A Turbo code probability domain decoding method based on a memory-computing integrated device

Abstract

The application discloses a Turbo code probability domain decoding method based on a memory-computing integrated device and particularly relates to the technical field of channel coding and decoding; original branch probabilities and complementary mirror probabilities are written into adjacent memory cells of the same word line in pairs, and a differential response is obtained by using forward reading bias and reverse reading bias to construct a mismatch-resistant branch metric; a state shielding set is introduced to control the limited participation of local aliasing states, and forward recursion and backward recursion of a first component code and a second component code are respectively completed; further, position-related re-labeling of the branch metric is performed by combining a state entropy sequence to generate a position credible weight; finally, a pseudo-convergence index is constructed by using posterior probability, component code difference, adjacent iteration hard decision stability and entropy collapse rate to realize identification and reverse correction of a pseudo-convergence state; the application can effectively inhibit small probability compression and error amplification problems caused by array mismatch of the memory-computing integrated device and improve decoding stability and accuracy.

Description

Technical Field

[0001] This invention relates to the field of channel coding and decoding technology, and specifically to a Turbo code probability domain decoding method based on a memory-based computing unit. Background Technology

[0002] In the scenario of emergency short packet telemetry transmission from high-altitude long-endurance UAVs, in order to reduce power consumption and size, the receiver often uses an in-memory computing unit to implement Turbo code probability domain decoding. However, in low-voltage arrays, after the residual conductance of adjacent cells, bit line voltage drop and low-confidence branch of punched code are coupled, the extremely small probability and its complementary probability are prone to mirror compression, causing abnormal collapse of local state entropy in the later stage of iteration, resulting in rare but serious pseudo-convergence. This causes a small number of control frames to be misinterpreted even when hard decision is stable, which is difficult to suppress with existing probability normalization or fixed stopping criteria. Summary of the Invention

[0003] The purpose of this invention is to provide a Turbo code probabilistic domain decoding method based on a memory computing device to address the shortcomings in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a Turbo code probability domain decoding method based on a memory-based computing device, comprising: S100: Obtain the received sequence y, interleaving mapping π, perforation parameter P and in-memory computing array mismatch parameter η, generate the original branch probability of each state transition, and construct the corresponding complementary mirror probability group Γ0. S200, Γ0 is written into the same word line according to the adjacent mapping of the original branch probability and the complementary mirror probability, the differential response is read to obtain the anti-mismatch branch metric Γ1, and the state mask set Ω is generated according to the probability span of Γ1; S300, based on Γ1 and Ω, perform forward and backward recursion of the first component code to obtain the first external information probability Q1 and the state entropy sequence H1; S400, interweave Q1 by π and combine it with H1 to form a positional confidence weight W, and use W to recalibrate Γ1 to obtain the updated branch metric Γ2; S500, based on Γ2 and Ω, perform forward and backward recursion of the second component code to obtain the second external information probability Q2; S600, the posterior probability R is obtained by deinterleaving Q2, and the pseudo-convergence index Z is constructed based on the difference between R and Q1, the hard decision stability of adjacent iterations and the entropy collapse rate of H1. S700: When Z is greater than the preset threshold, extract the mismatch sensitive bit in R, reverse correct Γ1, and return to S300 to continue iterating; when Z is less than or equal to the preset threshold, output the decoding result according to R.

[0005] Preferably, S100 includes: S110, the received sequence is backfilled with positions according to the perforation parameters, and the system bit observations and check bit observations are recombined into time-series corresponding observation pairs; S120, Based on the mismatch parameters of the in-memory computing array, establish a probabilistic correction model including cell conduction drift, word line voltage drop and bit line coupling disturbance, and generate the mismatch compensation factor corresponding to each observation pair column from the probabilistic correction model. S130, using the mismatch compensation factor to perform probability domain mapping on each observation pair column, and obtaining the original branch probability corresponding to each state transition in the code trellis diagram; S140, based on each original branch probability, generate complementary mirror probabilities according to the complementary constraints of the target event being true and the target event not being true under the same state transition, and associate the original branch probabilities with the corresponding complementary mirror probabilities according to the state transition number to form a complementary mirror probability group Γ0.

[0006] Preferably, S200 includes: S210, Write the original branch probabilities and corresponding complementary mirror probabilities in the complementary mirror probability group into adjacent memory cells on the same word line according to the correspondence of the same state transition number and the same time position. S220, apply a forward read bias and a reverse read bias to the adjacent memory cells after writing, obtain the difference between the two read current values, and correspond the read current difference value to the differential response of the state transition; S230, based on the degree of closeness between the differential response of each state transition and the preset reference differential partition, the original branch probability is recalibrated against mismatch, and the anti-mismatch branch metric Γ1 is generated; S240: Sort all anti-mismatch branch metrics at the same time, calculate the probability span between the maximum and the second largest values, and add the corresponding state transitions with probability spans less than the preset shielding threshold to the state shielding set Ω.

[0007] Preferably, S300 includes: S310, based on the state masking set, the anti-mismatch branch metric at each time position is screened to determine the open state transitions that are allowed to participate in the recursion and the masked state transitions that are restricted from participating in the recursion. S320, in forward time order, the mismatch-resistant branch metric of each open state transition is jointly accumulated with the forward state probability of the previous time position, and the restricted accumulation of each shielded state transition is performed in a weighted retention manner to obtain the first forward state probability sequence. S330, in reverse chronological order, the anti-mismatch branch metric of each open state transition is jointly accumulated with the backward state probability of the next time position, and the restricted accumulation is performed on each shielded state transition according to the weighted retention method to obtain the first backward state probability sequence. S340, based on the first forward state probability sequence, the first backward state probability sequence and the anti-mismatch branch metric at each time position, generate the first external information probability Q1, and generate the state entropy sequence H1 based on the probability dispersion of each state transition within the same time position.

[0008] Preferably, S400 includes: S410, rearrange the positions of the first external information probability Q1 according to the interleaving mapping π to obtain the interleaved external information sequence corresponding to the time position of the second component code; S420, based on the magnitude of the entropy value at each time position in the state entropy sequence H1 and the continuity of entropy change between adjacent time positions, the credibility level of each time position in the interwoven external information sequence is classified to obtain the position credibility level sequence. S430, jointly constrain the location credibility level sequence and the interleaved external information sequence at the same time location to generate the location credibility weight W corresponding to each time location; S440, using the location confidence weight W to perform location-related recalibration against mismatch branch metric Γ1, so that the mismatch-resistant branch metric of high confidence time location maintains its distinguishing advantage, and the mismatch-resistant branch metric of low confidence time location shrinks towards the neutral probability direction, generating an updated branch metric Γ2.

[0009] Preferably, S500 includes: S510, based on the updated branch metric Γ2 and the state masking set Ω, the state transitions at each time position of the second component code are screened to determine the second open state transition and the second masked state transition. S520, in forward time order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second forward state probability of the previous time position, and the second shielded state transition is subjected to restricted accumulation in a weighted retention manner to obtain the second forward state probability sequence. S530, in reverse chronological order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second backward state probability at the next time position, and the restricted accumulation is performed on each second shielded state transition according to the weighted retention method to obtain the second backward state probability sequence. S540, based on the second forward state probability sequence, the second backward state probability sequence and the update branch metric Γ2 at each time position, generate the second external information probability Q2.

[0010] Preferably, S600 includes: S610, the second external information probability Q2 is restored according to the inverse mapping of the interleaving mapping π to obtain the back-placed external information sequence corresponding to the time position of the first component code, and the posterior probability R is generated based on the back-placed external information sequence; S620, compare the probability shift between the posterior probability R and the first external information probability Q1 at the same time position to obtain the difference sequence at each time position; S630, perform hard decisions on the posterior probabilities R of the current iteration and the previous iteration respectively, compare the consistency of decisions at the same time position, and obtain the hard decision stability of adjacent iterations; S640, read the entropy value change trend at each time position in the state entropy sequence H1, determine the entropy collapse rate at each time position, and construct the pseudo-convergence index Z based on the difference sequence, the hard decision stability of adjacent iterations and the entropy collapse rate.

[0011] Preferably, S700 includes: S710, when the pseudo-convergence exponent Z is greater than the preset threshold, extract the mismatch sensitive positions that coincide with the high difference position and the high entropy collapse rate position from the posterior probability R to form a mismatch sensitive position set. S720, based on the distribution of the mismatch sensitive bit set at each time position, perform reverse correction on the state transition corresponding to the mismatch sensitive bit set in the adversarial mismatch branch metric Γ1, weaken the probability bias of the current direction and restore the competitive state; S730, regenerate the state mask set Ω corresponding to the time position based on the reversed mismatch branch metric Γ1, and return to step S300 to continue the iteration; S740: When the pseudo-convergence index Z is less than or equal to the preset threshold, the information bits at each time position are finally hard-determined based on the posterior probability R, and the decoding result is output.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention writes the original branch probability and complementary mirror probability in pairs into adjacent memory cells on the same word line, and extracts the differential response of state transition using forward read bias and reverse read bias. Then, based on the differential response, the original branch probability is recalibrated against mismatch. This effectively weakens the amplification effect of mismatched branch metrics in the in-memory compute array at the probability domain decoding front end. Compared to existing technologies that only use unified normalization, static compensation, or conventional probability correction methods, this invention does not directly correct all probabilities with the same amplitude. Instead, it utilizes the differential distinguishability of the original branch probability and complementary mirror probability under the same perturbation conditions to reconstruct a mismatched branch metric that more closely resembles the actual path competition relationship. Furthermore, it generates a state mask set through probability span to restrict participation control of local aliasing states, thereby improving the ability to identify the true preferred path during the recursion of the first and second component codes, and reducing the risk of local error propagation caused by the combined effects of word line voltage drop, cell conduction drift, and bit line coupling perturbation.

[0013] This invention constructs a path for identifying and reversing pseudo-convergence. Even when hard-decision surfaces are stable but the underlying support between component codes is not truly consistent, it can identify mismatch-sensitive bits and perform targeted reverse correction against mismatch branch metrics, thereby restoring effective competition between erroneous stable paths and non-optimal paths. Therefore, this invention not only avoids the premature termination problem caused by "stopping iteration when hard decisions remain unchanged" in existing technologies, but also improves the stability, accuracy, and anti-spurious convergence capability of Turbo code probability domain decoding results in application scenarios with high mismatch and easily compressible low-confidence branches. It is particularly suitable for low-power, high-reliability decoding scenarios sensitive to low-probability branch distortion under in-memory computing implementation conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] For examples, please refer to Figure 1 As shown in this embodiment, a Turbo code probabilistic domain decoding method based on a memory computing device includes: In one embodiment of this application, step S100 is used to obtain the received sequence y, the interleaving mapping π, the perforation parameter P, and the in-memory computing array mismatch parameter η, generate the original branch probabilities of each state transition, and construct the corresponding complementary mirror probability set Γ0. The specific steps are as follows: Specifically, in step S110, the position backfilling of the received sequence according to the perforation parameters refers to first reading the perforation period, perforation start position, and the order of retained and deleted bits at each time point recorded in the perforation parameters, and then scanning bit by bit according to the input order of the received sequence; for positions marked as retained bits in the perforation parameters, the current observation value in the received sequence is directly extracted; for positions marked as deleted bits in the perforation parameters, a neutral observation value is filled in at that position. The neutral observation value is preferably 0, corresponding to a state that is unbiased in both "support is true" and "support is not true". After the position backfilling is completed, the system bit observation value and the check bit observation value at the same time point are arranged into observation pairs in chronological order, wherein each observation pair contains a time position identifier, a system bit observation value, and a check bit observation value, thereby ensuring that the input data is consistent with the coding constraints in time sequence when performing probability domain mapping for each state transition.

[0018] In step S120, the probability correction model is constructed as follows: First, a preset high conduction state and a preset low conduction state are written into a calibration unit that has the same word line position and bit line position as the actual decoding write position. Reading is performed under the same read voltage, read pulse width and ambient temperature conditions as the decoding stage. The deviation between the read result and the preset conduction state is determined as the unit conduction drift amount. Then, the same conduction state is written into the beginning, middle and end of the same word line and read. The difference between the reading result at the beginning and the reading results at the middle and end is determined as the word line voltage drop amount. Then, under the condition that the target bit line is kept in normal reading, read perturbations consistent with the actual operation are applied to its adjacent bit lines in sequence. The change of the target bit line reading result before and after the perturbation is determined as the bit line coupling perturbation amount. After obtaining the cell conduction drift, word line voltage drop, and bit line coupling disturbance, these three are compared with pre-set calibration threshold ranges and converted into correction sub-values ​​within the 0-1 range in a segmented manner. The smaller the deviation, the closer the correction sub-value is to 1, indicating a weaker impact of the mismatch on the observed value; the larger the deviation, the closer the correction sub-value is to 0, indicating a stronger impact of the mismatch on the observed value. The segmentation method is preferably a three-level segmentation: a higher correction sub-value is assigned when the deviation is in the slight range; a medium correction sub-value is assigned when the deviation is in the moderate range; and a lower correction sub-value is assigned when the deviation is in the severe range. The probabilistic correction model refers to performing step-by-step corrections according to the error formation path. Specifically, firstly, the system bit observations and parity bit observations in the observation pair are subjected to a first confidence contraction using the correction sub-value corresponding to the cell conduction drift, so that the observations that are greatly affected by cell conduction drift move closer to the neutral observations, resulting in the first corrected observation pair. Secondly, the observations located at different word line positions in the first corrected observation pair are subjected to a second position correlation correction using the correction sub-value corresponding to the word line voltage drop, so that the observations far from the word line drive end move further closer to the neutral observations, resulting in the second corrected observation pair. Finally, the second corrected observation pair is subjected to a third adjacent line perturbation correction using the correction sub-value corresponding to the bit line coupling perturbation, so that the abnormal enhancement or abnormal weakening caused by the activity of adjacent bit lines is suppressed, resulting in the final corrected observation pair. The mismatch compensation factor is obtained as follows: the effective distinguishing ability retained by the final corrected observation pair relative to the original observation pair is used as the compensation result; if the observed value still deviates significantly from the neutral observed value after three levels of correction, the effective distinguishing ability of the observation pair is determined to be high, and a larger mismatch compensation factor is assigned; if the observed value has significantly shrunk to near the neutral observed value after three levels of correction, the observation pair is determined to be greatly affected by the mismatch parameters of the in-memory computing array, and a smaller mismatch compensation factor is assigned.

[0019] In step S130, the probability domain mapping of each observation pair using the mismatch compensation factor specifically involves: first, normalizing the system bit observations and parity bit observations in each observation pair according to the receiver output range; then, adjusting the amplitude of the normalized observations based on the corresponding mismatch compensation factor; when the mismatch compensation factor is small, shrinking the observations towards neutral observations to suppress false high confidence introduced by the mismatch parameters of the in-memory compute array; when the mismatch compensation factor is large, preserving the original discriminative power of the observations. Then, according to the output bit constraints corresponding to all possible state transitions at the current time, comparing the consistency between the corrected system bit observations and parity bit observations and the expected symbols of each state transition, and assigning corresponding probability values ​​through a pre-written hierarchical mapping table. The hierarchical mapping table is preferably divided into four levels: "strong consistency," "weak consistency," "weak conflict," and "strong conflict," ultimately obtaining the original branch probability corresponding to each state transition. The original branch probability is a value between 0 and 1, representing the degree to which the current observation pair supports the validity of the state transition.

[0020] In step S140, the complementary mirror probability is generated as follows: Based on the original branch probability corresponding to each state transition, the remaining portion of the original branch probability within the unit probability interval is determined as the complementary mirror probability. The original branch probability represents the degree of support for "target event is true," and the complementary mirror probability represents the degree of support for "target event is false." "Target event is true" means that the combination of information bits and check bits corresponding to the current state transition matches the observation pair after mismatch compensation. After generating the complementary mirror probability, the original branch probability and complementary mirror probability under the same time position and state transition number are associated in a fixed order, and the position identifier corresponding to the time position and the perforation parameter is also saved, forming a complementary mirror probability group. Preferably, the fixed order is original branch probability first, then complementary mirror probability, so that in subsequent step S200, adjacent memory cells on the same word line are written according to an adjacent mapping method.

[0021] In one embodiment of this application, step S200 is used to write Γ0 into the same word line according to the adjacent mapping of the original branch probability and the complementary mirror probability, read the differential response to obtain the mismatch-resistant branch metric Γ1, and generate the state mask set Ω based on the probability span of Γ1. The specific steps are as follows: In step S210, the original branch probabilities and their corresponding complementary mirror probabilities in the complementary mirror probability groups are written to adjacent memory cells on the same word line according to the same state transition number and the same time position. Specifically, the complementary mirror probability groups corresponding to each time point are extracted in chronological order, and the original branch probabilities and complementary mirror probabilities in each complementary mirror probability group are paired and addressed. The original branch probabilities are written to the first memory cell on the same word line, and the complementary mirror probabilities are written to the second memory cell physically adjacent to the first memory cell. The first and second memory cells are located on adjacent bit lines. This writing method ensures that the original branch probabilities and complementary mirror probabilities experience essentially the same word line voltage drop, local thermal disturbance, and neighboring read disturbance during subsequent readings, providing unified physical conditions for subsequent differential response extraction.

[0022] In the specific implementation of step S210, the target conductance state is first determined from a pre-established conductance state correspondence table based on the numerical values ​​of the original branch probability and the complementary mirror probability. Then, the first and second memory cells are written using the same write pulse width, the same write amplitude, and the same write order. The conductance state correspondence table is obtained through calibration, specifically by writing multiple known probability values ​​to multiple discrete conductance levels, reading the stable output results of each discrete conductance level, and establishing a mapping relationship between the known probability values ​​and the corresponding conductance levels. By using the same write conditions for paired writing, the difference between the original branch probability and the complementary mirror probability is mainly reflected as a difference in conductance state, rather than a difference in write conditions.

[0023] After writing is complete, the state transition numbers, time positions, and adjacency relationships corresponding to the first and second storage units are further recorded to form a storage mapping relationship consistent with the complementary mirror probability group. This storage mapping relationship ensures that during subsequent reads, adjacent storage units corresponding to the original branch probability and the complementary mirror probability can be accurately extracted according to the same state transition number and the same time position, avoiding mixed reads across state transition numbers or time positions.

[0024] In step S220, a forward read bias and a reverse read bias are sequentially applied to adjacent memory cells after writing, and the difference between the two read current values ​​is obtained. The difference between the read current values ​​is then used as the differential response of the state transition. Specifically, a forward read bias is first applied to the first and second memory cells, causing the read current to flow through the first and second memory cells in a first direction, and the forward read current is recorded. Subsequently, while keeping the read pulse width and read amplitude unchanged, a reverse read bias is applied to the same pair of adjacent memory cells, causing the read current to flow through the second and first memory cells in a second direction opposite to the first direction, and the reverse read current is recorded. The difference between the forward read current and the reverse read current is then determined as the differential response corresponding to the state transition.

[0025] The differential response characterizes the actual distinguishability between the original branch probability and the complementary mirror probability. Since the original branch probability and the complementary mirror probability are written to adjacent memory cells on the same word line, they exhibit high consistency in the word line voltage drop, local conduction fluctuations, and adjacent bit line perturbations experienced during the read process. Furthermore, by performing two reads in opposite directions on the same pair of adjacent memory cells, the consistency perturbations are weakened during the difference processing, while the read differences caused by the inherent differences between the original branch probability and the complementary mirror probability are retained. Therefore, the differential response can more realistically reflect the probabilistic tendency of the current state transition. A larger differential response value indicates a higher degree of distinguishability between the original branch probability and the complementary mirror probability; a smaller differential response value indicates more significant compression or aliasing between the two in the in-memory compute array.

[0026] In step S230, based on the proximity between the differential responses of each state transition and the preset reference differential interval, the original branch probabilities are recalibrated to resist mismatch, generating a mismatch-resistant branch metric. The preset reference differential interval is established through offline calibration, specifically: multiple sets of original branch probabilities and complementary mirror probabilities with known probability differences are selected and written into adjacent memory cells on the same word line as described in step S210, and then the corresponding differential responses are read as described in step S220; subsequently, based on the distribution of the differential response as the probability difference changes, the differential response is divided into a high-discrimination reference differential interval, a medium-discrimination reference differential interval, and a low-discrimination reference differential interval, thereby forming the preset reference differential interval.

[0027] After obtaining the differential response of the current state transition, the differential response is compared with a preset reference differential interval, and the original branch probability is recalibrated against mismatch based on the degree of closeness. When the differential response falls into the high-discrimination reference differential interval, the current original branch probability is determined to have high confidence, and its ability to optimize the target state transition is retained. When the differential response falls into the medium-discrimination reference differential interval, the current original branch probability is determined to be affected by array mismatch, and the original branch probability is contracted to the neutral probability direction for the first degree. When the differential response falls into the low-discrimination reference differential interval, it is determined that there has been significant compression between the current original branch probability and the complementary mirror probability, and the original branch probability is contracted to the neutral probability direction for the second degree, and the magnitude of the second degree of contraction is greater than that of the first degree of contraction. The neutral probability refers to the probability position corresponding to the equilibrium state of the probability of the target event being true and the probability of the target event not being true; the contraction to the neutral probability direction means reducing the deviation of the original branch probability from the equilibrium state to avoid the direct retention of artificially high confidence caused by array mismatch. The value obtained after the above processing is the mismatch-resistant branch metric for the corresponding state transition. The mismatch-resistant branch metric is a probability metric that is re-determined based on the original branch probability and combined with the actual distinguishing ability of the differential response. It retains the ability of the original branch probability to represent the optimal relationship of the state transition, while suppressing the amplification effect of mismatch of the in-memory computing array on the original branch probability.

[0028] In step S240, all anti-mismatch branch metrics at the same time are sorted, the probability span between the maximum and second-largest values ​​is calculated, and the corresponding state transitions with a probability span less than a preset masking threshold are added to the state masking set. Specifically, at each time position, all anti-mismatch branch metrics corresponding to state transitions are arranged in descending order of value, the anti-mismatch branch metric with the largest value and the second-largest value are determined, and the difference between the two is defined as the probability span. The probability span is used to characterize the distinguishability between the optimal state transition and the second-best state transition at that time position. The preset masking threshold is obtained through offline sample statistics. Specifically, under different signal-to-noise conditions and different array mismatch levels, the difference range of anti-mismatch branch metrics corresponding to the optimal state transition and the second-best state transition in the correctly decoded samples and the incorrectly decoded samples are statistically analyzed, and the boundary value that can distinguish between stable decision states and easily aliased states is selected from the boundary interval between the two as the preset masking threshold. When the probability span at a certain time position is less than the preset shielding threshold, it indicates that there is insufficient distinction between the optimal state transition and the suboptimal state transition at that time position, and there is a risk of enhanced local competition and expansion of pseudo-optimal paths.

[0029] When the probability span is less than a preset masking threshold, state transitions other than the first ranked state, whose difference in anti-mismatch branch metric from the first ranked state is less than the preset masking threshold, are added to the state masking set. The state masking set is used to mark state transitions with a high risk of aliasing at the current time position. State transitions added to the state masking set are not directly deleted, but their participation intensity is limited in subsequent forward and backward recursion processes, preventing them from continuously amplifying with high confidence, thereby suppressing the erroneous propagation of local competitive states to the globally preferred path.

[0030] In one embodiment of this application, step S300 is used to perform forward and backward recursion of the first component code based on Γ1 and Ω to obtain the first external information probability Q1 and the state entropy sequence H1. The specific steps are as follows: In step S310, the anti-mismatch branch metrics for each time position are screened according to the state masking set to determine open state transitions allowed to participate in the recursion and masked state transitions restricted from participating in the recursion. Specifically, within each time position, the anti-mismatch branch metrics corresponding to all state transitions at that time position are read one by one, and state transitions in the state masking set are marked as masked state transitions, while state transitions not in the state masking set are marked as open state transitions. The open state transitions are used to represent state transitions with a high degree of distinguishability at the current time position and that can directly participate in path competition; the masked state transitions are used to represent state transitions with a high risk of aliasing at the current time position and that are not suitable for participating in the recursion with full strength. Through this screening process, a state transition participation table corresponding one-to-one with time positions can be formed before the recursion begins.

[0031] In the specific implementation of step S310, to ensure the continuity of the subsequent recursive process, the shielded state transition is not directly deleted, but rather granted limited participation eligibility. Limited participation eligibility means that the shielded state transition retains its state connection relationship, but its contribution to state probability propagation is weaker than that of the open state transition. This approach is adopted because completely eliminating the shielded state transition might cut off the true path at local moments; allowing the shielded state transition to participate with its original strength might lead to excessive amplification of local aliasing paths. Therefore, by distinguishing between open and shielded state transitions, pseudo-optimal path expansion can be suppressed while maintaining path continuity.

[0032] In step S320, the anti-mismatch branch metric of each open state transition is jointly accumulated with the forward state probability of the previous time position in forward chronological order. Then, restricted accumulation is performed on each shielded state transition using a weighted retention method to obtain the first forward state probability sequence. Specifically, at the initial time position, the initial forward state probability is first set according to the starting state constraint of the first component code, where the initial allowed state is given a deterministic initial probability, and the other states are given a suppressive initial probability. Subsequently, processing is performed position by position in chronological order starting from the forward time start point. Within each time position, for all open state transitions, the forward state probability connected to the starting state of the open state transition in the previous time position is jointly accumulated with the anti-mismatch branch metric corresponding to the open state transition to form the candidate forward probability of the open state transition pointing to the termination state.

[0033] For cases where the same termination state may be reached through multiple open state transitions, the forward probabilities of each candidate are compared and merged, and the accumulated result with the highest probability contribution is retained as the forward state probability of the open path for that termination state. For masked state transitions, they do not directly participate in the competition as open state transitions, but instead undergo restricted accumulation using a weighted retention method. The weighted retention method refers to: first, further shrinking the anti-mismatch branch metric corresponding to the masked state transition towards the neutral probability direction, and then jointly accumulating the shrunken result with the forward state probability of the previous time position, so that the masked state transition only retains the path connectivity effect without excessively affecting the optimal result of the current termination state.

[0034] In step S320, the first forward state probability sequence refers to the set of state probabilities obtained position by position along the forward time direction. Each time position corresponds to a forward state probability set, and each forward state probability set contains the forward state probabilities of all states of the first component code at the current time position. To avoid excessive concentration of forward state probabilities during long sequence recursion, after completing open path accumulation and restricted path accumulation at each time position, normalization processing is performed on all forward state probabilities at the current time position to redistribute them into a unified probability interval and maintain comparability between different time positions.

[0035] In step S330, following the reverse chronological order, the anti-mismatch branch metric of each open state transition is jointly accumulated with the backward state probability of the next time position, and the restricted accumulation is performed on each shielded state transition according to the weighted retention method to obtain the first backward state probability sequence. Specifically, at the termination time position, the initial backward state probability is first set according to the termination state constraint of the first component code, wherein the termination allowed state is given a deterministic initial probability, and the other states are given a suppressive initial probability; then, starting from the backward time start point, it is processed position by position in reverse chronological order. Within each time position, for all open state transitions, the backward state probability connected to the termination state of the open state transition in the next time position is jointly accumulated with the anti-mismatch branch metric corresponding to the open state transition to form the candidate backward probability of the open state transition back to the starting state.

[0036] For cases where multiple open state transitions converge to the same initial state, the candidate backward probabilities are compared and merged, and the accumulated result with the highest probability contribution is retained as the open path backward state probability of that initial state. For masked state transitions, a weighted retention method is also used to perform restricted accumulation, ensuring that it only maintains backtracking connectivity and does not dominate backward probability propagation with high confidence. After completing the backward accumulation at each time point, all backward state probabilities at the current time point are normalized to obtain a first backward state probability sequence with the same dimensions as the forward state probabilities.

[0037] In step S340, a first external information probability is generated based on the first forward state probability sequence, the first backward state probability sequence, and the anti-mismatch branch metric at each time position. A state entropy sequence is also generated based on the probability dispersion of each state transition within the same time position. Specifically, within each time position, a set of state transitions corresponding to the validity of the information bit value and a set of state transitions corresponding to the invalidity of the information bit value are extracted. For the former type of state transition, the forward state probability corresponding to its initial state, the anti-mismatch branch metric corresponding to the state transition, and the backward state probability corresponding to its final state are jointly converged to obtain the valid path probability. For the latter type of state transition, the invalid path probability is obtained in the same way. Subsequently, the first external information probability at that time position is determined based on the relative magnitude between the valid path probability and the invalid path probability. The first external information probability refers to the additional probability information provided by the state recursion result for the current information bit after removing the influence of the known fixed constraints of the current component code itself. Because its input comes from a mismatch-resistant branch metric that has been recalibrated for mismatch resistance, and because a state masking set constraint is introduced during the recursion process, the first extrinsic information probability can better reflect the true path competition relationship and is less susceptible to amplification by local aliasing states. The state entropy sequence is used to characterize the dispersion of state competition at each time point. Specifically, at each time point, the joint path probability distribution corresponding to all state transitions is statistically analyzed; if the probability is mainly concentrated in a few state transitions, the state uncertainty at that time point is determined to be low, corresponding to a low state entropy; if the probability is dispersed among multiple state transitions, the state uncertainty at that time point is determined to be high, corresponding to a high state entropy. Arranging the state entropies at each time point in chronological order yields the state entropy sequence. This state entropy sequence is used in subsequent steps to identify local anomalous collapse phenomena during pseudo-convergence.

[0038] In one embodiment of this application, step S400 is used to interleave Q1 by π and combine it with H1 to form a position confidence weight W. The updated branch metric Γ2 is obtained by recalibrating Γ1 using W. The specific steps are as follows: In step S410, the first external information probability Q1 is rearranged according to the interleaving map π to obtain the interleaved external information sequence corresponding to the time position of the second component code. Specifically, according to the predefined correspondence between input and output positions in the interleaving map π, the probability values ​​of the first external information probability Q1 at each time position of the first component code are read one by one and moved to the corresponding time positions of the second component code specified by the interleaving map π, forming an interleaved external information sequence arranged in the order of the second component code. Each position in the interleaved external information sequence inherits the probability tendency of the first external information probability Q1 for the corresponding information bit to be true or false, but its position index has been changed from the time order of the first component code to the time order of the second component code.

[0039] In the specific implementation of step S410, to ensure the consistency of the meaning of information bits before and after interleaving, a mapping record between the original time position of the first external information probability Q1 and the time position after interleaving is synchronously saved during position rearrangement. This mapping record is used to subsequently convert the time position attribute corresponding to the state entropy sequence H1 to the time position of the second component code, preventing the state entropy sequence H1 from remaining at the first component code time position after the first external information probability Q1 has completed interleaving, thus avoiding distortion of the position correspondence. Through the above processing, a unified positional basis is established between the interleaved external information sequence and the subsequent trust grading process.

[0040] In step S420, based on the entropy values ​​at each time position in the state entropy sequence H1 and the continuity of entropy changes between adjacent time positions, the credibility of each time position in the interleaved external information sequence is classified to obtain a position credibility level sequence. Specifically, firstly, according to the mapping record saved in step S410, the state entropy corresponding to each time position in the state entropy sequence H1 is moved to the time position corresponding to the second component code to form an interleaved entropy sequence; then, the entropy value and local change trend of the interleaved entropy sequence are analyzed position by position. If the state entropy at a certain time position is low, and the state entropy change between it and the previous and next time positions is gradual, it is determined that the state competition corresponding to that time position has converged sufficiently and is given a high credibility level; if the state entropy at a certain time position is at an intermediate level, and the change amplitude between adjacent time positions is limited, it is given a medium credibility level; if the state entropy at a certain time position suddenly decreases or increases significantly, and there is a discontinuous jump between it and adjacent time positions, it is determined that there is a local abnormal convergence or a sudden increase in local competition at that time position and is given a low credibility level.

[0041] The entropy value is used to characterize the dispersion of the state competition at the current time position, and the entropy change continuity is used to characterize the stability of the evolution between the current time position and adjacent time positions. The reason for using both entropy value and entropy change continuity for credibility classification is that judging solely based on the level of state entropy may mistake local abnormal collapse for stable convergence; however, by simultaneously introducing the change continuity between adjacent time positions, abnormal entropy valleys caused by local pseudo-convergence can be identified. The resulting location credibility level sequence can more accurately distinguish between true stable locations and locally distorted locations.

[0042] In the specific implementation of step S420, the location confidence level sequence is preferably divided into three levels: high confidence level, medium confidence level, and low confidence level. The high confidence level indicates that the probability Q1 of the first external information formed by the first component code at that time position has high reference value; the medium confidence level indicates that the probability Q1 of the first external information can be adopted but should not be directly strengthened; and the low confidence level indicates that the probability Q1 of the first external information may be affected by local pseudo-convergence at that time position. Dividing the location confidence level sequence into discrete levels is beneficial for subsequent stable jointing with the interleaved external information sequence and avoids additional quantization inconsistencies when continuous quantities are used within the in-memory computing array.

[0043] In step S430, the location confidence level sequence and the interleaved external information sequence are jointly constrained at the same time position to generate location confidence weights W corresponding to each time position. Specifically, for each time position, the probability tendency of the interleaved external information sequence corresponding to that time position is first read, and then the location confidence level corresponding to the same time position is read. When the time position is assigned a high confidence level, the original probability tendency of the interleaved external information sequence is retained, and a larger location confidence weight W is generated. When the time position is assigned a medium confidence level, the original probability tendency of the interleaved external information sequence is moderately suppressed, and a medium-sized location confidence weight W is generated. When the time position is assigned a low confidence level, the original probability tendency of the interleaved external information sequence is significantly suppressed, and a smaller location confidence weight W is generated. The location confidence weight W is used to characterize the adjustment strength of the interleaved external information on the second component code branch metric at the corresponding time position. The larger the positional credibility weight W, the more suitable the first external information probability Q1 at that time position is to directly participate in the branch metric correction of the second component code after interleaving; the smaller the positional credibility weight W, the more suitable it is for the first external information probability Q1 at that time position to participate in the branch metric correction of the second component code with high intensity, even though it can retain directional reference. By jointly generating the positional credibility weight W from the positional credibility level sequence and the interleaved external information sequence, the second component code can adopt a differentiated absorption method for different time positions when receiving the recursive result of the first component code.

[0044] In step S440, position-related recalibration is performed using the position confidence weight W to counteract the mismatch branch metric Γ1. This ensures that the mismatch branch metric Γ1 at high-confidence time positions maintains its distinguishing advantage, while the mismatch branch metric Γ1 at low-confidence time positions shrinks towards the neutral probability direction, generating an updated branch metric Γ2. Specifically, within each time position of the second component code, the position confidence weight W and the mismatch branch metric Γ1 corresponding to that time position are read one by one. When the position confidence weight W is large, the relative magnitude relationship of the mismatch branch metric Γ1 with each state transition is maintained, and the state transitions supporting consistency are enhanced and maintained according to the probability direction corresponding to the interleaved external information sequence. When the position confidence weight W is at an intermediate level, the relative magnitude relationship of the mismatch branch metric Γ1 is retained, but the enhancement magnitude is weakened, so that it only reflects a mild correction. When the position confidence weight W is small, the mismatch branch metric Γ1 is shrunk towards the neutral probability direction, and the direct amplification effect of the interleaved external information sequence on it is suppressed. The state transition that supports consistency refers to a state transition in which the direction of the information bit value corresponding to the state transition is consistent with the probability tendency of the interleaved external information sequence at the current time position. The contraction towards the neutral probability direction refers to reducing the degree of difference between the optimal and suboptimal state transitions in the mismatch-resistant branch metric Γ1, so that the local bias introduced by low-confidence external information cannot be pre-fixed before the second component code is recursively applied. Through the above position-related recalibration, the updated branch metric Γ2 can retain the ability of the mismatch-resistant branch metric Γ1 to represent the real path competition relationship, and also absorb the external information constraints from the first component code that have been filtered for credibility.

[0045] In the specific implementation of step S440, the updated branch metric Γ2 is output sequentially according to its time position and maintains a one-to-one correspondence with the state transition number of the second component code. The updated branch metric Γ2 serves as the direct input for the forward and backward recursion of the second component code in the subsequent step S500. Since the updated branch metric Γ2 has undergone interleaved extrinsic information guidance and positional confidence weight W adjustment, the second component code can preferentially utilize the extrinsic information constraints of high-confidence positions during the recursion process, while suppressing the propagation of pseudo-convergence information from low-confidence positions.

[0046] In one embodiment of this application, step S500 performs forward and backward recursion of the second component code based on Γ2 and Ω to obtain the second external information probability Q2. The specific steps are as follows: In step S510, the state transitions at each time position of the second component code are screened based on the updated branch metric Γ2 and the state mask set Ω to determine the second open state transitions and the second masked state transitions. Specifically, within each time position of the second component code, the updated branch metric Γ2 corresponding to all state transitions at that time position is read one by one, and it is checked whether the corresponding state transition is in the state mask set Ω; state transitions in the state mask set Ω are marked as second masked state transitions, and state transitions not in the state mask set Ω are marked as second open state transitions. The second open state transition is used to represent a state transition that can participate in the path competition of the second component code with full strength at the current time position, and the second masked state transition is used to represent a state transition that only retains limited connectivity at the current time position and is not suitable for dominating probability propagation.

[0047] In the specific implementation of step S510, the second open state transition and the second masked state transition still cover all state connection relationships of the second component code. That is, the second masked state transition is not directly deleted, but participates in the accumulation in the subsequent recursion process using a weighted retention method. The reason for adopting this processing method is that the recursion of the second component code still needs to maintain complete state reachability. If the second masked state transition is completely removed, the real path may be accidentally truncated at a local moment; if the second masked state transition is allowed to participate completely without distinction, the local competition results of low-confidence positions may be amplified again in the second component code.

[0048] In step S520, following the forward chronological order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second forward state probability of the previous time position, and a restricted accumulation is performed on each second shielded state transition using a weighted retention method to obtain a second forward state probability sequence. Specifically, at the starting time position of the second component code, the initial second forward state probability is first set according to the starting state constraint of the second component code, where the initial allowed state is given a deterministic initial probability, and the other states are given a suppressive initial probability; then, it is processed position by position in the forward chronological order. Within each time position, for all second open state transitions, the second forward state probability connected to the starting state of the second open state transition in the previous time position is jointly accumulated with the update branch metric Γ2 corresponding to the second open state transition to form a candidate second forward probability pointing to the termination state of the second open state transition.

[0049] When the same termination state is reached by multiple second open states, the multiple candidate second forward probabilities are compared and merged, and the accumulated result with the highest probability contribution is retained as the open path second forward state probability of the termination state. For second shielded state transitions, the corresponding update branch metric Γ2 is first further shrunk towards the neutral probability direction, and then the shrunk result is jointly accumulated with the second forward state probability of the previous time position to obtain the restricted path second forward probability. Through this processing method, the second shielded state transition only retains the path connectivity effect, and cannot affect the optimal result of the termination state with the same strength as the second open state transition. After completing the accumulation at the current time position, all second forward state probabilities are normalized to form the second forward state probability set at the current time position. The second forward state probability sequence refers to the set of second component code state probabilities obtained position by position along the forward time direction. Each time position corresponds to a second forward state probability set, and each second forward state probability set contains the second forward state probabilities of all states of the second component code at the current time position. By normalizing the time position, we can avoid the second forward state probability from becoming overly concentrated or overly decayed during the long sequence recursion, thus ensuring that the basis for comparison between different time positions is consistent.

[0050] In step S530, following the reverse chronological order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second backward state probability at the next time position. Then, a restricted accumulation is performed on each second masked state transition according to the aforementioned weighted retention method to obtain a second backward state probability sequence. Specifically, at the termination time position of the second component code, an initial second backward state probability is first set based on the termination state constraint of the second component code, where a deterministic initial probability is assigned to the termination allowed state, and a suppressive initial probability is assigned to the other states. Subsequently, processing is performed position by position in the reverse chronological order. Within each time position, for all second open state transitions, the second backward state probability connected to the termination state of the second open state transition at the next time position is jointly accumulated with the update branch metric Γ2 corresponding to the second open state transition to form a candidate second backward probability for the second open state transition to trace back to the starting state.

[0051] When multiple second open states converge to form a single initial state, the candidate second backward probabilities are compared and merged, and the accumulated result with the highest probability contribution is retained as the second backward state probability of the open path for that initial state. For second shielded state transitions, a weighted retention method is also used to perform restricted accumulation, ensuring that backtracking connectivity is maintained while preventing high-confidence backpropagation from dominating backward probability propagation. After completing the backward accumulation at each time point, all second backward state probabilities are normalized to obtain a second backward state probability sequence with a consistent scale with the second forward state probabilities.

[0052] In step S540, a second external information probability Q2 is generated based on the second forward state probability sequence, the second backward state probability sequence, and the update branch metric Γ2 at each time position. Specifically, within each time position of the second component code, the state transition set corresponding to the information bit value being true and the state transition set corresponding to the information bit value being false are extracted respectively. For the former type of state transition, the second forward state probability corresponding to its starting state, the update branch metric Γ2 corresponding to the state transition, and the second backward state probability corresponding to its ending state are jointly converged to obtain the success path probability. For the latter type of state transition, the failure path probability is obtained in the same way. Then, based on the relative magnitude between the success path probability and the failure path probability, the second external information probability Q2 at that time position is determined. The second external information probability Q2 refers to the additional probability information formed by the second forward state probability sequence, the second backward state probability sequence, and the update branch metric Γ2 under the constraint of the second component code. Since the updated branch metric Γ2 has introduced the interleaved extrinsic information constraint of the first component code and has been adjusted by the position confidence weight W, the second extrinsic information probability Q2 not only includes the path competition result within the second component code, but also carries the confidence auxiliary information from the first component code.

[0053] In one embodiment of this application, step S600 is used to de-interleave Q2 to obtain the posterior probability R, and construct a pseudo-convergence index Z based on the difference between R and Q1, the hard decision stability of adjacent iterations, and the entropy collapse rate of H1. The specific steps are as follows: In step S610, the second external information probability Q2 is restored to its position according to the inverse mapping of the interleaving map π, resulting in a backfilled external information sequence corresponding to the time position of the first component code. A posterior probability R is then generated based on this backfilled external information sequence. Specifically, according to the inverse mapping of the interleaving map π, the probability values ​​of the second external information probability Q2 at each time position of the second component code are read one by one and restored to the original time position corresponding to the first component code, forming a backfilled external information sequence arranged in the time order of the first component code. Subsequently, within each original time position, the information bit establishment tendency represented by the backfilled external information sequence is jointly compared with the existing path support direction at that original time position to obtain a comprehensive probability result of whether the information bit is established or not at that time position, and this comprehensive probability result is determined as the posterior probability R. The posterior probability R refers to the probability result corresponding to the final establishment tendency of the information bit at each time position after fusing the recursive information of the first and second component codes within the current iteration round. The more the posterior probability R leans towards the true direction, the more likely the information bit at that time position is to take the true value; the more the posterior probability R leans towards the false direction, the more likely the information bit at that time position is to take the false value; the closer the posterior probability R is to the neutral probability, the more likely there is still significant competition for the information bit at that time position. The posterior probability R serves as the comprehensive judgment basis for this iteration and is used for subsequent difference analysis and hard decision stability analysis.

[0054] In step S620, the probability offset between the posterior probability R and the first external information probability Q1 at the same time position is compared to obtain a difference sequence for each time position. Specifically, within each time position, the posterior probability R and the first external information probability Q1 are read respectively, and their support strength for the direction of information bit establishment and their deviation from the neutral probability are compared. If the posterior probability R and the first external information probability Q1 are consistent in direction and their deviation is similar, the difference at that time position is determined to be small. If they are reversed in direction, or although they are consistent in direction, their deviation differs significantly, the difference at that time position is determined to be large. The comparison results at each time point are arranged in order of time position to form a difference sequence. The difference sequence is used to characterize the consistency between the output result of the second component code and the original external information of the first component code after being repositioned to the time position of the first component code. If, even when the hard decision has become stable, there are still multiple time positions with large differences in the difference sequence, it indicates that the support basis of the first component code and the second component code for the same information bit is not truly unified, and the current iteration may have only formed a superficially stable rather than truly consistent convergence state.

[0055] In step S630, hard decisions are made on the posterior probabilities R of the current iteration and the previous iteration, and the consistency of decisions at the same time position is compared to obtain the hard decision stability of adjacent iterations. Specifically, within the current iteration, according to the positional relationship between the posterior probability R and the neutral probability, a binary judgment of "true" or "false" is made on the information bits at each time position to obtain the hard decision result of the current iteration; at the same time, the hard decision results corresponding to the posterior probabilities saved in the previous iteration are read and compared one by one at the same time position. If a certain time position maintains the same decision in two consecutive iterations, the time position is counted as a stable position; if a certain time position experiences a decision reversal in two consecutive iterations, the time position is counted as an unstable position. The proportion of all stable positions in the total number of time positions is determined as the hard decision stability of adjacent iterations. The hard decision stability of adjacent iterations is used to characterize the surface stability of the current decoding process at the discrete decision level. A higher hard-decision stability between adjacent iterations indicates that the binary decision of the current iteration is closer to that of the previous iteration; a lower hard-decision stability between adjacent iterations indicates that the current decoding result is still fluctuating within a large range. However, in the scenario of local pseudo-convergence caused by in-memory compute array mismatch, the hard-decision stability between adjacent iterations may be high, but the probability support basis at some time points is still inconsistent. Therefore, the hard-decision stability between adjacent iterations cannot be used alone to determine true convergence, but must be used in conjunction with the difference sequence and entropy collapse rate.

[0056] In step S640, the entropy change trend of each time position in the state entropy sequence H1 is read to determine the entropy collapse rate of each time position, and a pseudo-convergence index Z is constructed based on the difference sequence, the hard decision stability of adjacent iterations, and the entropy collapse rate. Specifically, the entropy change of the state entropy sequence H1 at the corresponding time positions of the current iteration round and the previous iteration round is read one by one; if the state entropy at a certain time position decreases rapidly in two adjacent iterations, and the decrease is significantly greater than the average change level of adjacent time positions, it is determined that there is a high entropy collapse rate at that time position; if the state entropy at a certain time position changes slowly, it is determined that the entropy collapse rate at that time position is low. The entropy collapse rate is used to characterize the speed of contraction when state competition moves from dispersion to concentration, especially to identify abnormally rapid concentration phenomena caused by local aliasing.

[0057] In the specific implementation of step S640, when constructing the pseudo-convergence index Z, the number and distribution density of time positions with large differences in the difference sequence are used as the first criterion, the stability of hard decisions in adjacent iterations is used as the second criterion, and the number and distribution continuity of time positions with high entropy collapse rates are used as the third criterion. When the stability of hard decisions in adjacent iterations is high, but there are still many positions with large differences in the difference sequence, and these positions are accompanied by high entropy collapse rates, the current iteration is determined to have strong pseudo-convergence characteristics, and the pseudo-convergence index Z is increased accordingly. When the stability of hard decisions in adjacent iterations is high and the overall difference sequence is small, and the distribution of entropy collapse rates is gentle, the current iteration is determined to be closer to true convergence, and the pseudo-convergence index Z is decreased accordingly. The pseudo-convergence index Z is used to characterize the degree of deviation between "surface stability" and "true consistency" in the current iteration. The larger the value of the pseudo-convergence index Z, the more likely the current iteration is to be a pseudo-convergence formed under the dominance of local aliasing; the smaller the value of the pseudo-convergence index Z, the closer the current iteration is to true convergence. By simultaneously introducing the difference between the posterior probability R and the first external information probability Q1, the stability of hard decisions in adjacent iterations, and the entropy collapse rate of the state entropy sequence H1, the risk of misjudgment caused by prematurely terminating the iteration based solely on the hard decision no longer changing can be avoided.

[0058] In one embodiment of this application, step S700 is used to extract the mismatch sensitive bit in R and reverse correct Γ1 when Z is greater than a preset threshold, then return to S300 to continue iteration; when Z is less than or equal to the preset threshold, the decoding result is output according to R, and the specific steps are as follows: In step S710, when the pseudo-convergence index Z is greater than a preset threshold, mismatch sensitive bits that coincide with high dissimilarity positions and high entropy collapse rate positions are extracted from the posterior probability R to form a set of mismatch sensitive bits. Specifically, the dissimilarity sequence obtained in step S620 is read first, and the time positions where the dissimilarity is higher than the preset dissimilarity threshold are marked as high dissimilarity positions; then the entropy collapse rate distribution determined in step S640 is read, and the time positions where the entropy collapse rate is higher than the preset collapse threshold are marked as high entropy collapse rate positions; subsequently, the high dissimilarity positions and high entropy collapse rate positions are cross-compared, and the time positions that simultaneously meet both conditions are extracted as suspected pseudo-convergence positions. Then, the probability bias direction and deviation from the neutral probability of the posterior probability R at the suspected pseudo-convergence positions are read, and the information bits that show strong bias at the positions but whose support basis of the preceding and following component codes is inconsistent are determined as mismatch sensitive bits, and a set of mismatch sensitive bits is formed in time position order. The mismatch-sensitive bits refer to information bits that exhibit a strong posterior probability bias in the current iteration, but this bias is accompanied by a large component-to-component code difference and a rapid state entropy collapse rate. These information bits are typically not formed by the full convergence of the true path competition, but rather by local error stabilization points resulting from the combined effects of in-memory compute array mismatch, local state aliasing, and interleaving propagation. By filtering out mismatch-sensitive bits from all time positions, simultaneous intervention on all branch metrics can be avoided during subsequent backpropagation, thus ensuring that the correction range is concentrated on high-risk pseudo-convergence locations.

[0059] In the specific implementation of step S710, to improve the targeting of mismatch-sensitive bit localization, the continuous distribution of the suspected pseudo-convergence positions in adjacent time positions can be further checked. When multiple suspected pseudo-convergence positions appear consecutively in time, they are treated as mismatch-sensitive segments. When the suspected pseudo-convergence positions are discretely distributed, only the information bits corresponding to a single time position are treated as mismatch-sensitive bits. The reason for adopting this method is that pseudo-convergence caused by local disturbances in the in-memory computing array sometimes manifests as single-point anomalies and sometimes as continuous segment anomalies. By distinguishing between single-point and segment anomalies, subsequent reverse correction can better conform to the actual error propagation pattern.

[0060] In step S720, based on the distribution of the mismatch-sensitive bit set at each time position, a reverse correction is performed on the state transitions in the anti-mismatch branch metric Γ1 corresponding to the mismatch-sensitive bit set, weakening the probability bias of the current preferred direction and restoring the competitive state. Specifically, within the time position corresponding to each mismatch-sensitive bit, state transitions in the anti-mismatch branch metric Γ1 that are consistent with the bias direction of the current posterior probability R, and state transitions that are opposite to the bias direction of the current posterior probability R, are first read. Subsequently, a suppression process is performed on the anti-mismatch branch metric Γ1 of the former type of state transition, reducing its deviation from the neutral probability. A recovery process is performed on the anti-mismatch branch metric Γ1 of the latter type of state transition, causing it to rise back to the competitive visible state without exceeding the original anti-mismatch recalibration range. Through the above reverse correction, non-preferred paths that have been suppressed by local error stabilization regain the opportunity to participate in the competition.

[0061] The reverse correction is not a simple uniform reduction or increase of all state transitions, but rather a targeted processing based on the directional bias corresponding to the mismatch-sensitive bit set. Weakening the probability bias of the current preferred direction means reducing the deviation of the anti-mismatch branch metric Γ1 relative to the neutral probability for state transitions aligned with the current posterior probability R. Restoring the competitive state means appropriately increasing the comparability of the anti-mismatch branch metric Γ1 at the same time position for state transitions opposite to the current posterior probability R, enabling them to re-establish effective competition in subsequent forward and backward iterations. This approach avoids the reverse correction becoming an undirected overall smoothing, thus ensuring that the correction action is specifically targeted at the bias structure formed by pseudo-convergence.

[0062] In the specific implementation of step S720, when the mismatch-sensitive bit belongs to a mismatch-sensitive segment, the reverse correction intensity for each time position within the mismatch-sensitive segment is allocated in a manner that the intensity is high at the center of the segment and low at the edge of the segment; when the mismatch-sensitive bit is a discrete single point, only the corresponding single time position is subject to local reverse correction. This processing method can more fully disrupt continuous segment-type pseudo-convergence while avoiding excessive disturbance to adjacent normal time positions. The anti-mismatch branch metric Γ1 after reverse correction still maintains a one-to-one correspondence with the state transition number and time position, serving as the input basis for re-entering the next round of first component code recursion.

[0063] In step S730, the state mask set Ω for the corresponding time position is regenerated based on the reverse-corrected anti-mismatch branch metric Γ1, and the process returns to step S300 to continue iteration. Specifically, for time positions where reverse correction has been performed, all anti-mismatch branch metrics are reordered according to the method described in step S240, the probability span between the maximum and second-largest values ​​is recalculated, and the state mask set Ω corresponding to that time position is updated according to the preset masking threshold; for time positions where reverse correction has not been performed, the original state mask set Ω is retained unchanged. After the update is completed, the corrected anti-mismatch branch metric Γ1 and the updated state mask set Ω are re-inputted into step S300, and the forward and backward recursion of the first component code continues.

[0064] By synchronously updating the state mask set Ω after the reverse correction, the local competition state upon which subsequent iterations are based can be kept consistent with the corrected mismatch branch metric Γ1. If only the mismatch branch metric Γ1 is corrected without updating the state mask set Ω, the original state mask set Ω may still retain the competition constraint relationship from the previous round of pseudo-convergence, thus weakening the effect of the reverse correction. Therefore, performing the reverse correction in conjunction with the update of the state mask set Ω ensures that subsequent iterations are restarted on the basis of the new path competition.

[0065] In step S740, when the pseudo-convergence index Z is less than or equal to a preset threshold, a final hard decision is made on the information bits at each time position based on the posterior probability R, and the decoding result is output. Specifically, within each time position, the positional relationship between the posterior probability R and the neutral probability is compared; when the posterior probability R is biased towards the direction of the information bit being true, the information bit at that time position is determined to have a true value; when the posterior probability R is biased towards the direction of the information bit being false, the information bit at that time position is determined to have a false value; subsequently, the decision results of all time positions are arranged in chronological order to form the final decoded bit sequence, and the final decoded bit sequence is output as the decoding result.

[0066] The preset threshold is used to distinguish whether the current iteration round belongs to an acceptable convergence state or a pseudo-convergence state. The preset threshold is obtained through offline sample statistics, specifically: the distribution intervals of the pseudo-convergence index Z corresponding to the true convergence samples and pseudo-convergence samples are statistically analyzed, and a boundary value that can balance false judgment control and detection capability is selected from the boundary region of the two types of sample distributions as the preset threshold. When the pseudo-convergence index Z is less than or equal to the preset threshold, it indicates that the inter-component code differences, hard decision stability, and state entropy changes in the current iteration round have reached a relatively consistent state, and further iteration has limited improvement on the result; therefore, the decoding result can be directly output.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A Turbo code probabilistic domain decoding method based on in-memory computing, characterized in that: include: S100: Obtain the received sequence y, interleaving mapping π, perforation parameter P and in-memory computing array mismatch parameter η, generate the original branch probability of each state transition, and construct the corresponding complementary mirror probability group Γ0. S200, Γ0 is written into the same word line according to the adjacent mapping of the original branch probability and the complementary mirror probability, the differential response is read to obtain the anti-mismatch branch metric Γ1, and the state mask set Ω is generated according to the probability span of Γ1. S300, based on Γ1 and Ω, perform forward and backward recursion of the first component code to obtain the first external information probability Q1 and the state entropy sequence H1; S400, interweave Q1 by π and combine it with H1 to form a positional confidence weight W, and use W to recalibrate Γ1 to obtain the updated branch metric Γ2; S500, based on Γ2 and Ω, perform forward and backward recursion of the second component code to obtain the second external information probability Q2; S600, the posterior probability R is obtained by deinterleaving Q2, and the pseudo-convergence index Z is constructed based on the difference between R and Q1, the hard decision stability of adjacent iterations and the entropy collapse rate of H1. S700: When Z is greater than the preset threshold, extract the mismatch sensitive bit in R, reverse correct Γ1, and return to S300 to continue iterating; when Z is less than or equal to the preset threshold, output the decoding result according to R.

2. The Turbo code probability domain decoding method based on a memory computing device according to claim 1, characterized in that: S100 includes: S110, fill the position back into the received sequence according to the perforation parameters, and reassemble the system bit observation value and the check bit observation value into a time-series corresponding observation pair; S120, Based on the mismatch parameters of the in-memory computing array, establish a probabilistic correction model including cell conduction drift, word line voltage drop and bit line coupling disturbance, and generate the mismatch compensation factor corresponding to each observation pair column from the probabilistic correction model. S130, using the mismatch compensation factor to perform probability domain mapping on each observation pair column, and obtaining the original branch probability corresponding to each state transition in the code trellis diagram; S140, based on each original branch probability, generate complementary mirror probabilities according to the complementary constraints of the target event being true and the target event not being true under the same state transition, and associate the original branch probabilities with the corresponding complementary mirror probabilities according to the state transition number to form a complementary mirror probability group Γ0.

3. The Turbo code probability domain decoding method based on a memory computing device according to claim 1, characterized in that: S200 includes: S210, Write the original branch probabilities and corresponding complementary mirror probabilities in the complementary mirror probability group into adjacent memory cells on the same word line according to the correspondence of the same state transition number and the same time position. S220, apply a forward read bias and a reverse read bias to the adjacent memory cells after writing, obtain the difference between the two read current values, and correspond the read current difference value to the differential response of the state transition; S230, based on the degree of closeness between the differential response of each state transition and the preset reference differential partition, the original branch probability is recalibrated against mismatch, and the anti-mismatch branch metric Γ1 is generated; S240: Sort all anti-mismatch branch metrics at the same time, calculate the probability span between the maximum and the second largest values, and add the corresponding state transitions with probability spans less than the preset shielding threshold to the state shielding set Ω.

4. The Turbo code probability domain decoding method based on a memory-based computing device according to claim 1, characterized in that: The S300 includes: S310, based on the state masking set, the anti-mismatch branch metric at each time position is screened to determine the open state transitions that are allowed to participate in the recursion and the masked state transitions that are restricted from participating in the recursion. S320, in forward time order, the mismatch-resistant branch metric of each open state transition is jointly accumulated with the forward state probability of the previous time position, and the restricted accumulation of each shielded state transition is performed in a weighted retention manner to obtain the first forward state probability sequence. S330, in reverse chronological order, the anti-mismatch branch metric of each open state transition is jointly accumulated with the backward state probability of the next time position, and the restricted accumulation is performed on each shielded state transition according to the weighted retention method to obtain the first backward state probability sequence. S340, based on the first forward state probability sequence, the first backward state probability sequence and the anti-mismatch branch metric at each time position, generate the first external information probability Q1, and generate the state entropy sequence H1 based on the probability dispersion of each state transition within the same time position.

5. The Turbo code probability domain decoding method based on a memory computing device according to claim 1, characterized in that: The S400 includes: S410, rearrange the positions of the first external information probability Q1 according to the interleaving mapping π to obtain the interleaved external information sequence corresponding to the time position of the second component code; S420, based on the magnitude of the entropy value at each time position in the state entropy sequence H1 and the continuity of entropy change between adjacent time positions, the credibility level of each time position in the interwoven external information sequence is classified to obtain the position credibility level sequence. S430, jointly constrain the location credibility level sequence and the interleaved external information sequence at the same time location to generate the location credibility weight W corresponding to each time location; S440, using the location confidence weight W to perform location-related recalibration against mismatch branch metric Γ1, so that the mismatch-resistant branch metric of high confidence time location maintains its distinguishing advantage, and the mismatch-resistant branch metric of low confidence time location shrinks towards the neutral probability direction, generating an updated branch metric Γ2.

6. The Turbo code probability domain decoding method based on a memory computing device according to claim 1, characterized in that: The S500 includes: S510, based on the updated branch metric Γ2 and the state masking set Ω, the state transitions at each time position of the second component code are screened to determine the second open state transition and the second masked state transition. S520, in forward time order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second forward state probability of the previous time position, and the second shielded state transition is subjected to restricted accumulation in a weighted retention manner to obtain the second forward state probability sequence. S530, in reverse chronological order, the update branch metric Γ2 of each second open state transition is jointly accumulated with the second backward state probability at the next time position, and the restricted accumulation is performed on each second shielded state transition according to the weighted retention method to obtain the second backward state probability sequence. S540, based on the second forward state probability sequence, the second backward state probability sequence and the update branch metric Γ2 at each time position, generate the second external information probability Q2.

7. The Turbo code probability domain decoding method based on a memory computing device according to claim 1, characterized in that: The S600 includes: S610, the second external information probability Q2 is restored according to the inverse mapping of the interleaving mapping π to obtain the back-placed external information sequence corresponding to the time position of the first component code, and the posterior probability R is generated based on the back-placed external information sequence; S620, compare the probability shift between the posterior probability R and the first external information probability Q1 at the same time position to obtain the difference sequence at each time position; S630, perform hard decisions on the posterior probabilities R of the current iteration and the previous iteration respectively, compare the consistency of decisions at the same time position, and obtain the hard decision stability of adjacent iterations; S640, read the entropy value change trend at each time position in the state entropy sequence H1, determine the entropy collapse rate at each time position, and construct the pseudo-convergence index Z based on the difference sequence, the hard decision stability of adjacent iterations and the entropy collapse rate.

8. The Turbo code probability domain decoding method based on a memory computing unit according to claim 1, characterized in that: The S700 includes: S710, when the pseudo-convergence exponent Z is greater than the preset threshold, extract the mismatch sensitive positions that coincide with the high difference position and the high entropy collapse rate position from the posterior probability R to form a mismatch sensitive position set. S720, based on the distribution of the mismatch sensitive bit set at each time position, perform reverse correction on the state transition corresponding to the mismatch sensitive bit set in the adversarial mismatch branch metric Γ1, weaken the probability bias of the current direction and restore the competitive state; S730, regenerate the state mask set Ω corresponding to the time position based on the reversed mismatch branch metric Γ1, and return to step S300 to continue the iteration; S740: When the pseudo-convergence index Z is less than or equal to the preset threshold, the information bits at each time position are finally hard-determined based on the posterior probability R, and the decoding result is output.

Images