Weighing data denoising method, denoising model training method, equipment and storage medium

Abstract

The embodiment of the invention provides a weighing data denoising method, a denoising model training method, equipment and a storage medium. In the embodiment of the invention, a target denoising model is provided, the target denoising model comprises a short-time denoising network and a long-time denoising network, and the short-time denoising network extracts local spatial features of weighing data and carries out short-time denoising processing; the long-time denoising network learns a time sequence dependency relationship of the weighing data to carry out long-time denoising processing; and finally, based on the global noise reduction feature after long-time denoising processing, generating target weighing data of which the noise data is removed. Through combination of the short-time denoising network and the long-time denoising network, the model can capture short-term and long-term dependence characteristics in the weighing data, effective data and noise data are effectively separated from the weighing data, and the accuracy and robustness of noise removal by the model are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a weighing data denoising method and a denoising model training method, device and storage medium. Background Art

[0002] In industrial scenarios, weighing is an important part of the production and management process, and is closely related to production efficiency, quality control, cost control, etc. Weighing equipment, as a weighing tool for measuring the weight of an object, is widely used in various industrial scenarios. When the sensor of the weighing equipment is subjected to external force from an object, it will deform. The deformation of the sensor is converted into a measurable electrical signal, thereby completing the measurement of the mass of the object.

[0003] However, in industrial environments, various noises generated by complex environments, such as electromagnetic interference and temperature fluctuations, can lead to inaccurate weighing results and poor robustness. Therefore, how to effectively reduce noise during weighing is an urgent problem to be solved. Summary of the invention

[0004] Multiple aspects of the present application provide a weighing data denoising method and a denoising model training method, device and storage medium to reduce noise in the weighing process and improve the accuracy and robustness of the weighing results.

[0005] The embodiment of the present application provides a weighing data denoising method, comprising: collecting multiple original weighing data of a target object based on a weighing sensor, wherein the multiple original weighing data have a timestamp to form time series data, and the original weighing data includes noise data; performing frequency domain conversion on the time series data to obtain initial time-frequency distribution information of the original weighing data, wherein the initial time-frequency distribution information is a two-dimensional representation of the original weighing data in the time domain and the frequency domain; inputting the initial time-frequency distribution information into an encoder in an autoencoder, performing encoding denoising processing on the initial time-frequency distribution information, and obtaining an initial feature vector; inputting the initial feature vector into a short-time denoising network in a target denoising model for local denoising The spatial features are extracted and short-term denoising is performed based on the local spatial features to obtain local denoising features corresponding to the original weighing data, and the local denoising features are used to describe the short-term variation law of the original weighing data after the short-term denoising; the local denoising features are input into the long-term denoising network in the target denoising model, the long-term dependency between the local denoising features is learned and long-term denoising is performed based on the long-term dependency to obtain the global denoising features corresponding to the original weighing data; based on the global denoising features corresponding to the original weighing data, the target weighing data of the target object is generated, and the noise data in the original weighing data is removed from the target weighing data.

[0006] The embodiment of the present application also provides a denoising model training method, comprising: obtaining a sample data set, the sample data set comprising a plurality of sample data with sample noise and corresponding annotated data without sample noise, the plurality of sample data having a timestamp to form sample time series data; performing frequency domain conversion on the sample time series data to obtain sample time-frequency distribution information of the sample data, the sample time-frequency distribution information being a two-dimensional representation of the sample data in the time domain and the frequency domain; inputting the sample time-frequency distribution information into an encoder in an autoencoder, performing encoding denoising processing on the sample time-frequency distribution information, and obtaining a sample feature vector; inputting the sample feature vector into a short-time denoising network in an initial denoising model to extract local spatial features and performing short-time denoising processing based on the local spatial features to obtain the sample data. According to the corresponding first denoising feature, the first denoising feature is used to describe the short-term change law of the sample data after short-term denoising processing; the first denoising feature is input into the long-term denoising network in the initial denoising model, the long-term dependency relationship between the first denoising features is learned and long-term denoising processing is performed based on the long-term dependency relationship to obtain the second denoising feature corresponding to the sample data; based on the second denoising feature corresponding to the sample data, denoising result data corresponding to the sample data is generated, and the denoising result data removes the noise data in the sample data; according to the sample data and the denoising result data, a loss function is calculated; when the loss function meets the corresponding termination condition, the initial denoising model training is terminated to obtain the short-term denoising network and the long-term denoising network in the target denoising model.

[0007] An embodiment of the present application also provides an electronic device, comprising: a memory and a processor; the memory is used to store a computer program, and the processor is coupled to the memory and is used to execute the computer program in the memory to implement the steps in the above method.

[0008] An embodiment of the present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement the steps in the above-described method.

[0009] In an embodiment of the present application, a target denoising model is provided, which includes a short-time denoising network and a long-time denoising network. The short-time denoising network extracts the local spatial features of the weighing data and performs short-time denoising; the long-time denoising network learns the temporal dependency of the weighing data and performs long-time denoising; finally, based on the global denoising features after the long-time denoising, the target weighing data with noise removed is generated. By combining the short-time denoising network and the long-time denoising network, the model can capture the short-term and long-term dependency characteristics in the weighing data, effectively separate the valid data and the noise data from the weighing data, and improve the accuracy and robustness of the model in noise removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0011] Figure 1 A schematic diagram of the structure of a target denoising model provided by an exemplary embodiment of the present application;

[0012] Figure 2 A flow chart of a weighing data denoising method provided for another exemplary embodiment of the present application;

[0013] Figure 3 A flowchart of a denoising model training method provided for another exemplary embodiment of the present application;

[0014] Figure 4 A schematic structural diagram of an electronic device provided as yet another exemplary embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0016] It should be noted that, in the case where the embodiments of the present application involve user information, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) are in compliance with relevant laws and standards.

[0017] During the exploration process, the inventors of the present application used the following methods to remove the noise data in the initial weighing data collected by the sensor. For the convenience of description, in some embodiments of the present application, "noise data" may be referred to as "noise".

[0018] A1. A signal processing method based on a filtering algorithm removes noise data from the initial weighing data collected by the sensor and extracts effective weighing data. Effective weighing data refers to weighing data without noise data in the initial weighing data. The filtering algorithm mainly includes the following filtering algorithms: Low-pass filter: remove high-frequency noise data from the initial weighing data through a low-pass filter, and retain low-frequency effective weighing data. It is suitable for filtering out noise data generated by external environmental vibration in a short period of time. Kalman filter: The Kalman filter can dynamically estimate the state of the system, is adaptive, and is suitable for processing continuously changing initial weighing data. In the presence of high-frequency interference or large dynamic changes in the system, the Kalman filter exhibits good stability. Median filter: By selecting the median in the data window as the output, it can effectively remove short-term burst noise data.

[0019] A2. Extraction method based on statistical features. This method uses the statistical characteristics of the initial weighing data to extract valid weighing data from the initial weighing data, thereby removing the noise data in the initial weighing data. The method includes the following steps: mean and variance calculation: by calculating the mean, variance and other statistical indicators of the initial weighing data, remove obvious fluctuations and abnormal data; peak analysis: using a sliding window to calculate the peak of the initial weighing data, thereby identifying the stable interval and obtaining valid weighing data.

[0020] A3. Machine learning methods based on manual feature extraction can be combined with machine learning models, such as support vector machines (SVM), random forests, etc., in order to improve the extraction effect of effective weighing data. The machine learning method can include the following steps: Feature extraction: First, it is necessary to manually define the data features used for machine learning, such as the mean, variance, peak, frequency domain features, etc. of the time series. Training the machine learning model: The machine learning model is trained with the data after feature extraction to realize the identification and removal of noise data, thereby extracting effective weighing data. Among them, the machine learning method based on manual feature extraction can improve the accuracy of effective weighing data extraction to a certain extent.

[0021] A4. Self-correction method based on adaptive algorithm. The adaptive algorithm adapts to the changes in the weighing environment by continuously updating the system parameters during the weighing process. It may include the following steps: Adaptive control: Real-time analysis of the impact of the environment on the effective weighing data during the weighing process, and adjustment of the system through the self-correction mechanism to reduce noise and errors. For example, the adaptive recursive filtering algorithm can be used to dynamically adjust the filter parameters according to the historical weighing data to cope with environmental changes. Feedforward compensation: By collecting external environmental information (such as temperature, humidity, etc.), these external environmental information are used for feedforward compensation, thereby reducing the impact of the external environment on the extraction of effective weighing data. Among them, the advantage of the adaptive algorithm is that it can adjust the system parameters in real time according to the changes in the external environment, thereby improving the stability and accuracy of feature extraction of weighing data.

[0022] A5. Denoising method based on signal decomposition. Signal decomposition technology is also used to extract effective weighing data to separate effective weighing data and noise data from initial weighing data. The method may include the following steps: Wavelet transform: Decompose the initial weighing data into different frequency components through wavelet transform, so as to distinguish low-frequency effective weighing data and high-frequency noise data. This method can effectively remove multi-scale noise data. Empirical mode decomposition (EMD): Decompose the initial weighing data into several intrinsic mode functions (IMFs) through empirical mode decomposition, and then remove the IMF containing noise data components to retain the effective weighing data. Signal decomposition methods have advantages in processing multi-scale and nonlinear noise data.

[0023] Although the above methods A1-A5 can remove the noise data of weighing data, however, in an industrial environment, affected by complex environments, such as electromagnetic interference, temperature fluctuations, etc., various noises generated will cause the weighing results of the object weight to be inaccurate and have poor robustness. In view of the technical problem that the existing sensors are affected by noise, resulting in inaccurate weighing results and poor robustness, in an embodiment of the present application, a target denoising model is provided, and the target denoising model includes a short-time denoising network and a long-time denoising network. The short-time denoising network extracts the local spatial features of the weighing data and performs short-time denoising; the long-time denoising network learns the temporal dependency of the weighing data and performs long-time denoising; finally, based on the global denoising features after long-time denoising, the target weighing data with noise data removed is generated. By combining the short-time denoising network and the long-time denoising network, the model can capture the short-term and long-term dependency characteristics in the weighing data, effectively separate the valid data and the noise data from the weighing data, and improve the accuracy and robustness of the model for noise removal.

[0024] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0025] Figure 1A schematic diagram of a target denoising model provided by an exemplary embodiment of the present application. Figure 1 As shown, the target denoising model 10 includes: a preprocessing module 11, an autoencoder 12, a short-time denoising network 13 and a long-time denoising network 14.

[0026] In this implementation, there is no limitation on the deployment location of the target denoising model, for example, it can be directly deployed on a weighing device, which includes a weighing sensor. Alternatively, it can be deployed on a computer device, which can be any electronic device with computing, storage and communication functions, such as computers, mobile phones, tablets and other terminal devices, or traditional servers, cloud servers, server clusters or workstations and other server devices, without limitation.

[0027] In this embodiment, a plurality of original weighing data of a target object collected by a weighing sensor is obtained. The target object may be determined according to the weighing scenario and requirements, for example, it may be, but not limited to, raw materials, crops, medical equipment, engineering equipment, and living bodies (such as animals). Among them, this embodiment does not limit the method of obtaining weighing data. When the target denoising model is deployed on the weighing device, the original weighing data from the weighing sensor may be directly received; or, when the target denoising model is deployed on a computer device, the original weighing data collected by the weighing sensor may be sent to the computer device via a wired connection or a network.

[0028] The weighing sensor is affected by the external environment and / or itself, such as electrical interference, temperature fluctuation, etc., and the collected original weighing data includes noise data. This embodiment does not limit the type of weighing sensor, and the types of weighing sensors include but are not limited to: resistance strain gauge weighing sensor, piezoelectric weighing sensor, capacitive weighing sensor, hydraulic weighing sensor, magnetostrictive weighing sensor, torque weighing sensor, etc.

[0029] In this embodiment, multiple original weighing data of the target object have timestamps, thereby forming time series data, which refers to a collection of original weighing data arranged at set time intervals, each original weighing data corresponds to a timestamp, and the original weighing data can be recorded at various frequencies (for example, every day, every month, every year). At the same time, in time series prediction, "time step" refers to the time interval between one original weighing data and the next original weighing data in the time series data. It defines the sampling frequency and time scale of the time series data.

[0030] Further, the time series data of the original weighing data is input into the preprocessing module 11. The preprocessing module 11 performs frequency domain conversion on the time series data to convert the time series data from the time domain to the frequency domain, and obtains the initial time-frequency distribution information of the original weighing data. The initial time-frequency distribution information is a two-dimensional joint representation of the original weighing data in the time domain and the frequency domain, and captures the characteristics of the original weighing data in the time domain and the frequency domain at the same time, which is helpful for the subsequent removal of noise data from the original weighing data.

[0031] In this embodiment, the initial time-frequency distribution information output by the preprocessing module 11 is input into the encoder in the autoencoder 12 (autoencoder). The autoencoder 12 is an unsupervised learning model. By performing feature compression, dimensionality reduction and noise removal on the initial time-frequency distribution information time by time step, the initial time-frequency distribution information can be encoded into a low-dimensional feature representation, and a representative initial feature vector corresponding to each time step is extracted, thereby separating the noise data from the valid data in the input data. For example, if there is burst noise in the original weighing data, the autoencoder 12 can remove these burst noise data from the original weighing data by learning the structural characteristics of the original weighing data, which can provide higher quality input features for the subsequent processing process, thereby improving the processing performance of the subsequent deep learning model, and better eliminating noise and extracting valid data from the original weighing data.

[0032] Among them, each time step corresponds to an initial feature vector. Further, the initial feature vector corresponding to each time step is subjected to short-term denoising. Short-term denoising refers to removing short-term noise data in time series data. Short-term noise refers to noise data with a short duration and strong randomness in time series data. For example, the weighing sensor may be affected by lightning, electromagnetic radiation in high-voltage power grids, etc. to generate noise during the process of collecting the original weighing data of the target object. In the short-term denoising process, the initial feature vector corresponding to each time step is input into the short-term denoising network 13 in the target denoising model 10 to extract local spatial features, and obtain the local spatial features corresponding to each time step. Then, short-term denoising is performed based on the local spatial features to obtain local noise reduction features. The local noise reduction features can also be corresponding to each time step. The local noise reduction features are used to describe the short-term change law of the original weighing data after short-term denoising.

[0033] After obtaining the local denoising features, the local denoising features are input into the long-term denoising network 14 in the target denoising model. Each local denoising feature corresponds to a time step, so that a local denoising feature sequence can be obtained. The local denoising feature sequence can include multiple local denoising features. The long-term denoising network 14 is used to capture the time series characteristics between local denoising features, and is particularly good at capturing the long-term dependencies between local denoising features for long-term denoising processing.

[0034] Among them, long-term denoising processing refers to separating the trend characteristics and long-term noise data in the original weighing data. Among them, the trend characteristics refer to the characteristics of the time series data that gradually present a long-term change pattern over time. The long-term change pattern includes monotonically increasing, monotonically decreasing, and periodic changes, etc. In addition, long-term noise data refers to interference or errors with a longer time span. In this embodiment, the long-term denoising network can capture the changing trends of trend characteristics and long-term noise data, and effectively compensate for them. It can not only remove instantaneous noise data, but also analyze the characteristics of long-term noise data, thereby effectively removing long-term noise data from the original load-bearing data.

[0035] After long-term noise processing, the global noise reduction feature corresponding to the original weighing data is obtained. Further, based on the global noise reduction feature, the target weighing data of the target object is generated, and the target weighing data removes the noise data of the original weighing data. The execution subject of generating the target weighing data is not limited in this embodiment. Figure 1 The operation of generating target weighing data in the long-term denoising network is given, but it is not limited to this. The operation can be performed within the target denoising model.

[0036] In an embodiment of the present application, a target denoising model is provided, which includes a short-time denoising network and a long-time denoising network. The short-time denoising network extracts the local spatial features of the weighing data and performs short-time denoising; the long-time denoising network learns the temporal dependency of the weighing data and performs long-time denoising; finally, based on the global denoising features after the long-time denoising, the target weighing data with noise removed is generated. By combining the short-time denoising network and the long-time denoising network, the model can capture the short-term and long-term dependency characteristics in the weighing data, effectively separate the valid data and the noise data from the weighing data, and improve the accuracy and robustness of the model in noise removal.

[0037] In addition, compared with the traditional denoising method, the weighing data denoising method provided in the embodiment of the present application has the following advantages:

[0038] 1. No need to manually mark features, reducing development difficulty.

[0039] In traditional machine learning methods, such as the statistical feature-based method in A2 above, feature extraction relies on manual design, such as using expert experience to extract statistical features such as mean and variance. This process is not only time-consuming and labor-intensive, but may also be subjective and incomplete, resulting in poor model performance. In addition, feature extraction needs to be adjusted according to environmental changes, which makes it difficult to adapt to the changing industrial environment.

[0040] However, the denoising model of the technical solution of the present application adopts a combination of a short-time denoising network and a long-time denoising network. Through the short-time denoising network and the long-time denoising network, the model can effectively learn the local spatial features and temporal dependencies in the weighing data without human intervention, thereby replacing the traditional manual feature extraction.

[0041] 2. Powerful complex noise data processing capabilities.

[0042] The above-mentioned signal processing method based on filtering algorithm A1 and the machine learning method based on manual feature extraction A3 usually perform poorly when facing complex and diverse noise data. For example, low-pass filtering is prone to cause effective signal loss when processing non-Gaussian noise, while the method based on manual features is difficult to effectively distinguish and process different types of noise data.

[0043] However, the technical solution of the present application adopts a model structure that combines a short-term denoising network with a long-term denoising network. The short-term denoising network is used to extract the local spatial features (such as local patterns) of the weighing data, while the long-term denoising network is used to capture the temporal dependencies of the data (such as long-term dependencies). This combined structure allows the model to identify and remove noise from weighing data in both the short and long term. In addition, the introduction of an autoencoder for data dimensionality reduction and denoising can distinguish noise data from valid data.

[0044] 3. High precision and robustness.

[0045] The above-mentioned machine learning method based on manual feature extraction of A3 and the denoising method based on signal decomposition of A5 are difficult to provide such comprehensive feature expressions when dealing with complex noise data, resulting in reduced accuracy.

[0046] However, the denoising model of the technical solution of the present application uses a short-term denoising network to extract local spatial features and a long-term denoising network to extract time series features, so that the model can capture the short-term and long-term dependency characteristics in the data and provide a more accurate weighing data extraction effect. In addition, automatic denoising is also achieved through an autoencoder model to effectively separate useful signals and noise, thereby improving the robustness of the model to different noise environments.

[0047] 4. Reduce human intervention and reduce maintenance costs.

[0048] The above-mentioned machine learning method based on manual feature extraction of A3 and the self-correction method based on adaptive algorithm of A4 have complex parameter tuning processes and need to be adjusted regularly to cope with environmental changes.

[0049] However, in the technical solution of the present application, the target denoising model is used to automatically learn features and adjust parameters, thereby reducing the maintenance workload and the development and maintenance costs of the system.

[0050] 5. Fusion of multi-source information to improve stability.

[0051] The signal processing method based on the filtering algorithm of A1 mentioned above has a single problem. This method usually only processes noise data in a specific frequency range. For example, a low-pass filter can only remove high-frequency noise data, but is powerless for low-frequency noise data. A single feature extraction method often fails to capture all the important information in the weighing data, resulting in poor processing of noise data.

[0052] However, in the technical solution of the present application, by fusing multiple features such as the time domain and frequency domain and combining the capabilities of the short-time denoising network and the long-time denoising network, multi-level information is learned from the weighing data, further improving the model's resistance to noise interference and ensuring the stability of the generated target weighing data.

[0053] In an optional embodiment, before the time series data is converted into the frequency domain, the preprocessing module is further used to: standardize the time series data according to the mean and variance of the time series data to ensure that the input data of different scales are evenly distributed; and / or remove abnormal values ​​in the time series data according to the mean or preset value range corresponding to the time series data. The mean and the preset data range are set according to empirical values, and the values ​​of different weighing sensors may be different.

[0054] In an optional embodiment, when the preprocessing module performs frequency domain conversion on the time series data to obtain the initial time-frequency distribution information of the original weighing data, it includes: performing a short-time Fourier transform (STFT) on the time series data to obtain the initial time-frequency distribution of the original weighing data. Of course, this embodiment can also adopt other frequency domain conversion methods, such as but not limited to wavelet transformation, etc. Wavelet transformation refers to decomposing the weighing data into different frequency components through wavelet transformation, thereby distinguishing the low-frequency effective weighing data and high-frequency noise. This method can effectively remove multi-scale noise.

[0055] In an optional embodiment, the short-time denoising network is implemented using a convolutional neural network (CNN), which includes multiple cascaded convolutional layers and pooling layers. When the initial feature vector is input into the short-time denoising network in the target denoising model to extract local spatial features and perform short-time denoising processing based on the local spatial features to obtain local denoising features corresponding to the original weighing data, it includes: inputting the initial feature vector into the convolutional neural network, extracting and pooling local spatial features through cascaded convolutional layers and pooling layers in sequence, and using the result of the last level of pooling processing as the local denoising feature; wherein the input of each pooling layer is the output of the previous convolutional layer cascaded with it.

[0056] In this embodiment, the convolutional layer of the convolutional neural network can automatically adjust the weights through training, without manually designing features, so that the system can adapt to a changing input environment. Because the original weighing data is usually affected by environmental interference, such as mechanical vibration, temperature change, etc., the convolutional neural network can extract the effective features hidden in these interferences, improving the robustness and generalization ability of the model. Among them, in the convolution operation, the convolution kernel (filter) automatically extracts the local change pattern of the initial feature vector, which can effectively identify the local spatial features of the initial feature vector, and then the short-term noise data and the local spatial features can be separated to obtain local noise reduction features.

[0057] Among them, each time step of the time series data corresponds to a local noise reduction feature, thereby forming a local noise reduction feature sequence. The long-term denoising network can capture the long-term dependency between the local noise reduction features for long-term denoising processing. In an optional embodiment, the long-term denoising network can be implemented as a long short-term memory network (Long Short Term Memory Network, LSTM), which is used to capture the temporal dependency in the original weighing data, and is particularly good at capturing long-term dependencies. In the process of learning the long-term dependency between the local noise reduction features and performing long-term denoising based on the long-term dependency, the long-term memory network inputs the local noise reduction features into the memory unit in the long-term denoising network, and the memory unit is used to store the historical dependency information learned from the historical local noise reduction features. Further, the memory unit learns the long-term dependency between the local noise reduction features based on the gating mechanism; the long-term dependency and the local noise reduction features are input into the denoising module in the long-term denoising network, and the denoising module performs long-term denoising on the local noise reduction features based on the long-term dependency to obtain the global noise reduction features. By combining the processing capabilities of long short-term memory networks for time series data, the target denoising model can make more accurate judgments based on historical dependency information, which is particularly important for ensuring the reliability and consistency of weighing results (i.e., target weighing data).

[0058] In an optional embodiment, based on the global noise reduction feature corresponding to the original weighing data, generating the target weighing data of the target object includes: inputting the global noise reduction feature into the decoder in the autoencoder for decoding to obtain the target time-frequency distribution information; performing time domain conversion on the target time-frequency distribution information to obtain the target weighing data. Optionally, performing time domain conversion on the target time-frequency distribution information to obtain the target weighing data includes: performing short-time Fourier inverse transform on the target time-frequency distribution information to obtain the target weighing data.

[0059] Optionally, during the period of collecting multiple original weighing data of the target objects based on the weighing sensor, the number of target objects changes dynamically; accordingly, the original weighing data also changes dynamically, and the law of dynamic change is monotonically increasing, monotonically decreasing or periodic change; the method also includes: outputting the target weighing data, and the target weighing data changes with the dynamic change of the number of target objects.

[0060] In an optional embodiment, during the training process of the target denoising model, a sample data set is obtained, the sample data set includes multiple sample data with sample noise and corresponding labeled data without sample noise, and at least part of the multiple sample data is obtained by learning the labeled data without sample noise using a generative adversarial network (GAN) to improve the robustness and generalization ability of the model. Optionally, the generative adversarial network includes a generator and a discriminator, and the generator can generate labeled data with random noise to simulate possible interference in a real environment. The discriminator is used to determine whether the labeled data with random noise is real sample data or generated by the generator. Sample data with random noise and features are generated through a process of mutual game adversarial learning, which can enrich the sample data set and prevent overfitting in the model training process.

[0061] Further, sample feature vectors corresponding to multiple sample data are obtained. The method for obtaining the sample feature vectors can refer to the method for obtaining the initial feature vectors, which will not be repeated here. The obtained sample feature vectors are input into the short-time denoising network in the initial denoising model to extract local spatial features and perform short-time denoising based on the local spatial features to obtain the first denoising feature. The first denoising feature is input into the long-time denoising network in the initial denoising model to learn the long-term dependency between the first denoising features and perform long-term denoising based on the long-term dependency to obtain the second denoising feature; based on the second denoising feature, denoising result data corresponding to the sample data is generated. Then, based on the labeled data and the denoising result data, the loss function is calculated. The loss function may be, but is not limited to, mean squared error (MSE) and mean absolute error (MAE), etc. When the loss function satisfies the corresponding termination condition, the model training is terminated to obtain the target denoising model. In one example, the termination condition may be minimizing the loss function,

[0062] Optionally, the loss function can be implemented as a piecewise loss function. In this case, calculating the loss function includes: calculating a global loss function based on the labeled data and the denoising result data; calculating a local loss function based on the first denoising feature and the second denoising feature; and according to the global loss function and the local loss function, the termination condition corresponding to the model training is satisfied, and the model training is terminated to obtain the target denoising model. The termination conditions of the global loss function and the local loss function may be different, and there is no limitation on this.

[0063] In an optional embodiment, in the process of calculating the local loss function according to the first denoising feature and the second denoising feature, the loss function is calculated for the first denoising feature and the second denoising feature and the labeled data respectively, the loss function between the first denoising feature and the labeled data is called the first loss function, and the loss function between the second denoising feature and the labeled data is called the second loss function. The first loss function and the second loss function are weightedly summed as the local loss function.

[0064] After training with a large amount of data, it has strong generalization ability and is applicable to a variety of industrial weighing scenarios without the need to redesign features. In contrast, the manual feature extraction method A3 in the above embodiment has poor applicability in different application scenarios and needs to redesign features for each scenario, with low versatility and transferability.

[0065] In an optional embodiment, the initial denoising model also includes a result verification module, which is used to evaluate the accuracy and stability of the model and adjust the model online. During the model training process, cross-validation is performed through the result verification module to evaluate the generalization ability of the model and ensure its performance on unknown data. The accuracy and robustness of the model are detected by dividing the sample data set into a training set and a validation set.

[0066] In an optional embodiment, in the target denoising model, by introducing an online learning mechanism, the model can be adjusted according to real-time data after deployment to adapt to environmental changes. For example, the conditions in the industrial weighing environment may change frequently, such as the impact of changes in temperature and humidity on sensors. Through the online learning mechanism, the model can dynamically adjust parameters according to the newly input data, so that it can still maintain high performance in long-term operation, ensuring the stability and accuracy of the weighing results. In addition, the online learning mechanism enables the model to have the ability to adapt to environmental changes without manual intervention. For example, when a new product is launched or the weighing process changes, the model can be automatically updated to adapt to the new data distribution. This adaptability significantly improves the flexibility and practicality of the system and avoids frequent manual parameter adjustment. In contrast, the self-correction method based on the adaptive algorithm of A4 in the above embodiment, although it can adjust parameters, has limited ability to cope with environmental changes, especially when the changes are too fast, it is difficult to effectively cope with them.

[0067] Optionally, in order to adapt to different industrial environments, transfer learning technology is used. Specifically, the pre-trained target denoising model can be fine-tuned with a small amount of new data to adapt to new scenarios. This approach reduces the need for large amounts of data while improving the generalizability of the model.

[0068] Optionally, in order to meet the real-time processing requirements of weighing data at industrial sites, the model design optimizes the computational efficiency. For example, the real-time processing capability of the original weighing data can be achieved by optimizing the convolution calculation of the convolutional neural network and the recursive calculation of the long short-term memory network in the target denoising model. In addition, the pre-trained model has a fast reasoning speed and can complete the processing and feature extraction of the original weighing data in a very short time, ensuring that the system can respond quickly to external input. In addition, the target denoising model can be deployed on hardware computing resources specially designed for accelerating deep learning computing tasks, such as but not limited to GPU (Graphic Processing Unit) and TPU (Tensor Processing Unit), thereby further improving the processing speed and meeting the high real-time requirements for data processing in the industrial production process. In contrast, the signal processing method based on the filtering algorithm of A1 and the denoising method based on signal decomposition of A5 in the above-mentioned embodiment are usually highly computationally complex and difficult to achieve efficient real-time processing.

[0069] The present application also provides a weighing data denoising method, such as Figure 2 As shown, the method comprises:

[0070] S201: collecting a plurality of original weighing data of a target object based on a weighing sensor, wherein the plurality of original weighing data have a time stamp to form time series data, and the original weighing data include noise data;

[0071] S202: performing frequency domain conversion on the time series data to obtain initial time-frequency distribution information of the original weighing data, where the initial time-frequency distribution information is a two-dimensional representation of the original weighing data in the time domain and the frequency domain;

[0072] S203: inputting the initial time-frequency distribution information into an encoder in the autoencoder, performing encoding and denoising processing on the initial time-frequency distribution information to obtain an initial feature vector;

[0073] S204: Inputting the initial feature vector into the short-time denoising network in the target denoising model to extract local spatial features and perform short-time denoising based on the local spatial features to obtain local denoising features corresponding to the original weighing data. The local denoising features are used to describe the short-term variation law of the original weighing data after the short-time denoising process;

[0074] S205: inputting the local denoising features into the long-term denoising network in the target denoising model, learning the long-term dependency relationship between the local denoising features and performing long-term denoising processing based on the long-term dependency relationship, and obtaining the global denoising features corresponding to the original weighing data;

[0075] S206: Generate target weighing data of the target object based on the global noise reduction features corresponding to the original weighing data, where the target weighing data removes noise data in the original weighing data.

[0076] In an optional embodiment, target weighing data of the target object is generated based on the global noise reduction features corresponding to the original weighing data, including: inputting the global noise reduction features into a decoder in the encoder for decoding to obtain target time-frequency distribution information; performing time domain conversion on the target time-frequency distribution information to obtain the target weighing data.

[0077] In an optional embodiment, the initial time-frequency distribution information of the original weighing data is obtained by performing frequency domain conversion on the time series data, including: performing short-time Fourier transform on the time series data to obtain the initial time-frequency distribution of the original weighing data; performing time domain conversion on the target time-frequency distribution information to obtain the target weighing data, including: performing short-time Fourier inverse transform on the target time-frequency distribution information to obtain the target weighing data.

[0078] In an optional embodiment, the short-time denoising network is implemented using a convolutional neural network, which includes multiple cascaded convolutional layers and pooling layers. The initial feature vector is input into the short-time denoising network in the target denoising model to extract local spatial features and perform short-time denoising processing based on the local spatial features to obtain local denoising features corresponding to the original weighing data, including: inputting the initial feature vector into the convolutional neural network, extracting and pooling local spatial features through cascaded convolutional layers and pooling layers in sequence, and using the result of the last level of pooling processing as the local denoising feature; wherein the input of each pooling layer is the output of the previous convolutional layer cascaded with it.

[0079] In an optional embodiment, the local denoising features are input into a long-term denoising network in the target denoising model, the long-term dependencies between the local denoising features are learned, and long-term denoising processing is performed based on the long-term dependencies to obtain the global denoising features corresponding to the original weighing data, including: inputting the local denoising features into a memory unit in the long-term denoising network, and learning the long-term dependencies between the local denoising features based on a gating mechanism; inputting the long-term dependencies and the local denoising features into a denoising module in the long-term denoising network, and performing long-term denoising processing on the local denoising features based on the long-term dependencies to obtain the global denoising features; wherein the memory unit is used to store historical dependency information learned from historical local denoising features.

[0080] In an optional embodiment, before performing frequency domain conversion on the time series data, it also includes: standardizing the time series data according to the mean and variance of the time series data; and / or removing outliers in the time series data according to the mean or a preset numerical range corresponding to the time series data.

[0081] In an optional embodiment, during the period of collecting multiple original weighing data of the target objects based on the weighing sensor, the number of target objects changes dynamically; accordingly, the original weighing data also changes dynamically, and the law of the dynamic change is dynamic increase or dynamic decrease; the method also includes: outputting the target weighing data, and the target weighing data changes with the dynamic change of the number of target objects.

[0082] In an optional embodiment, the method further includes: obtaining a sample data set, the sample data set including: a plurality of sample data with sample noise and corresponding annotated data without sample noise, wherein at least part of the plurality of sample data is obtained by learning the annotated data without sample noise using a generative adversarial network; obtaining sample feature vectors corresponding to the plurality of sample data, inputting the sample feature vectors into a short-time denoising network in an initial denoising model to extract local spatial features and performing short-time denoising based on the local spatial features to obtain a first denoising feature; inputting the first denoising feature into a long-time denoising network in the initial denoising model, learning a long-term dependency relationship between the first denoising features and performing long-term denoising based on the long-term dependency to obtain a second denoising feature; generating denoising result data corresponding to the sample data based on the second denoising feature; calculating a loss function according to the annotated data and the denoising result data, wherein the loss function is a loss function between the annotated data and the denoising result data; and terminating model training when the loss function satisfies a corresponding termination condition to obtain the target denoising model.

[0083] The present application also provides a denoising model training method, such as Figure 3 As shown, the method comprises:

[0084] S301: Acquire a sample data set, where the sample data set includes a plurality of sample data with sample noise and corresponding annotated data without sample noise, where the plurality of sample data have timestamps to form sample time series data;

[0085] S302: Perform frequency domain conversion on the sample time series data to obtain sample time-frequency distribution information of the sample data, where the sample time-frequency distribution information is a two-dimensional representation of the sample data in the time domain and the frequency domain;

[0086] S303: Input the sample time-frequency distribution information into the encoder in the autoencoder, perform encoding and denoising processing on the sample time-frequency distribution information, and obtain a sample feature vector;

[0087] S304: Inputting the sample feature vector into the short-term denoising network in the initial denoising model to extract local spatial features and perform short-term denoising based on the local spatial features to obtain a first denoising feature corresponding to the sample data, the first denoising feature being used to describe the short-term variation law of the sample data after the short-term denoising process;

[0088] S305: inputting the first denoising features into the long-term denoising network in the initial denoising model, learning the long-term dependency relationship between the first denoising features and performing long-term denoising processing based on the long-term dependency relationship to obtain the second denoising features corresponding to the sample data;

[0089] S306: generating denoising result data corresponding to the sample data based on the second denoising feature corresponding to the sample data, where the denoising result data removes noise data in the sample data;

[0090] S307: Calculate the loss function based on the sample data and the denoising result data;

[0091] S308: When the loss function satisfies the corresponding termination condition, the initial denoising model training is terminated to obtain the short-term denoising network and the long-term denoising network in the target denoising model.

[0092] The detailed implementation and beneficial effects of each step in the method of this embodiment have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0093] It should be noted that the execution subject of each step of the method provided in the above embodiment can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 201 to 203 can be device A; for another example, the execution subject of steps 201 and 202 can be device A, and the execution subject of step 203 can be device B; and so on.

[0094] In addition, in some of the processes described in the above embodiments and the accompanying drawings, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel, and the sequence numbers of the operations, such as 201, 202, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0095] Figure 4 FIG. 1 is a schematic diagram of a structure of an electronic device provided by another exemplary embodiment of the present application. Figure 4 As shown, the electronic device includes: a memory 44 and a processor 45.

[0096] The memory 44 is used to store computer programs and can be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, contact data, phone book data, messages, pictures, videos, etc.

[0097] The memory 44 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 read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0098] The processor 45 is coupled to the memory 44 and is used to execute the computer program in the memory 44, so as to: collect multiple original weighing data of the target object based on the weighing sensor, wherein the multiple original weighing data have timestamps to form time series data, and the original weighing data includes noise data; perform frequency domain conversion on the time series data to obtain initial time-frequency distribution information of the original weighing data, wherein the initial time-frequency distribution information is a two-dimensional representation of the original weighing data in the time domain and the frequency domain; input the initial time-frequency distribution information into the encoder in the autoencoder, perform encoding and denoising processing on the initial time-frequency distribution information, and obtain an initial feature vector; input the initial feature vector into the short-time denoising in the target denoising model; The local denoising network extracts local spatial features and performs short-term denoising based on the local spatial features to obtain local denoising features corresponding to the original weighing data, and the local denoising features are used to describe the short-term variation law of the original weighing data after the short-term denoising; the local denoising features are input into the long-term denoising network in the target denoising model, the long-term dependency between the local denoising features is learned and long-term denoising is performed based on the long-term dependency to obtain the global denoising features corresponding to the original weighing data; based on the global denoising features corresponding to the original weighing data, the target weighing data of the target object is generated, and the noise data in the original weighing data is removed from the target weighing data.

[0099] In an optional embodiment, when the processor 45 generates the target weighing data of the target object based on the global noise reduction features corresponding to the original weighing data, it is specifically used to: input the global noise reduction features into the decoder in the encoder for decoding to obtain target time-frequency distribution information; perform time domain conversion on the target time-frequency distribution information to obtain the target weighing data.

[0100] In an optional embodiment, when the processor 45 performs frequency domain conversion on the time series data to obtain the initial time-frequency distribution information of the original weighing data, it is specifically used to: perform short-time Fourier transform on the time series data to obtain the initial time-frequency distribution of the original weighing data; perform time domain conversion on the target time-frequency distribution information to obtain the target weighing data, including: perform short-time Fourier inverse transform on the target time-frequency distribution information to obtain the target weighing data.

[0101] In an optional embodiment, the short-time denoising network is implemented using a convolutional neural network, which includes multiple cascaded convolutional layers and pooling layers. When the processor 45 inputs the initial feature vector into the short-time denoising network in the target denoising model to extract local spatial features and performs short-time denoising based on the local spatial features to obtain the local denoising features corresponding to the original weighing data, it is specifically used to: input the initial feature vector into the convolutional neural network, extract and pool the local spatial features through the cascaded convolutional layers and pooling layers in turn, and use the result of the last level of pooling as the local denoising features; wherein the input of each pooling layer is the output of the previous convolutional layer cascaded with it.

[0102] In an optional embodiment, when the processor 45 inputs the local denoising features into the long-term denoising network in the target denoising model, learns the long-term dependencies between the local denoising features and performs long-term denoising based on the long-term dependencies to obtain the global denoising features corresponding to the original weighing data, it is specifically used to: input the local denoising features into the memory unit in the long-term denoising network, and learn the long-term dependencies between the local denoising features based on the gating mechanism; input the long-term dependencies and the local denoising features into the denoising module in the long-term denoising network, and perform long-term denoising on the local denoising features based on the long-term dependencies to obtain the global denoising features; wherein the memory unit is used to store historical dependency information learned from historical local denoising features.

[0103] In an optional embodiment, before performing frequency domain conversion on the time series data, the processor 45 is further used to: standardize the time series data according to the mean and variance of the time series data; and / or remove outliers in the time series data according to the mean or a preset numerical range corresponding to the time series data.

[0104] In an optional embodiment, during the period of collecting multiple original weighing data of the target objects based on the weighing sensor, the number of target objects changes dynamically; accordingly, the original weighing data also changes dynamically, and the law of the dynamic change is dynamic increase or dynamic decrease; the processor 45 is also used to: output the target weighing data, and the target weighing data changes with the dynamic change of the number of target objects.

[0105] In an optional embodiment, the processor 45 is further used to: obtain a sample data set, the sample data set including: a plurality of sample data with sample noise and corresponding annotated data without sample noise, wherein at least part of the plurality of sample data is obtained by learning the annotated data without sample noise using a generative adversarial network; obtain sample feature vectors corresponding to the plurality of sample data, input the sample feature vectors into a short-time denoising network in an initial denoising model to extract local spatial features and perform short-time denoising based on the local spatial features to obtain a first denoising feature; input the first denoising feature into a long-time denoising network in the initial denoising model, learn the long-term dependency between the first denoising features and perform long-term denoising based on the long-term dependency to obtain a second denoising feature; generate denoising result data corresponding to the sample data based on the second denoising feature; calculate a loss function based on the annotated data and the denoising result data, wherein the loss function is a loss function between the annotated data and the denoising result data; and terminate model training when the loss function satisfies a corresponding termination condition to obtain the target denoising model.

[0106] Further, if Figure 4 As shown, the electronic device also includes: a communication component 46, a display 47, a power component 48, an audio component 49 and other components. Figure 4 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 4 In addition, Figure 4The components in the dashed box are optional components, not mandatory components, and the specific components depend on the product form of the electronic device. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, or an IOT device, or a server device such as a conventional server, a cloud server, or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it can include Figure 4 If the electronic device of this embodiment is implemented as a server device such as a conventional server, a cloud server or a server array, it may not include Figure 4 Components within the dashed box.

[0107] The embodiment of the present application also provides an electronic device, the implementation structure of the electronic device is Figure 4 The implementation structure of the electronic device shown is the same or similar, and can refer to Figure 4 The electronic device provided in this embodiment is implemented as follows: Figure 4 The difference between the electronic devices in the illustrated embodiments mainly lies in that the functions implemented by the processor in the electronic device executing the computer program stored in the memory are different. For the electronic device provided in this embodiment, its processor executes the computer program stored in the memory, which can be used to: obtain a sample data set, the sample data set includes multiple sample data with sample noise and its corresponding labeled data without sample noise, the multiple sample data have timestamps to form sample time series data; perform frequency domain conversion on the sample time series data to obtain sample time-frequency distribution information of the sample data, the sample time-frequency distribution information is a two-dimensional representation of the sample data in the time domain and the frequency domain; input the sample time-frequency distribution information into the encoder in the autoencoder, perform encoding and denoising processing on the sample time-frequency distribution information to obtain a sample feature vector; input the sample feature vector into the short-time denoising network in the initial denoising model to extract local spatial features and perform short-time denoising based on the local spatial features. , obtain a first denoising feature corresponding to the sample data, the first denoising feature is used to describe the short-term variation law of the sample data after short-term denoising processing; input the first denoising feature into the long-term denoising network in the initial denoising model, learn the long-term dependency relationship between the first denoising features and perform long-term denoising based on the long-term dependency relationship to obtain a second denoising feature corresponding to the sample data; based on the second denoising feature corresponding to the sample data, generate denoising result data corresponding to the sample data, the denoising result data removes the noise data in the sample data; calculate the loss function according to the sample data and the denoising result data; end the initial denoising model training when the loss function meets the corresponding termination condition to obtain the short-term denoising network and the long-term denoising network in the target denoising model.

[0108] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the processor is enabled to implement the steps in the above method.

[0109] An embodiment of the present application also provides a computer program product, which includes a computer program / instructions. When the computer program / instructions are executed by a processor, the processor is enabled to implement the steps in the above method embodiment.

[0110] The above-mentioned memory 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 read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0111] The above-mentioned communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wide Band (UWB) technology, Bluetooth (BT) technology and other technologies.

[0112] The above-mentioned display includes a screen, and the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0113] The power supply assembly provides power to various components of the device where the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply assembly is located.

[0114] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (Microphone, MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting an audio signal.

[0115] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-readable storage media (including but not limited to disk storage, compact disc read-only memory (Compact Disc Read-Only Memory, CD-ROM), optical storage, etc.) containing computer-usable program code.

[0116] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0119] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interface, network interface and memory.

[0120] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0121] Computer readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0122] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity 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, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0123] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A weighing data denoising method, characterized in that: include: Collecting a plurality of original weighing data of the target object based on a weighing sensor, wherein the plurality of original weighing data have a time stamp to form time series data, and the original weighing data includes noise data; Performing frequency domain conversion on the time series data to obtain initial time-frequency distribution information of the original weighing data, wherein the initial time-frequency distribution information is a two-dimensional representation of the original weighing data in the time domain and the frequency domain; Inputting the initial time-frequency distribution information into an encoder in an autoencoder, performing encoding and denoising processing on the initial time-frequency distribution information to obtain an initial feature vector; Input the initial feature vector into the short-time denoising network in the target denoising model to extract local spatial features and perform short-time denoising processing based on the local spatial features to obtain local denoising features corresponding to the original weighing data, and the local denoising features are used to describe the short-term variation law of the original weighing data after the short-time denoising processing; Inputting the local denoising features into the long-term denoising network in the target denoising model, learning the long-term dependency relationship between the local denoising features and performing long-term denoising processing based on the long-term dependency relationship to obtain the global denoising features corresponding to the original weighing data; Based on the global noise reduction features corresponding to the original weighing data, target weighing data of the target object is generated, and the target weighing data is obtained by removing noise data from the original weighing data.

2. The method according to claim 1, characterized in that Generating target weighing data of the target object based on the global noise reduction features corresponding to the original weighing data, including: Inputting the global noise reduction feature into a decoder in the autoencoder for decoding to obtain target time-frequency distribution information; The target time-frequency distribution information is converted into time domain to obtain the target weighing data.

3. The method according to claim 2, characterized in that According to the frequency domain conversion of the time series data, the initial time-frequency distribution information of the original weighing data is obtained, including: performing a short-time Fourier transform on the time series data to obtain the initial time-frequency distribution of the original weighing data; Performing time domain conversion on the target time-frequency distribution information to obtain the target weighing data includes: performing short-time Fourier inverse transform on the target time-frequency distribution information to obtain the target weighing data.

4. The method according to claim 1, characterized in that The short-time denoising network is implemented by a convolutional neural network, which includes a plurality of cascaded convolutional layers and pooling layers. The initial feature vector is input into the short-time denoising network in the target denoising model to extract local spatial features and perform short-time denoising based on the local spatial features to obtain local denoising features corresponding to the original weighing data, including: The initial feature vector is input into the convolutional neural network, and local spatial features are extracted and pooled through cascaded convolutional layers and pooling layers in sequence, and the result of the last level of pooling is used as the local denoising feature; wherein the input of each pooling layer is the output of the previous convolutional layer cascaded therewith.

5. The method according to claim 1, characterized in that Inputting the local denoising features into the long-term denoising network in the target denoising model, learning the long-term dependency between the local denoising features and performing long-term denoising processing based on the long-term dependency, and obtaining the global denoising features corresponding to the original weighing data, including: Inputting the local denoising features into the memory unit in the long-term denoising network, and learning the long-term dependency relationship between the local denoising features based on the gating mechanism; Inputting the long-term dependency and the local noise reduction feature into a denoising module in the long-term denoising network, and performing long-term denoising processing on the local noise reduction feature based on the long-term dependency to obtain the global noise reduction feature; The memory unit is used to store historical dependency information learned from historical local denoising features.

6. The method according to any one of claims 1 to 5, characterized in that: The method further comprises: Acquire a sample data set, the sample data set comprising: a plurality of sample data with sample noise and corresponding labeled data without sample noise, at least part of the plurality of sample data being obtained by learning the labeled data without sample noise using a generative adversarial network; Obtaining sample feature vectors corresponding to the plurality of sample data, inputting the sample feature vectors into a short-time denoising network in an initial denoising model to extract local spatial features and performing short-time denoising processing based on the local spatial features to obtain a first denoising feature; Inputting the first denoising features into a long-term denoising network in an initial denoising model, learning long-term dependencies between the first denoising features and performing long-term denoising processing based on the long-term dependencies to obtain second denoising features; Generate denoising result data corresponding to the sample data based on the second denoising feature; The loss function is calculated according to the labeled data and the denoising result data; and the model training is terminated when the loss function satisfies the corresponding termination condition to obtain the target denoising model.

7. The method according to any one of claim 6, characterized in that: Calculating a loss function according to the labeled data and the denoising result data includes: Calculating a global loss function based on the labeled data and the denoising result data; And calculating a local loss function according to the first denoising feature and the second denoising feature; When both the global loss function and the local loss function satisfy corresponding termination conditions, the model training is terminated to obtain the target denoising model.

8. A denoising model training method, characterized in that: include: Acquire a sample data set, the sample data set comprising a plurality of sample data with sample noise and corresponding annotated data without sample noise, the plurality of sample data having a timestamp to form sample time series data; Performing frequency domain conversion on the sample time series data to obtain sample time-frequency distribution information of the sample data, wherein the sample time-frequency distribution information is a two-dimensional representation of the sample data in the time domain and the frequency domain; Inputting the sample time-frequency distribution information into an encoder in an autoencoder, performing encoding and denoising processing on the sample time-frequency distribution information to obtain a sample feature vector; Inputting the sample feature vector into the short-time denoising network in the initial denoising model to extract local spatial features and perform short-time denoising processing based on the local spatial features to obtain a first denoising feature corresponding to the sample data, wherein the first denoising feature is used to describe the short-term variation law of the sample data after the short-time denoising processing; Inputting the first denoising features into a long-term denoising network in the initial denoising model, learning the long-term dependency relationship between the first denoising features and performing long-term denoising processing based on the long-term dependency relationship to obtain a second denoising feature corresponding to the sample data; Based on the second noise reduction feature corresponding to the sample data, generating denoising result data corresponding to the sample data, wherein the denoising result data removes noise data in the sample data; Calculating a loss function based on the sample data and the denoising result data; When the loss function satisfies the corresponding termination condition, the initial denoising model training is terminated to obtain a short-term denoising network and a long-term denoising network in the target denoising model.

9. An electronic device, characterized in that: include: Memory and processor; The memory is used to store a computer program, and the processor is coupled to the memory and is used to execute the computer program in the memory to implement the steps in the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program / instruction, characterized in that: When the computer program / instructions are executed by a processor, the processor is enabled to implement the steps in the method according to any one of claims 1 to 8.

Images