and feature memory and stores them in the local memory bank and the feature memory bank; autonomous code generation, generates code and goes to the instruction memory;

[0009] The central processing unit executes the code in the instruction memory for controlling, transaction processing and data processing of the intelligent terminal;

[0010] The knowledge base input module is an input module for the external input knowledge base of the intelligent terminal, generates knowledge base data, and senses the decision module; the input knowledge base includes a local knowledge base and a network input knowledge base;

[0011] The sensing module [1:n] is a group of sensing modules for identifying and detecting the multimodal sensing input information of the intelligent terminal, generating multidimensional sensing data, and sensing the decision module; the multimodal sensing input information includes the multimodal sensing input information outside the intelligent terminal and the multimodal sensing input information inside the intelligent terminal;

[0012] The feature memory library is a storage module for intelligent terminal feature memory;

[0013] The local memory is a storage module of the local memory of the intelligent terminal;

[0014] The instruction memory is a memory for the instruction code of the central processing unit, a memory for the instruction code generated by the autonomous code and a memory for the input external instruction code;

[0015] The input-output module is an input-output interface module of the central processing unit, including a circuit communication interface, a user input and a display output;

[0016] The control module is a control command generating module of the execution device of the intelligent terminal;

[0017] The execution result storage module is a storage module for the execution results of the central processing unit.

[0018] Further, the neural network processing unit includes: a perception decision module, a recall decision module, a code automatic generation module and a memory neural network processing unit;

[0019] Among them, the perception decision module realizes preprocessing of the multi-dimensional sensor data from the sensor module [1:n], the knowledge base data from the knowledge base input module, and the operation output data from the central processing unit, as well as prediction and decision making on the preprocessing results, generating numerical matrices / vectors, predictions / decisions, de-memorizing the neural network processing unit, generating control decisions, and de-controlling the module;

[0020] The recall decision module predicts and makes decisions on the output memory description from the memory neural network processing unit, generates control decisions, and goes to the control module; performs memory output processing, generates output memory, and goes to the central processing unit;

[0021] The code automatic generation module realizes the automatic generation of code and generates code to the instruction memory;

[0022] The memory neural network processing unit fuses, extracts features, and reconstructs the numerical matrix / vector and prediction / decision from the perception decision module to generate the feature memory and local memory of the intelligent terminal, and removes the feature memory library and the local memory library; it deduce, predict, and reconstruct the feature memory and local memory stored in the feature memory library and the local memory library to generate an output memory description in a public format, and removes the recall decision module, the perception decision module, and the automatic code generation module.

[0023] Furthermore, the automatic code generation module includes: a requirement analysis module, a task generation module, a code analysis module, a code preprocessing module, a code processing module, a code optimization module, a debugging module, a compilation module, an in-system test module, and a source code storage;

[0024] Wherein, the demand analysis module generates code generation modification requirements for the inputs from the central processing unit, the execution result storage module, and the sensor module [1:n] involving the requirements for the central processing unit code improvement, function modification, and function addition;

[0025] The task generation module converts the code generation modification requirements output by the requirement analysis module into coding task instructions, and the associated coding task instructions are combined to generate a coding task;

[0026] The code analysis module tokenizes the source code from the source code storage, performs structure analysis on the source code based on sequence and graph, obtains source code control flow and data flow, and outputs a sequence containing source code structure;

[0027] The code preprocessing module performs operations on the output of the code analysis module according to the coding task: cleaning the code; refactoring when necessary; searching the code, locating and marking the code snippets related to the coding task, and locating and marking the places where the code needs to be modified;

[0028] The code processing module processes the output of the code preprocessing module to complete the following functions:

[0029] Create new data structures, functions, classes, member functions, and methods;

[0030] Modify relevant data structures, functions, classes, member functions and methods;

[0031] Modify function calls, structure declarations, and object generation;

[0032] Modify code input / output, including interaction with the user, i.e. reading user input and displaying output;

[0033] Main function modification, error handling, and compilation instruction modification;

[0034] Modify modules, package files and other files;

[0035] The code optimization module evaluates and sorts the different code processing results generated by the code processing module, selects the optimal code, checks the logic errors or potential bugs in the code, analyzes the code security, optimizes the code, and performs syntax checking on the generated code.

[0036] The debugging module and the debugging simulation module perform code bug checks on the source code generated by the code optimization module to generate bug-free source code;

[0037] The compiling module compiles the output code processed by the debugging module to generate executable code;

[0038] The in-system test module runs the executable code generated by the compiling module in a central processing unit environment to check whether the running output of the central processing unit is consistent with the target;

[0039] The source code generated by the automatic code generation module is stored in the source code memory for automatic code generation of the next version; the executable code is stored in the instruction memory module and then executed in the central processing unit for controlling the intelligent terminal and performing transaction processing and data processing.

[0040] Furthermore, the intelligent terminal system architecture is used to implement an intelligent terminal memory structure, and the intelligent terminal memory structure includes: a perception decision module, a recall decision module, a memory neural network processing unit, a feature memory library, and a local memory library;

[0041] Wherein, the perception decision module pre-processes the multi-dimensional sensor data from the sensor module [1:n], the knowledge base data of the knowledge base input module, and the operation output data of the central processing unit, and predicts and makes decisions on the pre-processed results to generate numerical matrices / vectors, predictions / decisions, and memory-free neural network processing units;

[0042] The memory neural network processing unit reconstructs and extracts features from the output of the perception decision module based on the encoder-vector fusion-decoder framework, generates local memory and feature memory, and stores them in the local memory bank and the feature memory bank; performs feature memory deduction and prediction on the local memory and feature memory in the feature memory bank and the local memory bank, and reconstructs them based on the encoder-vector fusion-decoder framework to generate an output memory description in a public format;

[0043] The recall decision module predicts and makes decisions on the output memory description from the memory neural network processing unit, generates control decisions, and removes the control module; performs memory output processing, generates output memory, and removes the central processing unit;

[0044] The feature memory library is a storage module for intelligent terminal feature memory;

[0045] The local memory is a storage module of the local memory of the intelligent terminal.

[0046] Furthermore, the memory neural network processing unit is used for memory generation, and the modules involved in the memory generation process include: an encoder module, a decision memory module, a vector fusion module, a feature extraction module, an induction module, a decoder module, and a compression module;

[0047] The encoder module converts the sensing numerical matrix / vector and the knowledge base numerical matrix / vector from the perception decision module into an internal vector format;

[0048] The decision memory module predicts and makes decisions on the sensory numerical matrix / vector and the knowledge base numerical matrix / vector from the perception decision module;

[0049] The vector fusion module extracts and splices the output vectors and decisions of the encoder module and the decision memory module;

[0050] The decoder module decodes the output of the vector fusion module to generate a memory stream description in a private format inside the smart terminal;

[0051] The feature extraction module extracts information features from the output of the vector fusion module;

[0052] The output of the feature extraction module of the induction module is summarized to generate generalized and abstract feature memory, which is stored in the feature memory bank;

[0053] The compression module compresses the output of the decoder module, generates local memory, and stores it in the local memory bank.

[0054] Furthermore, the memory neural network processing unit is also used for memory output, and the modules involved in the memory output process include: a decompression module, a memory preprocessing module, an encoder module, a local memory prediction module, a deduction module, a feature memory prediction module, a memory retrieval module, a fusion module, and a decoder module;

[0055] Wherein, the decompression module decompresses data from the local memory in the local memory bank;

[0056] The memory preprocessing module performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the output of the decompression module to generate a local memory numerical matrix / vector;

[0057] The encoder module converts the local memory numerical matrix / vector from the memory preprocessing module into an internal vector;

[0058] The local memory prediction module predicts the local memory numerical matrix / vector from the memory preprocessing module to generate local memory prediction information;

[0059] The deduction module generates a specific memory description for the generalized and abstract feature memory, such as a generalized logical fact description, from the feature memory library;

[0060] The feature memory prediction module predicts the output of the deduction module to generate feature memory and prediction information;

[0061] The memory retrieval module receives the memory slice search item from the perception decision module and performs memory retrieval on the local memory bank and the feature memory bank;

[0062] The fusion module extracts the output vectors to be output from the encoder module, the local memory prediction module, and the feature memory prediction module and splices them;

[0063] The decoder module decodes the output of the fusion module and generates an output memory description in a public format.

[0064] Furthermore, feature memory and local memory participate in the decision-making and control work of the neural network processing unit.

[0065] The feature memory generates feature memory and prediction information through the deduction and prediction reasoning of the deduction module and the feature memory prediction module, and participates in the decision-making and control of the neural network processing unit. The local memory generates an output memory description in a public format through the prediction and reconstruction of the memory preprocessing module, the local memory prediction module, the encoder module, the fusion module, and the decoder module, and participates in the decision-making and control of the neural network processing unit.

[0066] Furthermore, local memory adopts a private format within the smart terminal; local memory can only be output to the outside of the smart terminal after being reconstructed and recalled by the local memory neural network processing unit based on the encoder-vector fusion-decoder framework, thereby improving the privacy of the smart terminal memory and the security of the smart terminal; the memory generated based on the personalized memory neural network processing unit has a private format characteristic.

[0067] The present invention has the following beneficial effects:

[0068] (1) The intelligent terminal system architecture disclosed in the present invention introduces automatic code generation, multi-level memory implementation, multi-dimensional sensor data input, and knowledge base data input, which solves the shortcomings of many current intelligent terminals that cannot automatically generate running codes for upgrade optimization, lack memory, cannot use memory for reasoning, and have weak personalization. It has relatively complete artificial intelligence and can better interact with humans. It can be widely used in various intelligent terminal fields such as robots and smart cars;

[0069] (2) The local memory in the smart terminal system architecture disclosed in the present invention is stored in a private format, which improves the privacy of the memory and the security of the smart terminal;

[0070] (3) The feature memory in the smart terminal system architecture disclosed in the present invention can be used in the reasoning of the smart terminal to improve the reasoning ability of the smart terminal and speed up the reasoning speed.

[0071] The invention is novel, highly operable, has a wide range of applications and great application significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1-1 This is a general diagram of the system architecture of the intelligent terminal provided by an embodiment of the present invention.

[0073] Figure 1-2 It is a structural diagram of the neural network processing unit of the intelligent terminal provided in an embodiment of the present invention.

[0074] Figure 2-1 It is a module composition structure diagram of the intelligent terminal code automatic generation function provided by an embodiment of the present invention.

[0075] Figure 2-2 The present invention provides a flowchart of automatic code generation for a smart terminal.

[0076] Figure 2-3 It is a functional block diagram of an intelligent terminal code automatic generation implementation provided by an embodiment of the present invention.

[0077] Figure 2-4 It is a functional block diagram of a code generation module implemented by an embodiment of the present invention.

[0078] Figure 3-1 It is a module composition structure diagram for realizing the memory generation function of the intelligent terminal provided by an embodiment of the present invention.

[0079] Figure 3-2 It is a functional block diagram of an intelligent terminal memory generation function implementation provided by an embodiment of the present invention.

[0080] Figure 4-1 It is a module composition structure diagram of the intelligent terminal memory output function provided by an embodiment of the present invention.

[0081] Figure 4-2 It is a functional block diagram of an intelligent terminal memory output function implementation provided by an embodiment of the present invention.

[0082] Figure 4-3 It is a schematic diagram of a private domain / shared domain memory of an intelligent terminal provided by an embodiment of the present invention.

[0083] Figure 5 This is a module composition structure diagram for implementing the decision control function of the intelligent terminal provided in the embodiment of the present invention.

[0084] Figure 6 This is a diagram of the architecture of a smart terminal system for a smart car provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0085] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0086] Embodiment 1

[0087] The present invention provides a system architecture of an intelligent terminal, such as Figure 1-1 As shown, it includes: a neural network processing unit 101, a central processing unit 102, a sensor module [1:n] 103, a knowledge base input module 104, a feature memory 105, a local memory 106, an instruction memory 108, an input and output module 110, a control module 107, an execution result storage module 109, etc.; wherein,

[0088] The neural network processing unit 101 performs prediction, decision-making and other processing on the multi-dimensional sensor data 113 from the sensor module [1:n] 103, the feature memory 115 from the feature memory library 105, the local memory 116 from the local memory library 106 and the operation output 120 from the central processing unit 102, generates a control decision 117, and sends it to the control module 107; at the same time, the multi-dimensional sensor data 113 from the sensor module [1:n] 103, the knowledge base data 114 from the knowledge base input module 104, and the operation output data 120 from the central processing unit 102 are predicted, decided, integrated, feature extracted, reconstructed, etc., to generate a local memory 116 and a feature memory 115, and store them in the local memory library 106 and the feature memory library 105; completes the autonomous code generation, and generates a code 118 to go to the instruction memory 108;

[0089] The central processing unit 102 is used to control the intelligent terminal, process transactions and data by running the code in the instruction memory 108;

[0090] The knowledge base input module 104 is an input module for the external input knowledge base 112 of the intelligent terminal, generating knowledge base data 114, and removing the perception decision module 1011 of the neural network processing unit 101; the input knowledge base 112 includes a local knowledge base and a network input knowledge base;

[0091] The sensor module [1:n] 103 is a group of sensor modules for multi-modal sensor input information 111 such as vision, audio, vibration, temperature, touch, and azimuth, generating multi-dimensional sensor data 113, and removing the perception decision module 1011 of the neural network processing unit 101; the multi-modal sensor input information includes multi-modal sensor input information outside the smart terminal and multi-modal sensor input information inside the smart terminal;

[0092] The feature memory 105 is a storage module for the intelligent terminal feature memory 115;

[0093] The local memory 106 is a storage module of the local memory 116 of the intelligent terminal;

[0094] The instruction memory 108 is a memory for the instruction code of the central processing unit 102, the instruction code 118 after the code is automatically generated, or the memory for inputting external instruction code;

[0095] The input and output module 110 is an input and output interface module of the central processing unit, including circuit communication interface, user input and display output, etc.;

[0096] The control module 107 is a control command generation module of the execution device of the intelligent terminal;

[0097] The execution result storage module 109 is a storage module for the central processing unit execution result 119 .

[0098] The neural network processing unit 101 is further described as follows: Figure 1-2 As shown, its composition structure includes: a perception decision module 1011, a recall decision module 1012, a code automatic generation module 1013, and a memory neural network processing unit 1014; wherein the memory neural network processing unit 1014 includes a feature memory processing unit 10141 and a local memory processing unit 10142.

[0099] Among them, the perception decision module 1011 realizes preprocessing of the multi-dimensional sensor data 113 from the sensor module [1:n] 103, the knowledge base data 114 of the knowledge base input module 104, and the operation output 120 of the central processing unit 102, as well as prediction and decision-making on the preprocessing results, generates numerical matrices / vectors, predictions / decisions, and memorizes the neural network processing unit 1014, generates control decisions 117, and controls the module 107 for controlling the execution device of the intelligent terminal.

[0100] The recall decision module 1012 predicts and makes decisions on the output memory description 121 of the memory neural network processing unit 1014, generates a control decision 117, and goes to the control module 107; performs memory output processing, generates output memory, goes to the central processing unit 102, and realizes the memory output of the intelligent terminal.

[0101] The code automatic generation module 1013 realizes code automatic generation, and the generated code 118 goes to the instruction memory 108 .

[0102] The memory neural network processing unit 1014 is composed of a feature memory processing module 10141 and a local memory processing module 10142. It fuses, extracts features, and reconstructs the numerical matrix / vector and prediction / decision of the perception decision module 1011 to generate the internal memory of the intelligent terminal, including the feature memory 115 and the local memory 116, and removes the feature memory library 105 and the local memory library 106 to realize the memory generation of the intelligent terminal; at the same time, it deduces, predicts, and reconstructs the feature memory 115 and the local memory 116 stored in the feature memory library 105 and the local memory library 106 to generate an output memory description 121 in a public format, and removes the recall decision module 1012, the perception decision module 1011, and the automatic code generation module 1013.

[0103] The feature memory processing module 10141 realizes the functions of memory feature extraction and induction during memory generation; and the functions of feature memory deduction and prediction during memory output.

[0104] The local memory processing module 10142 realizes functions such as memory reconstruction and compression during memory generation; and functions such as decompression, prediction, and reconstruction of local memory.

[0105] Embodiment 2

[0106] Figure 2-1 , which is a module composition structure diagram for realizing the automatic code generation function of the intelligent terminal, including: a neural network processing unit 201, a central processing unit 202, an execution result storage module 203, a knowledge base input module 204, a sensor module [1:n] 205, a feature memory library 206, a local memory library 207, an instruction memory 208, an input and output module 209, etc.; wherein, the neural network processing unit 201 includes an automatic code generation module 2011 and a memory neural network processing unit 2012, and the memory neural network processing unit 2012 includes a feature memory processing module 20121 and a local memory processing module 20122.

[0107] The code automatic generation module 2011 is triggered by the multi-dimensional sensor data 229 from the sensor module [1:n] 205, the operation output 210 from the central processing unit 202, etc., to generate code, and the generated executable code 225 is stored in the instruction memory module 208. The code is then executed in the central processing unit 202 to control the intelligent terminal, process transactions and data, etc.

[0108] The automatic code generation process of smart terminals is as follows Figure 2-2 As shown, including:

[0109] S201, demand analysis. The operation output 210 of the central processing unit, the execution result 230 of the central processing unit, and the multi-dimensional sensor data 229, i.e., the problem input and operation status feedback of the intelligent terminal, involve the requirements for the improvement of the central processing unit code, function modification, and function addition, and generate the modification requirements 2201, 2202, and 2203 for code generation.

[0110] S202 , code generation task instruction generation: The code generation modification requirements 2201 , 2202 , and 2203 are converted into task instructions 221 , which are used as inputs of the code generation module 213 .

[0111] S203, code generation. Clean and reconstruct the original source code according to the task instructions, call the corresponding code generation model to modify and add code functions, generate code snippets that meet the requirements, optimize the code, check the syntax, and complete the modification of the original source code; generate output source code 223.

[0112] S204, debugging. Check for code bugs; generate bug-free source code 224.

[0113] S205 , compile and optimize. Compile the code to generate executable code 225 . The generated executable code is stored in the instruction memory module 208 .

[0114] A functional block diagram of an intelligent terminal code automatic generation implementation, such as Figure 2-3 As shown, it includes: requirement analysis modules 2111, 2112, 2113, task generation module 212, code analysis module 214, code generation module 213, debugging module 215, compilation module 216, debugging simulation module 217, in-system testing module 218, source code memory 219, instruction memory 208, central processing unit 202, execution result storage module 203, sensor module [1:n] 205, code generation sample library 232, memory neural network processing unit 2012, and knowledge base input module 204. Figure 2-1 The code automatic generation module 2011 is composed of Figure 2-3 The system is composed of requirement analysis modules 2111, 2112, 2113, a task generation module 212, a code analysis module 214, a code generation module 213, a debugging module 215, a compilation module 216, a debugging simulation module 217, an in-system testing module 218, a source code storage 219, and a code generation sample library 232.

[0115] The demand analysis modules 2111, 2112, 2113 generate code generation modification requirements 2201, 2202, 2203 for the requirements from the central processing unit 202, the execution result storage module 203, and the sensor module [1:n] 205 that require the central processing unit code to be improved, the function to be modified, and the function to be added;

[0116] The task generation module 212 converts the modification requirements 2201, 2202, 2203 for code generation output by the requirement analysis modules 2111, 2112, 2113 into coding task instructions Code_Instr[1], Code_Instr[2], ..., Code_Instr[m], and the associated coding task instructions are combined to generate coding tasks 221, such as Code_Task[1], Code_Taks[2], ..., Code_Task[n], which are used as inputs of the code generation module 212;

[0117] The code analysis module 214 tokenizes the source code from the source code storage 219, performs structure analysis on the source code based on sequence and graph, obtains source code control flow and data flow, and outputs a code sequence 222 containing the source code structure;

[0118] The code generation module 213 performs code preprocessing, code processing, and code checking on the code sequence 222 containing the source code structure according to the code generation task instruction 221, combined with the sample data from the code automatic generation sample library 232, the knowledge base data 228 from the knowledge base input module 204, and the output memory description 231 from the memory neural network processing unit 2012, to generate an output source code 223;

[0119] The debugging module 215 and the debugging simulation module 217 perform bug checks on the code of the code generation module 213 and generate bug-free source code 224;

[0120] The compiling module 216 compiles the output code 224 processed by the debugging module 215 to generate an executable code 225;

[0121] The system test module 218 runs the generated executable code 225 in the central processing unit 202 environment to check whether the operation output of the central processing unit is consistent with the target;

[0122] A functional block diagram of the code generation module 213 is shown in FIG. Figure 2-4 As shown, it includes: a code preprocessing module 2131, a code processing module 2132, and a code optimization module 2133.

[0123] The code preprocessing module 2131 performs the following operations on the code sequence 222 of the code analysis module 214 according to the coding task 221: 1. Cleaning the code; 2. Reconstructing the code when necessary to facilitate subsequent code modification and generation; 3. Retrieving the code, locating and marking the code fragments related to the coding task, locating and marking the places where the code needs to be modified, as a reference for the code processing module;

[0124] The code processing module 2132 processes the output of the code preprocessing module 2131 and completes the modification and addition of code functions: 1. Establishing new data structures, functions, classes, member functions and methods; 2. Modifying related data structures, functions, classes, member functions and methods; 3. Modifying function calls, structure declarations and object generation; 4. Modifying code input / output, including interaction with the user, i.e. reading user input and displaying output; 5. Main function modification, error handling and compilation instruction modification; 6. Modifying modules, package files and other files;

[0125] The code optimization module 2133 evaluates and sorts the different code results generated by the code processing module 2132, selects the best code, checks the logic errors or potential bugs in the code, analyzes the code security, optimizes the code, performs syntax check on the generated code, and generates output source code 223;

[0126] The code automatic generation module can be implemented using a deep learning model, such as a long short-term memory LSTM network with an attention layer, an RNN network, or a Transformer architecture with a self-attention mechanism. A model with natural language processing (NLP) and basic code generation capabilities can be used as a pre-training basis. The pre-trained model has learned the implementation methods of functions, tasks, and functions in code automatic generation from a large amount of open source code. Further simplification is performed to generate a small parameter model suitable for smart terminal devices. The model is further trained for the specific code, specific library, and local and network input knowledge base of the smart terminal used in the function implementation of the smart terminal central processing unit to generate a model for the code automatic generation module of the smart terminal.

[0127] The automatic code generation module 2011 is composed of a neural network processor AI processor, and is also implemented by modules such as CPU, GPGPU, and ASIC accelerator. The central processing unit 202 is composed of a CPU, and is also implemented by modules such as NPU and ASIC accelerator.

[0128] Embodiment 3

[0129] Figure 3-1 , which is a module structure diagram for realizing the memory generation function of the intelligent terminal, including: a neural network processing unit 301, a feature memory library 302, a local memory library 303, a sensor module [1:n] 304, a knowledge base input module 305, and a central processing unit 306. Among them, the neural network processing unit 301 includes a perception decision module 3011 and a memory neural network processing unit 3012, and the memory neural network processing unit 3012 includes a feature memory processing module 30121 and a local memory processing module 30122.

[0130] The perception decision module 3011 preprocesses the multi-dimensional sensor data 309 from the sensor module [1:n] 304, the operation output data 320 of the central processing unit 306, and the knowledge base data 310 of the knowledge base input module 305, and predicts and makes decisions on the preprocessing results, generates numerical matrices / vectors 3211, 3212, 3213, predictions / decisions 322, and de-memorizes the neural network processing unit 3012.

[0131] The local memory processing module 30122 generates local memory 308 by reconstructing and compressing the numerical matrix / vector 3211, 3212, 3213 and prediction / decision 322 of the perception decision module 3011, and stores it in the local memory bank 303 to achieve local memory generation; the feature memory processing module 30121 generates feature memory 307 by extracting and summarizing the features of the numerical matrix / vector 3211, 3212, 3213 and prediction / decision 322 of the perception decision module, and stores it in the feature memory bank 302 to achieve feature memory generation.

[0132] Figure 3-2, which is a functional block diagram for realizing the memory generation function of the intelligent terminal, including: sensor module [1:n] 304, central processing unit 306, knowledge base input module 305, sensor preprocessing module 3121, operation output preprocessing module 3122, knowledge base preprocessing module 3123, encoder modules 3131, 3132, 3133, decision memory module 311, vector fusion module 314, feature extraction module 315, induction module 316, decoder module 317, compression module 318, feature memory module 302, local memory module 303. Among them, the sensor module [1:n] 304 is a group of sensor modules for identifying and detecting the multi-modal sensor input information of the intelligent terminal, and the multi-modal sensor input information includes the external multi-modal sensor input information of the intelligent terminal and the internal multi-modal sensor input information of the intelligent terminal, and generates multi-dimensional sensor data 309. The central processing unit 306 is used to control, process transactions and process data for the intelligent terminal, and generate operation output data 320. The knowledge base input module 305 is an input module for the external input of the knowledge base of the intelligent terminal, and generates knowledge base data 310. The sensor preprocessing module 3121 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the multi-dimensional sensor data 309 from the sensor module [1:n] 304, and generates a sensor numerical matrix / vector 3211. The operation output preprocessing module 3122 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the operation output data 320 from the central processing unit 306, and generates an operation output numerical matrix / vector 3212. The knowledge base preprocessing module 3123 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the knowledge base data 310 from the knowledge base input module 305, and generates a knowledge base numerical matrix / vector 3213. The encoder modules 3131, 3132, 3133 convert the sensing numerical matrix / vector 3211 from the sensing preprocessing module 3121, the operation output numerical matrix / vector 3212 of the operation output preprocessing module 3122, and the knowledge base numerical matrix / vector 3213 of the knowledge base preprocessing module 3123 into internal vector formats 3231, 3232, 3233. The decision memory module 311 predicts and makes decisions on the sensing numerical matrix / vector 3211 from the sensing preprocessing module 3121, the operation output numerical matrix / vector 3212 of the operation output preprocessing module 3122, and the knowledge base numerical matrix / vector 3213 of the knowledge base preprocessing module 3123, and generates predictions / decisions 322. The vector fusion module 314 extracts and concatenates the vectors 3231, 3232, 3233 of the encoder modules 3131, 3132, 3133 and the prediction / decision 322 of the decision memory module 311 to generate a fused vector 324. The decoder module 317 reconstructs the output 324 of the vector fusion module 314 to generate a memory stream description 325 in a private format inside the smart terminal.The feature extraction module 315 extracts information features from the output of the vector fusion module 314. The summarization module 316 summarizes the output 327 of the feature extraction module 315 to generate a generalized and abstract feature memory 307, such as a generalized logical fact description, and stores it in the feature memory 302. The compression module 318 compresses the output of the decoder module 317 to improve memory efficiency, generates a local memory 308, and stores it in the local memory 303.

[0133] Figure 3-1 The perception decision module 3011 is composed of Figure 3-2 It consists of a sensing preprocessing module 3121, an operation output preprocessing module 3122, a knowledge base preprocessing module 3123, and a decision memory module 311. Figure 3-1 The memory neural network processing unit 3012 is composed of a local memory processing module 30122 and a feature memory processing module 30121; wherein the local memory processing module 30122 is composed of encoder modules 3131, 3132, 3133, a vector fusion module 314, a decoder module 317, and a compression module 318; and the feature memory processing module 30121 is composed of a feature extraction module 315 and an induction module 316.

[0134] The internal memory of the intelligent terminal includes local memory 308 and feature memory 307. Local memory 308 is the image memory of the intelligent terminal, which is a mixed memory formed by the following memory contents, including: memory contents of the external input of the intelligent terminal reflected by the multidimensional sensor data 309 from the external sensor module 304 of the intelligent terminal, the knowledge base data 310 from the knowledge base input module 305, and the operation output data 320 from the central processing unit 306; memory contents of the internal operation state feedback of the intelligent terminal reflected by the multidimensional sensor data 309 from the internal sensor module 304 of the intelligent terminal; memory contents of the internal individual understanding, prediction, and decision of the intelligent terminal on the external input at that time reflected by the prediction / decision 322 from the decision memory module 311. Feature memory 307 is an abstract and generalized memory, which is a memory about memory features extracted from the memory data 326, reflecting a more concentrated and flexible knowledge representation than local memory, such as abstract logical facts. Feature memory 302 can also be mapped with local memory 303 to realize a fast retrieval function of local memory. The local memory compresses data when storing to save storage space and improve access speed. The feature memory 302 storage can be implemented by distributed storage or near neural network storage or fusion of neural network and storage or neural network with storage to improve access operation efficiency. The local memory 303 storage is implemented by centralized storage or block centralized storage to improve storage efficiency. As the intelligent terminal operates for a long time, the memory will increase. Some memories have expired, and subsequent memory output and reasoning are not used, which will reduce the memory processing speed of the intelligent terminal and hinder memory generation. At this time, the feature memory processing module 30121 and the local memory processing module 30122 delete and update the memory according to certain rules.

[0135] The memory neural network processing unit 3012 in this embodiment is implemented based on the encoder-vector fusion-decoder model. Under the encoder-vector fusion-decoder framework, the output of the perception decision module 1011 is subjected to feature extraction, reconstruction and other processing to generate the internal memory of the intelligent terminal, including feature memory and local memory. The encoder-vector fusion-decoder can adopt RNN, LSTM, GRU models, introduce attention mechanisms, and improve the accuracy of memory flow. The memory neural network processing unit 3012 has a self-learning function. Through an end-to-end learning memory generation mode, iterative training verification is performed based on memory data to optimize memory capture and generation, feature extraction, and form a personalized model.

[0136] Embodiment 4

[0137] Figure 4-1, which is a module structure diagram for realizing the memory output function of the intelligent terminal, including: a neural network processing unit 401, a central processing unit 402, a sensor module [1:n] 403, a feature memory library 405, a local memory library 406, a control module 407, etc. Among them, the neural network processing unit 401 includes a recall decision module 4011, a memory neural network processing unit 4012, and a perception decision module 4013. The recall decision module 4011 includes a memory output module 40111 and a prediction decision module 40112. The memory neural network processing unit 4012 includes a feature memory processing module 40121 and a local memory processing module 40122.

[0138] The perception decision module 4013 receives the stimulation of the multi-dimensional sensor data 425 from the sensor module [1:n] 403 or the operation output 422 of the central processing unit 402, generates the memory slice search item 426, removes the memory retrieval module 411 of the memory neural network processing unit 4012; generates the memory output control 427, removes the recall decision module 4011, and controls the working mode of the recall decision module 4011. The feature memory processing module 40121 deduce and predict the feature memory 408 of the feature memory bank 405; the local memory processing module 40122 predicts and reconstructs the local memory 409 of the local memory bank 406, and generates the output memory description 419 in the public format. The prediction decision module 40112 predicts and makes decisions on the output memory description 419 of the memory neural network processing unit 4012, generates the control decision 410, removes the control module 407; generates the output memory decision 428, removes the memory output module 40111. The memory output module 40111 performs memory output processing on the output memory description 419 from the memory neural network processing unit 4012 and the output memory decision 428 from the prediction decision module 40112 to generate output memory 429 and send it to the central processing unit 402.

[0139] There are three working modes for the output and use of memory, which are controlled by the memory output control 427 from the perception decision module 4013. The first working mode is memory output, which is stimulated by the multi-dimensional sensor data 425 from the sensor module [1:n] 403 or the operation output 422 of the central processing unit 402, and the feature memory 408 from the feature memory bank 405 and the ground memory 409 from the local memory bank 406 are processed by the feature memory processing module 40121, the local memory processing module 40122, and the recall decision module 4011 to generate the output memory 429, which is then output to the central processing unit 402. The second working mode is memory recall control, which is stimulated by the multi-dimensional sensor data 425 from the sensor module [1:n] 403 or the operation output 422 of the central processing unit 402, and the feature memory 408 from the feature memory bank 405 and the local memory 409 from the local memory bank 406 are processed by the feature memory processing module 40121, the local memory processing module 40122, and the recall decision module 4011 to generate a control decision 410 to the control module 407. The third mode is feature extraction of the local memory, which is stimulated by the multi-dimensional sensor data 425 from the sensor module [1:n] 403 or the operation output 422 of the central processing unit 402, and the local memory 409 from the local memory bank 406 are processed by the feature memory processing module 40121 to generate a feature memory 408 to the feature memory bank 405 for storage.

[0140] Figure 4-2, which is a functional block diagram for realizing the memory output function of the intelligent terminal, including: a local memory library module 406, a decompression module 412, a memory preprocessing module 413, an encoder module 414, a local memory prediction module 415, a feature memory library module 405, a deduction module 416, a feature memory prediction module 417, a memory retrieval module 411, a fusion module 420, a decoder module 421, a recall decision module 4011, a central processing unit 402, and a control module 407. Among them, the decompression module 412 decompresses the memory 409 stored in the local memory library 406. The memory preprocessing module 413 performs preprocessing operations such as data conversion, combination, and dynamic range scaling on the output 418 of the decompression module 412 to generate a local memory numerical matrix / vector 430. The encoder module 414 converts the local memory numerical matrix / vector 430 from the memory preprocessing module 413 into an internal vector 432. The local memory prediction module 415 predicts the local memory numerical matrix / vector 430 of the memory preprocessing module 413 and generates local memory prediction information 431. The deduction module 416 generates a specific memory description for the generalized and abstract feature memory 408 from the feature memory library 405, such as a generalized logical fact description. The feature memory prediction module 417 predicts the output 423 of the deduction module 416 and generates feature memory and prediction information 424. The memory retrieval module 411 receives the memory slice search item 426 from the perception decision module 4013 and performs memory retrieval on the local memory library 406 and the feature memory library 405. The fusion module 420 extracts the vectors to be output from the outputs of the encoder module 414, the local memory prediction module 415, and the feature memory prediction module 417 and splices them. The decoder module 421 decodes the output 433 of the fusion module 420 and generates an output memory description 419 in a public format. The recall decision module 4011 predicts and makes decisions on the output 419 of the decoder module 421, generates a control decision 410, and sends it to the control module 407 of the intelligent terminal; performs memory output processing and sends it to the central processing unit 402.

[0141] Figure 4-1 The memory neural network processing unit 4012 consists of Figure 4-2 It consists of a memory retrieval module 411, a decompression module 412, a memory preprocessing module 413, an encoder module 414, a local memory prediction module 415, a deduction module 416, a feature memory prediction module 417, a fusion module 420, and a decoder module 421.

[0142] The memory neural network processing unit 4012 in this embodiment is implemented based on the encoder-vector fusion-decoder model. Under the encoder-vector fusion-decoder framework, the feature memory 408 stored in the feature memory bank 405 and the local memory 409 of the local memory bank 406 are deduced, predicted, reconstructed, etc. to generate output memory 429. The encoder-vector fusion-decoder can adopt RNN, LSTM, GRU models, introduce attention mechanisms, and improve the accuracy of memory flow. The memory neural network processing unit 4012 has a self-learning function. Through an end-to-end learning memory generation mode, iterative training verification is performed based on memory data to optimize the generation, prediction, and decision-making of memories and form a personalized model.

[0143] Intelligent terminals remember the private domain and shared domain divisions, such as Figure 4-3 It can be seen that the generation and output of the memory of the intelligent terminal are divided into two domains: private domain 434 and shared domain 435. The multi-dimensional sensor data 309, the knowledge base data 310, and the central processing unit operation output 320 are generated by the perception decision module 3011 and the memory neural network processing unit 3012_4012. The feature memory 408 and the local memory 409 generated by the feature memory library 405 and the local memory 409 in the local memory library 406 are output by the memory neural network processing unit 3012_4012 and the recall decision module 4011 to generate the output memory description 419 and the output memory 429 in the public format. The feature memory 408 and the local memory 409 are in a private format and cannot be directly output outside the intelligent terminal. The memory neural network processing unit 4012 has a self-learning function. Through the end-to-end memory generation and recall generation mode, iterative training and verification are performed based on the memory data and recall data to form a personalized model. The memory generated based on the personalized memory neural network processing unit further has private format characteristics.

[0144] Embodiment 5

[0145] Figure 5 , which is a module structure diagram for realizing the decision control function of the intelligent terminal, including: neural network processing unit 501, central processing unit 502, knowledge base input module 504, sensor module [1:n] 503, feature memory 505, local memory 506, control module 507, etc. Among them, the neural network processing unit 501 includes a perception decision module 5011 and a memory neural network processing unit 5012. The perception decision module 5011 includes a prediction decision module 50111, a sensor preprocessing module 50112, a knowledge base preprocessing module 50113, and an operation output preprocessing module 50114. The memory neural network processing unit 5012 includes a feature memory processing module 50121 and a local memory processing module 50122.

[0146] The sensor preprocessing module 50112 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the multi-dimensional sensor data 512 from the sensor module [1:n] 503 to generate a sensor numerical matrix / vector 514. The knowledge base preprocessing module 50113 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the knowledge base data 513 from the knowledge base input module 504 to generate a knowledge base numerical matrix / vector 515. The operation output preprocessing module 50114 performs preprocessing operations such as data conversion, combination, dynamic range scaling, etc. on the operation output 511 from the central processing unit 502 to generate an operation output numerical matrix / vector 516. The prediction and decision module 50111 predicts and makes decision reasoning on the sensor value matrix / vector 514 from the sensor preprocessing module 50112 and the operation output value matrix / vector 516 from the operation output preprocessing module 50114, generates a control decision 510, and sends it to the control module 507; predicts and makes decision reasoning on the sensor value matrix / vector 514 from the sensor preprocessing module 50112 and the knowledge base value matrix / vector 515 from the knowledge base preprocessing module 50113, generates a central processing decision 517, and sends it to the central processing unit 502 for output. The feature memory processing module 50121 deduce and predict the feature memory 508 from the feature memory bank 505, and generates feature memory and prediction information 424; the local memory processing module 50122 decompresses, predicts, and reconstructs the local memory 509 from the local memory bank 506, and generates an output memory description 419, which is used for the prediction and decision reasoning work of the prediction and decision module 50111 in the perception decision module 5011.

[0147] In the decision-making and control process, the memory neural network processing unit 5012 can participate in the decision-making and control work of the neural network processing unit 501. When the multi-dimensional sensor data 512 from the sensor module [1:n] 503 is predicted and decided in the perception decision module 5011, the feature memory 508 and the local memory 509 can participate in the realization of its decision-making and control functions. The feature memory 508 in the feature memory library 505 is deduced by the deduction module 416, and then the output 423 of the deduction module 416 is predicted by the feature memory prediction module 417 to generate feature memory and prediction information 424, see Figure 4-2 , participates in the prediction and decision-making work of the perception decision module 5011 through the local memory processing module 50122. Further, when necessary, the local memory 509 in the local memory bank 506 generates an output memory description 419 through decompression, preprocessing, prediction, reconstruction and other processing by the decompression module 412, the memory preprocessing module 413, the local memory prediction module 415, the encoder module 414, the fusion module 420, and the decoder module 421, see Figure 4-2, participate in the prediction and decision-making work of the perception decision module 5011. Relatively speaking, the prediction and decision-making by referencing feature memory, through modules such as the feature memory library 505, the deduction module 416, and the feature memory prediction module 417, are fast; the prediction and decision-making by referencing local memory, through modules such as the local memory library 506, the decompression module 412, the memory preprocessing module 413, the local memory prediction module 415, the encoder module 414, the fusion module 420, and the decoder module 421, are slow.

[0148] Embodiment 6

[0149] The present invention is a universal intelligent terminal system architecture that can be used in the fields of autonomous driving. The efficient realization of autonomous driving requires turning the car into an intelligent car that continuously learns and interacts with the driver, has human-like thinking, communicates smoothly with the driver and other car drivers and pedestrians, has the most suitable driving habits and style for the driver, and performs autonomous driving according to the driver's intention, rather than just a driving control machine based on vision, radar and other sensors.

[0150] like Figure 6 , which is a system architecture diagram of the smart terminal used in smart cars, including:

[0151] Neural network processing unit 601, central processing unit 602, sensor module [1:n] 603, knowledge base input module 604, feature memory 605, local memory 606, instruction memory 608, input and output module 610, control module 607, execution result storage module 609, etc.; wherein,

[0152] The neural network processing unit 601 predicts and makes decisions on the multi-dimensional sensor data 613 from the sensor module [1:n] 603, the feature memory 615 from the feature memory bank 605, the local memory 616 of the local memory bank 606, and the operation output 620 from the central processing unit 602, and generates a control decision 617 to the control module 607; at the same time, the multi-dimensional sensor data 613 from the sensor module [1:n] 603 and the knowledge base data 614 from the knowledge base input module 604 are pre-processed, and then prediction, decision, feature extraction, and reconstruction are performed to generate feature memory 615 and local memory 616 and store them in the feature memory bank 605 and the local memory bank 606; autonomous code generation is completed, and the generated code 618 is sent to the instruction 623 memory 608;

[0153] The central processing unit 602 is used to control the intelligent terminal, process transactions and data by running the code in the instruction memory 608;

[0154] The knowledge base input module 604 is an input module for the external input knowledge base 612, generating knowledge base data 614, and removing the neural network processing unit 601; the input knowledge base includes a local knowledge base and a network input knowledge base, such as: road data, weather data, map data, humanities data, etc.;

[0155] The sensor module [1:n] 603 is a group of sensor modules for multimodal sensor input information 611, specifically including a visual sensor, a radar sensor, a temperature sensor, a vibration sensor, an automobile exterior / interior audio sensor, an air quality sensor, etc., to generate multidimensional sensor data 613, and remove the neural network processing unit 601;

[0156] The feature memory library 605 is a storage module for intelligent terminal feature memory;

[0157] The local memory 606 is a storage module for the local memory of the intelligent terminal;

[0158] The instruction memory 608 is a memory module for inputting the instruction code 618 or the external instruction code after the code is automatically generated;

[0159] The input and output module 610 is an input and output interface module of the central processing unit 602, including a system upgrade interface, a navigation data interface, a human interface, a communication interface, etc. of the intelligent terminal;

[0160] The control module 607 is a control command generation module of the execution device of the intelligent terminal. The control commands include steering control commands, throttle control commands, brake control commands, horn control commands, air conditioning control commands, door and window control commands, etc.

[0161] The execution result storage module 609 is a storage module for the central processing unit execution result 619 .

[0162] based on Figure 6 The smart car of the smart terminal system architecture not only has necessary sensors such as cameras and radars, but also has sound sensors to sense the sounds of external vehicles, pedestrians, surrounding environment and the vehicle itself, as well as vibration sensors to sense the ground bumps and tire noise of the vehicle, temperature and air quality sensors. These multi-dimensional sensor data not only strengthen the interaction between the car and the driver, but also provide more modal vehicle external information and feedback information for autonomous driving. The more multi-modal information is perceived, the more accurate and closer to reality the autonomous driving algorithm can be established. The smart car also has automatic code generation, which can realize autonomous learning, personalized functions and features. At the same time, the smart car has memory. In extreme cases, such as lack of real-time navigation data, weather, unexpected problems on the surrounding roads that affect the effective acquisition of certain sensor data, or certain sensor failures, it still relies on memory to perform autonomous driving or assist manual driving.

[0163] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. An intelligent terminal system architecture, characterized in that: include: A neural network processing unit, a central processing unit, a sensor module [1:n], a knowledge base input module, a feature memory, a local memory, an instruction memory, an input and output module, a control module, and an execution result storage module, wherein the neural network processing unit includes a code automatic generation module, a memory neural network processing unit, a perception decision module, and a recall decision module; Wherein, the perception decision module realizes preprocessing of the sensor data from the sensor module [1:n], the knowledge base data from the knowledge base input module, and the operation output data from the central processing unit, and predicts and makes decisions on the preprocessing results, generates numerical matrices / vectors, predictions / decisions, de-memorizes the neural network processing unit, generates control decisions, and de-controls the module; The memory neural network processing unit fuses, extracts features, and reconstructs the numerical matrix / vector, prediction / decision from the perception decision module to generate the local memory and feature memory of the intelligent terminal, and removes the local memory library and feature memory library; deduce, predict, and reconstruct the feature memory and local memory stored in the feature memory library and local memory library to generate an output memory description in a public format, and removes the recall decision module, the perception decision module, and the code automatic generation module; The recall decision module predicts and makes decisions on the output memory description from the memory neural network processing unit, generates a control decision, and removes it from the control module; performs memory output processing, generates output memory, and removes it from the central processing unit; The automatic code generation module realizes automatic code generation and generates code to go to the instruction memory; The central processing unit executes the code in the instruction memory for controlling, transaction processing and data processing of the intelligent terminal; The knowledge base input module is an input module for the external input knowledge base of the intelligent terminal, generates knowledge base data, and senses the decision module; the input knowledge base includes a local knowledge base and a network input knowledge base; The sensing module [1:n] is a group of sensing modules for identifying and detecting the multimodal sensing input information of the intelligent terminal, generating multidimensional sensing data, and sensing the decision module; the multimodal sensing input information includes the multimodal sensing input information outside the intelligent terminal and the multimodal sensing input information inside the intelligent terminal; The feature memory library is a storage module for intelligent terminal feature memory; The local memory is a storage module of the local memory of the intelligent terminal; The instruction memory is a memory for the instruction code of the central processing unit, a memory for the instruction code generated by the autonomous code and a memory for the input external instruction code; The input-output module is an input-output interface module of the central processing unit, including a circuit communication interface, a user input and a display output; The control module is a control command generating module of the execution device of the intelligent terminal; The execution result storage module is a storage module for the execution results of the central processing unit.

2. The intelligent terminal system architecture according to claim 1, characterized in that: The automatic code generation module includes: a requirement analysis module, a task generation module, a code analysis module, a code preprocessing module, a code processing module, a code optimization module, a debugging module, a debugging simulation module, a compilation module, an in-system testing module, and a source code storage; Wherein, the demand analysis module generates code generation modification requirements for the inputs from the central processing unit, the execution result storage module, and the sensor module [1:n] involving the requirements for the central processing unit code improvement, function modification, and function addition; The task generation module converts the code generation modification requirements output by the requirement analysis module into coding task instructions, and the associated coding task instructions are combined to generate a coding task; The code analysis module tokenizes the source code from the source code storage, performs structure analysis on the source code based on sequence and graph, obtains source code control flow and data flow, and outputs a sequence containing source code structure; The code preprocessing module performs operations on the output of the code analysis module according to the coding task: cleaning the code; refactoring when necessary; searching the code, locating and marking the code snippets related to the coding task, and locating and marking the places where the code needs to be modified; The code processing module processes the output of the code preprocessing module to complete the following functions: Create new data structures, functions, classes, member functions, and methods; Modify relevant data structures, functions, classes, member functions and methods; Modify function calls, structure declarations, and object generation; Modify code input / output, including interaction with the user, i.e. reading user input and displaying output; Main function modification, error handling, and compilation instruction modification; Modify modules, package files and other files; The code optimization module evaluates and sorts the different code results generated by the code processing module, selects the best code, checks the logic errors or potential bugs in the code, analyzes the code security, optimizes the code, and performs syntax checking on the generated code. The debugging module and the debugging simulation module perform code bug checks on the source code generated by the code optimization module to generate bug-free source code; The compiling module compiles the output code processed by the debugging module to generate executable code; The in-system test module runs the executable code generated by the compiling module in a central processing unit environment to check whether the running output of the central processing unit is consistent with the target; The source code generated by the automatic code generation module is stored in the source code memory for automatic code generation of the next version; the executable code is stored in the instruction memory module and then executed in the central processing unit for controlling the intelligent terminal and performing transaction processing and data processing.

3. The intelligent terminal system architecture according to claim 1, characterized in that: The intelligent terminal system architecture is used to realize an intelligent terminal memory structure, which includes: a perception decision module, a recall decision module, a memory neural network processing unit, a feature memory library, and a local memory library; Wherein, the perception decision module pre-processes the sensor data from the sensor module [1:n], the knowledge base data of the knowledge base input module, and the operation output data of the central processing unit, and predicts and makes decisions on the pre-processing results to generate numerical matrices / vectors, predictions / decisions, and de-memorize the neural network processing unit; The memory neural network processing unit reconstructs and extracts features of the output of the perception decision module based on the encoder-vector fusion-decoder framework, generates local memory and feature memory, and stores them in the local memory bank and the feature memory bank; performs feature memory deduction and prediction on the feature memory and local memory in the feature memory bank and the local memory bank, and reconstructs them based on the encoder-vector fusion-decoder framework to generate an output memory description in a public format; The recall decision module predicts and makes decisions on the output memory description from the memory neural network processing unit, generates control decisions, and removes the control module; performs memory output processing, generates output memory, and removes the central processing unit; The feature memory library is a storage module for intelligent terminal feature memory; The local memory is a storage module of the local memory of the intelligent terminal.

4. The intelligent terminal system architecture according to claim 3, characterized in that: The memory neural network processing unit is used for memory generation. The modules involved in the memory generation process include: encoder module, decision memory module, vector fusion module, feature extraction module, induction module, decoder module, and compression module. The encoder module converts the sensing numerical matrix / vector and the knowledge base numerical matrix / vector from the perception decision module into an internal vector format; The decision memory module predicts and makes decisions on the sensory numerical matrix / vector and the knowledge base numerical matrix / vector from the perception decision module; The vector fusion module extracts and splices the output vectors and decisions of the encoder module and the decision memory module; The decoder module decodes the output of the vector fusion module to generate a memory stream description in a private format inside the smart terminal; The feature extraction module extracts information features from the output of the vector fusion module; The induction module summarizes the output of the feature extraction module to generate generalized and abstract feature memory and stores it in the feature memory library; The compression module compresses the output of the decoder module, generates local memory, and stores it in the local memory bank.

5. The intelligent terminal system architecture according to claim 3, characterized in that: The memory neural network processing unit is also used for memory output. The modules involved in the memory output process include: decompression module, memory preprocessing module, encoder module, local memory prediction module, deduction module, feature memory prediction module, memory retrieval module, fusion module, and decoder module. Wherein, the decompression module decompresses data from the local memory in the local memory bank; The memory preprocessing module performs preprocessing operations including data conversion, combination, and dynamic range scaling on the output of the decompression module to generate a local memory numerical matrix / vector; The encoder module converts the local memory numerical matrix / vector from the memory preprocessing module into an internal vector; The local memory prediction module predicts the local memory numerical matrix / vector from the memory preprocessing module to generate local memory prediction information; The deduction module generates a specific memory description for the generalized and abstract feature memory from the feature memory library; The feature memory prediction module predicts the output of the deduction module to generate feature memory and prediction information; The memory retrieval module receives the memory slice search item from the perception decision module and performs memory retrieval on the local memory bank and the feature memory bank; The fusion module extracts the output vectors to be output from the encoder module, the local memory prediction module, and the feature memory prediction module and splices them; The decoder module decodes the output of the fusion module and generates an output memory description in a public format.

6. The intelligent terminal system architecture according to claim 1, characterized in that: Feature memory and local memory participate in the decision-making and control work of the neural network processing unit; The feature memory generates feature memory and prediction information through the deduction and predictive reasoning of the deduction module and the feature memory prediction module, and participates in the decision-making and control of the neural network processing unit; the local memory generates an output memory description in a public format through the prediction and reconstruction of the memory preprocessing module, the local memory prediction module, the encoder module, the fusion module, and the decoder module, and participates in the decision-making and control of the neural network processing unit.

7. The intelligent terminal system architecture according to any one of claims 3 to 6, characterized in that: The local memory adopts a private format inside the smart terminal; the local memory can only be output to the outside of the smart terminal after being reconstructed and recalled by the local memory neural network processing unit based on the encoder-vector fusion-decoder framework.

Images