RFID anti-drift positioning method and device based on double-cycle learning of unsupervised learning

Abstract

The application discloses an RFID anti-drift positioning method and device based on unsupervised double-cycle learning, and is applied to the field of artificial intelligence technology.The method comprises the following steps: collecting an RFID reading set corresponding to a current reading window; if signal drift is determined according to the RFID reading set, then performing inner cycle learning based on a shadow model of an RFID online positioning model to obtain a first candidate model; if the effectiveness verification of the first candidate model is passed, then uploading the first candidate model in the current stable version; if the inner cycle learning satisfies an outer cycle triggering condition, then triggering a computing side device to perform outer cycle learning of the RFID positioning model, and a second candidate model is obtained; the computing side device is an edge computing node or a terminal device; and if the new-old domain double verification of the second candidate model is passed, then performing version updating on the current stable version based on the second candidate model. The short-term drift can be quickly corrected, anti-forgetting learning can be realized, long-term positioning stability can be improved, and cost can be reduced.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an RFID anti-drift positioning method and device based on unsupervised double-loop learning. Background Technology

[0002] In large-scale indoor scenarios such as warehousing, manufacturing, and medical facilities, Radio Frequency Identification (RFID) readers are highly sensitive to environmental changes (shelf relocation, metal obstructions, temporary stacking, new tags, etc.), causing signal distribution domains to drift over time. The traditional approach is to build a fingerprint database and train a model offline all at once, then deploy it online for prediction. This combination of static maps and fixed models results in a significant performance degradation after several weeks to months, requiring map re-surveying and model retraining, leading to high maintenance costs and a significant risk of downtime. Meanwhile, directly fine-tuning the model with new data can result in catastrophic amnesia: performance improves in new scenarios but deteriorates in old ones. These issues make existing RFID positioning systems unreliable for long-term robust operation, requiring frequent manual maintenance. Summary of the Invention

[0003] At least one embodiment of this application provides an RFID anti-drift positioning method and apparatus based on unsupervised dual-loop learning, which solves the problem that existing RFID positioning systems are difficult to operate robustly for a long time and require frequent manual maintenance.

[0004] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide an RFID anti-drift positioning method based on unsupervised dual-loop learning, applied to a terminal device, comprising: Collect the set of RFID readings corresponding to the current reading window of the RFID reader / writer; If signal drift is determined to have occurred based on the RFID reading set, then an inner loop learning is performed on the shadow model of the RFID online positioning model to obtain the first candidate model after fine-tuning the shadow model; If the validity of the first candidate model is verified and the inner loop learning does not meet the outer loop triggering condition, the first candidate model is launched in the current stable version for RFID positioning and signal drift judgment. When the inner loop learning satisfies the outer loop triggering condition, the computing-side device is triggered to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and obtain the second candidate model obtained by the computing-side device in the outer loop learning. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. If the dual verification of the old and new domains of the second candidate model passes, the current stable version is updated based on the second candidate model.

[0005] Specifically, the RFID anti-drift positioning method based on unsupervised dual-loop learning, as described above, further includes: If the number of valid readings in the RFID reading set is less than a preset reading threshold, the current reading window is extended to obtain the RFID reading set; or, the current reading window is merged with the previous or subsequent reading window to obtain a new RFID reading set.

[0006] Furthermore, the RFID anti-drift positioning method based on unsupervised dual-loop learning, as described above, further includes: After obtaining the RFID reading set, the monitoring features of the RFID reading set are extracted based on the drift monitoring encoder frozen in the current stable version. Based on the monitoring characteristics, the first RFID observation distribution frequency shift index and the first positioning output uncertainty index are obtained; The first RFID observation distribution frequency shift index and the first positioning output uncertainty index are exponentially smoothed to obtain the second RFID observation distribution frequency shift index and the second positioning output uncertainty index. Based on the second RFID observation distribution frequency shift index and the second positioning output uncertainty index, signal drift is determined to identify the signal drift result corresponding to the RFID reading set.

[0007] Specifically, in the RFID anti-drift positioning method based on unsupervised double-loop learning as described above, the shadow model based on the RFID online positioning model undergoes inner loop learning to obtain a first candidate model after fine-tuning the shadow model, including: The model parameters of the RFID online positioning model are copied to obtain the shadow model; High-confidence target domain samples are obtained by filtering from the RFID reading set; If the number of samples in the high-confidence target domain is less than the preset reading threshold, the shadow model is updated by allowing only batch normalization (BN) statistical updates, entropy monitoring, or security degradation mode to obtain the first candidate model. If the number of samples in the high-confidence target domain is greater than or equal to the preset reading threshold, then old domain replay samples are extracted hierarchically from the replay buffer set. Based on the high-confidence target domain sample and the old domain replay sample, obtain the first domain alignment residual, the target domain average entropy, and the old domain preservation loss; Based on the alignment residual of the first domain, the average entropy of the target domain, and the preservation loss of the old domain, the preset optimization objective of the shadow model is optimized to obtain the first candidate model.

[0008] Furthermore, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, includes: The third localization output uncertainty index, old domain degradation amount, and inference time are obtained based on the first candidate model. If the preset conditions are met based on the old domain degradation amount, the third positioning output uncertainty index, the first domain alignment residual, and the inference time, then the validity verification of the first candidate model is determined to be successful.

[0009] Specifically, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, further includes: Based on the environmental distribution offset persistence index, the third positioning output uncertainty index, the old domain degradation amount, and the first domain alignment residual, the inner loop compensation residual is obtained; Based on the inner loop compensation residual and the preset outer loop triggering condition, determine whether to trigger the outer loop learning outer loop triggering result.

[0010] Specifically, in the RFID anti-drift positioning method based on unsupervised dual-loop learning as described above, the triggering computing device performs outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and obtains a second candidate model obtained by the computing device through the outer loop learning, including: A terminal request is generated and sent to the computing-side device. The terminal request includes at least one of the following: stable version number, device number, warehouse number, trigger window, indicator sequence, model summary, representative sample, weak label or pseudo label, replay summary, and resource status. The second candidate model is obtained by the computing device through training based on the new domain dataset, the old domain replay set, and the outer loop target.

[0011] Specifically, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, updates the current stable version based on the second candidate model, including: Register the second candidate model as the RFID online positioning model corresponding to the new stable version, and regenerate and freeze the drift monitoring encoder; Re-estimate the stability monitoring feature baseline and stability uncertainty index baseline corresponding to the stable version on the gray-scale stable window; Update the threshold configuration corresponding to the stable version based on the quantile of the grayscale stable window; The replay buffer set corresponding to the stable version is updated based on the new domain dataset, the replay buffer set, representative samples, difficult sample samples, coverage range, and capacity limit. Based on the newly registered RFID online positioning model, the drift monitoring encoder, the stable monitoring feature baseline, the stable uncertainty index baseline, the threshold configuration, and the updated playback buffer set, the updated stable version is obtained.

[0012] Secondly, another embodiment of this application provides a control device for RFID anti-drift positioning based on unsupervised dual-loop learning, applied to a terminal device, comprising: The data acquisition module is used to collect the set of RFID readings corresponding to the current reading window of the RFID reader; The inner loop learning module is used to perform inner loop learning based on the shadow model of the RFID online positioning model if signal drift is determined to have occurred according to the RFID reading set, so as to obtain the first candidate model after fine-tuning the shadow model. The first processing module is used to launch the first candidate model in the current stable version and perform RFID positioning and signal drift judgment when the validity verification of the first candidate model is passed and the inner loop learning does not meet the outer loop triggering condition. The outer loop learning module is used to trigger the computing-side device to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set when the inner loop learning satisfies the outer loop triggering condition, and to obtain the second candidate model obtained by the computing-side device through the outer loop learning. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. The second processing module is used to update the current stable version based on the second candidate model if the dual verification of the old and new domains of the second candidate model passes.

[0013] Thirdly, another embodiment of this application provides a terminal device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described above.

[0014] Fourthly, another embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described above.

[0015] Fifthly, another embodiment of this application provides a computer program product including computer instructions that, when executed by a processor, implement the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described above.

[0016] Compared with the prior art, the embodiments of this application provide an RFID anti-drift positioning method and apparatus based on unsupervised dual-loop learning. By collecting RFID readings based on the reading window, and when signal drift is determined to occur, an inner loop learning based on the shadow model is performed to safely probe and quickly correct short-term, local, and compensable drifts that occur during the positioning process. When the outer loop trigger condition is met, an outer loop learning is performed based on the new domain dataset and the old domain replay set to achieve anti-forgetting learning. This allows the system's drift benchmark to evolve with the stable environment, thereby improving long-term positioning stability and reducing the cost of simple model retraining or manual recalibration. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is one of the flowcharts of the RFID anti-drift positioning method based on unsupervised double-loop learning proposed in this application; Figure 2 This is the second flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 3 This is the third flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 4This is the fourth flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 5 This is the fifth flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 6 This is the sixth flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 7 This is the seventh flowchart illustrating the RFID anti-drift positioning method based on unsupervised dual-loop learning proposed in this application. Figure 8 This is a schematic diagram of the control device for RFID anti-drift positioning based on unsupervised dual-loop learning according to this application; Figure 9 This is a schematic diagram of the structure of the terminal device of this application. Detailed Implementation

[0018] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0019] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The terms “and / or” in the specification and claims indicate at least one of the connected objects.

[0020] Please refer to Figure 1 This application provides an RFID anti-drift positioning method based on unsupervised dual-loop learning, applied to a terminal device, comprising: Step S101: Collect the set of RFID readings corresponding to the current reading window of the RFID reader; Step S102: If signal drift is determined to have occurred based on the RFID reading set, then an inner loop learning is performed on the shadow model of the RFID online positioning model to obtain the first candidate model after fine-tuning the shadow model. Step S103: If the validity verification of the first candidate model is passed and the inner loop learning does not meet the outer loop triggering condition, the first candidate model is launched in the current stable version to perform RFID positioning and signal drift judgment. Step S104: When the inner loop learning satisfies the outer loop triggering condition, the computing-side device is triggered to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and a second candidate model obtained by the computing-side device in the outer loop learning is obtained. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. Step S105: If the dual verification of the old and new domains of the second candidate model passes, update the current stable version based on the second candidate model.

[0021] In this embodiment, when a terminal device, such as a robot, performs anti-drift positioning, it collects the set of RFID readings from the RFID reader within the current reading window, for example, in the... The set of RFID readings collected by each reading window can be represented as: ,in, The number of valid readings within the window; a single RFID reading. The reading window should include at least the Received Signal Strength Indicator (RSSI), phase, antenna port, RFID tag identifier (ID), number of reads, and timestamp. It may also include robot synchronization pose, speed, path segment number, warehouse area number, or reader operating frequency. This reading window can be segmented based on time, number of readings, or path distance. Due to the instantaneous fluctuations and tag sparsity of RFID readings, a single reading is insufficient to stably reflect environmental changes. By setting a reading window, subsequent distribution offsets, output uncertainties, and compensation residuals can be calculated on the same time scale, thus reducing the impact of occasional noise on triggering. After obtaining the RFID reading set, signal drift can be determined based on the RFID readings in the set. If the determination result is no drift, positioning output can be performed solely based on the RFID reading set. If the determination result is signal drift, the current RFID online positioning model needs to be fine-tuned based on the RFID reading set. To avoid abnormal sample contamination caused by directly adjusting the RFID online positioning model, in this embodiment, the shadow model based on the RFID online positioning model is preferentially subjected to controlled unsupervised adaptive learning within the current stable version to obtain the first candidate model. After its effectiveness is verified, it is deployed in the current stable version. This allows for safe testing and rapid correction of short-term, local, and compensable drift that occurs during positioning without contaminating the online stable model. The parameters to be corrected include, but are not limited to, RSSI changes, phase perturbations, multipath reflections, and tag occlusion.

[0022] In one specific embodiment, it is preferable to switch the first candidate model online via A / B slot switching. That is, the A / B slot achieves physical isolation between the current RFID online positioning model and the shadow model, and only performs inner loop learning on the shadow model. This ensures that the inner loop process does not affect the RFID online positioning model. Furthermore, if the validity verification of the first candidate model passes, it indicates that the corrected first candidate model can be deployed online and is superior to the current RFID online positioning model. Therefore, the A / B slot is switched, that is, the current RFID online positioning model is deactivated to the candidate slot in the A / B slot, and the first candidate model is deployed to the working slot in the A / B slot. This first candidate model is then used for RFID positioning and signal drift detection, thus achieving A / B slot switching. It should be noted that the inner loop learning can update the positioning head, batch normalized statistics, etc., but does not change the drift detection baseline of the current version.

[0023] Since the current stable version's internal loop learning cannot fully compensate for long-term, structural RFID positioning drift, and long-term internal loop learning is prone to catastrophic forgetting, this application also includes anti-forgetting learning. Therefore, it will determine whether the external loop triggering condition is met based on the internal loop learning to determine whether to trigger the external loop learning. If the internal loop learning does not meet the external loop triggering condition, the first candidate model will be directly launched. If the external loop triggering condition is met, the first candidate model will be abandoned and the computing device will be triggered to perform external loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set. Through this external loop learning, a second candidate model that can perform RFID positioning and signal drift judgment can be obtained. If the second candidate model passes the dual verification of the new and old domains, the current stable version will be updated based on the second candidate model. This version update includes updating the RFID positioning model and updating other parameters within the version. In another embodiment, the first candidate model can be directly launched if the validity verification of the first candidate model is passed. Then, the outer loop triggering condition is judged based on the inner loop learning. If the outer loop triggering condition is met, the computing device is triggered to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain playback set. Otherwise, the current process is directly terminated.

[0024] It should be noted that when performing outer loop learning, the subject performing outer loop learning can be the terminal device or an edge computing node, etc. Among them, anti-forgetting learning, i.e., outer loop learning, can only be performed through the terminal device if the local resource conditions allow for anti-forgetting learning.

[0025] Specifically, the new domain dataset in this embodiment includes data from multiple RFID reading sets of at least one of the terminal devices that have experienced signal drift, and the old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set.

[0026] In summary, this application collects RFID readings based on a reading window and performs inner loop learning based on a shadow model when signal drift is detected. This allows for safe probing and rapid correction of short-term, local, and compensable drifts that occur during positioning. When the outer loop trigger condition is met, outer loop learning is performed based on a new domain dataset and an old domain replay set to achieve anti-forgetting learning. This enables the system's drift baseline to evolve with the stable environment, thereby improving long-term positioning stability and reducing the cost of simple model retraining or manual recalibration.

[0027] It should be noted that the current stable version can be specifically represented as: ,in, Indicates a stable version; Indicates the version number; This represents the current RFID online positioning model. This refers to the localization head, used to transform feature vectors into the final position probability distribution. The feature extractor is used to convert raw RFID readings into high-dimensional feature vectors. To be consistent with the stable version Bound freeze monitoring encoder; and These are the monitoring characteristic values ​​and monitoring characteristic variances in the stable monitoring characteristic baseline, respectively. To stabilize the baseline of uncertainty; For playback buffering; Configure thresholds and gating.

[0028] Specifically, the RFID anti-drift positioning method based on unsupervised dual-loop learning, as described above, further includes: If the number of valid readings in the RFID reading set is less than a preset reading threshold, the current reading window is extended to obtain the RFID reading set; or, the current reading window is merged with the previous or subsequent reading window to obtain a new RFID reading set.

[0029] In this embodiment, after collecting the RFID reading set of the RFID reader in the current reading window, the number of valid readings in the RFID reading set is compared with a preset reading threshold. If the number of valid readings is greater than or equal to the preset reading threshold, the subsequent step of determining whether signal drift has occurred based on the RFID reading set is executed. If the number of valid readings is less than the preset reading threshold, it indicates that the current number of valid readings is small. To avoid inaccuracies or instability in subsequent calculations, such as covariance calculation, caused by a small sample size, the RFID reading set is redefined by extending the current reading window. Alternatively, the current reading window is merged with the previous or next reading window, and the RFID readings collected in the merged reading window constitute the current new RFID reading set, so as to ensure the accuracy and stability of subsequent processing steps.

[0030] In another embodiment, if the number of valid readings is less than a preset reading threshold, the RFID reading set can be ignored. That is, only the RFID reading set is used for positioning output, but no judgment is made on whether signal offset has occurred, nor is any subsequent model update performed. This can also avoid inaccuracies or instability in subsequent calculations (such as covariance calculation) caused by a small sample size.

[0031] See Figure 2 Furthermore, the RFID anti-drift positioning method based on unsupervised dual-loop learning, as described above, further includes: Step S201: After obtaining the RFID reading set, extract the monitoring features of the RFID reading set based on the drift monitoring encoder frozen in the current stable version. Step S202: Based on the monitoring characteristics, obtain the first RFID observation distribution frequency shift index and the first positioning output uncertainty index; Step S203: Perform exponential smoothing on the first RFID observation distribution frequency shift index and the first positioning output uncertainty index to obtain the second RFID observation distribution frequency shift index and the second positioning output uncertainty index. Step S204: Based on the second RFID observation distribution frequency shift index and the second positioning output uncertainty index, a signal drift judgment is made to determine the signal drift result corresponding to the RFID reading set.

[0032] This embodiment illustrates the signal drift judgment process in this application. After obtaining the RFID reading set for the current reading window, monitoring features of the RFID reading set are extracted based on the drift monitoring encoder frozen in the current stable version. During the inner loop, if the features of the RFID positioning model that can be updated within the inner loop are directly used for distribution offset, the feature space before and after the RFID positioning model update may differ, leading to incomparability of the smoothing index, i.e., the RFID observation distribution offset index. Therefore, this application freezes the drift monitoring encoder within the current stable version, ensuring that the RFID observation distribution offset of different reading windows is measured within the same monitoring space. At this time, the reading window is extracted... The formula for the monitoring characteristics of the corresponding RFID reading set can be expressed as follows: ,in, .

[0033] Then, based on this monitoring characteristic, the unsupervised drift health is calculated. In this embodiment, the unsupervised drift health is specifically calculated using the RFID observation distribution frequency shift index. (Used to characterize whether the RFID observation field has changed) and positioning output uncertainty index (Used to characterize whether the model is certain about the current output) is represented, where if only An increase may indicate environmental changes, but the model can still maintain stable positioning; only An increase in the output might indicate model output degradation without a significant change in input statistics. Combining both can more accurately characterize the unsupervised drift health of RFID positioning.

[0034] Specifically, in this embodiment, the frequency shift index of RFID observation distribution is standardized dimensionally. This can be specifically expressed as: ; in, To monitor the number of feature dimensions; Number the dimensions of the monitored features; Indicates the number of times within the current stable version The monitoring features corresponding to each monitoring feature dimension; Indicates the current reading window Inner The mean value of the monitoring features corresponding to each monitoring feature dimension can be specifically expressed as: ; Indicates the number of the current stable version The variance of the monitoring features corresponding to each monitoring feature dimension; representing Current reading window Inner The variance vector of the monitoring features corresponding to each monitoring feature dimension can be specifically represented as follows: ; Used to place fractions with a denominator of zero; These are preset parameters used to adjust the importance of mean shift and variance offset; the first term in the formula reflects whether the RFID observation center deviates from the stable region, and the second term reflects whether the RFID reading fluctuation structure has changed. Based on this, the first RFID observation distribution frequency shift index corresponding to the current reading window can be obtained.

[0035] Specifically, in this embodiment, the positioning output uncertainty index can be represented by normalized entropy, as follows: ; in, For the number of candidate locations, grids, or regions, Number the candidate location, grid, or region. Indicates the current reading window Inner The first of the readings The probability corresponding to each candidate location, grid, or region, where... The output value of the RFID online positioning model can be used if the model directly outputs continuous coordinates. Alternatively, the coordinate space can be discretized into candidate grids, or constructed from the model's output covariance, ensemble prediction variance, or neighborhood probabilistic results. .

[0036] Furthermore, since RFID readings are affected by perturbations such as instantaneous multipath propagation, temporary obstruction, and reader posture, exponential smoothing is used to suppress spikes and prevent frequent false triggers caused by a single window threshold. Hysteresis is used to avoid trigger boundary jitter and ensure the continuity of the continuous window drift. Specifically, the formula for exponential smoothing the RFID observation distribution frequency shift index and the positioning output uncertainty index can be expressed as follows: ; ; In this embodiment This indicates the frequency shift index of the first RFID observation distribution; This indicates the uncertainty index of the first positioning output; This indicates the frequency shift index of the second RFID observation distribution; This indicates the uncertainty index of the second positioning output; This indicates the frequency shift index of the second RFID observation distribution corresponding to the previous reading window; This indicates the uncertainty index of the second positioning output corresponding to the previous reading window; and These are the preset smoothing coefficients, and , ,and and The initial value can be the mean of the stable period or the average of the first few windows.

[0037] Then, signal drift can be determined based on the obtained second RFID observation distribution frequency shift index and second positioning output uncertainty index. Specifically, the determination can be based on the first-level triggering rule and the alarm release rule. The first-level triggering rule is as follows: ; The alarm release rules are as follows: ; in, This indicates the signal drift result corresponding to the current reading window. This indicates that signal drift has occurred. This indicates that no signal drift has occurred; and These represent the first and second thresholds of the pre-configured RFID observation distribution frequency shift index, respectively. ; and These represent the first and second thresholds of the pre-configured positioning output uncertainty index, respectively. ; To trigger the number of consecutive windows; To release consecutive window counts.

[0038] Using a dual-indicator approach—RFID distribution offset and output uncertainty—along with smoothing, hysteresis, and continuous windows to determine whether to perform inner loop learning, helps reduce false triggering caused by short-term occlusion, instantaneous multipath, and sample sparsity.

[0039] See Figure 3 Specifically, in the RFID anti-drift positioning method based on unsupervised double-loop learning as described above, the shadow model based on the RFID online positioning model undergoes inner loop learning to obtain a first candidate model after fine-tuning the shadow model, including: Step S301: Copy the model parameters of the RFID online positioning model to obtain the shadow model; Step S302: Select high-confidence target domain samples from the RFID reading set; Step S303: If the number of samples in the high-confidence target domain is less than the preset reading threshold, the shadow model is updated by allowing only batch normalized statistical updates, entropy monitoring, or security degradation mode to obtain the first candidate model. Step S304: If the number of high-confidence target domain samples is greater than or equal to the preset reading threshold, then old domain replay samples are extracted hierarchically from the replay buffer set. Step S305: Based on the high-confidence target domain sample and the old domain replay sample, obtain the first domain alignment residual, the target domain average entropy, and the old domain preservation loss. Step S306: Based on the first domain alignment residual, the target domain average entropy, and the old domain preservation loss, optimize the preset optimization objective of the shadow model to obtain the first candidate model.

[0040] In this embodiment, the step of performing inner loop learning based on the shadow model of the RFID online positioning model includes: copying the model parameters of the RFID online positioning model, and generating a shadow model for inner loop learning within the candidate slots based on the copied model parameters, i.e., the model parameters of the shadow model. , The model parameters represent the RFID online positioning model; Then, high-confidence target domain samples are selected from the RFID reading set. These high-confidence target domain samples are preferentially determined based on the maximum output value of each candidate location, grid, or region, a preset probability threshold, a positioning output uncertainty index, and a preset positioning output uncertainty index threshold. For example, when the positioning output uncertainty index corresponding to the sample is less than or equal to the preset positioning output uncertainty index threshold, and the maximum output value in the corresponding region is greater than or equal to the preset probability threshold, the calculation formula can be specifically expressed as follows: ; in, This represents the set of samples in the high-confidence target domain. Indicates a preset probability threshold; This indicates the threshold value for the preset positioning output uncertainty index.

[0041] If the number of high-confidence target domain samples in the current reading set is less than the preset reading threshold, then the pseudo-tag driven inner loop learning is not executed. Instead, the shadow model is updated by only allowing batch normalized statistical updates, entropy monitoring, or security degradation mode (SAFE degradation mode or fail-safe degradation mode) to obtain the first candidate model. This can limit and avoid erroneous pseudo-tags in the low-confidence window from contaminating the RFID online positioning model. If the number of high-confidence target domain samples in the current reading set is greater than or equal to the preset reading threshold, then it is determined that pseudo-tag-driven inner loop learning can be performed. At this time, old domain replay samples are extracted hierarchically from the replay buffer set. In a specific embodiment, when extracting old domain replay samples, they are extracted hierarchically according to warehouse area, antenna, tag density, error level and time period. Then, based on the high-confidence target domain samples and the old domain replay samples, the alignment residual of the first domain, the average entropy of the target domain, and the preservation loss of the old domain are calculated. The formula for obtaining the alignment residual of the first domain can be expressed as: ; in, Represents the first domain alignment residual of the shadow model; Indicates the number of samples in the high-confidence target domain; This indicates a sample for replaying the old domain. ; Indicates the number of old domain replay samples; This represents the feature extractor corresponding to the shadow model, where... The parameters refer to the shadow model. Since the high-confidence target domain samples and the old domain replay samples are calculated using the same localization model, there is no feature space inconsistency problem when used for inner loop compensation determination. If the computing power allows, the first domain alignment residual of the number of iterations can also be replaced by the maximum mean discrepancy (MMD), energy distance, or lightweight domain discriminator loss.

[0042] The formula for obtaining the average entropy of the target domain can be expressed as: ; in, Represents the average entropy of the target domain. This represents the output value of the shadow model; When obtaining the old domain retention loss, adaptive calculations can be performed based on the shadow model and the properties of the old domain replay samples. For example, whether weak localization labels are present or whether teacher consistency is used, etc. The specific formula can be expressed as: ; in, This refers to the old domain preservation loss for old domain replay samples under the shadow model.

[0043] Then, based on the first domain alignment residual, target domain average entropy, and old domain preservation loss obtained above, the preset optimization objective of the shadow model is optimized, wherein the preset optimization objective can be expressed as: ; in, , , and These are the corresponding preset weights; During optimization, only the controlled parameter subset in the shadow model is updated. This includes batch normalized statistics, the localization head, and several layers after the feature extractor. The specific formula is as follows: ,in, , This represents the number of controlled parameters. This allows us to obtain the first candidate model after updating the shadow model.

[0044] It should be noted that the low-level RFID basic feature freeze monitoring encoder and the stable monitoring feature baseline of the stable version are not updated in the inner loop learning.

[0045] It should also be noted that, in this embodiment, weak localization labels will be generated for all samples in the new domain data set and the old domain replay set. These labels can be generated based on at least one of the following sources: The actual coordinates obtained from initial database construction or sampling verification; Simultaneous localization and mapping (SLAM), odometry, ultra-wideband (UWB), and synchronous pose provided by orbital constraints or map path constraints; High confidence means the model outputs pseudo-labels that have been filtered by confidence threshold, temporal consistency and path reachability; Sparse calibration points were obtained by manually verifying a small number of key windows on the side.

[0046] See Figure 4 Furthermore, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, includes: Step S401: Obtain the third localization output uncertainty index, old domain degradation amount, and inference time based on the first candidate model; Step S402: If the preset conditions are met based on the old domain degradation amount, the third positioning output uncertainty index, the first domain alignment residual, and the inference time, then the validity verification of the first candidate model is determined to be successful.

[0047] In this embodiment, the effectiveness verification of the first candidate model is illustrated. First, based on the obtained first candidate model, the positioning output uncertainty index is recalculated to obtain the third positioning output uncertainty index. Based on the old domain replay samples, the RFID online positioning model, and the first candidate model, the old domain degradation amount of the first candidate model in old domain replay is obtained. Specifically, it can be expressed as: ,in, In this embodiment, the deviation can be determined according to actual needs, such as: average positioning error, region classification error rate, teacher consistency error, or prototype deviation. Furthermore, the inference time of the first candidate model is obtained through timing and other means.

[0048] Therefore, a judgment can be made based on preset conditions. Specifically, the preset conditions include: ; ; ; ; in, This indicates the uncertainty index of the third positioning output; This indicates the uncertainty index of the second positioning output; This indicates the percentage decrease in the first compensation level; Indicates the alignment residual of the first field; This represents the second-domain alignment residual before the inner loop learning; This indicates the percentage decrease in the second compensation. Indicates the amount of degradation in the old domain; This indicates the upper limit of acceptable degradation for the old domain; The reasoning process takes time; This indicates the preset maximum delay.

[0049] If all the above preset conditions are met, it indicates that the first candidate model has proven that the new domain compensation is effective and the old domain performance has not been damaged. At this time, the validity verification of the first candidate model is determined to be successful, and the first candidate model can be switched online through slot A / B. If at least one of the above preset conditions is not met, it is determined that the validity verification of the first candidate model has failed, and the first candidate model is prohibited from being online, while the current RFID online positioning model is retained.

[0050] See Figure 5 Specifically, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, further includes: Step S501: Obtain the inner loop compensation residual based on the environmental distribution offset persistence index, the third positioning output uncertainty index, the old domain degradation amount, and the first domain alignment residual. Step S502: Based on the inner loop compensation residual and the preset outer loop triggering condition, determine whether to trigger the outer loop learning outer loop triggering result.

[0051] This embodiment demonstrates whether the inner loop learning satisfies the outer loop triggering condition. First, based on the environmental distribution offset persistence index, the aforementioned third positioning output uncertainty index, the old domain degradation, and the first domain alignment residual, the inner loop compensation residual is obtained. Its calculation formula can be specifically expressed as: ; in, This indicates the residual of the inner loop compensation corresponding to the current reading window; This indicates the persistence of environmental distribution shift, specifically expressed as: In the formula Indicates the duration of the environmental distribution offset window; Representation domain alignment residual threshold; , , and These represent the corresponding weight parameters, all of which are greater than or equal to zero and can be normalized.

[0052] After obtaining the inner loop compensation residual, a judgment can be made based on the preset outer loop triggering conditions. These outer loop triggering conditions include: the number of consecutive times the inner loop compensation residual is greater than the compensation residual threshold is greater than a preset threshold; the time interval between this and the most recent outer loop learning point is greater than or equal to the minimum interval of the outer loop; and the number of inner loop compensation failures within a preset time period is greater than or equal to the minimum number of inner loop compensation failures. The corresponding formula can be specifically expressed as follows: ; ; ; in, Indicates the residual compensation threshold; Indicates the number of consecutive times the limit is exceeded; Indicates the minimum interval of the outer loop; This indicates the time point of the most recent external learning cycle; This indicates the number of times the internal loop compensation failed within a preset time period; This represents the minimum threshold number of failed inner loop compensations.

[0053] If all of the above outer loop triggering conditions are met, it indicates that outer loop learning needs to be triggered; if at least one of the above outer loop triggering conditions is not met, it is determined that outer loop learning will not be triggered. This can be specifically determined through... This indicates the result of the outer loop triggering the judgment. This indicates that the outer loop learning has been triggered; This indicates that the outer loop learning will not be triggered, and the terminal device will continue to observe, adopt a security degradation mode, or wait for more windows. Using the compensation residual from the inner loop learning as the trigger for the outer loop, continuous learning in the outer loop only begins when the inner loop compensation is insufficient. This helps reduce the computing power and network overhead of the computing-side devices and avoids ineffective retraining.

[0054] See Figure 6 Specifically, in the RFID anti-drift positioning method based on unsupervised double-loop learning as described above, the triggering computing device performs outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and obtains a second candidate model obtained by the computing device through the outer loop learning, including: Step S601: Generate a terminal request and send it to the computing-side device. The terminal request includes at least one of the following: stable version number, device number, warehouse number, trigger window, indicator sequence, model summary, representative sample, weak label or pseudo label, replay summary and resource status. Step S602: Obtain the second candidate model trained by the computing device based on the new domain dataset, the old domain replay set, and the outer loop target.

[0055] In this embodiment, the steps of performing outer loop learning described above are illustrated. After determining that outer loop learning is triggered, a terminal request is generated and sent to the computing-side device, so that the computing-side device performs outer loop learning according to the terminal request. When performing outer loop learning, the computing-side device aggregates the trigger windows reported by one or more terminal devices to obtain a new domain dataset. And extract old domain replay sets from the historical stable repository. The current RFID online positioning model is trained and optimized based on a preset outer loop learning objective to obtain a second candidate model. The outer loop learning objective is expressed as: ; in, Indicates the learning objective of the outer loop; This indicates location-based supervision, weak supervision, or false labeling loss. Represents RFID online positioning model Compared with the second candidate model In the old domain replay collection Distillation consistency loss on replay samples; Indicates the first The historical importance of each parameter is estimated, preferably by Fisher information diagonal approximation, historical gradient sensitivity, or validation set perturbation sensitivity. , and These are the weight parameters for the corresponding items. This represents the Elastic Weight Consolidation (EWC) weights. By replaying samples, distillation consistency, and EWC parameter importance constraints, the new model can maintain its ability to locate old domains while absorbing changes in the structure of new domains, effectively suppressing the forgetting of old domains.

[0056] Furthermore, when the computing-side device is a terminal device, the terminal requests the data to the local training scheduler of the terminal device, and the local training scheduler executes the aforementioned outer loop learning process; when the computing-side device is an edge computing node, the terminal requests the data to the edge computing node, and the edge computing node executes the aforementioned outer loop learning process.

[0057] See Figure 7 Specifically, the RFID anti-drift positioning method based on unsupervised double-loop learning, as described above, updates the current stable version based on the second candidate model, including: Step S701: Register the second candidate model as the RFID online positioning model corresponding to the new stable version, and regenerate and freeze the drift monitoring encoder; Step S702: Re-estimate the stability monitoring feature baseline and stability uncertainty index baseline corresponding to the stable version on the grayscale stabilization window; Step S703: Update the threshold configuration corresponding to the stable version according to the quantile of the grayscale stable window; Step S704: Update the replay buffer set corresponding to the stable version according to the new domain dataset, the replay buffer set, representative samples, difficult sample samples, coverage range and capacity limit; Step S705: Based on the newly registered RFID online positioning model, the drift monitoring encoder, the stable monitoring feature baseline, the stable uncertainty index baseline, the threshold configuration, and the updated playback buffer set, the updated stable version is obtained.

[0058] In this embodiment, after obtaining the second candidate model, the current stable version will be updated based on the second candidate model. Specifically, the second candidate model will be registered as the RFID online positioning model corresponding to the new stable version, and the drift monitoring encoder bound to the new stable version will be regenerated and frozen so that monitoring feature extraction and other steps can be performed in the new stable version. Furthermore, the baseline of stable monitoring features and the baseline of stable uncertainty index corresponding to the stable version will be re-estimated on the gray-scale stable window, which can be specifically expressed as: ; ; ; in, and These represent the monitoring characteristic values ​​and monitoring characteristic variances in the re-estimated stable monitoring characteristic baseline, respectively; This represents the baseline of the stability uncertainty index; Indicates a stable grayscale window; This indicates the newly generated drift monitoring encoder; This represents the uncertainty index of the positioning output of the newly registered RFID online positioning model within the grayscale stable window.

[0059] Furthermore, the threshold configuration corresponding to the stable version will be updated based on the quantile of the grayscale stable window. Its partial threshold can be specifically expressed as: ; ; in, and These represent the preset offsets of the preset RFID observation frequency shift index and the positioning output uncertainty index, respectively.

[0060] Furthermore, the replay buffer set corresponding to the stable version will be updated based on the new domain dataset, replay buffer set, representative samples, difficult examples, coverage, and capacity limit. The representative samples and difficult examples can be obtained during the acquisition of the old domain replay set; specifically, this can be represented as follows:

[0061] in, This indicates that representative samples close to the center of each stratum are retained; This indicates that difficult examples with high error, high entropy, or low confidence should be retained. This indicates the coverage area, such as warehouse area, passageway, antenna, tag density, and time period; Indicates the maximum capacity.

[0062] Finally, the updated stable version is obtained by combining the newly acquired RFID online positioning model, drift monitoring encoder, stable monitoring feature baseline, stable uncertainty index baseline, threshold configuration, and playback buffer set. The new RFID online positioning model synchronously updates the monitoring encoder, baseline statistics, threshold configuration, and playback buffer, ensuring that the next round of drift detection is consistent with the current stable environment.

[0063] In another embodiment of this application, for ease of understanding by those skilled in the art, an example is provided of the steps for determining whether the dual verification of the new and old domains of the second candidate model passes. This involves calculating the new domain gain and old domain fallback of the second candidate model relative to the RFID online positioning model based on the new domain verification set and the old domain replay verification set, respectively, to determine whether a preset verification condition is met. This preset verification condition is: the second candidate model... New Domain Validation Set Positioning error It should be less than or equal to the RFID online positioning model (old model) corresponding to the current stable version. In the same new domain verification set Positioning error Subtract the third threshold The difference; and the second candidate model Old domain replay verification set Positioning error It should be less than or equal to the RFID online positioning model (old model) corresponding to the current stable version. Replaying the verification set in the same old domain Positioning error Add a fourth threshold The sum of; specifically, it can be expressed as: ; ; in, This represents the localization error corresponding to the second candidate model in the new domain; This indicates the positioning error corresponding to the RFID online positioning model in the new domain; Indicates the third threshold; This represents the localization error corresponding to the second candidate model in the old domain; This indicates the positioning error corresponding to the RFID online positioning model in the old domain; This represents the fourth threshold.

[0064] In addition, to ensure the successful use of the second candidate model, deployment constraints will be imposed on it, such as model size. Must be less than or equal to the preset maximum size ,Right now Inference delay The delay must be less than or equal to the preset maximum delay. ,Right now .

[0065] Therefore, the second candidate model is only confirmed to have passed the dual verification of the old and new domains if it meets the above-mentioned preset verification conditions and deployment constraints. The step of updating the current stable version based on the second candidate model is then executed. Otherwise, the version update is rejected, the current stable version is maintained, and the reason for failure is recorded.

[0066] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0067] See Figure 8 Another embodiment of this application provides a control device for RFID anti-drift positioning based on unsupervised dual-loop learning, applied to a terminal device, comprising: The data acquisition module 801 is used to acquire the set of RFID readings corresponding to the current reading window of the RFID reader / writer; The inner loop learning module 802 is used to perform inner loop learning based on the shadow model of the RFID online positioning model if signal drift is determined to have occurred according to the RFID reading set, so as to obtain a first candidate model after fine-tuning the shadow model. The first processing module 803 is used to launch the first candidate model in the current stable version and perform RFID positioning and signal drift judgment when the validity verification of the first candidate model is passed and the inner loop learning does not meet the outer loop triggering condition. The outer loop learning module 804 is used to trigger the computing-side device to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set when the inner loop learning satisfies the outer loop triggering condition, and to obtain the second candidate model obtained by the computing-side device in the outer loop learning. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. The second processing module 805 is used to update the current stable version based on the second candidate model if the dual benchmark verification of the old and new domains of the second candidate model passes.

[0068] Specifically, the RFID anti-drift positioning control device based on unsupervised dual-loop learning, as described above, further includes: The third processing module is used to extend the current reading window to obtain the RFID reading set when the number of valid readings in the RFID reading set is less than a preset reading threshold, or to merge the current reading window with the previous or subsequent reading window to obtain a new RFID reading set.

[0069] Furthermore, the RFID anti-drift positioning control device based on unsupervised dual-loop learning, as described above, also includes: The feature extraction module is used to extract the monitoring features of the RFID reading set based on the drift monitoring encoder frozen in the current stable version after obtaining the RFID reading set. The indicator acquisition module is used to acquire a first RFID observation distribution frequency shift indicator and a first positioning output uncertainty indicator based on the monitoring characteristics. The index smoothing module is used to perform exponential smoothing on the first RFID observation distribution frequency shift index and the first positioning output uncertainty index to obtain the second RFID observation distribution frequency shift index and the second positioning output uncertainty index. The signal drift judgment module is used to judge the signal drift based on the second RFID observation distribution frequency shift index and the second positioning output uncertainty index, and to determine the signal drift result corresponding to the RFID reading set.

[0070] Specifically, in the RFID anti-drift positioning control device based on unsupervised dual-loop learning as described above, the inner loop learning module includes: The first processing unit is used to copy the model parameters of the RFID online positioning model to obtain the shadow model; The second processing unit is used to filter high-confidence target domain samples from the RFID reading set; The third processing unit is used to update the shadow model by allowing only batch normalized statistical updates, entropy monitoring, or security degradation mode if the number of samples in the high-confidence target domain is less than the preset reading threshold, so as to obtain the first candidate model. The fourth processing unit is used to extract old domain replay samples hierarchically from the replay buffer set if the number of samples in the high-confidence target domain is greater than or equal to the preset reading threshold. The fifth processing unit is used to obtain the first domain alignment residual, the target domain average entropy, and the old domain retention loss based on the high-confidence target domain sample and the old domain replay sample. The sixth processing unit is used to optimize the shadow model based on the first domain alignment residual, the target domain average entropy and the old domain preservation loss, to obtain the first candidate model.

[0071] Furthermore, the control device for RFID anti-drift positioning based on unsupervised dual-loop learning as described above includes: The fourth processing module is used to obtain the third localization output uncertainty index, old domain degradation amount, and inference time based on the first candidate model; The validity verification module is used to determine that the validity verification of the first candidate model is passed if the preset conditions are met based on the old domain degradation amount, the third positioning output uncertainty index, the first domain alignment residual, and the inference time.

[0072] Specifically, the RFID anti-drift positioning control device based on unsupervised dual-loop learning, as described above, further includes: The fifth processing module is used to obtain the inner loop compensation residual based on the environmental distribution offset persistence index, the third positioning output uncertainty index, the old domain degradation amount and the first domain alignment residual; The outer loop trigger judgment module is used to determine whether to trigger the outer loop trigger result of the outer loop learning based on the inner loop compensation residual and the preset outer loop trigger conditions.

[0073] Specifically, in the RFID anti-drift positioning control device based on unsupervised dual-loop learning as described above, the outer loop learning module includes: The seventh processing unit is used to generate a terminal request and send it to the computing-side device. The terminal request includes at least one of the following: stable version number, device number, warehouse number, trigger window, indicator sequence, model summary, representative sample, weak label or pseudo label, replay summary and resource status. The eighth processing unit is used to obtain the second candidate model trained by the computing device based on the new domain dataset, the old domain replay set, and the outer loop target.

[0074] Specifically, the second processing module of the RFID anti-drift positioning control device based on unsupervised dual-loop learning, as described above, includes: The ninth processing unit is used to register the second candidate model as the RFID online positioning model corresponding to the new stable version, and to regenerate and freeze the drift monitoring encoder. The tenth processing unit is used to re-estimate the stability monitoring feature baseline and stability uncertainty index baseline corresponding to the stable version on the gray-scale stabilization window; The eleventh processing unit is used to update the threshold configuration corresponding to the stable version based on the quantile of the grayscale stable window. The twelfth processing unit is used to update the replay buffer set corresponding to the stable version based on the new domain dataset, the replay buffer set, representative samples, difficult sample samples, coverage range and capacity limit. The thirteenth processing unit is used to obtain the updated stable version based on the newly registered RFID online positioning model, the drift monitoring encoder, the stable monitoring feature baseline, the stable uncertainty index baseline, the threshold configuration, and the updated playback buffer set.

[0075] The apparatus embodiments of this application are apparatuses corresponding to the embodiments of the methods described above. All implementation means in the method embodiments described above are applicable to the apparatus embodiments and can achieve the same technical effects. The apparatus provided in this application embodiments can implement all the method steps implemented in the method embodiments described above and can achieve the same technical effects. Therefore, the parts and beneficial effects that are the same as those in the method embodiments in this embodiment will not be described in detail here.

[0076] See Figure 9 Another embodiment of this application provides a terminal device 900, including: a processor 901, a memory 902, and a program stored in the memory 902 and executable on the processor 901. When the program is executed by the processor 901, it implements the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described above, and can achieve the same technical effect. Here, the parts and beneficial effects that are the same as those in the method embodiment will not be described in detail.

[0077] Another embodiment of this application provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the RFID anti-drift positioning method based on unsupervised dual-loop learning as described above, and achieves the same technical effect. Therefore, the parts and beneficial effects identical to those in the method embodiment will not be described in detail here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0078] Another embodiment of this application provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described above, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail again.

[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0081] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An RFID anti-drift positioning method based on unsupervised dual-loop learning, applied to terminal devices, characterized in that, include: Collect the set of RFID readings from the RFID reader / writer within the current reading window; If signal drift is determined to have occurred based on the RFID reading set, then an inner loop learning is performed on the shadow model of the RFID online positioning model to obtain the first candidate model after fine-tuning the shadow model; If the validity of the first candidate model is verified and the inner loop learning does not meet the outer loop triggering condition, the first candidate model is launched in the current stable version for RFID positioning and signal drift judgment. When the inner loop learning satisfies the outer loop triggering condition, the computing-side device is triggered to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and obtain the second candidate model obtained by the computing-side device in the outer loop learning. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. If the dual verification of the old and new domains of the second candidate model passes, the current stable version is updated based on the second candidate model. The shadow model based on the RFID online positioning model undergoes inner loop learning to obtain a first candidate model after fine-tuning the shadow model, including: The model parameters of the RFID online positioning model are copied to obtain the shadow model; High-confidence target domain samples are obtained by filtering from the RFID reading set; If the number of samples in the high-confidence target domain is less than a preset reading threshold, the shadow model is updated by allowing only batch normalized statistical updates, entropy monitoring, or security degradation mode to obtain the first candidate model. If the number of samples in the high-confidence target domain is greater than or equal to the preset reading threshold, then old domain replay samples are extracted hierarchically from the replay buffer set. Based on the high-confidence target domain sample and the old domain replay sample, obtain the first domain alignment residual, the target domain average entropy, and the old domain preservation loss; Based on the alignment residual of the first domain, the average entropy of the target domain, and the preservation loss of the old domain, the preset optimization objective of the shadow model is optimized to obtain the first candidate model.

2. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 1, characterized in that, The method further includes: If the number of valid readings in the RFID reading set is less than a preset reading threshold, the current reading window is extended to obtain the RFID reading set; or, the current reading window is merged with the previous or subsequent reading window to obtain a new RFID reading set.

3. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 1 or 2, characterized in that, The method further includes: After obtaining the RFID reading set, the monitoring features of the RFID reading set are extracted based on the drift monitoring encoder frozen in the current stable version. Based on the monitoring characteristics, the first RFID observation distribution frequency shift index and the first positioning output uncertainty index are obtained; The first RFID observation distribution frequency shift index and the first positioning output uncertainty index are exponentially smoothed to obtain the second RFID observation distribution frequency shift index and the second positioning output uncertainty index. Based on the second RFID observation distribution frequency shift index and the second positioning output uncertainty index, signal drift is determined to identify the signal drift result corresponding to the RFID reading set.

4. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 1, characterized in that, include: The third localization output uncertainty index, old domain degradation amount, and inference time are obtained based on the first candidate model. If the preset conditions are met based on the old domain degradation amount, the third positioning output uncertainty index, the first domain alignment residual, and the inference time, then the validity verification of the first candidate model is determined to be successful.

5. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 4, characterized in that, Also includes: Based on the environmental distribution offset persistence index, the third positioning output uncertainty index, the old domain degradation amount, and the first domain alignment residual, the inner loop compensation residual is obtained; Based on the inner loop compensation residual and the preset outer loop triggering condition, determine whether to trigger the outer loop learning outer loop triggering result.

6. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 1, characterized in that, The triggering computing device performs outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set, and obtains a second candidate model obtained by the computing device through the outer loop learning, including: A terminal request is generated and sent to the computing-side device. The terminal request includes at least one of the following: stable version number, device number, warehouse number, trigger window, indicator sequence, model summary, representative sample, weak label or pseudo label, replay summary, and resource status. The second candidate model is obtained by the computing device through training based on the new domain dataset, the old domain replay set, and the outer loop target.

7. The RFID anti-drift positioning method based on unsupervised dual-loop learning according to claim 1, characterized in that, Based on the second candidate model, the current stable version is updated, including: Register the second candidate model as the RFID online positioning model corresponding to the new stable version, and regenerate and freeze the drift monitoring encoder; Re-estimate the stability monitoring feature baseline and stability uncertainty index baseline corresponding to the stable version on the gray-scale stable window; Update the threshold configuration corresponding to the stable version based on the quantile of the grayscale stable window; The replay buffer set corresponding to the stable version is updated based on the new domain dataset, the replay buffer set, representative samples, difficult sample samples, coverage range, and capacity limit. Based on the newly registered RFID online positioning model, the drift monitoring encoder, the stable monitoring feature baseline, the stable uncertainty index baseline, the threshold configuration, and the updated playback buffer set, the updated stable version is obtained.

8. A control device for RFID anti-drift positioning based on unsupervised dual-loop learning, applied to terminal equipment, characterized in that, include: The data acquisition module is used to collect the set of RFID readings corresponding to the current reading window of the RFID reader; The inner loop learning module is used to perform inner loop learning based on the shadow model of the RFID online positioning model if signal drift is determined to have occurred according to the RFID reading set, so as to obtain the first candidate model after fine-tuning the shadow model. The first processing module is used to launch the first candidate model in the current stable version and perform RFID positioning and signal drift judgment when the validity verification of the first candidate model is passed and the inner loop learning does not meet the outer loop triggering condition. The outer loop learning module is used to trigger the computing-side device to perform outer loop learning of the RFID positioning model based on the new domain dataset and the old domain replay set when the inner loop learning satisfies the outer loop triggering condition, and to obtain the second candidate model obtained by the computing-side device through the outer loop learning. The computing-side device is an edge computing node or a terminal device. The new domain dataset includes data from multiple RFID reading sets of at least one terminal device that have experienced signal drift. The old domain replay set includes historical data extracted from the historical stable library and / or replay buffer set. The second processing module is used to update the current stable version based on the second candidate model if the dual verification of the old and new domains of the second candidate model passes. The inner loop learning module includes: The first processing unit is used to copy the model parameters of the RFID online positioning model to obtain the shadow model; The second processing unit is used to filter high-confidence target domain samples from the RFID reading set; The third processing unit is used to update the shadow model by only allowing batch normalized statistical updates, entropy monitoring, or security degradation mode if the number of samples in the high-confidence target domain is less than a preset reading threshold, so as to obtain the first candidate model. The fourth processing unit is used to extract old domain replay samples hierarchically from the replay buffer set if the number of samples in the high-confidence target domain is greater than or equal to the preset reading threshold. The fifth processing unit is used to obtain the first domain alignment residual, the target domain average entropy, and the old domain retention loss based on the high-confidence target domain sample and the old domain replay sample. The sixth processing unit is used to optimize the shadow model based on the first domain alignment residual, the target domain average entropy and the old domain preservation loss, to obtain the first candidate model.

9. A terminal device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the RFID anti-drift positioning method based on unsupervised double-loop learning as described in any one of claims 1 to 7.

11. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the RFID anti-drift positioning method based on unsupervised dual-loop learning as described in any one of claims 1 to 7.

Images