Machine learning model closed-loop iteration method, device and medium based on edge-cloud cooperation

Abstract

The application discloses a machine learning model closed-loop iteration method and device based on edge cloud cooperation and a medium, relates to the technical field of incremental learning, and comprises the following steps: copying a corresponding edge side machine learning model into three training models, and performing false positive incremental training, false negative incremental training and false positive and false negative joint incremental training to form a false positive candidate model, a false negative candidate model and a joint candidate model; comparing error correction conditions with original recognition maintenance conditions, and collecting candidate models meeting retention conditions into a candidate model set; implementing bypass reasoning on the candidate model set, comparing false positive improvement conditions, false negative improvement conditions and original recognition maintenance conditions of each candidate model, calculating competition results of each candidate model according to the false positive improvement conditions, the false negative improvement conditions and the original recognition maintenance conditions, and forming edge competition evaluation results. The application improves the stability of joint incremental training of the machine learning model through approximate orthogonal low-rank updating.

Description

Technical Field

[0001] This invention relates to the field of incremental learning technology, and in particular to a closed-loop iterative method, device and medium for machine learning models based on edge-cloud collaboration. Background Technology

[0002] With the continuous development of edge computing, cloud computing, and machine learning technologies, the deployment of machine learning models is gradually evolving from centralized training to long-term operation at the edge and collaborative updates in the cloud. Edge-side machine learning models perform real-time recognition for continuous sampling tasks, continuously generating recognition results, model versions, and sampling time sequence records during operation. The cloud relies on higher computing resources to undertake sample verification, parameter updates, and model version maintenance. Machine learning research for dynamic scenarios has gradually focused on technologies such as continuous learning, incremental learning, and efficient parameter fine-tuning. Among these, low-rank adaptation can adjust a limited number of parameters while keeping the main parameters unchanged, while side-channel inference can evaluate the actual recognition performance of different candidate models without changing the current online recognition state. These technologies have become an important research direction for the continuous evolution of edge intelligence.

[0003] In the long run, the false positive and false negative samples generated by existing machine learning models usually have different error causes and parameter correction directions. However, existing incremental training methods often involve false positive and false negative samples in parameter updates, lacking independent constraints for different error directions. When the parameter update directions corresponding to the two types of samples overlap or even conflict, joint training is prone to the phenomenon that one type of recognition error is corrected while the other type of recognition error increases in the opposite direction. This may also change the parameter representation corresponding to the original correctly recognized samples, making it difficult for the machine learning model to maintain a stable relationship between false positive correction, false negative correction and the original recognition ability during continuous iteration. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a closed-loop iterative method for machine learning models based on edge-cloud collaboration to solve the problem of simultaneously maintaining a stable relationship between false alarm correction, false negative correction, and the original recognition capability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a closed-loop iterative method for machine learning models based on edge-cloud collaboration, comprising, The identification results of the edge-side machine learning model are arranged in time sequence to form an edge running trajectory. Anomaly identification segments are collected from the edge running trajectory, and false positive and false negative evidence is accumulated. The anomaly identification segments are reviewed in the cloud, and the review results are divided into false positive samples and false negative samples to form an incremental training sample set. Based on the model version recorded in the incremental training sample set, the corresponding edge-side machine learning model is copied into three training models. Based on the independent low-rank adaptation parameters, the three training models are subjected to false positive incremental training, false negative incremental training, and joint false positive and false negative incremental training respectively, forming false positive candidate models, false negative candidate models, and joint candidate models. The error correction situation is compared with the original recognition preservation situation, and the candidate models that meet the preservation conditions are grouped into a candidate model set. Bypass inference is performed on the candidate model set to compare the improvement in false alarms and false alarms of each candidate model with the original recognition retention. The competition results of each candidate model are calculated according to the improvement in false alarms, false alarms and original recognition retention. The candidate models are ranked according to the competition results to form the marginal competition evaluation results. The candidate model ranked first in the edge competition evaluation results is selected. The training weights of false positive samples and false negative samples and the accumulation conditions of false positive and false negative evidence are adjusted according to the improvement of false positives and false negatives of the candidate model ranked first. The edge side machine learning model is replaced by the candidate model ranked first, forming a closed-loop iterative result of the machine learning model.

[0007] As a preferred embodiment of the closed-loop iterative method for the edge-cloud collaborative machine learning model described in this invention, wherein: the formation of the incremental training sample set specifically involves, The recognition results of the machine learning model along the edge are arranged in the recognition time sequence, and the recognition results of continuous sampling positions are recorded in sequence to form the edge running trajectory; Anomaly identification segments whose identification results change continuously are collected from the edge running trajectory. The identification results within the anomaly identification segments are accumulated according to the direction of false alarm and the direction of missed alarm, forming evidence of false alarm and missed alarm. By reviewing the anomaly identification sections involved in false positives and false negatives in the cloud, and comparing the review results with the identification results of the edge-side machine learning model, false positive samples and false negative samples are obtained respectively. False positive and false negative samples are arranged according to the recognition time sequence. The arranged false positive and false negative samples are then aggregated to form an incremental training sample set.

[0008] As a preferred embodiment of the closed-loop iterative method for the edge-cloud collaborative machine learning model described in this invention, the step of replicating the corresponding edge-side machine learning model into three training models specifically involves: Based on the model version recorded in the incremental training sample set, retrieve the model structure and model parameters of the edge-side machine learning model; The model structure and parameters of the edge-side machine learning model are copied to three training instances respectively, keeping the model structure and parameters of the three training instances consistent, forming a training instance group of the same version; Independent low-rank adaptation parameters are configured at the same trainable parameter positions in the same training instance group of the same version. The original model parameters of the three training instances are fixed, and the update paths of the independent low-rank adaptation parameters of the three training instances are separated to form three training models. The three training models include a false positive training model, a false negative training model, and a joint false positive and false negative training model.

[0009] As a preferred embodiment of the closed-loop iterative method for machine learning models based on edge-cloud collaboration described in this invention, the formation of false positive candidate models, false negative candidate models, and joint candidate models specifically involves: False positive samples and false negative samples are separated from the incremental training sample set. False positive samples are assigned to the false positive training model, false negative samples are assigned to the false negative training model, and false positive samples and false negative samples are jointly assigned to the false positive and false negative joint training model, forming a three-way incremental training relationship. The original model parameters of the false alarm training model are frozen, and the independent low-rank fitting parameters are updated using false alarm samples to form a false alarm candidate model. Based on the missed sample, the independent low-rank fitting parameters of the missed training model are incrementally updated while keeping the original model parameters unchanged, thus forming a missed candidate model. In the joint training model of false positives and false negatives in the three-way incremental training relationship, independent low-rank adaptation parameter update directions are established according to the false positive samples and the false negative samples respectively. The two independent low-rank adaptation parameter update directions are restricted to an approximately orthogonal low-rank update subspace. The independent low-rank adaptation parameters are updated according to the two independent low-rank adaptation parameter update directions to form a joint candidate model.

[0010] As a preferred embodiment of the closed-loop iterative method for machine learning models based on edge-cloud collaboration described in this invention, the independent low-rank adaptation parameters are a set of trainable parameters set at the same parameter layer position of the three training models and participating in the false positive incremental training, false negative incremental training and false positive and false negative joint incremental training in the form of a low-rank matrix respectively. The approximately orthogonal low-rank update subspace is an adjustment that orthogonals the update directions of independent low-rank adaptation parameters, and limits the update directions of two independent low-rank adaptation parameters to the low-rank parameter update range where the inner product is close to zero.

[0011] As a preferred embodiment of the closed-loop iterative method for the edge-cloud collaborative machine learning model described in this invention, wherein: the formation of the edge competition evaluation result specifically comprises, Bypass inference is performed on the candidate model set, and the recognition results of the edge-side machine learning model and each candidate model on the same edge recognition sample are recorded separately to form a bypass inference record. The bypass inference involves deploying each candidate model in the candidate model set at the bypass inference position corresponding to the edge-side machine learning model, synchronously sending the same edge recognition samples into the edge-side machine learning model and each candidate model, and performing recognition on the same edge recognition samples while maintaining consistency in preprocessing method, input size, recognition parameter configuration and result judgment method, and recording the recognition results of the edge-side machine learning model and each candidate model according to the sampling time sequence. In the side-channel inference record, compare the number of false alarms and false negatives of each candidate model with the edge-side machine learning model, check the number of correctly identified samples retained by each candidate model, and form the false alarm improvement, false negative improvement and original identification retention respectively, and compile them into the candidate model evaluation record. Based on the evaluation records of candidate models, Pareto dominance judgment is performed on each candidate model in pairs according to the improvement of false alarms, the improvement of false negatives, and the preservation of the original identification. The candidate models that are not dominated by the remaining candidate models are listed as the first non-dominated layer. Pareto dominance judgment is performed on the remaining candidate models step by step to form the candidate model non-dominated layer. In the non-dominated levels of the candidate models, the evaluation results with the highest improvement in false positives, the highest improvement in false negatives, and the highest retention of the original identification were selected respectively. The three highest evaluation results are combined as reference evaluation points. The improvement of false alarms, the improvement of false alarms, and the preservation of the original identification are normalized on the same scale. The competition distance is calculated according to the Euclidean distance between each candidate model and the reference evaluation point. The candidate models are arranged in order based on the non-dominated level of the candidate models. Within the same non-dominated level of the candidate models, they are arranged in ascending order of competition distance to form the candidate model competition results. The candidate models are ranked according to their competition results, forming the marginal competition evaluation results.

[0012] As a preferred embodiment of the closed-loop iterative method for machine learning models based on edge-cloud collaboration described in this invention, the edge recognition samples are real-time sample sets collected at the edge and used by the edge-side machine learning model and each candidate model to perform bypass inference under the same recognition conditions.

[0013] As a preferred embodiment of the edge-cloud collaborative machine learning model closed-loop iteration method described in this invention, wherein: the formation of the machine learning model closed-loop iteration result specifically comprises, From the edge competition evaluation results, the candidate model ranked first is selected according to the order of candidate models. The improvement of false alarms and the improvement of false alarms of the ranked candidate model are retained to form the evaluation record of the ranked candidate model. Based on the evaluation record of the first candidate model, the training weights of false positive samples and false negative samples are adjusted according to the improvement of false positives and the improvement of false negatives, respectively, to form the training weights for the next round. Based on the training weights for the next round, the cumulative requirements for false positive and false negative evidence in the cumulative conditions for false positive and false negative evidence are adjusted respectively to form the cumulative conditions for false positive and false negative evidence in the next round. The cumulative conditions for false positive and false negative evidence are that the cumulative situation of false positive and false negative evidence reaches the preset requirements for initiating cloud review and incremental training of the model. The edge-side machine learning model is replaced by the first-ranked candidate model, and the training weights for the next round and the cumulative conditions for false positives and false negatives are incorporated into subsequent iterations to form a closed-loop iterative result for the machine learning model.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the closed-loop iterative method for the edge-cloud collaborative machine learning model as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the closed-loop iterative method for the edge-cloud collaborative machine learning model as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By using approximately orthogonal low-rank updates, the stability of joint incremental training of machine learning models is improved. Constraining the update directions of the independent low-rank adaptation parameters corresponding to false positive samples and those corresponding to false negative samples within an approximately orthogonal low-rank update subspace reduces the overlap between the two update directions at the same parameter positions, weakens the directional interference caused by false positive correction to false negative correction, and keeps the original model parameters unchanged. This allows for targeted adjustments to both false positive and false negative identification capabilities, and reduces the disturbance of incremental training to the original correct identification capability. Therefore, it improves the error correction stability and original identification retention capability of the joint candidate model, providing a more stable machine learning model foundation for subsequent competitive evaluation of edge-side candidate models. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a closed-loop iterative method for edge-cloud collaborative machine learning models.

[0019] Figure 2 The graph is formed for the incremental training sample set.

[0020] Figure 3 The candidate model is used for incremental training.

[0021] Figure 4 This is a comparison chart of the evaluation results of the candidate models. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a closed-loop iterative method for a machine learning model based on edge-cloud collaboration, comprising the following steps: S1. Arrange the recognition results of the edge-side machine learning model in time sequence to form an edge running trajectory. Collect anomaly recognition segments from the edge running trajectory, accumulate false positive and false negative evidence, and review the anomaly recognition segments through the cloud. Divide the review results into false positive samples and false negative samples to form an incremental training sample set.

[0026] S1.1. The recognition results of the machine learning model along the edge are arranged in a temporal sequence. The recognition results of continuous sampling positions are recorded sequentially to form the edge running trajectory. Specifically, Following the recognition timeline of the edge-side machine learning model, edge recognition samples formed at consecutive sampling locations are sequentially recognized. The existence state of the target in the edge recognition sample is determined according to the recognition task currently undertaken by the edge-side machine learning model, and the corresponding target category is recorded synchronously when the recognition task includes target category differentiation. The recognition result of each sampling location is formed by the target existence state and the corresponding target category. According to the sampling order, the sampling location, recognition result, and the current model version identifier of the edge-side machine learning model are kept in the same temporal correspondence. The recognition results before and after are concatenated according to the consecutive sampling locations, so that the recognition results formed at each sampling location can continuously reflect the changes in the target recognition state with the sampling timeline. Then, the recognition results are collected according to the sequential relationship of the consecutive sampling locations to form the edge running trajectory.

[0027] S1.2. Collect anomaly identification segments from the edge trajectory where the identification results change continuously. Accumulate the identification results within the anomaly identification segments according to the direction of false alarms and the direction of missed alarms to form evidence of false alarms and missed alarms. Specifically, Starting from the initial sampling position of the edge trajectory, the adjacent recognition results are compared one by one. The sampling positions that maintain the same recognition result continuously are used as the recognition state reference. The sampling positions between the recognition result leaving the previous continuous state and continuously changing until re-entering the continuous and consistent state are grouped into anomaly recognition segments. The direction of change within the anomaly recognition segment is checked against the continuous and consistent recognition results before and after the anomaly recognition segment. Changes in which the edge-side machine learning model provides additional target recognition but no target recognition is provided in the continuous states before and after are classified as false alarm directions. Changes in which the edge-side machine learning model lacks target recognition but maintains target recognition in the continuous states before and after are classified as false alarm directions. The recognition results and sampling positions corresponding to the false alarm direction and the false alarm direction are accumulated according to the recognition time sequence to form false alarm and false alarm evidence.

[0028] S1.3. The abnormal identification sections involved in the false positives and false negatives are reviewed via the cloud. The review results are compared with the identification results of the edge-side machine learning model to obtain false positive samples and false negative samples, specifically: By retrieving the original identification samples corresponding to each sampling location from the false positive and false negative evidence in the cloud, and according to the identification sequence recorded by the edge running trajectory, the original identification samples are associated with the corresponding sampling locations and the identification results of the edge-side machine learning model item by item. Based on the sample labeling rules pre-used by the edge-side machine learning model for the current identification task, the existence status of the target in the original identification samples is verified and confirmed, and the existence status of the target confirmed according to the same sample labeling rules is recorded as the cloud verification result. The cloud verification result is compared with the identification result formed by the edge-side machine learning model at the same sampling location. The original identification samples in which the edge-side machine learning model identifies the target but the cloud verification result confirms that the target does not exist are classified as false positive samples. The original identification samples in which the edge-side machine learning model does not identify the target but the cloud verification result confirms that the target exists are classified as false negative samples. The original identification samples in which the two identification results are consistent are not classified as false positive samples or false negative samples. False positive samples and false negative samples are obtained separately. According to the correspondence between the false positive samples and false negative samples and the cloud verification result, the cloud verification result is saved together with the corresponding sample for supervision results in subsequent incremental training.

[0029] S1.4. Arrange the false alarm samples and missed alarm samples according to the recognition time sequence, and collect the arranged false alarm samples and missed alarm samples to form an incremental training sample set. Specifically, Based on the original recognition time sequence of false positive and false negative samples in the edge running trajectory, the sampling order is restored, and the false positive and false negative samples are arranged separately. The sampling positions corresponding to the false positive and false negative samples and the model version identifiers of the edge-side machine learning models are retained, so that the false positive and false negative samples generated by the same model version have a clear affiliation. The false positive and false negative samples that have completed the time sequence arrangement are placed in the same training sample set according to the model version identifier, so that subsequent false positive incremental training, false negative incremental training and false positive and false negative joint incremental training can directly correspond to the edge-side machine learning model versions that generate false positive and false negative samples, forming an incremental training sample set.

[0030] S2. Based on the model version recorded in the incremental training sample set, the corresponding edge-side machine learning model is copied into three training models. Based on the independent low-rank adaptation parameters, the three training models are subjected to false positive incremental training, false negative incremental training, and joint false positive and false negative incremental training respectively, forming false positive candidate models, false negative candidate models, and joint candidate models. The error correction situation is compared with the original recognition preservation situation, and the candidate models that meet the preservation conditions are grouped into a candidate model set.

[0031] Among them, the error correction situation is the change in the identification of false positives and false negatives after the candidate model re-identifies false positives and false negatives compared to the edge-side machine learning model; the original identification retention situation is the situation in which the candidate model maintains the same correct identification results after the samples correctly identified by the edge-side machine learning model are re-entered into the candidate model; the candidate model that meets the retention conditions is the candidate model whose error correction situation meets the preset correction requirements and whose original identification retention situation meets the preset retention range lower limit. S2.1. Based on the model version recorded in the incremental training sample set, retrieve the model structure and model parameters of the edge-side machine learning model. Specifically, The model version identifier, jointly recorded by false positive and false negative samples, is read from the incremental training sample set. Using the model version identifier as the version retrieval index, version entries with the same model version identifier are located in the edge-side machine learning model version records stored in the cloud. Consistency verification is performed according to the parameter layer name, parameter layer order, and parameter tensor size recorded in the version entry. Version entries with mismatched model version identifiers or inconsistent parameter layer records are marked as version anomalies, and the corresponding versions are stopped from being copied into subsequent training instances. For version entries that pass the consistency verification, the parameter layer order and parameter layer connection relationship are restored according to the model structure file associated with the version entry. According to the parameter layer name and parameter tensor position recorded in the model parameter file, each parameter tensor is restored to its corresponding parameter layer. The number of parameter layers and parameter tensor size after restoration are checked again. Thus, the model structure and model parameters of the edge-side machine learning model corresponding to the model version recorded in the incremental training sample set are obtained.

[0032] Among them, the edge-side machine learning model version record is a version record that is saved synchronously when the edge-side machine learning model is published to the edge side. It is used to maintain the correspondence between the model version identifier, the model structure file and the model parameter file.

[0033] S2.2 Copy the model structure and parameters of the edge-side machine learning model to three training instances respectively, keeping the model structure and parameters consistent across the three training instances to form a training instance group with the same version. Specifically, The model structure of the edge-side machine learning model is copied to three independent training instances according to the original parameter layer arrangement. Model parameters corresponding to the same model version are written to the three training instances according to parameter layer name and parameter tensor position. Consistency checks are performed on each of the three training instances based on the number of parameter layers, parameter layer names, parameter layer order, parameter tensor size, and initial parameter values ​​corresponding to the edge-side machine learning model. If all three training instances are consistent with the edge-side machine learning model in the above checks, the corresponding training instance is considered to have passed the consistency check. For training instances with inconsistencies, the corresponding model structure and model parameters are copied again based on the inconsistent parameter layer positions, and consistency checks are performed again according to the same checks until all three training instances pass the consistency check. Independent parameter storage areas are configured for each of the three training instances that have passed the consistency check, ensuring that subsequent changes to training parameters in any training instance are only retained within that specific training instance and not written to the other two. The results of the three training instances having a consistent initial model structure and consistent initial model parameters are aggregated to form a training instance group with the same version.

[0034] S2.3. Configure independent low-rank adaptation parameters at the same trainable parameter positions in the same training instance group of the same version, fix the original model parameters of the three training instances, and separate the update paths of the independent low-rank adaptation parameters of the three training instances to form three training models. The three training models include a false positive training model, a false negative training model, and a joint false positive and false negative training model. Specifically, Within the same training instance group, locate the trainable parameter positions for feature mapping using matrix weights. Use positions where the inner layer names, layer numbers, and parameter tensor sizes of the three training instances are completely identical as independent low-rank adaptation parameter configuration positions. Set a dimension-reduced low-rank matrix and an increased low-rank matrix at each configuration position, so that the incremental parameter matrix formed by multiplying the dimension-reduced low-rank matrix and the increased low-rank matrix has the same matrix size as the corresponding original model parameters. The low-rank order is lower than the minimum value of the number of rows and columns of the corresponding original model parameter matrix. For example, the low-rank order is 8. The low-rank order is determined based on the size of the original model parameter matrix and the number of parameters allowed to participate in incremental training.

[0035] The original model parameters in the three training instances are set to a frozen state, so that the original model parameters continue to participate in feature mapping during forward recognition but do not participate in backward parameter updates. The incremental parameter matrix is ​​superimposed on the original feature mapping results of the corresponding parameter layer. The three training instances save independent low-rank adaptation parameters respectively. Parameter updates are only written to the independent low-rank adaptation parameters of the corresponding training instance. The three training instances form a false positive training model, a false negative training model, and a false positive / false negative joint training model according to their uses of false positive incremental training, false negative incremental training, and false positive / false negative joint incremental training.

[0036] Independent low-rank fitting parameters are a set of trainable parameters composed of a reduced-dimensional low-rank matrix and an increased-dimensional low-rank matrix. They are used to restrict parameter changes generated by incremental training to the low-rank parameter space while keeping the original model parameters of the three trained models unchanged.

[0037] S2.4. Separate false positive samples and false negative samples from the incremental training sample set. Assign false positive samples to the false positive training model, and false negative samples to the false negative training model. Assign both false positive and false negative samples to the joint false positive and false negative training model, forming a three-way incremental training relationship. Specifically, Training samples are distributed from the incremental training sample set according to the category relationship between false positive and false negative samples. The cloud verification results corresponding to the false positive and false negative samples are retained as supervision labels. False positive samples are written into the training sequence corresponding to the false positive training model, false negative samples are written into the training sequence corresponding to the false negative training model, and false positive and false negative samples are jointly written into the training sequence corresponding to the false positive and false negative joint training model, while keeping the original category labels of false positive and false negative samples unchanged.

[0038] The three training sequences are respectively established with the false positive training model, the false negative training model, and the joint false positive and false negative training model. The first iteration organizes the training batch according to the actual number of samples in the incremental training sample set. Subsequent iterations adjust the proportion of false positive samples and false negative samples entering the training batch according to the next round training weight formed by S4, forming a three-way incremental training relationship.

[0039] Before the results of the previous round of edge competition evaluation are formed in the first closed-loop iteration of the machine learning model, the training weights for false positive samples and false negative samples are given the same initial weights. Based on the relationship that the sum of the two types of training weights is 1, the training weights for false positive samples and false negative samples are set to 0.5 respectively, so that the first incremental training does not pre-bias the direction of false positives and false negatives. The cumulative requirements for false positive evidence and false negative evidence in the first false positive and false negative evidence accumulation conditions are given the same initial accumulation quantity. In one embodiment, both the cumulative requirements for false positive evidence and false negative evidence are set to 5, so that single identification changes and a small number of discrete identification changes do not directly trigger cloud verification and incremental model training, and that continuously generated abnormal identification results enter the corresponding processing after reaching the initial accumulation quantity. After the first edge competition evaluation is completed, the first training weights are replaced with the next round training weights formed in S4, and the first evidence accumulation requirements are replaced with the next round false positive and false negative evidence accumulation conditions formed in S4, so that subsequent iterations adjust the training emphasis and evidence accumulation requirements for the two types of samples based on the improvement of false positives and false negatives in the previous round.

[0040] S2.5. Freeze the original model parameters of the false alarm training model, and update the independent low-rank fitting parameters using false alarm samples to form a false alarm candidate model. Specifically, False positive samples are fed into the false positive training model in training batches. The cloud verification results corresponding to the false positive samples are used as supervision labels. The deviation between the recognition output of the false positive training model and the supervision label is written into the supervision loss function used in the original training task of the edge machine learning model. Backpropagation is used to form the parameter update amount corresponding to the false positive samples. The false positive training model keeps the original model parameters unchanged in the frozen state set in S2.3. The parameter update amount is only written into the independent low-rank adaptation parameters of the false positive training model, so that the recognition correction generated by the false positive samples is concentrated in the independent low-rank adaptation parameters.

[0041] After completing the training traversal of false positive samples in the incremental training sample set and reaching the unified training stopping condition, the updated independent low-rank adaptation parameters and the original model parameters that remain frozen are retained to form a false positive candidate model.

[0042] Among them, the independent low-rank adaptation parameters are the set of trainable parameters set at the same parameter layer position in the three training models and participating in the false positive incremental training, false negative incremental training and false positive and false negative joint incremental training respectively in the form of low-rank matrices. S2.6. Based on the missed sample, incrementally update the independent low-rank fitting parameters of the missed training model while maintaining the original model parameters unchanged, forming a missed candidate model. Specifically, Using the cloud verification results corresponding to the missed samples as the target recognition results, the missed samples are fed into the missed sample training model to form the current recognition output. The deviation between the current recognition output and the target recognition result is backpropagated along the supervised loss function used in the original training task of the edge side machine learning model, so that the parameter update caused by the missed recognition deviation only affects the independent low-rank fitting parameters of the missed sample training model. The original model parameters of the missed sample training model are kept frozen, and the independent low-rank fitting parameters of the missed sample training model continue to absorb the parameter correction direction corresponding to the missed samples. After completing the training traversal of missed samples and reaching the unified training stopping condition, the updated independent low-rank fitting parameters and the original model parameters kept frozen are saved to form the missed sample candidate model.

[0043] S2.7. In the joint training model for false positives and false negatives in the three-way incremental training relationship, independent low-rank adaptation parameter update directions are established based on the false positive samples and false negative samples respectively. The update directions of the two independent low-rank adaptation parameters are restricted to an approximately orthogonal low-rank update subspace. The independent low-rank adaptation parameters are updated according to the update directions of the two independent low-rank adaptation parameters to form a joint candidate model. Specifically, At the start of each round of joint incremental training for false positives and false negatives, the independent low-rank adaptation parameters saved by the joint training model before the start of this round of parameter updates are used as the starting point for parameters shared by both false positive and false negative samples, and the state of the independent low-rank adaptation parameters corresponding to the parameter update directions of the two types of samples is kept consistent. False positive samples are identified and backpropagated from the common parameter starting point in batches, and the gradient direction formed by the false positive samples acting on the independent low-rank adaptation parameters is recorded, and the corresponding gradient direction is determined as the update direction of the independent low-rank adaptation parameters corresponding to the false positive samples. Similarly, false negative samples are identified and backpropagated from the same common parameter starting point in batches, and the gradient direction formed by the false negative samples acting on the independent low-rank adaptation parameters is recorded, and the corresponding gradient direction is determined as the update direction of the independent low-rank adaptation parameters corresponding to the false negative samples. This ensures that the update directions of the two types of independent low-rank adaptation parameters are generated from the same independent low-rank adaptation parameter state before this round of parameter updates, avoiding the parameter changes generated by the previous type of samples altering the parameter basis used when calculating the parameter update direction of the subsequent type of samples.

[0044] The update direction of the independent low-rank adaptation parameters corresponding to the false positive samples is taken as the first orthogonal direction, and the first orthogonal direction is normalized. According to the projection relationship between the first orthogonal direction and the update direction of the independent low-rank adaptation parameters corresponding to the false negative samples, the update direction of the independent low-rank adaptation parameters corresponding to the false negative samples is projected onto the first orthogonal direction, and the corresponding projection component is subtracted from the update direction of the independent low-rank adaptation parameters corresponding to the false negative samples. Then, the remaining direction after subtracting the projection component is normalized to form the second orthogonal direction. The same process is carried out layer by layer along the parameter layer where the independent low-rank adaptation parameters are located, and the degree of overlap between the two update directions is measured by the absolute value of the inner product of the two normalized independent low-rank adaptation parameter update directions. The absolute value of the inner product of the two normalized independent low-rank adaptation parameter update directions is recorded as the approximate orthogonal evaluation quantity.

[0045] An approximate orthogonality determination threshold is set according to the approximate orthogonality evaluation quantity. In one embodiment, the approximate orthogonality determination threshold is set to 0.05. When the absolute value of the inner product of the update directions of the two normalized independent low-rank adaptation parameters is not higher than 0.05, it is determined that the update directions of the two independent low-rank adaptation parameters meet the approximate orthogonality requirement. At this time, the deviation of the included angle between the two normalized independent low-rank adaptation parameter update directions from 90° does not exceed about 2.9°. When the absolute value of the inner product is higher than 0.05, the residual projection component formed along the first orthogonal direction is further subtracted from the second orthogonal direction, and normalization is performed again until the absolute value of the inner product of the update directions of the two normalized independent low-rank adaptation parameters enters the approximate orthogonality determination threshold range. The first orthogonal direction and the second orthogonal direction that meet the approximate orthogonality requirement together form an approximate orthogonal low-rank update subspace. Here, 0.05 is used to give an implementation example of the approximate orthogonality determination threshold and is not a fixed value.

[0046] The residual projection components are further subtracted and renormalized for the second orthogonal direction. The absolute value of the inner product is written into the joint update direction after entering the approximate orthogonality determination threshold. This completes the continuous update of the training batches corresponding to false positive samples and false negative samples, forming a joint candidate model.

[0047] The formula for the joint update direction is expressed as follows: ; In the formula, Indicates the direction of joint updates. Indicates the first orthogonal direction. Indicates the second orthogonal direction. This represents the training weights corresponding to the false positive samples. This represents the training weights corresponding to the missed samples.

[0048] The training process of the three training models uses false positive and false negative samples from the incremental training sample set as training samples, and uses the cloud verification results generated in S1.3 as supervision labels. The supervision loss function adopts the original training loss function in the training configuration record saved when the edge machine learning model forms the current model version. For example, when the edge machine learning model undertakes the classification and recognition task, the cross-entropy loss function is used. The parameter update adopts the optimization method saved in the training configuration record, such as the AdamW optimization method. Backpropagation only updates the independent low-rank adaptation parameters. The unified training stopping condition is set according to the change of training loss in consecutive training rounds. For example, the update of independent low-rank adaptation parameters is stopped when the change of training loss relative to the previous round for three consecutive rounds is no more than one percent of the initial training loss. The value is determined to avoid premature stopping caused by fluctuations in a single round of training and to limit invalid repeated updates. The training configuration record is the training parameter record saved synchronously when the edge machine learning model forms the model version. It is used to save the original training loss function, optimization method and training parameter settings.

[0049] The approximately orthogonal low-rank update subspace is an adjustment that orthogonals the update directions of independent low-rank fitting parameters, and limits the update directions of two independent low-rank fitting parameters to the low-rank parameter update range where the inner product is close to zero.

[0050] It should be noted that, based on the false positive candidate model formed in S2.5, the false negative candidate model formed in S2.6, and the joint candidate model formed in S2.7, the candidate retention evaluation of the three candidate models is carried out separately. The original recognition samples corresponding to the current model version are selected from the edge running trajectory. False positive samples, false negative samples, and original correctly recognized samples are distinguished according to the cloud verification results. False positive samples and false negative samples are sent to the three candidate models for re-identification. The number of false positives and false negatives after re-identification by each candidate model is compared with the original number of false positives and false negatives formed by the edge-side machine learning model for the same samples. The error correction status of each candidate model is formed according to the degree of reduction in the number of false positives and false negatives. Then, the original correctly recognized samples are sent to the three candidate models for re-identification. The original recognition retention status is formed according to the proportion of the number of samples that each candidate model continues to retain as correctly recognized to the total number of original correctly recognized samples. The three candidate models are checked together by the error correction status and the original recognition retention status. The candidate models that meet the retention conditions are grouped according to their corresponding model names to form a candidate model set.

[0051] Based on the error correction and original identification retention status of each candidate model, the preset correction requirements and preset retention ranges are clearly defined. The error correction ratio is defined as the ratio of the reduction in the sum of false positives and false negatives after re-identification by the candidate model to the original sum of false positives and false negatives, based on the original number of errors in the edge-side machine learning model on the same false positive and false negative samples. The preset correction requirement is the minimum error correction ratio pre-set before the current round of candidate retention evaluation. The original identification retention ratio is defined as the ratio of the number of correctly identified samples that the candidate model continues to correctly identify to the total number of correctly identified samples, based on the original identification retention ratio pre-set before the current round of candidate retention evaluation. The three candidate models are then checked against the same preset correction requirements and preset retention ranges. The candidate model whose error correction ratio meets the preset correction requirements and whose original identification retention ratio meets the lower limit of the preset retention range is determined as the candidate model that meets the retention conditions, and both criteria remain unchanged during the same round of candidate retention evaluation.

[0052] After completing the candidate retention evaluation of the three candidate models, the number of candidate models in the candidate model set is checked. When at least one candidate model meets the retention condition, the bypass inference and edge competition evaluation in S3 are continued according to the formed candidate model set. When the false positive candidate model, the false negative candidate model, and the joint candidate model do not meet the retention condition, the current edge-side machine learning model continues to maintain the original model version, and the current round of candidate model bypass inference and model version replacement are not performed. The false positive and false negative evidence in the subsequent edge running trajectory are accumulated according to the current false positive and false negative evidence accumulation condition. After the subsequent false positive and false negative evidence reaches the corresponding accumulation requirement again, the incremental training and candidate retention evaluation of the three training models are performed again through the newly formed incremental training sample set, so that the current round of iteration without forming a candidate model set returns to the subsequent false positive and false negative evidence accumulation stage.

[0053] For processing positions where both S2 and S3 use the original recognition retention status for evaluation, the corresponding evaluation samples are distinguished according to the different purposes of candidate retention evaluation and edge competition evaluation. In the candidate retention evaluation of S2, the original correctly identified samples corresponding to the current model version and confirmed by cloud verification as correctly identified by the edge-side machine learning model are used as the evaluation objects. The retention rate of each candidate model after re-identifying the original correctly identified samples is used to determine whether incremental training affects the correct recognition ability already possessed by the edge-side machine learning model. In the edge competition evaluation of S3, the edge recognition samples that are continuously collected on the edge side after the candidate model set is formed and whose actual recognition status is confirmed by cloud verification are used as the evaluation objects. Bypass inference is used to compare the false alarm improvement, false alarm improvement and original recognition retention status of each candidate model under the current edge operating conditions. The original recognition retention status in S2 is used for candidate model retention screening, and the original recognition retention status in S3 is used for competition evaluation between candidate models.

[0054] S3. Perform bypass inference on the candidate model set, compare the improvement of false alarms and false negatives of each candidate model with the original recognition retention, calculate the competition results of each candidate model according to the improvement of false alarms, false negatives and the original recognition retention, and arrange the candidate models according to the competition results to form the marginal competition evaluation results.

[0055] S3.1. Perform bypass inference on the candidate model set, and record the recognition results of the edge-side machine learning model and each candidate model on the same edge recognition sample, forming a bypass inference record. Specifically, Each candidate model in the candidate model set is deployed at the bypass inference position corresponding to the edge-side machine learning model. The edge recognition samples formed by real-time collection on the edge side are synchronously sent to the edge-side machine learning model and each candidate model. The preprocessing method, input size, recognition parameter configuration and result judgment method of the edge recognition samples are kept consistent, so that the edge-side machine learning model and each candidate model face the same recognition object.

[0056] During bypass inference, only the recognition results of the edge-side machine learning model are used to carry out the current recognition output. The recognition results generated by each candidate model only participate in the competitive evaluation and do not change the current running state of the edge-side machine learning model. The edge recognition samples, the recognition results of the edge-side machine learning model, the recognition results of each candidate model and the corresponding candidate model name are recorded according to the collection time sequence of the edge recognition samples to form the bypass inference record.

[0057] Edge recognition samples are real-time sample sets continuously collected by edge-side acquisition devices during bypass inference of the candidate model set. They are used to compare the actual recognition performance of the edge-side machine learning model with that of each candidate model under the same recognition conditions.

[0058] S3.2. Compare the number of false positives and false negatives of each candidate model with the edge-side machine learning model in the side-channel inference record. Verify the number of correctly identified samples retained by each candidate model, and separately form the false positive improvement, false negative improvement, and original recognition retention status. These are then compiled into a candidate model evaluation record. Specifically, Align the edge-side machine learning model identification results with the identification results of each candidate model according to the same edge identification samples from the bypass inference record, and use the cloud verification method in S1.3 to confirm the actual identification status of the edge identification samples participating in the evaluation; edge identification samples whose edge-side machine learning model identification results are consistent with the cloud verification results are recorded as the original correct identification samples, and edge identification samples whose edge-side machine learning model identification results are inconsistent with the cloud verification results retain the original error category according to the verification results of false alarms and omissions.

[0059] After each candidate model generates a recognition result for the same edge recognition sample, the number of false positives and false negatives for each candidate model are summarized and compared with the number of false positives and false negatives for the edge-side machine learning model. The decrease in the number of false positives represents the improvement in false positives, and the decrease in the number of false negatives represents the improvement in false negatives. Results where the number of false positives and false negatives increases are retained as negative changes. The recognition results of each candidate model are compared one by one for the original correctly recognized samples, and the number of samples that still maintain the correct recognition result is summarized to form the original recognition retention status. The false positive improvement status, false negative improvement status and original recognition retention status are merged with the corresponding candidate models to form the candidate model evaluation record.

[0060] Among them, the actual recognition status of the edge recognition samples participating in the competitive evaluation is confirmed by the cloud according to the abnormal recognition section verification method already adopted in S1.3, which is used to uniformly determine the false alarm results, false negative results and correct recognition results generated by the edge side machine learning model and each candidate model.

[0061] like Figure 4 As shown, under the same edge recognition samples and the same recognition conditions, the edge-side machine learning model, false positive candidate model, false negative candidate model, and joint candidate model are each subjected to bypass inference. The actual recognition status confirmed by cloud verification is used as the unified verification basis. The number of false positives, the number of false negatives, and the number of samples that were originally correctly recognized are counted for each model. The improvement of false positives is determined by the ratio of the number of false positives reduced by the candidate model relative to the edge-side machine learning model to the original number of false positives of the edge-side machine learning model. The improvement of false negatives is determined by the ratio of the number of false negatives reduced by the candidate model relative to the edge-side machine learning model to the original number of false negatives of the edge-side machine learning model. The retention of original recognition is determined by the ratio of the number of original correctly recognized samples that were correctly recognized by the candidate model to the total number of original correctly recognized samples. The three results are recorded according to the same candidate model.

[0062] Based on the false positive candidate model formed in S2.5, the false negative candidate model formed in S2.6, and the joint candidate model formed in S2.7, side-channel inference is performed on the edge-side machine learning model and the three candidate models under the same edge recognition samples and the same recognition conditions. The number of false positives and false negatives for the edge-side machine learning model and the three candidate models are then counted based on the cloud verification results. Samples correctly identified by the edge-side machine learning model in the same edge recognition samples are defined as the original correctly identified samples, and the number of samples that the three candidate models continue to correctly identify in relation to the original correctly identified samples is counted. The edge-side machine learning model is used as a comparison. The baseline is defined as follows: False positive improvement is determined by the ratio of the number of false positives reduced by the candidate model compared to the edge-side machine learning model, to the original number of false positives in the edge-side machine learning model; false negative improvement is determined by the ratio of the number of false negatives reduced by the candidate model compared to the original number of false negatives in the edge-side machine learning model; and original recognition retention is determined by the ratio of the number of original correctly identified samples that the candidate model continues to correctly identify, to the total number of original correctly identified samples. When the edge-side machine learning model is used as the comparison baseline, both false positive and false negative improvement are recorded as 0%, and original recognition retention is recorded as 100%. The false positive improvement, false negative improvement, and original recognition retention for the three candidate models are aggregated according to their model correspondence to form a candidate model evaluation result comparison table, and then... Figure 4The evaluation results of the edge-side machine learning model and the three candidate models are visualized to show the differences between the false positive candidate model, the false negative candidate model, and the joint candidate model in different error correction directions and the ability to maintain the original recognition. This also provides an evaluation basis for the subsequent division of the non-dominated level of the candidate models, calculation of the competition distance, and ranking. The specific results are shown in the table below: Table 1: Comparison of Evaluation Results of Candidate Models

[0063] Combined with Table 1 and Figure 4 The evaluation results of the candidate models show that, under the same evaluation conditions, the false positive candidate model achieves a 69% improvement in false positives, higher than the 19% improvement in false negatives. This indicates that after incrementally updating the independent low-rank fitting parameters of the false positive training model using false positive samples, the resulting parameter correction mainly affects the false positive identification direction. Similarly, the false negative candidate model achieves a 67% improvement in false negatives, higher than the 20% improvement in false positives. This indicates that after incrementally updating the independent low-rank fitting parameters of the false negative training model using false negative samples, the resulting parameter correction mainly affects the false negative identification direction. The joint candidate model achieves a 74% improvement in false positives and a 71% improvement in false negatives, while maintaining 96% of the original recognition accuracy. This indicates that restricting the update directions of the independent low-rank fitting parameters corresponding to false positive samples and false negative samples to approximately orthogonal low-rank update subspaces reduces the overlap of the two independent low-rank fitting parameter update directions at the same low-rank parameter position. This maintains a high level of original recognition capability while allowing false positive and false negative corrections to work together on the joint candidate model. The edge-side machine learning model serves as the benchmark for comparison among the three evaluation criteria. Its false positive and false negative improvements are both 0%, and its original recognition is maintained at 100%. This is used to characterize the degree of error correction and the degree of retention of the original correct recognition results of the three candidate models relative to the current edge-side machine learning model.

[0064] Based on Table 1 and Figure 4 The three evaluation results are used to perform Pareto dominance discrimination on the false alarm candidate model, the missed alarm candidate model, and the joint candidate model according to the improvement of false alarm, the improvement of missed alarm, and the preservation of the original identification. The non-dominated level of the candidate model is formed based on the Pareto dominance discrimination results. The highest evaluation result corresponding to the three evaluation results in the non-dominated level of the candidate model is selected as the reference evaluation point. After normalizing the three evaluation results on the same scale, the competition distance of each candidate model relative to the reference evaluation point is calculated. Then, the candidate models are ranked according to the non-dominated level of the candidate model and the competition distance, so that the determination of the first candidate model in the ranking takes into account the false alarm correction ability, the missed alarm correction ability, and the degree of preservation of the original correct identification results.

[0065] The evaluation results in Table 1 are used to illustrate the relative changes between the edge-side machine learning model and different candidate models in three evaluation dimensions: improvement in false positives, improvement in false negatives, and retention of original recognition. The values ​​listed in Table 1 correspond to the evaluation results of the bypass inference in this embodiment. In specific applications, based on the actual number of false positives, the number of false negatives, and the number of original correctly recognized samples that the candidate model continues to maintain correct recognition obtained from the current bypass inference record, the corresponding improvement in false positives, improvement in false negatives, and retention of original recognition are calculated respectively, and the current candidate model evaluation record is formed based on the actual calculation results.

[0066] S3.3. Based on the evaluation records of candidate models, Pareto dominance is applied pairwise to each candidate model according to the improvement in false positives, the improvement in false negatives, and the preservation of the original identification. Candidate models not dominated by the remaining candidate models are listed as the first non-dominated layer. Pareto dominance is then applied to the remaining candidate models level by level to form the candidate model non-dominated layer. Specifically... Using the improvement in false positives, the improvement in false negatives, and the preservation of the original identification in the candidate model evaluation records as three simultaneously increasing evaluation dimensions, Pareto dominance is performed on each pair of candidate models in the candidate model set.

[0067] If candidate model A's evaluation results in all three evaluation dimensions are no lower than those of candidate model B, and at least one evaluation dimension is higher than that of candidate model B, then candidate model A has a Pareto dominance relationship with candidate model B. If any evaluation result in any of the three evaluation dimensions is lower than that of candidate model B, then candidate model A does not have a Pareto dominance relationship with candidate model B. This avoids using the improvement advantage of a single false positive to cover up the decline in the improvement of false negatives and the decline in the original recognition.

[0068] From the candidate models that have completed pairwise Pareto dominance judgment, candidate models that are not dominated by the remaining candidate models are selected and included in the first non-dominated layer. The candidate models corresponding to the first non-dominated layer are temporarily excluded from the candidate model evaluation record. The same Pareto dominance judgment is performed among the remaining candidate models, and the candidate models that are not dominated by the current remaining candidate models are included in the second non-dominated layer. The candidate models to be judged are reduced layer by layer in the same way until all candidate models in the candidate model evaluation record have a corresponding non-dominated layer number, forming the candidate model non-dominated layer.

[0069] For example, if the false positive candidate model has the highest improvement in false positives but a significantly lower improvement in false negatives than the joint candidate model, and the joint candidate model has the advantage in both the improvement in false negatives and the preservation of the original recognition but a lower improvement in false positives than the false positive candidate model, then the two can enter the first non-dominated layer simultaneously when there is no overall dominance relationship between them. If the false negative candidate model is not higher than the joint candidate model in all three evaluation dimensions and is lower than the joint candidate model in at least one evaluation dimension, then the false negative candidate model enters the subsequent non-dominated layer.

[0070] S3.4. In the non-dominated levels of the candidate models, the evaluation results with the highest improvement in false positives, the highest improvement in false negatives, and the highest preservation of the original identification are selected respectively. Specifically, The highest evaluation values ​​for false positive improvement, false negative improvement, and original identification retention are retrieved from the evaluation records of candidate models with non-dominated hierarchical association. The three highest evaluation values ​​can come from different candidate models and are recorded in a fixed dimensional order of false positive improvement, false negative improvement, and original identification retention. If multiple highest evaluation values ​​appear for the same evaluation dimension, only the highest evaluation value itself is retained, and no evaluation dimension is added repeatedly. The three highest evaluation results used for subsequent distance evaluation are obtained.

[0071] S3.5. Combine the three highest evaluation results into a reference evaluation point. Normalize the improvement in false positives, false negatives, and the preservation of the original identification on the same scale. Calculate the competition distance based on the Euclidean distance between each candidate model and the reference evaluation point. Rank the candidate models sequentially based on their non-dominated levels, and within the same candidate model's non-dominated level, arrange them in ascending order of competition distance to form the candidate model competition results. Specifically... The improvement in false positives, the improvement in false negatives, and the preservation of original identification in the candidate model evaluation records are respectively subjected to min-max normalization according to the highest and lowest evaluation values ​​of the corresponding evaluation dimensions, so that the three evaluation dimensions are transformed to the same numerical scale.

[0072] The normalization expression is: ; ; ; In the formula, Representing candidate models Improvement of false alarms Representing candidate models Improvement of underreporting, Representing candidate models The original identification status is maintained. , and These represent the normalized evaluation results for the corresponding evaluation dimensions. , and These represent the highest rating values ​​for the corresponding evaluation dimensions. , These represent the lowest rating values ​​for the corresponding evaluation dimensions.

[0073] When the highest and lowest evaluation values ​​of a certain evaluation dimension are the same, the corresponding evaluation dimension does not have the function of distinguishing candidate models, and the normalization results of all candidate models in the corresponding evaluation dimension are uniformly recorded as 1.

[0074] The three highest evaluation results obtained from S3.4 are combined into a reference evaluation point. After normalization to the same scale, the reference evaluation point is represented as 1, 1, 1. The competition distance is calculated according to the Euclidean distance between the three normalized evaluation results of each candidate model and the reference evaluation point. The expression for the competition distance is: ; In the formula, Representing candidate models The smaller the competition distance, the better the candidate model. The closer the results are to the reference evaluation point formed by the three highest evaluation results in the three evaluation dimensions of improvement in false alarms, improvement in underreporting, and maintenance of the original identification, the better.

[0075] The candidate models are ranked primarily based on their non-dominated levels, with the first non-dominated level preceding the second. Subsequent non-dominated levels are ranked in ascending order of their layer numbers. Within the same candidate model's non-dominated level, they are ranked in ascending order of their competition distance. Candidate models with the same competition distance are compared first in terms of their original recognition retention. If the original recognition retention is still the same, they are compared in terms of false positive improvement. If the false positive improvement is still the same, they are compared in terms of false negative improvement. If all three evaluation dimensions are completely identical, they are retained in parallel competitive positions, thus forming the candidate model competition results.

[0076] S3.6. Rank the candidate models according to the competition results to form the marginal competition evaluation results, specifically as follows: The candidate models in the candidate model competition results are arranged according to the non-dominated level and competition distance of the candidate models. The false alarm improvement, false alarm improvement, original identification retention, candidate model non-dominated level and competition distance are retained in the corresponding arrangement position. This allows S4 to directly read the corresponding candidate model and the corresponding false alarm improvement and false alarm improvement from the first position of the arrangement, forming the marginal competition evaluation result.

[0077] S4. Select the candidate model ranked first from the edge competition evaluation results. Adjust the training weights of false positive samples and false negative samples and the accumulation conditions of false positive and false negative evidence according to the improvement of false positives and false negatives of the candidate model ranked first. Replace the edge side machine learning model with the candidate model ranked first to form the closed-loop iteration result of the machine learning model.

[0078] S4.1 From the marginal competition evaluation results, select the top-ranked candidate model according to the order of candidate models, and retain the improvement in false positives and false negatives of the top-ranked candidate model to form an evaluation record of the top-ranked candidate model. Specifically, The candidate model ranking order is read from the edge competition evaluation results. The candidate model with the highest ranking position is taken as the first-ranked candidate model. At the same time, the false alarm improvement, false alarm improvement, original identification retention, non-dominated level and competition distance of the first-ranked candidate model are retained in the candidate model evaluation record. When the first-ranked candidate model has a tie in the competition position, the ranking result of the same layer formed in S3.5 is used to determine the ranking position. The competition relationship of the candidate models is not changed. A fixed correspondence is established between the first-ranked candidate model and the corresponding false alarm improvement and false alarm improvement, forming the evaluation record of the first-ranked candidate model.

[0079] S4.2. Based on the evaluation record of the top-ranked candidate model, adjust the training weights of false positive samples and false negative samples according to the improvement in false positives and false negatives, respectively, to form the training weights for the next round. Specifically, The improvement in false positives and false negatives in the evaluation record of the first candidate model is used as the basis for adjusting the training focus of the next round. The same-scale normalization result formed by S3.5 is used to map the part with insufficient improvement in false positives to the basis for increasing the training weight of false positive samples, and the part with insufficient improvement in false negatives to the basis for increasing the training weight of false negative samples, so that the error direction with lower improvement will get a higher training ratio in the next round of incremental training.

[0080] The training weights for false positives and false negatives are determined according to the following formula: ; ; in: ; In the formula, Indicates the training weights of the false positive samples in the next round. Indicates the training weights for the next round of missed samples. This indicates the improvement in false positives after normalization of the first-ranked candidate model. This indicates the improvement in false negatives after normalization of the first-ranked candidate model. This represents a positive smoothing factor used to avoid training weights to be zero and denominators to be zero.

[0081] S4.3. Based on the weights for the next round of training, adjust the accumulation requirements for false positive and false negative evidence in the accumulation conditions for false positive and false negative evidence to form the accumulation conditions for false positive and false negative evidence in the next round. The accumulation conditions for false positive and false negative evidence are that the accumulation of false positive and false negative evidence reaches the preset requirements for initiating cloud-based verification and incremental model training. Specifically, The requirements for accumulating false positive and false negative evidence are retrieved from the current conditions for accumulating false positive and false negative evidence. Specifically, the requirement for accumulating false positive evidence is determined based on the number of anomaly identification results corresponding to the false positive direction in the edge trajectory that are required to initiate cloud-based verification of the corresponding anomaly identification segment and subsequent incremental training of false positives. The requirement for accumulating false negative evidence is determined based on the number of anomaly identification results corresponding to the false negative direction in the edge trajectory that are required to initiate cloud-based verification of the corresponding anomaly identification segment and subsequent incremental training of false negatives. Based on the training weights of the next round, the training weights of the false positive samples and the false negative samples of the next round are compared with the corresponding training weights of the current round. The requirements for accumulating false positive and false negative evidence are adjusted according to the correspondence that the requirements for accumulating evidence decrease when the training weight increases and increase when the training weight decreases. This ensures that erroneous directions with lower improvement levels and increased training weights in the next round can enter cloud-based verification and subsequent incremental training of the model with a lower number of accumulated evidence.

[0082] The cumulative requirements for false positive evidence and the cumulative requirements for unreported evidence in the next round shall be adjusted according to the following formulas: ; ; In the formula, and These represent the cumulative requirements for false alarm evidence and the cumulative requirements for missed alarm evidence in the current round, respectively. and These represent the cumulative requirements for false positive evidence and the cumulative requirements for unreported evidence, respectively, for the next round. and These represent the training weights of the false positive samples and the false negative samples in the current round, respectively. and These represent the training weights for false positive samples and false negative samples in the next round, respectively. This indicates taking the nearest integer. This is used to ensure that the cumulative number of evidence is not less than 1.

[0083] The adjustment results are converted into cumulative integers of not less than 1, and written into the cumulative requirements for false positive evidence and the cumulative requirements for false negative evidence respectively. When the training weight of the false positive sample in the next round is higher than that of the current false positive sample, the cumulative requirement for false positive evidence in the next round is reduced accordingly. When the training weight of the false positive sample in the next round is lower than that of the current false positive sample, the cumulative requirement for false positive evidence in the next round is increased accordingly. The same correspondence is used for the direction of false negatives.

[0084] The requirement for accumulating false positive evidence is that the number of anomaly identification results corresponding to the false positive direction in the edge running trajectory reaches a certain level that can trigger cloud-based review and subsequent incremental training of false positives. This is used to control the timing of starting slow closed-loop processing in the false positive direction. The requirement for accumulating false negative evidence is that the number of anomaly identification results corresponding to the false negative direction in the edge running trajectory reaches a certain level that can trigger cloud-based review and subsequent incremental training of false negatives. This is used to control the timing of starting slow closed-loop processing in the false negative direction.

[0085] If the cumulative false alarm evidence reaches the requirement for the next round of false alarm evidence accumulation, or if the cumulative missed alarm evidence reaches the requirement for the next round of missed alarm evidence accumulation, the cloud-based review of the corresponding anomaly identification segment and subsequent incremental model training can be initiated once either error direction reaches the corresponding cumulative requirement. The two adjusted cumulative requirements will together form the conditions for the next round of false alarm and missed alarm evidence accumulation.

[0086] S4.4 Replace the edge-side machine learning model with the top-ranked candidate model, and incorporate the training weights for the next round and the accumulation conditions for false positives and false negatives in subsequent iterations to form a closed-loop iterative result for the machine learning model. Specifically, The replacement version is composed of the model structure corresponding to the first-ranked candidate model, the original model parameters, and the trained independent low-rank adaptation parameters. The replacement version is loaded on the edge side, and a compatibility check is performed according to the input size, recognition parameter configuration, and result output format currently used by the edge side machine learning model. If the model structure can be loaded normally and the recognition result can be returned in the original output format, the first-ranked candidate model is set as the new edge side machine learning model. If the model structure fails to load, the parameter size is inconsistent, or the recognition result cannot be returned in the original output format, the original edge side machine learning model is retained and not written into the model version replacement result.

[0087] The next round of training weights and the conditions for accumulating false positive and false negative evidence are linked to the model version identifier of the new edge-side machine learning model. This allows the new edge-side machine learning model to accumulate false positive and false negative evidence according to the conditions for accumulating false positive and false negative evidence in the next round after entering S1. In the subsequent incremental training in S2, the training ratio of false positive and false negative samples is configured according to the training weights of the next round. The new edge-side running results continue to enter the incremental training sample set, candidate model competition, and model version replacement, forming a closed-loop iterative result of the machine learning model.

[0088] This embodiment also provides a computer device applicable to the closed-loop iterative method of machine learning model based on edge-cloud collaboration, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the closed-loop iterative method of machine learning model based on edge-cloud collaboration as proposed in the above embodiment.

[0089] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0090] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the closed-loop iterative method for implementing an edge-cloud collaborative machine learning model as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0091] In summary, this invention improves the authenticity and training relevance of false positives and false negatives by organizing the recognition results generated by the continuous operation of the edge-side machine learning model in a temporal sequence and combining the results of anomaly recognition segments with cloud verification results to form an incremental training sample set. Based on the same model version, it implements incremental training for false positives, false negatives, and a combined incremental training of false positives and false negatives, respectively. By using independent low-rank adaptation parameters and approximately orthogonal low-rank update subspaces to limit parameter overlap between the two error correction directions, it reduces the disturbance to the original correct recognition capability caused by incremental training. By combining the improvement of false positives, the improvement of false negatives, and the preservation of the original recognition through side-path inference, it completes the competitive evaluation of candidate models, selecting the candidate model with the best overall performance from multiple candidate models. Simultaneously, it dynamically adjusts the training weights and the accumulation conditions for false positive and false negative evidence based on the actual improvement results of the ranked candidate model, creating a continuous feedback relationship between edge operation, cloud verification, incremental training, candidate evaluation, and version replacement. This improves the recognition stability, adaptability, and continuous iteration efficiency of the machine learning model in a long-term operating environment.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A closed-loop iterative method for machine learning models based on edge-cloud collaboration, characterized by: include, The identification results of the edge-side machine learning model are arranged in time sequence to form an edge running trajectory. Anomaly identification segments are collected from the edge running trajectory, and false positive and false negative evidence is accumulated. The anomaly identification segments are reviewed in the cloud, and the review results are divided into false positive samples and false negative samples to form an incremental training sample set. Based on the model version recorded in the incremental training sample set, the corresponding edge-side machine learning model is copied into three training models. Based on the independent low-rank adaptation parameters, the three training models are subjected to false positive incremental training, false negative incremental training, and joint false positive and false negative incremental training respectively, forming false positive candidate models, false negative candidate models, and joint candidate models. The error correction situation is compared with the original recognition preservation situation, and the candidate models that meet the preservation conditions are grouped into a candidate model set. Bypass inference is performed on the candidate model set to compare the improvement in false alarms and false alarms of each candidate model with the original recognition retention. The competition results of each candidate model are calculated according to the improvement in false alarms, false alarms and original recognition retention. The candidate models are ranked according to the competition results to form the marginal competition evaluation results. The candidate model ranked first in the edge competition evaluation results is selected. The training weights of false positive samples and false negative samples and the accumulation conditions of false positive and false negative evidence are adjusted according to the improvement of false positives and false negatives of the candidate model ranked first. The edge side machine learning model is replaced by the candidate model ranked first, forming a closed-loop iterative result of the machine learning model.

2. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 1, characterized in that: The formation of the incremental training sample set specifically involves... The recognition results of the machine learning model along the edge are arranged in the recognition time sequence, and the recognition results of continuous sampling positions are recorded in sequence to form the edge running trajectory; Anomaly identification segments whose identification results change continuously are collected from the edge running trajectory. The identification results within the anomaly identification segments are accumulated according to the direction of false alarm and the direction of missed alarm, forming evidence of false alarm and missed alarm. By reviewing the anomaly identification sections involved in false positives and false negatives in the cloud, and comparing the review results with the identification results of the edge-side machine learning model, false positive samples and false negative samples are obtained respectively. False positive and false negative samples are arranged according to the recognition time sequence. The arranged false positive and false negative samples are then aggregated to form an incremental training sample set.

3. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 2, characterized in that: Specifically, the corresponding edge-side machine learning model is replicated into three training models. Based on the model version recorded in the incremental training sample set, retrieve the model structure and model parameters of the edge-side machine learning model; The model structure and parameters of the edge-side machine learning model are copied to three training instances respectively, keeping the model structure and parameters of the three training instances consistent, forming a training instance group of the same version; Independent low-rank adaptation parameters are configured at the same trainable parameter positions in the same training instance group of the same version. The original model parameters of the three training instances are fixed, and the update paths of the independent low-rank adaptation parameters of the three training instances are separated to form three training models. The three training models include a false positive training model, a false negative training model, and a joint false positive and false negative training model.

4. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 3, characterized in that: The formation of the false positive candidate model, the false negative candidate model, and the joint candidate model specifically involves... False positive samples and false negative samples are separated from the incremental training sample set. False positive samples are assigned to the false positive training model, false negative samples are assigned to the false negative training model, and false positive samples and false negative samples are jointly assigned to the false positive and false negative joint training model, forming a three-way incremental training relationship. The original model parameters of the false alarm training model are frozen, and the independent low-rank fitting parameters are updated using false alarm samples to form a false alarm candidate model. Based on the missed sample, the independent low-rank fitting parameters of the missed training model are incrementally updated while keeping the original model parameters unchanged, thus forming a missed candidate model. In the joint training model of false positives and false negatives in the three-way incremental training relationship, independent low-rank adaptation parameter update directions are established according to the false positive samples and the false negative samples respectively. The two independent low-rank adaptation parameter update directions are restricted to an approximately orthogonal low-rank update subspace. The independent low-rank adaptation parameters are updated according to the two independent low-rank adaptation parameter update directions to form a joint candidate model.

5. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 4, characterized in that: The independent low-rank adaptation parameters are a set of trainable parameters set at the same parameter layer position in the three training models and participating in the false positive incremental training, false negative incremental training, and false positive and false negative joint incremental training in the form of a low-rank matrix. The approximately orthogonal low-rank update subspace is an adjustment that orthogonals the update directions of independent low-rank adaptation parameters, and limits the update directions of two independent low-rank adaptation parameters to the low-rank parameter update range where the inner product is close to zero.

6. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 5, characterized in that: The formation of the edge competition evaluation result is specifically as follows: Bypass inference is performed on the candidate model set, and the recognition results of the edge-side machine learning model and each candidate model on the same edge recognition sample are recorded separately to form a bypass inference record. The bypass inference involves deploying each candidate model in the candidate model set at the bypass inference position corresponding to the edge-side machine learning model, synchronously sending the same edge recognition samples into the edge-side machine learning model and each candidate model, and performing recognition on the same edge recognition samples while maintaining consistency in preprocessing method, input size, recognition parameter configuration and result judgment method, and recording the recognition results of the edge-side machine learning model and each candidate model according to the sampling time sequence. In the side-channel inference record, compare the number of false alarms and false negatives of each candidate model with the edge-side machine learning model, check the number of correctly identified samples retained by each candidate model, and form the false alarm improvement, false negative improvement and original identification retention respectively, and compile them into the candidate model evaluation record. Based on the evaluation records of candidate models, Pareto dominance judgment is performed on each candidate model in pairs according to the improvement of false alarms, the improvement of false negatives, and the preservation of the original identification. The candidate models that are not dominated by the remaining candidate models are listed as the first non-dominated layer. Pareto dominance judgment is performed on the remaining candidate models step by step to form the candidate model non-dominated layer. In the non-dominated levels of the candidate models, the evaluation results with the highest improvement in false positives, the highest improvement in false negatives, and the highest retention of the original identification were selected respectively. The three highest evaluation results are combined as reference evaluation points. The improvement of false alarms, the improvement of false alarms, and the preservation of the original identification are normalized on the same scale. The competition distance is calculated according to the Euclidean distance between each candidate model and the reference evaluation point. The candidate models are arranged in order based on the non-dominated level of the candidate models. Within the same non-dominated level of the candidate models, they are arranged in ascending order of competition distance to form the candidate model competition results. The candidate models are ranked according to their competition results, forming the marginal competition evaluation results.

7. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 6, characterized in that: The edge recognition samples are real-time sample sets collected at the edge and used by the edge-side machine learning model and each candidate model to perform bypass inference under the same recognition conditions.

8. The closed-loop iterative method for machine learning models based on edge-cloud collaboration as described in claim 7, characterized in that: The formation of the closed-loop iterative result of the machine learning model specifically refers to... From the edge competition evaluation results, the candidate model ranked first is selected according to the order of candidate models. The improvement of false alarms and the improvement of false alarms of the ranked candidate model are retained to form the evaluation record of the ranked candidate model. Based on the evaluation record of the first candidate model, the training weights of false positive samples and false negative samples are adjusted according to the improvement of false positives and the improvement of false negatives, respectively, to form the training weights for the next round. Based on the training weights for the next round, the cumulative requirements for false positive and false negative evidence in the cumulative conditions for false positive and false negative evidence are adjusted respectively to form the cumulative conditions for false positive and false negative evidence in the next round. The cumulative conditions for false positive and false negative evidence are that the cumulative situation of false positive and false negative evidence reaches the preset requirements for initiating cloud review and incremental training of the model. The edge-side machine learning model is replaced by the first-ranked candidate model, and the training weights for the next round and the cumulative conditions for false positives and false negatives are incorporated into subsequent iterations to form a closed-loop iterative result for the machine learning model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the closed-loop iterative method for the machine learning model based on edge-cloud collaboration as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the closed-loop iterative method for the machine learning model based on edge-cloud collaboration as described in any one of claims 1 to 8.

Images