receiving, by an LPWAN (low-power wide-area network) gateway receiver, an uplink signal transmitted by a terminal node, and demodulating the uplink signal by using a soft demodulation algorithm, wherein for a demodulation result, under a model of an additive white Gaussian noise channel, a noise follows Gaussian distribution with zero mean N(0, σ²), and based on a maximum a posteriori criterion, there is a conditional probability P(Y|X_A^ω):

where Y denotes an actual signal received by the receiver, and X_A^ω denotes a modulated signal of the symbol A^ω∈Ω;

determining a bit log-likelihood ratio calculated according to the uplink signal by a Bayes formula

 wherein transmission probabilities p(x) of symbols are equal, formula so that the bit log-likelihood ratio of the k^th bit is as follows:

where Ω_k,0, Ω_k,1 denotes sets of symbols with the k^th bit equal to 0 and 1, respectively, and Ω={A¹, A², . . . , A^W} denotes a set of all possible symbols received;

thereby obtaining a set of bit log-likelihood ratios and realizing sensing of a wireless channel;

selecting, by a deep reinforcement learning model, key frequency points as pilot frequencies according to the bit log-likelihood ratio;

performing a channel estimation of the key frequency points based on the pilot frequencies according to a modulated channel model to obtain estimated channels of the key frequency points;

performing an original estimation of a complete channel according to the estimated channels of the key frequency points by an interpolation method to obtain an original estimated channel; and

performing N iterative estimations on the original estimated channel through N symbols acquired in a continuous time slice window, and completing reconstruction of a complete channel of a corresponding link during communication to obtain a reconstructed channel;

calculating a current optimal network configuration according to the reconstructed channel so as to allocate the current optimal network configuration to terminal nodes during the next downlink transmission;

before actual communication, performing offline training on the deep reinforcement learning model by using the following reward function (1), so that the model is used for the first round of communication during actual communication:

during the actual communication, performing online reinforcement learning training by using the following reward function (2), so that dynamically adaptive overall optimal communication transmission in a local LPWAN network is finally realized through continuous iterations:

where Ĥ denotes a reconstructed channel, H denotes a real channel generated by simulation in the dataset, p denotes a frequency point position, ∥H∥=N denotes a total number of frequency points, Ĥ(p) denotes the reconstructed channel corresponding to the frequency point p, H(p) denotes the real channel corresponding to the frequency point p, X_t−1 denotes communication performance parameters for the previous round t−1, X_t denotes communication performance parameters for the current round t, and X_t=(x_t,i), 0≤i≤∥X∥, ω_i, and α_i are all weight coefficients.