Intelligent joint control system based on Internet of Things

Abstract

The invention discloses an intelligent joint control system based on the Internet of Things, and the system comprises a sensor data collection unit which is used for obtaining current sensor collection data; the data protocol conversion unit is used for converting communication protocols of different sensors into a unified protocol; the sensor function abstraction unit is used for abstracting and packaging the data acquisition function of the sensor; the data transmission unit is used for transmitting data acquired by the sensor; the cloud data error analysis unit is used for predicting the sensor acquisition data through an LSTM neural network model, comparing an obtained predicted value with the current actual sensor acquisition data, and judging that an error exists if a difference value is greater than a preset threshold value; the data error compensation unit is used for adjusting an output result of an output layer of the LSTM neural network model by adopting a self-attention mechanism, and acquiring a predicted final output result as actual sensor acquisition data; and the cloud data storage unit is used for storing the data acquired by the sensor and the predicted value at the current moment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication data conversion, and particularly relates to an intelligent joint control system based on Internet of Things. BACKGROUND

[0002] Traditional sensors have many problems in outdoor environment deployment and management. The model, function, communication protocol and platform access management mode of the traditional sensors are excessively coupled, which greatly limits the scalability of the system in complex outdoor scenes. The outdoor environment often needs to flexibly increase or decrease sensor deployment according to regional characteristics, for example, in forest fire monitoring, an infrared thermal imaging sensor needs to be added, or in river monitoring, a water quality sensor needs to be supplemented. Due to the close coupling relationship of the traditional system, large-scale hardware modification and software adaptation of the entire outdoor monitoring network are required, which not only has high cost, but also causes the deployment cycle to be greatly prolonged due to the difficulty of outdoor construction.

[0003] Because the outdoor environment needs to flexibly adjust the sensor deployment according to the regional characteristics (such as adding an infrared thermal imaging sensor for forest fire prevention or supplementing a water quality sensor for river monitoring), and the model, function, communication protocol and platform access management of the traditional sensor have a close coupling relationship, when a new sensor is added or adjusted, large-scale hardware modification and software adaptation of the outdoor monitoring network are required. Due to the complex outdoor construction environment (such as terrain restrictions in mountainous areas and river courses), the modification cycle is greatly prolonged and the cost is significantly increased, ultimately resulting in poor scalability of the system. At the same time, the complex and variable outdoor natural conditions (such as extreme weather and geological activities) lead to frequent emergence of new monitoring demands (such as adding wind speed and direction sensors and vibration sensors), but under the traditional coupled architecture, the protocol adaptation and data connection of the new sensor and the existing system require a large amount of customized development, which cannot achieve rapid integration, making it difficult for the system to respond to outdoor dynamic business demands. In addition, the harsh outdoor environments such as high temperature, severe cold and heavy rain can accelerate the performance degradation of sensors, and the traditional sensors lack a targeted upgrade mechanism, which restricts the ability to optimize functions or repair performance degradation through remote methods due to the coupling relationship, and can only rely on on-site replacement, which not only increases the outdoor maintenance cost, but also causes the equipment to gradually fail after performance degradation, which cannot guarantee the long-term stability of outdoor monitoring. SUMMARY

[0004] In view of the problems of sensor performance degradation under the existing Internet of Things joint control and the inability to guarantee the long-term stability of outdoor monitoring, the application provides an intelligent joint control system based on Internet of Things.

[0005] To achieve the above technical purposes, the technical solutions adopted by the application are as follows: An intelligent joint control system based on Internet of Things, comprising a sensor data acquisition unit, a data protocol conversion unit, a sensor function abstraction unit, a data transmission unit, a cloud data error analysis unit, a data error compensation unit and a cloud data storage unit. The sensor data acquisition unit is used to acquire the data currently collected by the sensor. The data protocol conversion unit converts the communication protocols of different sensors into a unified internal protocol that the system can recognize; The sensor function abstraction unit is used to abstract and encapsulate the function of sensor data acquisition; Data transmission unit, used to transmit data collected by the sensor; The cloud-based data error analysis unit uses a trained LSTM neural network model to predict the sensor data collected at the current moment. The predicted value is compared with the actual sensor data collected at the current moment. If the difference is greater than a preset threshold, it is determined that there is an error in the actual sensor data collected at the current moment, and at the same time, it is determined that the sensor is aging and has abnormal errors. The data error compensation unit uses a self-attention mechanism to adjust the output of the output layer of the LSTM neural network model to obtain the final predicted output. The final predicted output is used as the actual sensor data collected before the sensor maintenance that caused the error is completed. The cloud-based data storage unit is used to store sensor-collected data and the predicted values ​​for the current moment obtained from the LSTM neural network model.

[0006] Furthermore, the sensor collects data including location data, temperature data, humidity data, weather data, geological activity data, and other data related to environmental and geological changes.

[0007] Furthermore, the data protocol conversion unit includes multiple protocol parsing modules and protocol conversion modules; A multi-protocol parsing module reads source protocol data from the sensor and parses the data read from the source protocol to determine the data type, format, and content; The protocol conversion module converts the parsed data into a format that the system can recognize using its internal protocols.

[0008] Furthermore, the sensor functional abstraction unit includes a functional abstraction module and an encapsulation module; The functional abstraction module maps the sensor source protocol transmission object and the target object, that is, it maps the source protocol transmission object address and the target object address, and ensures that the mapped address can be recognized by this protocol. The encapsulation module encapsulates and saves the converted sensor data and mapped addresses to the cloud data storage unit.

[0009] Furthermore, the data transmission unit includes a wired or wireless transmission channel for sensor-collected data and a wireless communication channel for uploading data to a cloud data storage unit.

[0010] Further, the LSTM neural network model comprises a forgetting gate F t , an input gate I t and an output gate O t . F t Discard unimportant information by forgetting, and the calculation formula is:

[0011] In the formula, is a Sigmoid function, W F is a weight, b F is a bias term, H t-1 is the unit output at the previous time, X t is the current input; The input gate calculates the update information N t based on the previous time hidden layer information and the current time input information, and the calculation formula is:

[0012] In the formula, Tanh is a state activation function; Then, the forgetting gate and the input gate are combined to calculate the updated cell state C t , and the calculation formula is:

[0013] Finally, the output gate outputs important information in the cell and multiplies it with the updated cell state after state activation to obtain the final output result H t , and the calculation formula is:

[0014] The LSTM neural network uses the forgetting gate to selectively remember the past neural cell state, and the forgetting gate discards unimportant information to reduce the overall calculation amount of the neural network, while effectively avoiding gradient explosion or gradient disappearance.

[0015] When the LSTM neural network performs early warning on time-series attribute data, it can only process point by point in time order, and lacks the processing ability of the same time data in different days. The self-attention mechanism (Self-Attention) can realize parallel computing, filter out deep information from time-series attribute data, and assign different weights according to the importance of information, so that the model pays more attention to information with higher weights.

[0016] Furthermore, before the self-attention mechanism in the data error compensation unit predicts the sensor data of a single category in the sensor acquisition data, it first performs a correlation analysis on the sensor acquisition data of other categories that affect the prediction result of the sensor acquisition data of a single category. The correlation analysis adopts the grey relational analysis method.

[0017] Furthermore, the steps for performing correlation analysis between other categories of sensor data and single-category sensor data using grey relational analysis are as follows: First, it is necessary to determine the main sequence (also known as the reference sequence or mother sequence) and the remaining interference sequences (also known as comparison sequences or sub-sequences). The calculation formula is as follows:

[0018] In the formula, n0 is the main sequence and n j This is an interference sequence.

[0019] Secondly, since the units of the parameters in the original dataset are different, in order to accurately analyze the correlation between the parameters, it is necessary to perform dimensionless quantization on the original dataset. The corresponding dimensionless quantization methods are initialization and mean quantization. Initialization means dividing each data point in a set of data by the first data point, while mean quantization means dividing each data point in a set of data by the average value of that set. The initialization calculation formula is:

[0020]

[0021] The formula for calculating the mean is:

[0022] In the formula, n i (k) represents the kth data point in this dataset.

[0023] Next, the correlation between the main sequence and the interfering sequences is calculated. First, the correlation parameter formula is used:

[0024]

[0025] The formula for recalculating the correlation degree is as follows:

[0026] In the formula, ∆ 0𝑖 (𝑘) = |𝑛0(𝑘) − 𝑛 𝑗 (𝑘)|,𝑛0(𝑖) and 𝑛 𝑖 (𝑖) represent the values ​​of the main term and the distractor term, respectively. ∆ 0𝑖 (x) is the absolute difference, x is the resolution coefficient, and y is the absolute difference.𝑖 is a correlation parameter.

[0027] Finally, the correlation between the main sequence and each interference sequence is calculated and sorted to determine the correlation of the single category sensor data with other category data.

[0028] Further, the self-attention mechanism first performs matrix operations on the query matrix (query, Q) and the key matrix (key, K) to obtain the corresponding weight, then uses size change and node processing, and then uses the Softmax function for normalization processing, and finally performs weighted summation with the data matrix (values, V) to obtain the final output result.

[0029] Further, the self-attention mechanism processing process:

[0030] In the formula, A is the input matrix, Q is the query matrix, K is the key matrix, V is the value matrix, is the dimension number of Q, K, and V, W q , W k , W v is the weight matrix corresponding to Q, K, and V.

[0031] Self-Attention can perform parallel processing of data, but excessive attention to information location can lead to local optimization. Multihead attention (Multihead attention) is based on Self-Attention, which performs multiple Self-Attention learning on input information (related data in the predicted category data in the sensor data, such as position data prediction, which is affected by temperature data, humidity data, weather data, geological activity data, and other environmental and geological change-related data). Then concatenate each Self-Attention learning result and perform linear transformation to get the final result, thereby greatly overcoming the shortcomings of Self-Attention. Multihead attention focuses on high-correlation feature information and continuously adjusts the attention weight during training, thereby improving the warning accuracy.

[0032] Multihead attention first performs linear transformation on the query matrix, key matrix, and value matrix, then calculates the result of each group using scaled dot-product attention, then performs result concatenation (Concat), and finally performs linear transformation and outputs the result.

[0033] Multihead attention mechanism calculation formula:

[0034] wherein W i Q ,W i k ,W i v respectively represent the weight matrix of the query matrix, the key matrix and the value matrix in the i-th head, W m represent the weight matrix of the Multihead attention mechanism.

[0035] Compared with the prior art, the present application has the following beneficial effects: Through the cloud data error analysis unit, the sensor collected data error is analyzed and the data error compensation unit is used to compensate the sensor collected data, so that the collected sensor data will not affect its accuracy with performance loss. At the same time, the data protocol conversion unit is used to convert the collected sensor data, so as to ensure that the system can identify the communication protocol and data of different sensors; the sensor function abstraction unit is used to abstract and encapsulate the function of the sensor collected data, so as to ensure that the intelligent joint control system based on the Internet of Things can quickly respond to outdoor dynamic business needs. When new sensors are added or adjusted, there is no need to make large-scale hardware modification and software adaptation to the outdoor monitoring network, and sensors of various communication protocols are adapted. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The present application provides an overall structure block diagram of an intelligent joint control system based on the Internet of Things.

[0037] The figure mark description: 10-sensor data acquisition unit, 20-data protocol conversion unit, 30-sensor function abstraction unit, 40-data transmission unit, 50-cloud data error analysis unit, 60-data error compensation unit, 70-cloud data storage unit, 201-multiple protocol analysis module, 202-protocol conversion module, 301-function abstraction module, 302-encapsulation module. DETAILED DESCRIPTION

[0038] In order to facilitate the understanding of those skilled in the art, the present application will be further described below in conjunction with the embodiments and the drawings. The content mentioned in the embodiments is not a limitation of the present application.

[0039] As Figure 1 shown, the present embodiment provides an intelligent joint control system based on the Internet of Things, which comprises a sensor data acquisition unit 10, a data protocol conversion unit 20, a sensor function abstraction unit 30, a data transmission unit 40, a cloud data error analysis unit 50, a data error compensation unit 60 and a cloud data storage unit 70. Sensor data acquisition unit 10 is used to acquire the current sensor data; The data protocol conversion unit 20 converts the communication protocols of different sensors into a unified internal protocol that the system can recognize. The multi-protocol parsing and conversion module parses the sensor data according to its protocol type when it arrives, and then converts it into a format that the system can process uniformly. For example, for sensor data using the Bluetooth protocol, the data is first extracted by the Bluetooth parsing module, and then encapsulated according to the internal protocol format for easier subsequent processing.

[0040] The sensor function abstraction unit 30 is used to abstract and encapsulate the function of the sensor to collect data; Data transmission unit 40 is used to transmit data collected by the sensor; The cloud-based data error analysis unit 50 uses a trained LSTM neural network model to predict the sensor data collected at the current moment. The predicted value is compared with the actual sensor data collected at the current moment. If the difference is greater than a preset threshold, it is determined that there is an error in the actual sensor data collected at the current moment, and the age limit is increased. If it exceeds the normal working life of the sensor, it is determined that the sensor is aging and has abnormal errors. The preset threshold is determined by the error range of normal sensor data collected by model training. It will increase with the age of the sensor. Alternatively, it can be predicted by another time-related neural network model to increase the error value range.

[0041] The LSTM neural network model is trained using a historical sensor dataset stored in cloud data storage unit 70. The data error compensation unit 60 uses a self-attention mechanism to adjust the output of the output layer of the LSTM neural network model to obtain the predicted final output. The predicted final output is used as the actual sensor data collected before the sensor maintenance that caused the error is completed. The cloud data storage unit 70 is used to store sensor-acquired data and the predicted value of the current moment obtained by the LSTM neural network model.

[0042] The sensor collects data including location data, temperature data, humidity data, weather data, geological activity data, and other data related to environmental and geological changes.

[0043] The data protocol conversion unit 20 includes a multi-protocol parsing module 201 and a protocol conversion module 202; The multi-protocol parsing module 201 reads source protocol data from the sensor and parses the data read from the source protocol to determine the data type, format, and content. The protocol conversion module 202 converts the parsed data into a format of an internal protocol recognizable by the system. The most core of the protocol conversion module 202 needs to carefully consider factors such as data type and address mapping.

[0044] The sensor function abstraction unit 30 includes a function abstraction module 301 and a packaging module 302. The function abstraction module 301 maps the sensor source protocol transmission object and the action object, that is, maps the source protocol transmission object address and the action object address, and ensures that the mapped address is recognizable by the protocol. The packaging module 302 encapsulates and saves the converted sensor collection data and the mapped address to the cloud data storage unit 70.

[0045] The data transmission unit 40 includes a sensor collection data through a wired or wireless transmission channel and an upload cloud data storage unit 70 wireless communication channel.

[0046] The LSTM neural network model includes a forgetting gate F t , an input gate I t and an output gate O t . F t discards unimportant information by forgetting, and its calculation formula is:

[0047] In the formula, is a Sigmoid function, W F is a weight, b F is a bias term, H t-1 is the output of the unit at the previous moment, and X t is the current input. The input gate calculates the update information N t based on the previous moment hidden layer information and the current moment input information, and its calculation formula is:

[0048] In the formula, Tanh is a state activation function, I t is the input gate value of the previous moment hidden layer information, W I is the previous moment hidden layer weight, b I is the previous moment hidden layer bias term; N t is the input information value at the current moment, W c is the input information weight at the current moment, and b C is the input information bias term at the current moment.

[0049] Then, the forgetting gate and the input gate are combined to calculate the updated unit state C t, the calculation formula is:

[0050] wherein C t-1 is the state of the unit at the previous time; Finally, the output gate outputs the important information in the unit and multiplies the updated unit state activated by the state to obtain the final output result H t , the calculation formula is:

[0051] wherein O t is the important information in the unit of the output gate, W o is the weight of the current unit information of the output gate, b O is the bias term of the current unit of the output gate; Tanh(C t ) is the updated unit state activated by the state.

[0052] The LSTM neural network uses the forget gate to selectively remember the past neural unit state, and the forget gate discards the unimportant information, reducing the overall calculation amount of the neural network, while effectively avoiding gradient explosion or gradient disappearance.

[0053] When the LSTM neural network is used for early warning of time series attribute data, it can only process point by point in time sequence, and lacks the ability to process data at the same time of different days. The self-attention mechanism (Self-Attention) can realize parallel computing, filter out deep information from time series attribute data, and assign different weights according to the importance of information, so that the model pays more attention to information with higher weights.

[0054] Before the self-attention mechanism in the data error compensation unit 60 predicts the single category sensor data in the sensor data, it first performs correlation analysis on the sensor data that has an impact on the prediction results of the single category sensor data. The correlation analysis uses a gray correlation analysis method.

[0055] The steps of using the gray correlation analysis method to analyze the correlation between the other category data and the single category sensor data in the sensor data are as follows: First of all, it is necessary to determine the main sequence (also known as the reference sequence or mother sequence) and the remaining interference sequence (also known as the comparison sequence or subsequence), and the calculation formula is:

[0056] In the formula, n0 is the main sequence and n j is the interference sequence.

[0057] Secondly, because the units of each parameter in the original data set are different, in order to accurately analyze the correlation between each parameter, it is necessary to non-dimensionalize the original data set, and the corresponding non-dimensional processing methods are initial value and mean value. The so-called initial value is that each data in a group of data is divided by the first data value, and the mean value is that each data in a group of data is divided by the average value of the group of data. The initial value calculation formula is:

[0058]

[0059] The mean value calculation formula is:

[0060] In the formula, n i (k) is the kth data in the group of data.

[0061] Then, the correlation between the main sequence and the interference sequence is calculated. First, the correlation parameter formula is calculated:

[0062]

[0063] Then, the correlation degree formula is calculated:

[0064] In the formula, Δ 0𝑖 (k) = |n0(k) - n 𝑗 (k)|, n0(i) and n 𝑖 (i) represent the values of the main item and the interference item, respectively, Δ 0𝑖 (k) is the absolute difference value, p is the resolution coefficient, and R 𝑖 is the correlation parameter.

[0065] Finally, by calculating the correlation between the main sequence and each interference sequence and sorting the correlation, the correlation of the single category sensor collected data and other category data is determined.

[0066] Self-Attention first performs matrix operation on the query matrix (query, Q) and the key matrix (key, K) to obtain the corresponding weight, then uses size change and node processing, and then uses the Softmax function for normalization processing, and finally performs weighted summation with the data matrix (values, V) to obtain the final output result.

[0067] Self-Attention mechanism processing process:

[0068] In the formula, A is an input matrix, Q is a query matrix, K is a key matrix, and V is a value matrix, Q, K, V are dimension numbers, and Wq, Wk, Wv are weight matrices corresponding to Q, K, and V.

[0069] Self-Attention can perform parallel processing of data, but its excessive attention to information position can lead to local optimization. Multihead attention is based on Self-Attention, and multiple sets of Self-Attention learning are performed on input information (other related data in sensor collected data, such as prediction category data, such as position data, which is affected by temperature data, humidity data, weather data, geological activity data and other environmental geological change related data), then the Self-Attention learning results of each set are spliced, and then linearly transformed to obtain the final result, thereby greatly overcoming the shortcomings of Self-Attention. Multihead attention focuses on high correlation feature information and continuously adjusts the attention weight in training, thereby improving the early warning accuracy.

[0070] Multihead attention first linearly transforms the query matrix, the key matrix and the value matrix, then calculates the result of each set by using the scaled dot-product attention, then splices (Concat) the results, and finally linearly transforms and outputs the results.

[0071] Multihead attention mechanism calculation formula:

[0072] In the formula, WQ, Wk and Wv represent the weight matrices of the query matrix, the key matrix and the value matrix in the i-th head, respectively, and W m represents the weight matrix of the Multihead attention mechanism.

[0073] Compared with the prior art, the present application has the following beneficial effects: The sensor acquisition data error analysis by the cloud data error analysis unit 50 and the compensation of the sensor acquisition data by the data error compensation unit 60 ensure that the collected sensor data will not be affected by the performance loss. At the same time, the protocol conversion of the collected sensor data by the data protocol conversion unit 20 ensures that the system can identify the communication protocol and data of different sensors; the abstraction and encapsulation of the function of the sensor acquisition data by the sensor function abstraction unit 30 ensure that the intelligent joint control system based on the Internet of Things can quickly respond to outdoor dynamic business needs. When adding or adjusting the sensor, there is no need to make large-scale hardware modification and software adaptation to the outdoor monitoring network, and the sensor of various communication protocols can be adapted.

[0074] The above describes in detail the intelligent joint control system based on the Internet of Things provided by the present application. The description of the specific embodiments is only used to help understand the method and its core idea. It should be pointed out that for those skilled in the art, without departing from the principles of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. An Internet of Things based intelligent joint control system, characterized in that, The system comprises a sensor data acquisition unit (10), a data protocol conversion unit (20), a sensor function abstraction unit (30), a data transmission unit (40), a cloud data error analysis unit (50), a data error compensation unit (60) and a cloud data storage unit (70); The sensor data acquisition unit (10) is used for acquiring current sensor acquisition data; The data protocol conversion unit (20) converts the communication protocols of different sensors into a unified internal protocol recognizable by the system; The sensor function abstraction unit (30) is used for abstracting and encapsulating the functions of sensor acquisition data; The data transmission unit (40) is used for transmitting sensor acquisition data; The cloud data error analysis unit (50) predicts the current sensor acquisition data through a trained LSTM neural network model, compares the predicted value with the current actual sensor acquisition data, and determines that there is an error in the current actual sensor acquisition data and the sensor is abnormally aged if the difference is greater than a preset threshold; The data error compensation unit (60) adjusts the output result of the output layer of the LSTM neural network model using a self-attention mechanism to obtain a final predicted output result, which is used as the actual sensor acquisition data before the error sensor is maintained; The cloud data storage unit (70) is used for storing sensor acquisition data and the predicted value of the current time obtained by the LSTM neural network model. 2.The smart joint control system based on Internet of Things according to claim 1, characterized in that, The data protocol conversion unit (20) comprises a plurality of protocol analysis modules (201) and a protocol conversion module (202); The plurality of protocol analysis modules (201) read source protocol data from the sensor, analyze the data read from the source protocol, and determine the data type, format and content; The protocol conversion module (202) converts the analyzed data into a format of a system-recognizable internal protocol. 3.The smart joint control system based on the Internet of Things according to claim 2, characterized in that, The sensor function abstraction unit (30) comprises a function abstraction module (301) and an encapsulation module (302); The function abstraction module (301) maps the sensor source protocol transmission object and the action object, i.e., maps the source protocol transmission object address and the action object address, and ensures that the mapped address is recognizable by the protocol; The encapsulation module (302) encapsulates and saves the converted sensor acquisition data and the mapped address to the cloud data storage unit (70). 4.The smart joint control system based on the Internet of Things according to claim 3, characterized in that, The LSTM neural network model comprises a forget gate F t , an input gate I t , and an output gate O t ; F t Discarding unimportant information by means of forgetting, whose formula is: where, is a Sigmoid function, W F is a weight, b F is a bias term, H t-1 is the cell output at the previous time step, X t is the current input; The input gate calculates the previous time hidden layer information and the current time input information to obtain update information N t The calculation formula is: wherein Tanh is a state activation function, I t is the input gate last time hidden layer information value, W I is the last time hidden layer weight, b I is the last time hidden layer bias term; N t is the current time input information value, W c is the current time input information weight, b C is the current time input information bias term; Then, the forget gate is combined with the input gate to calculate the updated cell state C t The calculation formula is: where C t-1 is the state of the cell at the previous time instant; Finally, the output gate will output the important information in the cell and multiply it with the state of the update gate which is activated by the state to get the final output H t The formula is: where, O t is the output gate current cell information, W o is the output gate current cell information weight, b O is the output gate current cell bias term; Tanh(C t ) is the state activated update cell state. 5.The smart joint control system based on Internet of Things according to claim 4, characterized in that, Before the self-attention mechanism in the data error compensation unit (60) predicts the single category of sensor acquisition data, the sensor acquisition data that has an influence on the prediction result of the single category of sensor acquisition data is subjected to correlation analysis, and the correlation analysis is performed based on a grey correlation analysis method. 6.The smart joint control system based on Internet of Things according to claim 5, characterized in that, The steps of performing correlation analysis on the other category data and the single category of sensor acquisition data in the sensor acquisition data based on the grey correlation analysis method are as follows: Firstly, the main sequence and the remaining interference sequences need to be determined; The correlation between various parameters is accurately analyzed by non-dimensionalization of the original data set, and the non-dimensionalization processing method includes initial value and mean value; Then, the correlation between the main sequence and the interference sequence is calculated; Finally, the correlation between the main sequence and each interference sequence is calculated and sorted, thereby determining the correlation between the single category sensor data and other category data. 7.The smart joint control system based on Internet of Things according to claim 6, characterized in that, The self-attention mechanism first performs matrix operation on the query matrix and the key matrix to obtain the corresponding weight, then uses size change and node processing, then uses the Softmax function for normalization processing, and finally performs weighted summation with the data matrix to obtain the final output result. 8.The smart joint control system based on Internet of Things according to claim 7, characterized in that, Self-attention mechanism processing process: where A is the input matrix, Q is the query matrix, K is the key matrix, V is the value matrix, is the dimension number of Q, K, V, W q , W k , W v is the weight matrix corresponding to Q, K, V. 9.The smart joint control system based on the Internet of Things according to claim 8, characterized in that, Through the multi-head attention mechanism, multiple self-attention mechanism learning is performed on the input sensor data; The multi-head attention first performs linear transformation on the query matrix, the key matrix and the value matrix, then calculates the result of each group by using the scaled dot-product attention, then performs result splicing, and finally performs linear transformation and outputs the result. 10.The smart joint control system based on Internet of Things according to claim 8, characterized in that, Mechanism calculation formula: where W i Q where W i k where W i v denote the weight matrix of the query matrix, key matrix and value matrix in the i-th head, respectively, W m denote the weight matrix of the Multihead attention mechanism.

Images