Identification model training method and device, storage medium and computer program product

Abstract

The invention provides an identification model training method and device, a storage medium and a computer program product. The method comprises the steps of obtaining motion data of a first device and channel feature data of a second device; constructing a pseudo label information data set based on the motion data and the channel characteristic data; determining model updating parameters based on the pseudo label information data set, and updating parameters of the pre-training model by using the model updating parameters to obtain an updated model; and performing fusion processing on the updated model and the pre-training model to obtain an identification model which is used for identifying user behaviors.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a recognition model training method, device, storage medium and computer program product. Background Art

[0002] In the retail store scenario, in order to ensure that clothing can be stocked in a timely manner and orders can be placed in advance based on the supply chain situation, it is necessary to count the popularity of different goods in the store, and accurately identify customer behaviors such as clothing being flipped, picked up, tried on, and checked for prices through labels. The behavioral data of each new product can reflect the preferences for new products in advance, thereby automatically calculating the effectiveness of goods in different store formats and accurately matching goods.

[0003] However, in the actual use of related technologies, the deep learning model training scheme for store customer behavior recognition has inaccurate prediction results due to the lack of open source and mature data sets for model training, and the models trained by the data sets used are difficult to adapt to the actual scene changes in the environment. Summary of the invention

[0004] In view of this, the present application hopes to provide a recognition model training method, device, storage medium and computer program product that can improve the accuracy of prediction results.

[0005] In a first aspect, the present application provides a recognition model training method, the method comprising:

[0006] Acquire motion data of the first device and channel characteristic data of the second device;

[0007] Based on motion data and channel feature data, a pseudo-label information dataset is constructed;

[0008] Based on the pseudo-label information data set, determine the model update parameters, and use the model update parameters to update the parameters of the pre-trained model to obtain an updated model;

[0009] The updated model and the pre-trained model are fused to obtain a recognition model, which is used to recognize user behavior.

[0010] In a second aspect, the present application provides a recognition model training device, the device comprising:

[0011] An acquiring unit, configured to acquire motion data of the first device and channel characteristic data of the second device;

[0012] A construction unit, used to construct a pseudo label information data set based on motion data and channel feature data;

[0013] A determination unit, used to determine a model update parameter based on the pseudo-label information data set, and update the parameters of the pre-trained model using the model update parameter to obtain an updated model;

[0014] The processing unit is used to fuse the updated model and the pre-trained model to obtain a recognition model, and the recognition model is used to recognize user behavior.

[0015] In a third aspect, the present application provides a recognition model training device, the device comprising: a processor and a memory; when the processor executes the running program stored in the memory, the recognition model training method as described above is implemented.

[0016] In a fourth aspect, the present application provides a storage medium having a computer program stored thereon, which implements the above-mentioned recognition model training method when executed by a processor.

[0017] In a fifth aspect, the present application provides a computer program product, including a computer program, which implements the above-mentioned recognition model training method when executed by a processor.

[0018] The present application provides a recognition model training method, device, storage medium and computer program product, the method comprising: obtaining motion data of a first device and channel feature data of a second device; constructing a pseudo-label information data set based on the motion data and the channel feature data; determining a model update parameter based on the pseudo-label information data set, and using the model update parameter to update the parameters of a pre-trained model to obtain an updated model; fusing the updated model with the pre-trained model to obtain a recognition model, which is used to identify user behavior. Using the above implementation scheme, in the process of training the recognition model, the parameters of the pre-trained model are updated by collecting motion data and channel feature data in an actual scene, combined with the collected motion data and channel feature data, because the collected motion data and channel feature data are collected data in an actual scene, the updated model obtained is more accurate, and after obtaining the updated model, the updated model is further optimized by further fusing the updated model with the pre-trained model, so that the prediction result of the recognition model is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a passive Bluetooth tag device with an integrated inertial measurement unit (IMU);

[0020] Figure 2 A schematic diagram of the three-axis acceleration data of an IMU sensor under different situations (such as no interference, trying on clothes, and clothes turning over) collected by an IMU;

[0021] Figure 3 A schematic diagram of RFID signal phase data under different situations of RFID collection (i.e., no interference, trying on clothes, and turning over clothes);

[0022] Figure 4 A flowchart of the training and reasoning phases of a deep learning model;

[0023] Figure 5 A flow chart of a recognition model training method provided in an embodiment of the present application;

[0024] Figure 6 A schematic diagram of a process for obtaining channel characteristic data and motion data and preprocessing the channel characteristic data and motion data provided in an embodiment of the present application;

[0025] Figure 7 A schematic diagram of a process for screening a pseudo-label information data set provided in an embodiment of the present application;

[0026] Figure 8 A schematic diagram of the structure of a decision tree provided in an embodiment of the present application;

[0027] Fig. 9 A schematic diagram of a training process of an initial pre-training model provided in an embodiment of the present application;

[0028] Fig.10 A schematic diagram of a process for preprocessing an RFID time series data set used in training an initial pre-training model in an embodiment of the present application;

[0029] Fig.11 A schematic diagram of a process for obtaining an updated model provided in an embodiment of the present application;

[0030] Fig.12 A schematic diagram of a process for iteratively optimizing a model provided in an embodiment of the present application;

[0031] Fig.13 A physical diagram of a tag device hardware used in a training phase provided in an embodiment of the present application;

[0032] Fig.14 A schematic diagram of the hardware architecture of a tag device hardware used in a training phase provided in an embodiment of the present application;

[0033] Fig.15 A schematic diagram of an architecture using common labels in the inference stage provided for an embodiment of the present application;

[0034] Fig.16 A schematic diagram of a main system end system architecture provided in an embodiment of the present application;

[0035] Fig.17 A schematic diagram of the overall process of a recognition model training phase and a reasoning phase provided in an embodiment of the present application;

[0036] Fig.18 A schematic diagram of actual collected motion data provided by an embodiment of the present application;

[0037] Fig.19 A schematic diagram of channel characteristic data actually collected according to an embodiment of the present application;

[0038] Fig. 20 A schematic diagram of the structure of a recognition model training device provided in an embodiment of the present application Figure 1 ;

[0039] Fig.21 A schematic diagram of the structure of a recognition model training device provided in an embodiment of the present application Figure 2 . DETAILED DESCRIPTION

[0040] In order to enable a more detailed understanding of the features and technical contents of the embodiments of the present application, the technical solution of the present application is further elaborated in detail below in combination with the drawings and specific embodiments of the specification. The attached drawings are for reference only and are not used to limit the embodiments of the present application.

[0041] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as those commonly understood by those skilled in the art to which the present application belongs. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0042] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict. It should also be noted that the terms "first / second / third" involved in the embodiments of the present application are only used to distinguish similar objects and do not represent a specific ordering of the objects. It is understandable that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0043] In the retail store scenario, in order to ensure that clothing can be stocked in a timely manner and to make advance orders based on the supply chain situation, it is necessary to count the popularity of different goods in the store, and accurately identify customer behaviors such as clothing being turned over, picked up, tried on, and checked for prices through labels. The behavioral data of each new product can reflect the customer's preference for the new product in advance, thereby automatically calculating the effectiveness of goods in different stores and accurately matching goods.

[0044] In one approach, using Figure 1 The passive Bluetooth tag device with integrated IMU shown in the figure mainly refers to the accelerometer, which can measure the three-axis motion acceleration of an object. The three-axis acceleration data of the IMU sensor under different situations (such as no interference, trying on clothes, and clothes turning over) collected by IMU are as follows: Figure 2 As shown. However, due to the high cost of embedding tags into sensors, market acceptance is generally not high. Compared with wireless signal sensing methods such as Wi-Fi and Bluetooth, the advantage of Radio Frequency Identification (RFID) is that RFID tags have the characteristics of large-scale deployment. RFID tag arrays form multiple receiving and transmitting paths in space through backscattering, which can obtain more fine-grained spatial features and have stronger anti-interference capabilities. In the smart store scenario, the use of RFID tags is low-cost. By deploying item tags on all shelves and warehouses, it can not only realize automatic inventory and positioning of goods, but also identify the behavior of store customers.

[0045] RFID-based wireless sensing identifies the sensing target through channel feature analysis. Due to the complexity of the actual store scene, situations such as pedestrians walking and metal obstructions will cause signal interference. Usually, a dual-tag method of sensing tags and reference tags is used to eliminate multipath interference. Generally, the two tags are deployed close to each other, one of which is bound to the object and the other is placed in the environment, which makes it difficult to implement actual store deployment. The single-tag method will make the rule-based machine learning method less accurate, so the deep supervision method is mainly used in the implementation plan. For each scenario, it is first necessary to collect a large amount of RFID channel feature data, such as received signal strength indication (RSSI), signal phase, Doppler frequency shift data, etc. Among them, the RFID signal phase data collected by RFID in three situations (i.e., no interference, trying on clothes, and turning over clothes) is as follows: Figure 3 As shown in the figure, it can be seen that the phase difference between trying on clothes and turning clothes is small, and feature extraction is more difficult. After collecting a large amount of RFID channel feature data, it can be used as follows Figure 4 The flowchart shown implements the training set application of the deep learning model. Figure 4Specific implementation process: Clean, preprocess and annotate the data to generate a labeled training data set. Then build a deep learning model based on time series, input a large number of labeled data sets into the model for training, output the estimated behavior category of the sequence (such as disturbance behavior, flipping behavior, try-on behavior, etc.), back propagate until the model converges, and then in the actual scenario (i.e., the inference stage), collect the RFID channel feature data, input it into the trained deep learning model to infer the disturbance behavior, flipping behavior, and try-on behavior.

[0046] Based on the above implementation, when using RFID technology to identify store customer behavior, the following problems still exist in the actual training of the deep learning model:

[0047] (1) There is currently no open source, mature data set for model training. In actual scenarios, the collection and labeling of training data will consume a lot of manpower and the initial investment cost is high.

[0048] (2) The collected training data is difficult to cover all the changing conditions of actual scenarios. RFID channel estimation is easily affected by ambient noise and interference, resulting in phase jumps and other situations. It is difficult to extract relevant behavioral features, which seriously affects the results of model reasoning. Therefore, the model training phase needs to run in the actual scenario for a period of time, so that the model needs to be continuously updated and optimized, which involves incremental data cleaning, preprocessing and labeling. In short, the above technical solutions are difficult to quickly adapt to changes in the new environment.

[0049] To solve the above problems, the present application provides a recognition model training method, such as Figure 5 As shown, the method may include:

[0050] S501. Obtain motion data of a first device and channel characteristic data of a second device.

[0051] In the embodiment of the present application, the first device is a micro electromechanical system inertial measurement unit (MEMSIMU) sensor.

[0052] In the embodiment of the present application, the second device is an RFID antenna.

[0053] In the embodiment of the present application, the motion data includes x, y, and z three-axis acceleration data collected by the MEMS IMU sensor.

[0054] In the embodiment of the present application, the channel characteristic data includes RSSI, phase data, Doppler frequency shift data, etc.

[0055] In an embodiment of the present application, during the training of the model for identifying store customer behavior, the first device and the second device are integrated into the tag hardware device, that is, the RFID antenna and the MEMSIMU sensor are integrated in the tag hardware device, so the RSSI, phase data, Doppler frequency shift data, etc. of the RFID antenna and the x, y, and z three-axis acceleration data of the MEMSIMU sensor can be obtained at the same time, thereby realizing online data collection of the two types of data.

[0056] In the embodiment of the present application, for the collection of motion data and channel characteristic data, a RFID tag hardware device with an integrated IMU sensor can be installed on a piece of clothing in an actual store scene, and the three-axis acceleration data of the IMU sensor and the RFID multi-dimensional data stream can be collected and reported through the tag hardware device. Specifically, the RFID channel characteristic data and the IMU motion data can be obtained by parsing protocols such as Message Queuing Telemetry Transport (MQTT).

[0057] In an embodiment of the present application, after the motion data and channel characteristic data are acquired, a preprocessing operation may be performed on the acquired motion data and channel characteristic data.

[0058] In an embodiment of the present application, for channel characteristic data, including RSSI, phase data, and Doppler frequency shift data, when performing preprocessing operations, the channel characteristic data is first cleaned, including processing missing data values, outliers, etc. Secondly, since the phase data is easily affected by multipath effects and other wireless signal interference, phase flipping is prone to occur, and phase flipping processing (also known as phase correction processing) is required; at the same time, since multiple groups of time series data are collected, and the length of multiple groups of time series data is not fixed, it is necessary to perform data alignment, segmentation and other operations to convert the data into a form that can be processed by the model.

[0059] In an embodiment of the present application, for the acquired motion data, a moving average filter is first performed on the motion data with a window of fixed length, that is, the average value of the motion data in the window is taken as the filtered result; then the data is aligned and split, etc.

[0060] In an embodiment of the present application, after preprocessing the channel characteristic data and the motion data, the preprocessed channel characteristic data and the preprocessed motion data can also be merged and packaged, that is, data containing both the channel characteristic data and the motion data can be obtained.

[0061] In the embodiment of the present application, the above-mentioned operations of acquiring channel characteristic data and motion data in real time and preprocessing the channel characteristic data and motion data can be implemented by a real-time acquisition module. The specific implementation process of the real-time acquisition module can be as follows: Figure 6 As shown, the specific steps include:

[0062] The relevant processing process of channel characteristic data is as follows:

[0063] S11, protocol analysis, obtaining channel characteristic data collected by RFID.

[0064] S12. Processing of missing data values ​​and outliers.

[0065] S13, phase flip.

[0066] S14. Data alignment and segmentation.

[0067] The relevant processing process of motion data is as follows:

[0068] S21. Protocol analysis to obtain MEMSIMU motion data.

[0069] S22, performing moving average filtering on the motion data using a window of fixed length.

[0070] S23. Data alignment and segmentation.

[0071] The process of merging and packaging the preprocessed motion data and the preprocessed channel feature data is as follows:

[0072] S3: Data packaging and processing.

[0073] S502: Construct a pseudo-label information data set based on the motion data and the channel characteristic data.

[0074] In an embodiment of the present application, a pseudo-label information data set is constructed based on motion data and channel characteristic data, which can be to determine the category confidence of the behavior category corresponding to the motion data; perform data conversion processing on the category confidence to obtain category soft label information; determine the credibility of the pseudo-label information based on the category soft label information and the category confidence; determine the pseudo-label information data set based on the pseudo-label information credibility, a preset pseudo-label screening threshold and channel characteristic data.

[0075] In the embodiment of the present application, the data conversion processing includes normalization processing, one-hot encoding conversion processing and Gaussian space conversion processing.

[0076] In an embodiment of the present application, the category confidence of the behavior category corresponding to the motion data can be determined using the motion data and a preset target algorithm, wherein the preset target algorithm can be a random forest algorithm, that is, the random forest algorithm can be used to calculate the category confidence.

[0077] In the embodiment of the present application, because the data set usually used in the laboratory does not completely contain all the actual situations, it is necessary to screen out a large amount of special data such as multiple people passing through interference or metal occlusion in the actual scene. Since the actual collected data is not labeled, the characteristics of the MEMSIMU sensor that is not affected by the environment can be used to preliminarily judge customer behavior based on machine learning methods, and then obtain the label information and credibility of the pseudo-label data based on uncertainty estimation.

[0078] In the embodiment of the present application, this part of the content can be processed by the data screening and automatic annotation module, and the corresponding workflow is as follows: Figure 7 As shown, the specific steps include:

[0079] S1. Random forest construction.

[0080] S2. Use random forest to calculate category confidence.

[0081] S3. Normalize the category confidence.

[0082] S4. Perform one-hot encoding conversion on the normalized category confidence.

[0083] S5. Perform Gaussian distribution space transformation on the one-hot encoding to obtain category soft label information.

[0084] S6. Determine the pseudo-label dataset.

[0085] Among them, the specific implementation steps of S1 are as follows:

[0086] In the embodiment of the present application, during the construction of the random forest, the best features are gradually selected and the data set is divided into subsets, and subtrees are recursively constructed until all instances in the subset belong to the same category. Repeat the above steps to construct N decision trees to ensure that the training set of each tree is randomly and with replacement extracted from the original training set, N = 1, 2, 3, ..., 100. Multiple decision trees are combined into a random forest.

[0087] For the classification problem in the embodiments of the present application, the output result of the random forest is usually obtained by voting.

[0088] In the embodiment of the present application, since the three-axis acceleration data of MEMSIMU is not easily disturbed by the external environment, the features selected in the embodiment of the present application are the three-axis acceleration values ​​and the threshold and direction of the data difference of the previous and next time frames, such as acc_x represents the acceleration in the x-axis direction, and diff_y represents the difference between two consecutive sampling results in the y-axis direction. Optionally, N complete decision trees (N=4) constituting the random forest are as follows: Figure 8 shown.

[0089] exist Figure 8 In the above, there are 4 decision trees. In decision tree 1, the y-axis acceleration difference diff_y>0.3 is used as the classification condition for step-by-step judgment. In decision tree 2, the y-axis acceleration difference diff_y>0.45 is used as the classification condition for step-by-step judgment. In decision tree 3, the x-axis acceleration acc_x<0.2 and the y-axis acceleration acc_y<0.25 are used as the classification conditions for step-by-step judgment. In decision tree 4, the x-axis acceleration acc_x<0.3 and the y-axis acceleration acc_y<0.3 are used as the classification conditions for step-by-step judgment.

[0090] In the embodiment of the present application, the specific implementation steps of S2 are as follows:

[0091] In the embodiment of the present application, the decision tree is used as a rule-based machine learning method. The node satisfies the association rule, that is, IF Z then X. The confidence of a specific node i in a decision tree is the ratio of the number of samples that simultaneously meet condition Z and condition Y to the number of samples that meet condition Z, that is, The meaning expressed is that the confidence calculation formula of the leaf node of the output A category can be calculated using the following formula (1), and the specific formula (1) is as follows:

[0092]

[0093] Wherein, f represents the number of samples that meet the corresponding conditions, N is the total number of samples, and p is the probability function of a small sample size. Unlike simple frequency-based estimates, more robust estimates are used for the probability functions p(Z∪Y) and p(Z|Y) of a small sample size. Since the association rules are used in applications with large-scale data sets in the embodiments of the present application, in this case, a simple frequency-based estimate of the probability function can be used.

[0094] In the embodiment of the present application, for the entire random forest, the confidence calculation of disturbance, picking up, trying on, and no behavior represented by each category A, B, C, and D is related to the category labels output by all decision trees. The category labels output by all decision trees are counted, and the number of votes is calculated for each category. The proportion of votes is regarded as a measure of category confidence. Assuming that 0, 1, ..., i trees output the classification result of the sample (i.e., the input motion data) as A, then the category confidence formula for the sample is category A as shown in formula (2):

[0095]

[0096] In the above formula (2), the numerator represents the confidence of tree 0, 1, ..., i that outputs the classification result of the sample as A. The denominator represents the accumulation of confidences of all trees N that output the classification results of the sample. This formula can be regarded as the precision of the evaluation index of the machine learning model (which can be expressed as Precision in English). However, it should be noted that the model classification result is not necessarily A. Here, A is taken as the actual classification result as an example.

[0097] By analogy, the category confidences corresponding to category B, category C, category D, etc. can also be calculated.

[0098] In the embodiment of the present application, a series of data conversion processes are performed on the obtained category confidence, that is, the following are performed in sequence: S3, normalizing the category confidence. S4, performing one-hot encoding conversion operation processing on the normalized category confidence. S5, performing Gaussian distribution space conversion on the one-hot encoding to obtain category soft label information.

[0099] In the embodiment of the present application, the specific implementation steps of S3 are as follows:

[0100] The confidence data of different categories are normalized to ensure that the sum of the confidence data of each category is 1, so as to further generate one-hot encoding. Taking the normalization operation of the category confidence corresponding to category A as an example, the category confidence corresponding to category A after normalization operation can be shown in the following formula (3):

[0101]

[0102] Among them, Conf A is the category confidence of category A, Conf B is the category confidence of category A, Conf C is the category confidence of category A, Conf D is the category confidence of category A.

[0103] In the embodiment of the present application, the above-mentioned method of normalizing the category confidence of category A is adopted to perform normalization operations on the category confidences corresponding to category B, category C, and category D, respectively, to obtain the category confidences corresponding to each category after normalization. In this way, the four category confidences in the initially acquired online data set (i.e., motion data) can be obtained.

[0104] In the embodiment of the present application, the specific implementation steps of S4 are as follows:

[0105] The normalized category confidence sequence is converted into one-hot encoding to complete the automatic data labeling. The calculation formula of one-hot encoding is shown in the following formula (4):

[0106]

[0107] Among them, Conf_norm i is the normalized category confidence corresponding to different categories (category i), and θ represents other parameters.

[0108] In the embodiment of the present application, the specific implementation steps of S5 are as follows:

[0109] The one-hot encoding obtained above is transformed into Gaussian distribution space to obtain category soft label information.

[0110] In the embodiment of the present application, when determining the pseudo-label data set, the pseudo-label information credibility can be determined based on the obtained category soft label information and the obtained category confidence. In the specific implementation process, the similarity between the category soft label information and the category confidence can be determined; and the similarity is determined as the pseudo-label information credibility.

[0111] In the embodiment of the present application, the cosine similarity between the category soft label information and the category confidence is first calculated, and the cosine similarity is used as the credibility of the pseudo label information. The specific calculation formula for calculating the cosine similarity between the category soft label information and the category confidence can be shown in the following formula (5):

[0112]

[0113] Among them, l represents the category soft label information, and f represents the category confidence.

[0114] In an embodiment of the present application, after obtaining the credibility of the pseudo label information, a pseudo label information data set is determined based on the credibility of the pseudo label information, a preset pseudo label screening threshold and the channel characteristic data. The specific implementation process is that when the credibility of the pseudo label information is less than the preset pseudo label screening threshold, the target channel characteristic data corresponding to the category soft label information is determined from the channel characteristic data; the target channel characteristic data and the corresponding category soft label information are determined as the pseudo label information data set.

[0115] In the embodiment of the present application, the preset pseudo-label screening threshold can be selected according to actual conditions, and is not specifically limited in the embodiment of the present application.

[0116] In an embodiment of the present application, by judging the size between the credibility of the pseudo tag information and the preset pseudo tag screening threshold, it is judged whether the credibility of the pseudo tag information is less than the preset pseudo tag screening threshold. If so, the corresponding RFID channel characteristic data and soft tag information are screened and filtered from the latest collected online data set to obtain a pseudo tag information data set.

[0117] S503: Based on the pseudo-label information data set, determine the model update parameters, and use the model update parameters to update the parameters of the pre-trained model to obtain an updated model.

[0118] In the embodiment of the present application, this step can be implemented by a model updating module.

[0119] In an embodiment of the present application, the model update parameters may be gradient and weight parameters.

[0120] In an embodiment of the present application, the pre-trained model may be an initial pre-trained model at the initial stage, and the initial pre-trained model is obtained by training a pre-training module using a small amount of laboratory data.

[0121] The specific training process of the initial pre-training model is as follows Fig. 9 As shown, the specific steps include:

[0122] S1. Data preparation.

[0123] S2. Manual labeling.

[0124] S3. Model construction.

[0125] S4. Model training.

[0126] The specific implementation process of S1 is as follows:

[0127] Prepare a small amount of RFID time series data sets in advance, including RSSI, phase data, and Doppler frequency shift data (calculated). Fig.10 The process shown implements the preprocessing operation of the data, specifically including:

[0128] A11. First, clean the RFID collected data, including processing missing values, outliers, etc.

[0129] A12. Secondly, since phase data is affected by multipath effects and interference from other wireless signals, phase flipping is prone to occur, and phase correction processing is required.

[0130] A13. Since multiple sets of time series data are of varying lengths, data alignment is required.

[0131] And A14 data segmentation, converted into a form that the model can process.

[0132] A15. Finally, the data is packaged and divided into training data set and verification data set.

[0133] The specific implementation process of S2 is as follows:

[0134] Each segment of data is manually labeled to a category, such as disturbance (which can be expressed as disturbance in English), pick up (which can be expressed as pick in English), try on (which can be expressed as dress in English), and no behavior (which can be expressed as none in English), so as to obtain an initial label set, which includes but is not limited to timestamp intervals, behavior categories, etc.

[0135] The specific implementation process of S3 is as follows:

[0136] After data preprocessing, a deep learning model is designed to process time series data, which usually consists of multiple layers, including input layer, hidden layer and output layer. It is necessary to choose a network structure suitable for time series data, such as recurrent neural network (RNN), long short-term memory network (LSTM) or multi-head attention mechanism.

[0137] The specific implementation process of S4 is as follows:

[0138] Model training minimizes the loss function by adjusting the model parameters to improve the prediction ability of the model. In the embodiment of the present application, a weighted cross entropy loss function for RFID store customer behavior recognition is designed according to the business's attention to each category, as shown in the following formula (6):

[0139]

[0140] Where m is the number of batch samples, j = 1, 2, 3, 4, corresponding to none (i.e., no behavior category), disturbance (i.e., disturbance category), pick (i.e., pick up category), dress (i.e., try on category), and the weight relationship of the sub-items is ω none ≤ω disturb ≤ω pick ≤ω dress In this business, we pay more attention to the conversion rate from try-on behavior to sales, and the probability of this category is lower than that of other categories, so the weight of this category is set to the maximum. After a certain number of N rounds of iterative training, the training is completed, and the model parameters are saved to a file to obtain the initial pre-trained model, so that it can be loaded in the subsequent training process.

[0141] In an embodiment of the present application, when updating the model, the model update module trains the online model based on the screened pseudo-label information data set, calculates the loss function based on the pseudo-label information and the forward propagation output result, and the loss function calculation is shown in formula (6). The gradient and online model weight parameters are updated by backpropagation, and the final updated model is obtained after a certain number of iterations.

[0142] In the embodiment of the present application, the main implementation process of obtaining the updated model is as follows: Fig.11As shown, the specific steps include:

[0143] S1. Input a batch of pseudo-label data (i.e., data in the selected pseudo-label data set) into the online model. The initial online model is the initial pre-trained model mentioned above.

[0144] S2. Get the output result through forward propagation.

[0145] S3. Calculate the category weighted loss function according to the custom formula (6) and modify the model update parameters.

[0146] S4, back propagation, update the gradient and weight parameters of the online model.

[0147] Return to step S1, input a new batch of pseudo-labeled data, and repeat the calculations of S2-S4.

[0148] Sn, until a certain number of iterative training is reached, and the updated model is obtained.

[0149] S504: Fusing the updated model with the pre-trained model to obtain a recognition model, which is used to recognize user behavior.

[0150] In an embodiment of the present application, after obtaining the updated model, in order to ensure that the updated model has higher accuracy, it is necessary to fuse the updated model and the pre-trained model to obtain the final recognition model.

[0151] Specifically, a first weight value corresponding to the updated model and a second weight value corresponding to the pre-trained model are obtained; and a recognition model is determined based on the first weight value, the second weight value, the updated model and the pre-trained model.

[0152] In the embodiment of the present application, this process can be processed by a model tuning module.

[0153] In an embodiment of the present application, the identification model may be determined by iterating based on the last online model and the updated model obtained after updating the parameters to obtain an optimized identification model.

[0154] In the embodiment of the present application, the last online model here may be an initial pre-trained model.

[0155] In the embodiment of the present application, the first weight value 1-β of the previous online model is set, and the second weight value β of the updated model is set.

[0156] In the embodiment of the present application, a weighted moving average algorithm is used to calculate the optimized online model (i.e., recognition model) as shown in the following formula (7):

[0157] M'cur =βM new +(1-β)M cur (7)

[0158] Among them, M / cur represents the iteratively optimized online model (i.e., recognition model), M cur Represents the online model before update (previous step). The online model in the first iteration is the initial pre-trained model, i.e., M / cur =M0.

[0159] In an embodiment of the present application, after obtaining the recognition model, the obtained recognition model can be used to identify the behavior of store customers.

[0160] It can be understood that in a recognition model training method provided in an embodiment of the present application, in the process of training the recognition model, the parameters of the pre-trained model are updated by collecting motion data and channel feature data in the actual scene and combining the collected motion data and channel feature data. Because the collected motion data and channel feature data are data collected in the actual scene, the updated model obtained is more accurate. After obtaining the updated model, the updated model is further optimized by further fusing the updated model with the pre-trained model. In this way, the prediction results of the recognition model are more accurate.

[0161] In the embodiment of the present application, the iterative optimization process includes a cyclic interaction between automatic data annotation and updating and tuning of the online deep model, such as Fig.12 As shown, the specific process includes:

[0162] E1-E3. After the initial data is automatically labeled, the first round of online deep model updating and tuning is performed.

[0163] E4-E6. After the updated model is tuned, return to perform the second round of automatic data labeling, and then perform the second round of online deep model update and tuning.

[0164] En, and so on, iterate a certain number of times until the i-th round of data automatic labeling and the i-th round of online deep model updating and tuning are completed, and the tuned recognition model is obtained.

[0165] After each round of tuning of the updated model is completed, a new round of training will be carried out based on the tuned online model to form a closed loop until the conditions for iteration termination are met. This condition is set as when the model reaches the preset accuracy requirements and shows a convergence trend, the iteration process will terminate and the final recognition model will be output, thereby realizing the self-evolution of the recognition model.

[0166] In one embodiment of the present application, after obtaining the recognition model, in the process of using the recognition model to predict user behavior, channel feature data at the first moment can be collected; based on the recognition model, the user behavior corresponding to the channel feature data at the first moment is identified.

[0167] In the embodiment of the present application, the first moment is any moment in the real-time acquisition process.

[0168] In the embodiment of the present application, in the actual application process, the recognition model can only use the channel characteristic data collected by RFID to identify the user behavior. That is, the channel characteristic data collected in real time is input into the trained recognition model, and the recognition model is used to identify the user behavior.

[0169] Based on the above embodiments, a tag device hardware used in the training phase is provided in the embodiments of the present application. In the actual training phase, an RFID tag device with an integrated MEMS IMU sensor is used for training, including a vibration sensing module MP213, a storage module, an antenna, a filtering network, a microcontroller unit (MCU) and a communication module, a power module, an energy storage and management module. Radio frequency energy collection and storage can be achieved, triggering the switching of the back-end chip from the sleep mode to the start-up mode. According to the rhythm of model training and model tuning, the RFID tag transmits back the IMU sensor data and RFID excitation signal at a specific frequency, and the data acquisition module on the main system side can pre-process and analyze the data.

[0170] Specifically, the tag device hardware used in the training phase is based on a multi-layer process control block (PCB) substrate process, integrating RFID antennas and MEMS IMUs to achieve vibration sensing functions as well as processing, communication, and storage functions. Fig.13 As shown in the figure, the hardware architecture diagram of the tag device hardware corresponding to the physical diagram is as follows Fig.14 As shown, it specifically includes a vibration sensing module MP213, a storage module, an RFID antenna, a filtering network, an MCU and a communication module, a photovoltaic power module, and an energy storage and management module.

[0171] In the embodiment of the present application, the hardware communicates with the platform through RFID low-power consumption. To ensure the continuity of the data, the real-time collected IMU motion data and RFID channel characteristic data are stored in the storage module of the tag in a cached manner. The storage capacity is 32MB, and the size of the User area is 180 words (expressed as Word in English). According to the size of each type of data, IMU data is 2Word*3, phase data is 1Word*2, RSSI data is 1Word, and Doppler frequency shift data is 1Word. It can be calculated that the time period △T that can be stored in the User area is approximately N*△T, △T is the sampling period of the sensor, and a total of N=18 data. After the integrated read-write device sends an excitation signal at a sampling frequency H to activate the tag, the tag will take 18 data from the storage area and report them to the reader for decoding, and transmit them to the platform through an optional Internet of Things communication protocol (such as MQTT, etc.) for data processing and training. The energy management module can store the received radio frequency energy and photovoltaic power energy to supply the back-end chip and sensor applications. Optionally, to ensure low power operation of the tag, the tag chip is in sleep mode when not in use. The lower RF energy can wake up the device and start the sensor's periodic collection work.

[0172] The message reporting data format and structure parameters are shown in Table 1 below:

[0173] Table 1

[0174]

[0175] The reported data frame structure parameters are explained as follows:

[0176] Frame Type: distinguishes the data frame type. The frame with sensor data is the sensor (ie, SEN) mark, and the frame without sensor data is the normal communication handshake protocol mark, such as the Acknowledge character (ACK) and the request (ie, REQ).

[0177] EPC ID: The electronic product code ID of the tag corresponding to this frame.

[0178] User Data: User data carried by this frame, including phase, RSSI, Doppler shift, IMU-x, IMU-y, and IMU-z data.

[0179] Data Length: User data length.

[0180] Frequency: Data reporting frequency, generally selected from 15Hz to 25Hz.

[0181] Check Code: Check code.

[0182] Based on the above embodiments, the embodiment of the present application adopts common labels in the reasoning stage. The structural diagram of common labels is as follows: Fig.15 As shown, it includes a storage module, a communication module / MCU, a filtering network, and an RFID antenna.

[0183] In the embodiment of the present application, the inference stage is the actual model application stage. The model has been trained in the early training stage. In this stage, only the MCU module with integrated RFID tag antenna needs to be encapsulated and hung on the clothes. The RFID reader-writer integrated device excites the tag at a certain time t=1 / H, obtains the M RFID data streams within the previous cache time △T′, and inputs them into the actual model to obtain the recognition result. Since there is no IMU data in the inference stage, the time period of the data that can be reported △T′>△T, and the collection efficiency is higher.

[0184] The message reporting data format and structure parameters are shown in Table 2 below:

[0185] Table 2

[0186]

[0187] Explanation of reported data frame structure parameters:

[0188] Frame Type: distinguishes data frame types. Frames with sensor data are marked with SEN, and frames without sensor data are marked with normal communication handshake protocols, such as ACK and REQ.

[0189] EPC ID: The tag ID corresponding to the frame.

[0190] User Data: User data carried by the frame, including Phase, RSSI signal strength, and DopplerShift data.

[0191] Data Length: User data length.

[0192] Check Code: Check code.

[0193] Based on the above embodiments, it can be seen that in the training phase, the pre-trained model is trained using a small amount of tagged RFID laboratory data sets, and the pre-trained model and online data are automatically annotated and the online deep model is updated and tuned using an iterative optimization method. In the inference phase, ordinary RFID tags are used to collect data at the same sampling frequency, and the behavior results are analyzed in real time.

[0194] In the embodiment of the present application, the above-mentioned related defects are avoided, and data automatic labeling and online updating of the recognition model can be realized quickly and at low cost, so that the recognition model can adapt to environmental conditions such as human interference and metal interference, and the accuracy of the algorithm for identifying customer behavior in stores using RFID data is improved, and the cost of manual labeling and human intervention are reduced. Moreover, in the actual application process, stores can reuse existing ordinary RFID tags in combination with the above-mentioned model to complete customer behavior identification, which greatly improves the intelligent and automated level of store management.

[0195] Based on the above embodiments, the core ideas in the embodiments of the present application can be summarized as follows:

[0196] In order to avoid manual labeling costs, the embodiment of this application adopts price tags to bind clothing, and the reader-writer excites the tag device, obtains the tag's motion data and RFID channel feature data through decoding, and reports it to the main system platform to complete model training and iterative optimization. The main system end system architecture diagram as the core technical idea is as follows Fig.16 As shown, specifically including:

[0197] A. Pre-training module: Use a small amount of labeled data to train the initial pre-training model.

[0198] B. Real-time acquisition module: outputs the real-time data collected online (i.e., motion data and channel characteristic data collected in real time), and performs pre-processing operations such as slicing on it.

[0199] C. Data screening and automatic labeling module: First, read the three-axis acceleration data of the IMU sensor, and use the feature selection-based machine learning method of random forest to calculate the behavior recognition result and category confidence of the time series data, and then obtain the credibility of the pseudo-label information of the sample. According to the interval where the pseudo-label information credibility lies, a pseudo-label information data set is constructed to obtain special data that cannot be obtained in the laboratory environment, and generate sample labels.

[0200] D. Model update module: Train the online model based on the filtered pseudo-label information dataset, calculate the loss function based on the pseudo-label information and the forward propagation output results, and back-propagate to update the gradient and online model weight parameters. After a certain number of iterations, the final updated model is obtained.

[0201] E. Model tuning module: Calculate the moving weighted average based on the weight parameters of the previous online model and the updated model, and iterate to obtain the most optimized online model. If it is the first time to tune, the initial pre-trained model and the updated model are used.

[0202] A specific embodiment is shown below to explain the implementation details in the embodiment of the present application, but the embodiment of the present application can solve but is not limited to this embodiment.

[0203] The implementation of the present application is divided into a training phase and an inference phase. The technical solution is as follows: Fig.17 shown. Fig.17 The implementation process is as follows:

[0204] The training phase uses the system architecture diagram proposed in the embodiment of the present application and the tag device hardware used for training proposed in the embodiment of the present application to complete the automatic labeling of RFID data in actual scenes and the updating and optimization of the online deep model through repeated iterations, thereby obtaining the final recognition model. The training phase mainly includes the pre-training phase and the actual training phase.

[0205] Among them, the pre-training stage is to run the pre-training module A, use a small amount of labeled training data, obtain the initial pre-training model M_0, calculate the weighted loss function loss, and back-propagate to update the parameters of the model M_0.

[0206] In the actual training phase, the real-time acquisition module B, data screening and automatic labeling module C, model update module D, and model tuning module E are run to synchronously collect IMU motion data and RFID channel feature data; the IMU motion data is automatically labeled based on the random forest algorithm to generate the unique hot encoding of the label and the confidence of the pseudo-label information (i.e., credibility); according to the credibility filter (i.e., the preset pseudo-label screening threshold), the highly reliable complex scene data is screened, combined with the collected RFID channel feature data, and packaged and transferred to the updated model M_new in batches for training, the weighted loss function loss is calculated, and the parameters of the model M_new are updated through back propagation. The online model iteration is based on the exponentially weighted moving average algorithm, which is calculated using the previous online model M_cur and the updated model M_new.

[0207] In the inference stage, conventional RFID tags are used to obtain online data at the same sampling frequency, and the behavior recognition results are output through the final online model M_cur_final.

[0208] In the embodiment of the present application, the actual test results and the platform interface of the embodiment are as follows: Fig.18 and Fig.19 As shown. Fig.18 The motion data sensed by the IMU is shown in Figure 1. It can be seen that during behavior recognition, the time from 165s to 239s is the try-on behavior, which lasts for 7.4s, from 239 to 330s is the try-on behavior, which lasts for 9.1s, from 381 to 518s is the try-on behavior, which lasts for 13.7s, from 569 to 633s is the pick-up behavior, which lasts for 6.4s, and from 633 to 720s is the try-on behavior, which lasts for 8.7s. Fig.19 The channel characteristic data perceived by RFID is shown in FIG.

[0209] Based on the above embodiments, the present application has at least the following technical advantages:

[0210] (1) Only a small number of labeled samples are required. During the training phase, real-time data is collected in actual scenarios and the model is updated online for adaptive learning. This allows the model to collect data from complex environments, realize autonomous evolution of the deep model, and reduce human intervention.

[0211] (2) The training phase and the inference phase use RFID+MCU hardware with and without sensors respectively. Stores can reuse existing common RFID tags in combination with the above models to complete customer behavior recognition, reducing the actual application tag cost. By customizing the data reporting protocol, finely managing the data traffic distribution in the training phase and the inference phase, and coordinating the rhythm of model training and model tuning, the efficiency of model training can be improved.

[0212] (3) It can quickly and cost-effectively realize automatic data labeling, screen out environmental conditions such as human interference and metal interference, and use the credibility of the tag to determine whether to further train or discard it, thereby improving the accuracy of the store customer behavior recognition algorithm using RFID data and reducing the cost of manual labeling.

[0213] (4) To address the high labor cost of data labeling, an online data collection method and an automatic data labeling method based on pseudo-labels are proposed. The random forest algorithm is used to generate pseudo-labels and their confidence levels, and data is screened based on a preset pseudo-label screening threshold to obtain complex environment data with high credibility. The data is then input into the online tuning model for iterative optimization, gradually reducing the error of the label data and obtaining a training data set with high accuracy.

[0214] Based on the above embodiment, in another embodiment of the present application, a recognition model training device 1 is provided, such as Fig. 20 As shown, the recognition model training device 1 includes:

[0215] The acquisition unit 10 is used to acquire the motion data of the first device and the channel characteristic data of the second device.

[0216] The construction unit 11 is used to construct a pseudo label information data set based on the motion data and the channel characteristic data.

[0217] The determination unit 12 is used to determine the model update parameters based on the pseudo-label information data set, and use the model update parameters to update the parameters of the pre-trained model to obtain an updated model.

[0218] The processing unit 13 is used to fuse the updated model and the pre-trained model to obtain a recognition model, and the recognition model is used to recognize user behavior.

[0219] In one embodiment, the determination unit 12 is further configured to determine a category confidence level of a behavior category corresponding to the motion data.

[0220] The processing unit 13 is further used to perform data conversion processing on the category confidence to obtain category soft label information.

[0221] The determination unit 12 is further configured to determine the credibility of the pseudo label information based on the category soft label information and the category confidence.

[0222] The determining unit 12 is further configured to determine a pseudo label information data set based on the pseudo label information credibility, a preset pseudo label screening threshold and the channel characteristic data.

[0223] In one embodiment, the determination unit 12 is further configured to determine the similarity between the category soft label information and the category confidence; and determine the similarity as the pseudo label information credibility.

[0224] In one embodiment, the determining unit 12 is further configured to determine target channel characteristic data corresponding to the category soft label information from the channel characteristic data when the pseudo label information credibility is less than a preset pseudo label screening threshold.

[0225] The determination unit 12 is further configured to determine the target channel characteristic data and the corresponding category soft label information as a pseudo label information data set.

[0226] In one embodiment, the acquisition unit 10 is further used to acquire a first weight value corresponding to the updated model and a second weight value corresponding to the pre-trained model.

[0227] The determination unit 12 is further used to determine the recognition model based on the first weight value, the second weight value, the updated model and the pre-trained model.

[0228] In one embodiment, the acquisition unit 10 is further configured to collect channel characteristic data at a first moment.

[0229] The processing unit 13 is further configured to identify the user behavior corresponding to the channel characteristic data at the first moment based on the recognition model.

[0230] An embodiment of the present application provides a recognition model training device, which obtains motion data of a first device and channel feature data of a second device; constructs a pseudo-label information data set based on the motion data and the channel feature data; determines a model update parameter based on the pseudo-label information data set, and uses the model update parameter to update the parameters of a pre-trained model to obtain an updated model; fuses the updated model and the pre-trained model to obtain a recognition model, which is used to identify user behavior. It can be seen that an recognition model training device proposed in an embodiment of the present application, in the process of training the recognition model, collects motion data and channel feature data in an actual scene, combines the collected motion data and channel feature data, and updates the parameters of the pre-trained model. Because the collected motion data and channel feature data are collected in the actual scene, the updated model obtained is more accurate, and after obtaining the updated model, the updated model is further optimized by further fusing the updated model with the pre-trained model. In this way, the prediction result of the recognition model is more accurate.

[0231] Fig.21 This is a schematic diagram of the composition structure of a recognition model training device 1 provided in an embodiment of the present application. In practical applications, based on the same public concept of the above embodiments, Fig.21 As shown, the recognition model training device 1 of the embodiment of the present application includes a processor 14, a memory 15 and a communication bus 16.

[0232] In the specific implementation process, the acquisition unit 10, the construction unit 11, the determination unit 12, and the processing unit 13 can be implemented by a processor 14 located on the recognition model training device 1, and the processor 14 can be a specific application integrated circuit (ASIC, Application Specific Integrated Circuit), a digital signal processor (DSP, Digital Signal Processor), a digital signal processing image processing device (DSPD, Digital Signal Processing Device), a programmable logic image processing device (PLD, Programmable Logic Device), a field programmable gate array (FPGA, Field Programmable Gate Array), a CPU, a controller, a microcontroller, and at least one of a microprocessor. It can be understood that for different devices, the electronic device used to implement the above-mentioned processor function can also be other, and the embodiments of the present application are not specifically limited.

[0233] In the embodiment of the present application, the communication bus 16 is used to realize the connection and communication between the processor 14 and the memory 15; when the processor 14 executes the running program stored in the memory 15, the following recognition model training method is implemented:

[0234] Acquire motion data of the first device and channel characteristic data of the second device; construct a pseudo-label information data set based on the motion data and the channel characteristic data; determine model update parameters based on the pseudo-label information data set, and use the model update parameters to update the parameters of the pre-trained model to obtain an updated model; fuse the updated model and the pre-trained model to obtain a recognition model, which is used to recognize user behavior.

[0235] In one embodiment, the processor 14 is further used to determine the category confidence of the behavior category corresponding to the motion data; perform data conversion processing on the category confidence to obtain category soft label information; determine the credibility of the pseudo label information based on the category soft label information and the category confidence; determine the pseudo label information data set based on the pseudo label information credibility, a preset pseudo label screening threshold and channel feature data.

[0236] In one embodiment, the processor 14 is further configured to determine the similarity between the category soft label information and the category confidence; and determine the similarity as the pseudo label information credibility.

[0237] In one embodiment, the processor 14 is further used to determine target channel characteristic data corresponding to the category soft label information from the channel characteristic data when the credibility of the pseudo label information is less than a preset pseudo label screening threshold; and determine the target channel characteristic data and the corresponding category soft label information as a pseudo label information data set.

[0238] In one embodiment, the processor 14 is further used to obtain a first weight value corresponding to the updated model and a second weight value corresponding to the pre-trained model; and determine the recognition model based on the first weight value, the second weight value, the updated model and the pre-trained model.

[0239] In one embodiment, the processor 14 is further configured to collect channel characteristic data at a first moment; and identify user behavior corresponding to the channel characteristic data at the first moment based on a recognition model.

[0240] Based on the above embodiments, an embodiment of the present application provides a storage medium on which a computer program is stored. The above computer-readable storage medium stores one or more programs. The above one or more programs can be executed by one or more processors and applied to a recognition model training device. The computer program implements the recognition model training method as described above.

[0241] Based on the above embodiments, an embodiment of the present application provides a computer program product, including a computer program, which can be executed by one or more processors and applied to a recognition model training device. The computer program implements the recognition model training method as described above.

[0242] It should be noted that in the embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0243] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), including a number of instructions for an image display device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the method described in each embodiment of the embodiment of the present application.

[0244] The above is only a specific implementation of the embodiment of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A recognition model training method, characterized in that: The method comprises: Acquire motion data of the first device and channel characteristic data of the second device; Based on the motion data and the channel characteristic data, construct a pseudo label information data set; Based on the pseudo-label information data set, determine the model update parameters, and use the model update parameters to update the parameters of the pre-trained model to obtain an updated model; The updated model and the pre-trained model are fused to obtain a recognition model, and the recognition model is used to recognize user behavior.

2. The method according to claim 1, characterized in that The constructing a pseudo label information data set based on the motion data and the channel characteristic data includes: Determining a category confidence of a behavior category corresponding to the motion data; Performing data conversion processing on the category confidence to obtain category soft label information; Determining the credibility of the pseudo label information based on the category soft label information and the category confidence; A pseudo label information data set is determined based on the pseudo label information credibility, a preset pseudo label screening threshold and the channel characteristic data.

3. The method according to claim 2, characterized in that The determining the credibility of the pseudo label information based on the category soft label information and the category confidence includes: Determining the similarity between the category soft label information and the category confidence; The similarity is determined as the pseudo-label information credibility.

4. The method according to claim 2, characterized in that: The determining of the pseudo label information data set based on the pseudo label information credibility, the preset pseudo label screening threshold and the channel characteristic data includes: When the pseudo label information credibility is less than a preset pseudo label screening threshold, determining target channel characteristic data corresponding to the category soft label information from the channel characteristic data; The target channel characteristic data and the corresponding category soft label information are determined as a pseudo label information data set.

5. The method according to claim 1, characterized in that The step of fusing the updated model with the pre-trained model to obtain a recognition model includes: Obtaining a first weight value corresponding to the updated model and a second weight value corresponding to the pre-trained model; The recognition model is determined based on the first weight value, the second weight value, the updated model and the pre-trained model.

6. The method according to claim 1, characterized in that After fusing the updated model with the pre-trained model to obtain a recognition model, the method further includes: Collecting channel characteristic data at a first moment; Based on the recognition model, the user behavior corresponding to the channel characteristic data at the first moment is recognized.

7. A recognition model training device, characterized in that: The device comprises: An acquiring unit, configured to acquire motion data of the first device and channel characteristic data of the second device; A construction unit, configured to construct a pseudo label information data set based on the motion data and the channel characteristic data; A determination unit, configured to determine a model update parameter based on the pseudo-label information data set, and update the parameters of the pre-trained model using the model update parameter to obtain an updated model; A processing unit is used to fuse the updated model and the pre-trained model to obtain a recognition model, and the recognition model is used to recognize user behavior.

8. A recognition model training device, characterized in that: The device comprises: a processor and a memory; when the processor executes the running program stored in the memory, the method according to any one of claims 1 to 6 is implemented.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 6.

Images