Offline computing method and device of intelligent interactive device and interactive device

Abstract

The application discloses an offline computing method and device of an intelligent interactive device and an interactive device, relates to the technical field of intelligent interaction, and comprises the following steps: monitoring a network connection state; in the case that the network connection state is an offline state, determining the computing power requirement of a computing subtask in a to-be-executed interactive task, and matching the computing power requirement with a local resource state; determining whether to broadcast a task transfer request signal to the surrounding environment of the location where the device is located according to the matching result; in response to a task undertaking signal returned by a terminal device in the surrounding environment based on the task transfer request signal, transmitting the computing subtask to the terminal device for execution; receiving the computing result of the computing subtask by the terminal device, and continuing to execute the to-be-executed interactive task according to the computing result. The application realizes the completion of a complex interactive task by using the computing power of the surrounding terminal device in the offline state, so that the interactive device can meet the complex interactive demand of the user at low cost even when there is no network connection.

Description

Technical Field

[0001] This application relates to the field of intelligent interaction technology, and in particular to an offline computing method, apparatus and interactive device for intelligent interactive devices. Background Technology

[0002] With the development of technology, intelligent interactive devices are widely used in life and work, such as smart speakers, door locks, and robots. These devices offer rich functionality and provide users with a convenient interactive experience. However, existing devices often rely on the computing power and data resources of cloud servers when performing complex interactive tasks, resulting in several limitations. For example, network dependence is a significant issue; without a network connection or with an unstable network signal, the functionality of intelligent interactive devices is limited, preventing them from performing complex interactive tasks such as high-precision image recognition or complex voice command processing. This greatly restricts their usage scenarios and application scope. While some solutions exist that allow intelligent interactive devices to perform offline computing without relying on cloud servers—such as storing pre-set simple instructions and models locally for handling common interactive tasks offline—these solutions still cannot meet users' interactive needs for complex tasks. Alternatively, to independently complete complex interactive tasks, high-performance processors are built into intelligent interactive devices, which increases the cost, power consumption, and size of the devices.

[0003] Therefore, how to meet users' complex interaction needs at low cost when smart interactive devices are offline is a problem that urgently needs to be solved. Summary of the Invention

[0004] The main objective of this application is to provide an offline computing method, apparatus, and interactive device for intelligent interactive devices, aiming to solve the technical problem of how to fully meet the various interactive needs of users when the intelligent interactive device is offline.

[0005] To achieve the above objectives, this application proposes an offline computing method for intelligent interactive devices, the offline computing method for intelligent interactive devices comprising: Monitor network connectivity; When the network connection status is offline, determine the computing power requirements of the computational subtasks in the interactive task to be executed, and match the computing power requirements with the local resource status; Based on the matching results, determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device; When a task transfer request signal is broadcast to the surrounding environment of the location of the intelligent interactive device based on the matching result, in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal, the computing subtask is transmitted to the terminal device for execution. The terminal device receives the calculation result of the calculation of the calculation subtask, and continues to execute the interactive task to be executed based on the calculation result.

[0006] In one embodiment, the step of determining the computing power requirements of the computational subtasks in the interactive task to be executed includes: Extract data features from computational subtasks within the interactive task to be executed; The data features are encoded and preprocessed to obtain the feature vector of the computational subtask; The feature vector is input into a pre-trained prediction model, and the pre-trained prediction model predicts the computing power requirement of the computational subtask based on the data features.

[0007] In one embodiment, the data features include data size and operation type, the feature vector is formed by concatenating the data size vector and the operation type vector, the data size vector is obtained by encoding and preprocessing the data size, the operation type vector is obtained by encoding and preprocessing the operation type, and the pre-trained prediction model sequentially includes an input layer, an embedding layer, a hidden layer and an output layer; The step of predicting the computing power requirement of the computational subtask based on the data features using the pre-trained prediction model includes: The feature vector is received through the input layer of the pre-trained prediction model, wherein the number of neurons in the input layer is equal to the dimension of the feature vector; In the embedding layer of the pre-trained prediction model, the operation type vector in the feature vector is mapped to a dense vector of the target dimension to obtain the mapped feature vector, wherein the mapped feature vector is composed of the data size vector and the dense vector; The mapped feature vector is nonlinearly transformed by the hidden layer of the pre-trained prediction model, and the nonlinear transformation result is output through the output layer of the pre-trained prediction model to obtain the computing power requirement of the computational subtask.

[0008] In one embodiment, the step of receiving the feature vector through the input layer of the pre-trained prediction model further includes: Obtain the predicted computing power requirement information of the historical computing subtasks corresponding to the computing subtask, and generate a historical computing power feature vector based on the predicted computing power requirement information; The historical computing power feature vector is fused with the feature vector to form an enhanced feature vector, and the feature vector is updated based on the enhanced feature vector so that the enhanced feature vector can be received by the input layer of the pre-trained prediction model.

[0009] In one embodiment, the embedded layer obtains a mapped feature vector as a mapped enhanced feature vector, which includes the data size vector, the density vector, and the historical computing power feature vector. The hidden layer of the pre-trained prediction model includes a fully connected network branch, a recurrent neural network branch, and a feature fusion layer. The step of performing a nonlinear transformation on the updated feature vector through the hidden layer of the pre-trained prediction model includes: The data size vector and the dense vector are nonlinearly transformed by the fully connected network branch to obtain the first output result; The historical computing power feature vector is processed by the recurrent neural network branch to obtain a second output result; The first output result and the second output result are fused by the feature fusion layer to obtain a fused output result, which is then output as the nonlinear transformation result through the output layer of the pre-trained prediction model.

[0010] In one embodiment, the step of determining the computing power requirements of the computational subtasks in the interactive task to be executed further includes: Obtain the task type of the computational subtask in the interactive task to be executed, and find the computational complexity of the computational subtask in the preset task mapping table according to the task type; If the computational complexity found is less than the preset judgment index, the computational subtask is computed locally to obtain the computational result of the computational subtask. If the computational complexity found is greater than or equal to the preset judgment index, or if no computational complexity is found, the step of determining the computational power requirement of the computational subtask in the interactive task to be executed is performed.

[0011] In one embodiment, the step of determining whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result includes: If the matching result indicates that the local resource status meets the computing power requirement, the computing subtask is computed locally to obtain the computation result of the computing subtask. If the matching result indicates that the local resource conditions do not meet the computing power requirements, a task transfer request signal is broadcast to the surrounding environment of the location of the intelligent interactive device.

[0012] In one embodiment, the step of transferring the computational subtask to the terminal device for execution in response to a task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal includes: In response to task acceptance signals returned by multiple terminal devices in the surrounding environment based on the task transfer request signal, the performance parameters and current load information of the terminal device corresponding to each task acceptance signal are determined; The computing power weight of each terminal device is determined based on its performance parameters and current load information. Based on the computing power weight of each terminal device, the computing subtask is divided into subtask segments corresponding to each terminal device, wherein the computing amount of any subtask segment is matched with the computing power weight of the corresponding terminal device. Each of the subtask segments is transmitted to the corresponding terminal device for execution. After receiving the segment calculation results obtained by each terminal device from the corresponding subtask segment, the calculation results of each segment are merged to obtain the calculation result of the calculation subtask.

[0013] Furthermore, to achieve the above objectives, this application also proposes an offline computing device for an intelligent interactive device, the offline computing device for the intelligent interactive device comprising: The monitoring module is used to monitor network connectivity. An evaluation module is used to determine the computing power requirements of the computational subtasks in the interactive task to be executed when the network connection status is offline, and to match the computing power requirements with the local resource status. The broadcast module is used to determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching results. The transfer module is used to transmit the computational subtask to the terminal device for execution when, based on the matching result, it determines that a task transfer request signal is to be broadcast to the surrounding environment of the location of the intelligent interactive device; in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal. The receiving module is used to receive the calculation result of the terminal device on the calculation subtask, and continue to execute the interactive task to be executed based on the calculation result.

[0014] In addition, to achieve the above objectives, this application also proposes an interactive device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the offline computing method of the intelligent interactive device as described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: This application first monitors network connectivity. When the network connection is offline, it determines the computing power requirements of the computational subtasks within the interactive task to be executed and matches these requirements with local resource availability. This allows for real-time detection of the network connectivity of the intelligent interactive device, enabling timely assessment of its online or offline status. This ensures that when the device's functionality is limited due to lack of network or poor network signal, it can accurately determine whether local resources can meet the computing power requirements, avoiding the inability to complete tasks due to insufficient computing power caused by blindly using local resources. This process rationally selects interactive tasks that can be processed locally and those requiring external assistance. Next, based on the matching results, it broadcasts a task transfer request signal to the surrounding environment to expand the scope of task assistance and find nearby terminal devices capable of handling the computational task. This breaks the limitation of a single device relying solely on itself or the cloud for interactive tasks. This approach addresses the limitations of traditional offline tasks by providing potential computing power support for the continued advancement of complex interactive tasks. Then, in response to a task acceptance signal returned by a terminal device in the surrounding environment based on the task transfer request signal, the computational subtask is transferred to the terminal device for execution. This rationally distributes computational tasks that are difficult to complete efficiently, cost-effectively, and with low power consumption locally to surrounding devices with the capability to handle them, leveraging external computing power to advance complex interactive tasks and broadening the boundaries of task processing capabilities in offline scenarios. Finally, the approach receives the computational results from the terminal device on the computational subtask and continues to execute the pending interactive task based on the results, thus successfully achieving task closure. This allows intelligent interactive devices in offline states to complete complex interactive tasks with the help of external terminal computing power, realizing interactive tasks that were originally impossible due to insufficient local computing power, greatly enhancing the practical application value and user experience of the device in offline scenarios.

[0016] In summary, this application monitors network connectivity, determines the computing power requirements of computational subtasks when offline, matches them with local resources, broadcasts task transfer requests to surrounding devices based on the matching results, transmits computational subtasks after receiving task acceptance signals from receiving terminals, and finally receives the computation results and executes interactive tasks. This avoids the problem of interactive devices being unable to execute complex interactive tasks due to insufficient local resources when offline, and also breaks the limitation of unstable network signals on devices executing complex interactive tasks. It enables the full utilization of the computing power of surrounding terminal devices to complete complex interactive tasks in offline states, expands the application scenarios of intelligent interactive devices, improves their interactive capabilities and user experience in offline scenarios, and allows intelligent interactive devices to successfully execute complex interactive tasks even without a network connection, thereby fully meeting the user's interactive needs. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the offline computing method for the intelligent interactive device of this application (Example 1). Figure 2 A flowchart illustrating the second embodiment of the offline computing method for the intelligent interactive device of this application; Figure 3 A simplified flowchart illustrating the offline computing method for the intelligent interactive device provided in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the module structure of the offline computing device of the intelligent interactive device according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment involved in the offline computing method of the intelligent interactive device in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is: monitoring network connection status; when the network connection status is offline, determining the computing power requirement of the computational subtask in the interactive task to be executed, and matching the computing power requirement with the local resource status; determining whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result; when it is determined to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result, in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal, transmitting the computational subtask to the terminal device for execution; receiving the calculation result of the terminal device on the computational subtask, and continuing to execute the interactive task to be executed based on the calculation result.

[0024] Because existing devices often rely on the computing power and data resources of cloud servers to perform complex interactive tasks, they suffer from numerous limitations. For example, network dependence is a significant issue; without a network connection or with an unstable signal, the functionality of smart interactive devices is limited, preventing them from performing complex interactive tasks such as high-precision image recognition or complex voice command processing. This greatly restricts their usage scenarios and application scope. While some solutions exist that allow smart interactive devices to perform offline computing without relying on cloud servers—such as storing pre-set simple instructions and models locally for handling common interactive tasks when the device is offline—these solutions still cannot meet users' interactive needs for complex tasks. Alternatively, to independently complete complex interactive tasks, high-performance processors need to be built into the smart interactive device, which increases the cost, power consumption, and size. Therefore, how to meet users' complex interactive needs at low cost when smart interactive devices are offline is a pressing problem that needs to be solved.

[0025] This application provides a solution that monitors network connectivity, determines the computing power requirements of computational subtasks when offline, matches them with local resources, broadcasts task transfer requests to surrounding devices based on the matching results, transmits computational subtasks after receiving task acceptance signals from receiving terminals, and finally receives the computation results and executes interactive tasks. This avoids the problem of interactive devices being unable to execute complex interactive tasks offline due to insufficient local resources, and also breaks the limitation of unstable network signals on devices executing complex interactive tasks. It enables the full utilization of the computing power of surrounding terminal devices to complete complex interactive tasks in offline states, expands the application scenarios of intelligent interactive devices, improves their interactive capabilities and user experience in offline scenarios, and allows intelligent interactive devices to successfully execute complex interactive tasks even without a network connection, thereby fully meeting the user's interactive needs.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or interactive device capable of performing the above functions. The following description uses an interactive device as an example to illustrate this embodiment and the subsequent embodiments.

[0027] Based on this, embodiments of this application provide an offline computing method, apparatus, and interactive device method for intelligent interactive devices, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method of this application for fully meeting various user interaction needs when the smart interactive device is offline.

[0028] In this embodiment, the offline computing method, apparatus, and interactive device method of the intelligent interactive device include steps S10 to S50: Step S10: Monitor network connectivity status; Step S20: When the network connection status is offline, determine the computing power requirement of the computational subtask in the interactive task to be executed, and match the computing power requirement with the local resource status. It should be noted that the interactive tasks to be executed refer to various interactive tasks that the intelligent interactive device needs to complete in offline mode, such as high-precision image recognition or complex voice command processing. These tasks often require high computing power to complete complex calculations and data processing. Computational subtasks are the various components that make up the interactive tasks to be executed. Each computational subtask has specific data characteristics and requires a certain amount of computing resources to complete, such as the feature extraction subtask in image recognition interactive tasks or the semantic interpretation subtask in complex voice command processing. Computational power requirements refer to the computing resources required by the computational subtasks during execution, including the processor's computing power, memory capacity, and storage resources. These are predicted based on the data characteristics and operation types of the computational subtasks through corresponding models or methods. Local resource status refers to the computing resources that the intelligent interactive device itself possesses, including processor performance, memory size, storage space, and battery power. These resources determine the complexity of the computational tasks that the device can independently complete in offline mode.

[0029] Additionally, it should be noted that the steps for determining the computational subtasks in the interactive task to be executed include: decomposing the interactive task into multiple subtasks according to the execution flow and functional requirements, and identifying the subtasks that require more computational resources than a preset threshold, so as to use the subtasks as computational subtasks.

[0030] Understandably, when existing devices are offline, they often use local storage to store simple instructions and models to handle common interactive tasks. However, for complex interactive tasks, it is impossible to accurately determine whether local resources can meet their computing power requirements, leading to task execution failure or poor results. Therefore, step S20 can avoid the problems of insufficient resources and task incompleteness caused by blindly executing complex interactive tasks locally. It can accurately identify which computing subtasks can be completed locally and which need to rely on external computing power, thus providing a basis for subsequent task transfer and improving the targeting and success rate of task execution.

[0031] For example, taking a smart speaker as an example, when the smart speaker is offline and the user requests it to perform a complex voice command processing task, the first step is to analyze the goal and requirements of the interaction task and identify the parts that require computational processing. For example, in a voice control task, the device needs to convert the voice signal into a command, involving a processing step that consumes high computational resources for semantic understanding; this is the computational subtask. The device's internal computational resource assessment module first analyzes this computational subtask, extracting data features such as the length of the voice command, the number of keywords contained, and semantic complexity. Using a preset computational power requirement assessment model (e.g., a model trained based on historical data, with voice command features as input and required computational power as output), it calculates the computational power requirement of the voice command processing task. At the same time, it obtains local resource status information such as local processor performance parameters, remaining memory space, and available storage capacity. Then, it compares the computational power requirement with the local resource status one by one to determine whether the local resources can meet the task execution requirements and obtain a matching result.

[0032] In one feasible implementation, the step of determining the computing power requirements of the computational subtasks in the interactive task to be executed in step S20 may include steps S21 to S23: Step S21: Extract the data features of the computational subtasks in the interactive task to be executed; It should be noted that data characteristics refer to the various attributes and information contained in the computational subtask. These attributes and information can reflect the complexity of the task and the demand for computing resources. For example, data size (refers to the scale or size of the data involved in the computational subtask, such as the resolution of an image, the duration of a voice command, etc.) and operation type (refers to the specific computational operation or algorithm required to complete the computational subtask, such as convolution operation in image recognition, Fourier transform in speech processing, etc.).

[0033] Understandably, the lack of precise data feature extraction when determining the computing power requirements of computational subtasks often leads to inaccurate assessments of computing power requirements. For example, a rough estimate based solely on task type fails to consider the specific data conditions of the task, such as the impact of differences in data size and operation type on computing power requirements. Therefore, step S21 avoids significant deviations in computing power requirement assessments due to inaccurate understanding of data features. By extracting the data features of computational subtasks, a comprehensive and detailed understanding of the key information of the task can be obtained, providing a reliable data foundation for accurate prediction of computing power requirements in the future.

[0034] For example, when a smart interactive device needs to perform a high-precision image recognition task, this task, as the interactive task to be performed, may include computational subtasks such as image preprocessing and feature extraction. The device's image processing module will extract data features of the computational subtasks, specifically including information such as image size (e.g., pixel size), number of color channels (e.g., RGB three channels), and number of object categories in the image (e.g., identifying different object categories such as people, vehicles, and buildings in a photo). These data features can reflect the complexity of the image recognition task and the computational resource requirements.

[0035] Step S22: Encode and preprocess the data features to obtain the feature vector of the computational subtask; It should be noted that a feature vector is a vector representation obtained by encoding and preprocessing data features. This representation quantifies the data features into a set of numerical values ​​that can be effectively received and processed by a pre-trained prediction model. The numerical value of each dimension corresponds to a certain aspect of the data features. For example, a vector formed by concatenating the data size vector and the operation type vector can comprehensively describe the key features of the computational subtask.

[0036] Understandably, since data features often exist in their raw form, they may have issues such as inconsistent dimensions and data sparsity, making them difficult to directly utilize by the prediction model or affecting the model's prediction of computing power requirements. Therefore, step S22 avoids the problem of inaccurate prediction of computing power requirements due to ineffective processing of data features. By encoding and preprocessing the data features, they are converted into feature vectors, which better meet the input requirements of the pre-trained prediction model. This enhances the model's ability to identify and process data features, providing a strong guarantee for accurate prediction of computing power requirements.

[0037] For example, after extracting the data features for the image recognition task, the device's data processing module encodes and preprocesses these features. First, the image size is standardized, converting it to a uniform pixel range (e.g., converting pixels from images of different sizes to a range of 0-1000) to eliminate dimensional differences. Then, the number of object categories is one-hot encoded, converting it into a binary vector. For example, if there are 5 object categories to be identified, one-hot encoding will result in a 5-dimensional binary vector. Finally, the processed image size values ​​and the one-hot encoded vector are concatenated to form a complete feature vector. This feature vector comprehensively describes the key features of the computational subtask in a numerical form, preparing the input for the subsequent prediction model.

[0038] Step S23: Input the feature vector into the pre-trained prediction model, and use the pre-trained prediction model to predict the computing power requirement of the computational subtask based on the data features.

[0039] It should be noted that a pre-trained prediction model is a machine learning model that has been trained on a large amount of historical data to learn the mapping relationship between data features and computing power requirements for different types of computing tasks. This model can accept feature vectors as input and output corresponding predicted computing power requirements, providing a basis for determining the computing power requirements of computing subtasks.

[0040] Understandably, existing solutions lack effective prediction models when forecasting computing power requirements, often relying on manual experience or simple rules for estimation. This method is highly subjective, has low accuracy, and cannot adapt to the complexities of different computing subtasks. Therefore, step S23 avoids the problem of inaccurate computing power requirement prediction caused by manual estimation or simple rule judgment. By using a pre-trained prediction model to predict computing power requirements based on feature vectors, the model can make full use of the patterns learned from a large amount of historical data to quickly and accurately obtain the computing power requirements of computing subtasks, providing a scientific basis for subsequent task allocation and resource scheduling.

[0041] For example, the pre-trained prediction model can be a regression model based on an artificial neural network. This model has been trained on a large amount of historical image recognition task data, learning the mapping relationship between the data features of the image recognition task and the computing power requirements. The feature vector obtained above is input into the input layer of the pre-trained prediction model. The hidden layer of the model performs a non-linear transformation on the input feature vector, extracting deep-level information from the feature vector through the calculation of weights and biases between neurons. Then, after processing by an activation function (such as the ReLU function), the model can learn complex patterns in the data. Finally, the output layer of the model outputs a numerical value representing the computing power requirements of the predicted computational subtask, such as predicting the number of CPU cycles and / or GPU computing power units required to complete the image recognition task, thus providing a basis for subsequent computing power allocation and task scheduling on the device.

[0042] In this embodiment, by employing techniques such as extracting data features from computational subtasks, encoding and preprocessing the data features to obtain feature vectors, and using pre-trained prediction models to predict computing power requirements, the problems of strong subjectivity and low accuracy in human experience-based assessment of computing power requirements, as well as the inability of simple rule-based judgments to adapt to the complexities of different tasks, are avoided. This achieves a more accurate, objective, and adaptable prediction of computing power requirements, further improving the scientific nature and accuracy of task allocation and resource scheduling.

[0043] In one feasible implementation, the data features include data size and operation type, the feature vector is formed by concatenating the data size vector and the operation type vector, the data size vector is obtained by encoding and preprocessing the data size, the operation type vector is obtained by encoding and preprocessing the operation type, and the pre-trained prediction model sequentially includes an input layer, an embedding layer, a hidden layer and an output layer. The step S23, which involves predicting the computational power requirement of the computational subtask based on the data features using the pre-trained prediction model, may include steps S231 to S233: Step S231: Receive the feature vector through the input layer of the pre-trained prediction model, wherein the number of neurons in the input layer is equal to the dimension of the feature vector; Understandably, existing prediction models may not have a well-suited input layer that matches the dimensionality of the feature vectors, leading to information loss or incomplete utilization of all information within the feature vectors. For example, when the feature vector dimension is high, insufficient input layer neurons may miss some features; conversely, when the feature vector dimension is low, excessive neurons can lead to resource waste and overfitting risks. Therefore, step S231 avoids information loss or resource waste caused by a mismatch between the number of input layer neurons and the feature vector dimension. By setting the number of input layer neurons to be equal to the feature vector dimension, each feature value in the feature vector can be accurately received and transmitted, providing complete and accurate input data for subsequent embedding and hidden layer processing, thus improving the model's efficiency in utilizing features.

[0044] For example, suppose an intelligent interactive device needs to process a complex speech recognition task. This task, as a computational subtask, has a feature vector composed of a data size vector (representing information such as the duration and sampling rate of the speech signal) and an operation type vector (representing the operation types involved in speech recognition, such as acoustic model calculations and language model calculations), with a dimension of 10. The input layer of the pre-trained prediction model has 10 neurons, each corresponding to one dimension of the feature vector. After the feature vector is generated, it is input element by element into the neurons of the input layer, completing the transmission of the feature vector to the model.

[0045] Step S232: In the embedding layer of the pre-trained prediction model, the operation type vector in the feature vector is mapped to a dense vector of the target dimension to obtain the mapped feature vector, wherein the mapped feature vector is composed of the data size vector and the dense vector. It should be noted that a dense vector is a vector in which each dimension has a value, as opposed to a sparse vector in which most dimensions have zero or empty values. In this step, the dense vector maps the operation type vector to a vector of the target dimension through the embedding layer, which can more effectively represent the information of the operation type and facilitate subsequent nonlinear transformations and feature learning. The mapped feature vector refers to the feature vector obtained after processing in the embedding layer of the pre-trained prediction model. It has mapped the operation type vector in the original feature vector to a dense vector and is concatenated or fused with other parts such as the data size vector. It contains richer and more expressive feature information and can better reflect the comprehensive characteristics of the computational subtask.

[0046] Understandably, since operation type vectors are typically sparse, high-dimensional one-hot encoded vectors, this representation suffers from the curse of dimensionality and fails to effectively capture the semantic relationships and similarities between operation types, leading to poor model performance when processing operation type features. Therefore, step S232 avoids the waste of computational resources and loss of semantic information caused by the high dimensionality and sparsity of operation type vectors, improving the representation quality of operation type features and the processing efficiency of the model. Thus, by mapping the operation type vectors to low-dimensional dense vectors through the embedding layer, the potential semantic information and similarities between operation types can be captured, reducing dimensionality while enhancing the expressive power of the features. The mapped feature vectors integrate the dense operation type vectors and the original data size vectors, forming a more comprehensive and meaningful feature representation, which is beneficial for subsequent hidden layers to perform deeper feature learning and nonlinear transformations.

[0047] For example, the embedding layer of the pre-trained prediction model uses a predefined embedding matrix (e.g., a 5×5 matrix obtained through training) to map the operation type vector (originally a 5-dimensional one-hot encoded vector) to a dense vector space with a target dimension of 3. For example, the operation type vector is [0,1,0,0,0], which is transformed into a dense vector [0.2,0.5,0.1] by linear transformation (matrix multiplication) of the embedding layer. This dense vector is then concatenated with the data size vector (assumed to be [100,200,300,400]) to form the mapped feature vector [100,200,300,400,0.2,0.5,0.1], with a dimension of 7.

[0048] Step S233: The mapped feature vector is nonlinearly transformed through the hidden layer of the pre-trained prediction model, and the nonlinear transformation result is output through the output layer of the pre-trained prediction model to obtain the computing power requirement of the computational subtask.

[0049] Understandably, existing prediction models may only employ simple linear transformations or shallow structures to process feature vectors, making it difficult to capture the complex nonlinear relationships between features. This results in inaccurate predictions of computing power requirements and an inability to adapt to the complex characteristics and diverse scenarios of different computing subtasks. Therefore, step S233 avoids the problem of inaccurate computing power requirement predictions caused by the simple model structure and inability to handle nonlinear feature relationships. By utilizing hidden layers and nonlinear transformations (such as neural network layers using activation functions), the complex interactions and nonlinear relationships between features in the updated feature vectors can be deeply explored, extracting higher-level feature representations. These nonlinearly transformed feature information are ultimately output as accurate computing power requirement predictions through the output layer, improving the accuracy and adaptability of the predictions. This enables the model to better cope with various complex computing subtask scenarios, providing a more reliable basis for offline computing task allocation and resource scheduling for intelligent interactive devices.

[0050] For example, the hidden layer of the pre-trained prediction model consists of two fully connected layers. The first fully connected layer has 10 neurons and uses the ReLU activation function; the second fully connected layer has 5 neurons and also uses the ReLU activation function. The mapped feature vector (7-dimensional) is input into the first fully connected layer. After linear transformation and ReLU activation, a 10-dimensional feature vector is obtained. This 10-dimensional feature vector is then input into the second fully connected layer. After linear transformation and ReLU activation, a 5-dimensional feature vector is obtained. Finally, this 5-dimensional feature vector is input into the output layer (one linear neuron). After linear transformation, a numerical value is output, which is the predicted computational power requirement for the subtask. For example, it is predicted that the number of CPU cycles required to complete the speech recognition task is 1200.

[0051] In this embodiment, by employing techniques such as adapting the feature vector dimension to the input layer of the pre-trained prediction model, mapping the operation type vector to a dense vector in the embedding layer, and performing nonlinear transformation in the hidden layer, the problems of feature vector mismatch with model input, information loss caused by sparse and high-dimensional operation type vectors, and the model's inability to effectively handle nonlinear feature relationships are avoided. This achieves complete input and efficient transmission of feature vectors, semantic enhancement and dimensional optimization of operation type features, and in-depth mining and accurate prediction of complex feature relationships, further improving the accuracy, adaptability, and robustness of computing power demand prediction.

[0052] Step S30: Determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result. It should be noted that the matching result is a conclusion obtained by comparing and analyzing the computing power requirements of the computing subtask with the local resource status, indicating whether the local resources can meet the execution requirements of the computing subtask, that is, whether the processing of the computing subtask can be completed locally; the task transfer request signal is a signal broadcast to the surrounding environment by the intelligent interactive device when it detects that the network connection status is offline and the local resources cannot meet the computing power requirements of the computing subtask. This signal contains relevant information about the computing subtask and aims to request the surrounding terminal devices to take over and execute the computing subtask.

[0053] Understandably, when existing devices lack sufficient local resources, they lack an effective mechanism to find and utilize computing resources in the surrounding environment, causing complex interactive tasks to be unable to proceed. Therefore, step S30 can prevent devices from being unable to perform complex interactive tasks due to local resource limitations when offline, and from being forced to wait for network recovery or be unable to complete the task. By broadcasting a task transfer request signal, the scope of task assistance is expanded, the chance of finding suitable computing resources is increased, and the task processing capability of the device in offline scenarios is enhanced.

[0054] For example, if a smart interactive device determines that local resources cannot meet the computing power requirements of voice command processing tasks, its built-in wireless communication module (such as Bluetooth or Wi-Fi module) will generate a task transfer request signal containing information such as task type, required computing power, and data volume, and broadcast the signal to the surrounding environment within a certain range so that nearby terminal devices that may have the capability to perform the task (such as mobile phones, tablets, etc.) can receive it and thus find external computing power support.

[0055] In one feasible implementation, step S30 may include steps S31-S32: Step S31: If the matching result indicates that the local resource status meets the computing power requirement, the computing subtask is calculated locally to obtain the calculation result of the computing subtask. Understandably, since intelligent interactive devices often lack a sophisticated mechanism for matching local resources with task computing power requirements when offline, they often blindly seek external computing power even when local resources are sufficient, or force local execution when local resources are insufficient, leading to task failure. Therefore, step S31 can avoid the waste of resources caused by relying on external computing power when local resources are sufficient, and the problem of tasks failing to be completed when local resources are insufficient. By accurately matching local resources with computing power requirements, local resources can be fully utilized to complete tasks in a timely manner, improving task execution efficiency, while reducing unnecessary dependence on external devices, and enhancing the autonomy and reliability of the device in offline mode.

[0056] For example, when a smart interactive device performs a high-precision image recognition task, it first calculates the computational requirements of the task through a resource evaluation module, while simultaneously detecting local resources (such as CPU performance and memory size). If local resources are sufficient, the resource evaluation module allocates the task to the local image processing module. The image processing module utilizes local computing resources, such as GPU-accelerated computation, to perform operations such as feature extraction and classification on the image, completing the computational sub-tasks and obtaining image recognition results, such as identifying the category and location of objects in the image.

[0057] Step S32: If the matching result indicates that the local resource status does not meet the computing power requirements, broadcast a task transfer request signal to the surrounding environment of the location.

[0058] Understandably, existing intelligent interactive devices typically lack effective task transfer strategies when offline. When local resources cannot meet computing power requirements, they often have to terminate tasks or wait for network recovery, unable to dynamically utilize computing resources in the surrounding environment. Therefore, step S32 avoids the problem of tasks failing due to insufficient local resources, forcing users to passively wait for network recovery or fail directly. By promptly broadcasting task transfer request signals to the surrounding environment, the device actively seeks available external computing resources, improving the executability and flexibility of tasks in offline states and expanding the functional scope and application scenarios of devices in network-free environments.

[0059] For example, when a smart interactive device performs a complex speech recognition task offline, it detects insufficient local resources. The device's communication module generates a task transfer request signal containing the task type (speech recognition), required computing power (such as the number of CPU cycles), and data volume (speech signal size). The communication module broadcasts this signal via Wi-Fi or Bluetooth to nearby terminal devices (such as mobile phones and tablets) within a certain range, requesting external computing power support to utilize the computing power of surrounding devices to complete the speech recognition task.

[0060] In this embodiment, by employing the technical means of executing computational subtasks locally when local resources meet the requirements and broadcasting task transfer request signals when they do not, the problem of wasting local resources and task failure due to insufficient resources is avoided, thereby further realizing the rational use of resources and efficient execution of tasks.

[0061] Step S40: When it is determined, based on the matching result, to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device, in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal, the computing subtask is transmitted to the terminal device for execution. It should be noted that the task acceptance signal is a signal returned by the surrounding terminal device to the intelligent interactive device after receiving the task transfer request signal, based on its own resource situation and task acceptance willingness, indicating that the terminal device is willing to accept and execute the corresponding computing sub-task.

[0062] It is understandable that, due to the lack of an effective task allocation and transmission mechanism between existing interactive devices and peripheral terminal devices, the computing power of peripheral devices cannot be fully utilized to complete complex interactive tasks. Therefore, step S40 can avoid errors, loss, or failure to execute computational subtasks during transmission, ensuring the accuracy and reliability of task transfer. This enables efficient and accurate transmission of computational subtasks between the device and peripheral terminals, allowing complex interactive tasks to continue to be executed with the help of external computing power, and further expanding the functionality of the device in offline mode.

[0063] For example, when a nearby terminal device (such as a mobile phone) receives a task transfer request signal broadcast by a smart interactive device, the corresponding application on the mobile phone will decide whether to accept the task based on its current performance status (such as processor load, battery level, etc.) and task acceptance strategy (such as prioritizing tasks with high computing power requirements). If the mobile phone decides to accept the task, it will send a task acceptance signal containing device identification information and its own performance parameters to the smart interactive device. After receiving the signal, the smart interactive device will transmit the data and execution instructions required for the voice command processing task to the mobile phone through an established secure data transmission channel (such as an encrypted Bluetooth connection). The mobile phone will then use its own computing power to process and calculate the voice commands and execute the corresponding computational subtasks.

[0064] In one feasible implementation, the step S40, which responds to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal, may include steps S41 to S44: Step S41: In response to the task acceptance signals returned by multiple terminal devices in the surrounding environment based on the task transfer request signal, determine the performance parameters and current load information of the terminal device corresponding to each task acceptance signal; It should be noted that performance parameters refer to the hardware configuration and computing power indicators of the terminal device, such as CPU clock speed, GPU performance, memory size, memory bandwidth, etc. These parameters determine the speed and efficiency of the device in processing computing tasks; current load information refers to the number of tasks currently running on the terminal device, CPU utilization, memory usage, etc., reflecting the current busyness of the device and the availability of computing resources.

[0065] Understandably, when multiple terminal devices are available to handle computing tasks, it is often impossible to fully understand the performance and load of surrounding terminal devices during task allocation. This can lead to unreasonable task allocation, with some devices potentially overloaded while others remain idle, affecting task execution efficiency. Therefore, step S41 avoids the problem of uneven task allocation caused by insufficient understanding of terminal device performance and load information. By obtaining the performance parameters and current load information of the terminal devices, the actual computing power of each device can be accurately assessed, providing a basis for subsequent reasonable task allocation.

[0066] For example, when a smart interactive device performs a complex video processing task, it broadcasts a task transfer request signal to its surrounding environment. Multiple terminal devices (such as mobile phones and tablets) receive the signal and, based on their own performance and load, return a task acceptance signal. Upon receiving these signals, the device obtains performance parameters (e.g., CPU clock speed of 2.5GHz, memory of 8GB, and a specific GPU model) and current load information (e.g., CPU utilization of 30%, memory usage of 40%, and the number of running applications of 3) from each returning terminal device via its network communication module.

[0067] Step S42: Determine the computing power weight of each terminal device based on the performance parameters and current load information of each terminal device. It should be noted that the computing power weight is a comprehensive index calculated based on the performance parameters and current load information of the terminal device. It is used to measure the relative computing power of the terminal device in the current state. The higher the weight, the stronger the computing power that the device can currently provide.

[0068] Understandably, existing solutions lack scientific methods for evaluating computing power, making it impossible to accurately measure the actual computing power of different terminal devices under different load conditions. This results in the inability to fully leverage the advantages of each device during task allocation. Therefore, step S42 avoids the problem of unreasonable task allocation caused by inaccurate computing power evaluation. By comprehensively considering performance parameters and current load information, the computing power weight of each terminal device can be determined, which can more accurately reflect the actual computing power of the devices and provide a scientific basis for subsequent task splitting and allocation.

[0069] For example, when performing complex video processing tasks, the intelligent interactive device collects performance parameters and current load information from multiple surrounding terminal devices. Based on this information, the task management module within the device calculates the computing power weight for each terminal device using a predefined weighting formula (e.g., assigning different weights based on CPU frequency, memory size, and GPU performance, then adjusting based on the current load). For instance, terminal device A has excellent performance parameters and a low current load, resulting in a weight of 0.5; terminal device B has slightly lower performance and a higher load, resulting in a weight of 0.2; and terminal device C has moderate performance and a moderate load, resulting in a weight of 0.3.

[0070] Step S43: Based on the computing power weight of each terminal device, the computing subtask is divided into subtask segments corresponding to each terminal device, wherein the computing amount of any subtask segment is matched with the computing power weight of the corresponding terminal device. It should be noted that a subtask segment refers to a computational subtask that is broken down into multiple smaller, independently executable task units according to certain rules. For example, the feature extraction subtask in an image recognition interactive task can be divided into feature extraction subtask segments for different regions based on the image region. Similarly, the semantic interpretation subtask in complex voice command processing can be divided into semantic interpretation subtask segments for different time segments based on time segments. The computational load of each subtask segment is matched with the computational capability weight of the corresponding terminal device so that each terminal device can efficiently complete its respective part of the task.

[0071] Understandably, when tasks are split, a simple average or random allocation method is often used, without fully considering the actual computing power of the terminal devices. This results in some devices having too heavy a workload while others have too light a workload, affecting the overall efficiency of task execution. Therefore, step S43 can avoid the problem of some devices being overloaded and others being idle due to unreasonable task splitting. By splitting tasks according to computing power weights, the computing power of each sub-task segment is matched with the capabilities of the terminal devices, which can give full play to the computing advantages of each device, achieve efficient parallel processing of tasks, and improve the overall task execution efficiency.

[0072] For example, when performing complex video processing tasks, intelligent interactive devices divide the task into three sub-task segments based on the computational power weights of each terminal device (e.g., device A with a weight of 0.5, device B with a weight of 0.2, and device C with a weight of 0.3). Device A undertakes the most computationally intensive part, such as processing special effects for 50% of the video frames; devices B and C undertake processing for 20% and 30% of the frames, respectively. The computational load of each sub-task segment is matched to the computational power weight of the corresponding terminal device, ensuring that each device can efficiently complete its respective task portion.

[0073] Step S44: Each of the subtask segments is transmitted to the corresponding terminal device for execution. After receiving the segment calculation results obtained by each terminal device from the corresponding subtask segment, the segment calculation results are fused to obtain the calculation result of the calculation subtask.

[0074] Understandably, existing solutions lack effective coordination and fusion mechanisms during task execution and result integration, leading to inconsistent task execution results or ineffective integration, which affects the final task completion quality and correctness. Therefore, step S44 can avoid task failure or result errors caused by incoordination between task execution and result integration. By transmitting subtask fragments to corresponding terminal devices for parallel execution and then fusing the calculation results of each fragment, distributed processing of complex computing tasks can be achieved, improving task execution speed while ensuring the integrity and correctness of the final result, and enhancing the robustness and fault tolerance of the system.

[0075] For example, when performing complex video processing tasks, the intelligent interactive device transmits the split sub-task segments to their respective terminal devices. For instance, sub-task segment A is sent to terminal device A via a secure Wi-Fi connection, sub-task segment B to terminal device B, and sub-task segment C to terminal device C. After each terminal device completes its sub-task processing, it returns the result (such as processed video frame data) to the intelligent interactive device. Upon receiving the calculation results of all segments, the device merges the calculation results according to predefined fusion rules (such as splicing according to video time sequence, color correction, etc.) to generate a complete video processing result, such as a complete video with added effects.

[0076] In this embodiment, a multi-device collaborative processing technique is adopted. If multiple terminal devices can undertake computing tasks, after receiving task acceptance signals from multiple terminal devices, the performance parameters and current load information of each device are further determined. Based on this information, the computing power weight of each device is determined, and the computing subtask is reasonably divided into multiple subtask segments and allocated to each terminal device for execution. Finally, the computing results of each segment are merged to obtain the final computing result. This avoids the problem of some devices being overloaded or idle due to uneven task allocation, as well as the inefficiency and inaccurate results that may result from a single device processing complex tasks. This achieves multi-device collaborative task processing, improves task execution efficiency and result accuracy, enhances the overall performance and reliability of the system, and makes full use of the computing resources of each device by reasonably allocating tasks.

[0077] Step S50: Receive the calculation result of the terminal device on the calculation subtask, and continue to execute the interactive task to be executed based on the calculation result.

[0078] It should be noted that the calculation results are the processing results obtained by the terminal device after receiving and executing the calculation sub-task. These results will be received by the intelligent interactive device and used as the basis for continuing to execute the subsequent interactive tasks to be executed after the calculation sub-task.

[0079] Understandably, even when existing devices are offline, they lack an effective mechanism for receiving and integrating computation results, even if they can distribute tasks to surrounding devices. This results in the inability to complete the final interactive task. Therefore, step S40 avoids the problem of interactive task interruption or failure due to the inability to effectively receive and utilize the computation results returned by surrounding devices. This achieves a complete closed loop from task distribution to result reception, enabling intelligent interactive devices to complete complex interactive tasks with the help of surrounding computing power in offline states, thereby improving the device's interactive capabilities and user experience in offline scenarios.

[0080] For example, after the terminal device completes the computational subtask of voice command processing, it sends the obtained computational results (such as the instruction content corresponding to the voice command, semantic understanding results, etc.) back to the smart interactive device (such as a smart speaker) through the aforementioned data transmission channel. After receiving the computational results, the smart speaker verifies and parses them, and then executes the corresponding interactive actions according to the results, such as playing music, querying information and making voice broadcasts, thereby completing the interactive task to be executed, which was initially initiated by the user, of processing the voice command.

[0081] This embodiment provides an offline computing method for intelligent interactive devices. By monitoring network connectivity, the computing power requirements of sub-tasks are determined and matched with local resources when offline. Based on the matching results, a task transfer request is broadcast to surrounding devices. After receiving the task acceptance signal from the terminal, the computing sub-task is transmitted. Finally, the computing results are received and the interactive task is executed. This avoids the problem of interactive devices being unable to execute complex interactive tasks due to insufficient local resources when offline. It also breaks the limitation of unstable network signals on the execution of complex interactive tasks. It enables the full utilization of the computing power of surrounding terminal devices to complete complex interactive tasks in offline mode, expands the application scenarios of intelligent interactive devices, improves their interactive capabilities and user experience in offline scenarios, and allows devices to successfully execute complex interactive tasks even when there is no network connection or poor signal, thereby fully meeting the user's interactive needs.

[0082] In one feasible implementation, steps S201-S202 may be included before step S231: Step S201: Obtain the predicted computing power requirement information of the historical computing subtasks corresponding to the computing subtask, and generate a historical computing power feature vector based on the predicted computing power requirement information. It should be noted that historical computational subtasks refer to computational tasks that the intelligent interactive device has previously executed that are similar in type or characteristics to the current computational subtask. For example, in image recognition tasks, historical computational subtasks could be computational subtasks of image recognition tasks for different images previously executed by the intelligent interactive device. Predicted computing power requirement information refers to the computing power required for the execution of historical computational subtasks as estimated in advance through a certain prediction model or method. It is usually expressed in specific values ​​or indicators, such as the number of CPU cycles or memory usage. The historical computing power feature vector is a vector representation obtained after encoding and processing the computing power requirement information of historical computational subtasks. This representation can reflect the characteristics of historical tasks in terms of computing power requirements and can be used to assist in predicting the computing power requirements of the current computational subtask.

[0083] Understandably, existing solutions often only consider the characteristics of the current task when predicting the computing power requirements of computational subtasks, ignoring useful information that may be contained in historical tasks. This may lead to inaccurate estimation of the computing power requirements of the current task by the prediction model, especially when the task features are similar to those of historical tasks. Therefore, step S201 can avoid the problem of inaccurate computing power requirement prediction caused by ignoring historical task information. By obtaining the predicted computing power requirement information of historical computational subtasks and generating historical computing power feature vectors, the computing power features of historical tasks can be integrated into the prediction of the current task, providing richer information for the pre-trained prediction model and thus improving the accuracy of the prediction of the computing power requirements of the current computational subtask.

[0084] For example, in a computational subtask of processing a complex speech signal, the intelligent interactive device divides the speech task into multiple steps, such as preprocessing, feature extraction, and language decoding. The device retrieves historical records similar to the current task from locally stored historical speech task data. These records contain previously processed speech tasks and their predicted computational requirements. Specifically, the device collects predicted computational requirements (such as CPU utilization and memory usage) for speech processing tasks with similar speech signal lengths, languages, and background noise types. These predicted values ​​are then normalized and feature-encoded to generate a historical computational requirement feature vector. For instance, if three historical tasks are selected, and their CPU utilization (80%, 75%, 78%) and memory usage (60%, 65%, 58%) are recorded, these values ​​are standardized and concatenated into a 6-dimensional historical computational requirement feature vector (e.g., [0.8, 0.75, 0.78, 0.6, 0.65, 0.58]) to assist in predicting the computational requirements of the current task.

[0085] Step S202: The historical computing power feature vector is fused with the feature vector to form an enhanced feature vector, and the feature vector is updated based on the enhanced feature vector so that the enhanced feature vector can be received by the input layer of the pre-trained prediction model.

[0086] It should be noted that the enhanced feature vector is a new feature vector formed by fusing the feature vector of the current computing subtask with the historical computing power feature vector. This vector integrates the features of the current task and historical computing power information, and can more comprehensively reflect the computing power requirements of the current computing subtask, which helps to improve the accuracy of prediction.

[0087] Understandably, since the feature vector may only contain the features of the current computational subtask and lacks reference to the computational power requirements of historical tasks, this limits the predictive model's comprehensive understanding of the task's computational power requirements. Especially when there is correlation between tasks, it may lead to bias in the prediction results. Therefore, step S202 can avoid the problem of incomplete estimation of computational power requirements by the predictive model due to incomplete feature vector information. By fusing historical computational power feature vectors with current feature vectors to form an enhanced feature vector, it can more comprehensively describe the features of the current computational subtask. This allows the pre-trained predictive model to consider both the features of the current task and the computational power information of historical tasks. This helps the model to more accurately predict the computational power requirements of the current task and improve the reliability and accuracy of the prediction results.

[0088] For example, in a speech signal processing task, the intelligent interactive device has already extracted the feature vector of the current task, which is formed by concatenating a data size vector and an operation type vector. For example, the data size vector is [10, 16000] (representing a speech signal duration of 10 seconds and a sampling rate of 16000Hz), and the operation type vector is [0, 1, 0, 0] (representing acoustic model calculation in speech recognition). The device obtains predicted computing power requirement information from historical tasks and generates historical computing power feature vectors, such as [0.8, 0.75, 0.78, 0.6, 0.65, 0.58]. Then, the historical computing power feature vector is fused with the feature vector of the current task, i.e., an enhanced feature vector [10, 16000, 0, 1, 0, 0, 0.8, 0.75, 0.78, 0.6, 0.65, 0.58] is formed by concatenation. Finally, this enhanced feature vector is input into the input layer of the pre-trained prediction model to replace the original feature vector for more accurate prediction of computing power requirements.

[0089] In this embodiment, by acquiring the predicted computing power requirement information of historical computing subtasks and generating historical computing power feature vectors, and then fusing them with the feature vector of the current computing subtask to form an enhanced feature vector, the problem of inaccurate predicted computing power requirement that may be caused by relying solely on the features of the current task is avoided. Especially when the features of the current task are incomplete or have special correlations, the integration of historical information can provide a more comprehensive reference. This realizes the use of computing power information from historical tasks to enhance the accuracy and robustness of the model in predicting the computing power requirement of the current task, making the prediction results more reliable and further improving the efficiency of the rational allocation and utilization of computing resources by intelligent interactive devices in offline states.

[0090] In one feasible implementation, the embedded layer obtains a mapped feature vector as a mapped enhanced feature vector, which includes the data size vector, the density vector, and the historical computing power feature vector. The hidden layer of the pre-trained prediction model includes a fully connected network branch, a recurrent neural network branch, and a feature fusion layer. After step S232, steps S203 to S205 may also be included: Step S203: Perform a nonlinear transformation on the data size vector and the dense vector through the fully connected network branch to obtain the first output result; It should be noted that a fully connected network branch refers to a structure in a neural network in which each neuron is fully connected to all neurons in the previous layer. It is mainly used to perform nonlinear transformations on input features to extract higher-order combinations of features. The first output result refers to the output feature vector obtained by the fully connected network branch after performing nonlinear transformations on the data size vector and the dense vector. This vector contains feature information of the data size and operation type after nonlinear processing.

[0091] Understandably, when processing data size vectors and dense vectors, simple linear transformations or single-layer neural networks are usually used, which makes it difficult to capture the complex nonlinear relationships between features, resulting in insufficient and inaccurate feature representations. Therefore, step S203 can avoid the problem of insufficient feature representation capabilities caused by insufficient feature transformation. Through multi-layer nonlinear transformations of fully connected network branches, it is possible to deeply explore the complex feature relationships in data size vectors and dense vectors, extract higher-level feature representations, and provide richer information for subsequent feature fusion.

[0092] For example, in a speech signal processing task, the intelligent interactive device extracts the feature vector of the current task. The data size vector is [10, 16000] (representing the speech signal duration of 10 seconds and sampling rate of 16000Hz), and the operation type vector is mapped to a dense vector [0.2, 0.5, 0.3] (representing the relevant features calculated by the acoustic model) through an embedding layer. The fully connected network branch contains two fully connected layers: the first layer has 64 neurons and uses the ReLU activation function; the second layer has 32 neurons and also uses the ReLU activation function. The data size vector and the dense vector are concatenated into a 5-dimensional vector [10, 16000, 0.2, 0.5, 0.3], which is then input into the first fully connected layer. After linear transformation and ReLU activation, a 64-dimensional feature vector is output. This feature vector is then input into the second fully connected layer, and after linear transformation and ReLU activation, a 32-dimensional first output result is obtained.

[0093] Step S204: Process the historical computing power feature vector through the recurrent neural network branch to obtain the second output result; It should be noted that a recurrent neural network branch is a neural network structure used to process sequential data. By introducing a recurrent structure, this branch can process the input historical computing power feature vector and capture its temporal characteristics and contextual information. The second output result refers to the output feature vector obtained by the recurrent neural network branch after processing the historical computing power feature vector. This vector contains the feature information of the historical computing power features after temporal processing.

[0094] Understandably, when processing historical computing power feature vectors, their time-series characteristics are usually ignored, and only simple methods are used, such as directly taking the average or maximum value. This cannot fully utilize the time-series information in historical computing power features, causing the prediction model to be unable to effectively utilize the dynamic change patterns in historical data. Therefore, step S204 can avoid the problem of information loss caused by ignoring the time-series characteristics of historical computing power feature vectors. By processing historical computing power feature vectors through recurrent neural network branches, the patterns and trends of historical computing power features changing over time can be captured, and high-order information with time-series characteristics can be extracted, providing historical dynamic feature support for subsequent feature fusion.

[0095] For example, if the historical computing power feature vector is [0.8, 0.75, 0.78, 0.6, 0.65, 0.58], it represents the CPU utilization and memory usage of three historical tasks. The recurrent neural network branch uses one LSTM layer with 32 hidden units. The historical computing power feature vector is reshaped into a sequence with a shape of 3×2 (3 time steps, 2 features per time step). The sequence is as follows: Time step 1: [0.8, 0.6] (CPU utilization 80%, memory usage 60%); Time step 2: [0.75, 0.65] (CPU utilization 75%, memory usage 65%); Time step 3: [0.78, 0.58] (CPU utilization 78%, memory usage 58%). This sequence is input into the LSTM layer. The LSTM layer processes the features of each time step and outputs the hidden state of the last time step as the second output result, with a shape of 32.

[0096] Step S205: The first output result and the second output result are fused through the feature fusion layer to obtain a fused output result, and the fused output result is output as the nonlinear transformation result through the output layer of the pre-trained prediction model.

[0097] It should be noted that the feature fusion layer is a neural network layer used to fuse features from different branches (such as fully connected network branches and recurrent neural network branches). Through specific fusion strategies (such as concatenation, weighted summation, etc.), the output results of different branches are combined to form a comprehensive feature representation. The fusion output result refers to the final feature vector obtained by the feature fusion layer after fusing the first output result of the fully connected network branch and the second output result of the recurrent neural network branch. This vector integrates information about data size, operation type, and historical computing power characteristics.

[0098] Understandably, when fusing features from different sources, improper fusion may lead to insufficient information interaction between features, failing to fully leverage the advantages of each feature and affecting the performance of the prediction model. Therefore, step S205 avoids the problem of limited model performance improvement due to insufficient feature fusion. This is achieved by deeply fusing the first output of the fully connected network branch and the second output of the recurrent neural network branch through a feature fusion layer. This integrates multiple aspects of information, including data size, operation type, and historical computing power characteristics, to form a comprehensive feature representation. This fused feature vector contains rich information, more accurately reflecting the computational requirements of the computational subtask and improving the prediction accuracy and generalization ability of the pre-trained prediction model.

[0099] For example, suppose the first output of the fully connected network branch is a 32-dimensional vector [x1,x2,...,x32], and the second output of the recurrent neural network branch is a 32-dimensional vector [y1,y2,...,y32]. The feature fusion layer uses a simple concatenation method to combine these two vectors into a 64-dimensional fused output [x1,x2,...,x32,y1,y2,...,y32]. This fused output contains comprehensive information about the data size vector, the density vector, and the historical computing power feature vector. The fused output is input into the output layer of the pre-trained prediction model (a linear layer that outputs a single value). After a linear transformation, the final predicted computing power requirement is obtained, for example, predicting that the number of CPU cycles required to complete the speech signal processing task is 1450.

[0100] In this embodiment, fully connected network branches and recurrent neural network branches are used to process the data size vector, dense vector, and historical computing power feature vector, respectively. The processing results are then fused through a feature fusion layer. This avoids the problem that a single network structure cannot fully explore the complex correlations and temporal information between different feature types, further improving the comprehensiveness and depth of feature extraction. This enables a more accurate comprehensive analysis of data features and historical information, thereby enhancing the model's prediction accuracy for computing power requirements of computational subtasks and its adaptability to diverse tasks.

[0101] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Before the step of determining the computing power requirements of the computational subtasks in the interactive task to be executed in step S20, the offline computing method of the intelligent interactive device may further include steps S01 to S03: Step S01: Obtain the task type of the computational subtask in the interactive task to be executed, and find the computational complexity of the computational subtask in the preset task mapping table according to the task type. It should be noted that task type refers to the category to which the computational subtask belongs, such as image recognition, speech processing, natural language processing, etc. Each task type has different characteristics and computational requirements. The preset task mapping table is a data table stored in the intelligent interactive device, which records the mapping relationship between different task types and their corresponding computational complexity. For example, image recognition tasks may be mapped to high complexity, while simple text query tasks may be mapped to low complexity. Computational complexity refers to the amount of computing resources required to complete a certain computational subtask, usually measured by time complexity or space complexity, reflecting the difficulty of task execution and the demand for computing resources.

[0102] Understandably, when processing computational subtasks, it is often impossible to quickly and accurately assess the computational complexity of a task. This may lead to the device being unable to allocate resources reasonably when processing tasks, affecting task execution efficiency. Therefore, step S01 can avoid the problem of unreasonable resource allocation caused by the inability to accurately assess the computational complexity of a task. By obtaining the task type and looking up the computational complexity in the preset task mapping table, the computational requirements of the task can be quickly and accurately assessed, providing a scientific basis for subsequent task allocation and resource scheduling.

[0103] For example, before executing a computational subtask, the intelligent interactive device first identifies the task type. The device has a built-in preset task mapping table that records the correspondence between different task types and computational complexity. For example, image recognition tasks are marked as high complexity, while text query tasks are marked as low complexity. If the computational subtask of the interactive task to be executed is an image recognition task, the device determines that the computational complexity of the current image recognition task is high by looking up this table.

[0104] Step S02: If the computational complexity found is less than the value of the preset judgment index, the computational subtask is calculated locally to obtain the computational result of the computational subtask. Understandably, devices often upload all tasks to the cloud for processing, or process them blindly locally, or directly predict and calculate computing power requirements without being able to allocate resources reasonably based on task complexity. This can lead to waste of local resources or excessive cloud load. Therefore, step S02 can avoid task processing delays caused by insufficient local processing power or excessive cloud load. By processing tasks with low computational complexity locally, local resources can be fully utilized, reducing excessive reliance on the cloud or mobile devices, improving task processing efficiency and device autonomy.

[0105] For example, when the computational subtask is a text query, the device searches the preset task mapping table and finds that its computational complexity is low, below the value of the preset judgment index. In this case, the device directly calls local resources, such as CPU and memory, to quickly complete the text query task without the need for complex computing power assessment or cloud processing, thus improving response speed.

[0106] Step S03: If the computational complexity found is greater than or equal to the value of the preset judgment index, or if no computational complexity is found, execute the step of determining the computing power requirement of the computational subtask in the interactive task to be executed. Understandably, when computational complexity is high, the device may directly attempt to process the task using local resources, leading to local resource overload and task execution failure, or it may directly wait for user intervention, affecting the timeliness of the task. When the task type is not defined, the device usually cannot process it effectively, resulting in task failure. Therefore, step S03 can avoid task processing failure or delays caused by insufficient local resources or unknown tasks. By dynamically assessing the computing power requirements of the task, the device can reasonably allocate resources or find external computing power support to ensure the smooth completion of the task, enhance the flexibility and versatility of the device, and improve the user experience.

[0107] For example, for computationally complex tasks (such as speech recognition) or tasks that are not matched, the device proceeds to the step of determining computing power requirements. For instance, due to the high complexity of image recognition tasks, the device needs to further assess their computing power requirements. If an unknown task type is encountered, the device also performs this step to allocate resources appropriately or seek external computing power support to ensure the task is completed efficiently.

[0108] In this embodiment, by first obtaining the task type of the computational subtask and matching its computational complexity, the execution method of the task is determined based on the size of the complexity or whether the complexity can be found. This avoids the waste of resources and processing delays caused by uniformly assessing the computing power requirements of all tasks, as well as the risk of failure caused by improper handling of high-complexity or unknown tasks. This enables hierarchical processing of tasks with different complexities, optimizes resource allocation, improves device response speed and task processing efficiency, and enhances the adaptability and flexibility of the device to various tasks.

[0109] For example, to help understand the implementation process of the offline computing method for the intelligent interactive device obtained by combining this embodiment with the above embodiment one, please refer to... Figure 3 , Figure 3 A simplified flowchart of an offline computing method for intelligent interactive devices is provided, specifically: When the intelligent interactive device is offline, the computational subtasks in the received interactive task are assessed for task type. If the assessment result is a low-complexity task, the computational subtask is executed locally to obtain a calculation result for executing the interactive task. If the assessment result is a high-complexity task or the task type cannot be identified, the computational power requirement of the task is evaluated to determine the computational power requirement, and this requirement is matched with local resource conditions. If the local resource conditions meet the computational power requirement, the match is successful, and the computational subtask is executed locally to obtain a calculation result for executing the interactive task. If the local resource conditions do not meet the computational power requirement, the match fails, and a task transfer process is initiated. This includes: broadcasting a task transfer request signal to the surrounding environment of the intelligent interactive device's location, and responding to a task acceptance signal returned by a terminal device in the surrounding environment based on the task transfer request signal, transferring the computational subtask to the terminal device for execution, and then receiving the calculation result of the computational subtask from the terminal device to complete the execution of the interactive task.

[0110] This application also provides an offline computing device for intelligent interactive devices; please refer to [reference needed]. Figure 4 The offline computing device of the intelligent interactive device includes: Monitoring module 10 is used to monitor network connection status; Evaluation module 20 is used to determine the computing power requirements of the computational subtasks in the interactive task to be executed when the network connection status is offline, and to match the computing power requirements with the local resource status. Broadcast module 30 is used to determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result; The transfer module 40 is used to transmit the computational subtask to the terminal device for execution when, based on the matching result, it determines that a task transfer request signal is to be broadcast to the surrounding environment of the location of the intelligent interactive device; in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal. The receiving module 50 is used to receive the calculation result of the terminal device on the calculation subtask, and continue to execute the interactive task to be executed according to the calculation result.

[0111] Optionally, the evaluation module 20 is further configured to: Extract data features from computational subtasks within the interactive task to be executed; The data features are encoded and preprocessed to obtain the feature vector of the computational subtask; The feature vector is input into a pre-trained prediction model, and the pre-trained prediction model predicts the computing power requirement of the computational subtask based on the data features.

[0112] Optionally, the data features include data size and operation type, the feature vector is formed by concatenating the data size vector and the operation type vector, the data size vector is obtained by encoding and preprocessing the data size, the operation type vector is obtained by encoding and preprocessing the operation type, and the pre-trained prediction model sequentially includes an input layer, an embedding layer, a hidden layer and an output layer; The evaluation module 20 is also used for: The feature vector is received through the input layer of the pre-trained prediction model, wherein the number of neurons in the input layer is equal to the dimension of the feature vector; In the embedding layer of the pre-trained prediction model, the operation type vector in the feature vector is mapped to a dense vector of the target dimension to obtain the mapped feature vector, wherein the mapped feature vector is composed of the data size vector and the dense vector; The mapped feature vector is nonlinearly transformed by the hidden layer of the pre-trained prediction model, and the nonlinear transformation result is output through the output layer of the pre-trained prediction model to obtain the computing power requirement of the computational subtask.

[0113] Optionally, the evaluation module 20 is further configured to: Obtain the predicted computing power requirement information of the historical computing subtasks corresponding to the computing subtask, and generate a historical computing power feature vector based on the predicted computing power requirement information; The historical computing power feature vector is fused with the feature vector to form an enhanced feature vector, and the feature vector is updated based on the enhanced feature vector so that the enhanced feature vector can be received by the input layer of the pre-trained prediction model.

[0114] Optionally, the embedded layer obtains a mapped feature vector as a mapped enhanced feature vector, which includes the data size vector, the density vector, and the historical computing power feature vector. The hidden layer of the pre-trained prediction model includes a fully connected network branch, a recurrent neural network branch, and a feature fusion layer. The evaluation module 20 is also used for: The data size vector and the dense vector are nonlinearly transformed by the fully connected network branch to obtain the first output result; The historical computing power feature vector is processed by the recurrent neural network branch to obtain a second output result; The first output result and the second output result are fused by the feature fusion layer to obtain a fused output result, which is then output as the nonlinear transformation result through the output layer of the pre-trained prediction model.

[0115] Optionally, the evaluation module 20 is further configured to: Obtain the task type of the computational subtask in the interactive task to be executed, and find the computational complexity of the computational subtask in the preset task mapping table according to the task type; If the computational complexity found is less than the preset judgment index, the computational subtask is computed locally to obtain the computational result of the computational subtask. If the computational complexity found is greater than or equal to the preset judgment index, or if no computational complexity is found, the step of determining the computational power requirement of the computational subtask in the interactive task to be executed is performed.

[0116] Optionally, the broadcast module 30 is further configured to: If the matching result indicates that the local resource status meets the computing power requirement, the computing subtask is computed locally to obtain the computation result of the computing subtask. If the matching result indicates that the local resource conditions do not meet the computing power requirements, a task transfer request signal is broadcast to the surrounding environment of the location of the intelligent interactive device.

[0117] Optionally, the transfer module 40 is further configured to: In response to task acceptance signals returned by multiple terminal devices in the surrounding environment based on the task transfer request signal, the performance parameters and current load information of the terminal device corresponding to each task acceptance signal are determined; The computing power weight of each terminal device is determined based on its performance parameters and current load information. Based on the computing power weight of each terminal device, the computing subtask is divided into subtask segments corresponding to each terminal device, wherein the computing amount of any subtask segment is matched with the computing power weight of the corresponding terminal device. Each of the subtask segments is transmitted to the corresponding terminal device for execution. After receiving the segment calculation results obtained by each terminal device from the corresponding subtask segment, the calculation results of each segment are merged to obtain the calculation result of the calculation subtask.

[0118] The offline computing device for intelligent interactive devices provided in this application, employing the offline computing method for intelligent interactive devices in the above embodiments, can solve the technical problem of how to fully meet the various interactive needs of users when the intelligent interactive device is offline. Compared with the prior art, the beneficial effects of the offline computing device for intelligent interactive devices provided in this application are the same as the beneficial effects of the offline computing method for intelligent interactive devices provided in the above embodiments, and other technical features in the offline computing device for intelligent interactive devices are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0119] This application provides an interactive device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the offline computing method of the intelligent interactive device in the first embodiment described above.

[0120] The following is for reference. Figure 5 The diagram illustrates a structural schematic of an interactive device suitable for implementing embodiments of this application. The interactive devices in these embodiments may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The interactive device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0121] like Figure 5As shown, the interactive device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory 1002 or a program loaded from storage device 1003 into random access memory 1004. Random access memory 1004 also stores various programs and data required for the operation of the interactive device. The processing unit 1001, read-only memory 1002, and random access memory 1004 are interconnected via bus 1005. Input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, microphones, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the interactive device to communicate wirelessly or wiredly with other devices to exchange data.

[0122] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0123] The interactive device provided in this application employs the offline computing method of the intelligent interactive device in the above embodiments, which can solve the technical problem of how to fully meet the various interactive needs of users when the intelligent interactive device is offline. Compared with the prior art, the beneficial effects of the interactive device provided in this application are the same as those of the offline computing method of the intelligent interactive device provided in the above embodiments, and other technical features in this interactive device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

Claims

1. An offline computing method for an intelligent interactive device, characterized in that, The offline computing method of the intelligent interactive device includes: Monitor network connectivity; When the network connection status is offline, determine the computing power requirement of the computational subtask in the interactive task to be executed, and match the computing power requirement with the local resource status; based on the matching result, determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device. When a task transfer request signal is broadcast to the surrounding environment of the location of the intelligent interactive device based on the matching result, in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal, the computing subtask is transmitted to the terminal device for execution. The terminal device receives the calculation result of the calculation subtask, and continues to execute the interactive task to be executed based on the calculation result; The step of determining the computing power requirements of the computational subtasks in the interactive task to be executed includes: Extract data features from computational subtasks within the interactive task to be executed; The data features are encoded and preprocessed to obtain the feature vector of the computational subtask; The feature vector is input into a pre-trained prediction model, and the pre-trained prediction model predicts the computing power requirement of the computational subtask based on the data features.

2. The offline computing method for the intelligent interactive device as described in claim 1, characterized in that, The data features include data size and operation type. The feature vector is formed by concatenating the data size vector and the operation type vector. The data size vector is obtained by encoding and preprocessing the data size. The operation type vector is obtained by encoding and preprocessing the operation type. The pre-trained prediction model sequentially includes an input layer, an embedding layer, a hidden layer, and an output layer. The step of predicting the computing power requirement of the computational subtask based on the data features using the pre-trained prediction model includes: The feature vector is received through the input layer of the pre-trained prediction model, wherein the number of neurons in the input layer is equal to the dimension of the feature vector; In the embedding layer of the pre-trained prediction model, the operation type vector in the feature vector is mapped to a dense vector of the target dimension to obtain the mapped feature vector, wherein the mapped feature vector is composed of the data size vector and the dense vector; The mapped feature vector is nonlinearly transformed by the hidden layer of the pre-trained prediction model, and the nonlinear transformation result is output through the output layer of the pre-trained prediction model to obtain the computing power requirement of the computational subtask.

3. The offline computing method for intelligent interactive devices as described in claim 2, characterized in that, The step of receiving the feature vector through the input layer of the pre-trained prediction model further includes: Obtain the predicted computing power requirement information of the historical computing subtasks corresponding to the computing subtask, and generate a historical computing power feature vector based on the predicted computing power requirement information; The historical computing power feature vector is fused with the feature vector to form an enhanced feature vector, and the feature vector is updated based on the enhanced feature vector so that the enhanced feature vector can be received by the input layer of the pre-trained prediction model.

4. The offline computing method for the intelligent interactive device as described in claim 3, characterized in that, The embedded layer obtains a mapped feature vector as a mapped enhanced feature vector, which includes the data size vector, the density vector, and the historical computing power feature vector. The hidden layer of the pre-trained prediction model includes a fully connected network branch, a recurrent neural network branch, and a feature fusion layer. After the step of mapping the operation type vector in the feature vector to a dense vector of the target dimension in the embedding layer of the pre-trained prediction model to obtain the mapped feature vector, the method further includes: The data size vector and the dense vector are nonlinearly transformed by the fully connected network branch to obtain the first output result; The historical computing power feature vector is processed by the recurrent neural network branch to obtain a second output result; The first output result and the second output result are fused by the feature fusion layer to obtain a fused output result, which is then output as the nonlinear transformation result through the output layer of the pre-trained prediction model.

5. The offline computing method for the intelligent interactive device as described in claim 1, characterized in that, Before the step of determining the computing power requirements of the subtasks in the interactive task to be executed, the following steps are also included: Obtain the task type of the computational subtask in the interactive task to be executed, and find the computational complexity of the computational subtask in the preset task mapping table according to the task type; If the computational complexity found is less than the preset judgment index, the computational subtask is computed locally to obtain the computational result of the computational subtask. If the computational complexity found is greater than or equal to the preset judgment index, or if no computational complexity is found, the step of determining the computational power requirement of the computational subtask in the interactive task to be executed is performed.

6. The offline computing method for the intelligent interactive device as described in claim 1, characterized in that, The step of determining whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching result includes: If the matching result indicates that the local resource status meets the computing power requirement, the computing subtask is computed locally to obtain the computation result of the computing subtask. If the matching result indicates that the local resource conditions do not meet the computing power requirements, a task transfer request signal is broadcast to the surrounding environment of the location of the intelligent interactive device.

7. The offline computing method for the intelligent interactive device as described in claim 1, characterized in that, The step of transmitting the computational subtask to the terminal device for execution in response to a task acceptance signal returned by the terminal device based on the task transfer request signal in the surrounding environment includes: In response to task acceptance signals returned by multiple terminal devices in the surrounding environment based on the task transfer request signal, the performance parameters and current load information of the terminal device corresponding to each task acceptance signal are determined; The computing power weight of each terminal device is determined based on its performance parameters and current load information. Based on the computing power weight of each terminal device, the computing subtask is divided into subtask segments corresponding to each terminal device, wherein the computing amount of any subtask segment is matched with the computing power weight of the corresponding terminal device. Each of the subtask segments is transmitted to the corresponding terminal device for execution. After receiving the segment calculation results obtained by each terminal device from the corresponding subtask segment, the calculation results of each segment are merged to obtain the calculation result of the calculation subtask.

8. An offline computing device for an intelligent interactive device, characterized in that, The offline computing device of the intelligent interactive device includes: The monitoring module is used to monitor network connectivity. An evaluation module is used to determine the computing power requirements of the computational subtasks in the interactive task to be executed when the network connection status is offline, and to match the computing power requirements with the local resource status. The broadcast module is used to determine whether to broadcast a task transfer request signal to the surrounding environment of the location of the intelligent interactive device based on the matching results. The transfer module is used to transmit the computational subtask to the terminal device for execution when, based on the matching result, it determines that a task transfer request signal is to be broadcast to the surrounding environment of the location of the intelligent interactive device; in response to the task acceptance signal returned by the terminal device in the surrounding environment based on the task transfer request signal. A receiving module is used to receive the calculation result of the terminal device on the calculation subtask, and continue to execute the interactive task to be executed based on the calculation result; The evaluation module is further configured to: extract data features of computational subtasks in the interactive task to be executed; encode and preprocess the data features to obtain feature vectors of the computational subtasks; input the feature vectors into a pre-trained prediction model, and predict the computing power requirements of the computational subtasks based on the data features using the pre-trained prediction model.

9. An interactive device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the offline computing method for the intelligent interactive device as described in any one of claims 1 to 7.

Images