CN115659184B

CN115659184B - Distributed group performance intelligent optimization method and system and electronic equipment

Info

Publication number: CN115659184B
Application number: CN202211602430.1A
Authority: CN
Inventors: 杨远达; 张梦瑶
Original assignee: Individual
Current assignee: Individual
Priority date: 2022-12-14
Filing date: 2022-12-14
Publication date: 2023-07-25
Anticipated expiration: 2042-12-14
Also published as: CN115659184A

Abstract

The application belongs to the technical field of intelligent optimization of groups, and discloses a distributed intelligent optimization method, a system and electronic equipment for group performance, wherein a plurality of individual nodes in a group perform further fine tuning training on an up-to-date group pre-training model to obtain individual optimization models, then all individual nodes score each individual optimization model, finally all scoring values obtained by each individual optimization model are synthesized, a better model is determined from the individual optimization models according to scoring results, so that updating operation of the group pre-training model is performed, and the method and the system are circulated for a plurality of times, so that the absolute performance and application generalization performance of the group pre-training model on the whole group are gradually improved, and finally the group pre-training model with better absolute performance and application generalization performance on the whole group is obtained; the method can automatically improve the generalization performance and the absolute performance of the group pre-training model, thereby being beneficial to improving the performance of the guest group.

Description

Distributed group performance intelligent optimization method and system and electronic equipment

Technical Field

The application relates to the technical field of intelligent optimization of groups, in particular to an intelligent optimization method and system for distributed group performance and electronic equipment.

Background

In the model optimization training process, a pre-training model technology is often used. The pre-training model is a model which is trained and separated from the appointed task and can learn general semantic characteristics of the object perfectly, and has the characteristics of certain absolute performance and generalization performance when the similar but different objects are subjected to non-appointed tasks. (taking the field of computer vision as an example, in the training of a deep learning model, by training a model in advance by ResNet50 (residual network 50) trained on a large amount of image data, the method can be applied to various visual scenes (image classification, object detection, semantic segmentation, instance segmentation and the like) only by combining targeted fine tuning and network design of different tasks, and has better effect).

The pre-training model with excellent performance is used for model training of individual objects (such as mechanical arms, voice recognition equipment, visual equipment and the like, without being limited to the mechanical arms, the voice recognition equipment, the visual equipment and the like), so that the process of obtaining the individual optimization model by the individual objects can be accelerated, and the attribute or performance corresponding to the application of the individual objects and the model can be rapidly improved. Similarly, when the study object is a group object (or referred to as a guest group, where the group object is composed of a plurality of individual objects of the same type, for example, but not limited to, a mechanical arm group, a voice recognition device group, a visual device group, etc.), if the group pre-training model with excellent performance is used in model training of the group object, the attribute or performance corresponding to the group object and the model application can be rapidly improved.

At present, individual training schemes or a mode of using a better performance pre-training model are generally used for improving individual object performance and group object overall performance. The individual training schemes consume a great deal of time and effort as the number of groups increases, and training costs and management costs are continually rising. The acquisition of the better performance pre-training model is also a difficult problem, in order to obtain the pre-training model with higher generalization performance and absolute performance for the group objects consisting of the same type of individual objects, a group large-scale data set mixed training scheme is generally adopted, a training data set is continuously expanded to improve the performance, however, the problems of rising data acquisition cost and continuously rising training cost are also brought, and if new data are added into the data set, the problems of various data duty ratios in the data set, retraining and the like are considered again, so that the intelligent degree is lower.

Disclosure of Invention

The purpose of the application is to provide a distributed group performance intelligent optimization method, a distributed group performance intelligent optimization system and electronic equipment, which can automatically improve the generalization performance and the absolute performance of a group pre-training model, and further are beneficial to improving the performance of a guest group.

In a first aspect, the present application provides a distributed group performance intelligent optimization method, which is characterized in that the method is applied to a central node to optimize a group pre-training model applied to a group; the group comprises a plurality of individual nodes forming a node network, the individual nodes are mutually connected in a communication way, and the central node is connected with all the individual nodes in a communication way and is used for managing the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model;

the intelligent optimization method for the distributed group performance comprises the steps of performing circularly for a plurality of times:

A1. the latest group pre-training model is sent to a node network for downloading by each individual node of the group, and the latest group pre-training model is subjected to fine tuning training to obtain a corresponding individual optimization model; in the first cycle, the latest group pre-training model is an initial group pre-training model;

A2. Obtaining scoring scores of all individual nodes in a population for each individual optimization model;

A3. determining a stage optimal group pre-training model from each individual optimization model according to the scoring values;

A4. and updating the group pre-training model according to the phase optimal group pre-training model.

Further fine tuning training is carried out on the latest group pre-training model by a plurality of individual nodes in the group to obtain individual optimization models, then each individual optimization model is scored by all individual nodes, finally all scoring scores obtained by each individual optimization model are integrated, a better model is determined from the individual optimization models according to scoring results, so that updating operation of the group pre-training model is carried out, the circulation is carried out for a plurality of times, the absolute performance and the application generalization performance of the group pre-training model for the whole group are gradually improved, and finally the group pre-training model with better absolute performance and application generalization performance for the whole group compared with the initial group pre-training model is obtained. The method can realize automatic optimization of the group pre-training model, so that the absolute performance and the application generalization performance of the group pre-training model for the whole group are improved, and the group pre-training model has certain self-adaptive capacity for group expansion; in addition, through the joint optimization of group multinode, optimizing speed is fast, through distributed training mode, and the training pressure is shared jointly to a plurality of individuals, and the used dataset of every individual training is less, trains with low costs.

Preferably, step A1 comprises:

and sending the latest group pre-training model to a node network for downloading by each individual node of the group, and performing fine tuning training on the latest group pre-training model to obtain a plurality of corresponding individual optimization models.

For a single individual node, the working model with high matching degree is not necessarily high for other individual nodes, each individual node is trained to obtain a plurality of individual optimization models, and the matching degree of the plurality of individual optimization models of the same individual node is different from that of each individual node, so that more candidate models can be obtained, and the screening to obtain a stage optimal group pre-training model with better absolute performance and application generalization performance is facilitated.

Preferably, step A3 comprises:

calculating a comprehensive score of each individual optimization model according to the scoring value of each individual optimization model;

and determining the individual optimization model with the highest comprehensive score as the phase optimal group pre-training model.

Preferably, step A4 comprises:

if the latest group pre-training model is an initial group pre-training model, replacing the latest group pre-training model by the phase optimal group pre-training model;

If the latest group pre-training model is not the initial group pre-training model, comparing the comprehensive score corresponding to the stage optimal group pre-training model with the comprehensive score corresponding to the latest group pre-training model;

when the comprehensive score corresponding to the phase optimal group pre-training model is higher than the comprehensive score corresponding to the latest group pre-training model, replacing the latest group pre-training model by the phase optimal group pre-training model; otherwise, discarding the phase optimal group pre-training model.

In some embodiments, the individual node is a robotic arm, a voice recognition device, or a vision device; the group pre-training model is correspondingly a friction force compensation model, a voice instruction recognition model or an image feature extraction model.

In a second aspect, the present application provides a distributed population performance intelligent optimization method applied to individual nodes to optimize a population pre-training model applied to a population; the group comprises a plurality of individual nodes forming a node network, all the individual nodes are in communication connection with each other, all the individual nodes are in communication connection with a central node, and the central node is used for managing the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model;

The intelligent optimization method for the distributed group performance comprises the following steps:

B1. when a central node is perceived to send a latest group pre-training model to a node network, downloading the latest group pre-training model, and performing fine tuning training on the latest group pre-training model to obtain an individual optimization model;

B2. scoring the individual optimization model and transmitting the individual optimization model to other individual nodes so that the other individual nodes score the individual optimization model;

B3. when receiving the individual optimization models sent by other individual nodes, scoring the received individual optimization models;

B4. and sending the scoring values of the individual optimization models to the central node so that the central node can determine a stage optimal group pre-training model from the individual optimization models according to the scoring values to update the group pre-training model.

Preferably, step B1 comprises:

performing fine tuning training on the latest group pre-training model to obtain a fine tuning model;

comparing the fine-tuning model with a currently used working model to judge whether the fine-tuning model is better than the currently used working model;

if yes, the fine tuning model is determined to be an individual optimization model.

In a third aspect, the present application provides a distributed group performance intelligent optimization system, configured to optimize a group pre-training model applied to a group, including a central node and the group, where the group includes a plurality of individual nodes that form a node network, each individual node is connected to each other in a communication manner, and each individual node is connected to the central node in a communication manner, and the central node is configured to manage the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model;

The central node is used for storing the latest group pre-training model and circularly executing for a plurality of times:

the latest group pre-training model is sent to the node network for downloading of each individual node of the group, and fine tuning training is carried out on the latest group pre-training model to obtain a corresponding individual optimization model; in the first cycle, the latest group pre-training model is an initial group pre-training model;

obtaining scoring scores of all the individual nodes for each individual optimization model;

determining a stage optimal group pre-training model from each individual optimization model according to the scoring values;

updating the group pre-training model according to the phase optimal group pre-training model;

the individual nodes are used for downloading the latest group pre-training model when the central node is perceived to send the latest group pre-training model to the node network, performing fine tuning training on the latest group pre-training model to obtain an individual optimization model, scoring the individual optimization model, sending the individual optimization model to other individual nodes so that the other individual nodes score the individual optimization model, scoring the received individual optimization model when the individual optimization model sent by the other individual nodes is received, and sending scoring values of the individual optimization models to the central node so that the central node determines a stage optimal group pre-training model from the individual optimization models according to the scoring values, thereby performing updating operation of the group pre-training model.

Further fine tuning training is carried out on the latest group pre-training model by a plurality of individual nodes in the group to obtain individual optimization models, then each individual optimization model is scored by all individual nodes, finally all scoring scores obtained by each individual optimization model are integrated, a better model is determined from the individual optimization models according to scoring results, so that updating operation of the group pre-training model is carried out, the circulation is carried out for a plurality of times, the absolute performance and the application generalization performance of the group pre-training model for the whole group are gradually improved, and finally the group pre-training model with better absolute performance and application generalization performance for the whole group compared with the initial group pre-training model is obtained. The automatic optimization of the group pre-training model can be realized, so that the absolute performance and the application generalization performance of the group pre-training model for the whole group are improved, and the group pre-training model has certain self-adaptive capacity for group expansion; in addition, through the joint optimization of group multinode, optimizing speed is fast, through distributed training mode, and the training pressure is shared jointly to a plurality of individuals, and the used dataset of every individual training is less, trains with low costs.

In a fourth aspect, the present application provides an electronic device, comprising a processor and a memory, the memory storing a computer program executable by the processor, when executing the computer program, running steps in a distributed population performance intelligent optimization method as described above.

The beneficial effects are that:

according to the distributed group performance intelligent optimization method, system and electronic equipment, the group pre-training model technology is used as a medium, the group pre-training model is optimized based on the group intelligent optimization thought, and the optimized pre-training model is applied to the group, so that the aim of automatically improving the group performance is fulfilled; specifically, the latest group pre-training model is subjected to further fine tuning training by a plurality of individual nodes in the group to obtain individual optimization models, then all individual nodes score each individual optimization model, finally all scoring scores obtained by each individual optimization model are integrated, a better model is determined from the individual optimization models according to scoring results, so that the updating operation of the group pre-training model is carried out, the circulation is carried out for a plurality of times, the absolute performance and the application generalization performance of the group pre-training model for the whole group are gradually improved, and finally the group pre-training model with better absolute performance and application generalization performance for the whole group compared with the initial group pre-training model is obtained. The method can realize automatic optimization of the group pre-training model, so that the absolute performance and the application generalization performance of the group pre-training model for the whole group are improved, and the group pre-training model has certain self-adaptive capacity for group expansion; in addition, through the joint optimization of group multinode, optimizing speed is fast, through distributed training mode, and the training pressure is shared jointly to a plurality of individuals, and the used dataset of every individual training is less, trains with low costs.

Drawings

Fig. 1 is a flowchart of a distributed group performance intelligent optimization method provided in an embodiment of the present application.

Fig. 2 is a flowchart of another intelligent optimization method for distributed group performance according to an embodiment of the present application.

Fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Fig. 4 is a schematic structural diagram of a distributed group performance intelligent optimization system according to an embodiment of the present application.

FIG. 5 is a schematic diagram of a distributed community performance intelligent optimization system of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. The components of the embodiments of the present application, which are generally described and illustrated in the figures herein, may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, as provided in the accompanying drawings, is not intended to limit the scope of the application, as claimed, but is merely representative of selected embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present application without making any inventive effort, are intended to be within the scope of the present application.

It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures. Meanwhile, in the description of the present application, the terms "first", "second", and the like are used only to distinguish the description, and are not to be construed as indicating or implying relative importance.

The method and the device are applied to the group working scene, the group pre-training model technology is used as a medium, the group pre-training model is optimized based on the group intelligent optimization idea, and the optimized pre-training model is applied to the group, so that the aim of automatically improving the group performance is achieved.

Referring to fig. 1, fig. 1 is a schematic diagram of a distributed group performance intelligent optimization method applied to a central node to optimize a group pre-training model applied to a group according to some embodiments of the present application; the group comprises a plurality of individual nodes forming a node network, the individual nodes are mutually connected in a communication way, and the central node is connected with all the individual nodes in a communication way and is used for managing the individual nodes; the group pre-training model has model parameters which can be optimized iteratively; each body node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each body node can perform fine tuning training on the group pre-training model and score the group pre-training model;

A2. obtaining scoring values of all individual nodes in the group on each individual optimization model;

A3. determining a stage optimal group pre-training model from the individual optimization models according to the scoring values;

A4. and updating the group pre-training model according to the stage optimal group pre-training model.

Wherein the central node may be, but is not limited to, a computer, a server, etc. The individual nodes are computing nodes with certain hardware computing resources and independent working capabilities, can be actual physical devices such as mechanical arms, voice recognition devices, visual devices and the like, and can also be virtualized isolated computing environments such as virtual machines, containers and the like, and the specific form is not limited to the above.

For example, in the case where the individual nodes are mechanical arms, each mechanical arm may be a mechanical arm of the same model, but each mechanical arm may be respectively responsible for different tasks (for example, responsible for different transportation of different workpieces, different processing procedures of the same product, etc.), and although the models are the same, manufacturing and assembly errors, measurement errors, and gear wear of each mechanical arm are all different, so there are differences in structural parameters, that is, structural parameters and responsible tasks are different.

For another example, in the case where the individual nodes are voice recognition devices, each voice recognition device may be a voice recognition device of the same model, but the users of each voice recognition device are different, the accents of the users are different (the pronunciation of the same instruction is different) and the background noise of the environment where the users are located is different, that is, the task objects are different.

Or for example, in the case that the individual nodes are vision apparatuses, each vision apparatus may be the same model of vision apparatus, but each vision apparatus may be respectively responsible for different tasks (such as classification, positioning, detection, image segmentation, etc. of the workpiece), that is, the responsible tasks are different.

Wherein the loop process is automatically ended when at least one of the following conditions is met: 1. the circulation times reach the preset maximum times; 2. the comprehensive score of the stage optimal group pre-training model converges (for example, the deviation of the comprehensive score of the stage optimal group pre-training model obtained by n times of continuous circulation and the absolute value of the comprehensive score of the stage optimal group pre-training model obtained by the last circulation is smaller than a preset minimum optimization threshold value, and n is a preset positive integer and can be set according to actual needs); 3. the total time of the optimization process is greater than a preset time interval; 4. the comprehensive scores of the phase optimal group pre-training models obtained by continuous m times of circulation are lower than the comprehensive scores of the latest group pre-training models, and m is a preset positive integer and can be set according to actual needs.

In some embodiments, the above-described cycling process is automatically initiated whenever at least one of the following conditions is met: 1. adding new individual nodes into the population; 2. the time interval from the last update of the group pre-training model in the central node reaches a preset duration threshold; 3. the working model with individual nodes exceeding a preset number threshold or a preset proportion threshold in the group is not applicable (the working model is defined below, the working model is not applicable, the effect index brought by the working model is smaller than the preset index threshold, the positive effect index can be set according to actual needs, for example, for a mechanical arm friction force compensation model, the effect index can be the improvement of position control precision, the compensation effect of the friction force compensation model which can bring about good compensation effect can be reduced or even negative effect due to continuous abrasion of mechanical arm joints, for a voice command recognition model, the effect index can be the voice command recognition accuracy, the voice command recognition accuracy of the original voice command recognition model is reduced due to the fact that a user or a use environment of the voice recognition device can be changed, and for an image feature extraction model, the effect index can be the accuracy of performing operations such as workpiece classification, positioning, detection, image segmentation and the like according to image features performed by the image feature extraction model, and the accuracy of the voice command recognition accuracy is reduced when an original image feature extraction model is used for executing a new task due to the fact that a specific task of a vision device can be changed.

The central node may be the central node 1 in the distributed group performance intelligent optimization system shown in fig. 4, and the individual node may be the individual node 2 in the distributed group performance intelligent optimization system shown in fig. 4.

The model herein refers to a mathematical model generated based on certain algorithms or algorithm combinations and having characteristics capable of optimizing parameters and performing iterative optimization on the optimized parameters, for example, the model can be divided from principle and structure and comprises a deep learning model, a reinforcement learning model, a support vector machine model, a classification tree model, a regression tree model and the like; functionally, friction compensation models, voice command recognition models, image feature extraction models, and the like can be included, but are not limited to. Training refers to an iterative optimization process of a model; the pre-training refers to an optimization method for extracting common characteristics of data as much as possible by using related data to participate in training, so as to enhance generalization capability and absolute performance of a model; the group pre-training model is a mathematical model which is generated by pre-training a group object based on certain algorithms or algorithm combinations and has the characteristics of optimizing parameters and iterative optimization of the optimizing parameters; the initial group pre-training model is a model which is only initialized by model parameters before the intelligent optimization method for the distributed group performance is used for optimization, and the initialization parameters can be derived from empirical values, random values and the like.

In some embodiments, models obtained after fine-tuning training of the latest group pre-training model by the individual nodes can be all determined to be individual optimization models.

In other preferred embodiments, the individual optimization model is a model that is better than the working model currently used by the individual node after the individual node performs fine-tuning training on the latest population pre-training model. The currently used working model refers to a model which is the same as the latest group pre-training model and is currently used by the individual node in actual working (for example, if the latest group pre-training model is a friction force compensation model, a voice instruction recognition model, an image feature extraction model and the like, the working model is correspondingly the friction force compensation model, the voice instruction recognition model, the image feature extraction model and the like which are currently used by the individual node in actual working); for a single individual node, only the model after fine tuning training is a model which is better than the currently used working model, the model is judged to be an individual optimization model, and only the individual optimization model triggers a scoring mechanism (which scores itself and is sent to other individual nodes for scoring), so that the applicability of the finally determined stage optimal group pre-training model is more beneficial to developing to a better direction.

Wherein each individual node can obtain a corresponding one or more individual optimization models through fine tuning training.

Thus, in some embodiments, step A1 comprises:

and sending the latest group pre-training model to the node network for downloading by each individual node of the group, and performing fine tuning training on the latest group pre-training model to obtain a corresponding individual optimization model.

In other embodiments, step A1 comprises:

and sending the latest group pre-training model to the node network for downloading by each individual node of the group, and performing fine tuning training on the latest group pre-training model to obtain a plurality of corresponding individual optimization models.

For a single individual node, the working model with high matching degree is not necessarily high for other individual nodes, each individual node is trained to obtain a plurality of individual optimization models, and the matching degree of the plurality of individual optimization models of the same individual node is different from that of each individual node, so that more candidate models can be obtained, and the screening to obtain a stage optimal group pre-training model with better absolute performance and application generalization performance is facilitated. The specific number of the individual optimization models obtained by each individual node is generally a preset number N, and the specific value of the preset number N may be set according to actual needs, for example, n=3, but is not limited thereto.

In practice, when the models obtained after the individual nodes perform fine tuning training on the latest group pre-training model are all determined to be the individual optimization models, each individual node can perform fine tuning training for corresponding times according to the number of the required individual optimization models, so that the corresponding number of the individual optimization models are obtained; if the fine tuning training is carried out for a plurality of times, carrying out the fine tuning training on the model after the last fine tuning training from the second fine tuning training; or if the fine tuning training is performed for multiple times, performing the fine tuning training on the latest group pre-training model obtained by downloading by using different training sets each time.

When the individual optimization model is a model which is better than the working model currently used by the individual node, the individual optimization model can be obtained by each fine tuning training, or the individual optimization model can be obtained by multiple fine tuning training, so that each individual node can circularly perform fine tuning training on the latest group pre-training model (if multiple fine tuning training is performed, the model after the last fine tuning training is performed from the second fine tuning training, or if multiple fine tuning training is performed, the latest group pre-training model obtained by downloading is subjected to fine tuning training by using different training sets each time until the required number of individual fine tuning optimization models are obtained, or until the number of fine tuning training reaches a preset number threshold (the fine tuning is stopped), and the number threshold can be set according to actual needs; thus, the optimization inefficiency caused by excessive times of fine tuning training is avoided.

Wherein the scoring score of each individual optimization model includes the scoring score of the individual node generating the individual optimization model (hereinafter, the individual node is referred to as a first individual node) to the individual optimization model, and the scoring score of the individual optimization model by the other individual nodes after the first individual node transmits the individual optimization model to the other individual nodes. The first individual node can send the generated individual optimization model to other individual nodes through the central node, and can also directly send the generated individual optimization model to other individual nodes through the node network.

Specifically, the individual nodes may score the individual optimization model through a preset scoring rule. The preset scoring rule can be set according to actual needs.

Further, step A3 includes:

calculating the comprehensive score of each body optimization model according to the score value of each body optimization model;

and determining the individual optimization model with the highest comprehensive score as a stage optimal group pre-training model.

For example, the average or weighted average of the individual score scores for the same individual optimization model may be calculated as the composite score for that individual optimization model; the average value or weighted average value of the score values of the same individual optimization model can be calculated as the first score of the individual optimization model, the variance of the score values of the individual optimization model is calculated, and finally the weighted sum of the first score and the variance is calculated as the comprehensive score of the individual optimization model. The method of calculating the composite score is not limited thereto.

Preferably, step A4 comprises:

if the latest group pre-training model is the initial group pre-training model, replacing the latest group pre-training model with the phase optimal group pre-training model;

if the latest group pre-training model is not the initial group pre-training model, comparing the comprehensive score corresponding to the optimal group pre-training model in the stage with the comprehensive score corresponding to the latest group pre-training model;

when the comprehensive score corresponding to the phase optimal group pre-training model is higher than the comprehensive score corresponding to the latest group pre-training model, replacing the latest group pre-training model by the phase optimal group pre-training model; otherwise, discarding the stage optimal population pre-training model.

The initial group pre-training model is an optimized group pre-training model which is not subjected to the distributed group performance intelligent optimization method, and the group pre-training model at the moment is not scored, so that the comprehensive score is not generated, and therefore after the stage optimal group pre-training model is obtained by using the distributed group performance intelligent optimization method for the first time, the initial group pre-training model is directly replaced by the stage optimal group pre-training model to serve as the latest group pre-training model, and the corresponding comprehensive score is recorded. And comparing the comprehensive score of the phase optimal group pre-training model obtained by optimizing each time with the latest group pre-training model currently stored at the central node, wherein the phase optimal group pre-training model is used for replacing the latest group pre-training model at the central node only when the comprehensive score of the phase optimal group pre-training model is higher so as to ensure that the absolute performance and the application generalization performance of the latest group pre-training model stored at the central node are better.

Preferably, after obtaining the new group pre-training model, the new group pre-training model may be sent to each body node, and the working model is updated after the fine-tuning training is performed by each body node (the new model after the fine-tuning training of each body node may be compared with the currently used working model to compensate the effect, if the new model is better, the currently used working model is replaced by the new model, otherwise, the new model is discarded). Thus, after ending the cycle, the steps may also be included: A5. and sending the current latest group pre-training model to each body node so as to enable each body node to update the working model.

The above-mentioned processes (multiple steps A1 to A4 and A5) can be executed once every fixed period (which can be set according to actual needs), so that the performance of the working model of each body node is gradually improved, and the performance of the individual optimization model obtained by each body node when performing fine tuning training is also better and better, so that the absolute performance of the working model in each body node is better and better, and the absolute performance and application generalization performance of the group pre-training model stored at the central node are better and better; when new individual nodes are put into use (i.e. new individual nodes are added into the population), the population pre-training model can be further adapted to the expanded population through the process, so that the population pre-training model has better self-adaptation capability to the continuously expanded population, and the population has self-optimization capability.

Referring to fig. 2, the present application provides a distributed group performance intelligent optimization method applied to individual nodes to optimize a group pre-training model applied to a group; the group comprises a plurality of individual nodes forming a node network, all the individual nodes are connected with each other in a communication way, all the individual nodes are connected with a central node in a communication way, and the central node is used for managing the individual nodes; the group pre-training model has model parameters which can be optimized iteratively; each body node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each body node can perform fine tuning training on the group pre-training model and score the group pre-training model;

B1. when the central node is perceived to send the latest group pre-training model to the node network, downloading the latest group pre-training model, and performing fine tuning training on the latest group pre-training model to obtain an individual optimization model;

B2. Scoring the individual optimization model, and transmitting the individual optimization model to other individual nodes so that the other individual nodes score the individual optimization model;

B4. and sending the score values of the individual optimization models to the central node so that the central node can determine the phase optimal group pre-training model from the individual optimization models according to the score values to update the group pre-training model.

Wherein the central node may be, but is not limited to, a computer, a server, etc. The individual nodes are computing nodes with certain hardware computing resources and independent working capabilities, can be actual physical devices such as mechanical arms, voice recognition devices, visual devices and the like, and can also be virtualized isolated computing environments such as virtual machines, containers and the like, and the specific form is not limited to the above. The central node may be central node 1 in the distributed group performance intelligent optimization system shown in fig. 4, and the individual node may be individual node 2 in the distributed group performance intelligent optimization system shown in fig. 4.

In some embodiments, models obtained after fine-tuning training of the latest group pre-training model by the individual nodes can be all determined to be individual optimization models. Thus, step B1 comprises:

the fine tuning model is determined as an individual optimization model.

In other preferred embodiments, the individual optimization model is a model that is better than the working model currently used by the individual node after the individual node performs fine-tuning training on the latest population pre-training model. Thus, step B1 comprises:

comparing the fine tuning model with the currently used working model to judge whether the fine tuning model is better than the currently used working model;

For a single individual node, only the model after fine tuning training is a model which is better than the currently used working model, the model is judged to be an individual optimization model, and only the individual optimization model triggers a scoring mechanism (which scores itself and is sent to other individual nodes for scoring), so that the absolute performance and the application generalization performance of the finally determined stage optimal group pre-training model are more beneficial to developing to a better direction.

The specific process of fine-tuning the latest population pre-training model depends on the specific role of the population pre-training model and the specific type of individual nodes.

The particular method of determining whether the fine-tuning model is better than the currently used working model depends on the particular role of the population pre-training model and the particular type of individual nodes.

Wherein each individual node can obtain a corresponding one or more individual optimization models through fine tuning training. Thus, in some embodiments, step B1 comprises: and performing fine tuning training on the latest group pre-training model to obtain an individual optimization model. In other embodiments, step B1 comprises: and performing fine tuning training on the latest group pre-training model to obtain a plurality of individual optimization models. For a single individual node, the working model with high matching degree (the better the compensation effect is, the higher the matching degree is) is, and for other individual nodes, the matching degree is not necessarily high, each individual node is trained to obtain a plurality of individual optimization models, and the matching degree of the plurality of individual optimization models of the same individual node is different from that of each individual node, so that more candidate models can be obtained, and the screening to obtain a stage optimal group pre-training model with better applicability is facilitated. The specific number of the individual optimization models obtained by each individual node is generally a preset number N, and the specific value of the preset number N may be set according to actual needs, for example, n=3, but is not limited thereto.

Wherein scoring the individual optimization model comprises: and scoring the individual optimization model through a preset scoring rule.

The preset scoring rule can be set according to actual needs.

In some embodiments, step B2 comprises: the individual optimization model is transmitted to other individual nodes through the central node, so that the other individual nodes score the individual optimization model. In other embodiments, step B2 comprises: and directly transmitting the individual optimization model to other individual nodes through the node network so that the other individual nodes score the individual optimization model.

Preferably, after obtaining a new group pre-training model, the central node can send the new group pre-training model to each body node, and the working model is updated after fine tuning training is carried out by each body node; thus, in some embodiments, after step B4, further comprising: B5. after receiving a current latest group pre-training model sent by a central node, updating a working model according to the current latest group pre-training model; specifically, fine tuning training is performed on the current latest group pre-training model to obtain a new model (refer to the previous in the specific training process), the new model is compared with the currently used working model to judge whether the new model is better than the currently used working model (refer to the previous in the specific judging process), if so, the currently used working model is replaced by the new model, otherwise, the new model is abandoned.

Referring to fig. 4, the present application provides a distributed group performance intelligent optimization system, which is used for optimizing a group pre-training model applied to a group, and comprises a central node 1 and a group, wherein the group comprises a plurality of individual nodes 2 forming a node network, the individual nodes 2 are mutually connected in a communication manner, the individual nodes 2 are all connected with the central node 1 in a communication manner, and the central node 1 is used for managing the individual nodes 2; the group pre-training model has model parameters which can be optimized iteratively; each body node 2 is a device of the same type having a difference in at least one of a structural parameter, a task to be charged and a task object, and each body node 2 is capable of performing fine-tuning training on a group pre-training model and scoring the group pre-training model;

the central node 1 is used for storing the latest group pre-training model, and circularly executes for a plurality of times:

the latest group pre-training model is sent to a node network for downloading by each individual node 2 of the group, and the latest group pre-training model is subjected to fine tuning training to obtain a corresponding individual optimization model; in the first cycle, the latest population pre-training model is the initial population pre-training model (see above for specific procedures);

Obtaining scoring values of all individual nodes 2 for each individual optimization model (specific process refers to the previous);

determining a stage optimal group pre-training model from the individual optimization models according to the scoring values (refer to the previous for a specific process);

updating the group pre-training model according to the optimal model (refer to the previous for the specific process);

the individual node 2 is configured to download the latest group pre-training model when it is perceived that the central node 1 transmits the latest group pre-training model to the node network, perform fine tuning training on the latest group pre-training model to obtain an individual optimization model, score the individual optimization model, transmit the individual optimization model to other individual nodes 2, so that the other individual nodes 2 score the individual optimization model, score the received individual optimization model when the individual optimization model transmitted by the other individual nodes 2 is received, and transmit the scoring score of each individual optimization model to the central node 1, so that the central node 1 determines the stage optimal group pre-training model from the individual optimization models according to the scoring score, so as to perform the updating operation of the group pre-training model (refer to the foregoing for specific process).

According to the distributed group performance intelligent optimization system, a plurality of individual nodes 2 in a group conduct further fine tuning training on a latest group pre-training model to obtain individual optimization models, then all individual nodes 2 score each optimization model, finally all score values obtained by each individual optimization model are integrated, a better model is determined from the individual optimization models according to score results, updating operation of the group pre-training model is conducted, the circulation is conducted for many times, absolute performance and application generalization performance of the group pre-training model on the whole group are gradually improved, and finally the group pre-training model with better absolute performance and application generalization performance compared with the initial group pre-training model on the whole group is obtained. The automatic optimization of the group pre-training model can be realized, so that the absolute performance and the application generalization performance of the group pre-training model for the whole group are improved, and the group pre-training model has certain self-adaptive capacity for group expansion; in addition, through the joint optimization of group multinode, optimizing speed is fast, through distributed training mode, and the training pressure is shared jointly to a plurality of individuals, and the used dataset of every individual training is less, trains with low costs.

In some embodiments, the above-described cycling process is automatically initiated whenever at least one of the following conditions is met: 1. adding new individual nodes into the population; 2. the time interval from the last update of the group pre-training model in the central node reaches a preset duration threshold; 3. the working model with individual nodes exceeding the preset number threshold or the preset proportion threshold in the group is not applicable (the definition of the working model is shown below, the inapplicability of the working model means that the effect index brought by the working model is smaller than the preset index threshold, and the positive effect index can be set according to actual needs, for example, for a mechanical arm friction force compensation model, the effect index is the improvement amount of position control precision, and as the mechanical arm joints are worn continuously, the compensation effect of the friction force compensation model which can bring about better compensation effect can be reduced, even negative effects occur).

Although fig. 4 shows a case with three individual nodes 2, the number of individual nodes 2 is not limited to three, and may be fewer or greater.

Specifically, the principle of the distributed group performance intelligent optimization system can refer to fig. 5, the central node performs normalization detection, and periodically detects whether an automatic group optimization starting condition (namely, the condition of the automatic start cycle process) is met according to a preset detection period (which can be set according to actual needs);

if the automatic group optimization starting condition is met, starting an automatic optimization circulation process: the center node distributes the latest group pre-training model to a node network for downloading by each body node, each body node carries out fine-tuning training on the downloaded group pre-training model to obtain an individual optimization model, each body node scores the individual optimization model of the center node, the individual optimization model is transmitted to other individual nodes so that the other individual nodes score the individual optimization model, and then the center node gathers the scoring scores of the individual optimization models; if the current rest state preservation process is in a rest state preservation process (namely, when the group system meets the automatic group optimization ending condition, a process for preserving the current rest state data of the group, wherein the rest state data comprises, but is not limited to, time data when the automatic optimization process is stopped for the last time, the number of individual nodes and scoring scores of each individual node to a working model of the individual nodes), the central node obtains and preserves the rest state data and then exits the rest state preservation process (wherein, when the rest state data is preserved for enabling the central node to perform normal detection, whether the automatic group optimization starting condition is met or not can be judged according to the preserved rest state data, for example, whether a time interval between the current moment and the rest moment of the last automatic optimization process reaches a preset duration threshold value is judged, whether the number of the individual nodes is increased is judged, and whether the working model of the individual nodes exceeding the preset number threshold value or the preset proportion threshold value is not applicable is judged; if the current state is not in the rest state preservation process, determining a stage optimal group pre-training model from the individual optimization models according to the summarized score values; if the score (herein referred to as the composite score) of the stage-optimal group pre-training model (the stage-optimal model in fig. 5 is simply referred to as the stage-optimal group pre-training model) is greater than the score (herein referred to as the composite score) of the global-optimal group pre-training model (i.e., the latest group pre-training model currently stored in the central node, the global-optimal model in fig. 5 is simply referred to as the global-optimal group pre-training model), the latest group pre-training model stored in the central node is updated (i.e., replaced) by the global-optimal group pre-training model, and whether the automatic group optimization ending condition (i.e., the condition of the automatic ending cycle process described above) is currently satisfied is judged; if the score of the stage optimal group pre-training model is not greater than the score of the global optimal group pre-training model, discarding the stage optimal group pre-training model, and judging whether the automatic group optimization ending condition is met currently; if the automatic group optimization ending condition is met, the group optimization process is stopped, the central node activates the stop state storage process, and if the automatic group optimization ending condition is not met, the latest group pre-training model is distributed to the node network again for downloading by each body node to continue the automatic optimization cycle.

In the process of carrying out the normalization detection on the central node, if the automatic group optimization starting condition is not met, judging whether the current state is in the rest state preservation process, if the current state is not in the rest state preservation process, continuing to carry out the normalization detection, and if the current state is in the rest state preservation process, further distributing the latest group pre-training model so as to acquire the rest state data.

The central node can be functionally and formally divided into two nodes to work together, and one node is used for a top-layer storage model, model issuing, storage rest state data and a normalized detection function; the other node is used for scoring summarization, acquiring and storing rest state data, updating a global optimal model and judging an automatic group optimization ending condition; the two nodes may be two independent devices or may be two functional modules of the same device.

Referring to fig. 3, fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present application, where the electronic device includes: processor 301 and memory 302, the processor 301 and memory 302 being interconnected and in communication with each other by a communication bus 303 and/or other form of connection mechanism (not shown), the memory 302 storing a computer program executable by the processor 301, the computer program being executable by the processor 301 when the electronic device is running to perform the distributed population performance intelligent optimization method in any of the alternative implementations of the above embodiments to implement the following functions:

The loop is executed for a plurality of times: the latest group pre-training model is sent to a node network for downloading by each individual node of the group, and the latest group pre-training model is subjected to fine tuning training to obtain a corresponding individual optimization model; in the first cycle, the latest group pre-training model is an initial group pre-training model; obtaining scoring values of all individual nodes in the group on each individual optimization model; determining a stage optimal group pre-training model from the individual optimization models according to the scoring values; updating the group pre-training model according to the phase optimal group pre-training model;

or when the central node is perceived to send the latest group pre-training model to the node network, downloading the latest group pre-training model, and performing fine tuning training on the latest group pre-training model to obtain an individual optimization model; scoring the individual optimization model, and transmitting the individual optimization model to other individual nodes so that the other individual nodes score the individual optimization model; when receiving the individual optimization models sent by other individual nodes, scoring the received individual optimization models; and sending the score values of the individual optimization models to the central node so that the central node can determine the phase optimal group pre-training model from the individual optimization models according to the score values to update the group pre-training model.

Example 1

In the first embodiment, the individual node is a mechanical arm, the group pre-training model is a friction compensation model (the friction compensation model may be, but is not limited to, a deep learning model, a reinforcement learning model, etc.), and the central node is a management central server. The group consisting of a plurality of mechanical arms may be a mechanical arm group actually used in a use place (such as a factory), or may be a mechanical arm group specially constructed in a laboratory to obtain an optimized group pre-training model.

The servo system actuating mechanism of the mechanical arm has friction force, the existence form of the friction force comprises a joint motor, friction between mechanical arm surfaces and the like, the friction force can reduce the control precision of the mechanical arm and the motion smoothness of the mechanical arm, and in order to reduce the negative influence caused by the friction force, a friction force compensation model can be added to control the motion of the mechanical arm so as to improve the control precision; in the mechanical arm individuals, even though the mechanical arms of the same model are not completely identical in assembly error, observation error and joint gear abrasion, and the actual working environment and working tasks are not completely identical, so that the motion conditions of all joints of the mechanical arms are different, and the friction force compensation model obtained through unified training does not necessarily completely meet the actual use requirements of the mechanical arms.

This problem can be solved by the distributed population performance intelligent optimization method described above. Wherein, for the mechanical arm group, the mechanical arms of each body node have the same joint number, and the construction design and the model are similar or the same, so that one pre-training model can be shared.

In the first embodiment, the initial group friction compensation model (i.e., the group friction compensation model not optimized by the distributed group performance intelligent optimization method) is stored at the management center server, and then the following processes are performed in a circulating manner for a plurality of times:

1. and the management center server sends the locally stored friction force compensation model to a mechanical arm group communication network, and each mechanical arm carries out fine tuning training on the group friction force compensation model after downloading the friction force compensation model to obtain a corresponding individual optimization model.

2. And each mechanical arm scores the individual optimization model obtained by training the mechanical arm and then sends the individual optimization model to other mechanical arms through the mechanical arm group communication network so that the other mechanical arms score the individual optimization model.

3. After the management center server collects the scoring values of all the mechanical arms to each individual optimization model, the group friction force compensation model with the optimal stage is determined from each individual optimization model according to the scoring values.

4. And updating the group friction force compensation model currently stored in the management center server according to the group friction force compensation model score at the optimal stage.

Wherein the loop process is automatically ended when at least one of the following conditions is met: 1. the circulation times reach the preset maximum times; 2. the comprehensive score of the phase optimal group friction force compensation model converges (for example, the deviation of the absolute value of the comprehensive score of the phase optimal group friction force compensation model obtained by n times of continuous circulation and the comprehensive score of the phase optimal group friction force compensation model obtained by the last circulation is smaller than a preset minimum optimization threshold value, and n is a preset positive integer which can be set according to actual needs); 3. the total time of the optimization process is greater than a preset time interval; 4. the comprehensive scores of the phase optimal group friction force compensation models obtained by continuous m times of circulation are lower than the comprehensive scores of the latest group friction force compensation models, and m is a preset positive integer and can be set according to actual needs.

After the cycle is completed, the above-described cycle process is automatically started each time at least one of the following conditions is satisfied: 1. adding a new mechanical arm into the population; 2. the time interval of the last updating of the group friction force compensation model in the management center server reaches a preset duration threshold; 3. the working model of the mechanical arm with the number of mechanical arms exceeding the preset number threshold or the preset proportion threshold in the group is not applicable (for example, the compensation effect of the friction force compensation model which can bring about better compensation effect may be reduced or even negative effect occurs due to continuous abrasion of joints of the mechanical arm).

The group friction force compensation model stored in the management center server is an optimized group friction force compensation model, has better generalization performance and absolute performance, and can be sent to each mechanical arm in the mechanical arm group for updating the working model, so that the friction force compensation performance of the mechanical arm group is improved.

Specific processes can refer to the intelligent optimization method for the distributed group performance, and the main difference is that:

the specific process of the mechanical arm for carrying out fine tuning training on the latest group pre-training model comprises the following steps: under the condition that a currently used working model (which is a friction compensation model) and a latest group pre-training model do not participate in a control process, a plurality of designated working tasks (namely tasks which are responsible) are executed, motion instruction data and motion feedback data (specific motion instruction data items and motion feedback data items are set according to the input requirements of the group pre-training model, for example, data which are required to be input when the group pre-training model is used comprises at least one of moment instruction data, speed instruction data and angle instruction data of each joint motor of a mechanical arm, the motion instruction data correspondingly comprises at least one of moment instruction data, speed instruction data and angle instruction data of each joint motor of the mechanical arm, and correspondingly, moment feedback data of each joint motor are required to be acquired) are collected as a sample, a fine tuning data set is obtained, the latest group pre-training model is trained according to the fine tuning data set (wherein, in the training process, the motion instruction data in the sample are used as input data of the group pre-training model, the motion feedback data are used for calculating a loss function, and a difference value between the moment feedback data and the moment instruction data is used as an output label of the group pre-training model), and the label is obtained by the model.

The mechanical arm may perform when comparing the fine tuning model with the currently used working model to determine whether the fine tuning model is better than the currently used working model:

executing a designated work task under the condition that a currently used work model (which is a friction compensation model) participates in a control process, and recording the accumulated error (the sum of absolute value deviations of the feedback value and the target value at each moment) between the feedback value and the target value of a target motion parameter (at least one of a position feedback value, a speed feedback value, an acceleration feedback value and a moment feedback value) as a reference accumulated error;

executing a designated work task under the condition that the fine tuning model participates in the control process, and recording the accumulated error between the feedback value and the target value of the target motion parameter, and recording the accumulated error as a comparison accumulated error;

calculating the average value or weighted average value of the accumulated errors of each reference to obtain a reference average error;

calculating the average value or weighted average value of the comparison accumulated errors to obtain comparison average errors;

if the comparison average error is smaller than the reference average error, the fine tuning model is judged to be better than the currently used working model, otherwise, the fine tuning model is judged not to be better than the currently used working model.

The step of scoring the individual optimization model by the mechanical arm comprises the following steps: and scoring the individual optimization model through a preset first scoring rule.

For example, the first scoring rule is:

calculating the accumulated error (the sum of absolute value deviation of the feedback value and the target value at each moment can be recorded as a first accumulated error) between the feedback value and the target value of the target motion parameter (at least one of a position feedback value, a speed feedback value, an acceleration feedback value and a moment feedback value) when the mechanical arm executes a designated work task under the condition that the individual optimization model participates in the control process;

a scoring score is calculated based on the accumulated error (i.e., the first accumulated error).

Wherein the scoring score may be calculated according to the following formula:

；

for scoring score->Is a proportional coefficient (a preset value),>is->Accumulated error (first accumulated error) corresponding to each target motion parameter,/for each target motion parameter>Is->Weight of->Is the total number of terms of the motion parameter of the target.

But are not limited to, calculating a scoring score using the above formula; the preset first scoring rule is not limited to the first scoring rule described above.

Example two

In the second embodiment, the individual node is a voice recognition device (such as a smart phone, a smart speaker, a smart television, and other smart home appliances that can wake up through voice instructions, but not limited thereto), the group pre-training model is a voice instruction recognition model (the voice instruction recognition model may be, but not limited to, a deep learning model, a reinforcement learning model, and the like), and the central node is a management central server. The group of a plurality of speech recognition devices may be a speech recognition device group actually used in a use place (for example, a home, a factory, an office building, etc.), or a speech recognition device group specially constructed in a laboratory to obtain an optimized group pre-training model.

With the development of speech recognition technology, devices having speech recognition functions gradually appear in home scenes, such as: the smart cat eidolon, the little colleague, the small degree, various smart home appliances with voice functions and smart phones are provided, and the phones of many mobile phone manufacturers at present have voice smart wake-up functions, such as Huacheng 'little art smart voice assistant', siri of apples and the like. Even if the voice recognition equipment of the same model is different in accent of a user, common use environments and the like, the recognition effect of a voice command recognition model by unified training is not satisfactory, but if the data of each user is used for independent personalized training, although better individual recognition precision can be obtained, in the use process, the user experience cannot be ensured, the situation that the training process is complicated and long exists, a great amount of time and energy are consumed, and the training cost is higher; if richer speaker and pronunciation scene data are adopted for unified training, the requirement on hardware and the training time cost of a large-scale data set in the training process are high. The training to obtain a pre-trained model using acquisition of a large-scale dataset would be very costly.

This problem can be solved by the distributed population performance intelligent optimization method described above. For the voice recognition equipment group, the voice recognition equipment of each body node is used for executing voice recognition tasks, and the voice instruction sets are the same, but the accents and/or the environments of recognition objects of the voice recognition equipment are different, so that a pre-training model can be shared.

In the second embodiment, the initial group voice command recognition model (i.e., the group voice command recognition model not optimized by the distributed group performance intelligent optimization method) is stored at the management center server, and then the following processes are performed in a loop for a plurality of times:

1. and the management center server sends the group voice command recognition model stored locally to a group communication network of voice recognition equipment, and each voice recognition equipment carries out fine tuning training on the group voice command recognition model after downloading the voice command recognition model to obtain a corresponding individual optimization model.

2. Each voice recognition device scores the individual optimization model trained by the voice recognition device and then sends the individual optimization model to other voice recognition devices through the voice recognition device group communication network so that the other voice recognition devices score the individual optimization model.

3. After the management center server collects the scoring values of all the voice recognition devices to each individual optimization model, the group voice instruction recognition model with the optimal stage is determined from each individual optimization model according to the scoring values.

4. And updating the current group voice command recognition model stored in the management center server according to the group voice command recognition model score which is optimal in the stage.

Wherein the loop process is automatically ended when at least one of the following conditions is met: 1. the circulation times reach the preset maximum times; 2. the comprehensive score of the stage optimal group voice instruction recognition model converges (for example, the deviation of the absolute value of the comprehensive score of the stage optimal group voice instruction recognition model obtained by n times of continuous circulation and the comprehensive score of the stage optimal group voice instruction recognition model obtained by the last circulation is smaller than a preset minimum optimization threshold value, and n is a preset positive integer which can be set according to actual needs); 3. the total time of the optimization process is greater than a preset time interval; 4. the comprehensive scores of the phase optimal group voice instruction recognition models obtained by continuous m times of circulation are lower than the comprehensive scores of the latest group voice instruction recognition models, and m is a preset positive integer and can be set according to actual needs.

After the cycle is completed, the above-described cycle process is automatically started each time at least one of the following conditions is satisfied: 1. adding a new speech recognition data set to the population; 2. the time interval of the last update of the group voice instruction recognition model in the management center server reaches a preset duration threshold; 3. the working model of the voice recognition device with the number exceeding the preset number threshold or the preset proportion threshold in the group is not applicable (for example, according to the use requirement, the voice recognition object instruction is modified in a large range, so that the original model cannot be used for voice recognition, and the optimization flow needs to be carried out again).

Finally, the group voice command recognition model stored in the management center server is an optimized group voice command recognition model, has better generalization performance and absolute performance, and can be sent to each voice recognition device in the voice recognition device group to update the working model, so that the voice recognition performance of the voice recognition device group is improved.

the specific process of the voice recognition equipment for carrying out fine tuning training on the latest group pre-training model comprises the following steps: and collecting multiple pronunciation data (including pronunciation data of wake-up instructions and pronunciation data of non-wake-up instructions) of voice instructions of users of the voice recognition equipment to form a fine tuning data set, and training the latest group pre-training model according to the fine tuning data set to obtain a fine tuning model.

The voice recognition apparatus may perform, when comparing the fine-tuning model with a currently used working model (which is a voice instruction recognition model) to determine whether the fine-tuning model is better than the currently used working model:

identifying a test data set (the test data set comprises a plurality of pronunciation data of voice instructions of a user of the voice recognition equipment, wherein the pronunciation data of the voice instructions and the pronunciation data of non-wake instructions are collected and stored in advance) pre-stored locally by using the working model and the fine tuning model respectively, and recording wake-up rates (also called recall rate, which is the ratio of the number of times of successful wake-up by the wake-up instructions to the total number of wake-up instructions) and false wake-up rates (which is the ratio of the number of times of successful wake-up to the number of times of successful wake-up by the non-wake-up instructions) corresponding to the working model and the fine tuning model respectively, as a first wake-up rate and a first false wake-up rate; calculating a first weighted sum of the first wake-up rate and the first false wake-up rate (namely, calculating the weighted sum of the first wake-up rate and the first false wake-up rate according to a preset first weight, wherein the first weight of the first wake-up rate and the first false wake-up rate can be set according to actual needs); and if the first weighted sum corresponding to the fine tuning model is larger than the first weighted sum of the working model, judging that the fine tuning model is better than the working model, otherwise, judging that the fine tuning model is not better than the working model.

The step of scoring the individual optimization model by the speech recognition device comprises: and scoring the individual optimization model through a preset second scoring rule.

For example, the second scoring rule is:

identifying a pre-stored local test data set by using the individual optimization model, recording a wake-up rate, a false wake-up rate, response time and power consumption, and recording the wake-up rate and the false wake-up rate as a second wake-up rate and a second false wake-up rate;

the scoring score is calculated according to the following formula:

；

for scoring score->For the second wake-up rate, +.>For the second false wake-up rate,>for response time +.>In order for the power to be consumed,、/>、/>、/>four second weights (all preset values). Wherein, the higher the record wake-up rate, the lower the false wake-up rate, the lower the response time and the lower the power consumption, the higher the scoring score, therefore +.>Is positive in number and is added with->、/>、/>All are negative numbers, and specific values can be set according to actual needs.

But are not limited to, calculating a scoring score using the above formula; the preset second scoring rule is not limited to the second scoring rule described above.

Example III

In the third embodiment, the individual node is a vision device, the group pre-training model is an image feature extraction model (the model is used for extracting features of an image, the extracted features are used for specifying tasks, image classification, image segmentation, target detection, and the like), and the central node is a management central server. The group of a plurality of visual devices may be a group of visual devices actually used in a place of use (for example, a factory), or may be a group of visual devices specially constructed in a laboratory to obtain an optimized group pretrained model.

In an intelligent factory, a plurality of robots are generally arranged to perform operations such as workpiece sorting or assembling, and the robots are equipped with visual equipment to perform visual tasks such as sorting, detection, image segmentation and the like on different workpieces. Each visual device performs a corresponding visual task through a corresponding recognition model consisting of an image feature extraction model (or backbone/backbone network model for extracting image features) and a header network (or network structure header for achieving specific uses, such as classification, detection, image segmentation, etc., based on the extracted image features), wherein the image feature extraction model is generic and the header network differs depending on the visual task and recognition object. Therefore, the image feature extraction model with good quality has positive influence on the classification, detection and image segmentation processing effects, and the pre-training model of the image feature extraction model is required to be optimized. Even if the visual devices of the same model are different in visual tasks and recognition objects, the same image feature extraction model is different in use effect, and if the image feature extraction model used by each visual device is subjected to independent personalized training aiming at each task and task object, a great amount of time and energy are consumed in the training process and subsequent maintenance of the model, so that the cost is high and the management is difficult; if the pre-trained model is derived from a model co-trained with a conventional hybrid dataset, the adaptation is not possible: if a new task is added in the visual equipment group and the task is different from other existing tasks, the new task data added data set is needed to be used for retraining, firstly, the size of the mixed data set is further enlarged, the problems of data set composition and the like are still faced, the hardware storage pressure, training duration and training difficulty are also increased, and the method is not intelligent enough; secondly, a great deal of repeated resource waste can be brought; thirdly, the method cannot automatically adapt to task changes of groups and is not intelligent enough; fourth, the difficulty level of data acquisition in different task environments is different, the size proportion of a plurality of data sets is difficult to balance, and the condition of non-ideal training results is easy to cause.

This problem can be solved by the distributed population performance intelligent optimization method described above. For the group of visual devices, the visual devices of each body node are used for executing visual tasks (such as image classification, object detection, image segmentation and the like), but the identification objects and/or specific visual tasks of each visual device are different, so that one image feature extraction model can be shared for combining with each head network to execute each task, and the image feature extraction model is optimized as a group pre-training model.

In the third embodiment, an initial group image feature extraction model (i.e., a group image feature extraction model that is not optimized by the distributed group performance intelligent optimization method) is stored at the management center server, and then the following processes are performed in a loop for a plurality of times:

1. and the management center server sends the group image feature extraction model stored locally to a group communication network of the visual equipment, each visual equipment downloads the group image feature extraction model, forms a model to be trained by using the group image feature extraction model and a head network of the visual equipment, performs fine tuning training on the model to be trained, and extracts the group image feature extraction model in the model to be trained after the fine tuning training to obtain an individual optimization model.

2. And after scoring the individual optimization model obtained by training the visual equipment, the individual optimization model is transmitted to other visual equipment through group communication of the visual equipment so that the other visual equipment scores the individual optimization model (when the visual equipment scores the individual optimization model, the individual optimization model and a head network of the individual optimization model are combined into a model to be scored, and the scoring value of the corresponding individual optimization model is obtained by evaluating the model to be scored).

3. After the management center server collects the scoring values of all the visual equipment to each individual optimization model, the group image feature extraction model with the optimal stage is determined from each individual optimization model according to the scoring values.

4. And updating the current stored group image feature extraction model in the management server according to the group image feature extraction model score which is optimal in the stage.

Wherein the loop process is automatically ended when at least one of the following conditions is met: 1. the circulation times reach the preset maximum times; 2. the comprehensive score of the phase optimal group image feature extraction model converges (for example, the deviation of the absolute value of the comprehensive score of the phase optimal group image feature extraction model obtained by n times of continuous circulation and the comprehensive score of the phase optimal group image feature extraction model obtained by the last circulation is smaller than a preset minimum optimization threshold value, and n is a preset positive integer which can be set according to actual needs); 3. the total time of the optimization process is greater than a preset time interval; 4. the comprehensive scores of the phase optimal group image feature extraction models obtained by continuous m times of circulation are lower than those of the latest group image feature extraction models, and m is a preset positive integer and can be set according to actual needs.

After the cycle is completed, the above-described cycle process is automatically started each time at least one of the following conditions is satisfied: 1. adding new vision equipment into the group; 2. the time interval of the last update of the group image feature extraction model in the management center server reaches a preset duration threshold; 3. the working model of the visual equipment with the number exceeding the preset number threshold or the preset proportion threshold in the group is not applicable (for example, due to the customer or the production requirement, the original work piece for identification is redesigned, the identification effect is reduced by using the original group image feature extraction model, and a large number of work pieces are erroneously identified).

Finally, the group image feature extraction model stored in the management center server is an optimized group image feature extraction model, has better generalization performance and absolute performance, and can be sent to each visual device in the visual device group for updating the working model, so that the image feature extraction performance of the visual device group is improved.

the specific process of the vision equipment for fine tuning training on the latest group pre-training model comprises the following steps: collecting a plurality of images of a task object (namely a workpiece to be identified) of a user of the visual equipment, forming a fine tuning data set, forming a group pre-training model and a head network of the visual equipment into a model to be trained, training the model to be trained by using the fine tuning data set, and finally extracting the group pre-training model (image feature extraction model) in the training model to be trained after training as the fine tuning model.

The vision apparatus may perform, when comparing the fine-tuning model with a currently used working model (which is an image feature extraction model) to determine whether the fine-tuning model is better than the currently used working model:

respectively using the fine tuning model, the working model and a head network stored locally to form a model to be tested and a reference model;

respectively executing corresponding visual tasks on a pre-stored test image set (the test image set comprises image data of a task object of the visual equipment, which is acquired and stored in advance) by using a to-be-detected model and a reference model, and recording task evaluation indexes (for example, for an image segmentation task, the task evaluation indexes are the ratio of average intersection ratio MIoU to overall average precision mAP, and for a target detection task and an image classification task, the task evaluation indexes are the overall average precision mAP, the specific calculation method of the average intersection ratio MIoU to the overall average precision mAP is the prior art, and the detailed description is omitted herein), and the task evaluation indexes are recorded as first task evaluation indexes; and if the first task evaluation index corresponding to the to-be-tested model and the first task evaluation index larger than the reference model are the same, judging that the fine-tuning model is better than the working model, otherwise, judging that the fine-tuning model is not better than the working model.

The step of scoring the individual optimization model by the vision apparatus includes: and scoring the individual optimization model through a preset third scoring rule.

For example, the third scoring rule is:

forming a model to be scored by using the individual optimization model and a head network of the visual equipment;

executing corresponding visual tasks aiming at a local test image set by using the model to be scored, recording task evaluation indexes, and marking the task evaluation indexes as second task evaluation indexes;

and calculating a scoring value according to the second task evaluation index.

For an image segmentation task, the task evaluation index is the ratio of the average intersection ratio MIoU to the overall average precision mAP; for a target detection task and an image classification task, task evaluation indexes are all average precision mAP; the specific calculation method of the average cross-over ratio MIoU and the full-class average accuracy mAP is the prior art, and is not described in detail here.

For example, the scoring score may be calculated according to the following formula:

；

for scoring score->For the second task evaluation index,/I>、/>Is a preset parameter (which can be set according to actual needs).

But are not limited to, calculating a scoring score using the above formula; the preset third scoring rule is not limited to the third scoring rule described above.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other manners. The above-described apparatus embodiments are merely illustrative, for example, the division of the units is merely a logical function division, and there may be other manners of division in actual implementation, and for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be through some communication interface, device or unit indirect coupling or communication connection, which may be in electrical, mechanical or other form.

Further, the units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, may be located in one place, or may be distributed over a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Furthermore, functional modules in various embodiments of the present application may be integrated together to form a single portion, or each module may exist alone, or two or more modules may be integrated to form a single portion.

In this document, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.

The foregoing is merely exemplary embodiments of the present application and is not intended to limit the scope of the present application, and various modifications and variations may be suggested to one skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. The intelligent optimization method for the distributed group performance is characterized by being applied to a central node to optimize a group pre-training model applied to a group; the group comprises a plurality of individual nodes forming a node network, the individual nodes are mutually connected in a communication way, and the central node is connected with all the individual nodes in a communication way and is used for managing the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model; the individual nodes are mechanical arms, voice recognition equipment or visual equipment; the group pre-training model is correspondingly a friction force compensation model, a voice instruction recognition model or an image feature extraction model;

A4. updating the group pre-training model according to the phase optimal group pre-training model;

step A4 includes:

When the comprehensive score corresponding to the stage optimal group pre-training model is higher than the comprehensive score corresponding to the latest group pre-training model, replacing the latest group pre-training model by the stage optimal group pre-training model; otherwise, discarding the phase optimal group pre-training model.

2. The intelligent optimization method of distributed group performance according to claim 1, wherein step A1 comprises:

3. The intelligent optimization method of distributed group performance according to claim 1, wherein step A3 comprises:

4. The intelligent optimization method for the distributed group performance is characterized by being applied to individual nodes to optimize a group pre-training model applied to a group; the group comprises a plurality of individual nodes forming a node network, all the individual nodes are in communication connection with each other, all the individual nodes are in communication connection with a central node, and the central node is used for managing the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model; the individual nodes are mechanical arms, voice recognition equipment or visual equipment; the group pre-training model is correspondingly a friction force compensation model, a voice instruction recognition model or an image feature extraction model;

5. The intelligent optimization method of distributed group performance according to claim 4, wherein step B1 comprises:

6. The distributed group performance intelligent optimization system is used for optimizing a group pre-training model applied to a group and is characterized by comprising a central node and the group, wherein the group comprises a plurality of individual nodes forming a node network, the individual nodes are mutually connected in a communication way, the individual nodes are all connected with the central node in a communication way, and the central node is used for managing the individual nodes; the group pre-training model is provided with model parameters which can be optimized iteratively; each individual node is of the same type with differences in at least one of structural parameters, responsible tasks and task objects, and each individual node is capable of performing fine-tuning training on the group pre-training model and scoring the group pre-training model; the individual nodes are mechanical arms, voice recognition equipment or visual equipment; the group pre-training model is correspondingly a friction force compensation model, a voice instruction recognition model or an image feature extraction model;

the central node executes when updating the group pre-training model according to the phase optimal group pre-training model:

if the latest group pre-training model is an initial group pre-training model, replacing the latest group pre-training model by the phase optimal group pre-training model; if the latest group pre-training model is not the initial group pre-training model, comparing the comprehensive score corresponding to the stage optimal group pre-training model with the comprehensive score corresponding to the latest group pre-training model; when the comprehensive score corresponding to the stage optimal group pre-training model is higher than the comprehensive score corresponding to the latest group pre-training model, replacing the latest group pre-training model by the stage optimal group pre-training model; otherwise, discarding the phase optimal group pre-training model;

7. An electronic device comprising a processor and a memory, the memory storing a computer program executable by the processor, when executing the computer program, running the steps in the distributed population performance intelligent optimization method of any one of claims 1-5.