Intelligent fusion terminal data transmission adaptive method and device based on deep learning

Abstract

The invention relates to the field of data transmission, in particular to an intelligent fusion terminal data transmission self-adaption method and device based on deep learning, and the method comprises the steps: a terminal collects multi-source data in real time, and constructs a sequence for model input; based on the sequence, utilizing a pre-trained prediction model to generate a prediction sequence of a future data trend; calculating a risk state vector based on the prediction sequence, calculating a network state vector based on the network state parameters in the sequence, and combining the risk state vector and the network state vector to form a state vector; and inputting the state vector into an improved SAC algorithm which dynamically suppresses Q value over-estimation during updating of the double-Q network through a serial correction mechanism and outputs a data transmission action parameter to execute data transmission. According to the method, the problem of Q value over-estimation in a real noise environment can be effectively solved, the stability of algorithm training and the reliability of a strategy are enhanced, and adaptive data transmission optimization under multi-target conflicts is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data transmission. More particularly, the present application relates to a deep learning-based intelligent fusion terminal data transmission adaptive method and device. BACKGROUND

[0002] In the Internet of Things scenario, the intelligent fusion terminal needs to efficiently coordinate the data transmission of multiple services under the condition of limited resources and dynamically changing network environment. To balance the real-time performance, reliability and terminal power consumption of data transmission, the existing technology often uses a deep reinforcement learning algorithm, especially the SAC (Soft Actor-Critic) algorithm, as the core learning framework of the adaptive transmission strategy.

[0003] The standard SAC algorithm explores the transmission action space (such as modulation, compression, and scheduling parameters) through a maximum entropy mechanism, and estimates the action value using a double Q network. However, the edge network environment in which the intelligent fusion terminal is located has significant observation noise such as channel measurement error and dynamic noise such as sudden interference, which causes the double Q network value estimation based on Bellman update to be severely inaccurate. This estimation error is constantly accumulated and amplified in the time series update, causing the policy network (Actor) to learn unstable or even oscillating transmission strategies, which cannot achieve accurate and stable balance among multiple targets, ultimately affecting the stability and efficiency of system transmission. SUMMARY

[0004] To solve the technical problem that the above-mentioned existing method based on the standard SAC cannot be reliably applied in a real noisy environment, the present application provides a solution in the following aspects.

[0005] In a first aspect, a deep learning-based intelligent fusion terminal data transmission adaptive method includes: The intelligent fusion terminal collects multi-source data including electrical parameters, environmental parameters, and network state parameters in real time, and pre-processes the collected multi-source data to construct a sequence for model input; Based on the sequence, a pre-trained prediction model is used to generate a prediction sequence of future data trends; a risk state vector is calculated based on the prediction sequence, and a network state vector is calculated based on the network state parameters in the sequence; the risk state vector and the network state vector are combined to form a state vector; The state vector is input into an improved SAC algorithm, which dynamically suppresses the overestimation of Q values during double Q network update through a serial correction mechanism, and outputs data transmission action parameters; wherein the serial correction mechanism generates a correction factor for adjusting the future reward discount term based on the network state vector and the conflict degree among multiple targets; perform data transmission on the sequence according to the data transmission action.

[0006] Preferably, the preprocessing of the collected multi-source data comprises interpolating and aligning the multi-source data, and normalizing the multi-source data, and intercepting the multi-source data by using a sliding window.

[0007] Preferably, the risk state vector comprises a voltage out-of-limit risk value, a current out-of-limit risk value, a power out-of-limit risk value, an overall fluctuation risk value, an environment drastic change risk value, and a network congestion risk value.

[0008] Preferably, the network state vector comprises a received signal strength indication mean value, a received signal strength indication standard deviation, and a packet loss rate change amount.

[0009] Preferably, the improved SAC algorithm uses an experience replay buffer to store interaction records, the records comprising a current state vector, a transmission action, a multi-objective reward, a real-time reward, a reliability reward, a power consumption reward, and a next state vector.

[0010] Preferably, the generation of the correction factor comprises: calculating a network fluctuation sensitivity based on the network state vector; calculating a first Pearson correlation coefficient between the real-time reward sequence and the reliability reward sequence, and a second Pearson correlation coefficient between the reliability reward sequence and the power consumption reward sequence, in the interaction records sampled from the experience replay buffer in the recent period; if the first Pearson correlation coefficient is negative, taking the absolute value of the negative value as a first correction term, otherwise, taking 0 as the first correction term; if the second Pearson correlation coefficient is negative, taking the absolute value of the corresponding negative value as a second correction term, otherwise, taking 0 as the second correction term; adding the sum of the first correction term and the second correction term to 1, and multiplying the added result by the network fluctuation sensitivity to obtain a multi-objective conflict degree; based on the multi-objective conflict degree, calculating the correction factor by exponential decay.

[0011] Preferably, the acquisition of the network fluctuation sensitivity comprises: dividing the received signal strength indication standard deviation in the network state vector by the received signal strength indication mean value to obtain an initial ratio value; multiplying the initial ratio value by an adjustment term to obtain the network fluctuation sensitivity; the adjustment term is the sum of 1 and the packet loss rate change amount in the network state vector.

[0012] Preferably, the acquisition of the network fluctuation sensitivity comprises: dividing the received signal strength indication standard deviation in the network state vector by the received signal strength indication mean value to obtain an initial ratio value; The initial ratio is multiplied by an adjustment term to obtain the network fluctuation sensitivity; the adjustment term is the greater of the change in the packet loss rate in the network state vector and the maximum packet loss rate in the prediction sequence.

[0013] Preferably, the future reward discount term is the product of an original discount factor and the correction factor.

[0014] In a second aspect, the application discloses a deep learning-based intelligent fusion terminal data transmission adaptive device, which comprises a processor and a memory, and the memory stores computer program instructions.

[0015] The application has the following advantages: The application introduces a serial correction mechanism to dynamically suppress the overestimation of the Q value of the double Q network in the SAC algorithm, and adaptively adjusts the reward discount in combination with the network state and the degree of multi-target conflict, thereby enhancing the training stability and policy reliability of the algorithm in a real noise environment, and improving the adaptive decision-making capability of the intelligent fusion terminal data transmission. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 FIG. 1 is a method flowchart of steps S1-S4 in the deep learning-based intelligent fusion terminal data transmission adaptive method according to an embodiment of the application.

[0017] Figure 2 FIG. 2 is an effect diagram of the improved SAC algorithm according to an embodiment of the application.

[0018] Figure 3 FIG. 3 is a structural schematic diagram of the deep learning-based intelligent fusion terminal data transmission adaptive device according to an embodiment of the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application.

[0020] The application takes an intelligent fusion terminal deployed in an edge scene such as power monitoring as an application object. In such a scene, the terminal needs to simultaneously ensure the real-time and reliable uploading of monitoring data and the long-term low-power operation of the device itself. To achieve this goal, the terminal must intelligently schedule its transmission behavior. The basis for scheduling is derived from the grasp of three types of core information: electrical parameters reflecting the essence of power system operation, environmental parameters affecting the stable operation of the device, and network state parameters determining the success or failure and efficiency of data transmission. Therefore, the collaborative processing and intelligent analysis of these three types of multi-source data are the cornerstone of realizing adaptive transmission decisions.

[0021] Referring to Figure 1 The intelligent fusion terminal data transmission adaptive method based on deep learning includes steps S1-S4, as follows: S1: The intelligent fusion terminal collects multi-source data containing electrical parameters, environmental parameters, and network state parameters in real time, and pre-processes the collected multi-source data to construct a sequence for model input.

[0022] As mentioned earlier, electrical parameters, environmental parameters, and network state parameters are the key inputs for decision-making. However, these multi-source data come from different sensors or modules, with different physical meanings, importance, and variation frequencies. For example, voltage and current may suddenly mutate and need to be captured at high frequency, while environmental temperature changes slowly and can be collected at low frequency. If these original, asynchronous, and heterogeneous signals are directly used for subsequent calculations, it will lead to distortion of system state description, and thus any intelligent decision-making model will be built on an unreliable basis. Therefore, the primary task is to convert the original data into a set of time-aligned and scale-unified sequences.

[0023] In one embodiment, the terminal starts the collection of the above three types of data in parallel and sets the collection frequency according to their variation characteristics.

[0024] Electrical parameters: collect voltage, current, and power at a frequency of 1 time / second to ensure that transient fluctuations and precursors of power system failures can be captured.

[0025] Environmental parameters: collect temperature and humidity at a frequency of 1 time / minute to match the slowness of environmental changes.

[0026] Network state parameters: collect received signal strength indicator (RSSI) and packet loss rate at a frequency of 1 time / 10 seconds to balance between real-time monitoring and system overhead.

[0027] Further, the highest sampling frequency (1 time / 1 second) among the three is selected as the unified reference sampling frequency.

[0028] For environmental parameters and network state parameters, linear interpolation algorithm is used between any two adjacent actual samples to calculate the estimated value at each reference sampling frequency, so that all data are strictly aligned on the same time axis.

[0029] Further, each multi-source data after time alignment is independently processed by minimum-maximum normalization.

[0030] After the time alignment and normalization of multi-source data, it is necessary to determine how much data should be used for each decision. If each decision is based only on the normalized data of a single sampling time, the state representation will be fragile due to the lack of historical information, and cannot reflect the trend changes of the data (such as the slow rise of voltage and the continuous attenuation of signal). On the contrary, if all historical data since the start of the device is used, it will contain a large amount of obsolete information irrelevant to the current decision, increase the computational burden, and may cause the strategy to respond slowly to recent changes.

[0031] Therefore, a sliding window containing the last 5 minutes of data is defined, and each time a transmission decision needs to be made, the aligned data in a complete window from the current time back to the past is intercepted to form a sequence.

[0032] Through the above series of operations, a high-quality sequence is finally obtained.

[0033] S2: Based on the sequence for model input, a pre-trained prediction model is used to generate a prediction sequence of future data trends; based on the prediction sequence, a risk state vector is calculated, and based on the network state parameters in the sequence for model input, a network state vector is calculated; the risk state vector and the network state vector are combined to form a state vector.

[0034] The sequence finally obtained by the above S1 describes a state from the past to the present, but a high-quality transmission strategy also needs to have a certain foresight, such as increasing the transmission priority before predicting that the voltage will exceed the limit, while also fitting the current network transmission capacity to avoid blindly sending a large amount of data when the network is congested.

[0035] Therefore, two types of features need to be extracted from the above sequence: one is a feature representing the future risk of the data itself, and the other is a feature depicting the current real-time status of the network.

[0036] In one embodiment, the sequence finally obtained by the above S1 is input into a pre-trained long short-term memory network model. The model has learned the time evolution law between electrical parameters and environmental parameters.

[0037] Then the long short-term memory network model outputs a prediction sequence reflecting the future data trend, which is the same length as the above sequence, i.e. the prediction sequence contains the predicted values of each data in the next 5 minutes.

[0038] Further, a risk state vector is extracted from the prediction sequence, and a network state vector is extracted from the sequence finally obtained by the above S1. The risk state vector aims to quantify the loss or risk that each data may bring if the transmission is delayed. The network state vector aims to objectively describe the health of the current transmission channel.

[0039] The aforementioned risk state vector includes voltage over-limit risk values, current over-limit risk values, power over-limit risk values, overall fluctuation risk values, environmental drastic change risk values, and network congestion risk values. The numerical values ​​for each dimension are obtained as follows: Calculate the difference between the maximum predicted voltage value in the prediction sequence and the preset safety threshold (i.e., the normalized upper limit of voltage measurement). Use the larger of this difference and 0 as the voltage over-limit risk value. Similarly, obtain the current over-limit risk value and the power over-limit risk value using the same method as obtaining the voltage over-limit risk value.

[0040] Calculate the coefficients of variation (standard deviation / mean) of voltage, current and power in the predicted sequence, and then average all the coefficients of variation to obtain the overall volatility risk value.

[0041] The differences in predicted temperature and humidity values ​​at two adjacent sampling times are calculated separately, along with the standard deviations of the predicted temperature and humidity values ​​in the prediction sequence. The first ratio is obtained by dividing the difference in predicted temperature values ​​at two adjacent sampling times by the corresponding standard deviations of the predicted temperature values, and the second ratio is obtained by dividing the difference in predicted humidity values ​​at two adjacent sampling times by the corresponding standard deviations of the predicted humidity values. The first and second ratios are added together and averaged over all sampling times to obtain the environmental drastic change risk value.

[0042] The maximum value of the predicted packet loss rate in the predicted sequence is taken as the network congestion risk value.

[0043] The aforementioned network state vector includes the mean of the received signal strength indication, the standard deviation of the received signal strength indication, and the packet loss rate variation (i.e., the difference between the maximum and minimum packet loss rates in the final sequence obtained from S1).

[0044] Finally, the 6-dimensional risk state vector and the 3-dimensional network state vector are concatenated to obtain a 9-dimensional state vector.

[0045] In summary, the final state vector contains both early warnings of key future risks and assessments of current transmission conditions. This allows subsequent decisions to be made efficiently based directly on the state vector without processing the original data.

[0046] S3: Input the state vector into the improved SAC algorithm, which dynamically suppresses the overestimation of Q value during the update of the dual-Q network through a serial correction mechanism and outputs the data transmission action parameters.

[0047] In real-world noisy environments encountered by terminals, when using the standard SAC algorithm for policy learning, the dual-Q network used to evaluate action value in the standard SAC algorithm suffers from Q-value overestimation due to the inherent maximization bias of temporal difference learning. This error accumulates and amplifies during temporal updates, causing the actions output by the policy network to produce meaningless and violent oscillations in parameters such as priority and frequency, making it impossible to achieve a stable balance among multiple objectives. Therefore, targeted improvements to the value update mechanism of the SAC algorithm are needed.

[0048] Specifically, the core components of the SAC algorithm are still retained: a policy network for generating actions (priority, compression rate, frequency) based on the state; two value networks for independently evaluating the value of state-action pairs; and an experience replay buffer for storing interaction records (current state, transmitted actions, multi-objective rewards, next state, and round end flag).

[0049] Furthermore, improvements are made to the existing core components: First, the original state in the policy network is replaced with the state vector obtained in S2 above. Second, a serial correction mechanism is introduced to improve the calculation method of the target Q value. Finally, the interaction record is stored, including the current state vector, transmission action, multi-objective reward, real-time reward, reliability reward, power consumption reward, next state vector, and round end flag.

[0050] The serial correction mechanism generates a correction factor for adjusting future reward discount terms based on the network state vector and the degree of conflict among multiple objectives. The core of this mechanism involves introducing three key intermediate parameters: network volatility sensitivity, multi-objective conflict degree, and the correction factor. Network volatility sensitivity directly quantifies the intensity of environmental noise, which is the physical root cause of Q-value overestimation. The multi-objective conflict degree assesses the amplification effect of multi-objective competition on overestimation error under a given noise level, connecting environmental noise and policy instability. The correction factor transforms the quantified risk into a regulatory signal that can be applied to the algorithm's core update, achieving adaptive conservative adjustment.

[0051] In summary, by concatenating network volatility sensitivity, multi-objective conflict level, and correction factor, this mechanism organically couples network volatility at the physical layer, objective conflict at the policy layer, and value update at the algorithm layer, thereby achieving precise and dynamic suppression of overestimation. After obtaining the correction factor, it is applied to the discount factor, precisely applying the suppression force to the most critical link in the value update chain (future reward discount term), thus achieving source suppression of Q-value overestimation and ultimately outputting a stable, smooth, and adaptive transmission strategy.

[0052] In one embodiment, the process of obtaining the above network sensitivity is as follows: First, divide the standard deviation of RSSI in the network state vector by the mean of RSSI to obtain an initial ratio (to avoid the denominator being 0, a very small constant can be set), which is the coefficient of variation of RSSI, to measure the relative fluctuation of the signal.

[0053] Then, the sum of 1 (representing the baseline or benchmark of network fluctuations) and the change in packet loss rate in the network state vector is calculated as an adjustment term.

[0054] Finally, multiplying the initial ratio by the adjustment term yields a network fluctuation sensitivity. An increase in the initial ratio indicates more severe fluctuations relative to its mean, leading to increased environmental instability. Similarly, an increase in the packet loss rate indicates more severe fluctuations in transmission quality, also resulting in increased environmental instability.

[0055] In one embodiment, the process of obtaining the degree of conflict among the multiple objectives is as follows: Calculate the first Pearson correlation coefficient between the real-time reward sequence and the reliability reward sequence, and the second Pearson correlation coefficient between the reliability reward sequence and the power consumption reward sequence, calculated from the interaction records recently sampled from the experience playback buffer. If the first Pearson correlation coefficient is negative, the absolute value of the negative value is used as the first correction term; otherwise, 0 is used as the first correction term. Similarly, if the second Pearson correlation coefficient is negative, the absolute value of the corresponding negative value is used as the second correction term; otherwise, 0 is used as the second correction term. Add the sum of the first and second correction terms to 1, and multiply the result by the network fluctuation sensitivity to obtain the degree of multi-objective conflict. The greater the network fluctuation sensitivity or the stronger the conflict between objectives, the greater the degree of multi-objective conflict, indicating that in a noisy and contradictory environment, the overestimation error will be further amplified.

[0056] The real-time performance bonus is the transmission delay normalized using minimum and maximum values; the reliability bonus is the calculated transmission success rate (i.e., the number of successfully acknowledged packets in the experience replay buffer divided by the total number of packets sent); and the power consumption bonus is the energy consumption normalized using minimum and maximum values.

[0057] The minimum value in the normalization of the above real-time reward is 0 (the theoretical ideal value), and the maximum value is the transmission timeout threshold set by the system (which can be set to 1 in the power monitoring scenario). That is, the normalization is performed by the ratio of the difference between the maximum value and the actual delay to the difference between the maximum value and the minimum value.

[0058] The criteria for determining the "successful confirmation packet" in the aforementioned reliability reward are as follows: For the TCP protocol: Success is defined as receiving an acknowledgment frame with the corresponding sequence number within the ACK timeout period. Successful retransmission after the timeout is still counted.

[0059] For the UDP protocol: Success is considered achieved if a correctly verified acknowledgment message is received within the application layer ACK timeout period. Timeouts or verification failures are not counted.

[0060] It should be noted that in the data transmission scenario of intelligent converged terminals, reliability is the core hub and the focal point of conflict between real-time performance and power consumption. Most conflicts between real-time performance and power consumption caused by policy adjustments do not occur directly, but are transmitted and manifested through reliability as an intermediate variable. Therefore, when calculating the degree of multi-objective conflict, the Pearson correlation coefficient between real-time performance reward and power consumption reward is not directly calculated.

[0061] In one embodiment, the multi-objective conflict level is calculated using exponential decay to obtain the aforementioned correction factor, which satisfies the following relationship: In the formula, As a correction factor, This is the scaling factor (which can be set empirically, such as 0.5). For the degree of multi-objective conflict, It is an exponential function with the natural number e as its base.

[0062] The above Increase The decrease indicates that when environmental noise is high and target conflict is strong, the system automatically adopts a highly conservative value update strategy, which strongly suppresses strategy oscillations caused by overly optimistic estimation.

[0063] However, when In very large cases, such as when the network is extremely unstable and there are intense conflicts between targets, the situation can be very challenging. The calculation result will approach 0 if If the value is directly equal to this value which is close to 0, then the discount factor will also change in the future. This means that when updating the Critic network, the algorithm may ignore the long-term rewards in the future and only focus on the immediate rewards.

[0064] Therefore, by using the max function to The lower limit is set to a threshold. (e.g., 0.1), ensuring that regardless of The discount factor must be at least 0.1 of the original value.

[0065] The product of the final determined correction factor and the fixed original discount factor is then used as the future reward discount term, i.e., the improved effective discount factor.

[0066] Based on all the improvements mentioned above, we substitute them into the original Q-value calculation formula. Specifically, we only replace the discount factor in the formula with the future reward discount term calculated above, while keeping all other parameters unchanged.

[0067] For example, the updated formula for calculating the Q value is: In the formula, For the updated Q value, This is an immediate, multi-objective comprehensive reward (i.e., including rewards for real-time performance, reliability, and power consumption). To improve the effective discount factor, , Represent the next state and the state determined by the current policy, respectively. The next action generated, This means taking the smaller value of the next state-action pair value estimate from the two Q-networks. This is a classic design of the SAC algorithm to mitigate overestimation. The entropy coefficient, The policy entropy term is a standard component of the SAC algorithm.

[0068] Furthermore, with the above improvements, when the network environment is noisy or the target collision is strong, the correction factor... Reduced, leading to This reduction leads to a proactive and dynamic decrease in the dependence of the target Q-value on the future state value estimate during Q-network updates. This weakens the strength of the backpropagation of the inherent error (overestimation) in the future Q-value estimate to the current Q-value, curbing the accumulation of overestimation at the source. It also provides the Q-network with a gradient update target that focuses more on immediate and reliable reward signals, making the training process more stable and thus outputting a more robust value assessment, avoiding the policy oscillations caused by it.

[0069] In another embodiment, a different process for obtaining network sensitivity is also provided: First, divide the standard deviation of RSSI in the network state vector by the mean of RSSI to obtain an initial ratio (to avoid the denominator being 0, a very small constant can be set), which is the coefficient of variation of RSSI, to measure the relative fluctuation of the signal.

[0070] Then, the larger of the change in packet loss rate in the network state vector and the maximum packet loss rate in the predicted sequence is used as an adjustment term.

[0071] Finally, the initial ratio is multiplied by the adjustment term to obtain a network fluctuation sensitivity.

[0072] S4: Perform data transmission on the sequence used as model input according to the data transmission action.

[0073] The terminal actuator performs the following steps on the sequence finally obtained in S1, based on the data transmission action parameters output in S3: Priority scheduling involves placing the sequence into the appropriate position in the sending queue.

[0074] Adaptive compression uses the compression ratio specified by the output data transmission action to compress and encode the data.

[0075] Rate control refers to sending data packets according to the time interval specified by the output data transmission action.

[0076] After the above operations are completed, the actual network transmission of the sequence is initiated.

[0077] In addition, considering that a static strategy cannot adapt to long-term drift of network conditions or changes in data distribution, a real-time closed loop is constructed from "decision-making" to "execution" and then to "learning," enabling the system to continuously optimize online using the feedback generated by its own actions, and ultimately obtain long-term adaptive capabilities.

[0078] In one embodiment, after the above transmission is completed, the actual performance indicators of the current transmission are collected, and the real-time reward, reliability reward and power consumption reward are calculated. Then, the normalized values ​​of the real-time reward, reliability reward and power consumption reward are added together to obtain the comprehensive reward, which corresponds to the multi-objective reward in the interaction record stored in the experience playback buffer.

[0079] The complete experience of the current interaction is stored in the experience replay buffer. This experience is used to update the parameters of the neural network in subsequent updates, thereby making the policy better and better.

[0080] Through the above series of operations, an example is given as follows: Figure 2 The diagram shows a comparison between the improved SAC algorithm of this invention and the standard SAC algorithm.

[0081] This invention also provides an adaptive data transmission device for intelligent fusion terminals based on deep learning. For example... Figure 3 As shown, the device includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the deep learning-based intelligent fusion terminal data transmission adaptive method according to the first aspect of the present invention.

[0082] The device also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the setup and functions of which are known in the art and will not be described in detail here.

[0083] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A deep learning-based intelligent fusion terminal data transmission adaptive method, characterized in that, include: The intelligent fusion terminal collects multi-source data, including electrical parameters, environmental parameters, and network status parameters, in real time, and preprocesses the collected multi-source data to construct a sequence for model input. Based on the sequence, a predicted sequence of future data trends is generated using a pre-trained prediction model; a risk state vector is calculated based on the predicted sequence, and a network state vector is calculated based on the network state parameters in the sequence; the risk state vector and the network state vector are combined to form a state vector. The state vector is input into the improved SAC algorithm, which dynamically suppresses the overestimation of Q value during the update of the dual-Q network through a serial correction mechanism and outputs data transmission action parameters; wherein, the serial correction mechanism generates a correction factor for adjusting the future reward discount term based on the network state vector and the degree of conflict between multiple objectives. Data transmission is performed on the sequence according to the data transmission action.

2. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 1, characterized in that, The preprocessing of the collected multi-source data includes interpolation and alignment of the multi-source data, normalization processing, and using a sliding window to extract the multi-source data.

3. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 1, characterized in that, The risk state vector includes voltage over-limit risk value, current over-limit risk value, power over-limit risk value, overall fluctuation risk value, environmental drastic change risk value, and network congestion risk value.

4. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 1, characterized in that, The network state vector includes the mean of received signal strength indication, the standard deviation of received signal strength indication, and the change in packet loss rate.

5. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 1, characterized in that, The improved SAC algorithm uses an experience replay buffer to store interaction records, which include the current state vector, transmission action, multi-objective reward, real-time reward, reliability reward, power consumption reward, and next state vector.

6. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 5, characterized in that, The generation of the correction factor includes: Calculate network fluctuation sensitivity based on network state vector; Calculate the first Pearson correlation coefficient between the real-time reward sequence and the reliability reward sequence, and the second Pearson correlation coefficient between the reliability reward sequence and the power consumption reward sequence, in the interaction records recently sampled from the experience playback buffer. If the first Pearson correlation coefficient is negative, the absolute value of the negative value is used as the first correction term; otherwise, 0 is used as the first correction term. Similarly, if the second Pearson correlation coefficient is negative, the absolute value of the corresponding negative value is used as the second correction term; otherwise, 0 is used as the second correction term. Add the sum of the first and second correction terms to 1, and multiply the sum by the network fluctuation sensitivity to obtain the degree of multi-objective conflict. The correction factor is calculated using exponential decay based on the degree of multi-objective conflict.

7. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 6, characterized in that, The acquisition of network fluctuation sensitivity includes: Divide the standard deviation of the received signal strength indication in the network state vector by the mean of the received signal strength indication to obtain an initial ratio. The network fluctuation sensitivity is obtained by multiplying the initial ratio by an adjustment term, which is the sum of 1 and the change in packet loss rate in the network state vector.

8. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 6, characterized in that, The acquisition of network fluctuation sensitivity includes: Divide the standard deviation of the received signal strength indication in the network state vector by the mean of the received signal strength indication to obtain an initial ratio. The network fluctuation sensitivity is obtained by multiplying the initial ratio by an adjustment term; the adjustment term is the larger of the change in packet loss rate in the network state vector and the maximum packet loss rate in the prediction sequence.

9. The adaptive data transmission method for intelligent fusion terminals based on deep learning according to claim 1, characterized in that, The future reward discount term is the product of the original discount factor and the correction factor.

10. A deep learning-based intelligent fusion terminal data transmission adaptive device, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the deep learning-based intelligent fusion terminal data transmission adaptive method according to any one of claims 1-9.

Images