Robot testing method and system based on big data and artificial intelligence

Abstract

The invention discloses a robot testing method and system based on big data and artificial intelligence, and particularly relates to the technical field of robot testing, and the testing method comprises the following steps: S1, constructing a multi-modal testing data collection framework, and collecting multi-dimensional data of a robot in a testing process in real time; s2, the collected multi-modal data is preprocessed; s3, constructing an improved deep neural network model; s4, optimizing the test coverage; s5, integrating the output results of the plurality of evaluation models by adopting a dynamic weight fusion algorithm to obtain a final test conclusion; according to the robot test method and system based on big data and artificial intelligence, the multi-modal data acquisition and fusion technology is adopted, the running state of the robot is comprehensively captured, the test comprehensiveness is improved, and compared with single-modal data, the test efficiency is improved. And multi-modal data fusion enables the test accuracy to be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot testing, and more particularly, to a robot testing method and system based on big data and artificial intelligence. BACKGROUND

[0002] With the rapid development of robot technology, the application scenarios of robots are increasingly widespread, from industrial production to home service, from medical care to education training, robots are gradually penetrating into various fields of social life. The performance and safety of robots are directly related to their application effect and user experience, so it is crucial to test robots comprehensively, efficiently and accurately.

[0003] Traditional robot testing methods mainly rely on manually designed test cases to verify the functions of robots by simulating specific scenarios. This method has the following shortcomings:

[0004] Limited test coverage: manually designed test cases are difficult to cover all possible scenarios, and potential problems are easily missed.

[0005] Poor adaptability: when the functions or application scenarios of robots change, a large number of test cases need to be redesigned, which is costly and inefficient.

[0006] Insufficient evaluation accuracy: traditional testing methods rely heavily on manual judgment, which is highly subjective and prone to misjudgment.

[0007] Difficult to handle multi-modal data: modern robots are equipped with multiple sensors, producing multi-modal data that is difficult to analyze effectively using traditional methods.

[0008] In recent years, the development of big data and artificial intelligence technology has provided new ideas for robot testing. Some research attempts to apply machine learning algorithms to robot testing, but existing methods still have problems such as insufficient model generalization ability, inflexible testing strategies, and poor multi-modal data fusion results.

[0009] Therefore, a robot testing method and system based on big data and artificial intelligence are proposed to solve the above problems. SUMMARY

[0010] In order to overcome the above-mentioned defects of the prior art, the present application provides a robot testing method and system based on big data and artificial intelligence to solve the problems raised in the background art.

[0011] To achieve the above-mentioned purposes, the present application provides the following technical solutions: a robot testing method based on big data and artificial intelligence, the testing method comprising the following steps:

[0012] S1. Construct a multi-modal test data acquisition framework to collect multi-modal data of the robot in real time during the test process;

[0013] S2. Preprocess the collected multi-modal data, including data cleaning, data standardization, and feature extraction;

[0014] S3. Construct an improved deep neural network model to evaluate the function and performance of the robot;

[0015] S4. Generate adaptive test cases based on reinforcement learning algorithms to optimize test coverage;

[0016] S5. Integrate the output results of the improved deep neural network model using a dynamic weight fusion algorithm to obtain the final test conclusion;

[0017] S6. Generate robot performance optimization suggestions based on the test conclusion.

[0018] Further, the multi-modal data in S1 includes visual data, motion data, environmental data, and system state data, and the collected multi-modal data is synchronized through timestamps and stored in a distributed data storage system. The data organization structure used when storing multi-modal data is: wherein is a multi-modal data set, is a timestamp, is visual data, is motion data, is environmental data, is system state data, is the total number of samples.

[0019] Further, the data cleaning in S2 uses an improved Z-score algorithm to detect outliers, and the missing value processing uses an LSTM-based time series prediction model for filling;

[0020] In data standardization, Z-score standardization is used for data following Gaussian distribution, min-max standardization is used for data with values in a limited interval, and image data is normalized to the [0, 1] interval;

[0021] Feature extraction uses an improved autoencoder for data denoising and feature extraction. The improved autoencoder includes an encoder and a decoder. In the encoder, the input data is mapped to a low-dimensional feature space to obtain a feature vector ; In the decoder, the feature vector is reconstructed to the original data space to obtain reconstructed data .

[0022] Further, the improved deep neural network model in S3 comprises:

[0023] An improved convolutional neural network for processing visual sensor data of a robot;

[0024] An improved recurrent neural network for processing time-series motion data of a robot;

[0025] An attention mechanism fusion module for integrating features of different modal data;

[0026] Among them, the improved convolutional neural network MC-CNN is for robot visual data processing, and the improved convolutional neural network with deformable convolution kernel is used to capture irregular features. The specific strategy is: the output feature map calculation formula of deformable convolution is: Wherein is the value of the output feature map at position ; is the size of the convolution kernel; is the th convolution weight; is the input feature map; , is a learnable offset parameter, which is obtained by predicting an additional convolution layer: Wherein is the weight of the offset prediction convolution layer, is an activation function, and the tanh function is used to limit the offset to the range of .

[0027] The network structure of MC-CNN is as follows:

[0028] Input layer: receiving 224×224×3 RGB image or depth image;

[0029] Deformable convolution block 1: 32 3×3 deformable convolution kernels, step 1, padding 1, connected with BatchNorm and ReLU;

[0030] Pooling layer 1: 2×2 max pooling, step 2;

[0031] Deformable convolution block 2: 64 3×3 deformable convolution kernels, step 1, padding 1, connected with BatchNorm and ReLU;

[0032] Pooling layer 2: 2×2 max pooling, step 2;

[0033] Deformable convolution block 3: 128 3×3 deformable convolution kernels, step 1, padding 1, connected with BatchNorm and ReLU;

[0034] Global average pooling layer: convert the feature map into a 128-dimensional feature vector;

[0035] Output layer: set the corresponding output dimension and activation function according to the specific task (such as object recognition, defect detection).

[0036] Further, the improved recurrent neural network MB-GRU adopts an improved recurrent neural network of bidirectional gated recurrent unit GRU for the time sequence characteristics of robot motion data, captures the past and future time sequence dependence, and the specific strategy is:

[0037] MB-GRU is composed of forward GRU and backward GRU, the forward GRU processes data from the beginning to the end of the sequence, the backward GRU processes data from the end to the beginning of the sequence, and finally the hidden states of the two directions are spliced as the output;

[0038] The update formula of a single GRU unit is:

[0039] The update gate controls whether to update the hidden state, specifically:

[0040] The reset gate controls whether to ignore the historical hidden state, specifically:

[0041] The candidate hidden state, specifically:

[0042] The final hidden state, specifically:

[0043] wherein, is the update gate; is the reset gate; is the current hidden state; is the sigmoid activation function; is the element-level multiplication; , , respectively input the weight matrix acting on the update gate, the reset gate and the candidate hidden state; , respectively are the historical hidden state the weight matrix acting on the update gate and the reset gate; is the historical state filtered by the reset gate the weight matrix acting on the candidate hidden state ; respectively are the bias vectors of the update gate, the reset gate and the candidate hidden state;

[0044] ​​​​​​The output of the bidirectional GRU is:

[0045] Forward GRU hidden state ;

[0046] Backward GRU hidden state ;

[0047] Concatenate the forward and backward hidden states: ;

[0048] where is the forward hidden state at the previous time step; is the backward hidden state at the next time step.

[0049] The network structure of the MB-GRU is as follows:

[0050] Input layer: receives a sequence of time-series feature vectors with dimension d

[0051] Bidirectional GRU layer 1: 64 hidden units, returns the complete sequence

[0052] Bidirectional GRU layer 2: 32 hidden units, returns the complete sequence

[0053] Temporal attention layer: assigns attention weights to different time steps in the sequence

[0054] Fully connected layer: maps the output to the corresponding dimension, used for robot motion stability, trajectory accuracy, etc.

[0055] Further, the attention mechanism fusion module, i.e. AMFM, integrates different modal data and learns the weights of different modal features, and the specific strategy of the attention mechanism fusion module is:

[0056] Let the feature vectors of modalities obtained after feature extraction be , and , where represents the feature vector of the th modality; is the dimension of the feature vector, represents a real vector of dimension ;

[0057] The attention mechanism fusion process is:

[0058] Feature mapping: map modal features of different dimensions to the same dimensional space: , where and are learnable parameters, is the unified feature dimension; is the The output feature vector after mapping of each modal feature; For the first The original feature vectors of each modality; This is the weight matrix; For one A bias vector of dimension; express A dimensional real vector;

[0059] Attention weight calculation: Weights are calculated based on the similarity between features and context vectors, specifically as follows: ,in For the first Attention weights for each modality feature; These are learnable context vectors; The cosine similarity function;

[0060] Feature fusion: Weighted summation yields the final fused features. Specifically: ;

[0061] Feature enhancement: Enhancing fused features through residual connections and nonlinear transformations Specifically: ,in, This is a residual connectivity feature; and These are learnable parameters.

[0062] Furthermore, the specific strategy for generating adaptive test cases using the reinforcement learning algorithm in S4 is as follows:

[0063] Modeling the reinforcement learning environment: Define the state space. This is represented as a combination of the robot's current state and the test environment state, specifically: ,in , , and These are feature extraction functions for visual, motion, environmental, and system state data, respectively. This is the current test coverage vector; , , and The proportion is Real-time visual, motion, environmental, and system state data; motion space A set of executable test operations. ,in For the first Each test action can be an operation, such as changing ambient lighting, adding distractions, or adjusting task difficulty. For the continuous action space, a Gaussian distribution is used to represent the action probability distribution.

[0064] reward function The settings are made for test coverage improvement, problem severity discovery and test efficiency, specifically: , wherein is the test coverage improvement amount; is the newly discovered problem set; is the severity of the problem ; is the test time consumption; , , is the weight coefficient, and ; is the state after the execution of the action ;

[0065] In S4, the reinforcement learning algorithm adopts an improved deep deterministic policy gradient algorithm, i.e., M-DDPG, which introduces a priority experience replay and an adaptive learning mechanism.

[0066] Further, the dynamic weight in the dynamic weight fusion algorithm used in S5 is adaptively adjusted according to the performance of each model.

[0067] Further, the performance optimization suggestion in S6 is based on the test evaluation results and the constructed test knowledge graph, and provides specific suggestions for the optimization of the hardware and software of the robot.

[0068] Further, a robot testing system based on big data and artificial intelligence comprises: a multi-modal data acquisition module for acquiring various sensor data and running state data of a robot;

[0069] a data preprocessing module for cleaning, standardizing and feature extracting the acquired data;

[0070] a deep learning evaluation module comprising multiple improved neural network models for evaluating various performances of the robot;

[0071] an intelligent test case generation module for dynamically generating test cases based on a reinforcement learning algorithm;

[0072] a result fusion module for integrating the evaluation results of multiple models using a dynamic weight fusion algorithm;

[0073] a feedback optimization module for generating performance optimization suggestions for the robot according to the test results;

[0074] The multi-modal data acquisition module comprises:

[0075] a visual data acquisition unit comprising a high-definition camera and a depth sensor for acquiring visual data;

[0076] Motion data acquisition unit, including accelerometer, gyroscope and force sensor for collecting motion data;

[0077] Environmental data acquisition unit, including temperature and humidity sensor, barometric pressure sensor and sound sensor for collecting environmental data;

[0078] System state acquisition unit for collecting CPU usage, memory usage and battery status of the robot.

[0079] Technical effects and advantages of the present application:

[0080] Compared with the prior art, the robot testing method and system based on big data and artificial intelligence comprehensively captures the running state of the robot by adopting multi-modal data acquisition and fusion technology, improves the comprehensiveness of the test, and compared with single modal data, the multi-modal data fusion improves the test accuracy;

[0081] The improved deep neural network model is innovatively proposed, including the introduction of convolutional neural network with deformable convolution kernel and bidirectional gated recurrent unit, which enhances the extraction ability of complex features. In the industrial robot weld quality detection task, the recognition accuracy is improved;

[0082] The adaptive test case generation mechanism based on reinforcement learning can dynamically adjust the test strategy according to the real-time performance of the robot, improving the test efficiency and coverage. In the test of household service robot, the test coverage is significantly improved compared with the traditional method;

[0083] The dynamic weight fusion algorithm can adaptively adjust the weight according to the real-time performance of each model, improving the accuracy and reliability of the evaluation result. In multi-model integrated testing, dynamic weight fusion improves the evaluation result;

[0084] The introduction of test knowledge graph provides knowledge support for the test process, enhancing the explainability and intelligent level of the system. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 The method flowchart of the present application;

[0086] Figure 2 The system flowchart of the present application.

[0087] The figure marks are: 1, multi-modal data acquisition module; 101, visual data acquisition unit; 102, motion data acquisition unit; 103, environmental data acquisition unit; 104, system state acquisition unit; 2, data preprocessing module; 3, deep learning evaluation module; 4, intelligent test case generation module; 5, result fusion module; 6, feedback optimization module. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0089] Embodiment 1

[0090] As shown in a robot testing method based on big data and artificial intelligence, the testing method comprises the following steps: Figure 1

[0091] S1. Constructing a multi-modal testing data acquisition framework to acquire multi-modal data of a robot in a testing process in real time;

[0092] The multi-modal data in S1 comprises visual data, motion data, environment data and system state data, the acquired multi-modal data is synchronized through a time stamp, and is stored in a distributed data storage system, and a data organization structure adopted when the multi-modal data is stored is: , wherein is a multi-modal data set, is a time stamp, is visual data, is motion data, is environment data, is system state data, is a total number of samples.

[0093] S2. Preprocessing the acquired multi-modal data, including data cleaning, data standardization and feature extraction;

[0094] The data cleaning in S2 adopts an improved Z-score algorithm to detect abnormal values, and a time series prediction model based on LSTM is adopted to fill in missing values;

[0095] In the data standardization, Z-score standardization is adopted for data conforming to a Gaussian distribution, min-max standardization is adopted for data with values in a limited interval, and pixel values are normalized to the interval [0, 1] for image data;

[0096] The feature extraction adopts an improved autoencoder to perform data denoising and feature extraction, the improved autoencoder comprises an encoder and a decoder, and in the encoder, the input data is mapped to a low-dimensional feature space to obtain a feature vector ; in the decoder, the feature vector is reconstructed to an original data space to obtain reconstructed data .​

[0097] S3. Constructing an improved deep neural network model to evaluate the function and performance of the robot;

[0098] The improved deep neural network model in S3 includes:

[0099] An improved convolutional neural network for processing visual sensor data of the robot;

[0100] An improved recurrent neural network for processing time-series motion data of the robot;

[0101] An attention mechanism fusion module for integrating features of different modal data;

[0102] Among them, the improved convolutional neural network MC-CNN is for robot visual data processing. The improved convolutional neural network with deformable convolution kernel captures irregular features. The specific strategy is: the output feature map calculation formula of deformable convolution is: wherein is the value of the output feature map at position is the size of the convolution kernel; is the th convolution weight; is the input feature map; , is a learnable offset parameter, which is obtained by predicting an additional convolution layer: wherein is the weight of the offset prediction convolution layer, is an activation function, which uses the tanh function to limit the offset to the range of .

[0103] The network structure of MC-CNN is as follows:

[0104] Input layer: receiving 224x224x3 RGB images or depth images;

[0105] Deformable convolution block 1: 32 3x3 deformable convolution kernels, step 1, padding 1, connected with BatchNorm and ReLU;

[0106] Pooling layer 1: 2x2 max pooling, step 2;

[0107] Deformable convolution block 2: 64 3x3 deformable convolution kernels, step 1, padding 1, connected with BatchNorm and ReLU;

[0108] Pooling layer 2: 2x2 max pooling, step 2;

[0109] ​Deformable Conv Block 3: 128 3x3 deformable conv kernels, stride 1, padding 1, followed by BatchNorm and ReLU;

[0110] Global Average Pooling: convert the feature map to a 128-dimensional feature vector;

[0111] Output Layer: set the output dimension and activation function according to the specific task (such as object recognition, defect detection).

[0112] The improved recurrent neural network MB-GRU is a bidirectional gated recurrent unit GRU improved recurrent neural network aiming at the time sequence characteristics of robot motion data, which captures the past and future time sequence dependence relationship, and the specific strategy is:

[0113] MB-GRU is composed of forward GRU and backward GRU, the forward GRU processes data from the beginning to the end of the sequence, and the backward GRU processes data from the end to the beginning of the sequence, and finally the hidden states of the two directions are spliced as the output;

[0114] The update formula of a single GRU unit is:

[0115] The update gate controls whether to update the hidden state, which is:

[0116] The reset gate controls whether to ignore the historical hidden state, which is:

[0117] The candidate hidden state is:

[0118] The final hidden state is:

[0119] wherein, is the update gate; is the reset gate; is the current hidden state; is the sigmoid activation function; is the element-level multiplication; , , respectively input the weight matrix acting on the update gate, the reset gate and the candidate hidden state; , respectively are the historical hidden states acting on the update gate and the reset gate; is the historical state filtered by the reset gate acting on the candidate hidden state ; , ,​​​​ bias vectors of the update gate, the reset gate and the candidate hidden state, respectively;

[0120] The output of the bidirectional GRU is:

[0121] The forward GRU hidden state

[0122] The backward GRU hidden state

[0123] Concatenate the forward and backward hidden states:

[0124] where is the forward hidden state at the previous time step; is the backward hidden state at the next time step.

[0125] The network structure of the MB-GRU is as follows:

[0126] Input layer: receives a sequence of time-series feature vectors with dimension d

[0127] Bidirectional GRU layer 1: 64 hidden units, returns the complete sequence

[0128] Bidirectional GRU layer 2: 32 hidden units, returns the complete sequence

[0129] Temporal attention layer: assigns attention weights to different time steps in the sequence

[0130] Fully connected layer: maps the output to the corresponding dimension, used for robot motion stability, trajectory accuracy, etc.

[0131] The attention mechanism fusion module, AMFM, integrates different modal data and learns the weights of different modal features. The specific strategy of the attention mechanism fusion module is:

[0132] Let the feature vectors of modalities obtained after feature extraction be , and , where represents the feature vector of the th modality; is the dimension of the feature vector, represents a real vector with dimension .

[0133] The attention mechanism fusion process is:

[0134] Feature mapping: map modal features of different dimensions to the same dimensional space: , where and are learnable parameters,​​​ unified feature dimension; is the output feature vector of the th modality feature after mapping; is the original feature vector of the th modality; is the weight matrix; is a dimensional bias vector; represents a real-valued vector;

[0135] Attention weight calculation: calculate the weight based on the similarity between features and context vectors, specifically: where is the attention weight of the th modality feature; is the learnable context vector; is the cosine similarity function;

[0136] Feature fusion: weighted sum to get the final fusion feature , specifically: ;

[0137] Feature enhancement: enhance the fusion feature through residual connection and nonlinear transformation , specifically: where, is the residual connection feature; and are learnable parameters.

[0138] S4. Generate adaptive test cases based on reinforcement learning algorithm to optimize test coverage;

[0139] The specific strategy for generating adaptive test cases by reinforcement learning algorithm in S4 is:

[0140] Modeling the reinforcement learning environment: let the state space represent the joint of the current state of the robot and the test environment state, specifically: where , , and are feature extraction functions of visual, motion, environment and system state data respectively; is the current test coverage vector; , , and are the visual, motion, environment and system state data at moment; action space is the set of executable test operations, wherein is the th test action; and can be operations such as changing the ambient light, increasing the interference, adjusting the task difficulty, etc. For continuous action space, Gaussian distribution is used to represent the action probability distribution;

[0141] reward function is set for test coverage improvement, problem severity and test efficiency, specifically: wherein is the test coverage improvement amount; is the newly discovered problem set; is the severity of the problem ; is the test time consumption; , , is the weight coefficient, and ; is the state after the action is executed;

[0142] In S4, the reinforcement learning algorithm uses an improved deep deterministic policy gradient algorithm, i.e., M-DDPG, to introduce a priority experience replay and an adaptive learning mechanism.

[0143] S5. The output results of the improved deep neural network model are integrated by using a dynamic weight fusion algorithm to obtain a final test conclusion;

[0144] In S5, the dynamic weights in the dynamic weight fusion algorithm are adaptively adjusted according to the performance of each model.

[0145] S6. A performance optimization suggestion for a robot is generated according to the test conclusion.

[0146] In S6, the performance optimization suggestion is based on the test evaluation results and the constructed test knowledge graph, and provides specific suggestions for the software and hardware optimization of the robot.

[0147] Embodiment 2

[0148] Referring to Figure 2 , a robot testing system based on big data and artificial intelligence, comprising: a multi-modal data acquisition module 1 for acquiring various sensor data and running state data of a robot;

[0149] a data preprocessing module 2 for cleaning, standardizing and feature extracting the acquired data;

[0150] a deep learning evaluation module 3 comprising a plurality of improved neural network models for evaluating various performances of the robot;

[0151] The intelligent test case generation module 4 dynamically generates test cases based on reinforcement learning algorithms;

[0152] The result fusion module 5 integrates the evaluation results of multiple models using a dynamic weight fusion algorithm;

[0153] The feedback optimization module 6 generates robot performance optimization suggestions based on test results;

[0154] The multi-modal data acquisition module 1 includes:

[0155] The visual data acquisition unit 101 includes a high-definition camera and a depth sensor for collecting visual data;

[0156] The motion data acquisition unit 102 includes an accelerometer, a gyroscope, and a force sensor for collecting motion data;

[0157] The environmental data acquisition unit 103 includes a temperature and humidity sensor, a barometric pressure sensor, and a sound sensor for collecting environmental data;

[0158] The system state acquisition unit 104 is used to collect the CPU usage, memory occupation, and battery status of the robot.

Claims

1. A robot testing method based on big data and artificial intelligence, characterized by, The test method comprises the following steps: S1. Construct a multi-modal test data acquisition framework to acquire multi-modal data of the robot in real time during the test process; S2. Preprocess the acquired multi-modal data, including data cleaning, data standardization, and feature extraction; S3. Construct an improved deep neural network model to evaluate the function and performance of the robot; S4. Generate adaptive test cases based on a reinforcement learning algorithm to optimize test coverage; The improved deep neural network model in S3 comprises: An improved convolutional neural network for processing visual sensor data of the robot; An improved recurrent neural network for processing time-series motion data of the robot; An attention mechanism fusion module for integrating features of different modal data; In the improved convolutional neural network, i.e., MC-CNN, a deformable convolution kernel is introduced to capture irregular features, and the specific strategy is as follows: the output feature map calculation formula of the deformable convolution is as follows: wherein is the value of the output feature map at position ; is the convolution kernel size; is the mth convolution weight; is the input feature map; , , is a learnable offset parameter, which is obtained by an additional convolution layer: wherein is the weight of the offset prediction convolution layer, is an activation function, and the tanh function is used to limit the offset to the range of . The improved recurrent neural network, namely MB-GRU, adopts a bidirectional gated recurrent unit, namely GRU, to capture past and future time-series dependencies based on the time-series characteristics of the robot motion data, and the specific strategy is as follows: The MB-GRU is composed of a forward GRU and a backward GRU, the forward GRU processes data from the beginning to the end of the sequence, and the backward GRU processes data from the end to the beginning of the sequence, and finally the hidden states of the two directions are spliced as the output; The update formula of a single GRU unit is as follows: The update gate controls whether to update the hidden state, specifically: ; Reset the door control whether to ignore the history hidden state, in particular: ; candidate hidden states, in particular: ; Final hidden state, in particular: ; wherein, is an update gate; is a reset gate; is a current hidden state; is a sigmoid activation function; is an element-wise multiplication; , , are input vectors are weight matrices acting on the update gate, the reset gate, and the candidate hidden state; , are history hidden states are weight matrices acting on the update gate and the reset gate; is a history state filtered by the reset gate is a weight matrix acting on the candidate hidden state ; , , are bias vectors for the update gate, the reset gate, and the candidate hidden state, respectively; The output of the bidirectional GRU is as follows: forward GRU hidden state ; backward GRU hidden state ; Concatenate the forward and backward hidden states: ; wherein is a forward hidden state at a previous time; is a backward hidden state at a next time; The attention mechanism fusion module, namely AMFM, integrates features of different modal data and learns weights of different modal features, and the specific strategy of the attention mechanism fusion module is as follows: Let the feature vector of the i-th modality after feature extraction be , and where denotes the feature vector of the i-th modality; is the dimension of the feature vector, denotes a real vector of dimension .​​ The attention mechanism fusion process is as follows: Feature mapping: mapping different dimensional modal features to the same dimensional space wherein and are learnable parameters, is the unified feature dimension; is the output feature vector of the th modal feature after mapping; is the original feature vector of the th modal; is the weight matrix; is a dimensional bias vector; denotes a dimensional real vector; Attention weight calculation: the weight is calculated based on the similarity between the feature and the context vector, specifically: wherein is the attention weight of the th modality feature; is the learnable context vector; is the cosine similarity function; Feature fusion: weighted sum to get the final fusion feature , in particular: ; Feature enhancement: enhance fused features by residual connection and nonlinear transformation , in particular: wherein, is a residual connection feature; and are learnable parameters; S5. Integrate the output results of the improved deep neural network model using a dynamic weight fusion algorithm to obtain the final test conclusion; S6. Generate robot performance optimization suggestions based on the test conclusion.

2. The robot testing method based on big data and artificial intelligence according to claim 1, characterized in that: The multi-modal data in the S1 includes visual data, motion data, environment data and system state data, and the collected multi-modal data is synchronized by a timestamp and stored in a distributed data storage system, and the data organization structure adopted when the multi-modal data is stored is: Wherein is a multi-modal data set, is a timestamp, is visual data, is motion data, is environment data, is system state data, is the total number of samples.

3. The robot testing method based on big data and artificial intelligence according to claim 2, characterized in that: The improved Z-score algorithm is used for detecting outliers in the data cleaning of S2, and the time-series prediction model based on LSTM is used for filling in missing values; In the data standardization, Z-score standardization is used for data following Gaussian distribution, min-max standardization is used for data with values in a limited interval, and pixel values are normalized to the interval [0, 1] for image data; The feature extraction adopts an improved auto-encoder to perform data noise reduction and feature extraction, the improved auto-encoder comprises an encoder and a decoder, and in the encoder, input data is mapped to a low-dimensional feature space to obtain a feature vector ​​ The feature vector is decoded in the decoder to the original data space, resulting in reconstructed data .

4. The robot testing method based on big data and artificial intelligence according to claim 3, characterized in that: The specific strategy of the reinforcement learning algorithm for generating adaptive test cases in S4 is as follows: Modeling the reinforcement learning environment: Let the state space be represented as a joint of the current state of the robot and the test environment state, in particular: where , , and are feature extraction functions for visual, motion, environment, and system state data, respectively; is the current test coverage vector; , , and are the proportions of instantaneous visual, motion, environment, and system state data; the action space is the set of executable test operations, where is the test action; reward function is set for test coverage improvement, severity of discovered problems and test efficiency, specifically: wherein is the improvement amount of test coverage; is the newly discovered problem set; is the severity of problem ; is the test time consumption; , , is the weight coefficient, and ; is the state after the execution of action ; The improved deep deterministic policy gradient algorithm, namely M-DDPG, is used in the reinforcement learning algorithm in S4 to introduce a priority experience replay and an adaptive learning mechanism.

5. The robot testing method based on big data and artificial intelligence according to claim 4, characterized in that: The dynamic weights in the dynamic weight fusion algorithm used in S5 are adaptively adjusted according to the performance of each model.

6. The robot testing method based on big data and artificial intelligence according to claim 5, characterized in that: The performance optimization suggestions in S6 are based on test evaluation results and a constructed test knowledge graph to provide specific suggestions for software and hardware optimization of the robot.

7. A big data and artificial intelligence based robotic testing system, characterized in that: It comprises: A multi-modal data acquisition module for acquiring various sensor data and running state data of the robot; A data preprocessing module for cleaning, standardizing, and feature extracting the acquired data; A deep learning evaluation module containing multiple improved neural network models for evaluating various performances of the robot; An intelligent test case generation module for dynamically generating test cases based on a reinforcement learning algorithm; An intelligent test case generation module for dynamically generating test cases based on a reinforcement learning algorithm; A result fusion module is configured to integrate the evaluation results of the multiple models by using a dynamic weight fusion algorithm. A feedback optimization module is configured to generate a performance optimization suggestion for the robot according to the test results. The multi-modal data acquisition module comprises: A visual data acquisition unit comprising a high-definition camera and a depth sensor for acquiring visual data; A motion data acquisition unit comprising an accelerometer, a gyroscope and a force sensor for acquiring motion data; An environmental data acquisition unit comprising a temperature and humidity sensor, a barometric pressure sensor and a sound sensor for acquiring environmental data; A system state acquisition unit for acquiring the CPU usage, memory occupation and battery status of the robot.

Images